arch/s390/mm/gup.c

   1 /*
   2  *  Lockless get_user_pages_fast for s390
   3  *
   4  *  Copyright IBM Corp. 2010
   5  *  Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
   6  */
   7 #include <linux/sched.h>
   8 #include <linux/mm.h>
   9 #include <linux/hugetlb.h>
  10 #include <linux/vmstat.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/rwsem.h>
  13 #include <asm/pgtable.h>
  14
  15 /*
  16  * The performance critical leaf functions are made noinline otherwise gcc
  17  * inlines everything into a single function which results in too much
  18  * register pressure.
  19  */
  20 static inline int gup_pte_range(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  21                 unsigned long end, int write, struct page **pages, int *nr)
  22 {
  23         struct page *head, *page;
  24         unsigned long mask;
  25         pte_t *ptep, pte;
  26
  27         mask = (write ? _PAGE_PROTECT : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
  28
  29         ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
  30         do {
  31                 pte = *ptep;
  32                 barrier();
  33                 /* Similar to the PMD case, NUMA hinting must take slow path */
  34                 if (pte_protnone(pte))
  35                         return 0;
  36                 if ((pte_val(pte) & mask) != 0)
  37                         return 0;
  38                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  39                 page = pte_page(pte);
  40                 head = compound_head(page);
  41                 if (!page_cache_get_speculative(head))
  42                         return 0;
  43                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
  44                         put_page(head);
  45                         return 0;
  46                 }
  47                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
  48                 pages[*nr] = page;
  49                 (*nr)++;
  50
  51         } while (ptep++, addr += PAGE_SIZE, addr != end);
  52
  53         return 1;
  54 }
  55
  56 static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  57                 unsigned long end, int write, struct page **pages, int *nr)
  58 {
  59         struct page *head, *page;
  60         unsigned long mask;
  61         int refs;
  62
  63         mask = (write ? _SEGMENT_ENTRY_PROTECT : 0) | _SEGMENT_ENTRY_INVALID;
  64         if ((pmd_val(pmd) & mask) != 0)
  65                 return 0;
  66         VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
  67
  68         refs = 0;
  69         head = pmd_page(pmd);
  70         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
  71         do {
  72                 VM_BUG_ON(compound_head(page) != head);
  73                 pages[*nr] = page;
  74                 (*nr)++;
  75                 page++;
  76                 refs++;
  77         } while (addr += PAGE_SIZE, addr != end);
  78
  79         if (!page_cache_add_speculative(head, refs)) {
  80                 *nr -= refs;
  81                 return 0;
  82         }
  83
  84         if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
  85                 *nr -= refs;
  86                 while (refs--)
  87                         put_page(head);
  88                 return 0;
  89         }
  90
  91         return 1;
  92 }
  93
  94
  95 static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
  96                 unsigned long end, int write, struct page **pages, int *nr)
  97 {
  98         unsigned long next;
  99         pmd_t *pmdp, pmd;
 100
 101         pmdp = (pmd_t *) pudp;
 102         if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
 103                 pmdp = (pmd_t *) pud_deref(pud);
 104         pmdp += pmd_index(addr);
 105         do {
 106                 pmd = *pmdp;
 107                 barrier();
 108                 next = pmd_addr_end(addr, end);
 109                 if (pmd_none(pmd))
 110                         return 0;
 111                 if (unlikely(pmd_large(pmd))) {
 112                         /*
 113                          * NUMA hinting faults need to be handled in the GUP
 114                          * slowpath for accounting purposes and so that they
 115                          * can be serialised against THP migration.
 116                          */
 117                         if (pmd_protnone(pmd))
 118                                 return 0;
 119                         if (!gup_huge_pmd(pmdp, pmd, addr, next,
 120                                           write, pages, nr))
 121                                 return 0;
 122                 } else if (!gup_pte_range(pmdp, pmd, addr, next,
 123                                           write, pages, nr))
 124                         return 0;
 125         } while (pmdp++, addr = next, addr != end);
 126
 127         return 1;
 128 }
 129
 130 static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr,
 131                 unsigned long end, int write, struct page **pages, int *nr)
 132 {
 133         struct page *head, *page;
 134         unsigned long mask;
 135         int refs;
 136
 137         mask = (write ? _REGION_ENTRY_PROTECT : 0) | _REGION_ENTRY_INVALID;
 138         if ((pud_val(pud) & mask) != 0)
 139                 return 0;
 140         VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
 141
 142         refs = 0;
 143         head = pud_page(pud);
 144         page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 145         do {
 146                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
 147                 pages[*nr] = page;
 148                 (*nr)++;
 149                 page++;
 150                 refs++;
 151         } while (addr += PAGE_SIZE, addr != end);
 152
 153         if (!page_cache_add_speculative(head, refs)) {
 154                 *nr -= refs;
 155                 return 0;
 156         }
 157
 158         if (unlikely(pud_val(pud) != pud_val(*pudp))) {
 159                 *nr -= refs;
 160                 while (refs--)
 161                         put_page(head);
 162                 return 0;
 163         }
 164
 165         return 1;
 166 }
 167
 168 static inline int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr,
 169                 unsigned long end, int write, struct page **pages, int *nr)
 170 {
 171         unsigned long next;
 172         pud_t *pudp, pud;
 173
 174         pudp = (pud_t *) p4dp;
 175         if ((p4d_val(p4d) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
 176                 pudp = (pud_t *) p4d_deref(p4d);
 177         pudp += pud_index(addr);
 178         do {
 179                 pud = *pudp;
 180                 barrier();
 181                 next = pud_addr_end(addr, end);
 182                 if (pud_none(pud))
 183                         return 0;
 184                 if (unlikely(pud_large(pud))) {
 185                         if (!gup_huge_pud(pudp, pud, addr, next, write, pages,
 186                                           nr))
 187                                 return 0;
 188                 } else if (!gup_pmd_range(pudp, pud, addr, next, write, pages,
 189                                           nr))
 190                         return 0;
 191         } while (pudp++, addr = next, addr != end);
 192
 193         return 1;
 194 }
 195
 196 static inline int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
 197                 unsigned long end, int write, struct page **pages, int *nr)
 198 {
 199         unsigned long next;
 200         p4d_t *p4dp, p4d;
 201
 202         p4dp = (p4d_t *) pgdp;
 203         if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R1)
 204                 p4dp = (p4d_t *) pgd_deref(pgd);
 205         p4dp += p4d_index(addr);
 206         do {
 207                 p4d = *p4dp;
 208                 barrier();
 209                 next = p4d_addr_end(addr, end);
 210                 if (p4d_none(p4d))
 211                         return 0;
 212                 if (!gup_pud_range(p4dp, p4d, addr, next, write, pages, nr))
 213                         return 0;
 214         } while (p4dp++, addr = next, addr != end);
 215
 216         return 1;
 217 }
 218
 219 /*
 220  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 221  * back to the regular GUP.
 222  */
 223 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 224                           struct page **pages)
 225 {
 226         struct mm_struct *mm = current->mm;
 227         unsigned long addr, len, end;
 228         unsigned long next, flags;
 229         pgd_t *pgdp, pgd;
 230         int nr = 0;
 231
 232         start &= PAGE_MASK;
 233         addr = start;
 234         len = (unsigned long) nr_pages << PAGE_SHIFT;
 235         end = start + len;
 236         if ((end <= start) || (end > mm->context.asce_limit))
 237                 return 0;
 238         /*
 239          * local_irq_save() doesn't prevent pagetable teardown, but does
 240          * prevent the pagetables from being freed on s390.
 241          *
 242          * So long as we atomically load page table pointers versus teardown,
 243          * we can follow the address down to the the page and take a ref on it.
 244          */
 245         local_irq_save(flags);
 246         pgdp = pgd_offset(mm, addr);
 247         do {
 248                 pgd = *pgdp;
 249                 barrier();
 250                 next = pgd_addr_end(addr, end);
 251                 if (pgd_none(pgd))
 252                         break;
 253                 if (!gup_p4d_range(pgdp, pgd, addr, next, write, pages, &nr))
 254                         break;
 255         } while (pgdp++, addr = next, addr != end);
 256         local_irq_restore(flags);
 257
 258         return nr;
 259 }
 260
 261 /**
 262  * get_user_pages_fast() - pin user pages in memory
 263  * @start:      starting user address
 264  * @nr_pages:   number of pages from start to pin
 265  * @write:      whether pages will be written to
 266  * @pages:      array that receives pointers to the pages pinned.
 267  *              Should be at least nr_pages long.
 268  *
 269  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 270  * If not successful, it will fall back to taking the lock and
 271  * calling get_user_pages().
 272  *
 273  * Returns number of pages pinned. This may be fewer than the number
 274  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 275  * were pinned, returns -errno.
 276  */
 277 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 278                         struct page **pages)
 279 {
 280         int nr, ret;
 281
 282         might_sleep();
 283         start &= PAGE_MASK;
 284         nr = __get_user_pages_fast(start, nr_pages, write, pages);
 285         if (nr == nr_pages)
 286                 return nr;
 287
 288         /* Try to get the remaining pages with get_user_pages */
 289         start += nr << PAGE_SHIFT;
 290         pages += nr;
 291         ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
 292                                       write ? FOLL_WRITE : 0);
 293         /* Have to be a bit careful with return values */
 294         if (nr > 0)
 295                 ret = (ret < 0) ? nr : ret + nr;
 296         return ret;
 297 }