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 noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
 112                 unsigned long end, int write, struct page **pages, int *nr)
 113 {
 114         unsigned long mask;
 115         pte_t pte = *(pte_t *)&pmd;
 116         struct page *head, *page;
 117         int refs;
 118
 119         mask = _PAGE_PRESENT|_PAGE_USER;
 120         if (write)
 121                 mask |= _PAGE_RW;
 122         if ((pte_flags(pte) & mask) != mask)
 123                 return 0;
 124         /* hugepages are never "special" */
 125         VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
 126         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 127
 128         refs = 0;
 129         head = pte_page(pte);
 130         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 131         do {
 132                 VM_BUG_ON(compound_head(page) != head);
 133                 pages[*nr] = page;
 134                 if (PageTail(page))
 135                         get_huge_page_tail(page);
 136                 (*nr)++;
 137                 page++;
 138                 refs++;
 139         } while (addr += PAGE_SIZE, addr != end);
 140         get_head_page_multiple(head, refs);
 141
 142         return 1;
 143 }
 144
 145 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 146                 int write, struct page **pages, int *nr)
 147 {
 148         unsigned long next;
 149         pmd_t *pmdp;
 150
 151         pmdp = pmd_offset(&pud, addr);
 152         do {
 153                 pmd_t pmd = *pmdp;
 154
 155                 next = pmd_addr_end(addr, end);
 156                 /*
 157                  * The pmd_trans_splitting() check below explains why
 158                  * pmdp_splitting_flush has to flush the tlb, to stop
 159                  * this gup-fast code from running while we set the
 160                  * splitting bit in the pmd. Returning zero will take
 161                  * the slow path that will call wait_split_huge_page()
 162                  * if the pmd is still in splitting state. gup-fast
 163                  * can't because it has irq disabled and
 164                  * wait_split_huge_page() would never return as the
 165                  * tlb flush IPI wouldn't run.
 166                  */
 167                 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
 168                         return 0;
 169                 if (unlikely(pmd_large(pmd))) {
 170                         if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
 171                                 return 0;
 172                 } else {
 173                         if (!gup_pte_range(pmd, addr, next, write, pages, nr))
 174                                 return 0;
 175                 }
 176         } while (pmdp++, addr = next, addr != end);
 177
 178         return 1;
 179 }
 180
 181 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
 182                 unsigned long end, int write, struct page **pages, int *nr)
 183 {
 184         unsigned long mask;
 185         pte_t pte = *(pte_t *)&pud;
 186         struct page *head, *page;
 187         int refs;
 188
 189         mask = _PAGE_PRESENT|_PAGE_USER;
 190         if (write)
 191                 mask |= _PAGE_RW;
 192         if ((pte_flags(pte) & mask) != mask)
 193                 return 0;
 194         /* hugepages are never "special" */
 195         VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
 196         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 197
 198         refs = 0;
 199         head = pte_page(pte);
 200         page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 201         do {
 202                 VM_BUG_ON(compound_head(page) != head);
 203                 pages[*nr] = page;
 204                 (*nr)++;
 205                 page++;
 206                 refs++;
 207         } while (addr += PAGE_SIZE, addr != end);
 208         get_head_page_multiple(head, refs);
 209
 210         return 1;
 211 }
 212
 213 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 214                         int write, struct page **pages, int *nr)
 215 {
 216         unsigned long next;
 217         pud_t *pudp;
 218
 219         pudp = pud_offset(&pgd, addr);
 220         do {
 221                 pud_t pud = *pudp;
 222
 223                 next = pud_addr_end(addr, end);
 224                 if (pud_none(pud))
 225                         return 0;
 226                 if (unlikely(pud_large(pud))) {
 227                         if (!gup_huge_pud(pud, addr, next, write, pages, nr))
 228                                 return 0;
 229                 } else {
 230                         if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 231                                 return 0;
 232                 }
 233         } while (pudp++, addr = next, addr != end);
 234
 235         return 1;
 236 }
 237
 238 /*
 239  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 240  * back to the regular GUP.
 241  */
 242 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 243                           struct page **pages)
 244 {
 245         struct mm_struct *mm = current->mm;
 246         unsigned long addr, len, end;
 247         unsigned long next;
 248         unsigned long flags;
 249         pgd_t *pgdp;
 250         int nr = 0;
 251
 252         start &= PAGE_MASK;
 253         addr = start;
 254         len = (unsigned long) nr_pages << PAGE_SHIFT;
 255         end = start + len;
 256         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
 257                                         (void __user *)start, len)))
 258                 return 0;
 259
 260         /*
 261          * XXX: batch / limit 'nr', to avoid large irq off latency
 262          * needs some instrumenting to determine the common sizes used by
 263          * important workloads (eg. DB2), and whether limiting the batch size
 264          * will decrease performance.
 265          *
 266          * It seems like we're in the clear for the moment. Direct-IO is
 267          * the main guy that batches up lots of get_user_pages, and even
 268          * they are limited to 64-at-a-time which is not so many.
 269          */
 270         /*
 271          * This doesn't prevent pagetable teardown, but does prevent
 272          * the pagetables and pages from being freed on x86.
 273          *
 274          * So long as we atomically load page table pointers versus teardown
 275          * (which we do on x86, with the above PAE exception), we can follow the
 276          * address down to the the page and take a ref on it.
 277          */
 278         local_irq_save(flags);
 279         pgdp = pgd_offset(mm, addr);
 280         do {
 281                 pgd_t pgd = *pgdp;
 282
 283                 next = pgd_addr_end(addr, end);
 284                 if (pgd_none(pgd))
 285                         break;
 286                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 287                         break;
 288         } while (pgdp++, addr = next, addr != end);
 289         local_irq_restore(flags);
 290
 291         return nr;
 292 }
 293
 294 /**
 295  * get_user_pages_fast() - pin user pages in memory
 296  * @start:      starting user address
 297  * @nr_pages:   number of pages from start to pin
 298  * @write:      whether pages will be written to
 299  * @pages:      array that receives pointers to the pages pinned.
 300  *              Should be at least nr_pages long.
 301  *
 302  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 303  * If not successful, it will fall back to taking the lock and
 304  * calling get_user_pages().
 305  *
 306  * Returns number of pages pinned. This may be fewer than the number
 307  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 308  * were pinned, returns -errno.
 309  */
 310 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 311                         struct page **pages)
 312 {
 313         struct mm_struct *mm = current->mm;
 314         unsigned long addr, len, end;
 315         unsigned long next;
 316         pgd_t *pgdp;
 317         int nr = 0;
 318
 319         start &= PAGE_MASK;
 320         addr = start;
 321         len = (unsigned long) nr_pages << PAGE_SHIFT;
 322
 323         end = start + len;
 324         if (end < start)
 325                 goto slow_irqon;
 326
 327 #ifdef CONFIG_X86_64
 328         if (end >> __VIRTUAL_MASK_SHIFT)
 329                 goto slow_irqon;
 330 #endif
 331
 332         /*
 333          * XXX: batch / limit 'nr', to avoid large irq off latency
 334          * needs some instrumenting to determine the common sizes used by
 335          * important workloads (eg. DB2), and whether limiting the batch size
 336          * will decrease performance.
 337          *
 338          * It seems like we're in the clear for the moment. Direct-IO is
 339          * the main guy that batches up lots of get_user_pages, and even
 340          * they are limited to 64-at-a-time which is not so many.
 341          */
 342         /*
 343          * This doesn't prevent pagetable teardown, but does prevent
 344          * the pagetables and pages from being freed on x86.
 345          *
 346          * So long as we atomically load page table pointers versus teardown
 347          * (which we do on x86, with the above PAE exception), we can follow the
 348          * address down to the the page and take a ref on it.
 349          */
 350         local_irq_disable();
 351         pgdp = pgd_offset(mm, addr);
 352         do {
 353                 pgd_t pgd = *pgdp;
 354
 355                 next = pgd_addr_end(addr, end);
 356                 if (pgd_none(pgd))
 357                         goto slow;
 358                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 359                         goto slow;
 360         } while (pgdp++, addr = next, addr != end);
 361         local_irq_enable();
 362
 363         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 364         return nr;
 365
 366         {
 367                 int ret;
 368
 369 slow:
 370                 local_irq_enable();
 371 slow_irqon:
 372                 /* Try to get the remaining pages with get_user_pages */
 373                 start += nr << PAGE_SHIFT;
 374                 pages += nr;
 375
 376                 down_read(&mm->mmap_sem);
 377                 ret = get_user_pages(current, mm, start,
 378                         (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
 379                 up_read(&mm->mmap_sem);
 380
 381                 /* Have to be a bit careful with return values */
 382                 if (nr > 0) {
 383                         if (ret < 0)
 384                                 ret = nr;
 385                         else
 386                                 ret += nr;
 387                 }
 388
 389                 return ret;
 390         }
 391 }