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 * 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;
112 ptep = pte_offset_map(&pmd, addr);
114 pte_t pte = gup_get_pte(ptep);
117 /* Similar to the PMD case, NUMA hinting must take slow path */
118 if (pte_protnone(pte)) {
123 if (pte_devmap(pte)) {
124 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
125 if (unlikely(!pgmap)) {
126 undo_dev_pagemap(nr, nr_start, pages);
130 } else if (!pte_allows_gup(pte_val(pte), write) ||
135 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
136 page = pte_page(pte);
138 put_dev_pagemap(pgmap);
139 SetPageReferenced(page);
143 } while (ptep++, addr += PAGE_SIZE, addr != end);
149 static inline void get_head_page_multiple(struct page *page, int nr)
151 VM_BUG_ON_PAGE(page != compound_head(page), page);
152 VM_BUG_ON_PAGE(page_count(page) == 0, page);
153 page_ref_add(page, nr);
154 SetPageReferenced(page);
157 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
158 unsigned long end, struct page **pages, int *nr)
161 struct dev_pagemap *pgmap = NULL;
164 struct page *page = pfn_to_page(pfn);
166 pgmap = get_dev_pagemap(pfn, pgmap);
167 if (unlikely(!pgmap)) {
168 undo_dev_pagemap(nr, nr_start, pages);
171 SetPageReferenced(page);
174 put_dev_pagemap(pgmap);
177 } while (addr += PAGE_SIZE, addr != end);
181 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
182 unsigned long end, struct page **pages, int *nr)
184 unsigned long fault_pfn;
186 fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
187 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
190 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
191 unsigned long end, struct page **pages, int *nr)
193 unsigned long fault_pfn;
195 fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
196 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
199 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
200 unsigned long end, int write, struct page **pages, int *nr)
202 struct page *head, *page;
205 if (!pte_allows_gup(pmd_val(pmd), write))
208 VM_BUG_ON(!pfn_valid(pmd_pfn(pmd)));
210 return __gup_device_huge_pmd(pmd, addr, end, pages, nr);
212 /* hugepages are never "special" */
213 VM_BUG_ON(pmd_flags(pmd) & _PAGE_SPECIAL);
216 head = pmd_page(pmd);
217 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
219 VM_BUG_ON_PAGE(compound_head(page) != head, page);
224 } while (addr += PAGE_SIZE, addr != end);
225 get_head_page_multiple(head, refs);
230 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
231 int write, struct page **pages, int *nr)
236 pmdp = pmd_offset(&pud, addr);
240 next = pmd_addr_end(addr, end);
243 if (unlikely(pmd_large(pmd) || !pmd_present(pmd))) {
245 * NUMA hinting faults need to be handled in the GUP
246 * slowpath for accounting purposes and so that they
247 * can be serialised against THP migration.
249 if (pmd_protnone(pmd))
251 if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
254 if (!gup_pte_range(pmd, addr, next, write, pages, nr))
257 } while (pmdp++, addr = next, addr != end);
262 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
263 unsigned long end, int write, struct page **pages, int *nr)
265 struct page *head, *page;
268 if (!pte_allows_gup(pud_val(pud), write))
271 VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
273 return __gup_device_huge_pud(pud, addr, end, pages, nr);
275 /* hugepages are never "special" */
276 VM_BUG_ON(pud_flags(pud) & _PAGE_SPECIAL);
279 head = pud_page(pud);
280 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
282 VM_BUG_ON_PAGE(compound_head(page) != head, page);
287 } while (addr += PAGE_SIZE, addr != end);
288 get_head_page_multiple(head, refs);
293 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
294 int write, struct page **pages, int *nr)
299 pudp = pud_offset(&pgd, addr);
303 next = pud_addr_end(addr, end);
306 if (unlikely(pud_large(pud))) {
307 if (!gup_huge_pud(pud, addr, next, write, pages, nr))
310 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
313 } while (pudp++, addr = next, addr != end);
319 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
320 * back to the regular GUP.
322 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
325 struct mm_struct *mm = current->mm;
326 unsigned long addr, len, end;
334 len = (unsigned long) nr_pages << PAGE_SHIFT;
336 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
337 (void __user *)start, len)))
341 * XXX: batch / limit 'nr', to avoid large irq off latency
342 * needs some instrumenting to determine the common sizes used by
343 * important workloads (eg. DB2), and whether limiting the batch size
344 * will decrease performance.
346 * It seems like we're in the clear for the moment. Direct-IO is
347 * the main guy that batches up lots of get_user_pages, and even
348 * they are limited to 64-at-a-time which is not so many.
351 * This doesn't prevent pagetable teardown, but does prevent
352 * the pagetables and pages from being freed on x86.
354 * So long as we atomically load page table pointers versus teardown
355 * (which we do on x86, with the above PAE exception), we can follow the
356 * address down to the the page and take a ref on it.
358 local_irq_save(flags);
359 pgdp = pgd_offset(mm, addr);
363 next = pgd_addr_end(addr, end);
366 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
368 } while (pgdp++, addr = next, addr != end);
369 local_irq_restore(flags);
375 * get_user_pages_fast() - pin user pages in memory
376 * @start: starting user address
377 * @nr_pages: number of pages from start to pin
378 * @write: whether pages will be written to
379 * @pages: array that receives pointers to the pages pinned.
380 * Should be at least nr_pages long.
382 * Attempt to pin user pages in memory without taking mm->mmap_sem.
383 * If not successful, it will fall back to taking the lock and
384 * calling get_user_pages().
386 * Returns number of pages pinned. This may be fewer than the number
387 * requested. If nr_pages is 0 or negative, returns 0. If no pages
388 * were pinned, returns -errno.
390 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
393 struct mm_struct *mm = current->mm;
394 unsigned long addr, len, end;
401 len = (unsigned long) nr_pages << PAGE_SHIFT;
408 if (end >> __VIRTUAL_MASK_SHIFT)
413 * XXX: batch / limit 'nr', to avoid large irq off latency
414 * needs some instrumenting to determine the common sizes used by
415 * important workloads (eg. DB2), and whether limiting the batch size
416 * will decrease performance.
418 * It seems like we're in the clear for the moment. Direct-IO is
419 * the main guy that batches up lots of get_user_pages, and even
420 * they are limited to 64-at-a-time which is not so many.
423 * This doesn't prevent pagetable teardown, but does prevent
424 * the pagetables and pages from being freed on x86.
426 * So long as we atomically load page table pointers versus teardown
427 * (which we do on x86, with the above PAE exception), we can follow the
428 * address down to the the page and take a ref on it.
431 pgdp = pgd_offset(mm, addr);
435 next = pgd_addr_end(addr, end);
438 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
440 } while (pgdp++, addr = next, addr != end);
443 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
452 /* Try to get the remaining pages with get_user_pages */
453 start += nr << PAGE_SHIFT;
456 ret = get_user_pages_unlocked(start,
457 (end - start) >> PAGE_SHIFT,
458 pages, write ? FOLL_WRITE : 0);
460 /* Have to be a bit careful with return values */