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