arch/x86/mm/gup.c

   1 /*
   2  * Lockless get_user_pages_fast for x86
   3  *
   4  * Copyright (C) 2008 Nick Piggin
   5  * Copyright (C) 2008 Novell Inc.
   6  */
   7 #include <linux/sched.h>
   8 #include <linux/mm.h>
   9 #include <linux/vmstat.h>
  10 #include <linux/highmem.h>
  11 #include <linux/swap.h>
  12
  13 #include <asm/pgtable.h>
  14
  15 static inline pte_t gup_get_pte(pte_t *ptep)
  16 {
  17 #ifndef CONFIG_X86_PAE
  18         return ACCESS_ONCE(*ptep);
  19 #else
  20         /*
  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.
  28          *
  29          * Setting ptes from not present to present goes:
  30          * ptep->pte_high = h;
  31          * smp_wmb();
  32          * ptep->pte_low = l;
  33          *
  34          * And present to not present goes:
  35          * ptep->pte_low = 0;
  36          * smp_wmb();
  37          * ptep->pte_high = 0;
  38          *
  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
  46          * we're safe.
  47          *
  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.
  51          */
  52         pte_t pte;
  53
  54 retry:
  55         pte.pte_low = ptep->pte_low;
  56         smp_rmb();
  57         pte.pte_high = ptep->pte_high;
  58         smp_rmb();
  59         if (unlikely(pte.pte_low != ptep->pte_low))
  60                 goto retry;
  61
  62         return pte;
  63 #endif
  64 }
  65
  66 /*
  67  * The performance critical leaf functions are made noinline otherwise gcc
  68  * inlines everything into a single function which results in too much
  69  * register pressure.
  70  */
  71 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
  72                 unsigned long end, int write, struct page **pages, int *nr)
  73 {
  74         unsigned long mask;
  75         pte_t *ptep;
  76
  77         mask = _PAGE_PRESENT|_PAGE_USER;
  78         if (write)
  79                 mask |= _PAGE_RW;
  80
  81         ptep = pte_offset_map(&pmd, addr);
  82         do {
  83                 pte_t pte = gup_get_pte(ptep);
  84                 struct page *page;
  85
  86                 if ((pte_flags(pte) & (mask | _PAGE_SPECIAL)) != mask) {
  87                         pte_unmap(ptep);
  88                         return 0;
  89                 }
  90                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  91                 page = pte_page(pte);
  92                 get_page(page);
  93                 SetPageReferenced(page);
  94                 pages[*nr] = page;
  95                 (*nr)++;
  96
  97         } while (ptep++, addr += PAGE_SIZE, addr != end);
  98         pte_unmap(ptep - 1);
  99
 100         return 1;
 101 }
 102
 103 static inline void get_head_page_multiple(struct page *page, int nr)
 104 {
 105         VM_BUG_ON(page != compound_head(page));
 106         VM_BUG_ON(page_count(page) == 0);
 107         atomic_add(nr, &page->_count);
 108         SetPageReferenced(page);
 109 }
 110
 111 static inline void get_huge_page_tail(struct page *page)
 112 {
 113         /*
 114          * __split_huge_page_refcount() cannot run
 115          * from under us.
 116          */
 117         VM_BUG_ON(page_mapcount(page) < 0);
 118         VM_BUG_ON(atomic_read(&page->_count) != 0);
 119         atomic_inc(&page->_mapcount);
 120 }
 121
 122 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
 123                 unsigned long end, int write, struct page **pages, int *nr)
 124 {
 125         unsigned long mask;
 126         pte_t pte = *(pte_t *)&pmd;
 127         struct page *head, *page;
 128         int refs;
 129
 130         mask = _PAGE_PRESENT|_PAGE_USER;
 131         if (write)
 132                 mask |= _PAGE_RW;
 133         if ((pte_flags(pte) & mask) != mask)
 134                 return 0;
 135         /* hugepages are never "special" */
 136         VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
 137         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 138
 139         refs = 0;
 140         head = pte_page(pte);
 141         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 142         do {
 143                 VM_BUG_ON(compound_head(page) != head);
 144                 pages[*nr] = page;
 145                 if (PageTail(page))
 146                         get_huge_page_tail(page);
 147                 (*nr)++;
 148                 page++;
 149                 refs++;
 150         } while (addr += PAGE_SIZE, addr != end);
 151         get_head_page_multiple(head, refs);
 152
 153         return 1;
 154 }
 155
 156 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 157                 int write, struct page **pages, int *nr)
 158 {
 159         unsigned long next;
 160         pmd_t *pmdp;
 161
 162         pmdp = pmd_offset(&pud, addr);
 163         do {
 164                 pmd_t pmd = *pmdp;
 165
 166                 next = pmd_addr_end(addr, end);
 167                 /*
 168                  * The pmd_trans_splitting() check below explains why
 169                  * pmdp_splitting_flush has to flush the tlb, to stop
 170                  * this gup-fast code from running while we set the
 171                  * splitting bit in the pmd. Returning zero will take
 172                  * the slow path that will call wait_split_huge_page()
 173                  * if the pmd is still in splitting state. gup-fast
 174                  * can't because it has irq disabled and
 175                  * wait_split_huge_page() would never return as the
 176                  * tlb flush IPI wouldn't run.
 177                  */
 178                 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
 179                         return 0;
 180                 if (unlikely(pmd_large(pmd))) {
 181                         if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
 182                                 return 0;
 183                 } else {
 184                         if (!gup_pte_range(pmd, addr, next, write, pages, nr))
 185                                 return 0;
 186                 }
 187         } while (pmdp++, addr = next, addr != end);
 188
 189         return 1;
 190 }
 191
 192 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
 193                 unsigned long end, int write, struct page **pages, int *nr)
 194 {
 195         unsigned long mask;
 196         pte_t pte = *(pte_t *)&pud;
 197         struct page *head, *page;
 198         int refs;
 199
 200         mask = _PAGE_PRESENT|_PAGE_USER;
 201         if (write)
 202                 mask |= _PAGE_RW;
 203         if ((pte_flags(pte) & mask) != mask)
 204                 return 0;
 205         /* hugepages are never "special" */
 206         VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
 207         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 208
 209         refs = 0;
 210         head = pte_page(pte);
 211         page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 212         do {
 213                 VM_BUG_ON(compound_head(page) != head);
 214                 pages[*nr] = page;
 215                 (*nr)++;
 216                 page++;
 217                 refs++;
 218         } while (addr += PAGE_SIZE, addr != end);
 219         get_head_page_multiple(head, refs);
 220
 221         return 1;
 222 }
 223
 224 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 225                         int write, struct page **pages, int *nr)
 226 {
 227         unsigned long next;
 228         pud_t *pudp;
 229
 230         pudp = pud_offset(&pgd, addr);
 231         do {
 232                 pud_t pud = *pudp;
 233
 234                 next = pud_addr_end(addr, end);
 235                 if (pud_none(pud))
 236                         return 0;
 237                 if (unlikely(pud_large(pud))) {
 238                         if (!gup_huge_pud(pud, addr, next, write, pages, nr))
 239                                 return 0;
 240                 } else {
 241                         if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 242                                 return 0;
 243                 }
 244         } while (pudp++, addr = next, addr != end);
 245
 246         return 1;
 247 }
 248
 249 /*
 250  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 251  * back to the regular GUP.
 252  */
 253 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 254                           struct page **pages)
 255 {
 256         struct mm_struct *mm = current->mm;
 257         unsigned long addr, len, end;
 258         unsigned long next;
 259         unsigned long flags;
 260         pgd_t *pgdp;
 261         int nr = 0;
 262
 263         start &= PAGE_MASK;
 264         addr = start;
 265         len = (unsigned long) nr_pages << PAGE_SHIFT;
 266         end = start + len;
 267         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
 268                                         (void __user *)start, len)))
 269                 return 0;
 270
 271         /*
 272          * XXX: batch / limit 'nr', to avoid large irq off latency
 273          * needs some instrumenting to determine the common sizes used by
 274          * important workloads (eg. DB2), and whether limiting the batch size
 275          * will decrease performance.
 276          *
 277          * It seems like we're in the clear for the moment. Direct-IO is
 278          * the main guy that batches up lots of get_user_pages, and even
 279          * they are limited to 64-at-a-time which is not so many.
 280          */
 281         /*
 282          * This doesn't prevent pagetable teardown, but does prevent
 283          * the pagetables and pages from being freed on x86.
 284          *
 285          * So long as we atomically load page table pointers versus teardown
 286          * (which we do on x86, with the above PAE exception), we can follow the
 287          * address down to the the page and take a ref on it.
 288          */
 289         local_irq_save(flags);
 290         pgdp = pgd_offset(mm, addr);
 291         do {
 292                 pgd_t pgd = *pgdp;
 293
 294                 next = pgd_addr_end(addr, end);
 295                 if (pgd_none(pgd))
 296                         break;
 297                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 298                         break;
 299         } while (pgdp++, addr = next, addr != end);
 300         local_irq_restore(flags);
 301
 302         return nr;
 303 }
 304
 305 /**
 306  * get_user_pages_fast() - pin user pages in memory
 307  * @start:      starting user address
 308  * @nr_pages:   number of pages from start to pin
 309  * @write:      whether pages will be written to
 310  * @pages:      array that receives pointers to the pages pinned.
 311  *              Should be at least nr_pages long.
 312  *
 313  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 314  * If not successful, it will fall back to taking the lock and
 315  * calling get_user_pages().
 316  *
 317  * Returns number of pages pinned. This may be fewer than the number
 318  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 319  * were pinned, returns -errno.
 320  */
 321 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 322                         struct page **pages)
 323 {
 324         struct mm_struct *mm = current->mm;
 325         unsigned long addr, len, end;
 326         unsigned long next;
 327         pgd_t *pgdp;
 328         int nr = 0;
 329
 330         start &= PAGE_MASK;
 331         addr = start;
 332         len = (unsigned long) nr_pages << PAGE_SHIFT;
 333
 334         end = start + len;
 335         if (end < start)
 336                 goto slow_irqon;
 337
 338 #ifdef CONFIG_X86_64
 339         if (end >> __VIRTUAL_MASK_SHIFT)
 340                 goto slow_irqon;
 341 #endif
 342
 343         /*
 344          * XXX: batch / limit 'nr', to avoid large irq off latency
 345          * needs some instrumenting to determine the common sizes used by
 346          * important workloads (eg. DB2), and whether limiting the batch size
 347          * will decrease performance.
 348          *
 349          * It seems like we're in the clear for the moment. Direct-IO is
 350          * the main guy that batches up lots of get_user_pages, and even
 351          * they are limited to 64-at-a-time which is not so many.
 352          */
 353         /*
 354          * This doesn't prevent pagetable teardown, but does prevent
 355          * the pagetables and pages from being freed on x86.
 356          *
 357          * So long as we atomically load page table pointers versus teardown
 358          * (which we do on x86, with the above PAE exception), we can follow the
 359          * address down to the the page and take a ref on it.
 360          */
 361         local_irq_disable();
 362         pgdp = pgd_offset(mm, addr);
 363         do {
 364                 pgd_t pgd = *pgdp;
 365
 366                 next = pgd_addr_end(addr, end);
 367                 if (pgd_none(pgd))
 368                         goto slow;
 369                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 370                         goto slow;
 371         } while (pgdp++, addr = next, addr != end);
 372         local_irq_enable();
 373
 374         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 375         return nr;
 376
 377         {
 378                 int ret;
 379
 380 slow:
 381                 local_irq_enable();
 382 slow_irqon:
 383                 /* Try to get the remaining pages with get_user_pages */
 384                 start += nr << PAGE_SHIFT;
 385                 pages += nr;
 386
 387                 down_read(&mm->mmap_sem);
 388                 ret = get_user_pages(current, mm, start,
 389                         (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
 390                 up_read(&mm->mmap_sem);
 391
 392                 /* Have to be a bit careful with return values */
 393                 if (nr > 0) {
 394                         if (ret < 0)
 395                                 ret = nr;
 396                         else
 397                                 ret += nr;
 398                 }
 399
 400                 return ret;
 401         }
 402 }