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/mmu_context.h>
  15 #include <asm/pgtable.h>
  16
  17 static inline pte_t gup_get_pte(pte_t *ptep)
  18 {
  19 #ifndef CONFIG_X86_PAE
  20         return READ_ONCE(*ptep);
  21 #else
  22         /*
  23          * With get_user_pages_fast, we walk down the pagetables without taking
  24          * any locks.  For this we would like to load the pointers atomically,
  25          * but that is not possible (without expensive cmpxchg8b) on PAE.  What
  26          * we do have is the guarantee that a pte will only either go from not
  27          * present to present, or present to not present or both -- it will not
  28          * switch to a completely different present page without a TLB flush in
  29          * between; something that we are blocking by holding interrupts off.
  30          *
  31          * Setting ptes from not present to present goes:
  32          * ptep->pte_high = h;
  33          * smp_wmb();
  34          * ptep->pte_low = l;
  35          *
  36          * And present to not present goes:
  37          * ptep->pte_low = 0;
  38          * smp_wmb();
  39          * ptep->pte_high = 0;
  40          *
  41          * We must ensure here that the load of pte_low sees l iff pte_high
  42          * sees h. We load pte_high *after* loading pte_low, which ensures we
  43          * don't see an older value of pte_high.  *Then* we recheck pte_low,
  44          * which ensures that we haven't picked up a changed pte high. We might
  45          * have got rubbish values from pte_low and pte_high, but we are
  46          * guaranteed that pte_low will not have the present bit set *unless*
  47          * it is 'l'. And get_user_pages_fast only operates on present ptes, so
  48          * we're safe.
  49          *
  50          * gup_get_pte should not be used or copied outside gup.c without being
  51          * very careful -- it does not atomically load the pte or anything that
  52          * is likely to be useful for you.
  53          */
  54         pte_t pte;
  55
  56 retry:
  57         pte.pte_low = ptep->pte_low;
  58         smp_rmb();
  59         pte.pte_high = ptep->pte_high;
  60         smp_rmb();
  61         if (unlikely(pte.pte_low != ptep->pte_low))
  62                 goto retry;
  63
  64         return pte;
  65 #endif
  66 }
  67
  68 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
  69 {
  70         while ((*nr) - nr_start) {
  71                 struct page *page = pages[--(*nr)];
  72
  73                 ClearPageReferenced(page);
  74                 put_page(page);
  75         }
  76 }
  77
  78 /*
  79  * 'pteval' can come from a pte, pmd or pud.  We only check
  80  * _PAGE_PRESENT, _PAGE_USER, and _PAGE_RW in here which are the
  81  * same value on all 3 types.
  82  */
  83 static inline int pte_allows_gup(unsigned long pteval, int write)
  84 {
  85         unsigned long need_pte_bits = _PAGE_PRESENT|_PAGE_USER;
  86
  87         if (write)
  88                 need_pte_bits |= _PAGE_RW;
  89
  90         if ((pteval & need_pte_bits) != need_pte_bits)
  91                 return 0;
  92
  93         /* Check memory protection keys permissions. */
  94         if (!__pkru_allows_pkey(pte_flags_pkey(pteval), write))
  95                 return 0;
  96
  97         return 1;
  98 }
  99
 100 /*
 101  * The performance critical leaf functions are made noinline otherwise gcc
 102  * inlines everything into a single function which results in too much
 103  * register pressure.
 104  */
 105 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
 106                 unsigned long end, int write, struct page **pages, int *nr)
 107 {
 108         struct dev_pagemap *pgmap = NULL;
 109         int nr_start = *nr, ret = 0;
 110         pte_t *ptep, *ptem;
 111
 112         /*
 113          * Keep the original mapped PTE value (ptem) around since we
 114          * might increment ptep off the end of the page when finishing
 115          * our loop iteration.
 116          */
 117         ptem = ptep = pte_offset_map(&pmd, addr);
 118         do {
 119                 pte_t pte = gup_get_pte(ptep);
 120                 struct page *page;
 121
 122                 /* Similar to the PMD case, NUMA hinting must take slow path */
 123                 if (pte_protnone(pte))
 124                         break;
 125
 126                 if (!pte_allows_gup(pte_val(pte), write))
 127                         break;
 128
 129                 if (pte_devmap(pte)) {
 130                         pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
 131                         if (unlikely(!pgmap)) {
 132                                 undo_dev_pagemap(nr, nr_start, pages);
 133                                 break;
 134                         }
 135                 } else if (pte_special(pte))
 136                         break;
 137
 138                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 139                 page = pte_page(pte);
 140                 get_page(page);
 141                 put_dev_pagemap(pgmap);
 142                 SetPageReferenced(page);
 143                 pages[*nr] = page;
 144                 (*nr)++;
 145
 146         } while (ptep++, addr += PAGE_SIZE, addr != end);
 147         if (addr == end)
 148                 ret = 1;
 149         pte_unmap(ptem);
 150
 151         return ret;
 152 }
 153
 154 static inline void get_head_page_multiple(struct page *page, int nr)
 155 {
 156         VM_BUG_ON_PAGE(page != compound_head(page), page);
 157         VM_BUG_ON_PAGE(page_count(page) == 0, page);
 158         page_ref_add(page, nr);
 159         SetPageReferenced(page);
 160 }
 161
 162 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
 163                 unsigned long end, struct page **pages, int *nr)
 164 {
 165         int nr_start = *nr;
 166         struct dev_pagemap *pgmap = NULL;
 167
 168         do {
 169                 struct page *page = pfn_to_page(pfn);
 170
 171                 pgmap = get_dev_pagemap(pfn, pgmap);
 172                 if (unlikely(!pgmap)) {
 173                         undo_dev_pagemap(nr, nr_start, pages);
 174                         return 0;
 175                 }
 176                 SetPageReferenced(page);
 177                 pages[*nr] = page;
 178                 get_page(page);
 179                 put_dev_pagemap(pgmap);
 180                 (*nr)++;
 181                 pfn++;
 182         } while (addr += PAGE_SIZE, addr != end);
 183         return 1;
 184 }
 185
 186 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
 187                 unsigned long end, struct page **pages, int *nr)
 188 {
 189         unsigned long fault_pfn;
 190
 191         fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 192         return __gup_device_huge(fault_pfn, addr, end, pages, nr);
 193 }
 194
 195 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
 196                 unsigned long end, struct page **pages, int *nr)
 197 {
 198         unsigned long fault_pfn;
 199
 200         fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 201         return __gup_device_huge(fault_pfn, addr, end, pages, nr);
 202 }
 203
 204 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
 205                 unsigned long end, int write, struct page **pages, int *nr)
 206 {
 207         struct page *head, *page;
 208         int refs;
 209
 210         if (!pte_allows_gup(pmd_val(pmd), write))
 211                 return 0;
 212
 213         VM_BUG_ON(!pfn_valid(pmd_pfn(pmd)));
 214         if (pmd_devmap(pmd))
 215                 return __gup_device_huge_pmd(pmd, addr, end, pages, nr);
 216
 217         /* hugepages are never "special" */
 218         VM_BUG_ON(pmd_flags(pmd) & _PAGE_SPECIAL);
 219
 220         refs = 0;
 221         head = pmd_page(pmd);
 222         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 223         do {
 224                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
 225                 pages[*nr] = page;
 226                 (*nr)++;
 227                 page++;
 228                 refs++;
 229         } while (addr += PAGE_SIZE, addr != end);
 230         get_head_page_multiple(head, refs);
 231
 232         return 1;
 233 }
 234
 235 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 236                 int write, struct page **pages, int *nr)
 237 {
 238         unsigned long next;
 239         pmd_t *pmdp;
 240
 241         pmdp = pmd_offset(&pud, addr);
 242         do {
 243                 pmd_t pmd = *pmdp;
 244
 245                 next = pmd_addr_end(addr, end);
 246                 if (pmd_none(pmd))
 247                         return 0;
 248                 if (unlikely(pmd_large(pmd) || !pmd_present(pmd))) {
 249                         /*
 250                          * NUMA hinting faults need to be handled in the GUP
 251                          * slowpath for accounting purposes and so that they
 252                          * can be serialised against THP migration.
 253                          */
 254                         if (pmd_protnone(pmd))
 255                                 return 0;
 256                         if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
 257                                 return 0;
 258                 } else {
 259                         if (!gup_pte_range(pmd, addr, next, write, pages, nr))
 260                                 return 0;
 261                 }
 262         } while (pmdp++, addr = next, addr != end);
 263
 264         return 1;
 265 }
 266
 267 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
 268                 unsigned long end, int write, struct page **pages, int *nr)
 269 {
 270         struct page *head, *page;
 271         int refs;
 272
 273         if (!pte_allows_gup(pud_val(pud), write))
 274                 return 0;
 275
 276         VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
 277         if (pud_devmap(pud))
 278                 return __gup_device_huge_pud(pud, addr, end, pages, nr);
 279
 280         /* hugepages are never "special" */
 281         VM_BUG_ON(pud_flags(pud) & _PAGE_SPECIAL);
 282
 283         refs = 0;
 284         head = pud_page(pud);
 285         page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 286         do {
 287                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
 288                 pages[*nr] = page;
 289                 (*nr)++;
 290                 page++;
 291                 refs++;
 292         } while (addr += PAGE_SIZE, addr != end);
 293         get_head_page_multiple(head, refs);
 294
 295         return 1;
 296 }
 297
 298 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 299                         int write, struct page **pages, int *nr)
 300 {
 301         unsigned long next;
 302         pud_t *pudp;
 303
 304         pudp = pud_offset(&pgd, addr);
 305         do {
 306                 pud_t pud = *pudp;
 307
 308                 next = pud_addr_end(addr, end);
 309                 if (pud_none(pud))
 310                         return 0;
 311                 if (unlikely(pud_large(pud))) {
 312                         if (!gup_huge_pud(pud, addr, next, write, pages, nr))
 313                                 return 0;
 314                 } else {
 315                         if (!gup_pmd_range(pud, addr, next, write, pages, nr))
 316                                 return 0;
 317                 }
 318         } while (pudp++, addr = next, addr != end);
 319
 320         return 1;
 321 }
 322
 323 /*
 324  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 325  * back to the regular GUP.
 326  */
 327 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 328                           struct page **pages)
 329 {
 330         struct mm_struct *mm = current->mm;
 331         unsigned long addr, len, end;
 332         unsigned long next;
 333         unsigned long flags;
 334         pgd_t *pgdp;
 335         int nr = 0;
 336
 337         start &= PAGE_MASK;
 338         addr = start;
 339         len = (unsigned long) nr_pages << PAGE_SHIFT;
 340         end = start + len;
 341         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
 342                                         (void __user *)start, len)))
 343                 return 0;
 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_save(flags);
 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                         break;
 371                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 372                         break;
 373         } while (pgdp++, addr = next, addr != end);
 374         local_irq_restore(flags);
 375
 376         return nr;
 377 }
 378
 379 /**
 380  * get_user_pages_fast() - pin user pages in memory
 381  * @start:      starting user address
 382  * @nr_pages:   number of pages from start to pin
 383  * @write:      whether pages will be written to
 384  * @pages:      array that receives pointers to the pages pinned.
 385  *              Should be at least nr_pages long.
 386  *
 387  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 388  * If not successful, it will fall back to taking the lock and
 389  * calling get_user_pages().
 390  *
 391  * Returns number of pages pinned. This may be fewer than the number
 392  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 393  * were pinned, returns -errno.
 394  */
 395 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 396                         struct page **pages)
 397 {
 398         struct mm_struct *mm = current->mm;
 399         unsigned long addr, len, end;
 400         unsigned long next;
 401         pgd_t *pgdp;
 402         int nr = 0;
 403
 404         start &= PAGE_MASK;
 405         addr = start;
 406         len = (unsigned long) nr_pages << PAGE_SHIFT;
 407
 408         end = start + len;
 409         if (end < start)
 410                 goto slow_irqon;
 411
 412 #ifdef CONFIG_X86_64
 413         if (end >> __VIRTUAL_MASK_SHIFT)
 414                 goto slow_irqon;
 415 #endif
 416
 417         /*
 418          * XXX: batch / limit 'nr', to avoid large irq off latency
 419          * needs some instrumenting to determine the common sizes used by
 420          * important workloads (eg. DB2), and whether limiting the batch size
 421          * will decrease performance.
 422          *
 423          * It seems like we're in the clear for the moment. Direct-IO is
 424          * the main guy that batches up lots of get_user_pages, and even
 425          * they are limited to 64-at-a-time which is not so many.
 426          */
 427         /*
 428          * This doesn't prevent pagetable teardown, but does prevent
 429          * the pagetables and pages from being freed on x86.
 430          *
 431          * So long as we atomically load page table pointers versus teardown
 432          * (which we do on x86, with the above PAE exception), we can follow the
 433          * address down to the the page and take a ref on it.
 434          */
 435         local_irq_disable();
 436         pgdp = pgd_offset(mm, addr);
 437         do {
 438                 pgd_t pgd = *pgdp;
 439
 440                 next = pgd_addr_end(addr, end);
 441                 if (pgd_none(pgd))
 442                         goto slow;
 443                 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
 444                         goto slow;
 445         } while (pgdp++, addr = next, addr != end);
 446         local_irq_enable();
 447
 448         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
 449         return nr;
 450
 451         {
 452                 int ret;
 453
 454 slow:
 455                 local_irq_enable();
 456 slow_irqon:
 457                 /* Try to get the remaining pages with get_user_pages */
 458                 start += nr << PAGE_SHIFT;
 459                 pages += nr;
 460
 461                 ret = get_user_pages_unlocked(start,
 462                                               (end - start) >> PAGE_SHIFT,
 463                                               pages, write ? FOLL_WRITE : 0);
 464
 465                 /* Have to be a bit careful with return values */
 466                 if (nr > 0) {
 467                         if (ret < 0)
 468                                 ret = nr;
 469                         else
 470                                 ret += nr;
 471                 }
 472
 473                 return ret;
 474         }
 475 }