1 // SPDX-License-Identifier: GPL-2.0
3 * Page table allocation functions
5 * Copyright IBM Corp. 2016
6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
9 #include <linux/sysctl.h>
10 #include <linux/slab.h>
12 #include <asm/mmu_context.h>
13 #include <asm/pgalloc.h>
16 #include <asm/tlbflush.h>
20 int page_table_allocate_pgste = 0;
21 EXPORT_SYMBOL(page_table_allocate_pgste);
23 static struct ctl_table page_table_sysctl[] = {
25 .procname = "allocate_pgste",
26 .data = &page_table_allocate_pgste,
27 .maxlen = sizeof(int),
28 .mode = S_IRUGO | S_IWUSR,
29 .proc_handler = proc_dointvec_minmax,
30 .extra1 = SYSCTL_ZERO,
36 static struct ctl_table page_table_sysctl_dir[] = {
41 .child = page_table_sysctl,
46 static int __init page_table_register_sysctl(void)
48 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
50 __initcall(page_table_register_sysctl);
52 #endif /* CONFIG_PGSTE */
54 unsigned long *crst_table_alloc(struct mm_struct *mm)
56 struct page *page = alloc_pages(GFP_KERNEL, 2);
60 arch_set_page_dat(page, 2);
61 return (unsigned long *) page_to_virt(page);
64 void crst_table_free(struct mm_struct *mm, unsigned long *table)
66 free_pages((unsigned long) table, 2);
69 static void __crst_table_upgrade(void *arg)
71 struct mm_struct *mm = arg;
73 /* change all active ASCEs to avoid the creation of new TLBs */
74 if (current->active_mm == mm) {
75 S390_lowcore.user_asce = mm->context.asce;
76 __ctl_load(S390_lowcore.user_asce, 7, 7);
81 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
83 unsigned long *pgd = NULL, *p4d = NULL, *__pgd;
84 unsigned long asce_limit = mm->context.asce_limit;
86 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
87 VM_BUG_ON(asce_limit < _REGION2_SIZE);
89 if (end <= asce_limit)
92 if (asce_limit == _REGION2_SIZE) {
93 p4d = crst_table_alloc(mm);
96 crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
98 if (end > _REGION1_SIZE) {
99 pgd = crst_table_alloc(mm);
102 crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
105 spin_lock_bh(&mm->page_table_lock);
108 * This routine gets called with mmap_lock lock held and there is
109 * no reason to optimize for the case of otherwise. However, if
110 * that would ever change, the below check will let us know.
112 VM_BUG_ON(asce_limit != mm->context.asce_limit);
115 __pgd = (unsigned long *) mm->pgd;
116 p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd);
117 mm->pgd = (pgd_t *) p4d;
118 mm->context.asce_limit = _REGION1_SIZE;
119 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
120 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
124 __pgd = (unsigned long *) mm->pgd;
125 pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd);
126 mm->pgd = (pgd_t *) pgd;
127 mm->context.asce_limit = TASK_SIZE_MAX;
128 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
129 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
132 spin_unlock_bh(&mm->page_table_lock);
134 on_each_cpu(__crst_table_upgrade, mm, 0);
139 crst_table_free(mm, p4d);
144 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
146 unsigned int old, new;
149 old = atomic_read(v);
151 } while (atomic_cmpxchg(v, old, new) != old);
157 struct page *page_table_alloc_pgste(struct mm_struct *mm)
162 page = alloc_page(GFP_KERNEL);
164 table = (u64 *)page_to_virt(page);
165 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
166 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
171 void page_table_free_pgste(struct page *page)
176 #endif /* CONFIG_PGSTE */
179 * page table entry allocation/free routines.
181 unsigned long *page_table_alloc(struct mm_struct *mm)
183 unsigned long *table;
185 unsigned int mask, bit;
187 /* Try to get a fragment of a 4K page as a 2K page table */
188 if (!mm_alloc_pgste(mm)) {
190 spin_lock_bh(&mm->context.lock);
191 if (!list_empty(&mm->context.pgtable_list)) {
192 page = list_first_entry(&mm->context.pgtable_list,
194 mask = atomic_read(&page->_refcount) >> 24;
195 mask = (mask | (mask >> 4)) & 3;
197 table = (unsigned long *) page_to_virt(page);
198 bit = mask & 1; /* =1 -> second 2K */
200 table += PTRS_PER_PTE;
201 atomic_xor_bits(&page->_refcount,
203 list_del(&page->lru);
206 spin_unlock_bh(&mm->context.lock);
210 /* Allocate a fresh page */
211 page = alloc_page(GFP_KERNEL);
214 if (!pgtable_pte_page_ctor(page)) {
218 arch_set_page_dat(page, 0);
219 /* Initialize page table */
220 table = (unsigned long *) page_to_virt(page);
221 if (mm_alloc_pgste(mm)) {
222 /* Return 4K page table with PGSTEs */
223 atomic_xor_bits(&page->_refcount, 3 << 24);
224 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
225 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
227 /* Return the first 2K fragment of the page */
228 atomic_xor_bits(&page->_refcount, 1 << 24);
229 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
230 spin_lock_bh(&mm->context.lock);
231 list_add(&page->lru, &mm->context.pgtable_list);
232 spin_unlock_bh(&mm->context.lock);
237 void page_table_free(struct mm_struct *mm, unsigned long *table)
240 unsigned int bit, mask;
242 page = virt_to_page(table);
243 if (!mm_alloc_pgste(mm)) {
244 /* Free 2K page table fragment of a 4K page */
245 bit = ((unsigned long) table & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
246 spin_lock_bh(&mm->context.lock);
247 mask = atomic_xor_bits(&page->_refcount, 1U << (bit + 24));
250 list_add(&page->lru, &mm->context.pgtable_list);
252 list_del(&page->lru);
253 spin_unlock_bh(&mm->context.lock);
257 atomic_xor_bits(&page->_refcount, 3U << 24);
260 pgtable_pte_page_dtor(page);
264 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
265 unsigned long vmaddr)
267 struct mm_struct *mm;
269 unsigned int bit, mask;
272 page = virt_to_page(table);
273 if (mm_alloc_pgste(mm)) {
274 gmap_unlink(mm, table, vmaddr);
275 table = (unsigned long *) ((unsigned long)table | 3);
276 tlb_remove_table(tlb, table);
279 bit = ((unsigned long) table & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
280 spin_lock_bh(&mm->context.lock);
281 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
284 list_add_tail(&page->lru, &mm->context.pgtable_list);
286 list_del(&page->lru);
287 spin_unlock_bh(&mm->context.lock);
288 table = (unsigned long *) ((unsigned long) table | (1U << bit));
289 tlb_remove_table(tlb, table);
292 void __tlb_remove_table(void *_table)
294 unsigned int mask = (unsigned long) _table & 3;
295 void *table = (void *)((unsigned long) _table ^ mask);
296 struct page *page = virt_to_page(table);
299 case 0: /* pmd, pud, or p4d */
300 free_pages((unsigned long) table, 2);
302 case 1: /* lower 2K of a 4K page table */
303 case 2: /* higher 2K of a 4K page table */
304 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
309 case 3: /* 4K page table with pgstes */
311 atomic_xor_bits(&page->_refcount, 3 << 24);
312 pgtable_pte_page_dtor(page);
319 * Base infrastructure required to generate basic asces, region, segment,
320 * and page tables that do not make use of enhanced features like EDAT1.
323 static struct kmem_cache *base_pgt_cache;
325 static unsigned long base_pgt_alloc(void)
329 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
331 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
332 return (unsigned long) table;
335 static void base_pgt_free(unsigned long table)
337 kmem_cache_free(base_pgt_cache, (void *) table);
340 static unsigned long base_crst_alloc(unsigned long val)
344 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
346 crst_table_init((unsigned long *)table, val);
350 static void base_crst_free(unsigned long table)
352 free_pages(table, CRST_ALLOC_ORDER);
355 #define BASE_ADDR_END_FUNC(NAME, SIZE) \
356 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
359 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
361 return (next - 1) < (end - 1) ? next : end; \
364 BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
365 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
366 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
367 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
368 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
370 static inline unsigned long base_lra(unsigned long address)
376 : "=d" (real) : "a" (address) : "cc");
380 static int base_page_walk(unsigned long origin, unsigned long addr,
381 unsigned long end, int alloc)
383 unsigned long *pte, next;
387 pte = (unsigned long *) origin;
388 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
390 next = base_page_addr_end(addr, end);
391 *pte = base_lra(addr);
392 } while (pte++, addr = next, addr < end);
396 static int base_segment_walk(unsigned long origin, unsigned long addr,
397 unsigned long end, int alloc)
399 unsigned long *ste, next, table;
402 ste = (unsigned long *) origin;
403 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
405 next = base_segment_addr_end(addr, end);
406 if (*ste & _SEGMENT_ENTRY_INVALID) {
409 table = base_pgt_alloc();
412 *ste = table | _SEGMENT_ENTRY;
414 table = *ste & _SEGMENT_ENTRY_ORIGIN;
415 rc = base_page_walk(table, addr, next, alloc);
419 base_pgt_free(table);
421 } while (ste++, addr = next, addr < end);
425 static int base_region3_walk(unsigned long origin, unsigned long addr,
426 unsigned long end, int alloc)
428 unsigned long *rtte, next, table;
431 rtte = (unsigned long *) origin;
432 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
434 next = base_region3_addr_end(addr, end);
435 if (*rtte & _REGION_ENTRY_INVALID) {
438 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
441 *rtte = table | _REGION3_ENTRY;
443 table = *rtte & _REGION_ENTRY_ORIGIN;
444 rc = base_segment_walk(table, addr, next, alloc);
448 base_crst_free(table);
449 } while (rtte++, addr = next, addr < end);
453 static int base_region2_walk(unsigned long origin, unsigned long addr,
454 unsigned long end, int alloc)
456 unsigned long *rste, next, table;
459 rste = (unsigned long *) origin;
460 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
462 next = base_region2_addr_end(addr, end);
463 if (*rste & _REGION_ENTRY_INVALID) {
466 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
469 *rste = table | _REGION2_ENTRY;
471 table = *rste & _REGION_ENTRY_ORIGIN;
472 rc = base_region3_walk(table, addr, next, alloc);
476 base_crst_free(table);
477 } while (rste++, addr = next, addr < end);
481 static int base_region1_walk(unsigned long origin, unsigned long addr,
482 unsigned long end, int alloc)
484 unsigned long *rfte, next, table;
487 rfte = (unsigned long *) origin;
488 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
490 next = base_region1_addr_end(addr, end);
491 if (*rfte & _REGION_ENTRY_INVALID) {
494 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
497 *rfte = table | _REGION1_ENTRY;
499 table = *rfte & _REGION_ENTRY_ORIGIN;
500 rc = base_region2_walk(table, addr, next, alloc);
504 base_crst_free(table);
505 } while (rfte++, addr = next, addr < end);
510 * base_asce_free - free asce and tables returned from base_asce_alloc()
511 * @asce: asce to be freed
513 * Frees all region, segment, and page tables that were allocated with a
514 * corresponding base_asce_alloc() call.
516 void base_asce_free(unsigned long asce)
518 unsigned long table = asce & _ASCE_ORIGIN;
522 switch (asce & _ASCE_TYPE_MASK) {
523 case _ASCE_TYPE_SEGMENT:
524 base_segment_walk(table, 0, _REGION3_SIZE, 0);
526 case _ASCE_TYPE_REGION3:
527 base_region3_walk(table, 0, _REGION2_SIZE, 0);
529 case _ASCE_TYPE_REGION2:
530 base_region2_walk(table, 0, _REGION1_SIZE, 0);
532 case _ASCE_TYPE_REGION1:
533 base_region1_walk(table, 0, TASK_SIZE_MAX, 0);
536 base_crst_free(table);
539 static int base_pgt_cache_init(void)
541 static DEFINE_MUTEX(base_pgt_cache_mutex);
542 unsigned long sz = _PAGE_TABLE_SIZE;
546 mutex_lock(&base_pgt_cache_mutex);
548 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
549 mutex_unlock(&base_pgt_cache_mutex);
550 return base_pgt_cache ? 0 : -ENOMEM;
554 * base_asce_alloc - create kernel mapping without enhanced DAT features
555 * @addr: virtual start address of kernel mapping
556 * @num_pages: number of consecutive pages
558 * Generate an asce, including all required region, segment and page tables,
559 * that can be used to access the virtual kernel mapping. The difference is
560 * that the returned asce does not make use of any enhanced DAT features like
561 * e.g. large pages. This is required for some I/O functions that pass an
562 * asce, like e.g. some service call requests.
564 * Note: the returned asce may NEVER be attached to any cpu. It may only be
565 * used for I/O requests. tlb entries that might result because the
566 * asce was attached to a cpu won't be cleared.
568 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
570 unsigned long asce, table, end;
573 if (base_pgt_cache_init())
575 end = addr + num_pages * PAGE_SIZE;
576 if (end <= _REGION3_SIZE) {
577 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
580 rc = base_segment_walk(table, addr, end, 1);
581 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
582 } else if (end <= _REGION2_SIZE) {
583 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
586 rc = base_region3_walk(table, addr, end, 1);
587 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
588 } else if (end <= _REGION1_SIZE) {
589 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
592 rc = base_region2_walk(table, addr, end, 1);
593 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
595 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
598 rc = base_region1_walk(table, addr, end, 1);
599 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
602 base_asce_free(asce);