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 #include <linux/memremap.h>
  13
  14 #include <asm/pgtable.h>
  15
  16 static inline pte_t gup_get_pte(pte_t *ptep)
  17 {
  18 #ifndef CONFIG_X86_PAE
  19         return READ_ONCE(*ptep);
  20 #else
  21         /*
  22          * With get_user_pages_fast, we walk down the pagetables without taking
  23          * any locks.  For this we would like to load the pointers atomically,
  24          * but that is not possible (without expensive cmpxchg8b) on PAE.  What
  25          * we do have is the guarantee that a pte will only either go from not
  26          * present to present, or present to not present or both -- it will not
  27          * switch to a completely different present page without a TLB flush in
  28          * between; something that we are blocking by holding interrupts off.
  29          *
  30          * Setting ptes from not present to present goes:
  31          * ptep->pte_high = h;
  32          * smp_wmb();
  33          * ptep->pte_low = l;
  34          *
  35          * And present to not present goes:
  36          * ptep->pte_low = 0;
  37          * smp_wmb();
  38          * ptep->pte_high = 0;
  39          *
  40          * We must ensure here that the load of pte_low sees l iff pte_high
  41          * sees h. We load pte_high *after* loading pte_low, which ensures we
  42          * don't see an older value of pte_high.  *Then* we recheck pte_low,
  43          * which ensures that we haven't picked up a changed pte high. We might
  44          * have got rubbish values from pte_low and pte_high, but we are
  45          * guaranteed that pte_low will not have the present bit set *unless*
  46          * it is 'l'. And get_user_pages_fast only operates on present ptes, so
  47          * we're safe.
  48          *
  49          * gup_get_pte should not be used or copied outside gup.c without being
  50          * very careful -- it does not atomically load the pte or anything that
  51          * is likely to be useful for you.
  52          */
  53         pte_t pte;
  54
  55 retry:
  56         pte.pte_low = ptep->pte_low;
  57         smp_rmb();
  58         pte.pte_high = ptep->pte_high;
  59         smp_rmb();
  60         if (unlikely(pte.pte_low != ptep->pte_low))
  61                 goto retry;
  62
  63         return pte;
  64 #endif
  65 }
  66
  67 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
  68 {
  69         while ((*nr) - nr_start) {
  70                 struct page *page = pages[--(*nr)];
  71
  72                 ClearPageReferenced(page);
  73                 put_page(page);
  74         }
  75 }
  76
  77 /*
  78  * The performance critical leaf functions are made noinline otherwise gcc
  79  * inlines everything into a single function which results in too much
  80  * register pressure.
  81  */
  82 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
  83                 unsigned long end, int write, struct page **pages, int *nr)
  84 {
  85         struct dev_pagemap *pgmap = NULL;
  86         unsigned long mask;
  87         int nr_start = *nr;
  88         pte_t *ptep;
  89
  90         mask = _PAGE_PRESENT|_PAGE_USER;
  91         if (write)
  92                 mask |= _PAGE_RW;
  93
  94         ptep = pte_offset_map(&pmd, addr);
  95         do {
  96                 pte_t pte = gup_get_pte(ptep);
  97                 struct page *page;
  98
  99                 /* Similar to the PMD case, NUMA hinting must take slow path */
 100                 if (pte_protnone(pte)) {
 101                         pte_unmap(ptep);
 102                         return 0;
 103                 }
 104
 105                 if (pte_devmap(pte)) {
 106                         pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
 107                         if (unlikely(!pgmap)) {
 108                                 undo_dev_pagemap(nr, nr_start, pages);
 109                                 pte_unmap(ptep);
 110                                 return 0;
 111                         }
 112                 } else if ((pte_flags(pte) & (mask | _PAGE_SPECIAL)) != mask) {
 113                         pte_unmap(ptep);
 114                         return 0;
 115                 }
 116                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 117                 page = pte_page(pte);
 118                 get_page(page);
 119                 put_dev_pagemap(pgmap);
 120                 SetPageReferenced(page);
 121                 pages[*nr] = page;
 122                 (*nr)++;
 123
 124         } while (ptep++, addr += PAGE_SIZE, addr != end);
 125         pte_unmap(ptep - 1);
 126
 127         return 1;
 128 }
 129
 130 static inline void get_head_page_multiple(struct page *page, int nr)
 131 {
 132         VM_BUG_ON_PAGE(page != compound_head(page), page);
 133         VM_BUG_ON_PAGE(page_count(page) == 0, page);
 134         atomic_add(nr, &page->_count);
 135         SetPageReferenced(page);
 136 }
 137
 138 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
 139                 unsigned long end, struct page **pages, int *nr)
 140 {
 141         int nr_start = *nr;
 142         unsigned long pfn = pmd_pfn(pmd);
 143         struct dev_pagemap *pgmap = NULL;
 144
 145         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 146         do {
 147                 struct page *page = pfn_to_page(pfn);
 148
 149                 pgmap = get_dev_pagemap(pfn, pgmap);
 150                 if (unlikely(!pgmap)) {
 151                         undo_dev_pagemap(nr, nr_start, pages);
 152                         return 0;
 153                 }
 154                 SetPageReferenced(page);
 155                 pages[*nr] = page;
 156                 get_page(page);
 157                 put_dev_pagemap(pgmap);
 158                 (*nr)++;
 159                 pfn++;
 160         } while (addr += PAGE_SIZE, addr != end);
 161         return 1;
 162 }
 163
 164 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
 165                 unsigned long end, int write, struct page **pages, int *nr)
 166 {
 167         unsigned long mask;
 168         struct page *head, *page;
 169         int refs;
 170
 171         mask = _PAGE_PRESENT|_PAGE_USER;
 172         if (write)
 173                 mask |= _PAGE_RW;
 174         if ((pmd_flags(pmd) & mask) != mask)
 175                 return 0;
 176
 177         VM_BUG_ON(!pfn_valid(pmd_pfn(pmd)));
 178         if (pmd_devmap(pmd))
 179                 return __gup_device_huge_pmd(pmd, addr, end, pages, nr);
 180
 181         /* hugepages are never "special" */
 182         VM_BUG_ON(pmd_flags(pmd) & _PAGE_SPECIAL);
 183
 184         refs = 0;
 185         head = pmd_page(pmd);
 186         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 187         do {
 188                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
 189                 pages[*nr] = page;
 190                 (*nr)++;
 191                 page++;
 192                 refs++;
 193         } while (addr += PAGE_SIZE, addr != end);
 194         get_head_page_multiple(head, refs);
 195
 196         return 1;
 197 }
 198
 199 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 200                 int write, struct page **pages, int *nr)
 201 {
 202         unsigned long next;
 203         pmd_t *pmdp;
 204
 205         pmdp = pmd_offset(&pud, addr);
 206         do {
 207                 pmd_t pmd = *pmdp;
 208
 209                 next = pmd_addr_end(addr, end);
 210                 if (pmd_none(pmd))
 211                         return 0;
 212                 if (unlikely(pmd_large(pmd) || !pmd_present(pmd))) {
 213                         /*
 214                          * NUMA hinting faults need to be handled in the GUP
 215                          * slowpath for accounting purposes and so that they
 216                          * can be serialised against THP migration.
 217                          */
 218                         if (pmd_protnone(pmd))
 219                                 return 0;
 220                         if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
 221                                 return 0;
 222                 } else {
 223                         if (!gup_pte_range(pmd, addr, next, write, pages, nr))
 224                                 return 0;
 225                 }
 226         } while (pmdp++, addr = next, addr != end);
 227
 228         return 1;
 229 }
 230
 231 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
 232                 unsigned long end, int write, struct page **pages, int *nr)
 233 {
 234         unsigned long mask;
 235         struct page *head, *page;
 236         int refs;
 237
 238         mask = _PAGE_PRESENT|_PAGE_USER;
 239         if (write)
 240                 mask |= _PAGE_RW;
 241         if ((pud_flags(pud) & mask) != mask)
 242                 return 0;
 243         /* hugepages are never "special" */
 244         VM_BUG_ON(pud_flags(pud) & _PAGE_SPECIAL);
 245         VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
 246
 247         refs = 0;
 248         head = pud_page(pud);
 249         page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 250         do {
 251                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
 252                 pages[*nr] = page;
 253                 (*nr)++;
 254                 page++;
 255                 refs++;
 256         } while (addr += PAGE_SIZE, addr != end);
 257         get_head_page_multiple(head, refs);
 258
 259         return 1;
 260 }
 261
 262 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 263                         int write, struct page **pages, int *nr)
 264 {
 265         unsigned long next;
 266         pud_t *pudp;
 267
 268         pudp = pud_offset(&pgd, addr);
 269         do {
 270                 pud_t pud = *pudp;
 271
 272                 next = pud_addr_end(addr, end);
 273                 if (pud_none(pud))
 274                         return 0;
 275                 if (unlikely(pud_large(pud))) {
 276                         if (!gup_huge_pud(pud, addr, next, write, pages, nr))
 277                                 return 0;
 278                 } else {
 279                         if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 280                                 return 0;
 281                 }
 282         } while (pudp++, addr = next, addr != end);
 283
 284         return 1;
 285 }
 286
 287 /*
 288  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 289  * back to the regular GUP.
 290  */
 291 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 292                           struct page **pages)
 293 {
 294         struct mm_struct *mm = current->mm;
 295         unsigned long addr, len, end;
 296         unsigned long next;
 297         unsigned long flags;
 298         pgd_t *pgdp;
 299         int nr = 0;
 300
 301         start &= PAGE_MASK;
 302         addr = start;
 303         len = (unsigned long) nr_pages << PAGE_SHIFT;
 304         end = start + len;
 305         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
 306                                         (void __user *)start, len)))
 307                 return 0;
 308
 309         /*
 310          * XXX: batch / limit 'nr', to avoid large irq off latency
 311          * needs some instrumenting to determine the common sizes used by
 312          * important workloads (eg. DB2), and whether limiting the batch size
 313          * will decrease performance.
 314          *
 315          * It seems like we're in the clear for the moment. Direct-IO is
 316          * the main guy that batches up lots of get_user_pages, and even
 317          * they are limited to 64-at-a-time which is not so many.
 318          */
 319         /*
 320          * This doesn't prevent pagetable teardown, but does prevent
 321          * the pagetables and pages from being freed on x86.
 322          *
 323          * So long as we atomically load page table pointers versus teardown
 324          * (which we do on x86, with the above PAE exception), we can follow the
 325          * address down to the the page and take a ref on it.
 326          */
 327         local_irq_save(flags);
 328         pgdp = pgd_offset(mm, addr);
 329         do {
 330                 pgd_t pgd = *pgdp;
 331
 332                 next = pgd_addr_end(addr, end);
 333                 if (pgd_none(pgd))
 334                         break;
 335                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 336                         break;
 337         } while (pgdp++, addr = next, addr != end);
 338         local_irq_restore(flags);
 339
 340         return nr;
 341 }
 342
 343 /**
 344  * get_user_pages_fast() - pin user pages in memory
 345  * @start:      starting user address
 346  * @nr_pages:   number of pages from start to pin
 347  * @write:      whether pages will be written to
 348  * @pages:      array that receives pointers to the pages pinned.
 349  *              Should be at least nr_pages long.
 350  *
 351  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 352  * If not successful, it will fall back to taking the lock and
 353  * calling get_user_pages().
 354  *
 355  * Returns number of pages pinned. This may be fewer than the number
 356  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 357  * were pinned, returns -errno.
 358  */
 359 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 360                         struct page **pages)
 361 {
 362         struct mm_struct *mm = current->mm;
 363         unsigned long addr, len, end;
 364         unsigned long next;
 365         pgd_t *pgdp;
 366         int nr = 0;
 367
 368         start &= PAGE_MASK;
 369         addr = start;
 370         len = (unsigned long) nr_pages << PAGE_SHIFT;
 371
 372         end = start + len;
 373         if (end < start)
 374                 goto slow_irqon;
 375
 376 #ifdef CONFIG_X86_64
 377         if (end >> __VIRTUAL_MASK_SHIFT)
 378                 goto slow_irqon;
 379 #endif
 380
 381         /*
 382          * XXX: batch / limit 'nr', to avoid large irq off latency
 383          * needs some instrumenting to determine the common sizes used by
 384          * important workloads (eg. DB2), and whether limiting the batch size
 385          * will decrease performance.
 386          *
 387          * It seems like we're in the clear for the moment. Direct-IO is
 388          * the main guy that batches up lots of get_user_pages, and even
 389          * they are limited to 64-at-a-time which is not so many.
 390          */
 391         /*
 392          * This doesn't prevent pagetable teardown, but does prevent
 393          * the pagetables and pages from being freed on x86.
 394          *
 395          * So long as we atomically load page table pointers versus teardown
 396          * (which we do on x86, with the above PAE exception), we can follow the
 397          * address down to the the page and take a ref on it.
 398          */
 399         local_irq_disable();
 400         pgdp = pgd_offset(mm, addr);
 401         do {
 402                 pgd_t pgd = *pgdp;
 403
 404                 next = pgd_addr_end(addr, end);
 405                 if (pgd_none(pgd))
 406                         goto slow;
 407                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 408                         goto slow;
 409         } while (pgdp++, addr = next, addr != end);
 410         local_irq_enable();
 411
 412         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 413         return nr;
 414
 415         {
 416                 int ret;
 417
 418 slow:
 419                 local_irq_enable();
 420 slow_irqon:
 421                 /* Try to get the remaining pages with get_user_pages */
 422                 start += nr << PAGE_SHIFT;
 423                 pages += nr;
 424
 425                 ret = get_user_pages_unlocked(current, mm, start,
 426                                               (end - start) >> PAGE_SHIFT,
 427                                               write, 0, pages);
 428
 429                 /* Have to be a bit careful with return values */
 430                 if (nr > 0) {
 431                         if (ret < 0)
 432                                 ret = nr;
 433                         else
 434                                 ret += nr;
 435                 }
 436
 437                 return ret;
 438         }
 439 }