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, pud or p4d. We only check
80 * _PAGE_PRESENT, _PAGE_USER, and _PAGE_RW in here which are the
81 * same value on all 4 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 * The performance critical leaf functions are made noinline otherwise gcc
102 * inlines everything into a single function which results in too much
105 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
106 unsigned long end, int write, struct page **pages, int *nr)
108 struct dev_pagemap *pgmap = NULL;
109 int nr_start = *nr, ret = 0;
113 * Keep the original mapped PTE value (ptem) around since we
114 * might increment ptep off the end of the page when finishing
115 * our loop iteration.
117 ptem = ptep = pte_offset_map(&pmd, addr);
119 pte_t pte = gup_get_pte(ptep);
122 /* Similar to the PMD case, NUMA hinting must take slow path */
123 if (pte_protnone(pte))
126 if (!pte_allows_gup(pte_val(pte), write))
129 if (pte_devmap(pte)) {
130 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
131 if (unlikely(!pgmap)) {
132 undo_dev_pagemap(nr, nr_start, pages);
135 } else if (pte_special(pte))
138 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
139 page = pte_page(pte);
141 put_dev_pagemap(pgmap);
142 SetPageReferenced(page);
146 } while (ptep++, addr += PAGE_SIZE, addr != end);
154 static inline void get_head_page_multiple(struct page *page, int nr)
156 VM_BUG_ON_PAGE(page != compound_head(page), page);
157 VM_BUG_ON_PAGE(page_count(page) == 0, page);
158 page_ref_add(page, nr);
159 SetPageReferenced(page);
162 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
163 unsigned long end, struct page **pages, int *nr)
166 struct dev_pagemap *pgmap = NULL;
169 struct page *page = pfn_to_page(pfn);
171 pgmap = get_dev_pagemap(pfn, pgmap);
172 if (unlikely(!pgmap)) {
173 undo_dev_pagemap(nr, nr_start, pages);
176 SetPageReferenced(page);
179 put_dev_pagemap(pgmap);
182 } while (addr += PAGE_SIZE, addr != end);
186 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
187 unsigned long end, struct page **pages, int *nr)
189 unsigned long fault_pfn;
191 fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
192 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
195 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
196 unsigned long end, struct page **pages, int *nr)
198 unsigned long fault_pfn;
200 fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
201 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
204 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
205 unsigned long end, int write, struct page **pages, int *nr)
207 struct page *head, *page;
210 if (!pte_allows_gup(pmd_val(pmd), write))
213 VM_BUG_ON(!pfn_valid(pmd_pfn(pmd)));
215 return __gup_device_huge_pmd(pmd, addr, end, pages, nr);
217 /* hugepages are never "special" */
218 VM_BUG_ON(pmd_flags(pmd) & _PAGE_SPECIAL);
221 head = pmd_page(pmd);
222 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
224 VM_BUG_ON_PAGE(compound_head(page) != head, page);
229 } while (addr += PAGE_SIZE, addr != end);
230 get_head_page_multiple(head, refs);
235 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
236 int write, struct page **pages, int *nr)
241 pmdp = pmd_offset(&pud, addr);
245 next = pmd_addr_end(addr, end);
248 if (unlikely(pmd_large(pmd) || !pmd_present(pmd))) {
250 * NUMA hinting faults need to be handled in the GUP
251 * slowpath for accounting purposes and so that they
252 * can be serialised against THP migration.
254 if (pmd_protnone(pmd))
256 if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
259 if (!gup_pte_range(pmd, addr, next, write, pages, nr))
262 } while (pmdp++, addr = next, addr != end);
267 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
268 unsigned long end, int write, struct page **pages, int *nr)
270 struct page *head, *page;
273 if (!pte_allows_gup(pud_val(pud), write))
276 VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
278 return __gup_device_huge_pud(pud, addr, end, pages, nr);
280 /* hugepages are never "special" */
281 VM_BUG_ON(pud_flags(pud) & _PAGE_SPECIAL);
284 head = pud_page(pud);
285 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
287 VM_BUG_ON_PAGE(compound_head(page) != head, page);
292 } while (addr += PAGE_SIZE, addr != end);
293 get_head_page_multiple(head, refs);
298 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
299 int write, struct page **pages, int *nr)
304 pudp = pud_offset(&p4d, addr);
308 next = pud_addr_end(addr, end);
311 if (unlikely(pud_large(pud))) {
312 if (!gup_huge_pud(pud, addr, next, write, pages, nr))
315 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
318 } while (pudp++, addr = next, addr != end);
323 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
324 int write, struct page **pages, int *nr)
329 p4dp = p4d_offset(&pgd, addr);
333 next = p4d_addr_end(addr, end);
336 BUILD_BUG_ON(p4d_large(p4d));
337 if (!gup_pud_range(p4d, addr, next, write, pages, nr))
339 } while (p4dp++, addr = next, addr != end);
345 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
346 * back to the regular GUP.
348 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
351 struct mm_struct *mm = current->mm;
352 unsigned long addr, len, end;
360 len = (unsigned long) nr_pages << PAGE_SHIFT;
362 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
363 (void __user *)start, len)))
367 * XXX: batch / limit 'nr', to avoid large irq off latency
368 * needs some instrumenting to determine the common sizes used by
369 * important workloads (eg. DB2), and whether limiting the batch size
370 * will decrease performance.
372 * It seems like we're in the clear for the moment. Direct-IO is
373 * the main guy that batches up lots of get_user_pages, and even
374 * they are limited to 64-at-a-time which is not so many.
377 * This doesn't prevent pagetable teardown, but does prevent
378 * the pagetables and pages from being freed on x86.
380 * So long as we atomically load page table pointers versus teardown
381 * (which we do on x86, with the above PAE exception), we can follow the
382 * address down to the the page and take a ref on it.
384 local_irq_save(flags);
385 pgdp = pgd_offset(mm, addr);
389 next = pgd_addr_end(addr, end);
392 if (!gup_p4d_range(pgd, addr, next, write, pages, &nr))
394 } while (pgdp++, addr = next, addr != end);
395 local_irq_restore(flags);
401 * get_user_pages_fast() - pin user pages in memory
402 * @start: starting user address
403 * @nr_pages: number of pages from start to pin
404 * @write: whether pages will be written to
405 * @pages: array that receives pointers to the pages pinned.
406 * Should be at least nr_pages long.
408 * Attempt to pin user pages in memory without taking mm->mmap_sem.
409 * If not successful, it will fall back to taking the lock and
410 * calling get_user_pages().
412 * Returns number of pages pinned. This may be fewer than the number
413 * requested. If nr_pages is 0 or negative, returns 0. If no pages
414 * were pinned, returns -errno.
416 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
419 struct mm_struct *mm = current->mm;
420 unsigned long addr, len, end;
427 len = (unsigned long) nr_pages << PAGE_SHIFT;
434 if (end >> __VIRTUAL_MASK_SHIFT)
439 * XXX: batch / limit 'nr', to avoid large irq off latency
440 * needs some instrumenting to determine the common sizes used by
441 * important workloads (eg. DB2), and whether limiting the batch size
442 * will decrease performance.
444 * It seems like we're in the clear for the moment. Direct-IO is
445 * the main guy that batches up lots of get_user_pages, and even
446 * they are limited to 64-at-a-time which is not so many.
449 * This doesn't prevent pagetable teardown, but does prevent
450 * the pagetables and pages from being freed on x86.
452 * So long as we atomically load page table pointers versus teardown
453 * (which we do on x86, with the above PAE exception), we can follow the
454 * address down to the the page and take a ref on it.
457 pgdp = pgd_offset(mm, addr);
461 next = pgd_addr_end(addr, end);
464 if (!gup_p4d_range(pgd, addr, next, write, 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 */