Merge tag 'riscv-for-linus-6.6-rc6' of git://git.kernel.org/pub/scm/linux/kernel...
[platform/kernel/linux-rpi.git] / mm / hugetlb.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Generic hugetlb support.
4  * (C) Nadia Yvette Chambers, April 2004
5  */
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
37 #include <linux/mm_inline.h>
38
39 #include <asm/page.h>
40 #include <asm/pgalloc.h>
41 #include <asm/tlb.h>
42
43 #include <linux/io.h>
44 #include <linux/hugetlb.h>
45 #include <linux/hugetlb_cgroup.h>
46 #include <linux/node.h>
47 #include <linux/page_owner.h>
48 #include "internal.h"
49 #include "hugetlb_vmemmap.h"
50
51 int hugetlb_max_hstate __read_mostly;
52 unsigned int default_hstate_idx;
53 struct hstate hstates[HUGE_MAX_HSTATE];
54
55 #ifdef CONFIG_CMA
56 static struct cma *hugetlb_cma[MAX_NUMNODES];
57 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
58 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
59 {
60         return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
61                                 1 << order);
62 }
63 #else
64 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
65 {
66         return false;
67 }
68 #endif
69 static unsigned long hugetlb_cma_size __initdata;
70
71 __initdata LIST_HEAD(huge_boot_pages);
72
73 /* for command line parsing */
74 static struct hstate * __initdata parsed_hstate;
75 static unsigned long __initdata default_hstate_max_huge_pages;
76 static bool __initdata parsed_valid_hugepagesz = true;
77 static bool __initdata parsed_default_hugepagesz;
78 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
79
80 /*
81  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
82  * free_huge_pages, and surplus_huge_pages.
83  */
84 DEFINE_SPINLOCK(hugetlb_lock);
85
86 /*
87  * Serializes faults on the same logical page.  This is used to
88  * prevent spurious OOMs when the hugepage pool is fully utilized.
89  */
90 static int num_fault_mutexes;
91 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
92
93 /* Forward declaration */
94 static int hugetlb_acct_memory(struct hstate *h, long delta);
95 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
96 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
97 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
98 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
99                 unsigned long start, unsigned long end);
100
101 static inline bool subpool_is_free(struct hugepage_subpool *spool)
102 {
103         if (spool->count)
104                 return false;
105         if (spool->max_hpages != -1)
106                 return spool->used_hpages == 0;
107         if (spool->min_hpages != -1)
108                 return spool->rsv_hpages == spool->min_hpages;
109
110         return true;
111 }
112
113 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
114                                                 unsigned long irq_flags)
115 {
116         spin_unlock_irqrestore(&spool->lock, irq_flags);
117
118         /* If no pages are used, and no other handles to the subpool
119          * remain, give up any reservations based on minimum size and
120          * free the subpool */
121         if (subpool_is_free(spool)) {
122                 if (spool->min_hpages != -1)
123                         hugetlb_acct_memory(spool->hstate,
124                                                 -spool->min_hpages);
125                 kfree(spool);
126         }
127 }
128
129 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
130                                                 long min_hpages)
131 {
132         struct hugepage_subpool *spool;
133
134         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
135         if (!spool)
136                 return NULL;
137
138         spin_lock_init(&spool->lock);
139         spool->count = 1;
140         spool->max_hpages = max_hpages;
141         spool->hstate = h;
142         spool->min_hpages = min_hpages;
143
144         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
145                 kfree(spool);
146                 return NULL;
147         }
148         spool->rsv_hpages = min_hpages;
149
150         return spool;
151 }
152
153 void hugepage_put_subpool(struct hugepage_subpool *spool)
154 {
155         unsigned long flags;
156
157         spin_lock_irqsave(&spool->lock, flags);
158         BUG_ON(!spool->count);
159         spool->count--;
160         unlock_or_release_subpool(spool, flags);
161 }
162
163 /*
164  * Subpool accounting for allocating and reserving pages.
165  * Return -ENOMEM if there are not enough resources to satisfy the
166  * request.  Otherwise, return the number of pages by which the
167  * global pools must be adjusted (upward).  The returned value may
168  * only be different than the passed value (delta) in the case where
169  * a subpool minimum size must be maintained.
170  */
171 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
172                                       long delta)
173 {
174         long ret = delta;
175
176         if (!spool)
177                 return ret;
178
179         spin_lock_irq(&spool->lock);
180
181         if (spool->max_hpages != -1) {          /* maximum size accounting */
182                 if ((spool->used_hpages + delta) <= spool->max_hpages)
183                         spool->used_hpages += delta;
184                 else {
185                         ret = -ENOMEM;
186                         goto unlock_ret;
187                 }
188         }
189
190         /* minimum size accounting */
191         if (spool->min_hpages != -1 && spool->rsv_hpages) {
192                 if (delta > spool->rsv_hpages) {
193                         /*
194                          * Asking for more reserves than those already taken on
195                          * behalf of subpool.  Return difference.
196                          */
197                         ret = delta - spool->rsv_hpages;
198                         spool->rsv_hpages = 0;
199                 } else {
200                         ret = 0;        /* reserves already accounted for */
201                         spool->rsv_hpages -= delta;
202                 }
203         }
204
205 unlock_ret:
206         spin_unlock_irq(&spool->lock);
207         return ret;
208 }
209
210 /*
211  * Subpool accounting for freeing and unreserving pages.
212  * Return the number of global page reservations that must be dropped.
213  * The return value may only be different than the passed value (delta)
214  * in the case where a subpool minimum size must be maintained.
215  */
216 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
217                                        long delta)
218 {
219         long ret = delta;
220         unsigned long flags;
221
222         if (!spool)
223                 return delta;
224
225         spin_lock_irqsave(&spool->lock, flags);
226
227         if (spool->max_hpages != -1)            /* maximum size accounting */
228                 spool->used_hpages -= delta;
229
230          /* minimum size accounting */
231         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
232                 if (spool->rsv_hpages + delta <= spool->min_hpages)
233                         ret = 0;
234                 else
235                         ret = spool->rsv_hpages + delta - spool->min_hpages;
236
237                 spool->rsv_hpages += delta;
238                 if (spool->rsv_hpages > spool->min_hpages)
239                         spool->rsv_hpages = spool->min_hpages;
240         }
241
242         /*
243          * If hugetlbfs_put_super couldn't free spool due to an outstanding
244          * quota reference, free it now.
245          */
246         unlock_or_release_subpool(spool, flags);
247
248         return ret;
249 }
250
251 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
252 {
253         return HUGETLBFS_SB(inode->i_sb)->spool;
254 }
255
256 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
257 {
258         return subpool_inode(file_inode(vma->vm_file));
259 }
260
261 /*
262  * hugetlb vma_lock helper routines
263  */
264 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
265 {
266         if (__vma_shareable_lock(vma)) {
267                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
268
269                 down_read(&vma_lock->rw_sema);
270         }
271 }
272
273 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
274 {
275         if (__vma_shareable_lock(vma)) {
276                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
277
278                 up_read(&vma_lock->rw_sema);
279         }
280 }
281
282 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
283 {
284         if (__vma_shareable_lock(vma)) {
285                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
286
287                 down_write(&vma_lock->rw_sema);
288         }
289 }
290
291 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
292 {
293         if (__vma_shareable_lock(vma)) {
294                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
295
296                 up_write(&vma_lock->rw_sema);
297         }
298 }
299
300 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
301 {
302         struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
303
304         if (!__vma_shareable_lock(vma))
305                 return 1;
306
307         return down_write_trylock(&vma_lock->rw_sema);
308 }
309
310 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
311 {
312         if (__vma_shareable_lock(vma)) {
313                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
314
315                 lockdep_assert_held(&vma_lock->rw_sema);
316         }
317 }
318
319 void hugetlb_vma_lock_release(struct kref *kref)
320 {
321         struct hugetlb_vma_lock *vma_lock = container_of(kref,
322                         struct hugetlb_vma_lock, refs);
323
324         kfree(vma_lock);
325 }
326
327 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
328 {
329         struct vm_area_struct *vma = vma_lock->vma;
330
331         /*
332          * vma_lock structure may or not be released as a result of put,
333          * it certainly will no longer be attached to vma so clear pointer.
334          * Semaphore synchronizes access to vma_lock->vma field.
335          */
336         vma_lock->vma = NULL;
337         vma->vm_private_data = NULL;
338         up_write(&vma_lock->rw_sema);
339         kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
340 }
341
342 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
343 {
344         if (__vma_shareable_lock(vma)) {
345                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
346
347                 __hugetlb_vma_unlock_write_put(vma_lock);
348         }
349 }
350
351 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
352 {
353         /*
354          * Only present in sharable vmas.
355          */
356         if (!vma || !__vma_shareable_lock(vma))
357                 return;
358
359         if (vma->vm_private_data) {
360                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
361
362                 down_write(&vma_lock->rw_sema);
363                 __hugetlb_vma_unlock_write_put(vma_lock);
364         }
365 }
366
367 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
368 {
369         struct hugetlb_vma_lock *vma_lock;
370
371         /* Only establish in (flags) sharable vmas */
372         if (!vma || !(vma->vm_flags & VM_MAYSHARE))
373                 return;
374
375         /* Should never get here with non-NULL vm_private_data */
376         if (vma->vm_private_data)
377                 return;
378
379         vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
380         if (!vma_lock) {
381                 /*
382                  * If we can not allocate structure, then vma can not
383                  * participate in pmd sharing.  This is only a possible
384                  * performance enhancement and memory saving issue.
385                  * However, the lock is also used to synchronize page
386                  * faults with truncation.  If the lock is not present,
387                  * unlikely races could leave pages in a file past i_size
388                  * until the file is removed.  Warn in the unlikely case of
389                  * allocation failure.
390                  */
391                 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
392                 return;
393         }
394
395         kref_init(&vma_lock->refs);
396         init_rwsem(&vma_lock->rw_sema);
397         vma_lock->vma = vma;
398         vma->vm_private_data = vma_lock;
399 }
400
401 /* Helper that removes a struct file_region from the resv_map cache and returns
402  * it for use.
403  */
404 static struct file_region *
405 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
406 {
407         struct file_region *nrg;
408
409         VM_BUG_ON(resv->region_cache_count <= 0);
410
411         resv->region_cache_count--;
412         nrg = list_first_entry(&resv->region_cache, struct file_region, link);
413         list_del(&nrg->link);
414
415         nrg->from = from;
416         nrg->to = to;
417
418         return nrg;
419 }
420
421 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
422                                               struct file_region *rg)
423 {
424 #ifdef CONFIG_CGROUP_HUGETLB
425         nrg->reservation_counter = rg->reservation_counter;
426         nrg->css = rg->css;
427         if (rg->css)
428                 css_get(rg->css);
429 #endif
430 }
431
432 /* Helper that records hugetlb_cgroup uncharge info. */
433 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
434                                                 struct hstate *h,
435                                                 struct resv_map *resv,
436                                                 struct file_region *nrg)
437 {
438 #ifdef CONFIG_CGROUP_HUGETLB
439         if (h_cg) {
440                 nrg->reservation_counter =
441                         &h_cg->rsvd_hugepage[hstate_index(h)];
442                 nrg->css = &h_cg->css;
443                 /*
444                  * The caller will hold exactly one h_cg->css reference for the
445                  * whole contiguous reservation region. But this area might be
446                  * scattered when there are already some file_regions reside in
447                  * it. As a result, many file_regions may share only one css
448                  * reference. In order to ensure that one file_region must hold
449                  * exactly one h_cg->css reference, we should do css_get for
450                  * each file_region and leave the reference held by caller
451                  * untouched.
452                  */
453                 css_get(&h_cg->css);
454                 if (!resv->pages_per_hpage)
455                         resv->pages_per_hpage = pages_per_huge_page(h);
456                 /* pages_per_hpage should be the same for all entries in
457                  * a resv_map.
458                  */
459                 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
460         } else {
461                 nrg->reservation_counter = NULL;
462                 nrg->css = NULL;
463         }
464 #endif
465 }
466
467 static void put_uncharge_info(struct file_region *rg)
468 {
469 #ifdef CONFIG_CGROUP_HUGETLB
470         if (rg->css)
471                 css_put(rg->css);
472 #endif
473 }
474
475 static bool has_same_uncharge_info(struct file_region *rg,
476                                    struct file_region *org)
477 {
478 #ifdef CONFIG_CGROUP_HUGETLB
479         return rg->reservation_counter == org->reservation_counter &&
480                rg->css == org->css;
481
482 #else
483         return true;
484 #endif
485 }
486
487 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
488 {
489         struct file_region *nrg, *prg;
490
491         prg = list_prev_entry(rg, link);
492         if (&prg->link != &resv->regions && prg->to == rg->from &&
493             has_same_uncharge_info(prg, rg)) {
494                 prg->to = rg->to;
495
496                 list_del(&rg->link);
497                 put_uncharge_info(rg);
498                 kfree(rg);
499
500                 rg = prg;
501         }
502
503         nrg = list_next_entry(rg, link);
504         if (&nrg->link != &resv->regions && nrg->from == rg->to &&
505             has_same_uncharge_info(nrg, rg)) {
506                 nrg->from = rg->from;
507
508                 list_del(&rg->link);
509                 put_uncharge_info(rg);
510                 kfree(rg);
511         }
512 }
513
514 static inline long
515 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
516                      long to, struct hstate *h, struct hugetlb_cgroup *cg,
517                      long *regions_needed)
518 {
519         struct file_region *nrg;
520
521         if (!regions_needed) {
522                 nrg = get_file_region_entry_from_cache(map, from, to);
523                 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
524                 list_add(&nrg->link, rg);
525                 coalesce_file_region(map, nrg);
526         } else
527                 *regions_needed += 1;
528
529         return to - from;
530 }
531
532 /*
533  * Must be called with resv->lock held.
534  *
535  * Calling this with regions_needed != NULL will count the number of pages
536  * to be added but will not modify the linked list. And regions_needed will
537  * indicate the number of file_regions needed in the cache to carry out to add
538  * the regions for this range.
539  */
540 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
541                                      struct hugetlb_cgroup *h_cg,
542                                      struct hstate *h, long *regions_needed)
543 {
544         long add = 0;
545         struct list_head *head = &resv->regions;
546         long last_accounted_offset = f;
547         struct file_region *iter, *trg = NULL;
548         struct list_head *rg = NULL;
549
550         if (regions_needed)
551                 *regions_needed = 0;
552
553         /* In this loop, we essentially handle an entry for the range
554          * [last_accounted_offset, iter->from), at every iteration, with some
555          * bounds checking.
556          */
557         list_for_each_entry_safe(iter, trg, head, link) {
558                 /* Skip irrelevant regions that start before our range. */
559                 if (iter->from < f) {
560                         /* If this region ends after the last accounted offset,
561                          * then we need to update last_accounted_offset.
562                          */
563                         if (iter->to > last_accounted_offset)
564                                 last_accounted_offset = iter->to;
565                         continue;
566                 }
567
568                 /* When we find a region that starts beyond our range, we've
569                  * finished.
570                  */
571                 if (iter->from >= t) {
572                         rg = iter->link.prev;
573                         break;
574                 }
575
576                 /* Add an entry for last_accounted_offset -> iter->from, and
577                  * update last_accounted_offset.
578                  */
579                 if (iter->from > last_accounted_offset)
580                         add += hugetlb_resv_map_add(resv, iter->link.prev,
581                                                     last_accounted_offset,
582                                                     iter->from, h, h_cg,
583                                                     regions_needed);
584
585                 last_accounted_offset = iter->to;
586         }
587
588         /* Handle the case where our range extends beyond
589          * last_accounted_offset.
590          */
591         if (!rg)
592                 rg = head->prev;
593         if (last_accounted_offset < t)
594                 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
595                                             t, h, h_cg, regions_needed);
596
597         return add;
598 }
599
600 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
601  */
602 static int allocate_file_region_entries(struct resv_map *resv,
603                                         int regions_needed)
604         __must_hold(&resv->lock)
605 {
606         LIST_HEAD(allocated_regions);
607         int to_allocate = 0, i = 0;
608         struct file_region *trg = NULL, *rg = NULL;
609
610         VM_BUG_ON(regions_needed < 0);
611
612         /*
613          * Check for sufficient descriptors in the cache to accommodate
614          * the number of in progress add operations plus regions_needed.
615          *
616          * This is a while loop because when we drop the lock, some other call
617          * to region_add or region_del may have consumed some region_entries,
618          * so we keep looping here until we finally have enough entries for
619          * (adds_in_progress + regions_needed).
620          */
621         while (resv->region_cache_count <
622                (resv->adds_in_progress + regions_needed)) {
623                 to_allocate = resv->adds_in_progress + regions_needed -
624                               resv->region_cache_count;
625
626                 /* At this point, we should have enough entries in the cache
627                  * for all the existing adds_in_progress. We should only be
628                  * needing to allocate for regions_needed.
629                  */
630                 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
631
632                 spin_unlock(&resv->lock);
633                 for (i = 0; i < to_allocate; i++) {
634                         trg = kmalloc(sizeof(*trg), GFP_KERNEL);
635                         if (!trg)
636                                 goto out_of_memory;
637                         list_add(&trg->link, &allocated_regions);
638                 }
639
640                 spin_lock(&resv->lock);
641
642                 list_splice(&allocated_regions, &resv->region_cache);
643                 resv->region_cache_count += to_allocate;
644         }
645
646         return 0;
647
648 out_of_memory:
649         list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
650                 list_del(&rg->link);
651                 kfree(rg);
652         }
653         return -ENOMEM;
654 }
655
656 /*
657  * Add the huge page range represented by [f, t) to the reserve
658  * map.  Regions will be taken from the cache to fill in this range.
659  * Sufficient regions should exist in the cache due to the previous
660  * call to region_chg with the same range, but in some cases the cache will not
661  * have sufficient entries due to races with other code doing region_add or
662  * region_del.  The extra needed entries will be allocated.
663  *
664  * regions_needed is the out value provided by a previous call to region_chg.
665  *
666  * Return the number of new huge pages added to the map.  This number is greater
667  * than or equal to zero.  If file_region entries needed to be allocated for
668  * this operation and we were not able to allocate, it returns -ENOMEM.
669  * region_add of regions of length 1 never allocate file_regions and cannot
670  * fail; region_chg will always allocate at least 1 entry and a region_add for
671  * 1 page will only require at most 1 entry.
672  */
673 static long region_add(struct resv_map *resv, long f, long t,
674                        long in_regions_needed, struct hstate *h,
675                        struct hugetlb_cgroup *h_cg)
676 {
677         long add = 0, actual_regions_needed = 0;
678
679         spin_lock(&resv->lock);
680 retry:
681
682         /* Count how many regions are actually needed to execute this add. */
683         add_reservation_in_range(resv, f, t, NULL, NULL,
684                                  &actual_regions_needed);
685
686         /*
687          * Check for sufficient descriptors in the cache to accommodate
688          * this add operation. Note that actual_regions_needed may be greater
689          * than in_regions_needed, as the resv_map may have been modified since
690          * the region_chg call. In this case, we need to make sure that we
691          * allocate extra entries, such that we have enough for all the
692          * existing adds_in_progress, plus the excess needed for this
693          * operation.
694          */
695         if (actual_regions_needed > in_regions_needed &&
696             resv->region_cache_count <
697                     resv->adds_in_progress +
698                             (actual_regions_needed - in_regions_needed)) {
699                 /* region_add operation of range 1 should never need to
700                  * allocate file_region entries.
701                  */
702                 VM_BUG_ON(t - f <= 1);
703
704                 if (allocate_file_region_entries(
705                             resv, actual_regions_needed - in_regions_needed)) {
706                         return -ENOMEM;
707                 }
708
709                 goto retry;
710         }
711
712         add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
713
714         resv->adds_in_progress -= in_regions_needed;
715
716         spin_unlock(&resv->lock);
717         return add;
718 }
719
720 /*
721  * Examine the existing reserve map and determine how many
722  * huge pages in the specified range [f, t) are NOT currently
723  * represented.  This routine is called before a subsequent
724  * call to region_add that will actually modify the reserve
725  * map to add the specified range [f, t).  region_chg does
726  * not change the number of huge pages represented by the
727  * map.  A number of new file_region structures is added to the cache as a
728  * placeholder, for the subsequent region_add call to use. At least 1
729  * file_region structure is added.
730  *
731  * out_regions_needed is the number of regions added to the
732  * resv->adds_in_progress.  This value needs to be provided to a follow up call
733  * to region_add or region_abort for proper accounting.
734  *
735  * Returns the number of huge pages that need to be added to the existing
736  * reservation map for the range [f, t).  This number is greater or equal to
737  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
738  * is needed and can not be allocated.
739  */
740 static long region_chg(struct resv_map *resv, long f, long t,
741                        long *out_regions_needed)
742 {
743         long chg = 0;
744
745         spin_lock(&resv->lock);
746
747         /* Count how many hugepages in this range are NOT represented. */
748         chg = add_reservation_in_range(resv, f, t, NULL, NULL,
749                                        out_regions_needed);
750
751         if (*out_regions_needed == 0)
752                 *out_regions_needed = 1;
753
754         if (allocate_file_region_entries(resv, *out_regions_needed))
755                 return -ENOMEM;
756
757         resv->adds_in_progress += *out_regions_needed;
758
759         spin_unlock(&resv->lock);
760         return chg;
761 }
762
763 /*
764  * Abort the in progress add operation.  The adds_in_progress field
765  * of the resv_map keeps track of the operations in progress between
766  * calls to region_chg and region_add.  Operations are sometimes
767  * aborted after the call to region_chg.  In such cases, region_abort
768  * is called to decrement the adds_in_progress counter. regions_needed
769  * is the value returned by the region_chg call, it is used to decrement
770  * the adds_in_progress counter.
771  *
772  * NOTE: The range arguments [f, t) are not needed or used in this
773  * routine.  They are kept to make reading the calling code easier as
774  * arguments will match the associated region_chg call.
775  */
776 static void region_abort(struct resv_map *resv, long f, long t,
777                          long regions_needed)
778 {
779         spin_lock(&resv->lock);
780         VM_BUG_ON(!resv->region_cache_count);
781         resv->adds_in_progress -= regions_needed;
782         spin_unlock(&resv->lock);
783 }
784
785 /*
786  * Delete the specified range [f, t) from the reserve map.  If the
787  * t parameter is LONG_MAX, this indicates that ALL regions after f
788  * should be deleted.  Locate the regions which intersect [f, t)
789  * and either trim, delete or split the existing regions.
790  *
791  * Returns the number of huge pages deleted from the reserve map.
792  * In the normal case, the return value is zero or more.  In the
793  * case where a region must be split, a new region descriptor must
794  * be allocated.  If the allocation fails, -ENOMEM will be returned.
795  * NOTE: If the parameter t == LONG_MAX, then we will never split
796  * a region and possibly return -ENOMEM.  Callers specifying
797  * t == LONG_MAX do not need to check for -ENOMEM error.
798  */
799 static long region_del(struct resv_map *resv, long f, long t)
800 {
801         struct list_head *head = &resv->regions;
802         struct file_region *rg, *trg;
803         struct file_region *nrg = NULL;
804         long del = 0;
805
806 retry:
807         spin_lock(&resv->lock);
808         list_for_each_entry_safe(rg, trg, head, link) {
809                 /*
810                  * Skip regions before the range to be deleted.  file_region
811                  * ranges are normally of the form [from, to).  However, there
812                  * may be a "placeholder" entry in the map which is of the form
813                  * (from, to) with from == to.  Check for placeholder entries
814                  * at the beginning of the range to be deleted.
815                  */
816                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
817                         continue;
818
819                 if (rg->from >= t)
820                         break;
821
822                 if (f > rg->from && t < rg->to) { /* Must split region */
823                         /*
824                          * Check for an entry in the cache before dropping
825                          * lock and attempting allocation.
826                          */
827                         if (!nrg &&
828                             resv->region_cache_count > resv->adds_in_progress) {
829                                 nrg = list_first_entry(&resv->region_cache,
830                                                         struct file_region,
831                                                         link);
832                                 list_del(&nrg->link);
833                                 resv->region_cache_count--;
834                         }
835
836                         if (!nrg) {
837                                 spin_unlock(&resv->lock);
838                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
839                                 if (!nrg)
840                                         return -ENOMEM;
841                                 goto retry;
842                         }
843
844                         del += t - f;
845                         hugetlb_cgroup_uncharge_file_region(
846                                 resv, rg, t - f, false);
847
848                         /* New entry for end of split region */
849                         nrg->from = t;
850                         nrg->to = rg->to;
851
852                         copy_hugetlb_cgroup_uncharge_info(nrg, rg);
853
854                         INIT_LIST_HEAD(&nrg->link);
855
856                         /* Original entry is trimmed */
857                         rg->to = f;
858
859                         list_add(&nrg->link, &rg->link);
860                         nrg = NULL;
861                         break;
862                 }
863
864                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
865                         del += rg->to - rg->from;
866                         hugetlb_cgroup_uncharge_file_region(resv, rg,
867                                                             rg->to - rg->from, true);
868                         list_del(&rg->link);
869                         kfree(rg);
870                         continue;
871                 }
872
873                 if (f <= rg->from) {    /* Trim beginning of region */
874                         hugetlb_cgroup_uncharge_file_region(resv, rg,
875                                                             t - rg->from, false);
876
877                         del += t - rg->from;
878                         rg->from = t;
879                 } else {                /* Trim end of region */
880                         hugetlb_cgroup_uncharge_file_region(resv, rg,
881                                                             rg->to - f, false);
882
883                         del += rg->to - f;
884                         rg->to = f;
885                 }
886         }
887
888         spin_unlock(&resv->lock);
889         kfree(nrg);
890         return del;
891 }
892
893 /*
894  * A rare out of memory error was encountered which prevented removal of
895  * the reserve map region for a page.  The huge page itself was free'ed
896  * and removed from the page cache.  This routine will adjust the subpool
897  * usage count, and the global reserve count if needed.  By incrementing
898  * these counts, the reserve map entry which could not be deleted will
899  * appear as a "reserved" entry instead of simply dangling with incorrect
900  * counts.
901  */
902 void hugetlb_fix_reserve_counts(struct inode *inode)
903 {
904         struct hugepage_subpool *spool = subpool_inode(inode);
905         long rsv_adjust;
906         bool reserved = false;
907
908         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
909         if (rsv_adjust > 0) {
910                 struct hstate *h = hstate_inode(inode);
911
912                 if (!hugetlb_acct_memory(h, 1))
913                         reserved = true;
914         } else if (!rsv_adjust) {
915                 reserved = true;
916         }
917
918         if (!reserved)
919                 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
920 }
921
922 /*
923  * Count and return the number of huge pages in the reserve map
924  * that intersect with the range [f, t).
925  */
926 static long region_count(struct resv_map *resv, long f, long t)
927 {
928         struct list_head *head = &resv->regions;
929         struct file_region *rg;
930         long chg = 0;
931
932         spin_lock(&resv->lock);
933         /* Locate each segment we overlap with, and count that overlap. */
934         list_for_each_entry(rg, head, link) {
935                 long seg_from;
936                 long seg_to;
937
938                 if (rg->to <= f)
939                         continue;
940                 if (rg->from >= t)
941                         break;
942
943                 seg_from = max(rg->from, f);
944                 seg_to = min(rg->to, t);
945
946                 chg += seg_to - seg_from;
947         }
948         spin_unlock(&resv->lock);
949
950         return chg;
951 }
952
953 /*
954  * Convert the address within this vma to the page offset within
955  * the mapping, in pagecache page units; huge pages here.
956  */
957 static pgoff_t vma_hugecache_offset(struct hstate *h,
958                         struct vm_area_struct *vma, unsigned long address)
959 {
960         return ((address - vma->vm_start) >> huge_page_shift(h)) +
961                         (vma->vm_pgoff >> huge_page_order(h));
962 }
963
964 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
965                                      unsigned long address)
966 {
967         return vma_hugecache_offset(hstate_vma(vma), vma, address);
968 }
969 EXPORT_SYMBOL_GPL(linear_hugepage_index);
970
971 /**
972  * vma_kernel_pagesize - Page size granularity for this VMA.
973  * @vma: The user mapping.
974  *
975  * Folios in this VMA will be aligned to, and at least the size of the
976  * number of bytes returned by this function.
977  *
978  * Return: The default size of the folios allocated when backing a VMA.
979  */
980 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
981 {
982         if (vma->vm_ops && vma->vm_ops->pagesize)
983                 return vma->vm_ops->pagesize(vma);
984         return PAGE_SIZE;
985 }
986 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
987
988 /*
989  * Return the page size being used by the MMU to back a VMA. In the majority
990  * of cases, the page size used by the kernel matches the MMU size. On
991  * architectures where it differs, an architecture-specific 'strong'
992  * version of this symbol is required.
993  */
994 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
995 {
996         return vma_kernel_pagesize(vma);
997 }
998
999 /*
1000  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
1001  * bits of the reservation map pointer, which are always clear due to
1002  * alignment.
1003  */
1004 #define HPAGE_RESV_OWNER    (1UL << 0)
1005 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1006 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1007
1008 /*
1009  * These helpers are used to track how many pages are reserved for
1010  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1011  * is guaranteed to have their future faults succeed.
1012  *
1013  * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1014  * the reserve counters are updated with the hugetlb_lock held. It is safe
1015  * to reset the VMA at fork() time as it is not in use yet and there is no
1016  * chance of the global counters getting corrupted as a result of the values.
1017  *
1018  * The private mapping reservation is represented in a subtly different
1019  * manner to a shared mapping.  A shared mapping has a region map associated
1020  * with the underlying file, this region map represents the backing file
1021  * pages which have ever had a reservation assigned which this persists even
1022  * after the page is instantiated.  A private mapping has a region map
1023  * associated with the original mmap which is attached to all VMAs which
1024  * reference it, this region map represents those offsets which have consumed
1025  * reservation ie. where pages have been instantiated.
1026  */
1027 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1028 {
1029         return (unsigned long)vma->vm_private_data;
1030 }
1031
1032 static void set_vma_private_data(struct vm_area_struct *vma,
1033                                                         unsigned long value)
1034 {
1035         vma->vm_private_data = (void *)value;
1036 }
1037
1038 static void
1039 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1040                                           struct hugetlb_cgroup *h_cg,
1041                                           struct hstate *h)
1042 {
1043 #ifdef CONFIG_CGROUP_HUGETLB
1044         if (!h_cg || !h) {
1045                 resv_map->reservation_counter = NULL;
1046                 resv_map->pages_per_hpage = 0;
1047                 resv_map->css = NULL;
1048         } else {
1049                 resv_map->reservation_counter =
1050                         &h_cg->rsvd_hugepage[hstate_index(h)];
1051                 resv_map->pages_per_hpage = pages_per_huge_page(h);
1052                 resv_map->css = &h_cg->css;
1053         }
1054 #endif
1055 }
1056
1057 struct resv_map *resv_map_alloc(void)
1058 {
1059         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1060         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1061
1062         if (!resv_map || !rg) {
1063                 kfree(resv_map);
1064                 kfree(rg);
1065                 return NULL;
1066         }
1067
1068         kref_init(&resv_map->refs);
1069         spin_lock_init(&resv_map->lock);
1070         INIT_LIST_HEAD(&resv_map->regions);
1071
1072         resv_map->adds_in_progress = 0;
1073         /*
1074          * Initialize these to 0. On shared mappings, 0's here indicate these
1075          * fields don't do cgroup accounting. On private mappings, these will be
1076          * re-initialized to the proper values, to indicate that hugetlb cgroup
1077          * reservations are to be un-charged from here.
1078          */
1079         resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1080
1081         INIT_LIST_HEAD(&resv_map->region_cache);
1082         list_add(&rg->link, &resv_map->region_cache);
1083         resv_map->region_cache_count = 1;
1084
1085         return resv_map;
1086 }
1087
1088 void resv_map_release(struct kref *ref)
1089 {
1090         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1091         struct list_head *head = &resv_map->region_cache;
1092         struct file_region *rg, *trg;
1093
1094         /* Clear out any active regions before we release the map. */
1095         region_del(resv_map, 0, LONG_MAX);
1096
1097         /* ... and any entries left in the cache */
1098         list_for_each_entry_safe(rg, trg, head, link) {
1099                 list_del(&rg->link);
1100                 kfree(rg);
1101         }
1102
1103         VM_BUG_ON(resv_map->adds_in_progress);
1104
1105         kfree(resv_map);
1106 }
1107
1108 static inline struct resv_map *inode_resv_map(struct inode *inode)
1109 {
1110         /*
1111          * At inode evict time, i_mapping may not point to the original
1112          * address space within the inode.  This original address space
1113          * contains the pointer to the resv_map.  So, always use the
1114          * address space embedded within the inode.
1115          * The VERY common case is inode->mapping == &inode->i_data but,
1116          * this may not be true for device special inodes.
1117          */
1118         return (struct resv_map *)(&inode->i_data)->private_data;
1119 }
1120
1121 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1122 {
1123         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1124         if (vma->vm_flags & VM_MAYSHARE) {
1125                 struct address_space *mapping = vma->vm_file->f_mapping;
1126                 struct inode *inode = mapping->host;
1127
1128                 return inode_resv_map(inode);
1129
1130         } else {
1131                 return (struct resv_map *)(get_vma_private_data(vma) &
1132                                                         ~HPAGE_RESV_MASK);
1133         }
1134 }
1135
1136 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1137 {
1138         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1139         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1140
1141         set_vma_private_data(vma, (get_vma_private_data(vma) &
1142                                 HPAGE_RESV_MASK) | (unsigned long)map);
1143 }
1144
1145 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1146 {
1147         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1148         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1149
1150         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1151 }
1152
1153 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1154 {
1155         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1156
1157         return (get_vma_private_data(vma) & flag) != 0;
1158 }
1159
1160 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1161 {
1162         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1163         /*
1164          * Clear vm_private_data
1165          * - For shared mappings this is a per-vma semaphore that may be
1166          *   allocated in a subsequent call to hugetlb_vm_op_open.
1167          *   Before clearing, make sure pointer is not associated with vma
1168          *   as this will leak the structure.  This is the case when called
1169          *   via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1170          *   been called to allocate a new structure.
1171          * - For MAP_PRIVATE mappings, this is the reserve map which does
1172          *   not apply to children.  Faults generated by the children are
1173          *   not guaranteed to succeed, even if read-only.
1174          */
1175         if (vma->vm_flags & VM_MAYSHARE) {
1176                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1177
1178                 if (vma_lock && vma_lock->vma != vma)
1179                         vma->vm_private_data = NULL;
1180         } else
1181                 vma->vm_private_data = NULL;
1182 }
1183
1184 /*
1185  * Reset and decrement one ref on hugepage private reservation.
1186  * Called with mm->mmap_lock writer semaphore held.
1187  * This function should be only used by move_vma() and operate on
1188  * same sized vma. It should never come here with last ref on the
1189  * reservation.
1190  */
1191 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1192 {
1193         /*
1194          * Clear the old hugetlb private page reservation.
1195          * It has already been transferred to new_vma.
1196          *
1197          * During a mremap() operation of a hugetlb vma we call move_vma()
1198          * which copies vma into new_vma and unmaps vma. After the copy
1199          * operation both new_vma and vma share a reference to the resv_map
1200          * struct, and at that point vma is about to be unmapped. We don't
1201          * want to return the reservation to the pool at unmap of vma because
1202          * the reservation still lives on in new_vma, so simply decrement the
1203          * ref here and remove the resv_map reference from this vma.
1204          */
1205         struct resv_map *reservations = vma_resv_map(vma);
1206
1207         if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1208                 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1209                 kref_put(&reservations->refs, resv_map_release);
1210         }
1211
1212         hugetlb_dup_vma_private(vma);
1213 }
1214
1215 /* Returns true if the VMA has associated reserve pages */
1216 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1217 {
1218         if (vma->vm_flags & VM_NORESERVE) {
1219                 /*
1220                  * This address is already reserved by other process(chg == 0),
1221                  * so, we should decrement reserved count. Without decrementing,
1222                  * reserve count remains after releasing inode, because this
1223                  * allocated page will go into page cache and is regarded as
1224                  * coming from reserved pool in releasing step.  Currently, we
1225                  * don't have any other solution to deal with this situation
1226                  * properly, so add work-around here.
1227                  */
1228                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1229                         return true;
1230                 else
1231                         return false;
1232         }
1233
1234         /* Shared mappings always use reserves */
1235         if (vma->vm_flags & VM_MAYSHARE) {
1236                 /*
1237                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
1238                  * be a region map for all pages.  The only situation where
1239                  * there is no region map is if a hole was punched via
1240                  * fallocate.  In this case, there really are no reserves to
1241                  * use.  This situation is indicated if chg != 0.
1242                  */
1243                 if (chg)
1244                         return false;
1245                 else
1246                         return true;
1247         }
1248
1249         /*
1250          * Only the process that called mmap() has reserves for
1251          * private mappings.
1252          */
1253         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1254                 /*
1255                  * Like the shared case above, a hole punch or truncate
1256                  * could have been performed on the private mapping.
1257                  * Examine the value of chg to determine if reserves
1258                  * actually exist or were previously consumed.
1259                  * Very Subtle - The value of chg comes from a previous
1260                  * call to vma_needs_reserves().  The reserve map for
1261                  * private mappings has different (opposite) semantics
1262                  * than that of shared mappings.  vma_needs_reserves()
1263                  * has already taken this difference in semantics into
1264                  * account.  Therefore, the meaning of chg is the same
1265                  * as in the shared case above.  Code could easily be
1266                  * combined, but keeping it separate draws attention to
1267                  * subtle differences.
1268                  */
1269                 if (chg)
1270                         return false;
1271                 else
1272                         return true;
1273         }
1274
1275         return false;
1276 }
1277
1278 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1279 {
1280         int nid = folio_nid(folio);
1281
1282         lockdep_assert_held(&hugetlb_lock);
1283         VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1284
1285         list_move(&folio->lru, &h->hugepage_freelists[nid]);
1286         h->free_huge_pages++;
1287         h->free_huge_pages_node[nid]++;
1288         folio_set_hugetlb_freed(folio);
1289 }
1290
1291 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1292                                                                 int nid)
1293 {
1294         struct folio *folio;
1295         bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1296
1297         lockdep_assert_held(&hugetlb_lock);
1298         list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1299                 if (pin && !folio_is_longterm_pinnable(folio))
1300                         continue;
1301
1302                 if (folio_test_hwpoison(folio))
1303                         continue;
1304
1305                 list_move(&folio->lru, &h->hugepage_activelist);
1306                 folio_ref_unfreeze(folio, 1);
1307                 folio_clear_hugetlb_freed(folio);
1308                 h->free_huge_pages--;
1309                 h->free_huge_pages_node[nid]--;
1310                 return folio;
1311         }
1312
1313         return NULL;
1314 }
1315
1316 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1317                                                         int nid, nodemask_t *nmask)
1318 {
1319         unsigned int cpuset_mems_cookie;
1320         struct zonelist *zonelist;
1321         struct zone *zone;
1322         struct zoneref *z;
1323         int node = NUMA_NO_NODE;
1324
1325         zonelist = node_zonelist(nid, gfp_mask);
1326
1327 retry_cpuset:
1328         cpuset_mems_cookie = read_mems_allowed_begin();
1329         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1330                 struct folio *folio;
1331
1332                 if (!cpuset_zone_allowed(zone, gfp_mask))
1333                         continue;
1334                 /*
1335                  * no need to ask again on the same node. Pool is node rather than
1336                  * zone aware
1337                  */
1338                 if (zone_to_nid(zone) == node)
1339                         continue;
1340                 node = zone_to_nid(zone);
1341
1342                 folio = dequeue_hugetlb_folio_node_exact(h, node);
1343                 if (folio)
1344                         return folio;
1345         }
1346         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1347                 goto retry_cpuset;
1348
1349         return NULL;
1350 }
1351
1352 static unsigned long available_huge_pages(struct hstate *h)
1353 {
1354         return h->free_huge_pages - h->resv_huge_pages;
1355 }
1356
1357 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1358                                 struct vm_area_struct *vma,
1359                                 unsigned long address, int avoid_reserve,
1360                                 long chg)
1361 {
1362         struct folio *folio = NULL;
1363         struct mempolicy *mpol;
1364         gfp_t gfp_mask;
1365         nodemask_t *nodemask;
1366         int nid;
1367
1368         /*
1369          * A child process with MAP_PRIVATE mappings created by their parent
1370          * have no page reserves. This check ensures that reservations are
1371          * not "stolen". The child may still get SIGKILLed
1372          */
1373         if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1374                 goto err;
1375
1376         /* If reserves cannot be used, ensure enough pages are in the pool */
1377         if (avoid_reserve && !available_huge_pages(h))
1378                 goto err;
1379
1380         gfp_mask = htlb_alloc_mask(h);
1381         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1382
1383         if (mpol_is_preferred_many(mpol)) {
1384                 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1385                                                         nid, nodemask);
1386
1387                 /* Fallback to all nodes if page==NULL */
1388                 nodemask = NULL;
1389         }
1390
1391         if (!folio)
1392                 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1393                                                         nid, nodemask);
1394
1395         if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1396                 folio_set_hugetlb_restore_reserve(folio);
1397                 h->resv_huge_pages--;
1398         }
1399
1400         mpol_cond_put(mpol);
1401         return folio;
1402
1403 err:
1404         return NULL;
1405 }
1406
1407 /*
1408  * common helper functions for hstate_next_node_to_{alloc|free}.
1409  * We may have allocated or freed a huge page based on a different
1410  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1411  * be outside of *nodes_allowed.  Ensure that we use an allowed
1412  * node for alloc or free.
1413  */
1414 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1415 {
1416         nid = next_node_in(nid, *nodes_allowed);
1417         VM_BUG_ON(nid >= MAX_NUMNODES);
1418
1419         return nid;
1420 }
1421
1422 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1423 {
1424         if (!node_isset(nid, *nodes_allowed))
1425                 nid = next_node_allowed(nid, nodes_allowed);
1426         return nid;
1427 }
1428
1429 /*
1430  * returns the previously saved node ["this node"] from which to
1431  * allocate a persistent huge page for the pool and advance the
1432  * next node from which to allocate, handling wrap at end of node
1433  * mask.
1434  */
1435 static int hstate_next_node_to_alloc(struct hstate *h,
1436                                         nodemask_t *nodes_allowed)
1437 {
1438         int nid;
1439
1440         VM_BUG_ON(!nodes_allowed);
1441
1442         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1443         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1444
1445         return nid;
1446 }
1447
1448 /*
1449  * helper for remove_pool_huge_page() - return the previously saved
1450  * node ["this node"] from which to free a huge page.  Advance the
1451  * next node id whether or not we find a free huge page to free so
1452  * that the next attempt to free addresses the next node.
1453  */
1454 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1455 {
1456         int nid;
1457
1458         VM_BUG_ON(!nodes_allowed);
1459
1460         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1461         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1462
1463         return nid;
1464 }
1465
1466 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1467         for (nr_nodes = nodes_weight(*mask);                            \
1468                 nr_nodes > 0 &&                                         \
1469                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1470                 nr_nodes--)
1471
1472 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1473         for (nr_nodes = nodes_weight(*mask);                            \
1474                 nr_nodes > 0 &&                                         \
1475                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1476                 nr_nodes--)
1477
1478 /* used to demote non-gigantic_huge pages as well */
1479 static void __destroy_compound_gigantic_folio(struct folio *folio,
1480                                         unsigned int order, bool demote)
1481 {
1482         int i;
1483         int nr_pages = 1 << order;
1484         struct page *p;
1485
1486         atomic_set(&folio->_entire_mapcount, 0);
1487         atomic_set(&folio->_nr_pages_mapped, 0);
1488         atomic_set(&folio->_pincount, 0);
1489
1490         for (i = 1; i < nr_pages; i++) {
1491                 p = folio_page(folio, i);
1492                 p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1493                 p->mapping = NULL;
1494                 clear_compound_head(p);
1495                 if (!demote)
1496                         set_page_refcounted(p);
1497         }
1498
1499         __folio_clear_head(folio);
1500 }
1501
1502 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1503                                         unsigned int order)
1504 {
1505         __destroy_compound_gigantic_folio(folio, order, true);
1506 }
1507
1508 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1509 static void destroy_compound_gigantic_folio(struct folio *folio,
1510                                         unsigned int order)
1511 {
1512         __destroy_compound_gigantic_folio(folio, order, false);
1513 }
1514
1515 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1516 {
1517         /*
1518          * If the page isn't allocated using the cma allocator,
1519          * cma_release() returns false.
1520          */
1521 #ifdef CONFIG_CMA
1522         int nid = folio_nid(folio);
1523
1524         if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1525                 return;
1526 #endif
1527
1528         free_contig_range(folio_pfn(folio), 1 << order);
1529 }
1530
1531 #ifdef CONFIG_CONTIG_ALLOC
1532 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1533                 int nid, nodemask_t *nodemask)
1534 {
1535         struct page *page;
1536         unsigned long nr_pages = pages_per_huge_page(h);
1537         if (nid == NUMA_NO_NODE)
1538                 nid = numa_mem_id();
1539
1540 #ifdef CONFIG_CMA
1541         {
1542                 int node;
1543
1544                 if (hugetlb_cma[nid]) {
1545                         page = cma_alloc(hugetlb_cma[nid], nr_pages,
1546                                         huge_page_order(h), true);
1547                         if (page)
1548                                 return page_folio(page);
1549                 }
1550
1551                 if (!(gfp_mask & __GFP_THISNODE)) {
1552                         for_each_node_mask(node, *nodemask) {
1553                                 if (node == nid || !hugetlb_cma[node])
1554                                         continue;
1555
1556                                 page = cma_alloc(hugetlb_cma[node], nr_pages,
1557                                                 huge_page_order(h), true);
1558                                 if (page)
1559                                         return page_folio(page);
1560                         }
1561                 }
1562         }
1563 #endif
1564
1565         page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1566         return page ? page_folio(page) : NULL;
1567 }
1568
1569 #else /* !CONFIG_CONTIG_ALLOC */
1570 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1571                                         int nid, nodemask_t *nodemask)
1572 {
1573         return NULL;
1574 }
1575 #endif /* CONFIG_CONTIG_ALLOC */
1576
1577 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1578 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1579                                         int nid, nodemask_t *nodemask)
1580 {
1581         return NULL;
1582 }
1583 static inline void free_gigantic_folio(struct folio *folio,
1584                                                 unsigned int order) { }
1585 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1586                                                 unsigned int order) { }
1587 #endif
1588
1589 static inline void __clear_hugetlb_destructor(struct hstate *h,
1590                                                 struct folio *folio)
1591 {
1592         lockdep_assert_held(&hugetlb_lock);
1593
1594         folio_clear_hugetlb(folio);
1595 }
1596
1597 /*
1598  * Remove hugetlb folio from lists.
1599  * If vmemmap exists for the folio, update dtor so that the folio appears
1600  * as just a compound page.  Otherwise, wait until after allocating vmemmap
1601  * to update dtor.
1602  *
1603  * A reference is held on the folio, except in the case of demote.
1604  *
1605  * Must be called with hugetlb lock held.
1606  */
1607 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1608                                                         bool adjust_surplus,
1609                                                         bool demote)
1610 {
1611         int nid = folio_nid(folio);
1612
1613         VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1614         VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1615
1616         lockdep_assert_held(&hugetlb_lock);
1617         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1618                 return;
1619
1620         list_del(&folio->lru);
1621
1622         if (folio_test_hugetlb_freed(folio)) {
1623                 h->free_huge_pages--;
1624                 h->free_huge_pages_node[nid]--;
1625         }
1626         if (adjust_surplus) {
1627                 h->surplus_huge_pages--;
1628                 h->surplus_huge_pages_node[nid]--;
1629         }
1630
1631         /*
1632          * We can only clear the hugetlb destructor after allocating vmemmap
1633          * pages.  Otherwise, someone (memory error handling) may try to write
1634          * to tail struct pages.
1635          */
1636         if (!folio_test_hugetlb_vmemmap_optimized(folio))
1637                 __clear_hugetlb_destructor(h, folio);
1638
1639          /*
1640           * In the case of demote we do not ref count the page as it will soon
1641           * be turned into a page of smaller size.
1642          */
1643         if (!demote)
1644                 folio_ref_unfreeze(folio, 1);
1645
1646         h->nr_huge_pages--;
1647         h->nr_huge_pages_node[nid]--;
1648 }
1649
1650 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1651                                                         bool adjust_surplus)
1652 {
1653         __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1654 }
1655
1656 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1657                                                         bool adjust_surplus)
1658 {
1659         __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1660 }
1661
1662 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1663                              bool adjust_surplus)
1664 {
1665         int zeroed;
1666         int nid = folio_nid(folio);
1667
1668         VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1669
1670         lockdep_assert_held(&hugetlb_lock);
1671
1672         INIT_LIST_HEAD(&folio->lru);
1673         h->nr_huge_pages++;
1674         h->nr_huge_pages_node[nid]++;
1675
1676         if (adjust_surplus) {
1677                 h->surplus_huge_pages++;
1678                 h->surplus_huge_pages_node[nid]++;
1679         }
1680
1681         folio_set_hugetlb(folio);
1682         folio_change_private(folio, NULL);
1683         /*
1684          * We have to set hugetlb_vmemmap_optimized again as above
1685          * folio_change_private(folio, NULL) cleared it.
1686          */
1687         folio_set_hugetlb_vmemmap_optimized(folio);
1688
1689         /*
1690          * This folio is about to be managed by the hugetlb allocator and
1691          * should have no users.  Drop our reference, and check for others
1692          * just in case.
1693          */
1694         zeroed = folio_put_testzero(folio);
1695         if (unlikely(!zeroed))
1696                 /*
1697                  * It is VERY unlikely soneone else has taken a ref
1698                  * on the folio.  In this case, we simply return as
1699                  * free_huge_folio() will be called when this other ref
1700                  * is dropped.
1701                  */
1702                 return;
1703
1704         arch_clear_hugepage_flags(&folio->page);
1705         enqueue_hugetlb_folio(h, folio);
1706 }
1707
1708 static void __update_and_free_hugetlb_folio(struct hstate *h,
1709                                                 struct folio *folio)
1710 {
1711         bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio);
1712
1713         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1714                 return;
1715
1716         /*
1717          * If we don't know which subpages are hwpoisoned, we can't free
1718          * the hugepage, so it's leaked intentionally.
1719          */
1720         if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1721                 return;
1722
1723         if (hugetlb_vmemmap_restore(h, &folio->page)) {
1724                 spin_lock_irq(&hugetlb_lock);
1725                 /*
1726                  * If we cannot allocate vmemmap pages, just refuse to free the
1727                  * page and put the page back on the hugetlb free list and treat
1728                  * as a surplus page.
1729                  */
1730                 add_hugetlb_folio(h, folio, true);
1731                 spin_unlock_irq(&hugetlb_lock);
1732                 return;
1733         }
1734
1735         /*
1736          * Move PageHWPoison flag from head page to the raw error pages,
1737          * which makes any healthy subpages reusable.
1738          */
1739         if (unlikely(folio_test_hwpoison(folio)))
1740                 folio_clear_hugetlb_hwpoison(folio);
1741
1742         /*
1743          * If vmemmap pages were allocated above, then we need to clear the
1744          * hugetlb destructor under the hugetlb lock.
1745          */
1746         if (clear_dtor) {
1747                 spin_lock_irq(&hugetlb_lock);
1748                 __clear_hugetlb_destructor(h, folio);
1749                 spin_unlock_irq(&hugetlb_lock);
1750         }
1751
1752         /*
1753          * Non-gigantic pages demoted from CMA allocated gigantic pages
1754          * need to be given back to CMA in free_gigantic_folio.
1755          */
1756         if (hstate_is_gigantic(h) ||
1757             hugetlb_cma_folio(folio, huge_page_order(h))) {
1758                 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1759                 free_gigantic_folio(folio, huge_page_order(h));
1760         } else {
1761                 __free_pages(&folio->page, huge_page_order(h));
1762         }
1763 }
1764
1765 /*
1766  * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1767  * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1768  * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1769  * the vmemmap pages.
1770  *
1771  * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1772  * freed and frees them one-by-one. As the page->mapping pointer is going
1773  * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1774  * structure of a lockless linked list of huge pages to be freed.
1775  */
1776 static LLIST_HEAD(hpage_freelist);
1777
1778 static void free_hpage_workfn(struct work_struct *work)
1779 {
1780         struct llist_node *node;
1781
1782         node = llist_del_all(&hpage_freelist);
1783
1784         while (node) {
1785                 struct page *page;
1786                 struct hstate *h;
1787
1788                 page = container_of((struct address_space **)node,
1789                                      struct page, mapping);
1790                 node = node->next;
1791                 page->mapping = NULL;
1792                 /*
1793                  * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1794                  * folio_hstate() is going to trigger because a previous call to
1795                  * remove_hugetlb_folio() will clear the hugetlb bit, so do
1796                  * not use folio_hstate() directly.
1797                  */
1798                 h = size_to_hstate(page_size(page));
1799
1800                 __update_and_free_hugetlb_folio(h, page_folio(page));
1801
1802                 cond_resched();
1803         }
1804 }
1805 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1806
1807 static inline void flush_free_hpage_work(struct hstate *h)
1808 {
1809         if (hugetlb_vmemmap_optimizable(h))
1810                 flush_work(&free_hpage_work);
1811 }
1812
1813 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1814                                  bool atomic)
1815 {
1816         if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1817                 __update_and_free_hugetlb_folio(h, folio);
1818                 return;
1819         }
1820
1821         /*
1822          * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1823          *
1824          * Only call schedule_work() if hpage_freelist is previously
1825          * empty. Otherwise, schedule_work() had been called but the workfn
1826          * hasn't retrieved the list yet.
1827          */
1828         if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1829                 schedule_work(&free_hpage_work);
1830 }
1831
1832 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1833 {
1834         struct page *page, *t_page;
1835         struct folio *folio;
1836
1837         list_for_each_entry_safe(page, t_page, list, lru) {
1838                 folio = page_folio(page);
1839                 update_and_free_hugetlb_folio(h, folio, false);
1840                 cond_resched();
1841         }
1842 }
1843
1844 struct hstate *size_to_hstate(unsigned long size)
1845 {
1846         struct hstate *h;
1847
1848         for_each_hstate(h) {
1849                 if (huge_page_size(h) == size)
1850                         return h;
1851         }
1852         return NULL;
1853 }
1854
1855 void free_huge_folio(struct folio *folio)
1856 {
1857         /*
1858          * Can't pass hstate in here because it is called from the
1859          * compound page destructor.
1860          */
1861         struct hstate *h = folio_hstate(folio);
1862         int nid = folio_nid(folio);
1863         struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1864         bool restore_reserve;
1865         unsigned long flags;
1866
1867         VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1868         VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1869
1870         hugetlb_set_folio_subpool(folio, NULL);
1871         if (folio_test_anon(folio))
1872                 __ClearPageAnonExclusive(&folio->page);
1873         folio->mapping = NULL;
1874         restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1875         folio_clear_hugetlb_restore_reserve(folio);
1876
1877         /*
1878          * If HPageRestoreReserve was set on page, page allocation consumed a
1879          * reservation.  If the page was associated with a subpool, there
1880          * would have been a page reserved in the subpool before allocation
1881          * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1882          * reservation, do not call hugepage_subpool_put_pages() as this will
1883          * remove the reserved page from the subpool.
1884          */
1885         if (!restore_reserve) {
1886                 /*
1887                  * A return code of zero implies that the subpool will be
1888                  * under its minimum size if the reservation is not restored
1889                  * after page is free.  Therefore, force restore_reserve
1890                  * operation.
1891                  */
1892                 if (hugepage_subpool_put_pages(spool, 1) == 0)
1893                         restore_reserve = true;
1894         }
1895
1896         spin_lock_irqsave(&hugetlb_lock, flags);
1897         folio_clear_hugetlb_migratable(folio);
1898         hugetlb_cgroup_uncharge_folio(hstate_index(h),
1899                                      pages_per_huge_page(h), folio);
1900         hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1901                                           pages_per_huge_page(h), folio);
1902         if (restore_reserve)
1903                 h->resv_huge_pages++;
1904
1905         if (folio_test_hugetlb_temporary(folio)) {
1906                 remove_hugetlb_folio(h, folio, false);
1907                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1908                 update_and_free_hugetlb_folio(h, folio, true);
1909         } else if (h->surplus_huge_pages_node[nid]) {
1910                 /* remove the page from active list */
1911                 remove_hugetlb_folio(h, folio, true);
1912                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1913                 update_and_free_hugetlb_folio(h, folio, true);
1914         } else {
1915                 arch_clear_hugepage_flags(&folio->page);
1916                 enqueue_hugetlb_folio(h, folio);
1917                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1918         }
1919 }
1920
1921 /*
1922  * Must be called with the hugetlb lock held
1923  */
1924 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1925 {
1926         lockdep_assert_held(&hugetlb_lock);
1927         h->nr_huge_pages++;
1928         h->nr_huge_pages_node[nid]++;
1929 }
1930
1931 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1932 {
1933         hugetlb_vmemmap_optimize(h, &folio->page);
1934         INIT_LIST_HEAD(&folio->lru);
1935         folio_set_hugetlb(folio);
1936         hugetlb_set_folio_subpool(folio, NULL);
1937         set_hugetlb_cgroup(folio, NULL);
1938         set_hugetlb_cgroup_rsvd(folio, NULL);
1939 }
1940
1941 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1942 {
1943         __prep_new_hugetlb_folio(h, folio);
1944         spin_lock_irq(&hugetlb_lock);
1945         __prep_account_new_huge_page(h, nid);
1946         spin_unlock_irq(&hugetlb_lock);
1947 }
1948
1949 static bool __prep_compound_gigantic_folio(struct folio *folio,
1950                                         unsigned int order, bool demote)
1951 {
1952         int i, j;
1953         int nr_pages = 1 << order;
1954         struct page *p;
1955
1956         __folio_clear_reserved(folio);
1957         for (i = 0; i < nr_pages; i++) {
1958                 p = folio_page(folio, i);
1959
1960                 /*
1961                  * For gigantic hugepages allocated through bootmem at
1962                  * boot, it's safer to be consistent with the not-gigantic
1963                  * hugepages and clear the PG_reserved bit from all tail pages
1964                  * too.  Otherwise drivers using get_user_pages() to access tail
1965                  * pages may get the reference counting wrong if they see
1966                  * PG_reserved set on a tail page (despite the head page not
1967                  * having PG_reserved set).  Enforcing this consistency between
1968                  * head and tail pages allows drivers to optimize away a check
1969                  * on the head page when they need know if put_page() is needed
1970                  * after get_user_pages().
1971                  */
1972                 if (i != 0)     /* head page cleared above */
1973                         __ClearPageReserved(p);
1974                 /*
1975                  * Subtle and very unlikely
1976                  *
1977                  * Gigantic 'page allocators' such as memblock or cma will
1978                  * return a set of pages with each page ref counted.  We need
1979                  * to turn this set of pages into a compound page with tail
1980                  * page ref counts set to zero.  Code such as speculative page
1981                  * cache adding could take a ref on a 'to be' tail page.
1982                  * We need to respect any increased ref count, and only set
1983                  * the ref count to zero if count is currently 1.  If count
1984                  * is not 1, we return an error.  An error return indicates
1985                  * the set of pages can not be converted to a gigantic page.
1986                  * The caller who allocated the pages should then discard the
1987                  * pages using the appropriate free interface.
1988                  *
1989                  * In the case of demote, the ref count will be zero.
1990                  */
1991                 if (!demote) {
1992                         if (!page_ref_freeze(p, 1)) {
1993                                 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1994                                 goto out_error;
1995                         }
1996                 } else {
1997                         VM_BUG_ON_PAGE(page_count(p), p);
1998                 }
1999                 if (i != 0)
2000                         set_compound_head(p, &folio->page);
2001         }
2002         __folio_set_head(folio);
2003         /* we rely on prep_new_hugetlb_folio to set the destructor */
2004         folio_set_order(folio, order);
2005         atomic_set(&folio->_entire_mapcount, -1);
2006         atomic_set(&folio->_nr_pages_mapped, 0);
2007         atomic_set(&folio->_pincount, 0);
2008         return true;
2009
2010 out_error:
2011         /* undo page modifications made above */
2012         for (j = 0; j < i; j++) {
2013                 p = folio_page(folio, j);
2014                 if (j != 0)
2015                         clear_compound_head(p);
2016                 set_page_refcounted(p);
2017         }
2018         /* need to clear PG_reserved on remaining tail pages  */
2019         for (; j < nr_pages; j++) {
2020                 p = folio_page(folio, j);
2021                 __ClearPageReserved(p);
2022         }
2023         return false;
2024 }
2025
2026 static bool prep_compound_gigantic_folio(struct folio *folio,
2027                                                         unsigned int order)
2028 {
2029         return __prep_compound_gigantic_folio(folio, order, false);
2030 }
2031
2032 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2033                                                         unsigned int order)
2034 {
2035         return __prep_compound_gigantic_folio(folio, order, true);
2036 }
2037
2038 /*
2039  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2040  * transparent huge pages.  See the PageTransHuge() documentation for more
2041  * details.
2042  */
2043 int PageHuge(struct page *page)
2044 {
2045         struct folio *folio;
2046
2047         if (!PageCompound(page))
2048                 return 0;
2049         folio = page_folio(page);
2050         return folio_test_hugetlb(folio);
2051 }
2052 EXPORT_SYMBOL_GPL(PageHuge);
2053
2054 /*
2055  * Find and lock address space (mapping) in write mode.
2056  *
2057  * Upon entry, the page is locked which means that page_mapping() is
2058  * stable.  Due to locking order, we can only trylock_write.  If we can
2059  * not get the lock, simply return NULL to caller.
2060  */
2061 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2062 {
2063         struct address_space *mapping = page_mapping(hpage);
2064
2065         if (!mapping)
2066                 return mapping;
2067
2068         if (i_mmap_trylock_write(mapping))
2069                 return mapping;
2070
2071         return NULL;
2072 }
2073
2074 pgoff_t hugetlb_basepage_index(struct page *page)
2075 {
2076         struct page *page_head = compound_head(page);
2077         pgoff_t index = page_index(page_head);
2078         unsigned long compound_idx;
2079
2080         if (compound_order(page_head) > MAX_ORDER)
2081                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2082         else
2083                 compound_idx = page - page_head;
2084
2085         return (index << compound_order(page_head)) + compound_idx;
2086 }
2087
2088 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2089                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2090                 nodemask_t *node_alloc_noretry)
2091 {
2092         int order = huge_page_order(h);
2093         struct page *page;
2094         bool alloc_try_hard = true;
2095         bool retry = true;
2096
2097         /*
2098          * By default we always try hard to allocate the page with
2099          * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
2100          * a loop (to adjust global huge page counts) and previous allocation
2101          * failed, do not continue to try hard on the same node.  Use the
2102          * node_alloc_noretry bitmap to manage this state information.
2103          */
2104         if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2105                 alloc_try_hard = false;
2106         gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2107         if (alloc_try_hard)
2108                 gfp_mask |= __GFP_RETRY_MAYFAIL;
2109         if (nid == NUMA_NO_NODE)
2110                 nid = numa_mem_id();
2111 retry:
2112         page = __alloc_pages(gfp_mask, order, nid, nmask);
2113
2114         /* Freeze head page */
2115         if (page && !page_ref_freeze(page, 1)) {
2116                 __free_pages(page, order);
2117                 if (retry) {    /* retry once */
2118                         retry = false;
2119                         goto retry;
2120                 }
2121                 /* WOW!  twice in a row. */
2122                 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2123                 page = NULL;
2124         }
2125
2126         /*
2127          * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2128          * indicates an overall state change.  Clear bit so that we resume
2129          * normal 'try hard' allocations.
2130          */
2131         if (node_alloc_noretry && page && !alloc_try_hard)
2132                 node_clear(nid, *node_alloc_noretry);
2133
2134         /*
2135          * If we tried hard to get a page but failed, set bit so that
2136          * subsequent attempts will not try as hard until there is an
2137          * overall state change.
2138          */
2139         if (node_alloc_noretry && !page && alloc_try_hard)
2140                 node_set(nid, *node_alloc_noretry);
2141
2142         if (!page) {
2143                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2144                 return NULL;
2145         }
2146
2147         __count_vm_event(HTLB_BUDDY_PGALLOC);
2148         return page_folio(page);
2149 }
2150
2151 /*
2152  * Common helper to allocate a fresh hugetlb page. All specific allocators
2153  * should use this function to get new hugetlb pages
2154  *
2155  * Note that returned page is 'frozen':  ref count of head page and all tail
2156  * pages is zero.
2157  */
2158 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2159                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2160                 nodemask_t *node_alloc_noretry)
2161 {
2162         struct folio *folio;
2163         bool retry = false;
2164
2165 retry:
2166         if (hstate_is_gigantic(h))
2167                 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2168         else
2169                 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2170                                 nid, nmask, node_alloc_noretry);
2171         if (!folio)
2172                 return NULL;
2173         if (hstate_is_gigantic(h)) {
2174                 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2175                         /*
2176                          * Rare failure to convert pages to compound page.
2177                          * Free pages and try again - ONCE!
2178                          */
2179                         free_gigantic_folio(folio, huge_page_order(h));
2180                         if (!retry) {
2181                                 retry = true;
2182                                 goto retry;
2183                         }
2184                         return NULL;
2185                 }
2186         }
2187         prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2188
2189         return folio;
2190 }
2191
2192 /*
2193  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2194  * manner.
2195  */
2196 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2197                                 nodemask_t *node_alloc_noretry)
2198 {
2199         struct folio *folio;
2200         int nr_nodes, node;
2201         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2202
2203         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2204                 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2205                                         nodes_allowed, node_alloc_noretry);
2206                 if (folio) {
2207                         free_huge_folio(folio); /* free it into the hugepage allocator */
2208                         return 1;
2209                 }
2210         }
2211
2212         return 0;
2213 }
2214
2215 /*
2216  * Remove huge page from pool from next node to free.  Attempt to keep
2217  * persistent huge pages more or less balanced over allowed nodes.
2218  * This routine only 'removes' the hugetlb page.  The caller must make
2219  * an additional call to free the page to low level allocators.
2220  * Called with hugetlb_lock locked.
2221  */
2222 static struct page *remove_pool_huge_page(struct hstate *h,
2223                                                 nodemask_t *nodes_allowed,
2224                                                  bool acct_surplus)
2225 {
2226         int nr_nodes, node;
2227         struct page *page = NULL;
2228         struct folio *folio;
2229
2230         lockdep_assert_held(&hugetlb_lock);
2231         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2232                 /*
2233                  * If we're returning unused surplus pages, only examine
2234                  * nodes with surplus pages.
2235                  */
2236                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2237                     !list_empty(&h->hugepage_freelists[node])) {
2238                         page = list_entry(h->hugepage_freelists[node].next,
2239                                           struct page, lru);
2240                         folio = page_folio(page);
2241                         remove_hugetlb_folio(h, folio, acct_surplus);
2242                         break;
2243                 }
2244         }
2245
2246         return page;
2247 }
2248
2249 /*
2250  * Dissolve a given free hugepage into free buddy pages. This function does
2251  * nothing for in-use hugepages and non-hugepages.
2252  * This function returns values like below:
2253  *
2254  *  -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2255  *           when the system is under memory pressure and the feature of
2256  *           freeing unused vmemmap pages associated with each hugetlb page
2257  *           is enabled.
2258  *  -EBUSY:  failed to dissolved free hugepages or the hugepage is in-use
2259  *           (allocated or reserved.)
2260  *       0:  successfully dissolved free hugepages or the page is not a
2261  *           hugepage (considered as already dissolved)
2262  */
2263 int dissolve_free_huge_page(struct page *page)
2264 {
2265         int rc = -EBUSY;
2266         struct folio *folio = page_folio(page);
2267
2268 retry:
2269         /* Not to disrupt normal path by vainly holding hugetlb_lock */
2270         if (!folio_test_hugetlb(folio))
2271                 return 0;
2272
2273         spin_lock_irq(&hugetlb_lock);
2274         if (!folio_test_hugetlb(folio)) {
2275                 rc = 0;
2276                 goto out;
2277         }
2278
2279         if (!folio_ref_count(folio)) {
2280                 struct hstate *h = folio_hstate(folio);
2281                 if (!available_huge_pages(h))
2282                         goto out;
2283
2284                 /*
2285                  * We should make sure that the page is already on the free list
2286                  * when it is dissolved.
2287                  */
2288                 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2289                         spin_unlock_irq(&hugetlb_lock);
2290                         cond_resched();
2291
2292                         /*
2293                          * Theoretically, we should return -EBUSY when we
2294                          * encounter this race. In fact, we have a chance
2295                          * to successfully dissolve the page if we do a
2296                          * retry. Because the race window is quite small.
2297                          * If we seize this opportunity, it is an optimization
2298                          * for increasing the success rate of dissolving page.
2299                          */
2300                         goto retry;
2301                 }
2302
2303                 remove_hugetlb_folio(h, folio, false);
2304                 h->max_huge_pages--;
2305                 spin_unlock_irq(&hugetlb_lock);
2306
2307                 /*
2308                  * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2309                  * before freeing the page.  update_and_free_hugtlb_folio will fail to
2310                  * free the page if it can not allocate required vmemmap.  We
2311                  * need to adjust max_huge_pages if the page is not freed.
2312                  * Attempt to allocate vmemmmap here so that we can take
2313                  * appropriate action on failure.
2314                  */
2315                 rc = hugetlb_vmemmap_restore(h, &folio->page);
2316                 if (!rc) {
2317                         update_and_free_hugetlb_folio(h, folio, false);
2318                 } else {
2319                         spin_lock_irq(&hugetlb_lock);
2320                         add_hugetlb_folio(h, folio, false);
2321                         h->max_huge_pages++;
2322                         spin_unlock_irq(&hugetlb_lock);
2323                 }
2324
2325                 return rc;
2326         }
2327 out:
2328         spin_unlock_irq(&hugetlb_lock);
2329         return rc;
2330 }
2331
2332 /*
2333  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2334  * make specified memory blocks removable from the system.
2335  * Note that this will dissolve a free gigantic hugepage completely, if any
2336  * part of it lies within the given range.
2337  * Also note that if dissolve_free_huge_page() returns with an error, all
2338  * free hugepages that were dissolved before that error are lost.
2339  */
2340 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2341 {
2342         unsigned long pfn;
2343         struct page *page;
2344         int rc = 0;
2345         unsigned int order;
2346         struct hstate *h;
2347
2348         if (!hugepages_supported())
2349                 return rc;
2350
2351         order = huge_page_order(&default_hstate);
2352         for_each_hstate(h)
2353                 order = min(order, huge_page_order(h));
2354
2355         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2356                 page = pfn_to_page(pfn);
2357                 rc = dissolve_free_huge_page(page);
2358                 if (rc)
2359                         break;
2360         }
2361
2362         return rc;
2363 }
2364
2365 /*
2366  * Allocates a fresh surplus page from the page allocator.
2367  */
2368 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2369                                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2370 {
2371         struct folio *folio = NULL;
2372
2373         if (hstate_is_gigantic(h))
2374                 return NULL;
2375
2376         spin_lock_irq(&hugetlb_lock);
2377         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2378                 goto out_unlock;
2379         spin_unlock_irq(&hugetlb_lock);
2380
2381         folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2382         if (!folio)
2383                 return NULL;
2384
2385         spin_lock_irq(&hugetlb_lock);
2386         /*
2387          * We could have raced with the pool size change.
2388          * Double check that and simply deallocate the new page
2389          * if we would end up overcommiting the surpluses. Abuse
2390          * temporary page to workaround the nasty free_huge_folio
2391          * codeflow
2392          */
2393         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2394                 folio_set_hugetlb_temporary(folio);
2395                 spin_unlock_irq(&hugetlb_lock);
2396                 free_huge_folio(folio);
2397                 return NULL;
2398         }
2399
2400         h->surplus_huge_pages++;
2401         h->surplus_huge_pages_node[folio_nid(folio)]++;
2402
2403 out_unlock:
2404         spin_unlock_irq(&hugetlb_lock);
2405
2406         return folio;
2407 }
2408
2409 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2410                                      int nid, nodemask_t *nmask)
2411 {
2412         struct folio *folio;
2413
2414         if (hstate_is_gigantic(h))
2415                 return NULL;
2416
2417         folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2418         if (!folio)
2419                 return NULL;
2420
2421         /* fresh huge pages are frozen */
2422         folio_ref_unfreeze(folio, 1);
2423         /*
2424          * We do not account these pages as surplus because they are only
2425          * temporary and will be released properly on the last reference
2426          */
2427         folio_set_hugetlb_temporary(folio);
2428
2429         return folio;
2430 }
2431
2432 /*
2433  * Use the VMA's mpolicy to allocate a huge page from the buddy.
2434  */
2435 static
2436 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2437                 struct vm_area_struct *vma, unsigned long addr)
2438 {
2439         struct folio *folio = NULL;
2440         struct mempolicy *mpol;
2441         gfp_t gfp_mask = htlb_alloc_mask(h);
2442         int nid;
2443         nodemask_t *nodemask;
2444
2445         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2446         if (mpol_is_preferred_many(mpol)) {
2447                 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2448
2449                 gfp &=  ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2450                 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2451
2452                 /* Fallback to all nodes if page==NULL */
2453                 nodemask = NULL;
2454         }
2455
2456         if (!folio)
2457                 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2458         mpol_cond_put(mpol);
2459         return folio;
2460 }
2461
2462 /* folio migration callback function */
2463 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2464                 nodemask_t *nmask, gfp_t gfp_mask)
2465 {
2466         spin_lock_irq(&hugetlb_lock);
2467         if (available_huge_pages(h)) {
2468                 struct folio *folio;
2469
2470                 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2471                                                 preferred_nid, nmask);
2472                 if (folio) {
2473                         spin_unlock_irq(&hugetlb_lock);
2474                         return folio;
2475                 }
2476         }
2477         spin_unlock_irq(&hugetlb_lock);
2478
2479         return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2480 }
2481
2482 /* mempolicy aware migration callback */
2483 struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma,
2484                 unsigned long address)
2485 {
2486         struct mempolicy *mpol;
2487         nodemask_t *nodemask;
2488         struct folio *folio;
2489         gfp_t gfp_mask;
2490         int node;
2491
2492         gfp_mask = htlb_alloc_mask(h);
2493         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2494         folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
2495         mpol_cond_put(mpol);
2496
2497         return folio;
2498 }
2499
2500 /*
2501  * Increase the hugetlb pool such that it can accommodate a reservation
2502  * of size 'delta'.
2503  */
2504 static int gather_surplus_pages(struct hstate *h, long delta)
2505         __must_hold(&hugetlb_lock)
2506 {
2507         LIST_HEAD(surplus_list);
2508         struct folio *folio, *tmp;
2509         int ret;
2510         long i;
2511         long needed, allocated;
2512         bool alloc_ok = true;
2513
2514         lockdep_assert_held(&hugetlb_lock);
2515         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2516         if (needed <= 0) {
2517                 h->resv_huge_pages += delta;
2518                 return 0;
2519         }
2520
2521         allocated = 0;
2522
2523         ret = -ENOMEM;
2524 retry:
2525         spin_unlock_irq(&hugetlb_lock);
2526         for (i = 0; i < needed; i++) {
2527                 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2528                                 NUMA_NO_NODE, NULL);
2529                 if (!folio) {
2530                         alloc_ok = false;
2531                         break;
2532                 }
2533                 list_add(&folio->lru, &surplus_list);
2534                 cond_resched();
2535         }
2536         allocated += i;
2537
2538         /*
2539          * After retaking hugetlb_lock, we need to recalculate 'needed'
2540          * because either resv_huge_pages or free_huge_pages may have changed.
2541          */
2542         spin_lock_irq(&hugetlb_lock);
2543         needed = (h->resv_huge_pages + delta) -
2544                         (h->free_huge_pages + allocated);
2545         if (needed > 0) {
2546                 if (alloc_ok)
2547                         goto retry;
2548                 /*
2549                  * We were not able to allocate enough pages to
2550                  * satisfy the entire reservation so we free what
2551                  * we've allocated so far.
2552                  */
2553                 goto free;
2554         }
2555         /*
2556          * The surplus_list now contains _at_least_ the number of extra pages
2557          * needed to accommodate the reservation.  Add the appropriate number
2558          * of pages to the hugetlb pool and free the extras back to the buddy
2559          * allocator.  Commit the entire reservation here to prevent another
2560          * process from stealing the pages as they are added to the pool but
2561          * before they are reserved.
2562          */
2563         needed += allocated;
2564         h->resv_huge_pages += delta;
2565         ret = 0;
2566
2567         /* Free the needed pages to the hugetlb pool */
2568         list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2569                 if ((--needed) < 0)
2570                         break;
2571                 /* Add the page to the hugetlb allocator */
2572                 enqueue_hugetlb_folio(h, folio);
2573         }
2574 free:
2575         spin_unlock_irq(&hugetlb_lock);
2576
2577         /*
2578          * Free unnecessary surplus pages to the buddy allocator.
2579          * Pages have no ref count, call free_huge_folio directly.
2580          */
2581         list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2582                 free_huge_folio(folio);
2583         spin_lock_irq(&hugetlb_lock);
2584
2585         return ret;
2586 }
2587
2588 /*
2589  * This routine has two main purposes:
2590  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2591  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2592  *    to the associated reservation map.
2593  * 2) Free any unused surplus pages that may have been allocated to satisfy
2594  *    the reservation.  As many as unused_resv_pages may be freed.
2595  */
2596 static void return_unused_surplus_pages(struct hstate *h,
2597                                         unsigned long unused_resv_pages)
2598 {
2599         unsigned long nr_pages;
2600         struct page *page;
2601         LIST_HEAD(page_list);
2602
2603         lockdep_assert_held(&hugetlb_lock);
2604         /* Uncommit the reservation */
2605         h->resv_huge_pages -= unused_resv_pages;
2606
2607         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2608                 goto out;
2609
2610         /*
2611          * Part (or even all) of the reservation could have been backed
2612          * by pre-allocated pages. Only free surplus pages.
2613          */
2614         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2615
2616         /*
2617          * We want to release as many surplus pages as possible, spread
2618          * evenly across all nodes with memory. Iterate across these nodes
2619          * until we can no longer free unreserved surplus pages. This occurs
2620          * when the nodes with surplus pages have no free pages.
2621          * remove_pool_huge_page() will balance the freed pages across the
2622          * on-line nodes with memory and will handle the hstate accounting.
2623          */
2624         while (nr_pages--) {
2625                 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2626                 if (!page)
2627                         goto out;
2628
2629                 list_add(&page->lru, &page_list);
2630         }
2631
2632 out:
2633         spin_unlock_irq(&hugetlb_lock);
2634         update_and_free_pages_bulk(h, &page_list);
2635         spin_lock_irq(&hugetlb_lock);
2636 }
2637
2638
2639 /*
2640  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2641  * are used by the huge page allocation routines to manage reservations.
2642  *
2643  * vma_needs_reservation is called to determine if the huge page at addr
2644  * within the vma has an associated reservation.  If a reservation is
2645  * needed, the value 1 is returned.  The caller is then responsible for
2646  * managing the global reservation and subpool usage counts.  After
2647  * the huge page has been allocated, vma_commit_reservation is called
2648  * to add the page to the reservation map.  If the page allocation fails,
2649  * the reservation must be ended instead of committed.  vma_end_reservation
2650  * is called in such cases.
2651  *
2652  * In the normal case, vma_commit_reservation returns the same value
2653  * as the preceding vma_needs_reservation call.  The only time this
2654  * is not the case is if a reserve map was changed between calls.  It
2655  * is the responsibility of the caller to notice the difference and
2656  * take appropriate action.
2657  *
2658  * vma_add_reservation is used in error paths where a reservation must
2659  * be restored when a newly allocated huge page must be freed.  It is
2660  * to be called after calling vma_needs_reservation to determine if a
2661  * reservation exists.
2662  *
2663  * vma_del_reservation is used in error paths where an entry in the reserve
2664  * map was created during huge page allocation and must be removed.  It is to
2665  * be called after calling vma_needs_reservation to determine if a reservation
2666  * exists.
2667  */
2668 enum vma_resv_mode {
2669         VMA_NEEDS_RESV,
2670         VMA_COMMIT_RESV,
2671         VMA_END_RESV,
2672         VMA_ADD_RESV,
2673         VMA_DEL_RESV,
2674 };
2675 static long __vma_reservation_common(struct hstate *h,
2676                                 struct vm_area_struct *vma, unsigned long addr,
2677                                 enum vma_resv_mode mode)
2678 {
2679         struct resv_map *resv;
2680         pgoff_t idx;
2681         long ret;
2682         long dummy_out_regions_needed;
2683
2684         resv = vma_resv_map(vma);
2685         if (!resv)
2686                 return 1;
2687
2688         idx = vma_hugecache_offset(h, vma, addr);
2689         switch (mode) {
2690         case VMA_NEEDS_RESV:
2691                 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2692                 /* We assume that vma_reservation_* routines always operate on
2693                  * 1 page, and that adding to resv map a 1 page entry can only
2694                  * ever require 1 region.
2695                  */
2696                 VM_BUG_ON(dummy_out_regions_needed != 1);
2697                 break;
2698         case VMA_COMMIT_RESV:
2699                 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2700                 /* region_add calls of range 1 should never fail. */
2701                 VM_BUG_ON(ret < 0);
2702                 break;
2703         case VMA_END_RESV:
2704                 region_abort(resv, idx, idx + 1, 1);
2705                 ret = 0;
2706                 break;
2707         case VMA_ADD_RESV:
2708                 if (vma->vm_flags & VM_MAYSHARE) {
2709                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2710                         /* region_add calls of range 1 should never fail. */
2711                         VM_BUG_ON(ret < 0);
2712                 } else {
2713                         region_abort(resv, idx, idx + 1, 1);
2714                         ret = region_del(resv, idx, idx + 1);
2715                 }
2716                 break;
2717         case VMA_DEL_RESV:
2718                 if (vma->vm_flags & VM_MAYSHARE) {
2719                         region_abort(resv, idx, idx + 1, 1);
2720                         ret = region_del(resv, idx, idx + 1);
2721                 } else {
2722                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2723                         /* region_add calls of range 1 should never fail. */
2724                         VM_BUG_ON(ret < 0);
2725                 }
2726                 break;
2727         default:
2728                 BUG();
2729         }
2730
2731         if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2732                 return ret;
2733         /*
2734          * We know private mapping must have HPAGE_RESV_OWNER set.
2735          *
2736          * In most cases, reserves always exist for private mappings.
2737          * However, a file associated with mapping could have been
2738          * hole punched or truncated after reserves were consumed.
2739          * As subsequent fault on such a range will not use reserves.
2740          * Subtle - The reserve map for private mappings has the
2741          * opposite meaning than that of shared mappings.  If NO
2742          * entry is in the reserve map, it means a reservation exists.
2743          * If an entry exists in the reserve map, it means the
2744          * reservation has already been consumed.  As a result, the
2745          * return value of this routine is the opposite of the
2746          * value returned from reserve map manipulation routines above.
2747          */
2748         if (ret > 0)
2749                 return 0;
2750         if (ret == 0)
2751                 return 1;
2752         return ret;
2753 }
2754
2755 static long vma_needs_reservation(struct hstate *h,
2756                         struct vm_area_struct *vma, unsigned long addr)
2757 {
2758         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2759 }
2760
2761 static long vma_commit_reservation(struct hstate *h,
2762                         struct vm_area_struct *vma, unsigned long addr)
2763 {
2764         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2765 }
2766
2767 static void vma_end_reservation(struct hstate *h,
2768                         struct vm_area_struct *vma, unsigned long addr)
2769 {
2770         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2771 }
2772
2773 static long vma_add_reservation(struct hstate *h,
2774                         struct vm_area_struct *vma, unsigned long addr)
2775 {
2776         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2777 }
2778
2779 static long vma_del_reservation(struct hstate *h,
2780                         struct vm_area_struct *vma, unsigned long addr)
2781 {
2782         return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2783 }
2784
2785 /*
2786  * This routine is called to restore reservation information on error paths.
2787  * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2788  * and the hugetlb mutex should remain held when calling this routine.
2789  *
2790  * It handles two specific cases:
2791  * 1) A reservation was in place and the folio consumed the reservation.
2792  *    hugetlb_restore_reserve is set in the folio.
2793  * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2794  *    not set.  However, alloc_hugetlb_folio always updates the reserve map.
2795  *
2796  * In case 1, free_huge_folio later in the error path will increment the
2797  * global reserve count.  But, free_huge_folio does not have enough context
2798  * to adjust the reservation map.  This case deals primarily with private
2799  * mappings.  Adjust the reserve map here to be consistent with global
2800  * reserve count adjustments to be made by free_huge_folio.  Make sure the
2801  * reserve map indicates there is a reservation present.
2802  *
2803  * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2804  */
2805 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2806                         unsigned long address, struct folio *folio)
2807 {
2808         long rc = vma_needs_reservation(h, vma, address);
2809
2810         if (folio_test_hugetlb_restore_reserve(folio)) {
2811                 if (unlikely(rc < 0))
2812                         /*
2813                          * Rare out of memory condition in reserve map
2814                          * manipulation.  Clear hugetlb_restore_reserve so
2815                          * that global reserve count will not be incremented
2816                          * by free_huge_folio.  This will make it appear
2817                          * as though the reservation for this folio was
2818                          * consumed.  This may prevent the task from
2819                          * faulting in the folio at a later time.  This
2820                          * is better than inconsistent global huge page
2821                          * accounting of reserve counts.
2822                          */
2823                         folio_clear_hugetlb_restore_reserve(folio);
2824                 else if (rc)
2825                         (void)vma_add_reservation(h, vma, address);
2826                 else
2827                         vma_end_reservation(h, vma, address);
2828         } else {
2829                 if (!rc) {
2830                         /*
2831                          * This indicates there is an entry in the reserve map
2832                          * not added by alloc_hugetlb_folio.  We know it was added
2833                          * before the alloc_hugetlb_folio call, otherwise
2834                          * hugetlb_restore_reserve would be set on the folio.
2835                          * Remove the entry so that a subsequent allocation
2836                          * does not consume a reservation.
2837                          */
2838                         rc = vma_del_reservation(h, vma, address);
2839                         if (rc < 0)
2840                                 /*
2841                                  * VERY rare out of memory condition.  Since
2842                                  * we can not delete the entry, set
2843                                  * hugetlb_restore_reserve so that the reserve
2844                                  * count will be incremented when the folio
2845                                  * is freed.  This reserve will be consumed
2846                                  * on a subsequent allocation.
2847                                  */
2848                                 folio_set_hugetlb_restore_reserve(folio);
2849                 } else if (rc < 0) {
2850                         /*
2851                          * Rare out of memory condition from
2852                          * vma_needs_reservation call.  Memory allocation is
2853                          * only attempted if a new entry is needed.  Therefore,
2854                          * this implies there is not an entry in the
2855                          * reserve map.
2856                          *
2857                          * For shared mappings, no entry in the map indicates
2858                          * no reservation.  We are done.
2859                          */
2860                         if (!(vma->vm_flags & VM_MAYSHARE))
2861                                 /*
2862                                  * For private mappings, no entry indicates
2863                                  * a reservation is present.  Since we can
2864                                  * not add an entry, set hugetlb_restore_reserve
2865                                  * on the folio so reserve count will be
2866                                  * incremented when freed.  This reserve will
2867                                  * be consumed on a subsequent allocation.
2868                                  */
2869                                 folio_set_hugetlb_restore_reserve(folio);
2870                 } else
2871                         /*
2872                          * No reservation present, do nothing
2873                          */
2874                          vma_end_reservation(h, vma, address);
2875         }
2876 }
2877
2878 /*
2879  * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2880  * the old one
2881  * @h: struct hstate old page belongs to
2882  * @old_folio: Old folio to dissolve
2883  * @list: List to isolate the page in case we need to
2884  * Returns 0 on success, otherwise negated error.
2885  */
2886 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2887                         struct folio *old_folio, struct list_head *list)
2888 {
2889         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2890         int nid = folio_nid(old_folio);
2891         struct folio *new_folio;
2892         int ret = 0;
2893
2894         /*
2895          * Before dissolving the folio, we need to allocate a new one for the
2896          * pool to remain stable.  Here, we allocate the folio and 'prep' it
2897          * by doing everything but actually updating counters and adding to
2898          * the pool.  This simplifies and let us do most of the processing
2899          * under the lock.
2900          */
2901         new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2902         if (!new_folio)
2903                 return -ENOMEM;
2904         __prep_new_hugetlb_folio(h, new_folio);
2905
2906 retry:
2907         spin_lock_irq(&hugetlb_lock);
2908         if (!folio_test_hugetlb(old_folio)) {
2909                 /*
2910                  * Freed from under us. Drop new_folio too.
2911                  */
2912                 goto free_new;
2913         } else if (folio_ref_count(old_folio)) {
2914                 bool isolated;
2915
2916                 /*
2917                  * Someone has grabbed the folio, try to isolate it here.
2918                  * Fail with -EBUSY if not possible.
2919                  */
2920                 spin_unlock_irq(&hugetlb_lock);
2921                 isolated = isolate_hugetlb(old_folio, list);
2922                 ret = isolated ? 0 : -EBUSY;
2923                 spin_lock_irq(&hugetlb_lock);
2924                 goto free_new;
2925         } else if (!folio_test_hugetlb_freed(old_folio)) {
2926                 /*
2927                  * Folio's refcount is 0 but it has not been enqueued in the
2928                  * freelist yet. Race window is small, so we can succeed here if
2929                  * we retry.
2930                  */
2931                 spin_unlock_irq(&hugetlb_lock);
2932                 cond_resched();
2933                 goto retry;
2934         } else {
2935                 /*
2936                  * Ok, old_folio is still a genuine free hugepage. Remove it from
2937                  * the freelist and decrease the counters. These will be
2938                  * incremented again when calling __prep_account_new_huge_page()
2939                  * and enqueue_hugetlb_folio() for new_folio. The counters will
2940                  * remain stable since this happens under the lock.
2941                  */
2942                 remove_hugetlb_folio(h, old_folio, false);
2943
2944                 /*
2945                  * Ref count on new_folio is already zero as it was dropped
2946                  * earlier.  It can be directly added to the pool free list.
2947                  */
2948                 __prep_account_new_huge_page(h, nid);
2949                 enqueue_hugetlb_folio(h, new_folio);
2950
2951                 /*
2952                  * Folio has been replaced, we can safely free the old one.
2953                  */
2954                 spin_unlock_irq(&hugetlb_lock);
2955                 update_and_free_hugetlb_folio(h, old_folio, false);
2956         }
2957
2958         return ret;
2959
2960 free_new:
2961         spin_unlock_irq(&hugetlb_lock);
2962         /* Folio has a zero ref count, but needs a ref to be freed */
2963         folio_ref_unfreeze(new_folio, 1);
2964         update_and_free_hugetlb_folio(h, new_folio, false);
2965
2966         return ret;
2967 }
2968
2969 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2970 {
2971         struct hstate *h;
2972         struct folio *folio = page_folio(page);
2973         int ret = -EBUSY;
2974
2975         /*
2976          * The page might have been dissolved from under our feet, so make sure
2977          * to carefully check the state under the lock.
2978          * Return success when racing as if we dissolved the page ourselves.
2979          */
2980         spin_lock_irq(&hugetlb_lock);
2981         if (folio_test_hugetlb(folio)) {
2982                 h = folio_hstate(folio);
2983         } else {
2984                 spin_unlock_irq(&hugetlb_lock);
2985                 return 0;
2986         }
2987         spin_unlock_irq(&hugetlb_lock);
2988
2989         /*
2990          * Fence off gigantic pages as there is a cyclic dependency between
2991          * alloc_contig_range and them. Return -ENOMEM as this has the effect
2992          * of bailing out right away without further retrying.
2993          */
2994         if (hstate_is_gigantic(h))
2995                 return -ENOMEM;
2996
2997         if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
2998                 ret = 0;
2999         else if (!folio_ref_count(folio))
3000                 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3001
3002         return ret;
3003 }
3004
3005 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3006                                     unsigned long addr, int avoid_reserve)
3007 {
3008         struct hugepage_subpool *spool = subpool_vma(vma);
3009         struct hstate *h = hstate_vma(vma);
3010         struct folio *folio;
3011         long map_chg, map_commit;
3012         long gbl_chg;
3013         int ret, idx;
3014         struct hugetlb_cgroup *h_cg = NULL;
3015         bool deferred_reserve;
3016
3017         idx = hstate_index(h);
3018         /*
3019          * Examine the region/reserve map to determine if the process
3020          * has a reservation for the page to be allocated.  A return
3021          * code of zero indicates a reservation exists (no change).
3022          */
3023         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3024         if (map_chg < 0)
3025                 return ERR_PTR(-ENOMEM);
3026
3027         /*
3028          * Processes that did not create the mapping will have no
3029          * reserves as indicated by the region/reserve map. Check
3030          * that the allocation will not exceed the subpool limit.
3031          * Allocations for MAP_NORESERVE mappings also need to be
3032          * checked against any subpool limit.
3033          */
3034         if (map_chg || avoid_reserve) {
3035                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3036                 if (gbl_chg < 0) {
3037                         vma_end_reservation(h, vma, addr);
3038                         return ERR_PTR(-ENOSPC);
3039                 }
3040
3041                 /*
3042                  * Even though there was no reservation in the region/reserve
3043                  * map, there could be reservations associated with the
3044                  * subpool that can be used.  This would be indicated if the
3045                  * return value of hugepage_subpool_get_pages() is zero.
3046                  * However, if avoid_reserve is specified we still avoid even
3047                  * the subpool reservations.
3048                  */
3049                 if (avoid_reserve)
3050                         gbl_chg = 1;
3051         }
3052
3053         /* If this allocation is not consuming a reservation, charge it now.
3054          */
3055         deferred_reserve = map_chg || avoid_reserve;
3056         if (deferred_reserve) {
3057                 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3058                         idx, pages_per_huge_page(h), &h_cg);
3059                 if (ret)
3060                         goto out_subpool_put;
3061         }
3062
3063         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3064         if (ret)
3065                 goto out_uncharge_cgroup_reservation;
3066
3067         spin_lock_irq(&hugetlb_lock);
3068         /*
3069          * glb_chg is passed to indicate whether or not a page must be taken
3070          * from the global free pool (global change).  gbl_chg == 0 indicates
3071          * a reservation exists for the allocation.
3072          */
3073         folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3074         if (!folio) {
3075                 spin_unlock_irq(&hugetlb_lock);
3076                 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3077                 if (!folio)
3078                         goto out_uncharge_cgroup;
3079                 spin_lock_irq(&hugetlb_lock);
3080                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3081                         folio_set_hugetlb_restore_reserve(folio);
3082                         h->resv_huge_pages--;
3083                 }
3084                 list_add(&folio->lru, &h->hugepage_activelist);
3085                 folio_ref_unfreeze(folio, 1);
3086                 /* Fall through */
3087         }
3088
3089         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3090         /* If allocation is not consuming a reservation, also store the
3091          * hugetlb_cgroup pointer on the page.
3092          */
3093         if (deferred_reserve) {
3094                 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3095                                                   h_cg, folio);
3096         }
3097
3098         spin_unlock_irq(&hugetlb_lock);
3099
3100         hugetlb_set_folio_subpool(folio, spool);
3101
3102         map_commit = vma_commit_reservation(h, vma, addr);
3103         if (unlikely(map_chg > map_commit)) {
3104                 /*
3105                  * The page was added to the reservation map between
3106                  * vma_needs_reservation and vma_commit_reservation.
3107                  * This indicates a race with hugetlb_reserve_pages.
3108                  * Adjust for the subpool count incremented above AND
3109                  * in hugetlb_reserve_pages for the same page.  Also,
3110                  * the reservation count added in hugetlb_reserve_pages
3111                  * no longer applies.
3112                  */
3113                 long rsv_adjust;
3114
3115                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3116                 hugetlb_acct_memory(h, -rsv_adjust);
3117                 if (deferred_reserve)
3118                         hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3119                                         pages_per_huge_page(h), folio);
3120         }
3121         return folio;
3122
3123 out_uncharge_cgroup:
3124         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3125 out_uncharge_cgroup_reservation:
3126         if (deferred_reserve)
3127                 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3128                                                     h_cg);
3129 out_subpool_put:
3130         if (map_chg || avoid_reserve)
3131                 hugepage_subpool_put_pages(spool, 1);
3132         vma_end_reservation(h, vma, addr);
3133         return ERR_PTR(-ENOSPC);
3134 }
3135
3136 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3137         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3138 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3139 {
3140         struct huge_bootmem_page *m = NULL; /* initialize for clang */
3141         int nr_nodes, node;
3142
3143         /* do node specific alloc */
3144         if (nid != NUMA_NO_NODE) {
3145                 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3146                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3147                 if (!m)
3148                         return 0;
3149                 goto found;
3150         }
3151         /* allocate from next node when distributing huge pages */
3152         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3153                 m = memblock_alloc_try_nid_raw(
3154                                 huge_page_size(h), huge_page_size(h),
3155                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3156                 /*
3157                  * Use the beginning of the huge page to store the
3158                  * huge_bootmem_page struct (until gather_bootmem
3159                  * puts them into the mem_map).
3160                  */
3161                 if (!m)
3162                         return 0;
3163                 goto found;
3164         }
3165
3166 found:
3167         /* Put them into a private list first because mem_map is not up yet */
3168         INIT_LIST_HEAD(&m->list);
3169         list_add(&m->list, &huge_boot_pages);
3170         m->hstate = h;
3171         return 1;
3172 }
3173
3174 /*
3175  * Put bootmem huge pages into the standard lists after mem_map is up.
3176  * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3177  */
3178 static void __init gather_bootmem_prealloc(void)
3179 {
3180         struct huge_bootmem_page *m;
3181
3182         list_for_each_entry(m, &huge_boot_pages, list) {
3183                 struct page *page = virt_to_page(m);
3184                 struct folio *folio = page_folio(page);
3185                 struct hstate *h = m->hstate;
3186
3187                 VM_BUG_ON(!hstate_is_gigantic(h));
3188                 WARN_ON(folio_ref_count(folio) != 1);
3189                 if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3190                         WARN_ON(folio_test_reserved(folio));
3191                         prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3192                         free_huge_folio(folio); /* add to the hugepage allocator */
3193                 } else {
3194                         /* VERY unlikely inflated ref count on a tail page */
3195                         free_gigantic_folio(folio, huge_page_order(h));
3196                 }
3197
3198                 /*
3199                  * We need to restore the 'stolen' pages to totalram_pages
3200                  * in order to fix confusing memory reports from free(1) and
3201                  * other side-effects, like CommitLimit going negative.
3202                  */
3203                 adjust_managed_page_count(page, pages_per_huge_page(h));
3204                 cond_resched();
3205         }
3206 }
3207 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3208 {
3209         unsigned long i;
3210         char buf[32];
3211
3212         for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3213                 if (hstate_is_gigantic(h)) {
3214                         if (!alloc_bootmem_huge_page(h, nid))
3215                                 break;
3216                 } else {
3217                         struct folio *folio;
3218                         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3219
3220                         folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3221                                         &node_states[N_MEMORY], NULL);
3222                         if (!folio)
3223                                 break;
3224                         free_huge_folio(folio); /* free it into the hugepage allocator */
3225                 }
3226                 cond_resched();
3227         }
3228         if (i == h->max_huge_pages_node[nid])
3229                 return;
3230
3231         string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3232         pr_warn("HugeTLB: allocating %u of page size %s failed node%d.  Only allocated %lu hugepages.\n",
3233                 h->max_huge_pages_node[nid], buf, nid, i);
3234         h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3235         h->max_huge_pages_node[nid] = i;
3236 }
3237
3238 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3239 {
3240         unsigned long i;
3241         nodemask_t *node_alloc_noretry;
3242         bool node_specific_alloc = false;
3243
3244         /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3245         if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3246                 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3247                 return;
3248         }
3249
3250         /* do node specific alloc */
3251         for_each_online_node(i) {
3252                 if (h->max_huge_pages_node[i] > 0) {
3253                         hugetlb_hstate_alloc_pages_onenode(h, i);
3254                         node_specific_alloc = true;
3255                 }
3256         }
3257
3258         if (node_specific_alloc)
3259                 return;
3260
3261         /* below will do all node balanced alloc */
3262         if (!hstate_is_gigantic(h)) {
3263                 /*
3264                  * Bit mask controlling how hard we retry per-node allocations.
3265                  * Ignore errors as lower level routines can deal with
3266                  * node_alloc_noretry == NULL.  If this kmalloc fails at boot
3267                  * time, we are likely in bigger trouble.
3268                  */
3269                 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3270                                                 GFP_KERNEL);
3271         } else {
3272                 /* allocations done at boot time */
3273                 node_alloc_noretry = NULL;
3274         }
3275
3276         /* bit mask controlling how hard we retry per-node allocations */
3277         if (node_alloc_noretry)
3278                 nodes_clear(*node_alloc_noretry);
3279
3280         for (i = 0; i < h->max_huge_pages; ++i) {
3281                 if (hstate_is_gigantic(h)) {
3282                         if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3283                                 break;
3284                 } else if (!alloc_pool_huge_page(h,
3285                                          &node_states[N_MEMORY],
3286                                          node_alloc_noretry))
3287                         break;
3288                 cond_resched();
3289         }
3290         if (i < h->max_huge_pages) {
3291                 char buf[32];
3292
3293                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3294                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
3295                         h->max_huge_pages, buf, i);
3296                 h->max_huge_pages = i;
3297         }
3298         kfree(node_alloc_noretry);
3299 }
3300
3301 static void __init hugetlb_init_hstates(void)
3302 {
3303         struct hstate *h, *h2;
3304
3305         for_each_hstate(h) {
3306                 /* oversize hugepages were init'ed in early boot */
3307                 if (!hstate_is_gigantic(h))
3308                         hugetlb_hstate_alloc_pages(h);
3309
3310                 /*
3311                  * Set demote order for each hstate.  Note that
3312                  * h->demote_order is initially 0.
3313                  * - We can not demote gigantic pages if runtime freeing
3314                  *   is not supported, so skip this.
3315                  * - If CMA allocation is possible, we can not demote
3316                  *   HUGETLB_PAGE_ORDER or smaller size pages.
3317                  */
3318                 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3319                         continue;
3320                 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3321                         continue;
3322                 for_each_hstate(h2) {
3323                         if (h2 == h)
3324                                 continue;
3325                         if (h2->order < h->order &&
3326                             h2->order > h->demote_order)
3327                                 h->demote_order = h2->order;
3328                 }
3329         }
3330 }
3331
3332 static void __init report_hugepages(void)
3333 {
3334         struct hstate *h;
3335
3336         for_each_hstate(h) {
3337                 char buf[32];
3338
3339                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3340                 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3341                         buf, h->free_huge_pages);
3342                 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3343                         hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3344         }
3345 }
3346
3347 #ifdef CONFIG_HIGHMEM
3348 static void try_to_free_low(struct hstate *h, unsigned long count,
3349                                                 nodemask_t *nodes_allowed)
3350 {
3351         int i;
3352         LIST_HEAD(page_list);
3353
3354         lockdep_assert_held(&hugetlb_lock);
3355         if (hstate_is_gigantic(h))
3356                 return;
3357
3358         /*
3359          * Collect pages to be freed on a list, and free after dropping lock
3360          */
3361         for_each_node_mask(i, *nodes_allowed) {
3362                 struct page *page, *next;
3363                 struct list_head *freel = &h->hugepage_freelists[i];
3364                 list_for_each_entry_safe(page, next, freel, lru) {
3365                         if (count >= h->nr_huge_pages)
3366                                 goto out;
3367                         if (PageHighMem(page))
3368                                 continue;
3369                         remove_hugetlb_folio(h, page_folio(page), false);
3370                         list_add(&page->lru, &page_list);
3371                 }
3372         }
3373
3374 out:
3375         spin_unlock_irq(&hugetlb_lock);
3376         update_and_free_pages_bulk(h, &page_list);
3377         spin_lock_irq(&hugetlb_lock);
3378 }
3379 #else
3380 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3381                                                 nodemask_t *nodes_allowed)
3382 {
3383 }
3384 #endif
3385
3386 /*
3387  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
3388  * balanced by operating on them in a round-robin fashion.
3389  * Returns 1 if an adjustment was made.
3390  */
3391 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3392                                 int delta)
3393 {
3394         int nr_nodes, node;
3395
3396         lockdep_assert_held(&hugetlb_lock);
3397         VM_BUG_ON(delta != -1 && delta != 1);
3398
3399         if (delta < 0) {
3400                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3401                         if (h->surplus_huge_pages_node[node])
3402                                 goto found;
3403                 }
3404         } else {
3405                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3406                         if (h->surplus_huge_pages_node[node] <
3407                                         h->nr_huge_pages_node[node])
3408                                 goto found;
3409                 }
3410         }
3411         return 0;
3412
3413 found:
3414         h->surplus_huge_pages += delta;
3415         h->surplus_huge_pages_node[node] += delta;
3416         return 1;
3417 }
3418
3419 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3420 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3421                               nodemask_t *nodes_allowed)
3422 {
3423         unsigned long min_count, ret;
3424         struct page *page;
3425         LIST_HEAD(page_list);
3426         NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3427
3428         /*
3429          * Bit mask controlling how hard we retry per-node allocations.
3430          * If we can not allocate the bit mask, do not attempt to allocate
3431          * the requested huge pages.
3432          */
3433         if (node_alloc_noretry)
3434                 nodes_clear(*node_alloc_noretry);
3435         else
3436                 return -ENOMEM;
3437
3438         /*
3439          * resize_lock mutex prevents concurrent adjustments to number of
3440          * pages in hstate via the proc/sysfs interfaces.
3441          */
3442         mutex_lock(&h->resize_lock);
3443         flush_free_hpage_work(h);
3444         spin_lock_irq(&hugetlb_lock);
3445
3446         /*
3447          * Check for a node specific request.
3448          * Changing node specific huge page count may require a corresponding
3449          * change to the global count.  In any case, the passed node mask
3450          * (nodes_allowed) will restrict alloc/free to the specified node.
3451          */
3452         if (nid != NUMA_NO_NODE) {
3453                 unsigned long old_count = count;
3454
3455                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3456                 /*
3457                  * User may have specified a large count value which caused the
3458                  * above calculation to overflow.  In this case, they wanted
3459                  * to allocate as many huge pages as possible.  Set count to
3460                  * largest possible value to align with their intention.
3461                  */
3462                 if (count < old_count)
3463                         count = ULONG_MAX;
3464         }
3465
3466         /*
3467          * Gigantic pages runtime allocation depend on the capability for large
3468          * page range allocation.
3469          * If the system does not provide this feature, return an error when
3470          * the user tries to allocate gigantic pages but let the user free the
3471          * boottime allocated gigantic pages.
3472          */
3473         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3474                 if (count > persistent_huge_pages(h)) {
3475                         spin_unlock_irq(&hugetlb_lock);
3476                         mutex_unlock(&h->resize_lock);
3477                         NODEMASK_FREE(node_alloc_noretry);
3478                         return -EINVAL;
3479                 }
3480                 /* Fall through to decrease pool */
3481         }
3482
3483         /*
3484          * Increase the pool size
3485          * First take pages out of surplus state.  Then make up the
3486          * remaining difference by allocating fresh huge pages.
3487          *
3488          * We might race with alloc_surplus_hugetlb_folio() here and be unable
3489          * to convert a surplus huge page to a normal huge page. That is
3490          * not critical, though, it just means the overall size of the
3491          * pool might be one hugepage larger than it needs to be, but
3492          * within all the constraints specified by the sysctls.
3493          */
3494         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3495                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3496                         break;
3497         }
3498
3499         while (count > persistent_huge_pages(h)) {
3500                 /*
3501                  * If this allocation races such that we no longer need the
3502                  * page, free_huge_folio will handle it by freeing the page
3503                  * and reducing the surplus.
3504                  */
3505                 spin_unlock_irq(&hugetlb_lock);
3506
3507                 /* yield cpu to avoid soft lockup */
3508                 cond_resched();
3509
3510                 ret = alloc_pool_huge_page(h, nodes_allowed,
3511                                                 node_alloc_noretry);
3512                 spin_lock_irq(&hugetlb_lock);
3513                 if (!ret)
3514                         goto out;
3515
3516                 /* Bail for signals. Probably ctrl-c from user */
3517                 if (signal_pending(current))
3518                         goto out;
3519         }
3520
3521         /*
3522          * Decrease the pool size
3523          * First return free pages to the buddy allocator (being careful
3524          * to keep enough around to satisfy reservations).  Then place
3525          * pages into surplus state as needed so the pool will shrink
3526          * to the desired size as pages become free.
3527          *
3528          * By placing pages into the surplus state independent of the
3529          * overcommit value, we are allowing the surplus pool size to
3530          * exceed overcommit. There are few sane options here. Since
3531          * alloc_surplus_hugetlb_folio() is checking the global counter,
3532          * though, we'll note that we're not allowed to exceed surplus
3533          * and won't grow the pool anywhere else. Not until one of the
3534          * sysctls are changed, or the surplus pages go out of use.
3535          */
3536         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3537         min_count = max(count, min_count);
3538         try_to_free_low(h, min_count, nodes_allowed);
3539
3540         /*
3541          * Collect pages to be removed on list without dropping lock
3542          */
3543         while (min_count < persistent_huge_pages(h)) {
3544                 page = remove_pool_huge_page(h, nodes_allowed, 0);
3545                 if (!page)
3546                         break;
3547
3548                 list_add(&page->lru, &page_list);
3549         }
3550         /* free the pages after dropping lock */
3551         spin_unlock_irq(&hugetlb_lock);
3552         update_and_free_pages_bulk(h, &page_list);
3553         flush_free_hpage_work(h);
3554         spin_lock_irq(&hugetlb_lock);
3555
3556         while (count < persistent_huge_pages(h)) {
3557                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3558                         break;
3559         }
3560 out:
3561         h->max_huge_pages = persistent_huge_pages(h);
3562         spin_unlock_irq(&hugetlb_lock);
3563         mutex_unlock(&h->resize_lock);
3564
3565         NODEMASK_FREE(node_alloc_noretry);
3566
3567         return 0;
3568 }
3569
3570 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3571 {
3572         int i, nid = folio_nid(folio);
3573         struct hstate *target_hstate;
3574         struct page *subpage;
3575         struct folio *inner_folio;
3576         int rc = 0;
3577
3578         target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3579
3580         remove_hugetlb_folio_for_demote(h, folio, false);
3581         spin_unlock_irq(&hugetlb_lock);
3582
3583         rc = hugetlb_vmemmap_restore(h, &folio->page);
3584         if (rc) {
3585                 /* Allocation of vmemmmap failed, we can not demote folio */
3586                 spin_lock_irq(&hugetlb_lock);
3587                 folio_ref_unfreeze(folio, 1);
3588                 add_hugetlb_folio(h, folio, false);
3589                 return rc;
3590         }
3591
3592         /*
3593          * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3594          * sizes as it will not ref count folios.
3595          */
3596         destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3597
3598         /*
3599          * Taking target hstate mutex synchronizes with set_max_huge_pages.
3600          * Without the mutex, pages added to target hstate could be marked
3601          * as surplus.
3602          *
3603          * Note that we already hold h->resize_lock.  To prevent deadlock,
3604          * use the convention of always taking larger size hstate mutex first.
3605          */
3606         mutex_lock(&target_hstate->resize_lock);
3607         for (i = 0; i < pages_per_huge_page(h);
3608                                 i += pages_per_huge_page(target_hstate)) {
3609                 subpage = folio_page(folio, i);
3610                 inner_folio = page_folio(subpage);
3611                 if (hstate_is_gigantic(target_hstate))
3612                         prep_compound_gigantic_folio_for_demote(inner_folio,
3613                                                         target_hstate->order);
3614                 else
3615                         prep_compound_page(subpage, target_hstate->order);
3616                 folio_change_private(inner_folio, NULL);
3617                 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3618                 free_huge_folio(inner_folio);
3619         }
3620         mutex_unlock(&target_hstate->resize_lock);
3621
3622         spin_lock_irq(&hugetlb_lock);
3623
3624         /*
3625          * Not absolutely necessary, but for consistency update max_huge_pages
3626          * based on pool changes for the demoted page.
3627          */
3628         h->max_huge_pages--;
3629         target_hstate->max_huge_pages +=
3630                 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3631
3632         return rc;
3633 }
3634
3635 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3636         __must_hold(&hugetlb_lock)
3637 {
3638         int nr_nodes, node;
3639         struct folio *folio;
3640
3641         lockdep_assert_held(&hugetlb_lock);
3642
3643         /* We should never get here if no demote order */
3644         if (!h->demote_order) {
3645                 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3646                 return -EINVAL;         /* internal error */
3647         }
3648
3649         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3650                 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
3651                         if (folio_test_hwpoison(folio))
3652                                 continue;
3653                         return demote_free_hugetlb_folio(h, folio);
3654                 }
3655         }
3656
3657         /*
3658          * Only way to get here is if all pages on free lists are poisoned.
3659          * Return -EBUSY so that caller will not retry.
3660          */
3661         return -EBUSY;
3662 }
3663
3664 #define HSTATE_ATTR_RO(_name) \
3665         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3666
3667 #define HSTATE_ATTR_WO(_name) \
3668         static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3669
3670 #define HSTATE_ATTR(_name) \
3671         static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3672
3673 static struct kobject *hugepages_kobj;
3674 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3675
3676 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3677
3678 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3679 {
3680         int i;
3681
3682         for (i = 0; i < HUGE_MAX_HSTATE; i++)
3683                 if (hstate_kobjs[i] == kobj) {
3684                         if (nidp)
3685                                 *nidp = NUMA_NO_NODE;
3686                         return &hstates[i];
3687                 }
3688
3689         return kobj_to_node_hstate(kobj, nidp);
3690 }
3691
3692 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3693                                         struct kobj_attribute *attr, char *buf)
3694 {
3695         struct hstate *h;
3696         unsigned long nr_huge_pages;
3697         int nid;
3698
3699         h = kobj_to_hstate(kobj, &nid);
3700         if (nid == NUMA_NO_NODE)
3701                 nr_huge_pages = h->nr_huge_pages;
3702         else
3703                 nr_huge_pages = h->nr_huge_pages_node[nid];
3704
3705         return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3706 }
3707
3708 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3709                                            struct hstate *h, int nid,
3710                                            unsigned long count, size_t len)
3711 {
3712         int err;
3713         nodemask_t nodes_allowed, *n_mask;
3714
3715         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3716                 return -EINVAL;
3717
3718         if (nid == NUMA_NO_NODE) {
3719                 /*
3720                  * global hstate attribute
3721                  */
3722                 if (!(obey_mempolicy &&
3723                                 init_nodemask_of_mempolicy(&nodes_allowed)))
3724                         n_mask = &node_states[N_MEMORY];
3725                 else
3726                         n_mask = &nodes_allowed;
3727         } else {
3728                 /*
3729                  * Node specific request.  count adjustment happens in
3730                  * set_max_huge_pages() after acquiring hugetlb_lock.
3731                  */
3732                 init_nodemask_of_node(&nodes_allowed, nid);
3733                 n_mask = &nodes_allowed;
3734         }
3735
3736         err = set_max_huge_pages(h, count, nid, n_mask);
3737
3738         return err ? err : len;
3739 }
3740
3741 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3742                                          struct kobject *kobj, const char *buf,
3743                                          size_t len)
3744 {
3745         struct hstate *h;
3746         unsigned long count;
3747         int nid;
3748         int err;
3749
3750         err = kstrtoul(buf, 10, &count);
3751         if (err)
3752                 return err;
3753
3754         h = kobj_to_hstate(kobj, &nid);
3755         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3756 }
3757
3758 static ssize_t nr_hugepages_show(struct kobject *kobj,
3759                                        struct kobj_attribute *attr, char *buf)
3760 {
3761         return nr_hugepages_show_common(kobj, attr, buf);
3762 }
3763
3764 static ssize_t nr_hugepages_store(struct kobject *kobj,
3765                struct kobj_attribute *attr, const char *buf, size_t len)
3766 {
3767         return nr_hugepages_store_common(false, kobj, buf, len);
3768 }
3769 HSTATE_ATTR(nr_hugepages);
3770
3771 #ifdef CONFIG_NUMA
3772
3773 /*
3774  * hstate attribute for optionally mempolicy-based constraint on persistent
3775  * huge page alloc/free.
3776  */
3777 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3778                                            struct kobj_attribute *attr,
3779                                            char *buf)
3780 {
3781         return nr_hugepages_show_common(kobj, attr, buf);
3782 }
3783
3784 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3785                struct kobj_attribute *attr, const char *buf, size_t len)
3786 {
3787         return nr_hugepages_store_common(true, kobj, buf, len);
3788 }
3789 HSTATE_ATTR(nr_hugepages_mempolicy);
3790 #endif
3791
3792
3793 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3794                                         struct kobj_attribute *attr, char *buf)
3795 {
3796         struct hstate *h = kobj_to_hstate(kobj, NULL);
3797         return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3798 }
3799
3800 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3801                 struct kobj_attribute *attr, const char *buf, size_t count)
3802 {
3803         int err;
3804         unsigned long input;
3805         struct hstate *h = kobj_to_hstate(kobj, NULL);
3806
3807         if (hstate_is_gigantic(h))
3808                 return -EINVAL;
3809
3810         err = kstrtoul(buf, 10, &input);
3811         if (err)
3812                 return err;
3813
3814         spin_lock_irq(&hugetlb_lock);
3815         h->nr_overcommit_huge_pages = input;
3816         spin_unlock_irq(&hugetlb_lock);
3817
3818         return count;
3819 }
3820 HSTATE_ATTR(nr_overcommit_hugepages);
3821
3822 static ssize_t free_hugepages_show(struct kobject *kobj,
3823                                         struct kobj_attribute *attr, char *buf)
3824 {
3825         struct hstate *h;
3826         unsigned long free_huge_pages;
3827         int nid;
3828
3829         h = kobj_to_hstate(kobj, &nid);
3830         if (nid == NUMA_NO_NODE)
3831                 free_huge_pages = h->free_huge_pages;
3832         else
3833                 free_huge_pages = h->free_huge_pages_node[nid];
3834
3835         return sysfs_emit(buf, "%lu\n", free_huge_pages);
3836 }
3837 HSTATE_ATTR_RO(free_hugepages);
3838
3839 static ssize_t resv_hugepages_show(struct kobject *kobj,
3840                                         struct kobj_attribute *attr, char *buf)
3841 {
3842         struct hstate *h = kobj_to_hstate(kobj, NULL);
3843         return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3844 }
3845 HSTATE_ATTR_RO(resv_hugepages);
3846
3847 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3848                                         struct kobj_attribute *attr, char *buf)
3849 {
3850         struct hstate *h;
3851         unsigned long surplus_huge_pages;
3852         int nid;
3853
3854         h = kobj_to_hstate(kobj, &nid);
3855         if (nid == NUMA_NO_NODE)
3856                 surplus_huge_pages = h->surplus_huge_pages;
3857         else
3858                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3859
3860         return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3861 }
3862 HSTATE_ATTR_RO(surplus_hugepages);
3863
3864 static ssize_t demote_store(struct kobject *kobj,
3865                struct kobj_attribute *attr, const char *buf, size_t len)
3866 {
3867         unsigned long nr_demote;
3868         unsigned long nr_available;
3869         nodemask_t nodes_allowed, *n_mask;
3870         struct hstate *h;
3871         int err;
3872         int nid;
3873
3874         err = kstrtoul(buf, 10, &nr_demote);
3875         if (err)
3876                 return err;
3877         h = kobj_to_hstate(kobj, &nid);
3878
3879         if (nid != NUMA_NO_NODE) {
3880                 init_nodemask_of_node(&nodes_allowed, nid);
3881                 n_mask = &nodes_allowed;
3882         } else {
3883                 n_mask = &node_states[N_MEMORY];
3884         }
3885
3886         /* Synchronize with other sysfs operations modifying huge pages */
3887         mutex_lock(&h->resize_lock);
3888         spin_lock_irq(&hugetlb_lock);
3889
3890         while (nr_demote) {
3891                 /*
3892                  * Check for available pages to demote each time thorough the
3893                  * loop as demote_pool_huge_page will drop hugetlb_lock.
3894                  */
3895                 if (nid != NUMA_NO_NODE)
3896                         nr_available = h->free_huge_pages_node[nid];
3897                 else
3898                         nr_available = h->free_huge_pages;
3899                 nr_available -= h->resv_huge_pages;
3900                 if (!nr_available)
3901                         break;
3902
3903                 err = demote_pool_huge_page(h, n_mask);
3904                 if (err)
3905                         break;
3906
3907                 nr_demote--;
3908         }
3909
3910         spin_unlock_irq(&hugetlb_lock);
3911         mutex_unlock(&h->resize_lock);
3912
3913         if (err)
3914                 return err;
3915         return len;
3916 }
3917 HSTATE_ATTR_WO(demote);
3918
3919 static ssize_t demote_size_show(struct kobject *kobj,
3920                                         struct kobj_attribute *attr, char *buf)
3921 {
3922         struct hstate *h = kobj_to_hstate(kobj, NULL);
3923         unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3924
3925         return sysfs_emit(buf, "%lukB\n", demote_size);
3926 }
3927
3928 static ssize_t demote_size_store(struct kobject *kobj,
3929                                         struct kobj_attribute *attr,
3930                                         const char *buf, size_t count)
3931 {
3932         struct hstate *h, *demote_hstate;
3933         unsigned long demote_size;
3934         unsigned int demote_order;
3935
3936         demote_size = (unsigned long)memparse(buf, NULL);
3937
3938         demote_hstate = size_to_hstate(demote_size);
3939         if (!demote_hstate)
3940                 return -EINVAL;
3941         demote_order = demote_hstate->order;
3942         if (demote_order < HUGETLB_PAGE_ORDER)
3943                 return -EINVAL;
3944
3945         /* demote order must be smaller than hstate order */
3946         h = kobj_to_hstate(kobj, NULL);
3947         if (demote_order >= h->order)
3948                 return -EINVAL;
3949
3950         /* resize_lock synchronizes access to demote size and writes */
3951         mutex_lock(&h->resize_lock);
3952         h->demote_order = demote_order;
3953         mutex_unlock(&h->resize_lock);
3954
3955         return count;
3956 }
3957 HSTATE_ATTR(demote_size);
3958
3959 static struct attribute *hstate_attrs[] = {
3960         &nr_hugepages_attr.attr,
3961         &nr_overcommit_hugepages_attr.attr,
3962         &free_hugepages_attr.attr,
3963         &resv_hugepages_attr.attr,
3964         &surplus_hugepages_attr.attr,
3965 #ifdef CONFIG_NUMA
3966         &nr_hugepages_mempolicy_attr.attr,
3967 #endif
3968         NULL,
3969 };
3970
3971 static const struct attribute_group hstate_attr_group = {
3972         .attrs = hstate_attrs,
3973 };
3974
3975 static struct attribute *hstate_demote_attrs[] = {
3976         &demote_size_attr.attr,
3977         &demote_attr.attr,
3978         NULL,
3979 };
3980
3981 static const struct attribute_group hstate_demote_attr_group = {
3982         .attrs = hstate_demote_attrs,
3983 };
3984
3985 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3986                                     struct kobject **hstate_kobjs,
3987                                     const struct attribute_group *hstate_attr_group)
3988 {
3989         int retval;
3990         int hi = hstate_index(h);
3991
3992         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3993         if (!hstate_kobjs[hi])
3994                 return -ENOMEM;
3995
3996         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3997         if (retval) {
3998                 kobject_put(hstate_kobjs[hi]);
3999                 hstate_kobjs[hi] = NULL;
4000                 return retval;
4001         }
4002
4003         if (h->demote_order) {
4004                 retval = sysfs_create_group(hstate_kobjs[hi],
4005                                             &hstate_demote_attr_group);
4006                 if (retval) {
4007                         pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4008                         sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4009                         kobject_put(hstate_kobjs[hi]);
4010                         hstate_kobjs[hi] = NULL;
4011                         return retval;
4012                 }
4013         }
4014
4015         return 0;
4016 }
4017
4018 #ifdef CONFIG_NUMA
4019 static bool hugetlb_sysfs_initialized __ro_after_init;
4020
4021 /*
4022  * node_hstate/s - associate per node hstate attributes, via their kobjects,
4023  * with node devices in node_devices[] using a parallel array.  The array
4024  * index of a node device or _hstate == node id.
4025  * This is here to avoid any static dependency of the node device driver, in
4026  * the base kernel, on the hugetlb module.
4027  */
4028 struct node_hstate {
4029         struct kobject          *hugepages_kobj;
4030         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
4031 };
4032 static struct node_hstate node_hstates[MAX_NUMNODES];
4033
4034 /*
4035  * A subset of global hstate attributes for node devices
4036  */
4037 static struct attribute *per_node_hstate_attrs[] = {
4038         &nr_hugepages_attr.attr,
4039         &free_hugepages_attr.attr,
4040         &surplus_hugepages_attr.attr,
4041         NULL,
4042 };
4043
4044 static const struct attribute_group per_node_hstate_attr_group = {
4045         .attrs = per_node_hstate_attrs,
4046 };
4047
4048 /*
4049  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4050  * Returns node id via non-NULL nidp.
4051  */
4052 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4053 {
4054         int nid;
4055
4056         for (nid = 0; nid < nr_node_ids; nid++) {
4057                 struct node_hstate *nhs = &node_hstates[nid];
4058                 int i;
4059                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4060                         if (nhs->hstate_kobjs[i] == kobj) {
4061                                 if (nidp)
4062                                         *nidp = nid;
4063                                 return &hstates[i];
4064                         }
4065         }
4066
4067         BUG();
4068         return NULL;
4069 }
4070
4071 /*
4072  * Unregister hstate attributes from a single node device.
4073  * No-op if no hstate attributes attached.
4074  */
4075 void hugetlb_unregister_node(struct node *node)
4076 {
4077         struct hstate *h;
4078         struct node_hstate *nhs = &node_hstates[node->dev.id];
4079
4080         if (!nhs->hugepages_kobj)
4081                 return;         /* no hstate attributes */
4082
4083         for_each_hstate(h) {
4084                 int idx = hstate_index(h);
4085                 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4086
4087                 if (!hstate_kobj)
4088                         continue;
4089                 if (h->demote_order)
4090                         sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4091                 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4092                 kobject_put(hstate_kobj);
4093                 nhs->hstate_kobjs[idx] = NULL;
4094         }
4095
4096         kobject_put(nhs->hugepages_kobj);
4097         nhs->hugepages_kobj = NULL;
4098 }
4099
4100
4101 /*
4102  * Register hstate attributes for a single node device.
4103  * No-op if attributes already registered.
4104  */
4105 void hugetlb_register_node(struct node *node)
4106 {
4107         struct hstate *h;
4108         struct node_hstate *nhs = &node_hstates[node->dev.id];
4109         int err;
4110
4111         if (!hugetlb_sysfs_initialized)
4112                 return;
4113
4114         if (nhs->hugepages_kobj)
4115                 return;         /* already allocated */
4116
4117         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4118                                                         &node->dev.kobj);
4119         if (!nhs->hugepages_kobj)
4120                 return;
4121
4122         for_each_hstate(h) {
4123                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4124                                                 nhs->hstate_kobjs,
4125                                                 &per_node_hstate_attr_group);
4126                 if (err) {
4127                         pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4128                                 h->name, node->dev.id);
4129                         hugetlb_unregister_node(node);
4130                         break;
4131                 }
4132         }
4133 }
4134
4135 /*
4136  * hugetlb init time:  register hstate attributes for all registered node
4137  * devices of nodes that have memory.  All on-line nodes should have
4138  * registered their associated device by this time.
4139  */
4140 static void __init hugetlb_register_all_nodes(void)
4141 {
4142         int nid;
4143
4144         for_each_online_node(nid)
4145                 hugetlb_register_node(node_devices[nid]);
4146 }
4147 #else   /* !CONFIG_NUMA */
4148
4149 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4150 {
4151         BUG();
4152         if (nidp)
4153                 *nidp = -1;
4154         return NULL;
4155 }
4156
4157 static void hugetlb_register_all_nodes(void) { }
4158
4159 #endif
4160
4161 #ifdef CONFIG_CMA
4162 static void __init hugetlb_cma_check(void);
4163 #else
4164 static inline __init void hugetlb_cma_check(void)
4165 {
4166 }
4167 #endif
4168
4169 static void __init hugetlb_sysfs_init(void)
4170 {
4171         struct hstate *h;
4172         int err;
4173
4174         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4175         if (!hugepages_kobj)
4176                 return;
4177
4178         for_each_hstate(h) {
4179                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4180                                          hstate_kobjs, &hstate_attr_group);
4181                 if (err)
4182                         pr_err("HugeTLB: Unable to add hstate %s", h->name);
4183         }
4184
4185 #ifdef CONFIG_NUMA
4186         hugetlb_sysfs_initialized = true;
4187 #endif
4188         hugetlb_register_all_nodes();
4189 }
4190
4191 #ifdef CONFIG_SYSCTL
4192 static void hugetlb_sysctl_init(void);
4193 #else
4194 static inline void hugetlb_sysctl_init(void) { }
4195 #endif
4196
4197 static int __init hugetlb_init(void)
4198 {
4199         int i;
4200
4201         BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4202                         __NR_HPAGEFLAGS);
4203
4204         if (!hugepages_supported()) {
4205                 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4206                         pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4207                 return 0;
4208         }
4209
4210         /*
4211          * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some
4212          * architectures depend on setup being done here.
4213          */
4214         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4215         if (!parsed_default_hugepagesz) {
4216                 /*
4217                  * If we did not parse a default huge page size, set
4218                  * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4219                  * number of huge pages for this default size was implicitly
4220                  * specified, set that here as well.
4221                  * Note that the implicit setting will overwrite an explicit
4222                  * setting.  A warning will be printed in this case.
4223                  */
4224                 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4225                 if (default_hstate_max_huge_pages) {
4226                         if (default_hstate.max_huge_pages) {
4227                                 char buf[32];
4228
4229                                 string_get_size(huge_page_size(&default_hstate),
4230                                         1, STRING_UNITS_2, buf, 32);
4231                                 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4232                                         default_hstate.max_huge_pages, buf);
4233                                 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4234                                         default_hstate_max_huge_pages);
4235                         }
4236                         default_hstate.max_huge_pages =
4237                                 default_hstate_max_huge_pages;
4238
4239                         for_each_online_node(i)
4240                                 default_hstate.max_huge_pages_node[i] =
4241                                         default_hugepages_in_node[i];
4242                 }
4243         }
4244
4245         hugetlb_cma_check();
4246         hugetlb_init_hstates();
4247         gather_bootmem_prealloc();
4248         report_hugepages();
4249
4250         hugetlb_sysfs_init();
4251         hugetlb_cgroup_file_init();
4252         hugetlb_sysctl_init();
4253
4254 #ifdef CONFIG_SMP
4255         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4256 #else
4257         num_fault_mutexes = 1;
4258 #endif
4259         hugetlb_fault_mutex_table =
4260                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4261                               GFP_KERNEL);
4262         BUG_ON(!hugetlb_fault_mutex_table);
4263
4264         for (i = 0; i < num_fault_mutexes; i++)
4265                 mutex_init(&hugetlb_fault_mutex_table[i]);
4266         return 0;
4267 }
4268 subsys_initcall(hugetlb_init);
4269
4270 /* Overwritten by architectures with more huge page sizes */
4271 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4272 {
4273         return size == HPAGE_SIZE;
4274 }
4275
4276 void __init hugetlb_add_hstate(unsigned int order)
4277 {
4278         struct hstate *h;
4279         unsigned long i;
4280
4281         if (size_to_hstate(PAGE_SIZE << order)) {
4282                 return;
4283         }
4284         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4285         BUG_ON(order == 0);
4286         h = &hstates[hugetlb_max_hstate++];
4287         mutex_init(&h->resize_lock);
4288         h->order = order;
4289         h->mask = ~(huge_page_size(h) - 1);
4290         for (i = 0; i < MAX_NUMNODES; ++i)
4291                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4292         INIT_LIST_HEAD(&h->hugepage_activelist);
4293         h->next_nid_to_alloc = first_memory_node;
4294         h->next_nid_to_free = first_memory_node;
4295         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4296                                         huge_page_size(h)/SZ_1K);
4297
4298         parsed_hstate = h;
4299 }
4300
4301 bool __init __weak hugetlb_node_alloc_supported(void)
4302 {
4303         return true;
4304 }
4305
4306 static void __init hugepages_clear_pages_in_node(void)
4307 {
4308         if (!hugetlb_max_hstate) {
4309                 default_hstate_max_huge_pages = 0;
4310                 memset(default_hugepages_in_node, 0,
4311                         sizeof(default_hugepages_in_node));
4312         } else {
4313                 parsed_hstate->max_huge_pages = 0;
4314                 memset(parsed_hstate->max_huge_pages_node, 0,
4315                         sizeof(parsed_hstate->max_huge_pages_node));
4316         }
4317 }
4318
4319 /*
4320  * hugepages command line processing
4321  * hugepages normally follows a valid hugepagsz or default_hugepagsz
4322  * specification.  If not, ignore the hugepages value.  hugepages can also
4323  * be the first huge page command line  option in which case it implicitly
4324  * specifies the number of huge pages for the default size.
4325  */
4326 static int __init hugepages_setup(char *s)
4327 {
4328         unsigned long *mhp;
4329         static unsigned long *last_mhp;
4330         int node = NUMA_NO_NODE;
4331         int count;
4332         unsigned long tmp;
4333         char *p = s;
4334
4335         if (!parsed_valid_hugepagesz) {
4336                 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4337                 parsed_valid_hugepagesz = true;
4338                 return 1;
4339         }
4340
4341         /*
4342          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4343          * yet, so this hugepages= parameter goes to the "default hstate".
4344          * Otherwise, it goes with the previously parsed hugepagesz or
4345          * default_hugepagesz.
4346          */
4347         else if (!hugetlb_max_hstate)
4348                 mhp = &default_hstate_max_huge_pages;
4349         else
4350                 mhp = &parsed_hstate->max_huge_pages;
4351
4352         if (mhp == last_mhp) {
4353                 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4354                 return 1;
4355         }
4356
4357         while (*p) {
4358                 count = 0;
4359                 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4360                         goto invalid;
4361                 /* Parameter is node format */
4362                 if (p[count] == ':') {
4363                         if (!hugetlb_node_alloc_supported()) {
4364                                 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4365                                 return 1;
4366                         }
4367                         if (tmp >= MAX_NUMNODES || !node_online(tmp))
4368                                 goto invalid;
4369                         node = array_index_nospec(tmp, MAX_NUMNODES);
4370                         p += count + 1;
4371                         /* Parse hugepages */
4372                         if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4373                                 goto invalid;
4374                         if (!hugetlb_max_hstate)
4375                                 default_hugepages_in_node[node] = tmp;
4376                         else
4377                                 parsed_hstate->max_huge_pages_node[node] = tmp;
4378                         *mhp += tmp;
4379                         /* Go to parse next node*/
4380                         if (p[count] == ',')
4381                                 p += count + 1;
4382                         else
4383                                 break;
4384                 } else {
4385                         if (p != s)
4386                                 goto invalid;
4387                         *mhp = tmp;
4388                         break;
4389                 }
4390         }
4391
4392         /*
4393          * Global state is always initialized later in hugetlb_init.
4394          * But we need to allocate gigantic hstates here early to still
4395          * use the bootmem allocator.
4396          */
4397         if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4398                 hugetlb_hstate_alloc_pages(parsed_hstate);
4399
4400         last_mhp = mhp;
4401
4402         return 1;
4403
4404 invalid:
4405         pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4406         hugepages_clear_pages_in_node();
4407         return 1;
4408 }
4409 __setup("hugepages=", hugepages_setup);
4410
4411 /*
4412  * hugepagesz command line processing
4413  * A specific huge page size can only be specified once with hugepagesz.
4414  * hugepagesz is followed by hugepages on the command line.  The global
4415  * variable 'parsed_valid_hugepagesz' is used to determine if prior
4416  * hugepagesz argument was valid.
4417  */
4418 static int __init hugepagesz_setup(char *s)
4419 {
4420         unsigned long size;
4421         struct hstate *h;
4422
4423         parsed_valid_hugepagesz = false;
4424         size = (unsigned long)memparse(s, NULL);
4425
4426         if (!arch_hugetlb_valid_size(size)) {
4427                 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4428                 return 1;
4429         }
4430
4431         h = size_to_hstate(size);
4432         if (h) {
4433                 /*
4434                  * hstate for this size already exists.  This is normally
4435                  * an error, but is allowed if the existing hstate is the
4436                  * default hstate.  More specifically, it is only allowed if
4437                  * the number of huge pages for the default hstate was not
4438                  * previously specified.
4439                  */
4440                 if (!parsed_default_hugepagesz ||  h != &default_hstate ||
4441                     default_hstate.max_huge_pages) {
4442                         pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4443                         return 1;
4444                 }
4445
4446                 /*
4447                  * No need to call hugetlb_add_hstate() as hstate already
4448                  * exists.  But, do set parsed_hstate so that a following
4449                  * hugepages= parameter will be applied to this hstate.
4450                  */
4451                 parsed_hstate = h;
4452                 parsed_valid_hugepagesz = true;
4453                 return 1;
4454         }
4455
4456         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4457         parsed_valid_hugepagesz = true;
4458         return 1;
4459 }
4460 __setup("hugepagesz=", hugepagesz_setup);
4461
4462 /*
4463  * default_hugepagesz command line input
4464  * Only one instance of default_hugepagesz allowed on command line.
4465  */
4466 static int __init default_hugepagesz_setup(char *s)
4467 {
4468         unsigned long size;
4469         int i;
4470
4471         parsed_valid_hugepagesz = false;
4472         if (parsed_default_hugepagesz) {
4473                 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4474                 return 1;
4475         }
4476
4477         size = (unsigned long)memparse(s, NULL);
4478
4479         if (!arch_hugetlb_valid_size(size)) {
4480                 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4481                 return 1;
4482         }
4483
4484         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4485         parsed_valid_hugepagesz = true;
4486         parsed_default_hugepagesz = true;
4487         default_hstate_idx = hstate_index(size_to_hstate(size));
4488
4489         /*
4490          * The number of default huge pages (for this size) could have been
4491          * specified as the first hugetlb parameter: hugepages=X.  If so,
4492          * then default_hstate_max_huge_pages is set.  If the default huge
4493          * page size is gigantic (> MAX_ORDER), then the pages must be
4494          * allocated here from bootmem allocator.
4495          */
4496         if (default_hstate_max_huge_pages) {
4497                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4498                 for_each_online_node(i)
4499                         default_hstate.max_huge_pages_node[i] =
4500                                 default_hugepages_in_node[i];
4501                 if (hstate_is_gigantic(&default_hstate))
4502                         hugetlb_hstate_alloc_pages(&default_hstate);
4503                 default_hstate_max_huge_pages = 0;
4504         }
4505
4506         return 1;
4507 }
4508 __setup("default_hugepagesz=", default_hugepagesz_setup);
4509
4510 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4511 {
4512 #ifdef CONFIG_NUMA
4513         struct mempolicy *mpol = get_task_policy(current);
4514
4515         /*
4516          * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4517          * (from policy_nodemask) specifically for hugetlb case
4518          */
4519         if (mpol->mode == MPOL_BIND &&
4520                 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4521                  cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4522                 return &mpol->nodes;
4523 #endif
4524         return NULL;
4525 }
4526
4527 static unsigned int allowed_mems_nr(struct hstate *h)
4528 {
4529         int node;
4530         unsigned int nr = 0;
4531         nodemask_t *mbind_nodemask;
4532         unsigned int *array = h->free_huge_pages_node;
4533         gfp_t gfp_mask = htlb_alloc_mask(h);
4534
4535         mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4536         for_each_node_mask(node, cpuset_current_mems_allowed) {
4537                 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4538                         nr += array[node];
4539         }
4540
4541         return nr;
4542 }
4543
4544 #ifdef CONFIG_SYSCTL
4545 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4546                                           void *buffer, size_t *length,
4547                                           loff_t *ppos, unsigned long *out)
4548 {
4549         struct ctl_table dup_table;
4550
4551         /*
4552          * In order to avoid races with __do_proc_doulongvec_minmax(), we
4553          * can duplicate the @table and alter the duplicate of it.
4554          */
4555         dup_table = *table;
4556         dup_table.data = out;
4557
4558         return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4559 }
4560
4561 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4562                          struct ctl_table *table, int write,
4563                          void *buffer, size_t *length, loff_t *ppos)
4564 {
4565         struct hstate *h = &default_hstate;
4566         unsigned long tmp = h->max_huge_pages;
4567         int ret;
4568
4569         if (!hugepages_supported())
4570                 return -EOPNOTSUPP;
4571
4572         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4573                                              &tmp);
4574         if (ret)
4575                 goto out;
4576
4577         if (write)
4578                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4579                                                   NUMA_NO_NODE, tmp, *length);
4580 out:
4581         return ret;
4582 }
4583
4584 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4585                           void *buffer, size_t *length, loff_t *ppos)
4586 {
4587
4588         return hugetlb_sysctl_handler_common(false, table, write,
4589                                                         buffer, length, ppos);
4590 }
4591
4592 #ifdef CONFIG_NUMA
4593 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4594                           void *buffer, size_t *length, loff_t *ppos)
4595 {
4596         return hugetlb_sysctl_handler_common(true, table, write,
4597                                                         buffer, length, ppos);
4598 }
4599 #endif /* CONFIG_NUMA */
4600
4601 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4602                 void *buffer, size_t *length, loff_t *ppos)
4603 {
4604         struct hstate *h = &default_hstate;
4605         unsigned long tmp;
4606         int ret;
4607
4608         if (!hugepages_supported())
4609                 return -EOPNOTSUPP;
4610
4611         tmp = h->nr_overcommit_huge_pages;
4612
4613         if (write && hstate_is_gigantic(h))
4614                 return -EINVAL;
4615
4616         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4617                                              &tmp);
4618         if (ret)
4619                 goto out;
4620
4621         if (write) {
4622                 spin_lock_irq(&hugetlb_lock);
4623                 h->nr_overcommit_huge_pages = tmp;
4624                 spin_unlock_irq(&hugetlb_lock);
4625         }
4626 out:
4627         return ret;
4628 }
4629
4630 static struct ctl_table hugetlb_table[] = {
4631         {
4632                 .procname       = "nr_hugepages",
4633                 .data           = NULL,
4634                 .maxlen         = sizeof(unsigned long),
4635                 .mode           = 0644,
4636                 .proc_handler   = hugetlb_sysctl_handler,
4637         },
4638 #ifdef CONFIG_NUMA
4639         {
4640                 .procname       = "nr_hugepages_mempolicy",
4641                 .data           = NULL,
4642                 .maxlen         = sizeof(unsigned long),
4643                 .mode           = 0644,
4644                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
4645         },
4646 #endif
4647         {
4648                 .procname       = "hugetlb_shm_group",
4649                 .data           = &sysctl_hugetlb_shm_group,
4650                 .maxlen         = sizeof(gid_t),
4651                 .mode           = 0644,
4652                 .proc_handler   = proc_dointvec,
4653         },
4654         {
4655                 .procname       = "nr_overcommit_hugepages",
4656                 .data           = NULL,
4657                 .maxlen         = sizeof(unsigned long),
4658                 .mode           = 0644,
4659                 .proc_handler   = hugetlb_overcommit_handler,
4660         },
4661         { }
4662 };
4663
4664 static void hugetlb_sysctl_init(void)
4665 {
4666         register_sysctl_init("vm", hugetlb_table);
4667 }
4668 #endif /* CONFIG_SYSCTL */
4669
4670 void hugetlb_report_meminfo(struct seq_file *m)
4671 {
4672         struct hstate *h;
4673         unsigned long total = 0;
4674
4675         if (!hugepages_supported())
4676                 return;
4677
4678         for_each_hstate(h) {
4679                 unsigned long count = h->nr_huge_pages;
4680
4681                 total += huge_page_size(h) * count;
4682
4683                 if (h == &default_hstate)
4684                         seq_printf(m,
4685                                    "HugePages_Total:   %5lu\n"
4686                                    "HugePages_Free:    %5lu\n"
4687                                    "HugePages_Rsvd:    %5lu\n"
4688                                    "HugePages_Surp:    %5lu\n"
4689                                    "Hugepagesize:   %8lu kB\n",
4690                                    count,
4691                                    h->free_huge_pages,
4692                                    h->resv_huge_pages,
4693                                    h->surplus_huge_pages,
4694                                    huge_page_size(h) / SZ_1K);
4695         }
4696
4697         seq_printf(m, "Hugetlb:        %8lu kB\n", total / SZ_1K);
4698 }
4699
4700 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4701 {
4702         struct hstate *h = &default_hstate;
4703
4704         if (!hugepages_supported())
4705                 return 0;
4706
4707         return sysfs_emit_at(buf, len,
4708                              "Node %d HugePages_Total: %5u\n"
4709                              "Node %d HugePages_Free:  %5u\n"
4710                              "Node %d HugePages_Surp:  %5u\n",
4711                              nid, h->nr_huge_pages_node[nid],
4712                              nid, h->free_huge_pages_node[nid],
4713                              nid, h->surplus_huge_pages_node[nid]);
4714 }
4715
4716 void hugetlb_show_meminfo_node(int nid)
4717 {
4718         struct hstate *h;
4719
4720         if (!hugepages_supported())
4721                 return;
4722
4723         for_each_hstate(h)
4724                 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4725                         nid,
4726                         h->nr_huge_pages_node[nid],
4727                         h->free_huge_pages_node[nid],
4728                         h->surplus_huge_pages_node[nid],
4729                         huge_page_size(h) / SZ_1K);
4730 }
4731
4732 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4733 {
4734         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4735                    K(atomic_long_read(&mm->hugetlb_usage)));
4736 }
4737
4738 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4739 unsigned long hugetlb_total_pages(void)
4740 {
4741         struct hstate *h;
4742         unsigned long nr_total_pages = 0;
4743
4744         for_each_hstate(h)
4745                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4746         return nr_total_pages;
4747 }
4748
4749 static int hugetlb_acct_memory(struct hstate *h, long delta)
4750 {
4751         int ret = -ENOMEM;
4752
4753         if (!delta)
4754                 return 0;
4755
4756         spin_lock_irq(&hugetlb_lock);
4757         /*
4758          * When cpuset is configured, it breaks the strict hugetlb page
4759          * reservation as the accounting is done on a global variable. Such
4760          * reservation is completely rubbish in the presence of cpuset because
4761          * the reservation is not checked against page availability for the
4762          * current cpuset. Application can still potentially OOM'ed by kernel
4763          * with lack of free htlb page in cpuset that the task is in.
4764          * Attempt to enforce strict accounting with cpuset is almost
4765          * impossible (or too ugly) because cpuset is too fluid that
4766          * task or memory node can be dynamically moved between cpusets.
4767          *
4768          * The change of semantics for shared hugetlb mapping with cpuset is
4769          * undesirable. However, in order to preserve some of the semantics,
4770          * we fall back to check against current free page availability as
4771          * a best attempt and hopefully to minimize the impact of changing
4772          * semantics that cpuset has.
4773          *
4774          * Apart from cpuset, we also have memory policy mechanism that
4775          * also determines from which node the kernel will allocate memory
4776          * in a NUMA system. So similar to cpuset, we also should consider
4777          * the memory policy of the current task. Similar to the description
4778          * above.
4779          */
4780         if (delta > 0) {
4781                 if (gather_surplus_pages(h, delta) < 0)
4782                         goto out;
4783
4784                 if (delta > allowed_mems_nr(h)) {
4785                         return_unused_surplus_pages(h, delta);
4786                         goto out;
4787                 }
4788         }
4789
4790         ret = 0;
4791         if (delta < 0)
4792                 return_unused_surplus_pages(h, (unsigned long) -delta);
4793
4794 out:
4795         spin_unlock_irq(&hugetlb_lock);
4796         return ret;
4797 }
4798
4799 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4800 {
4801         struct resv_map *resv = vma_resv_map(vma);
4802
4803         /*
4804          * HPAGE_RESV_OWNER indicates a private mapping.
4805          * This new VMA should share its siblings reservation map if present.
4806          * The VMA will only ever have a valid reservation map pointer where
4807          * it is being copied for another still existing VMA.  As that VMA
4808          * has a reference to the reservation map it cannot disappear until
4809          * after this open call completes.  It is therefore safe to take a
4810          * new reference here without additional locking.
4811          */
4812         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4813                 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4814                 kref_get(&resv->refs);
4815         }
4816
4817         /*
4818          * vma_lock structure for sharable mappings is vma specific.
4819          * Clear old pointer (if copied via vm_area_dup) and allocate
4820          * new structure.  Before clearing, make sure vma_lock is not
4821          * for this vma.
4822          */
4823         if (vma->vm_flags & VM_MAYSHARE) {
4824                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4825
4826                 if (vma_lock) {
4827                         if (vma_lock->vma != vma) {
4828                                 vma->vm_private_data = NULL;
4829                                 hugetlb_vma_lock_alloc(vma);
4830                         } else
4831                                 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4832                 } else
4833                         hugetlb_vma_lock_alloc(vma);
4834         }
4835 }
4836
4837 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4838 {
4839         struct hstate *h = hstate_vma(vma);
4840         struct resv_map *resv;
4841         struct hugepage_subpool *spool = subpool_vma(vma);
4842         unsigned long reserve, start, end;
4843         long gbl_reserve;
4844
4845         hugetlb_vma_lock_free(vma);
4846
4847         resv = vma_resv_map(vma);
4848         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4849                 return;
4850
4851         start = vma_hugecache_offset(h, vma, vma->vm_start);
4852         end = vma_hugecache_offset(h, vma, vma->vm_end);
4853
4854         reserve = (end - start) - region_count(resv, start, end);
4855         hugetlb_cgroup_uncharge_counter(resv, start, end);
4856         if (reserve) {
4857                 /*
4858                  * Decrement reserve counts.  The global reserve count may be
4859                  * adjusted if the subpool has a minimum size.
4860                  */
4861                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4862                 hugetlb_acct_memory(h, -gbl_reserve);
4863         }
4864
4865         kref_put(&resv->refs, resv_map_release);
4866 }
4867
4868 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4869 {
4870         if (addr & ~(huge_page_mask(hstate_vma(vma))))
4871                 return -EINVAL;
4872
4873         /*
4874          * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4875          * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4876          * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4877          */
4878         if (addr & ~PUD_MASK) {
4879                 /*
4880                  * hugetlb_vm_op_split is called right before we attempt to
4881                  * split the VMA. We will need to unshare PMDs in the old and
4882                  * new VMAs, so let's unshare before we split.
4883                  */
4884                 unsigned long floor = addr & PUD_MASK;
4885                 unsigned long ceil = floor + PUD_SIZE;
4886
4887                 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4888                         hugetlb_unshare_pmds(vma, floor, ceil);
4889         }
4890
4891         return 0;
4892 }
4893
4894 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4895 {
4896         return huge_page_size(hstate_vma(vma));
4897 }
4898
4899 /*
4900  * We cannot handle pagefaults against hugetlb pages at all.  They cause
4901  * handle_mm_fault() to try to instantiate regular-sized pages in the
4902  * hugepage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
4903  * this far.
4904  */
4905 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4906 {
4907         BUG();
4908         return 0;
4909 }
4910
4911 /*
4912  * When a new function is introduced to vm_operations_struct and added
4913  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4914  * This is because under System V memory model, mappings created via
4915  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4916  * their original vm_ops are overwritten with shm_vm_ops.
4917  */
4918 const struct vm_operations_struct hugetlb_vm_ops = {
4919         .fault = hugetlb_vm_op_fault,
4920         .open = hugetlb_vm_op_open,
4921         .close = hugetlb_vm_op_close,
4922         .may_split = hugetlb_vm_op_split,
4923         .pagesize = hugetlb_vm_op_pagesize,
4924 };
4925
4926 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4927                                 int writable)
4928 {
4929         pte_t entry;
4930         unsigned int shift = huge_page_shift(hstate_vma(vma));
4931
4932         if (writable) {
4933                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4934                                          vma->vm_page_prot)));
4935         } else {
4936                 entry = huge_pte_wrprotect(mk_huge_pte(page,
4937                                            vma->vm_page_prot));
4938         }
4939         entry = pte_mkyoung(entry);
4940         entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4941
4942         return entry;
4943 }
4944
4945 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4946                                    unsigned long address, pte_t *ptep)
4947 {
4948         pte_t entry;
4949
4950         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4951         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4952                 update_mmu_cache(vma, address, ptep);
4953 }
4954
4955 bool is_hugetlb_entry_migration(pte_t pte)
4956 {
4957         swp_entry_t swp;
4958
4959         if (huge_pte_none(pte) || pte_present(pte))
4960                 return false;
4961         swp = pte_to_swp_entry(pte);
4962         if (is_migration_entry(swp))
4963                 return true;
4964         else
4965                 return false;
4966 }
4967
4968 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4969 {
4970         swp_entry_t swp;
4971
4972         if (huge_pte_none(pte) || pte_present(pte))
4973                 return false;
4974         swp = pte_to_swp_entry(pte);
4975         if (is_hwpoison_entry(swp))
4976                 return true;
4977         else
4978                 return false;
4979 }
4980
4981 static void
4982 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4983                       struct folio *new_folio, pte_t old, unsigned long sz)
4984 {
4985         pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
4986
4987         __folio_mark_uptodate(new_folio);
4988         hugepage_add_new_anon_rmap(new_folio, vma, addr);
4989         if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
4990                 newpte = huge_pte_mkuffd_wp(newpte);
4991         set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
4992         hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4993         folio_set_hugetlb_migratable(new_folio);
4994 }
4995
4996 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4997                             struct vm_area_struct *dst_vma,
4998                             struct vm_area_struct *src_vma)
4999 {
5000         pte_t *src_pte, *dst_pte, entry;
5001         struct folio *pte_folio;
5002         unsigned long addr;
5003         bool cow = is_cow_mapping(src_vma->vm_flags);
5004         struct hstate *h = hstate_vma(src_vma);
5005         unsigned long sz = huge_page_size(h);
5006         unsigned long npages = pages_per_huge_page(h);
5007         struct mmu_notifier_range range;
5008         unsigned long last_addr_mask;
5009         int ret = 0;
5010
5011         if (cow) {
5012                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5013                                         src_vma->vm_start,
5014                                         src_vma->vm_end);
5015                 mmu_notifier_invalidate_range_start(&range);
5016                 vma_assert_write_locked(src_vma);
5017                 raw_write_seqcount_begin(&src->write_protect_seq);
5018         } else {
5019                 /*
5020                  * For shared mappings the vma lock must be held before
5021                  * calling hugetlb_walk() in the src vma. Otherwise, the
5022                  * returned ptep could go away if part of a shared pmd and
5023                  * another thread calls huge_pmd_unshare.
5024                  */
5025                 hugetlb_vma_lock_read(src_vma);
5026         }
5027
5028         last_addr_mask = hugetlb_mask_last_page(h);
5029         for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5030                 spinlock_t *src_ptl, *dst_ptl;
5031                 src_pte = hugetlb_walk(src_vma, addr, sz);
5032                 if (!src_pte) {
5033                         addr |= last_addr_mask;
5034                         continue;
5035                 }
5036                 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5037                 if (!dst_pte) {
5038                         ret = -ENOMEM;
5039                         break;
5040                 }
5041
5042                 /*
5043                  * If the pagetables are shared don't copy or take references.
5044                  *
5045                  * dst_pte == src_pte is the common case of src/dest sharing.
5046                  * However, src could have 'unshared' and dst shares with
5047                  * another vma. So page_count of ptep page is checked instead
5048                  * to reliably determine whether pte is shared.
5049                  */
5050                 if (page_count(virt_to_page(dst_pte)) > 1) {
5051                         addr |= last_addr_mask;
5052                         continue;
5053                 }
5054
5055                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5056                 src_ptl = huge_pte_lockptr(h, src, src_pte);
5057                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5058                 entry = huge_ptep_get(src_pte);
5059 again:
5060                 if (huge_pte_none(entry)) {
5061                         /*
5062                          * Skip if src entry none.
5063                          */
5064                         ;
5065                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5066                         if (!userfaultfd_wp(dst_vma))
5067                                 entry = huge_pte_clear_uffd_wp(entry);
5068                         set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5069                 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5070                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
5071                         bool uffd_wp = pte_swp_uffd_wp(entry);
5072
5073                         if (!is_readable_migration_entry(swp_entry) && cow) {
5074                                 /*
5075                                  * COW mappings require pages in both
5076                                  * parent and child to be set to read.
5077                                  */
5078                                 swp_entry = make_readable_migration_entry(
5079                                                         swp_offset(swp_entry));
5080                                 entry = swp_entry_to_pte(swp_entry);
5081                                 if (userfaultfd_wp(src_vma) && uffd_wp)
5082                                         entry = pte_swp_mkuffd_wp(entry);
5083                                 set_huge_pte_at(src, addr, src_pte, entry, sz);
5084                         }
5085                         if (!userfaultfd_wp(dst_vma))
5086                                 entry = huge_pte_clear_uffd_wp(entry);
5087                         set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5088                 } else if (unlikely(is_pte_marker(entry))) {
5089                         pte_marker marker = copy_pte_marker(
5090                                 pte_to_swp_entry(entry), dst_vma);
5091
5092                         if (marker)
5093                                 set_huge_pte_at(dst, addr, dst_pte,
5094                                                 make_pte_marker(marker), sz);
5095                 } else {
5096                         entry = huge_ptep_get(src_pte);
5097                         pte_folio = page_folio(pte_page(entry));
5098                         folio_get(pte_folio);
5099
5100                         /*
5101                          * Failing to duplicate the anon rmap is a rare case
5102                          * where we see pinned hugetlb pages while they're
5103                          * prone to COW. We need to do the COW earlier during
5104                          * fork.
5105                          *
5106                          * When pre-allocating the page or copying data, we
5107                          * need to be without the pgtable locks since we could
5108                          * sleep during the process.
5109                          */
5110                         if (!folio_test_anon(pte_folio)) {
5111                                 page_dup_file_rmap(&pte_folio->page, true);
5112                         } else if (page_try_dup_anon_rmap(&pte_folio->page,
5113                                                           true, src_vma)) {
5114                                 pte_t src_pte_old = entry;
5115                                 struct folio *new_folio;
5116
5117                                 spin_unlock(src_ptl);
5118                                 spin_unlock(dst_ptl);
5119                                 /* Do not use reserve as it's private owned */
5120                                 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5121                                 if (IS_ERR(new_folio)) {
5122                                         folio_put(pte_folio);
5123                                         ret = PTR_ERR(new_folio);
5124                                         break;
5125                                 }
5126                                 ret = copy_user_large_folio(new_folio,
5127                                                             pte_folio,
5128                                                             addr, dst_vma);
5129                                 folio_put(pte_folio);
5130                                 if (ret) {
5131                                         folio_put(new_folio);
5132                                         break;
5133                                 }
5134
5135                                 /* Install the new hugetlb folio if src pte stable */
5136                                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5137                                 src_ptl = huge_pte_lockptr(h, src, src_pte);
5138                                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5139                                 entry = huge_ptep_get(src_pte);
5140                                 if (!pte_same(src_pte_old, entry)) {
5141                                         restore_reserve_on_error(h, dst_vma, addr,
5142                                                                 new_folio);
5143                                         folio_put(new_folio);
5144                                         /* huge_ptep of dst_pte won't change as in child */
5145                                         goto again;
5146                                 }
5147                                 hugetlb_install_folio(dst_vma, dst_pte, addr,
5148                                                       new_folio, src_pte_old, sz);
5149                                 spin_unlock(src_ptl);
5150                                 spin_unlock(dst_ptl);
5151                                 continue;
5152                         }
5153
5154                         if (cow) {
5155                                 /*
5156                                  * No need to notify as we are downgrading page
5157                                  * table protection not changing it to point
5158                                  * to a new page.
5159                                  *
5160                                  * See Documentation/mm/mmu_notifier.rst
5161                                  */
5162                                 huge_ptep_set_wrprotect(src, addr, src_pte);
5163                                 entry = huge_pte_wrprotect(entry);
5164                         }
5165
5166                         if (!userfaultfd_wp(dst_vma))
5167                                 entry = huge_pte_clear_uffd_wp(entry);
5168
5169                         set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5170                         hugetlb_count_add(npages, dst);
5171                 }
5172                 spin_unlock(src_ptl);
5173                 spin_unlock(dst_ptl);
5174         }
5175
5176         if (cow) {
5177                 raw_write_seqcount_end(&src->write_protect_seq);
5178                 mmu_notifier_invalidate_range_end(&range);
5179         } else {
5180                 hugetlb_vma_unlock_read(src_vma);
5181         }
5182
5183         return ret;
5184 }
5185
5186 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5187                           unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5188                           unsigned long sz)
5189 {
5190         struct hstate *h = hstate_vma(vma);
5191         struct mm_struct *mm = vma->vm_mm;
5192         spinlock_t *src_ptl, *dst_ptl;
5193         pte_t pte;
5194
5195         dst_ptl = huge_pte_lock(h, mm, dst_pte);
5196         src_ptl = huge_pte_lockptr(h, mm, src_pte);
5197
5198         /*
5199          * We don't have to worry about the ordering of src and dst ptlocks
5200          * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5201          */
5202         if (src_ptl != dst_ptl)
5203                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5204
5205         pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5206         set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5207
5208         if (src_ptl != dst_ptl)
5209                 spin_unlock(src_ptl);
5210         spin_unlock(dst_ptl);
5211 }
5212
5213 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5214                              struct vm_area_struct *new_vma,
5215                              unsigned long old_addr, unsigned long new_addr,
5216                              unsigned long len)
5217 {
5218         struct hstate *h = hstate_vma(vma);
5219         struct address_space *mapping = vma->vm_file->f_mapping;
5220         unsigned long sz = huge_page_size(h);
5221         struct mm_struct *mm = vma->vm_mm;
5222         unsigned long old_end = old_addr + len;
5223         unsigned long last_addr_mask;
5224         pte_t *src_pte, *dst_pte;
5225         struct mmu_notifier_range range;
5226         bool shared_pmd = false;
5227
5228         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5229                                 old_end);
5230         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5231         /*
5232          * In case of shared PMDs, we should cover the maximum possible
5233          * range.
5234          */
5235         flush_cache_range(vma, range.start, range.end);
5236
5237         mmu_notifier_invalidate_range_start(&range);
5238         last_addr_mask = hugetlb_mask_last_page(h);
5239         /* Prevent race with file truncation */
5240         hugetlb_vma_lock_write(vma);
5241         i_mmap_lock_write(mapping);
5242         for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5243                 src_pte = hugetlb_walk(vma, old_addr, sz);
5244                 if (!src_pte) {
5245                         old_addr |= last_addr_mask;
5246                         new_addr |= last_addr_mask;
5247                         continue;
5248                 }
5249                 if (huge_pte_none(huge_ptep_get(src_pte)))
5250                         continue;
5251
5252                 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5253                         shared_pmd = true;
5254                         old_addr |= last_addr_mask;
5255                         new_addr |= last_addr_mask;
5256                         continue;
5257                 }
5258
5259                 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5260                 if (!dst_pte)
5261                         break;
5262
5263                 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5264         }
5265
5266         if (shared_pmd)
5267                 flush_hugetlb_tlb_range(vma, range.start, range.end);
5268         else
5269                 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5270         mmu_notifier_invalidate_range_end(&range);
5271         i_mmap_unlock_write(mapping);
5272         hugetlb_vma_unlock_write(vma);
5273
5274         return len + old_addr - old_end;
5275 }
5276
5277 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5278                                    unsigned long start, unsigned long end,
5279                                    struct page *ref_page, zap_flags_t zap_flags)
5280 {
5281         struct mm_struct *mm = vma->vm_mm;
5282         unsigned long address;
5283         pte_t *ptep;
5284         pte_t pte;
5285         spinlock_t *ptl;
5286         struct page *page;
5287         struct hstate *h = hstate_vma(vma);
5288         unsigned long sz = huge_page_size(h);
5289         unsigned long last_addr_mask;
5290         bool force_flush = false;
5291
5292         WARN_ON(!is_vm_hugetlb_page(vma));
5293         BUG_ON(start & ~huge_page_mask(h));
5294         BUG_ON(end & ~huge_page_mask(h));
5295
5296         /*
5297          * This is a hugetlb vma, all the pte entries should point
5298          * to huge page.
5299          */
5300         tlb_change_page_size(tlb, sz);
5301         tlb_start_vma(tlb, vma);
5302
5303         last_addr_mask = hugetlb_mask_last_page(h);
5304         address = start;
5305         for (; address < end; address += sz) {
5306                 ptep = hugetlb_walk(vma, address, sz);
5307                 if (!ptep) {
5308                         address |= last_addr_mask;
5309                         continue;
5310                 }
5311
5312                 ptl = huge_pte_lock(h, mm, ptep);
5313                 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5314                         spin_unlock(ptl);
5315                         tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5316                         force_flush = true;
5317                         address |= last_addr_mask;
5318                         continue;
5319                 }
5320
5321                 pte = huge_ptep_get(ptep);
5322                 if (huge_pte_none(pte)) {
5323                         spin_unlock(ptl);
5324                         continue;
5325                 }
5326
5327                 /*
5328                  * Migrating hugepage or HWPoisoned hugepage is already
5329                  * unmapped and its refcount is dropped, so just clear pte here.
5330                  */
5331                 if (unlikely(!pte_present(pte))) {
5332                         /*
5333                          * If the pte was wr-protected by uffd-wp in any of the
5334                          * swap forms, meanwhile the caller does not want to
5335                          * drop the uffd-wp bit in this zap, then replace the
5336                          * pte with a marker.
5337                          */
5338                         if (pte_swp_uffd_wp_any(pte) &&
5339                             !(zap_flags & ZAP_FLAG_DROP_MARKER))
5340                                 set_huge_pte_at(mm, address, ptep,
5341                                                 make_pte_marker(PTE_MARKER_UFFD_WP),
5342                                                 sz);
5343                         else
5344                                 huge_pte_clear(mm, address, ptep, sz);
5345                         spin_unlock(ptl);
5346                         continue;
5347                 }
5348
5349                 page = pte_page(pte);
5350                 /*
5351                  * If a reference page is supplied, it is because a specific
5352                  * page is being unmapped, not a range. Ensure the page we
5353                  * are about to unmap is the actual page of interest.
5354                  */
5355                 if (ref_page) {
5356                         if (page != ref_page) {
5357                                 spin_unlock(ptl);
5358                                 continue;
5359                         }
5360                         /*
5361                          * Mark the VMA as having unmapped its page so that
5362                          * future faults in this VMA will fail rather than
5363                          * looking like data was lost
5364                          */
5365                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5366                 }
5367
5368                 pte = huge_ptep_get_and_clear(mm, address, ptep);
5369                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5370                 if (huge_pte_dirty(pte))
5371                         set_page_dirty(page);
5372                 /* Leave a uffd-wp pte marker if needed */
5373                 if (huge_pte_uffd_wp(pte) &&
5374                     !(zap_flags & ZAP_FLAG_DROP_MARKER))
5375                         set_huge_pte_at(mm, address, ptep,
5376                                         make_pte_marker(PTE_MARKER_UFFD_WP),
5377                                         sz);
5378                 hugetlb_count_sub(pages_per_huge_page(h), mm);
5379                 page_remove_rmap(page, vma, true);
5380
5381                 spin_unlock(ptl);
5382                 tlb_remove_page_size(tlb, page, huge_page_size(h));
5383                 /*
5384                  * Bail out after unmapping reference page if supplied
5385                  */
5386                 if (ref_page)
5387                         break;
5388         }
5389         tlb_end_vma(tlb, vma);
5390
5391         /*
5392          * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5393          * could defer the flush until now, since by holding i_mmap_rwsem we
5394          * guaranteed that the last refernece would not be dropped. But we must
5395          * do the flushing before we return, as otherwise i_mmap_rwsem will be
5396          * dropped and the last reference to the shared PMDs page might be
5397          * dropped as well.
5398          *
5399          * In theory we could defer the freeing of the PMD pages as well, but
5400          * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5401          * detect sharing, so we cannot defer the release of the page either.
5402          * Instead, do flush now.
5403          */
5404         if (force_flush)
5405                 tlb_flush_mmu_tlbonly(tlb);
5406 }
5407
5408 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5409                           struct vm_area_struct *vma, unsigned long start,
5410                           unsigned long end, struct page *ref_page,
5411                           zap_flags_t zap_flags)
5412 {
5413         hugetlb_vma_lock_write(vma);
5414         i_mmap_lock_write(vma->vm_file->f_mapping);
5415
5416         /* mmu notification performed in caller */
5417         __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5418
5419         if (zap_flags & ZAP_FLAG_UNMAP) {       /* final unmap */
5420                 /*
5421                  * Unlock and free the vma lock before releasing i_mmap_rwsem.
5422                  * When the vma_lock is freed, this makes the vma ineligible
5423                  * for pmd sharing.  And, i_mmap_rwsem is required to set up
5424                  * pmd sharing.  This is important as page tables for this
5425                  * unmapped range will be asynchrously deleted.  If the page
5426                  * tables are shared, there will be issues when accessed by
5427                  * someone else.
5428                  */
5429                 __hugetlb_vma_unlock_write_free(vma);
5430                 i_mmap_unlock_write(vma->vm_file->f_mapping);
5431         } else {
5432                 i_mmap_unlock_write(vma->vm_file->f_mapping);
5433                 hugetlb_vma_unlock_write(vma);
5434         }
5435 }
5436
5437 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5438                           unsigned long end, struct page *ref_page,
5439                           zap_flags_t zap_flags)
5440 {
5441         struct mmu_notifier_range range;
5442         struct mmu_gather tlb;
5443
5444         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5445                                 start, end);
5446         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5447         mmu_notifier_invalidate_range_start(&range);
5448         tlb_gather_mmu(&tlb, vma->vm_mm);
5449
5450         __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5451
5452         mmu_notifier_invalidate_range_end(&range);
5453         tlb_finish_mmu(&tlb);
5454 }
5455
5456 /*
5457  * This is called when the original mapper is failing to COW a MAP_PRIVATE
5458  * mapping it owns the reserve page for. The intention is to unmap the page
5459  * from other VMAs and let the children be SIGKILLed if they are faulting the
5460  * same region.
5461  */
5462 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5463                               struct page *page, unsigned long address)
5464 {
5465         struct hstate *h = hstate_vma(vma);
5466         struct vm_area_struct *iter_vma;
5467         struct address_space *mapping;
5468         pgoff_t pgoff;
5469
5470         /*
5471          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5472          * from page cache lookup which is in HPAGE_SIZE units.
5473          */
5474         address = address & huge_page_mask(h);
5475         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5476                         vma->vm_pgoff;
5477         mapping = vma->vm_file->f_mapping;
5478
5479         /*
5480          * Take the mapping lock for the duration of the table walk. As
5481          * this mapping should be shared between all the VMAs,
5482          * __unmap_hugepage_range() is called as the lock is already held
5483          */
5484         i_mmap_lock_write(mapping);
5485         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5486                 /* Do not unmap the current VMA */
5487                 if (iter_vma == vma)
5488                         continue;
5489
5490                 /*
5491                  * Shared VMAs have their own reserves and do not affect
5492                  * MAP_PRIVATE accounting but it is possible that a shared
5493                  * VMA is using the same page so check and skip such VMAs.
5494                  */
5495                 if (iter_vma->vm_flags & VM_MAYSHARE)
5496                         continue;
5497
5498                 /*
5499                  * Unmap the page from other VMAs without their own reserves.
5500                  * They get marked to be SIGKILLed if they fault in these
5501                  * areas. This is because a future no-page fault on this VMA
5502                  * could insert a zeroed page instead of the data existing
5503                  * from the time of fork. This would look like data corruption
5504                  */
5505                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5506                         unmap_hugepage_range(iter_vma, address,
5507                                              address + huge_page_size(h), page, 0);
5508         }
5509         i_mmap_unlock_write(mapping);
5510 }
5511
5512 /*
5513  * hugetlb_wp() should be called with page lock of the original hugepage held.
5514  * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5515  * cannot race with other handlers or page migration.
5516  * Keep the pte_same checks anyway to make transition from the mutex easier.
5517  */
5518 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5519                        unsigned long address, pte_t *ptep, unsigned int flags,
5520                        struct folio *pagecache_folio, spinlock_t *ptl)
5521 {
5522         const bool unshare = flags & FAULT_FLAG_UNSHARE;
5523         pte_t pte = huge_ptep_get(ptep);
5524         struct hstate *h = hstate_vma(vma);
5525         struct folio *old_folio;
5526         struct folio *new_folio;
5527         int outside_reserve = 0;
5528         vm_fault_t ret = 0;
5529         unsigned long haddr = address & huge_page_mask(h);
5530         struct mmu_notifier_range range;
5531
5532         /*
5533          * Never handle CoW for uffd-wp protected pages.  It should be only
5534          * handled when the uffd-wp protection is removed.
5535          *
5536          * Note that only the CoW optimization path (in hugetlb_no_page())
5537          * can trigger this, because hugetlb_fault() will always resolve
5538          * uffd-wp bit first.
5539          */
5540         if (!unshare && huge_pte_uffd_wp(pte))
5541                 return 0;
5542
5543         /*
5544          * hugetlb does not support FOLL_FORCE-style write faults that keep the
5545          * PTE mapped R/O such as maybe_mkwrite() would do.
5546          */
5547         if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5548                 return VM_FAULT_SIGSEGV;
5549
5550         /* Let's take out MAP_SHARED mappings first. */
5551         if (vma->vm_flags & VM_MAYSHARE) {
5552                 set_huge_ptep_writable(vma, haddr, ptep);
5553                 return 0;
5554         }
5555
5556         old_folio = page_folio(pte_page(pte));
5557
5558         delayacct_wpcopy_start();
5559
5560 retry_avoidcopy:
5561         /*
5562          * If no-one else is actually using this page, we're the exclusive
5563          * owner and can reuse this page.
5564          */
5565         if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5566                 if (!PageAnonExclusive(&old_folio->page))
5567                         page_move_anon_rmap(&old_folio->page, vma);
5568                 if (likely(!unshare))
5569                         set_huge_ptep_writable(vma, haddr, ptep);
5570
5571                 delayacct_wpcopy_end();
5572                 return 0;
5573         }
5574         VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5575                        PageAnonExclusive(&old_folio->page), &old_folio->page);
5576
5577         /*
5578          * If the process that created a MAP_PRIVATE mapping is about to
5579          * perform a COW due to a shared page count, attempt to satisfy
5580          * the allocation without using the existing reserves. The pagecache
5581          * page is used to determine if the reserve at this address was
5582          * consumed or not. If reserves were used, a partial faulted mapping
5583          * at the time of fork() could consume its reserves on COW instead
5584          * of the full address range.
5585          */
5586         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5587                         old_folio != pagecache_folio)
5588                 outside_reserve = 1;
5589
5590         folio_get(old_folio);
5591
5592         /*
5593          * Drop page table lock as buddy allocator may be called. It will
5594          * be acquired again before returning to the caller, as expected.
5595          */
5596         spin_unlock(ptl);
5597         new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
5598
5599         if (IS_ERR(new_folio)) {
5600                 /*
5601                  * If a process owning a MAP_PRIVATE mapping fails to COW,
5602                  * it is due to references held by a child and an insufficient
5603                  * huge page pool. To guarantee the original mappers
5604                  * reliability, unmap the page from child processes. The child
5605                  * may get SIGKILLed if it later faults.
5606                  */
5607                 if (outside_reserve) {
5608                         struct address_space *mapping = vma->vm_file->f_mapping;
5609                         pgoff_t idx;
5610                         u32 hash;
5611
5612                         folio_put(old_folio);
5613                         /*
5614                          * Drop hugetlb_fault_mutex and vma_lock before
5615                          * unmapping.  unmapping needs to hold vma_lock
5616                          * in write mode.  Dropping vma_lock in read mode
5617                          * here is OK as COW mappings do not interact with
5618                          * PMD sharing.
5619                          *
5620                          * Reacquire both after unmap operation.
5621                          */
5622                         idx = vma_hugecache_offset(h, vma, haddr);
5623                         hash = hugetlb_fault_mutex_hash(mapping, idx);
5624                         hugetlb_vma_unlock_read(vma);
5625                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5626
5627                         unmap_ref_private(mm, vma, &old_folio->page, haddr);
5628
5629                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
5630                         hugetlb_vma_lock_read(vma);
5631                         spin_lock(ptl);
5632                         ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5633                         if (likely(ptep &&
5634                                    pte_same(huge_ptep_get(ptep), pte)))
5635                                 goto retry_avoidcopy;
5636                         /*
5637                          * race occurs while re-acquiring page table
5638                          * lock, and our job is done.
5639                          */
5640                         delayacct_wpcopy_end();
5641                         return 0;
5642                 }
5643
5644                 ret = vmf_error(PTR_ERR(new_folio));
5645                 goto out_release_old;
5646         }
5647
5648         /*
5649          * When the original hugepage is shared one, it does not have
5650          * anon_vma prepared.
5651          */
5652         if (unlikely(anon_vma_prepare(vma))) {
5653                 ret = VM_FAULT_OOM;
5654                 goto out_release_all;
5655         }
5656
5657         if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
5658                 ret = VM_FAULT_HWPOISON_LARGE;
5659                 goto out_release_all;
5660         }
5661         __folio_mark_uptodate(new_folio);
5662
5663         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
5664                                 haddr + huge_page_size(h));
5665         mmu_notifier_invalidate_range_start(&range);
5666
5667         /*
5668          * Retake the page table lock to check for racing updates
5669          * before the page tables are altered
5670          */
5671         spin_lock(ptl);
5672         ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5673         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5674                 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
5675
5676                 /* Break COW or unshare */
5677                 huge_ptep_clear_flush(vma, haddr, ptep);
5678                 page_remove_rmap(&old_folio->page, vma, true);
5679                 hugepage_add_new_anon_rmap(new_folio, vma, haddr);
5680                 if (huge_pte_uffd_wp(pte))
5681                         newpte = huge_pte_mkuffd_wp(newpte);
5682                 set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
5683                 folio_set_hugetlb_migratable(new_folio);
5684                 /* Make the old page be freed below */
5685                 new_folio = old_folio;
5686         }
5687         spin_unlock(ptl);
5688         mmu_notifier_invalidate_range_end(&range);
5689 out_release_all:
5690         /*
5691          * No restore in case of successful pagetable update (Break COW or
5692          * unshare)
5693          */
5694         if (new_folio != old_folio)
5695                 restore_reserve_on_error(h, vma, haddr, new_folio);
5696         folio_put(new_folio);
5697 out_release_old:
5698         folio_put(old_folio);
5699
5700         spin_lock(ptl); /* Caller expects lock to be held */
5701
5702         delayacct_wpcopy_end();
5703         return ret;
5704 }
5705
5706 /*
5707  * Return whether there is a pagecache page to back given address within VMA.
5708  */
5709 static bool hugetlbfs_pagecache_present(struct hstate *h,
5710                         struct vm_area_struct *vma, unsigned long address)
5711 {
5712         struct address_space *mapping = vma->vm_file->f_mapping;
5713         pgoff_t idx = vma_hugecache_offset(h, vma, address);
5714         struct folio *folio;
5715
5716         folio = filemap_get_folio(mapping, idx);
5717         if (IS_ERR(folio))
5718                 return false;
5719         folio_put(folio);
5720         return true;
5721 }
5722
5723 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
5724                            pgoff_t idx)
5725 {
5726         struct inode *inode = mapping->host;
5727         struct hstate *h = hstate_inode(inode);
5728         int err;
5729
5730         __folio_set_locked(folio);
5731         err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5732
5733         if (unlikely(err)) {
5734                 __folio_clear_locked(folio);
5735                 return err;
5736         }
5737         folio_clear_hugetlb_restore_reserve(folio);
5738
5739         /*
5740          * mark folio dirty so that it will not be removed from cache/file
5741          * by non-hugetlbfs specific code paths.
5742          */
5743         folio_mark_dirty(folio);
5744
5745         spin_lock(&inode->i_lock);
5746         inode->i_blocks += blocks_per_huge_page(h);
5747         spin_unlock(&inode->i_lock);
5748         return 0;
5749 }
5750
5751 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5752                                                   struct address_space *mapping,
5753                                                   pgoff_t idx,
5754                                                   unsigned int flags,
5755                                                   unsigned long haddr,
5756                                                   unsigned long addr,
5757                                                   unsigned long reason)
5758 {
5759         u32 hash;
5760         struct vm_fault vmf = {
5761                 .vma = vma,
5762                 .address = haddr,
5763                 .real_address = addr,
5764                 .flags = flags,
5765
5766                 /*
5767                  * Hard to debug if it ends up being
5768                  * used by a callee that assumes
5769                  * something about the other
5770                  * uninitialized fields... same as in
5771                  * memory.c
5772                  */
5773         };
5774
5775         /*
5776          * vma_lock and hugetlb_fault_mutex must be dropped before handling
5777          * userfault. Also mmap_lock could be dropped due to handling
5778          * userfault, any vma operation should be careful from here.
5779          */
5780         hugetlb_vma_unlock_read(vma);
5781         hash = hugetlb_fault_mutex_hash(mapping, idx);
5782         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5783         return handle_userfault(&vmf, reason);
5784 }
5785
5786 /*
5787  * Recheck pte with pgtable lock.  Returns true if pte didn't change, or
5788  * false if pte changed or is changing.
5789  */
5790 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5791                                pte_t *ptep, pte_t old_pte)
5792 {
5793         spinlock_t *ptl;
5794         bool same;
5795
5796         ptl = huge_pte_lock(h, mm, ptep);
5797         same = pte_same(huge_ptep_get(ptep), old_pte);
5798         spin_unlock(ptl);
5799
5800         return same;
5801 }
5802
5803 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5804                         struct vm_area_struct *vma,
5805                         struct address_space *mapping, pgoff_t idx,
5806                         unsigned long address, pte_t *ptep,
5807                         pte_t old_pte, unsigned int flags)
5808 {
5809         struct hstate *h = hstate_vma(vma);
5810         vm_fault_t ret = VM_FAULT_SIGBUS;
5811         int anon_rmap = 0;
5812         unsigned long size;
5813         struct folio *folio;
5814         pte_t new_pte;
5815         spinlock_t *ptl;
5816         unsigned long haddr = address & huge_page_mask(h);
5817         bool new_folio, new_pagecache_folio = false;
5818         u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5819
5820         /*
5821          * Currently, we are forced to kill the process in the event the
5822          * original mapper has unmapped pages from the child due to a failed
5823          * COW/unsharing. Warn that such a situation has occurred as it may not
5824          * be obvious.
5825          */
5826         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5827                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5828                            current->pid);
5829                 goto out;
5830         }
5831
5832         /*
5833          * Use page lock to guard against racing truncation
5834          * before we get page_table_lock.
5835          */
5836         new_folio = false;
5837         folio = filemap_lock_folio(mapping, idx);
5838         if (IS_ERR(folio)) {
5839                 size = i_size_read(mapping->host) >> huge_page_shift(h);
5840                 if (idx >= size)
5841                         goto out;
5842                 /* Check for page in userfault range */
5843                 if (userfaultfd_missing(vma)) {
5844                         /*
5845                          * Since hugetlb_no_page() was examining pte
5846                          * without pgtable lock, we need to re-test under
5847                          * lock because the pte may not be stable and could
5848                          * have changed from under us.  Try to detect
5849                          * either changed or during-changing ptes and retry
5850                          * properly when needed.
5851                          *
5852                          * Note that userfaultfd is actually fine with
5853                          * false positives (e.g. caused by pte changed),
5854                          * but not wrong logical events (e.g. caused by
5855                          * reading a pte during changing).  The latter can
5856                          * confuse the userspace, so the strictness is very
5857                          * much preferred.  E.g., MISSING event should
5858                          * never happen on the page after UFFDIO_COPY has
5859                          * correctly installed the page and returned.
5860                          */
5861                         if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5862                                 ret = 0;
5863                                 goto out;
5864                         }
5865
5866                         return hugetlb_handle_userfault(vma, mapping, idx, flags,
5867                                                         haddr, address,
5868                                                         VM_UFFD_MISSING);
5869                 }
5870
5871                 folio = alloc_hugetlb_folio(vma, haddr, 0);
5872                 if (IS_ERR(folio)) {
5873                         /*
5874                          * Returning error will result in faulting task being
5875                          * sent SIGBUS.  The hugetlb fault mutex prevents two
5876                          * tasks from racing to fault in the same page which
5877                          * could result in false unable to allocate errors.
5878                          * Page migration does not take the fault mutex, but
5879                          * does a clear then write of pte's under page table
5880                          * lock.  Page fault code could race with migration,
5881                          * notice the clear pte and try to allocate a page
5882                          * here.  Before returning error, get ptl and make
5883                          * sure there really is no pte entry.
5884                          */
5885                         if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5886                                 ret = vmf_error(PTR_ERR(folio));
5887                         else
5888                                 ret = 0;
5889                         goto out;
5890                 }
5891                 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
5892                 __folio_mark_uptodate(folio);
5893                 new_folio = true;
5894
5895                 if (vma->vm_flags & VM_MAYSHARE) {
5896                         int err = hugetlb_add_to_page_cache(folio, mapping, idx);
5897                         if (err) {
5898                                 /*
5899                                  * err can't be -EEXIST which implies someone
5900                                  * else consumed the reservation since hugetlb
5901                                  * fault mutex is held when add a hugetlb page
5902                                  * to the page cache. So it's safe to call
5903                                  * restore_reserve_on_error() here.
5904                                  */
5905                                 restore_reserve_on_error(h, vma, haddr, folio);
5906                                 folio_put(folio);
5907                                 goto out;
5908                         }
5909                         new_pagecache_folio = true;
5910                 } else {
5911                         folio_lock(folio);
5912                         if (unlikely(anon_vma_prepare(vma))) {
5913                                 ret = VM_FAULT_OOM;
5914                                 goto backout_unlocked;
5915                         }
5916                         anon_rmap = 1;
5917                 }
5918         } else {
5919                 /*
5920                  * If memory error occurs between mmap() and fault, some process
5921                  * don't have hwpoisoned swap entry for errored virtual address.
5922                  * So we need to block hugepage fault by PG_hwpoison bit check.
5923                  */
5924                 if (unlikely(folio_test_hwpoison(folio))) {
5925                         ret = VM_FAULT_HWPOISON_LARGE |
5926                                 VM_FAULT_SET_HINDEX(hstate_index(h));
5927                         goto backout_unlocked;
5928                 }
5929
5930                 /* Check for page in userfault range. */
5931                 if (userfaultfd_minor(vma)) {
5932                         folio_unlock(folio);
5933                         folio_put(folio);
5934                         /* See comment in userfaultfd_missing() block above */
5935                         if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5936                                 ret = 0;
5937                                 goto out;
5938                         }
5939                         return hugetlb_handle_userfault(vma, mapping, idx, flags,
5940                                                         haddr, address,
5941                                                         VM_UFFD_MINOR);
5942                 }
5943         }
5944
5945         /*
5946          * If we are going to COW a private mapping later, we examine the
5947          * pending reservations for this page now. This will ensure that
5948          * any allocations necessary to record that reservation occur outside
5949          * the spinlock.
5950          */
5951         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5952                 if (vma_needs_reservation(h, vma, haddr) < 0) {
5953                         ret = VM_FAULT_OOM;
5954                         goto backout_unlocked;
5955                 }
5956                 /* Just decrements count, does not deallocate */
5957                 vma_end_reservation(h, vma, haddr);
5958         }
5959
5960         ptl = huge_pte_lock(h, mm, ptep);
5961         ret = 0;
5962         /* If pte changed from under us, retry */
5963         if (!pte_same(huge_ptep_get(ptep), old_pte))
5964                 goto backout;
5965
5966         if (anon_rmap)
5967                 hugepage_add_new_anon_rmap(folio, vma, haddr);
5968         else
5969                 page_dup_file_rmap(&folio->page, true);
5970         new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
5971                                 && (vma->vm_flags & VM_SHARED)));
5972         /*
5973          * If this pte was previously wr-protected, keep it wr-protected even
5974          * if populated.
5975          */
5976         if (unlikely(pte_marker_uffd_wp(old_pte)))
5977                 new_pte = huge_pte_mkuffd_wp(new_pte);
5978         set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
5979
5980         hugetlb_count_add(pages_per_huge_page(h), mm);
5981         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5982                 /* Optimization, do the COW without a second fault */
5983                 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl);
5984         }
5985
5986         spin_unlock(ptl);
5987
5988         /*
5989          * Only set hugetlb_migratable in newly allocated pages.  Existing pages
5990          * found in the pagecache may not have hugetlb_migratable if they have
5991          * been isolated for migration.
5992          */
5993         if (new_folio)
5994                 folio_set_hugetlb_migratable(folio);
5995
5996         folio_unlock(folio);
5997 out:
5998         hugetlb_vma_unlock_read(vma);
5999         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6000         return ret;
6001
6002 backout:
6003         spin_unlock(ptl);
6004 backout_unlocked:
6005         if (new_folio && !new_pagecache_folio)
6006                 restore_reserve_on_error(h, vma, haddr, folio);
6007
6008         folio_unlock(folio);
6009         folio_put(folio);
6010         goto out;
6011 }
6012
6013 #ifdef CONFIG_SMP
6014 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6015 {
6016         unsigned long key[2];
6017         u32 hash;
6018
6019         key[0] = (unsigned long) mapping;
6020         key[1] = idx;
6021
6022         hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6023
6024         return hash & (num_fault_mutexes - 1);
6025 }
6026 #else
6027 /*
6028  * For uniprocessor systems we always use a single mutex, so just
6029  * return 0 and avoid the hashing overhead.
6030  */
6031 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6032 {
6033         return 0;
6034 }
6035 #endif
6036
6037 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6038                         unsigned long address, unsigned int flags)
6039 {
6040         pte_t *ptep, entry;
6041         spinlock_t *ptl;
6042         vm_fault_t ret;
6043         u32 hash;
6044         pgoff_t idx;
6045         struct folio *folio = NULL;
6046         struct folio *pagecache_folio = NULL;
6047         struct hstate *h = hstate_vma(vma);
6048         struct address_space *mapping;
6049         int need_wait_lock = 0;
6050         unsigned long haddr = address & huge_page_mask(h);
6051
6052         /* TODO: Handle faults under the VMA lock */
6053         if (flags & FAULT_FLAG_VMA_LOCK) {
6054                 vma_end_read(vma);
6055                 return VM_FAULT_RETRY;
6056         }
6057
6058         /*
6059          * Serialize hugepage allocation and instantiation, so that we don't
6060          * get spurious allocation failures if two CPUs race to instantiate
6061          * the same page in the page cache.
6062          */
6063         mapping = vma->vm_file->f_mapping;
6064         idx = vma_hugecache_offset(h, vma, haddr);
6065         hash = hugetlb_fault_mutex_hash(mapping, idx);
6066         mutex_lock(&hugetlb_fault_mutex_table[hash]);
6067
6068         /*
6069          * Acquire vma lock before calling huge_pte_alloc and hold
6070          * until finished with ptep.  This prevents huge_pmd_unshare from
6071          * being called elsewhere and making the ptep no longer valid.
6072          */
6073         hugetlb_vma_lock_read(vma);
6074         ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6075         if (!ptep) {
6076                 hugetlb_vma_unlock_read(vma);
6077                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6078                 return VM_FAULT_OOM;
6079         }
6080
6081         entry = huge_ptep_get(ptep);
6082         if (huge_pte_none_mostly(entry)) {
6083                 if (is_pte_marker(entry)) {
6084                         pte_marker marker =
6085                                 pte_marker_get(pte_to_swp_entry(entry));
6086
6087                         if (marker & PTE_MARKER_POISONED) {
6088                                 ret = VM_FAULT_HWPOISON_LARGE;
6089                                 goto out_mutex;
6090                         }
6091                 }
6092
6093                 /*
6094                  * Other PTE markers should be handled the same way as none PTE.
6095                  *
6096                  * hugetlb_no_page will drop vma lock and hugetlb fault
6097                  * mutex internally, which make us return immediately.
6098                  */
6099                 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6100                                       entry, flags);
6101         }
6102
6103         ret = 0;
6104
6105         /*
6106          * entry could be a migration/hwpoison entry at this point, so this
6107          * check prevents the kernel from going below assuming that we have
6108          * an active hugepage in pagecache. This goto expects the 2nd page
6109          * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6110          * properly handle it.
6111          */
6112         if (!pte_present(entry)) {
6113                 if (unlikely(is_hugetlb_entry_migration(entry))) {
6114                         /*
6115                          * Release the hugetlb fault lock now, but retain
6116                          * the vma lock, because it is needed to guard the
6117                          * huge_pte_lockptr() later in
6118                          * migration_entry_wait_huge(). The vma lock will
6119                          * be released there.
6120                          */
6121                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6122                         migration_entry_wait_huge(vma, ptep);
6123                         return 0;
6124                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6125                         ret = VM_FAULT_HWPOISON_LARGE |
6126                             VM_FAULT_SET_HINDEX(hstate_index(h));
6127                 goto out_mutex;
6128         }
6129
6130         /*
6131          * If we are going to COW/unshare the mapping later, we examine the
6132          * pending reservations for this page now. This will ensure that any
6133          * allocations necessary to record that reservation occur outside the
6134          * spinlock. Also lookup the pagecache page now as it is used to
6135          * determine if a reservation has been consumed.
6136          */
6137         if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6138             !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6139                 if (vma_needs_reservation(h, vma, haddr) < 0) {
6140                         ret = VM_FAULT_OOM;
6141                         goto out_mutex;
6142                 }
6143                 /* Just decrements count, does not deallocate */
6144                 vma_end_reservation(h, vma, haddr);
6145
6146                 pagecache_folio = filemap_lock_folio(mapping, idx);
6147                 if (IS_ERR(pagecache_folio))
6148                         pagecache_folio = NULL;
6149         }
6150
6151         ptl = huge_pte_lock(h, mm, ptep);
6152
6153         /* Check for a racing update before calling hugetlb_wp() */
6154         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6155                 goto out_ptl;
6156
6157         /* Handle userfault-wp first, before trying to lock more pages */
6158         if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6159             (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6160                 struct vm_fault vmf = {
6161                         .vma = vma,
6162                         .address = haddr,
6163                         .real_address = address,
6164                         .flags = flags,
6165                 };
6166
6167                 spin_unlock(ptl);
6168                 if (pagecache_folio) {
6169                         folio_unlock(pagecache_folio);
6170                         folio_put(pagecache_folio);
6171                 }
6172                 hugetlb_vma_unlock_read(vma);
6173                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6174                 return handle_userfault(&vmf, VM_UFFD_WP);
6175         }
6176
6177         /*
6178          * hugetlb_wp() requires page locks of pte_page(entry) and
6179          * pagecache_folio, so here we need take the former one
6180          * when folio != pagecache_folio or !pagecache_folio.
6181          */
6182         folio = page_folio(pte_page(entry));
6183         if (folio != pagecache_folio)
6184                 if (!folio_trylock(folio)) {
6185                         need_wait_lock = 1;
6186                         goto out_ptl;
6187                 }
6188
6189         folio_get(folio);
6190
6191         if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6192                 if (!huge_pte_write(entry)) {
6193                         ret = hugetlb_wp(mm, vma, address, ptep, flags,
6194                                          pagecache_folio, ptl);
6195                         goto out_put_page;
6196                 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6197                         entry = huge_pte_mkdirty(entry);
6198                 }
6199         }
6200         entry = pte_mkyoung(entry);
6201         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6202                                                 flags & FAULT_FLAG_WRITE))
6203                 update_mmu_cache(vma, haddr, ptep);
6204 out_put_page:
6205         if (folio != pagecache_folio)
6206                 folio_unlock(folio);
6207         folio_put(folio);
6208 out_ptl:
6209         spin_unlock(ptl);
6210
6211         if (pagecache_folio) {
6212                 folio_unlock(pagecache_folio);
6213                 folio_put(pagecache_folio);
6214         }
6215 out_mutex:
6216         hugetlb_vma_unlock_read(vma);
6217         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6218         /*
6219          * Generally it's safe to hold refcount during waiting page lock. But
6220          * here we just wait to defer the next page fault to avoid busy loop and
6221          * the page is not used after unlocked before returning from the current
6222          * page fault. So we are safe from accessing freed page, even if we wait
6223          * here without taking refcount.
6224          */
6225         if (need_wait_lock)
6226                 folio_wait_locked(folio);
6227         return ret;
6228 }
6229
6230 #ifdef CONFIG_USERFAULTFD
6231 /*
6232  * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6233  * with modifications for hugetlb pages.
6234  */
6235 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6236                              struct vm_area_struct *dst_vma,
6237                              unsigned long dst_addr,
6238                              unsigned long src_addr,
6239                              uffd_flags_t flags,
6240                              struct folio **foliop)
6241 {
6242         struct mm_struct *dst_mm = dst_vma->vm_mm;
6243         bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6244         bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6245         struct hstate *h = hstate_vma(dst_vma);
6246         struct address_space *mapping = dst_vma->vm_file->f_mapping;
6247         pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6248         unsigned long size;
6249         int vm_shared = dst_vma->vm_flags & VM_SHARED;
6250         pte_t _dst_pte;
6251         spinlock_t *ptl;
6252         int ret = -ENOMEM;
6253         struct folio *folio;
6254         int writable;
6255         bool folio_in_pagecache = false;
6256
6257         if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6258                 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6259
6260                 /* Don't overwrite any existing PTEs (even markers) */
6261                 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6262                         spin_unlock(ptl);
6263                         return -EEXIST;
6264                 }
6265
6266                 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6267                 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6268                                 huge_page_size(h));
6269
6270                 /* No need to invalidate - it was non-present before */
6271                 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6272
6273                 spin_unlock(ptl);
6274                 return 0;
6275         }
6276
6277         if (is_continue) {
6278                 ret = -EFAULT;
6279                 folio = filemap_lock_folio(mapping, idx);
6280                 if (IS_ERR(folio))
6281                         goto out;
6282                 folio_in_pagecache = true;
6283         } else if (!*foliop) {
6284                 /* If a folio already exists, then it's UFFDIO_COPY for
6285                  * a non-missing case. Return -EEXIST.
6286                  */
6287                 if (vm_shared &&
6288                     hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6289                         ret = -EEXIST;
6290                         goto out;
6291                 }
6292
6293                 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6294                 if (IS_ERR(folio)) {
6295                         ret = -ENOMEM;
6296                         goto out;
6297                 }
6298
6299                 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6300                                            false);
6301
6302                 /* fallback to copy_from_user outside mmap_lock */
6303                 if (unlikely(ret)) {
6304                         ret = -ENOENT;
6305                         /* Free the allocated folio which may have
6306                          * consumed a reservation.
6307                          */
6308                         restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6309                         folio_put(folio);
6310
6311                         /* Allocate a temporary folio to hold the copied
6312                          * contents.
6313                          */
6314                         folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6315                         if (!folio) {
6316                                 ret = -ENOMEM;
6317                                 goto out;
6318                         }
6319                         *foliop = folio;
6320                         /* Set the outparam foliop and return to the caller to
6321                          * copy the contents outside the lock. Don't free the
6322                          * folio.
6323                          */
6324                         goto out;
6325                 }
6326         } else {
6327                 if (vm_shared &&
6328                     hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6329                         folio_put(*foliop);
6330                         ret = -EEXIST;
6331                         *foliop = NULL;
6332                         goto out;
6333                 }
6334
6335                 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6336                 if (IS_ERR(folio)) {
6337                         folio_put(*foliop);
6338                         ret = -ENOMEM;
6339                         *foliop = NULL;
6340                         goto out;
6341                 }
6342                 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6343                 folio_put(*foliop);
6344                 *foliop = NULL;
6345                 if (ret) {
6346                         folio_put(folio);
6347                         goto out;
6348                 }
6349         }
6350
6351         /*
6352          * The memory barrier inside __folio_mark_uptodate makes sure that
6353          * preceding stores to the page contents become visible before
6354          * the set_pte_at() write.
6355          */
6356         __folio_mark_uptodate(folio);
6357
6358         /* Add shared, newly allocated pages to the page cache. */
6359         if (vm_shared && !is_continue) {
6360                 size = i_size_read(mapping->host) >> huge_page_shift(h);
6361                 ret = -EFAULT;
6362                 if (idx >= size)
6363                         goto out_release_nounlock;
6364
6365                 /*
6366                  * Serialization between remove_inode_hugepages() and
6367                  * hugetlb_add_to_page_cache() below happens through the
6368                  * hugetlb_fault_mutex_table that here must be hold by
6369                  * the caller.
6370                  */
6371                 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6372                 if (ret)
6373                         goto out_release_nounlock;
6374                 folio_in_pagecache = true;
6375         }
6376
6377         ptl = huge_pte_lock(h, dst_mm, dst_pte);
6378
6379         ret = -EIO;
6380         if (folio_test_hwpoison(folio))
6381                 goto out_release_unlock;
6382
6383         /*
6384          * We allow to overwrite a pte marker: consider when both MISSING|WP
6385          * registered, we firstly wr-protect a none pte which has no page cache
6386          * page backing it, then access the page.
6387          */
6388         ret = -EEXIST;
6389         if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6390                 goto out_release_unlock;
6391
6392         if (folio_in_pagecache)
6393                 page_dup_file_rmap(&folio->page, true);
6394         else
6395                 hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr);
6396
6397         /*
6398          * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6399          * with wp flag set, don't set pte write bit.
6400          */
6401         if (wp_enabled || (is_continue && !vm_shared))
6402                 writable = 0;
6403         else
6404                 writable = dst_vma->vm_flags & VM_WRITE;
6405
6406         _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6407         /*
6408          * Always mark UFFDIO_COPY page dirty; note that this may not be
6409          * extremely important for hugetlbfs for now since swapping is not
6410          * supported, but we should still be clear in that this page cannot be
6411          * thrown away at will, even if write bit not set.
6412          */
6413         _dst_pte = huge_pte_mkdirty(_dst_pte);
6414         _dst_pte = pte_mkyoung(_dst_pte);
6415
6416         if (wp_enabled)
6417                 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6418
6419         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6420
6421         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6422
6423         /* No need to invalidate - it was non-present before */
6424         update_mmu_cache(dst_vma, dst_addr, dst_pte);
6425
6426         spin_unlock(ptl);
6427         if (!is_continue)
6428                 folio_set_hugetlb_migratable(folio);
6429         if (vm_shared || is_continue)
6430                 folio_unlock(folio);
6431         ret = 0;
6432 out:
6433         return ret;
6434 out_release_unlock:
6435         spin_unlock(ptl);
6436         if (vm_shared || is_continue)
6437                 folio_unlock(folio);
6438 out_release_nounlock:
6439         if (!folio_in_pagecache)
6440                 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6441         folio_put(folio);
6442         goto out;
6443 }
6444 #endif /* CONFIG_USERFAULTFD */
6445
6446 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6447                                       unsigned long address, unsigned int flags,
6448                                       unsigned int *page_mask)
6449 {
6450         struct hstate *h = hstate_vma(vma);
6451         struct mm_struct *mm = vma->vm_mm;
6452         unsigned long haddr = address & huge_page_mask(h);
6453         struct page *page = NULL;
6454         spinlock_t *ptl;
6455         pte_t *pte, entry;
6456         int ret;
6457
6458         hugetlb_vma_lock_read(vma);
6459         pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6460         if (!pte)
6461                 goto out_unlock;
6462
6463         ptl = huge_pte_lock(h, mm, pte);
6464         entry = huge_ptep_get(pte);
6465         if (pte_present(entry)) {
6466                 page = pte_page(entry);
6467
6468                 if (!huge_pte_write(entry)) {
6469                         if (flags & FOLL_WRITE) {
6470                                 page = NULL;
6471                                 goto out;
6472                         }
6473
6474                         if (gup_must_unshare(vma, flags, page)) {
6475                                 /* Tell the caller to do unsharing */
6476                                 page = ERR_PTR(-EMLINK);
6477                                 goto out;
6478                         }
6479                 }
6480
6481                 page += ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
6482
6483                 /*
6484                  * Note that page may be a sub-page, and with vmemmap
6485                  * optimizations the page struct may be read only.
6486                  * try_grab_page() will increase the ref count on the
6487                  * head page, so this will be OK.
6488                  *
6489                  * try_grab_page() should always be able to get the page here,
6490                  * because we hold the ptl lock and have verified pte_present().
6491                  */
6492                 ret = try_grab_page(page, flags);
6493
6494                 if (WARN_ON_ONCE(ret)) {
6495                         page = ERR_PTR(ret);
6496                         goto out;
6497                 }
6498
6499                 *page_mask = (1U << huge_page_order(h)) - 1;
6500         }
6501 out:
6502         spin_unlock(ptl);
6503 out_unlock:
6504         hugetlb_vma_unlock_read(vma);
6505
6506         /*
6507          * Fixup retval for dump requests: if pagecache doesn't exist,
6508          * don't try to allocate a new page but just skip it.
6509          */
6510         if (!page && (flags & FOLL_DUMP) &&
6511             !hugetlbfs_pagecache_present(h, vma, address))
6512                 page = ERR_PTR(-EFAULT);
6513
6514         return page;
6515 }
6516
6517 long hugetlb_change_protection(struct vm_area_struct *vma,
6518                 unsigned long address, unsigned long end,
6519                 pgprot_t newprot, unsigned long cp_flags)
6520 {
6521         struct mm_struct *mm = vma->vm_mm;
6522         unsigned long start = address;
6523         pte_t *ptep;
6524         pte_t pte;
6525         struct hstate *h = hstate_vma(vma);
6526         long pages = 0, psize = huge_page_size(h);
6527         bool shared_pmd = false;
6528         struct mmu_notifier_range range;
6529         unsigned long last_addr_mask;
6530         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6531         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6532
6533         /*
6534          * In the case of shared PMDs, the area to flush could be beyond
6535          * start/end.  Set range.start/range.end to cover the maximum possible
6536          * range if PMD sharing is possible.
6537          */
6538         mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6539                                 0, mm, start, end);
6540         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6541
6542         BUG_ON(address >= end);
6543         flush_cache_range(vma, range.start, range.end);
6544
6545         mmu_notifier_invalidate_range_start(&range);
6546         hugetlb_vma_lock_write(vma);
6547         i_mmap_lock_write(vma->vm_file->f_mapping);
6548         last_addr_mask = hugetlb_mask_last_page(h);
6549         for (; address < end; address += psize) {
6550                 spinlock_t *ptl;
6551                 ptep = hugetlb_walk(vma, address, psize);
6552                 if (!ptep) {
6553                         if (!uffd_wp) {
6554                                 address |= last_addr_mask;
6555                                 continue;
6556                         }
6557                         /*
6558                          * Userfaultfd wr-protect requires pgtable
6559                          * pre-allocations to install pte markers.
6560                          */
6561                         ptep = huge_pte_alloc(mm, vma, address, psize);
6562                         if (!ptep) {
6563                                 pages = -ENOMEM;
6564                                 break;
6565                         }
6566                 }
6567                 ptl = huge_pte_lock(h, mm, ptep);
6568                 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6569                         /*
6570                          * When uffd-wp is enabled on the vma, unshare
6571                          * shouldn't happen at all.  Warn about it if it
6572                          * happened due to some reason.
6573                          */
6574                         WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6575                         pages++;
6576                         spin_unlock(ptl);
6577                         shared_pmd = true;
6578                         address |= last_addr_mask;
6579                         continue;
6580                 }
6581                 pte = huge_ptep_get(ptep);
6582                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6583                         /* Nothing to do. */
6584                 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6585                         swp_entry_t entry = pte_to_swp_entry(pte);
6586                         struct page *page = pfn_swap_entry_to_page(entry);
6587                         pte_t newpte = pte;
6588
6589                         if (is_writable_migration_entry(entry)) {
6590                                 if (PageAnon(page))
6591                                         entry = make_readable_exclusive_migration_entry(
6592                                                                 swp_offset(entry));
6593                                 else
6594                                         entry = make_readable_migration_entry(
6595                                                                 swp_offset(entry));
6596                                 newpte = swp_entry_to_pte(entry);
6597                                 pages++;
6598                         }
6599
6600                         if (uffd_wp)
6601                                 newpte = pte_swp_mkuffd_wp(newpte);
6602                         else if (uffd_wp_resolve)
6603                                 newpte = pte_swp_clear_uffd_wp(newpte);
6604                         if (!pte_same(pte, newpte))
6605                                 set_huge_pte_at(mm, address, ptep, newpte, psize);
6606                 } else if (unlikely(is_pte_marker(pte))) {
6607                         /* No other markers apply for now. */
6608                         WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6609                         if (uffd_wp_resolve)
6610                                 /* Safe to modify directly (non-present->none). */
6611                                 huge_pte_clear(mm, address, ptep, psize);
6612                 } else if (!huge_pte_none(pte)) {
6613                         pte_t old_pte;
6614                         unsigned int shift = huge_page_shift(hstate_vma(vma));
6615
6616                         old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6617                         pte = huge_pte_modify(old_pte, newprot);
6618                         pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6619                         if (uffd_wp)
6620                                 pte = huge_pte_mkuffd_wp(pte);
6621                         else if (uffd_wp_resolve)
6622                                 pte = huge_pte_clear_uffd_wp(pte);
6623                         huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6624                         pages++;
6625                 } else {
6626                         /* None pte */
6627                         if (unlikely(uffd_wp))
6628                                 /* Safe to modify directly (none->non-present). */
6629                                 set_huge_pte_at(mm, address, ptep,
6630                                                 make_pte_marker(PTE_MARKER_UFFD_WP),
6631                                                 psize);
6632                 }
6633                 spin_unlock(ptl);
6634         }
6635         /*
6636          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6637          * may have cleared our pud entry and done put_page on the page table:
6638          * once we release i_mmap_rwsem, another task can do the final put_page
6639          * and that page table be reused and filled with junk.  If we actually
6640          * did unshare a page of pmds, flush the range corresponding to the pud.
6641          */
6642         if (shared_pmd)
6643                 flush_hugetlb_tlb_range(vma, range.start, range.end);
6644         else
6645                 flush_hugetlb_tlb_range(vma, start, end);
6646         /*
6647          * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
6648          * downgrading page table protection not changing it to point to a new
6649          * page.
6650          *
6651          * See Documentation/mm/mmu_notifier.rst
6652          */
6653         i_mmap_unlock_write(vma->vm_file->f_mapping);
6654         hugetlb_vma_unlock_write(vma);
6655         mmu_notifier_invalidate_range_end(&range);
6656
6657         return pages > 0 ? (pages << h->order) : pages;
6658 }
6659
6660 /* Return true if reservation was successful, false otherwise.  */
6661 bool hugetlb_reserve_pages(struct inode *inode,
6662                                         long from, long to,
6663                                         struct vm_area_struct *vma,
6664                                         vm_flags_t vm_flags)
6665 {
6666         long chg = -1, add = -1;
6667         struct hstate *h = hstate_inode(inode);
6668         struct hugepage_subpool *spool = subpool_inode(inode);
6669         struct resv_map *resv_map;
6670         struct hugetlb_cgroup *h_cg = NULL;
6671         long gbl_reserve, regions_needed = 0;
6672
6673         /* This should never happen */
6674         if (from > to) {
6675                 VM_WARN(1, "%s called with a negative range\n", __func__);
6676                 return false;
6677         }
6678
6679         /*
6680          * vma specific semaphore used for pmd sharing and fault/truncation
6681          * synchronization
6682          */
6683         hugetlb_vma_lock_alloc(vma);
6684
6685         /*
6686          * Only apply hugepage reservation if asked. At fault time, an
6687          * attempt will be made for VM_NORESERVE to allocate a page
6688          * without using reserves
6689          */
6690         if (vm_flags & VM_NORESERVE)
6691                 return true;
6692
6693         /*
6694          * Shared mappings base their reservation on the number of pages that
6695          * are already allocated on behalf of the file. Private mappings need
6696          * to reserve the full area even if read-only as mprotect() may be
6697          * called to make the mapping read-write. Assume !vma is a shm mapping
6698          */
6699         if (!vma || vma->vm_flags & VM_MAYSHARE) {
6700                 /*
6701                  * resv_map can not be NULL as hugetlb_reserve_pages is only
6702                  * called for inodes for which resv_maps were created (see
6703                  * hugetlbfs_get_inode).
6704                  */
6705                 resv_map = inode_resv_map(inode);
6706
6707                 chg = region_chg(resv_map, from, to, &regions_needed);
6708         } else {
6709                 /* Private mapping. */
6710                 resv_map = resv_map_alloc();
6711                 if (!resv_map)
6712                         goto out_err;
6713
6714                 chg = to - from;
6715
6716                 set_vma_resv_map(vma, resv_map);
6717                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6718         }
6719
6720         if (chg < 0)
6721                 goto out_err;
6722
6723         if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6724                                 chg * pages_per_huge_page(h), &h_cg) < 0)
6725                 goto out_err;
6726
6727         if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6728                 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6729                  * of the resv_map.
6730                  */
6731                 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6732         }
6733
6734         /*
6735          * There must be enough pages in the subpool for the mapping. If
6736          * the subpool has a minimum size, there may be some global
6737          * reservations already in place (gbl_reserve).
6738          */
6739         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6740         if (gbl_reserve < 0)
6741                 goto out_uncharge_cgroup;
6742
6743         /*
6744          * Check enough hugepages are available for the reservation.
6745          * Hand the pages back to the subpool if there are not
6746          */
6747         if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6748                 goto out_put_pages;
6749
6750         /*
6751          * Account for the reservations made. Shared mappings record regions
6752          * that have reservations as they are shared by multiple VMAs.
6753          * When the last VMA disappears, the region map says how much
6754          * the reservation was and the page cache tells how much of
6755          * the reservation was consumed. Private mappings are per-VMA and
6756          * only the consumed reservations are tracked. When the VMA
6757          * disappears, the original reservation is the VMA size and the
6758          * consumed reservations are stored in the map. Hence, nothing
6759          * else has to be done for private mappings here
6760          */
6761         if (!vma || vma->vm_flags & VM_MAYSHARE) {
6762                 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6763
6764                 if (unlikely(add < 0)) {
6765                         hugetlb_acct_memory(h, -gbl_reserve);
6766                         goto out_put_pages;
6767                 } else if (unlikely(chg > add)) {
6768                         /*
6769                          * pages in this range were added to the reserve
6770                          * map between region_chg and region_add.  This
6771                          * indicates a race with alloc_hugetlb_folio.  Adjust
6772                          * the subpool and reserve counts modified above
6773                          * based on the difference.
6774                          */
6775                         long rsv_adjust;
6776
6777                         /*
6778                          * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6779                          * reference to h_cg->css. See comment below for detail.
6780                          */
6781                         hugetlb_cgroup_uncharge_cgroup_rsvd(
6782                                 hstate_index(h),
6783                                 (chg - add) * pages_per_huge_page(h), h_cg);
6784
6785                         rsv_adjust = hugepage_subpool_put_pages(spool,
6786                                                                 chg - add);
6787                         hugetlb_acct_memory(h, -rsv_adjust);
6788                 } else if (h_cg) {
6789                         /*
6790                          * The file_regions will hold their own reference to
6791                          * h_cg->css. So we should release the reference held
6792                          * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6793                          * done.
6794                          */
6795                         hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6796                 }
6797         }
6798         return true;
6799
6800 out_put_pages:
6801         /* put back original number of pages, chg */
6802         (void)hugepage_subpool_put_pages(spool, chg);
6803 out_uncharge_cgroup:
6804         hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6805                                             chg * pages_per_huge_page(h), h_cg);
6806 out_err:
6807         hugetlb_vma_lock_free(vma);
6808         if (!vma || vma->vm_flags & VM_MAYSHARE)
6809                 /* Only call region_abort if the region_chg succeeded but the
6810                  * region_add failed or didn't run.
6811                  */
6812                 if (chg >= 0 && add < 0)
6813                         region_abort(resv_map, from, to, regions_needed);
6814         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6815                 kref_put(&resv_map->refs, resv_map_release);
6816         return false;
6817 }
6818
6819 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6820                                                                 long freed)
6821 {
6822         struct hstate *h = hstate_inode(inode);
6823         struct resv_map *resv_map = inode_resv_map(inode);
6824         long chg = 0;
6825         struct hugepage_subpool *spool = subpool_inode(inode);
6826         long gbl_reserve;
6827
6828         /*
6829          * Since this routine can be called in the evict inode path for all
6830          * hugetlbfs inodes, resv_map could be NULL.
6831          */
6832         if (resv_map) {
6833                 chg = region_del(resv_map, start, end);
6834                 /*
6835                  * region_del() can fail in the rare case where a region
6836                  * must be split and another region descriptor can not be
6837                  * allocated.  If end == LONG_MAX, it will not fail.
6838                  */
6839                 if (chg < 0)
6840                         return chg;
6841         }
6842
6843         spin_lock(&inode->i_lock);
6844         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6845         spin_unlock(&inode->i_lock);
6846
6847         /*
6848          * If the subpool has a minimum size, the number of global
6849          * reservations to be released may be adjusted.
6850          *
6851          * Note that !resv_map implies freed == 0. So (chg - freed)
6852          * won't go negative.
6853          */
6854         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6855         hugetlb_acct_memory(h, -gbl_reserve);
6856
6857         return 0;
6858 }
6859
6860 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6861 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6862                                 struct vm_area_struct *vma,
6863                                 unsigned long addr, pgoff_t idx)
6864 {
6865         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6866                                 svma->vm_start;
6867         unsigned long sbase = saddr & PUD_MASK;
6868         unsigned long s_end = sbase + PUD_SIZE;
6869
6870         /* Allow segments to share if only one is marked locked */
6871         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
6872         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
6873
6874         /*
6875          * match the virtual addresses, permission and the alignment of the
6876          * page table page.
6877          *
6878          * Also, vma_lock (vm_private_data) is required for sharing.
6879          */
6880         if (pmd_index(addr) != pmd_index(saddr) ||
6881             vm_flags != svm_flags ||
6882             !range_in_vma(svma, sbase, s_end) ||
6883             !svma->vm_private_data)
6884                 return 0;
6885
6886         return saddr;
6887 }
6888
6889 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6890 {
6891         unsigned long start = addr & PUD_MASK;
6892         unsigned long end = start + PUD_SIZE;
6893
6894 #ifdef CONFIG_USERFAULTFD
6895         if (uffd_disable_huge_pmd_share(vma))
6896                 return false;
6897 #endif
6898         /*
6899          * check on proper vm_flags and page table alignment
6900          */
6901         if (!(vma->vm_flags & VM_MAYSHARE))
6902                 return false;
6903         if (!vma->vm_private_data)      /* vma lock required for sharing */
6904                 return false;
6905         if (!range_in_vma(vma, start, end))
6906                 return false;
6907         return true;
6908 }
6909
6910 /*
6911  * Determine if start,end range within vma could be mapped by shared pmd.
6912  * If yes, adjust start and end to cover range associated with possible
6913  * shared pmd mappings.
6914  */
6915 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6916                                 unsigned long *start, unsigned long *end)
6917 {
6918         unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6919                 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6920
6921         /*
6922          * vma needs to span at least one aligned PUD size, and the range
6923          * must be at least partially within in.
6924          */
6925         if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6926                 (*end <= v_start) || (*start >= v_end))
6927                 return;
6928
6929         /* Extend the range to be PUD aligned for a worst case scenario */
6930         if (*start > v_start)
6931                 *start = ALIGN_DOWN(*start, PUD_SIZE);
6932
6933         if (*end < v_end)
6934                 *end = ALIGN(*end, PUD_SIZE);
6935 }
6936
6937 /*
6938  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6939  * and returns the corresponding pte. While this is not necessary for the
6940  * !shared pmd case because we can allocate the pmd later as well, it makes the
6941  * code much cleaner. pmd allocation is essential for the shared case because
6942  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
6943  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
6944  * bad pmd for sharing.
6945  */
6946 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6947                       unsigned long addr, pud_t *pud)
6948 {
6949         struct address_space *mapping = vma->vm_file->f_mapping;
6950         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6951                         vma->vm_pgoff;
6952         struct vm_area_struct *svma;
6953         unsigned long saddr;
6954         pte_t *spte = NULL;
6955         pte_t *pte;
6956
6957         i_mmap_lock_read(mapping);
6958         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6959                 if (svma == vma)
6960                         continue;
6961
6962                 saddr = page_table_shareable(svma, vma, addr, idx);
6963                 if (saddr) {
6964                         spte = hugetlb_walk(svma, saddr,
6965                                             vma_mmu_pagesize(svma));
6966                         if (spte) {
6967                                 get_page(virt_to_page(spte));
6968                                 break;
6969                         }
6970                 }
6971         }
6972
6973         if (!spte)
6974                 goto out;
6975
6976         spin_lock(&mm->page_table_lock);
6977         if (pud_none(*pud)) {
6978                 pud_populate(mm, pud,
6979                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6980                 mm_inc_nr_pmds(mm);
6981         } else {
6982                 put_page(virt_to_page(spte));
6983         }
6984         spin_unlock(&mm->page_table_lock);
6985 out:
6986         pte = (pte_t *)pmd_alloc(mm, pud, addr);
6987         i_mmap_unlock_read(mapping);
6988         return pte;
6989 }
6990
6991 /*
6992  * unmap huge page backed by shared pte.
6993  *
6994  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
6995  * indicated by page_count > 1, unmap is achieved by clearing pud and
6996  * decrementing the ref count. If count == 1, the pte page is not shared.
6997  *
6998  * Called with page table lock held.
6999  *
7000  * returns: 1 successfully unmapped a shared pte page
7001  *          0 the underlying pte page is not shared, or it is the last user
7002  */
7003 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7004                                         unsigned long addr, pte_t *ptep)
7005 {
7006         pgd_t *pgd = pgd_offset(mm, addr);
7007         p4d_t *p4d = p4d_offset(pgd, addr);
7008         pud_t *pud = pud_offset(p4d, addr);
7009
7010         i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7011         hugetlb_vma_assert_locked(vma);
7012         BUG_ON(page_count(virt_to_page(ptep)) == 0);
7013         if (page_count(virt_to_page(ptep)) == 1)
7014                 return 0;
7015
7016         pud_clear(pud);
7017         put_page(virt_to_page(ptep));
7018         mm_dec_nr_pmds(mm);
7019         return 1;
7020 }
7021
7022 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7023
7024 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7025                       unsigned long addr, pud_t *pud)
7026 {
7027         return NULL;
7028 }
7029
7030 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7031                                 unsigned long addr, pte_t *ptep)
7032 {
7033         return 0;
7034 }
7035
7036 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7037                                 unsigned long *start, unsigned long *end)
7038 {
7039 }
7040
7041 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7042 {
7043         return false;
7044 }
7045 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7046
7047 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7048 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7049                         unsigned long addr, unsigned long sz)
7050 {
7051         pgd_t *pgd;
7052         p4d_t *p4d;
7053         pud_t *pud;
7054         pte_t *pte = NULL;
7055
7056         pgd = pgd_offset(mm, addr);
7057         p4d = p4d_alloc(mm, pgd, addr);
7058         if (!p4d)
7059                 return NULL;
7060         pud = pud_alloc(mm, p4d, addr);
7061         if (pud) {
7062                 if (sz == PUD_SIZE) {
7063                         pte = (pte_t *)pud;
7064                 } else {
7065                         BUG_ON(sz != PMD_SIZE);
7066                         if (want_pmd_share(vma, addr) && pud_none(*pud))
7067                                 pte = huge_pmd_share(mm, vma, addr, pud);
7068                         else
7069                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7070                 }
7071         }
7072
7073         if (pte) {
7074                 pte_t pteval = ptep_get_lockless(pte);
7075
7076                 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7077         }
7078
7079         return pte;
7080 }
7081
7082 /*
7083  * huge_pte_offset() - Walk the page table to resolve the hugepage
7084  * entry at address @addr
7085  *
7086  * Return: Pointer to page table entry (PUD or PMD) for
7087  * address @addr, or NULL if a !p*d_present() entry is encountered and the
7088  * size @sz doesn't match the hugepage size at this level of the page
7089  * table.
7090  */
7091 pte_t *huge_pte_offset(struct mm_struct *mm,
7092                        unsigned long addr, unsigned long sz)
7093 {
7094         pgd_t *pgd;
7095         p4d_t *p4d;
7096         pud_t *pud;
7097         pmd_t *pmd;
7098
7099         pgd = pgd_offset(mm, addr);
7100         if (!pgd_present(*pgd))
7101                 return NULL;
7102         p4d = p4d_offset(pgd, addr);
7103         if (!p4d_present(*p4d))
7104                 return NULL;
7105
7106         pud = pud_offset(p4d, addr);
7107         if (sz == PUD_SIZE)
7108                 /* must be pud huge, non-present or none */
7109                 return (pte_t *)pud;
7110         if (!pud_present(*pud))
7111                 return NULL;
7112         /* must have a valid entry and size to go further */
7113
7114         pmd = pmd_offset(pud, addr);
7115         /* must be pmd huge, non-present or none */
7116         return (pte_t *)pmd;
7117 }
7118
7119 /*
7120  * Return a mask that can be used to update an address to the last huge
7121  * page in a page table page mapping size.  Used to skip non-present
7122  * page table entries when linearly scanning address ranges.  Architectures
7123  * with unique huge page to page table relationships can define their own
7124  * version of this routine.
7125  */
7126 unsigned long hugetlb_mask_last_page(struct hstate *h)
7127 {
7128         unsigned long hp_size = huge_page_size(h);
7129
7130         if (hp_size == PUD_SIZE)
7131                 return P4D_SIZE - PUD_SIZE;
7132         else if (hp_size == PMD_SIZE)
7133                 return PUD_SIZE - PMD_SIZE;
7134         else
7135                 return 0UL;
7136 }
7137
7138 #else
7139
7140 /* See description above.  Architectures can provide their own version. */
7141 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7142 {
7143 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7144         if (huge_page_size(h) == PMD_SIZE)
7145                 return PUD_SIZE - PMD_SIZE;
7146 #endif
7147         return 0UL;
7148 }
7149
7150 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7151
7152 /*
7153  * These functions are overwritable if your architecture needs its own
7154  * behavior.
7155  */
7156 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7157 {
7158         bool ret = true;
7159
7160         spin_lock_irq(&hugetlb_lock);
7161         if (!folio_test_hugetlb(folio) ||
7162             !folio_test_hugetlb_migratable(folio) ||
7163             !folio_try_get(folio)) {
7164                 ret = false;
7165                 goto unlock;
7166         }
7167         folio_clear_hugetlb_migratable(folio);
7168         list_move_tail(&folio->lru, list);
7169 unlock:
7170         spin_unlock_irq(&hugetlb_lock);
7171         return ret;
7172 }
7173
7174 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7175 {
7176         int ret = 0;
7177
7178         *hugetlb = false;
7179         spin_lock_irq(&hugetlb_lock);
7180         if (folio_test_hugetlb(folio)) {
7181                 *hugetlb = true;
7182                 if (folio_test_hugetlb_freed(folio))
7183                         ret = 0;
7184                 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7185                         ret = folio_try_get(folio);
7186                 else
7187                         ret = -EBUSY;
7188         }
7189         spin_unlock_irq(&hugetlb_lock);
7190         return ret;
7191 }
7192
7193 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7194                                 bool *migratable_cleared)
7195 {
7196         int ret;
7197
7198         spin_lock_irq(&hugetlb_lock);
7199         ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7200         spin_unlock_irq(&hugetlb_lock);
7201         return ret;
7202 }
7203
7204 void folio_putback_active_hugetlb(struct folio *folio)
7205 {
7206         spin_lock_irq(&hugetlb_lock);
7207         folio_set_hugetlb_migratable(folio);
7208         list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7209         spin_unlock_irq(&hugetlb_lock);
7210         folio_put(folio);
7211 }
7212
7213 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7214 {
7215         struct hstate *h = folio_hstate(old_folio);
7216
7217         hugetlb_cgroup_migrate(old_folio, new_folio);
7218         set_page_owner_migrate_reason(&new_folio->page, reason);
7219
7220         /*
7221          * transfer temporary state of the new hugetlb folio. This is
7222          * reverse to other transitions because the newpage is going to
7223          * be final while the old one will be freed so it takes over
7224          * the temporary status.
7225          *
7226          * Also note that we have to transfer the per-node surplus state
7227          * here as well otherwise the global surplus count will not match
7228          * the per-node's.
7229          */
7230         if (folio_test_hugetlb_temporary(new_folio)) {
7231                 int old_nid = folio_nid(old_folio);
7232                 int new_nid = folio_nid(new_folio);
7233
7234                 folio_set_hugetlb_temporary(old_folio);
7235                 folio_clear_hugetlb_temporary(new_folio);
7236
7237
7238                 /*
7239                  * There is no need to transfer the per-node surplus state
7240                  * when we do not cross the node.
7241                  */
7242                 if (new_nid == old_nid)
7243                         return;
7244                 spin_lock_irq(&hugetlb_lock);
7245                 if (h->surplus_huge_pages_node[old_nid]) {
7246                         h->surplus_huge_pages_node[old_nid]--;
7247                         h->surplus_huge_pages_node[new_nid]++;
7248                 }
7249                 spin_unlock_irq(&hugetlb_lock);
7250         }
7251 }
7252
7253 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7254                                    unsigned long start,
7255                                    unsigned long end)
7256 {
7257         struct hstate *h = hstate_vma(vma);
7258         unsigned long sz = huge_page_size(h);
7259         struct mm_struct *mm = vma->vm_mm;
7260         struct mmu_notifier_range range;
7261         unsigned long address;
7262         spinlock_t *ptl;
7263         pte_t *ptep;
7264
7265         if (!(vma->vm_flags & VM_MAYSHARE))
7266                 return;
7267
7268         if (start >= end)
7269                 return;
7270
7271         flush_cache_range(vma, start, end);
7272         /*
7273          * No need to call adjust_range_if_pmd_sharing_possible(), because
7274          * we have already done the PUD_SIZE alignment.
7275          */
7276         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7277                                 start, end);
7278         mmu_notifier_invalidate_range_start(&range);
7279         hugetlb_vma_lock_write(vma);
7280         i_mmap_lock_write(vma->vm_file->f_mapping);
7281         for (address = start; address < end; address += PUD_SIZE) {
7282                 ptep = hugetlb_walk(vma, address, sz);
7283                 if (!ptep)
7284                         continue;
7285                 ptl = huge_pte_lock(h, mm, ptep);
7286                 huge_pmd_unshare(mm, vma, address, ptep);
7287                 spin_unlock(ptl);
7288         }
7289         flush_hugetlb_tlb_range(vma, start, end);
7290         i_mmap_unlock_write(vma->vm_file->f_mapping);
7291         hugetlb_vma_unlock_write(vma);
7292         /*
7293          * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7294          * Documentation/mm/mmu_notifier.rst.
7295          */
7296         mmu_notifier_invalidate_range_end(&range);
7297 }
7298
7299 /*
7300  * This function will unconditionally remove all the shared pmd pgtable entries
7301  * within the specific vma for a hugetlbfs memory range.
7302  */
7303 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7304 {
7305         hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7306                         ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7307 }
7308
7309 #ifdef CONFIG_CMA
7310 static bool cma_reserve_called __initdata;
7311
7312 static int __init cmdline_parse_hugetlb_cma(char *p)
7313 {
7314         int nid, count = 0;
7315         unsigned long tmp;
7316         char *s = p;
7317
7318         while (*s) {
7319                 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7320                         break;
7321
7322                 if (s[count] == ':') {
7323                         if (tmp >= MAX_NUMNODES)
7324                                 break;
7325                         nid = array_index_nospec(tmp, MAX_NUMNODES);
7326
7327                         s += count + 1;
7328                         tmp = memparse(s, &s);
7329                         hugetlb_cma_size_in_node[nid] = tmp;
7330                         hugetlb_cma_size += tmp;
7331
7332                         /*
7333                          * Skip the separator if have one, otherwise
7334                          * break the parsing.
7335                          */
7336                         if (*s == ',')
7337                                 s++;
7338                         else
7339                                 break;
7340                 } else {
7341                         hugetlb_cma_size = memparse(p, &p);
7342                         break;
7343                 }
7344         }
7345
7346         return 0;
7347 }
7348
7349 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7350
7351 void __init hugetlb_cma_reserve(int order)
7352 {
7353         unsigned long size, reserved, per_node;
7354         bool node_specific_cma_alloc = false;
7355         int nid;
7356
7357         cma_reserve_called = true;
7358
7359         if (!hugetlb_cma_size)
7360                 return;
7361
7362         for (nid = 0; nid < MAX_NUMNODES; nid++) {
7363                 if (hugetlb_cma_size_in_node[nid] == 0)
7364                         continue;
7365
7366                 if (!node_online(nid)) {
7367                         pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7368                         hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7369                         hugetlb_cma_size_in_node[nid] = 0;
7370                         continue;
7371                 }
7372
7373                 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7374                         pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7375                                 nid, (PAGE_SIZE << order) / SZ_1M);
7376                         hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7377                         hugetlb_cma_size_in_node[nid] = 0;
7378                 } else {
7379                         node_specific_cma_alloc = true;
7380                 }
7381         }
7382
7383         /* Validate the CMA size again in case some invalid nodes specified. */
7384         if (!hugetlb_cma_size)
7385                 return;
7386
7387         if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7388                 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7389                         (PAGE_SIZE << order) / SZ_1M);
7390                 hugetlb_cma_size = 0;
7391                 return;
7392         }
7393
7394         if (!node_specific_cma_alloc) {
7395                 /*
7396                  * If 3 GB area is requested on a machine with 4 numa nodes,
7397                  * let's allocate 1 GB on first three nodes and ignore the last one.
7398                  */
7399                 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7400                 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7401                         hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7402         }
7403
7404         reserved = 0;
7405         for_each_online_node(nid) {
7406                 int res;
7407                 char name[CMA_MAX_NAME];
7408
7409                 if (node_specific_cma_alloc) {
7410                         if (hugetlb_cma_size_in_node[nid] == 0)
7411                                 continue;
7412
7413                         size = hugetlb_cma_size_in_node[nid];
7414                 } else {
7415                         size = min(per_node, hugetlb_cma_size - reserved);
7416                 }
7417
7418                 size = round_up(size, PAGE_SIZE << order);
7419
7420                 snprintf(name, sizeof(name), "hugetlb%d", nid);
7421                 /*
7422                  * Note that 'order per bit' is based on smallest size that
7423                  * may be returned to CMA allocator in the case of
7424                  * huge page demotion.
7425                  */
7426                 res = cma_declare_contiguous_nid(0, size, 0,
7427                                                 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7428                                                  0, false, name,
7429                                                  &hugetlb_cma[nid], nid);
7430                 if (res) {
7431                         pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7432                                 res, nid);
7433                         continue;
7434                 }
7435
7436                 reserved += size;
7437                 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7438                         size / SZ_1M, nid);
7439
7440                 if (reserved >= hugetlb_cma_size)
7441                         break;
7442         }
7443
7444         if (!reserved)
7445                 /*
7446                  * hugetlb_cma_size is used to determine if allocations from
7447                  * cma are possible.  Set to zero if no cma regions are set up.
7448                  */
7449                 hugetlb_cma_size = 0;
7450 }
7451
7452 static void __init hugetlb_cma_check(void)
7453 {
7454         if (!hugetlb_cma_size || cma_reserve_called)
7455                 return;
7456
7457         pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7458 }
7459
7460 #endif /* CONFIG_CMA */