2 * Lockless get_user_pages_fast for x86
4 * Copyright (C) 2008 Nick Piggin
5 * Copyright (C) 2008 Novell Inc.
7 #include <linux/sched.h>
9 #include <linux/vmstat.h>
10 #include <linux/highmem.h>
11 #include <linux/swap.h>
12 #include <linux/memremap.h>
14 #include <asm/mmu_context.h>
15 #include <asm/pgtable.h>
17 static inline pte_t gup_get_pte(pte_t *ptep)
19 #ifndef CONFIG_X86_PAE
20 return READ_ONCE(*ptep);
23 * With get_user_pages_fast, we walk down the pagetables without taking
24 * any locks. For this we would like to load the pointers atomically,
25 * but that is not possible (without expensive cmpxchg8b) on PAE. What
26 * we do have is the guarantee that a pte will only either go from not
27 * present to present, or present to not present or both -- it will not
28 * switch to a completely different present page without a TLB flush in
29 * between; something that we are blocking by holding interrupts off.
31 * Setting ptes from not present to present goes:
36 * And present to not present goes:
41 * We must ensure here that the load of pte_low sees l iff pte_high
42 * sees h. We load pte_high *after* loading pte_low, which ensures we
43 * don't see an older value of pte_high. *Then* we recheck pte_low,
44 * which ensures that we haven't picked up a changed pte high. We might
45 * have got rubbish values from pte_low and pte_high, but we are
46 * guaranteed that pte_low will not have the present bit set *unless*
47 * it is 'l'. And get_user_pages_fast only operates on present ptes, so
50 * gup_get_pte should not be used or copied outside gup.c without being
51 * very careful -- it does not atomically load the pte or anything that
52 * is likely to be useful for you.
57 pte.pte_low = ptep->pte_low;
59 pte.pte_high = ptep->pte_high;
61 if (unlikely(pte.pte_low != ptep->pte_low))
68 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
70 while ((*nr) - nr_start) {
71 struct page *page = pages[--(*nr)];
73 ClearPageReferenced(page);
79 * 'pteval' can come from a pte, pmd or pud. We only check
80 * _PAGE_PRESENT, _PAGE_USER, and _PAGE_RW in here which are the
81 * same value on all 3 types.
83 static inline int pte_allows_gup(unsigned long pteval, int write)
85 unsigned long need_pte_bits = _PAGE_PRESENT|_PAGE_USER;
88 need_pte_bits |= _PAGE_RW;
90 if ((pteval & need_pte_bits) != need_pte_bits)
93 /* Check memory protection keys permissions. */
94 if (!__pkru_allows_pkey(pte_flags_pkey(pteval), write))
101 * Return the compund head page with ref appropriately incremented,
102 * or NULL if that failed.
104 static inline struct page *try_get_compound_head(struct page *page, int refs)
106 struct page *head = compound_head(page);
107 if (WARN_ON_ONCE(page_ref_count(head) < 0))
109 if (unlikely(!page_cache_add_speculative(head, refs)))
115 * The performance critical leaf functions are made noinline otherwise gcc
116 * inlines everything into a single function which results in too much
119 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
120 unsigned long end, int write, struct page **pages, int *nr)
122 struct dev_pagemap *pgmap = NULL;
126 ptep = pte_offset_map(&pmd, addr);
128 pte_t pte = gup_get_pte(ptep);
129 struct page *head, *page;
131 /* Similar to the PMD case, NUMA hinting must take slow path */
132 if (pte_protnone(pte)) {
137 if (!pte_allows_gup(pte_val(pte), write)) {
142 if (pte_devmap(pte)) {
143 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
144 if (unlikely(!pgmap)) {
145 undo_dev_pagemap(nr, nr_start, pages);
149 } else if (pte_special(pte)) {
153 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
154 page = pte_page(pte);
156 head = try_get_compound_head(page, 1);
158 put_dev_pagemap(pgmap);
163 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
165 put_dev_pagemap(pgmap);
170 put_dev_pagemap(pgmap);
171 SetPageReferenced(page);
175 } while (ptep++, addr += PAGE_SIZE, addr != end);
181 static inline void get_head_page_multiple(struct page *page, int nr)
183 VM_BUG_ON_PAGE(page != compound_head(page), page);
184 VM_BUG_ON_PAGE(page_count(page) == 0, page);
185 page_ref_add(page, nr);
186 SetPageReferenced(page);
189 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
190 unsigned long end, struct page **pages, int *nr)
193 unsigned long pfn = pmd_pfn(pmd);
194 struct dev_pagemap *pgmap = NULL;
196 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
198 struct page *page = pfn_to_page(pfn);
200 pgmap = get_dev_pagemap(pfn, pgmap);
201 if (unlikely(!pgmap)) {
202 undo_dev_pagemap(nr, nr_start, pages);
205 if (unlikely(!try_get_page(page))) {
206 put_dev_pagemap(pgmap);
209 SetPageReferenced(page);
211 put_dev_pagemap(pgmap);
214 } while (addr += PAGE_SIZE, addr != end);
218 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
219 unsigned long end, int write, struct page **pages, int *nr)
221 struct page *head, *page;
224 if (!pte_allows_gup(pmd_val(pmd), write))
227 VM_BUG_ON(!pfn_valid(pmd_pfn(pmd)));
229 return __gup_device_huge_pmd(pmd, addr, end, pages, nr);
231 /* hugepages are never "special" */
232 VM_BUG_ON(pmd_flags(pmd) & _PAGE_SPECIAL);
235 head = pmd_page(pmd);
236 if (WARN_ON_ONCE(page_ref_count(head) <= 0))
238 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
240 VM_BUG_ON_PAGE(compound_head(page) != head, page);
245 } while (addr += PAGE_SIZE, addr != end);
246 get_head_page_multiple(head, refs);
251 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
252 int write, struct page **pages, int *nr)
257 pmdp = pmd_offset(&pud, addr);
261 next = pmd_addr_end(addr, end);
264 if (unlikely(pmd_large(pmd) || !pmd_present(pmd))) {
266 * NUMA hinting faults need to be handled in the GUP
267 * slowpath for accounting purposes and so that they
268 * can be serialised against THP migration.
270 if (pmd_protnone(pmd))
272 if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
275 if (!gup_pte_range(pmd, addr, next, write, pages, nr))
278 } while (pmdp++, addr = next, addr != end);
283 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
284 unsigned long end, int write, struct page **pages, int *nr)
286 struct page *head, *page;
289 if (!pte_allows_gup(pud_val(pud), write))
291 /* hugepages are never "special" */
292 VM_BUG_ON(pud_flags(pud) & _PAGE_SPECIAL);
293 VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
296 head = pud_page(pud);
297 if (WARN_ON_ONCE(page_ref_count(head) <= 0))
299 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
301 VM_BUG_ON_PAGE(compound_head(page) != head, page);
306 } while (addr += PAGE_SIZE, addr != end);
307 get_head_page_multiple(head, refs);
312 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
313 int write, struct page **pages, int *nr)
318 pudp = pud_offset(&pgd, addr);
322 next = pud_addr_end(addr, end);
325 if (unlikely(pud_large(pud))) {
326 if (!gup_huge_pud(pud, addr, next, write, pages, nr))
329 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
332 } while (pudp++, addr = next, addr != end);
338 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
339 * back to the regular GUP.
341 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
344 struct mm_struct *mm = current->mm;
345 unsigned long addr, len, end;
353 len = (unsigned long) nr_pages << PAGE_SHIFT;
355 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
356 (void __user *)start, len)))
360 * XXX: batch / limit 'nr', to avoid large irq off latency
361 * needs some instrumenting to determine the common sizes used by
362 * important workloads (eg. DB2), and whether limiting the batch size
363 * will decrease performance.
365 * It seems like we're in the clear for the moment. Direct-IO is
366 * the main guy that batches up lots of get_user_pages, and even
367 * they are limited to 64-at-a-time which is not so many.
370 * This doesn't prevent pagetable teardown, but does prevent
371 * the pagetables and pages from being freed on x86.
373 * So long as we atomically load page table pointers versus teardown
374 * (which we do on x86, with the above PAE exception), we can follow the
375 * address down to the the page and take a ref on it.
377 local_irq_save(flags);
378 pgdp = pgd_offset(mm, addr);
382 next = pgd_addr_end(addr, end);
385 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
387 } while (pgdp++, addr = next, addr != end);
388 local_irq_restore(flags);
394 * get_user_pages_fast() - pin user pages in memory
395 * @start: starting user address
396 * @nr_pages: number of pages from start to pin
397 * @write: whether pages will be written to
398 * @pages: array that receives pointers to the pages pinned.
399 * Should be at least nr_pages long.
401 * Attempt to pin user pages in memory without taking mm->mmap_sem.
402 * If not successful, it will fall back to taking the lock and
403 * calling get_user_pages().
405 * Returns number of pages pinned. This may be fewer than the number
406 * requested. If nr_pages is 0 or negative, returns 0. If no pages
407 * were pinned, returns -errno.
409 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
412 struct mm_struct *mm = current->mm;
413 unsigned long addr, len, end;
420 len = (unsigned long) nr_pages << PAGE_SHIFT;
427 if (end >> __VIRTUAL_MASK_SHIFT)
432 * XXX: batch / limit 'nr', to avoid large irq off latency
433 * needs some instrumenting to determine the common sizes used by
434 * important workloads (eg. DB2), and whether limiting the batch size
435 * will decrease performance.
437 * It seems like we're in the clear for the moment. Direct-IO is
438 * the main guy that batches up lots of get_user_pages, and even
439 * they are limited to 64-at-a-time which is not so many.
442 * This doesn't prevent pagetable teardown, but does prevent
443 * the pagetables and pages from being freed on x86.
445 * So long as we atomically load page table pointers versus teardown
446 * (which we do on x86, with the above PAE exception), we can follow the
447 * address down to the the page and take a ref on it.
450 pgdp = pgd_offset(mm, addr);
454 next = pgd_addr_end(addr, end);
458 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
459 * because get_user_pages() may need to cause an early COW in
460 * order to avoid confusing the normal COW routines. So only
461 * targets that are already writable are safe to do by just
462 * looking at the page tables.
464 if (!gup_pud_range(pgd, addr, next, 1, pages, &nr))
466 } while (pgdp++, addr = next, addr != end);
469 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
478 /* Try to get the remaining pages with get_user_pages */
479 start += nr << PAGE_SHIFT;
482 ret = get_user_pages_unlocked(start,
483 (end - start) >> PAGE_SHIFT,
484 pages, write ? FOLL_WRITE : 0);
486 /* Have to be a bit careful with return values */