1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright 2013 Red Hat Inc.
5 * Authors: Jérôme Glisse <jglisse@redhat.com>
8 * Refer to include/linux/hmm.h for information about heterogeneous memory
9 * management or HMM for short.
11 #include <linux/pagewalk.h>
12 #include <linux/hmm.h>
13 #include <linux/init.h>
14 #include <linux/rmap.h>
15 #include <linux/swap.h>
16 #include <linux/slab.h>
17 #include <linux/sched.h>
18 #include <linux/mmzone.h>
19 #include <linux/pagemap.h>
20 #include <linux/swapops.h>
21 #include <linux/hugetlb.h>
22 #include <linux/memremap.h>
23 #include <linux/sched/mm.h>
24 #include <linux/jump_label.h>
25 #include <linux/dma-mapping.h>
26 #include <linux/mmu_notifier.h>
27 #include <linux/memory_hotplug.h>
30 struct hmm_range *range;
35 HMM_NEED_FAULT = 1 << 0,
36 HMM_NEED_WRITE_FAULT = 1 << 1,
37 HMM_NEED_ALL_BITS = HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT,
40 static int hmm_pfns_fill(unsigned long addr, unsigned long end,
41 struct hmm_range *range, unsigned long cpu_flags)
43 unsigned long i = (addr - range->start) >> PAGE_SHIFT;
45 for (; addr < end; addr += PAGE_SIZE, i++)
46 range->hmm_pfns[i] = cpu_flags;
51 * hmm_vma_fault() - fault in a range lacking valid pmd or pte(s)
52 * @addr: range virtual start address (inclusive)
53 * @end: range virtual end address (exclusive)
54 * @required_fault: HMM_NEED_* flags
55 * @walk: mm_walk structure
56 * Return: -EBUSY after page fault, or page fault error
58 * This function will be called whenever pmd_none() or pte_none() returns true,
59 * or whenever there is no page directory covering the virtual address range.
61 static int hmm_vma_fault(unsigned long addr, unsigned long end,
62 unsigned int required_fault, struct mm_walk *walk)
64 struct hmm_vma_walk *hmm_vma_walk = walk->private;
65 struct vm_area_struct *vma = walk->vma;
66 unsigned int fault_flags = FAULT_FLAG_REMOTE;
68 WARN_ON_ONCE(!required_fault);
69 hmm_vma_walk->last = addr;
71 if (required_fault & HMM_NEED_WRITE_FAULT) {
72 if (!(vma->vm_flags & VM_WRITE))
74 fault_flags |= FAULT_FLAG_WRITE;
77 for (; addr < end; addr += PAGE_SIZE)
78 if (handle_mm_fault(vma, addr, fault_flags, NULL) &
84 static unsigned int hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
85 unsigned long pfn_req_flags,
86 unsigned long cpu_flags)
88 struct hmm_range *range = hmm_vma_walk->range;
91 * So we not only consider the individual per page request we also
92 * consider the default flags requested for the range. The API can
93 * be used 2 ways. The first one where the HMM user coalesces
94 * multiple page faults into one request and sets flags per pfn for
95 * those faults. The second one where the HMM user wants to pre-
96 * fault a range with specific flags. For the latter one it is a
97 * waste to have the user pre-fill the pfn arrays with a default
100 pfn_req_flags &= range->pfn_flags_mask;
101 pfn_req_flags |= range->default_flags;
103 /* We aren't ask to do anything ... */
104 if (!(pfn_req_flags & HMM_PFN_REQ_FAULT))
107 /* Need to write fault ? */
108 if ((pfn_req_flags & HMM_PFN_REQ_WRITE) &&
109 !(cpu_flags & HMM_PFN_WRITE))
110 return HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT;
112 /* If CPU page table is not valid then we need to fault */
113 if (!(cpu_flags & HMM_PFN_VALID))
114 return HMM_NEED_FAULT;
119 hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
120 const unsigned long hmm_pfns[], unsigned long npages,
121 unsigned long cpu_flags)
123 struct hmm_range *range = hmm_vma_walk->range;
124 unsigned int required_fault = 0;
128 * If the default flags do not request to fault pages, and the mask does
129 * not allow for individual pages to be faulted, then
130 * hmm_pte_need_fault() will always return 0.
132 if (!((range->default_flags | range->pfn_flags_mask) &
136 for (i = 0; i < npages; ++i) {
137 required_fault |= hmm_pte_need_fault(hmm_vma_walk, hmm_pfns[i],
139 if (required_fault == HMM_NEED_ALL_BITS)
140 return required_fault;
142 return required_fault;
145 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
146 __always_unused int depth, struct mm_walk *walk)
148 struct hmm_vma_walk *hmm_vma_walk = walk->private;
149 struct hmm_range *range = hmm_vma_walk->range;
150 unsigned int required_fault;
151 unsigned long i, npages;
152 unsigned long *hmm_pfns;
154 i = (addr - range->start) >> PAGE_SHIFT;
155 npages = (end - addr) >> PAGE_SHIFT;
156 hmm_pfns = &range->hmm_pfns[i];
158 hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0);
162 return hmm_pfns_fill(addr, end, range, HMM_PFN_ERROR);
165 return hmm_vma_fault(addr, end, required_fault, walk);
166 return hmm_pfns_fill(addr, end, range, 0);
169 static inline unsigned long hmm_pfn_flags_order(unsigned long order)
171 return order << HMM_PFN_ORDER_SHIFT;
174 static inline unsigned long pmd_to_hmm_pfn_flags(struct hmm_range *range,
177 if (pmd_protnone(pmd))
179 return (pmd_write(pmd) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
181 hmm_pfn_flags_order(PMD_SHIFT - PAGE_SHIFT);
184 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
185 static int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
186 unsigned long end, unsigned long hmm_pfns[],
189 struct hmm_vma_walk *hmm_vma_walk = walk->private;
190 struct hmm_range *range = hmm_vma_walk->range;
191 unsigned long pfn, npages, i;
192 unsigned int required_fault;
193 unsigned long cpu_flags;
195 npages = (end - addr) >> PAGE_SHIFT;
196 cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
198 hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, cpu_flags);
200 return hmm_vma_fault(addr, end, required_fault, walk);
202 pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
203 for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
204 hmm_pfns[i] = pfn | cpu_flags;
207 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
208 /* stub to allow the code below to compile */
209 int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
210 unsigned long end, unsigned long hmm_pfns[], pmd_t pmd);
211 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
213 static inline bool hmm_is_device_private_entry(struct hmm_range *range,
216 return is_device_private_entry(entry) &&
217 device_private_entry_to_page(entry)->pgmap->owner ==
218 range->dev_private_owner;
221 static inline unsigned long pte_to_hmm_pfn_flags(struct hmm_range *range,
224 if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
226 return pte_write(pte) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
229 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
230 unsigned long end, pmd_t *pmdp, pte_t *ptep,
231 unsigned long *hmm_pfn)
233 struct hmm_vma_walk *hmm_vma_walk = walk->private;
234 struct hmm_range *range = hmm_vma_walk->range;
235 unsigned int required_fault;
236 unsigned long cpu_flags;
238 uint64_t pfn_req_flags = *hmm_pfn;
242 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
249 if (!pte_present(pte)) {
250 swp_entry_t entry = pte_to_swp_entry(pte);
253 * Never fault in device private pages, but just report
254 * the PFN even if not present.
256 if (hmm_is_device_private_entry(range, entry)) {
257 cpu_flags = HMM_PFN_VALID;
258 if (is_write_device_private_entry(entry))
259 cpu_flags |= HMM_PFN_WRITE;
260 *hmm_pfn = device_private_entry_to_pfn(entry) |
266 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
267 if (!required_fault) {
272 if (!non_swap_entry(entry))
275 if (is_migration_entry(entry)) {
277 hmm_vma_walk->last = addr;
278 migration_entry_wait(walk->mm, pmdp, addr);
282 /* Report error for everything else */
287 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
289 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
294 * Bypass devmap pte such as DAX page when all pfn requested
295 * flags(pfn_req_flags) are fulfilled.
296 * Since each architecture defines a struct page for the zero page, just
297 * fall through and treat it like a normal page.
299 if (pte_special(pte) && !pte_devmap(pte) &&
300 !is_zero_pfn(pte_pfn(pte))) {
301 if (hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0)) {
305 *hmm_pfn = HMM_PFN_ERROR;
309 *hmm_pfn = pte_pfn(pte) | cpu_flags;
314 /* Fault any virtual address we were asked to fault */
315 return hmm_vma_fault(addr, end, required_fault, walk);
318 static int hmm_vma_walk_pmd(pmd_t *pmdp,
321 struct mm_walk *walk)
323 struct hmm_vma_walk *hmm_vma_walk = walk->private;
324 struct hmm_range *range = hmm_vma_walk->range;
325 unsigned long *hmm_pfns =
326 &range->hmm_pfns[(start - range->start) >> PAGE_SHIFT];
327 unsigned long npages = (end - start) >> PAGE_SHIFT;
328 unsigned long addr = start;
333 pmd = READ_ONCE(*pmdp);
335 return hmm_vma_walk_hole(start, end, -1, walk);
337 if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
338 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0)) {
339 hmm_vma_walk->last = addr;
340 pmd_migration_entry_wait(walk->mm, pmdp);
343 return hmm_pfns_fill(start, end, range, 0);
346 if (!pmd_present(pmd)) {
347 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
349 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
352 if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
354 * No need to take pmd_lock here, even if some other thread
355 * is splitting the huge pmd we will get that event through
356 * mmu_notifier callback.
358 * So just read pmd value and check again it's a transparent
359 * huge or device mapping one and compute corresponding pfn
362 pmd = pmd_read_atomic(pmdp);
364 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
367 return hmm_vma_handle_pmd(walk, addr, end, hmm_pfns, pmd);
371 * We have handled all the valid cases above ie either none, migration,
372 * huge or transparent huge. At this point either it is a valid pmd
373 * entry pointing to pte directory or it is a bad pmd that will not
377 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
379 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
382 ptep = pte_offset_map(pmdp, addr);
383 for (; addr < end; addr += PAGE_SIZE, ptep++, hmm_pfns++) {
386 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, hmm_pfns);
388 /* hmm_vma_handle_pte() did pte_unmap() */
396 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
397 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
398 static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range,
401 if (!pud_present(pud))
403 return (pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
405 hmm_pfn_flags_order(PUD_SHIFT - PAGE_SHIFT);
408 static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
409 struct mm_walk *walk)
411 struct hmm_vma_walk *hmm_vma_walk = walk->private;
412 struct hmm_range *range = hmm_vma_walk->range;
413 unsigned long addr = start;
416 spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
421 /* Normally we don't want to split the huge page */
422 walk->action = ACTION_CONTINUE;
424 pud = READ_ONCE(*pudp);
427 return hmm_vma_walk_hole(start, end, -1, walk);
430 if (pud_huge(pud) && pud_devmap(pud)) {
431 unsigned long i, npages, pfn;
432 unsigned int required_fault;
433 unsigned long *hmm_pfns;
434 unsigned long cpu_flags;
436 if (!pud_present(pud)) {
438 return hmm_vma_walk_hole(start, end, -1, walk);
441 i = (addr - range->start) >> PAGE_SHIFT;
442 npages = (end - addr) >> PAGE_SHIFT;
443 hmm_pfns = &range->hmm_pfns[i];
445 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
446 required_fault = hmm_range_need_fault(hmm_vma_walk, hmm_pfns,
448 if (required_fault) {
450 return hmm_vma_fault(addr, end, required_fault, walk);
453 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
454 for (i = 0; i < npages; ++i, ++pfn)
455 hmm_pfns[i] = pfn | cpu_flags;
459 /* Ask for the PUD to be split */
460 walk->action = ACTION_SUBTREE;
467 #define hmm_vma_walk_pud NULL
470 #ifdef CONFIG_HUGETLB_PAGE
471 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
472 unsigned long start, unsigned long end,
473 struct mm_walk *walk)
475 unsigned long addr = start, i, pfn;
476 struct hmm_vma_walk *hmm_vma_walk = walk->private;
477 struct hmm_range *range = hmm_vma_walk->range;
478 struct vm_area_struct *vma = walk->vma;
479 unsigned int required_fault;
480 unsigned long pfn_req_flags;
481 unsigned long cpu_flags;
485 ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
486 entry = huge_ptep_get(pte);
488 i = (start - range->start) >> PAGE_SHIFT;
489 pfn_req_flags = range->hmm_pfns[i];
490 cpu_flags = pte_to_hmm_pfn_flags(range, entry) |
491 hmm_pfn_flags_order(huge_page_order(hstate_vma(vma)));
493 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
494 if (required_fault) {
496 return hmm_vma_fault(addr, end, required_fault, walk);
499 pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
500 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
501 range->hmm_pfns[i] = pfn | cpu_flags;
507 #define hmm_vma_walk_hugetlb_entry NULL
508 #endif /* CONFIG_HUGETLB_PAGE */
510 static int hmm_vma_walk_test(unsigned long start, unsigned long end,
511 struct mm_walk *walk)
513 struct hmm_vma_walk *hmm_vma_walk = walk->private;
514 struct hmm_range *range = hmm_vma_walk->range;
515 struct vm_area_struct *vma = walk->vma;
517 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP | VM_MIXEDMAP)) &&
518 vma->vm_flags & VM_READ)
522 * vma ranges that don't have struct page backing them or map I/O
523 * devices directly cannot be handled by hmm_range_fault().
525 * If the vma does not allow read access, then assume that it does not
526 * allow write access either. HMM does not support architectures that
527 * allow write without read.
529 * If a fault is requested for an unsupported range then it is a hard
532 if (hmm_range_need_fault(hmm_vma_walk,
534 ((start - range->start) >> PAGE_SHIFT),
535 (end - start) >> PAGE_SHIFT, 0))
538 hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
540 /* Skip this vma and continue processing the next vma. */
544 static const struct mm_walk_ops hmm_walk_ops = {
545 .pud_entry = hmm_vma_walk_pud,
546 .pmd_entry = hmm_vma_walk_pmd,
547 .pte_hole = hmm_vma_walk_hole,
548 .hugetlb_entry = hmm_vma_walk_hugetlb_entry,
549 .test_walk = hmm_vma_walk_test,
553 * hmm_range_fault - try to fault some address in a virtual address range
554 * @range: argument structure
556 * Returns 0 on success or one of the following error codes:
558 * -EINVAL: Invalid arguments or mm or virtual address is in an invalid vma
559 * (e.g., device file vma).
560 * -ENOMEM: Out of memory.
561 * -EPERM: Invalid permission (e.g., asking for write and range is read
563 * -EBUSY: The range has been invalidated and the caller needs to wait for
564 * the invalidation to finish.
565 * -EFAULT: A page was requested to be valid and could not be made valid
566 * ie it has no backing VMA or it is illegal to access
568 * This is similar to get_user_pages(), except that it can read the page tables
569 * without mutating them (ie causing faults).
571 int hmm_range_fault(struct hmm_range *range)
573 struct hmm_vma_walk hmm_vma_walk = {
575 .last = range->start,
577 struct mm_struct *mm = range->notifier->mm;
580 mmap_assert_locked(mm);
583 /* If range is no longer valid force retry. */
584 if (mmu_interval_check_retry(range->notifier,
585 range->notifier_seq))
587 ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
588 &hmm_walk_ops, &hmm_vma_walk);
590 * When -EBUSY is returned the loop restarts with
591 * hmm_vma_walk.last set to an address that has not been stored
592 * in pfns. All entries < last in the pfn array are set to their
593 * output, and all >= are still at their input values.
595 } while (ret == -EBUSY);
598 EXPORT_SYMBOL(hmm_range_fault);