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 READ_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                 /* Similar to the PMD case, NUMA hinting must take slow path */
  87                 if (pte_protnone(pte)) {
  88                         pte_unmap(ptep);
  89                         return 0;
  90                 }
  91
  92                 if ((pte_flags(pte) & (mask | _PAGE_SPECIAL)) != mask) {
  93                         pte_unmap(ptep);
  94                         return 0;
  95                 }
  96                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  97                 page = pte_page(pte);
  98                 get_page(page);
  99                 SetPageReferenced(page);
 100                 pages[*nr] = page;
 101                 (*nr)++;
 102
 103         } while (ptep++, addr += PAGE_SIZE, addr != end);
 104         pte_unmap(ptep - 1);
 105
 106         return 1;
 107 }
 108
 109 static inline void get_head_page_multiple(struct page *page, int nr)
 110 {
 111         VM_BUG_ON_PAGE(page != compound_head(page), page);
 112         VM_BUG_ON_PAGE(page_count(page) == 0, page);
 113         atomic_add(nr, &page->_count);
 114         SetPageReferenced(page);
 115 }
 116
 117 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
 118                 unsigned long end, int write, struct page **pages, int *nr)
 119 {
 120         unsigned long mask;
 121         struct page *head, *page;
 122         int refs;
 123
 124         mask = _PAGE_PRESENT|_PAGE_USER;
 125         if (write)
 126                 mask |= _PAGE_RW;
 127         if ((pmd_flags(pmd) & mask) != mask)
 128                 return 0;
 129         /* hugepages are never "special" */
 130         VM_BUG_ON(pmd_flags(pmd) & _PAGE_SPECIAL);
 131         VM_BUG_ON(!pfn_valid(pmd_pfn(pmd)));
 132
 133         refs = 0;
 134         head = pmd_page(pmd);
 135         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 136         do {
 137                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
 138                 pages[*nr] = page;
 139                 if (PageTail(page))
 140                         get_huge_page_tail(page);
 141                 (*nr)++;
 142                 page++;
 143                 refs++;
 144         } while (addr += PAGE_SIZE, addr != end);
 145         get_head_page_multiple(head, refs);
 146
 147         return 1;
 148 }
 149
 150 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 151                 int write, struct page **pages, int *nr)
 152 {
 153         unsigned long next;
 154         pmd_t *pmdp;
 155
 156         pmdp = pmd_offset(&pud, addr);
 157         do {
 158                 pmd_t pmd = *pmdp;
 159
 160                 next = pmd_addr_end(addr, end);
 161                 /*
 162                  * The pmd_trans_splitting() check below explains why
 163                  * pmdp_splitting_flush has to flush the tlb, to stop
 164                  * this gup-fast code from running while we set the
 165                  * splitting bit in the pmd. Returning zero will take
 166                  * the slow path that will call wait_split_huge_page()
 167                  * if the pmd is still in splitting state. gup-fast
 168                  * can't because it has irq disabled and
 169                  * wait_split_huge_page() would never return as the
 170                  * tlb flush IPI wouldn't run.
 171                  */
 172                 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
 173                         return 0;
 174                 if (unlikely(pmd_large(pmd) || !pmd_present(pmd))) {
 175                         /*
 176                          * NUMA hinting faults need to be handled in the GUP
 177                          * slowpath for accounting purposes and so that they
 178                          * can be serialised against THP migration.
 179                          */
 180                         if (pmd_protnone(pmd))
 181                                 return 0;
 182                         if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
 183                                 return 0;
 184                 } else {
 185                         if (!gup_pte_range(pmd, addr, next, write, pages, nr))
 186                                 return 0;
 187                 }
 188         } while (pmdp++, addr = next, addr != end);
 189
 190         return 1;
 191 }
 192
 193 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
 194                 unsigned long end, int write, struct page **pages, int *nr)
 195 {
 196         unsigned long mask;
 197         struct page *head, *page;
 198         int refs;
 199
 200         mask = _PAGE_PRESENT|_PAGE_USER;
 201         if (write)
 202                 mask |= _PAGE_RW;
 203         if ((pud_flags(pud) & mask) != mask)
 204                 return 0;
 205         /* hugepages are never "special" */
 206         VM_BUG_ON(pud_flags(pud) & _PAGE_SPECIAL);
 207         VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
 208
 209         refs = 0;
 210         head = pud_page(pud);
 211         page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 212         do {
 213                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
 214                 pages[*nr] = page;
 215                 if (PageTail(page))
 216                         get_huge_page_tail(page);
 217                 (*nr)++;
 218                 page++;
 219                 refs++;
 220         } while (addr += PAGE_SIZE, addr != end);
 221         get_head_page_multiple(head, refs);
 222
 223         return 1;
 224 }
 225
 226 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 227                         int write, struct page **pages, int *nr)
 228 {
 229         unsigned long next;
 230         pud_t *pudp;
 231
 232         pudp = pud_offset(&pgd, addr);
 233         do {
 234                 pud_t pud = *pudp;
 235
 236                 next = pud_addr_end(addr, end);
 237                 if (pud_none(pud))
 238                         return 0;
 239                 if (unlikely(pud_large(pud))) {
 240                         if (!gup_huge_pud(pud, addr, next, write, pages, nr))
 241                                 return 0;
 242                 } else {
 243                         if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 244                                 return 0;
 245                 }
 246         } while (pudp++, addr = next, addr != end);
 247
 248         return 1;
 249 }
 250
 251 /*
 252  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 253  * back to the regular GUP.
 254  */
 255 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 256                           struct page **pages)
 257 {
 258         struct mm_struct *mm = current->mm;
 259         unsigned long addr, len, end;
 260         unsigned long next;
 261         unsigned long flags;
 262         pgd_t *pgdp;
 263         int nr = 0;
 264
 265         start &= PAGE_MASK;
 266         addr = start;
 267         len = (unsigned long) nr_pages << PAGE_SHIFT;
 268         end = start + len;
 269         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
 270                                         (void __user *)start, len)))
 271                 return 0;
 272
 273         /*
 274          * XXX: batch / limit 'nr', to avoid large irq off latency
 275          * needs some instrumenting to determine the common sizes used by
 276          * important workloads (eg. DB2), and whether limiting the batch size
 277          * will decrease performance.
 278          *
 279          * It seems like we're in the clear for the moment. Direct-IO is
 280          * the main guy that batches up lots of get_user_pages, and even
 281          * they are limited to 64-at-a-time which is not so many.
 282          */
 283         /*
 284          * This doesn't prevent pagetable teardown, but does prevent
 285          * the pagetables and pages from being freed on x86.
 286          *
 287          * So long as we atomically load page table pointers versus teardown
 288          * (which we do on x86, with the above PAE exception), we can follow the
 289          * address down to the the page and take a ref on it.
 290          */
 291         local_irq_save(flags);
 292         pgdp = pgd_offset(mm, addr);
 293         do {
 294                 pgd_t pgd = *pgdp;
 295
 296                 next = pgd_addr_end(addr, end);
 297                 if (pgd_none(pgd))
 298                         break;
 299                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 300                         break;
 301         } while (pgdp++, addr = next, addr != end);
 302         local_irq_restore(flags);
 303
 304         return nr;
 305 }
 306
 307 /**
 308  * get_user_pages_fast() - pin user pages in memory
 309  * @start:      starting user address
 310  * @nr_pages:   number of pages from start to pin
 311  * @write:      whether pages will be written to
 312  * @pages:      array that receives pointers to the pages pinned.
 313  *              Should be at least nr_pages long.
 314  *
 315  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 316  * If not successful, it will fall back to taking the lock and
 317  * calling get_user_pages().
 318  *
 319  * Returns number of pages pinned. This may be fewer than the number
 320  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 321  * were pinned, returns -errno.
 322  */
 323 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 324                         struct page **pages)
 325 {
 326         struct mm_struct *mm = current->mm;
 327         unsigned long addr, len, end;
 328         unsigned long next;
 329         pgd_t *pgdp;
 330         int nr = 0;
 331
 332         start &= PAGE_MASK;
 333         addr = start;
 334         len = (unsigned long) nr_pages << PAGE_SHIFT;
 335
 336         end = start + len;
 337         if (end < start)
 338                 goto slow_irqon;
 339
 340 #ifdef CONFIG_X86_64
 341         if (end >> __VIRTUAL_MASK_SHIFT)
 342                 goto slow_irqon;
 343 #endif
 344
 345         /*
 346          * XXX: batch / limit 'nr', to avoid large irq off latency
 347          * needs some instrumenting to determine the common sizes used by
 348          * important workloads (eg. DB2), and whether limiting the batch size
 349          * will decrease performance.
 350          *
 351          * It seems like we're in the clear for the moment. Direct-IO is
 352          * the main guy that batches up lots of get_user_pages, and even
 353          * they are limited to 64-at-a-time which is not so many.
 354          */
 355         /*
 356          * This doesn't prevent pagetable teardown, but does prevent
 357          * the pagetables and pages from being freed on x86.
 358          *
 359          * So long as we atomically load page table pointers versus teardown
 360          * (which we do on x86, with the above PAE exception), we can follow the
 361          * address down to the the page and take a ref on it.
 362          */
 363         local_irq_disable();
 364         pgdp = pgd_offset(mm, addr);
 365         do {
 366                 pgd_t pgd = *pgdp;
 367
 368                 next = pgd_addr_end(addr, end);
 369                 if (pgd_none(pgd))
 370                         goto slow;
 371                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 372                         goto slow;
 373         } while (pgdp++, addr = next, addr != end);
 374         local_irq_enable();
 375
 376         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 377         return nr;
 378
 379         {
 380                 int ret;
 381
 382 slow:
 383                 local_irq_enable();
 384 slow_irqon:
 385                 /* Try to get the remaining pages with get_user_pages */
 386                 start += nr << PAGE_SHIFT;
 387                 pages += nr;
 388
 389                 ret = get_user_pages_unlocked(current, mm, start,
 390                                               (end - start) >> PAGE_SHIFT,
 391                                               write, 0, pages);
 392
 393                 /* Have to be a bit careful with return values */
 394                 if (nr > 0) {
 395                         if (ret < 0)
 396                                 ret = nr;
 397                         else
 398                                 ret += nr;
 399                 }
 400
 401                 return ret;
 402         }
 403 }