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>
13 #include <asm/pgtable.h>
15 static inline pte_t gup_get_pte(pte_t *ptep)
17 #ifndef CONFIG_X86_PAE
18 return READ_ONCE(*ptep);
21 * With get_user_pages_fast, we walk down the pagetables without taking
22 * any locks. For this we would like to load the pointers atomically,
23 * but that is not possible (without expensive cmpxchg8b) on PAE. What
24 * we do have is the guarantee that a pte will only either go from not
25 * present to present, or present to not present or both -- it will not
26 * switch to a completely different present page without a TLB flush in
27 * between; something that we are blocking by holding interrupts off.
29 * Setting ptes from not present to present goes:
34 * And present to not present goes:
39 * We must ensure here that the load of pte_low sees l iff pte_high
40 * sees h. We load pte_high *after* loading pte_low, which ensures we
41 * don't see an older value of pte_high. *Then* we recheck pte_low,
42 * which ensures that we haven't picked up a changed pte high. We might
43 * have got rubbish values from pte_low and pte_high, but we are
44 * guaranteed that pte_low will not have the present bit set *unless*
45 * it is 'l'. And get_user_pages_fast only operates on present ptes, so
48 * gup_get_pte should not be used or copied outside gup.c without being
49 * very careful -- it does not atomically load the pte or anything that
50 * is likely to be useful for you.
55 pte.pte_low = ptep->pte_low;
57 pte.pte_high = ptep->pte_high;
59 if (unlikely(pte.pte_low != ptep->pte_low))
67 * The performance critical leaf functions are made noinline otherwise gcc
68 * inlines everything into a single function which results in too much
71 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
72 unsigned long end, int write, struct page **pages, int *nr)
77 mask = _PAGE_PRESENT|_PAGE_USER;
81 ptep = pte_offset_map(&pmd, addr);
83 pte_t pte = gup_get_pte(ptep);
86 /* Similar to the PMD case, NUMA hinting must take slow path */
87 if (pte_protnone(pte)) {
92 if ((pte_flags(pte) & (mask | _PAGE_SPECIAL)) != mask) {
96 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
98 if (unlikely(!try_get_page(page))) {
102 SetPageReferenced(page);
106 } while (ptep++, addr += PAGE_SIZE, addr != end);
112 static inline void get_head_page_multiple(struct page *page, int nr)
114 VM_BUG_ON_PAGE(page != compound_head(page), page);
115 VM_BUG_ON_PAGE(page_count(page) == 0, page);
116 atomic_add(nr, &page->_count);
117 SetPageReferenced(page);
120 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
121 unsigned long end, int write, struct page **pages, int *nr)
124 struct page *head, *page;
127 mask = _PAGE_PRESENT|_PAGE_USER;
130 if ((pmd_flags(pmd) & mask) != mask)
132 /* hugepages are never "special" */
133 VM_BUG_ON(pmd_flags(pmd) & _PAGE_SPECIAL);
134 VM_BUG_ON(!pfn_valid(pmd_pfn(pmd)));
137 head = pmd_page(pmd);
138 if (WARN_ON_ONCE(page_ref_count(head) <= 0))
140 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
142 VM_BUG_ON_PAGE(compound_head(page) != head, page);
145 get_huge_page_tail(page);
149 } while (addr += PAGE_SIZE, addr != end);
150 get_head_page_multiple(head, refs);
155 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
156 int write, struct page **pages, int *nr)
161 pmdp = pmd_offset(&pud, addr);
165 next = pmd_addr_end(addr, end);
167 * The pmd_trans_splitting() check below explains why
168 * pmdp_splitting_flush has to flush the tlb, to stop
169 * this gup-fast code from running while we set the
170 * splitting bit in the pmd. Returning zero will take
171 * the slow path that will call wait_split_huge_page()
172 * if the pmd is still in splitting state. gup-fast
173 * can't because it has irq disabled and
174 * wait_split_huge_page() would never return as the
175 * tlb flush IPI wouldn't run.
177 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
179 if (unlikely(pmd_large(pmd) || !pmd_present(pmd))) {
181 * NUMA hinting faults need to be handled in the GUP
182 * slowpath for accounting purposes and so that they
183 * can be serialised against THP migration.
185 if (pmd_protnone(pmd))
187 if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
190 if (!gup_pte_range(pmd, addr, next, write, pages, nr))
193 } while (pmdp++, addr = next, addr != end);
198 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
199 unsigned long end, int write, struct page **pages, int *nr)
202 struct page *head, *page;
205 mask = _PAGE_PRESENT|_PAGE_USER;
208 if ((pud_flags(pud) & mask) != mask)
210 /* hugepages are never "special" */
211 VM_BUG_ON(pud_flags(pud) & _PAGE_SPECIAL);
212 VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
215 head = pud_page(pud);
216 if (WARN_ON_ONCE(page_ref_count(head) <= 0))
218 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
220 VM_BUG_ON_PAGE(compound_head(page) != head, page);
223 get_huge_page_tail(page);
227 } while (addr += PAGE_SIZE, addr != end);
228 get_head_page_multiple(head, refs);
233 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
234 int write, struct page **pages, int *nr)
239 pudp = pud_offset(&pgd, addr);
243 next = pud_addr_end(addr, end);
246 if (unlikely(pud_large(pud))) {
247 if (!gup_huge_pud(pud, addr, next, write, pages, nr))
250 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
253 } while (pudp++, addr = next, addr != end);
259 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
260 * back to the regular GUP.
262 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
265 struct mm_struct *mm = current->mm;
266 unsigned long addr, len, end;
274 len = (unsigned long) nr_pages << PAGE_SHIFT;
276 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
277 (void __user *)start, len)))
281 * XXX: batch / limit 'nr', to avoid large irq off latency
282 * needs some instrumenting to determine the common sizes used by
283 * important workloads (eg. DB2), and whether limiting the batch size
284 * will decrease performance.
286 * It seems like we're in the clear for the moment. Direct-IO is
287 * the main guy that batches up lots of get_user_pages, and even
288 * they are limited to 64-at-a-time which is not so many.
291 * This doesn't prevent pagetable teardown, but does prevent
292 * the pagetables and pages from being freed on x86.
294 * So long as we atomically load page table pointers versus teardown
295 * (which we do on x86, with the above PAE exception), we can follow the
296 * address down to the the page and take a ref on it.
298 local_irq_save(flags);
299 pgdp = pgd_offset(mm, addr);
303 next = pgd_addr_end(addr, end);
306 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
308 } while (pgdp++, addr = next, addr != end);
309 local_irq_restore(flags);
315 * get_user_pages_fast() - pin user pages in memory
316 * @start: starting user address
317 * @nr_pages: number of pages from start to pin
318 * @write: whether pages will be written to
319 * @pages: array that receives pointers to the pages pinned.
320 * Should be at least nr_pages long.
322 * Attempt to pin user pages in memory without taking mm->mmap_sem.
323 * If not successful, it will fall back to taking the lock and
324 * calling get_user_pages().
326 * Returns number of pages pinned. This may be fewer than the number
327 * requested. If nr_pages is 0 or negative, returns 0. If no pages
328 * were pinned, returns -errno.
330 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
333 struct mm_struct *mm = current->mm;
334 unsigned long addr, len, end;
341 len = (unsigned long) nr_pages << PAGE_SHIFT;
348 if (end >> __VIRTUAL_MASK_SHIFT)
353 * XXX: batch / limit 'nr', to avoid large irq off latency
354 * needs some instrumenting to determine the common sizes used by
355 * important workloads (eg. DB2), and whether limiting the batch size
356 * will decrease performance.
358 * It seems like we're in the clear for the moment. Direct-IO is
359 * the main guy that batches up lots of get_user_pages, and even
360 * they are limited to 64-at-a-time which is not so many.
363 * This doesn't prevent pagetable teardown, but does prevent
364 * the pagetables and pages from being freed on x86.
366 * So long as we atomically load page table pointers versus teardown
367 * (which we do on x86, with the above PAE exception), we can follow the
368 * address down to the the page and take a ref on it.
371 pgdp = pgd_offset(mm, addr);
375 next = pgd_addr_end(addr, end);
378 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
380 } while (pgdp++, addr = next, addr != end);
383 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
392 /* Try to get the remaining pages with get_user_pages */
393 start += nr << PAGE_SHIFT;
396 ret = get_user_pages_unlocked(current, mm, start,
397 (end - start) >> PAGE_SHIFT,
398 pages, write ? FOLL_WRITE : 0);
400 /* Have to be a bit careful with return values */