arch/s390/mm/gup.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  Lockless get_user_pages_fast for s390
   4  *
   5  *  Copyright IBM Corp. 2010
   6  *  Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
   7  */
   8 #include <linux/sched.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/vmstat.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/rwsem.h>
  14 #include <asm/pgtable.h>
  15
  16 /*
  17  * The performance critical leaf functions are made noinline otherwise gcc
  18  * inlines everything into a single function which results in too much
  19  * register pressure.
  20  */
  21 static inline int gup_pte_range(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  22                 unsigned long end, int write, struct page **pages, int *nr)
  23 {
  24         struct page *head, *page;
  25         unsigned long mask;
  26         pte_t *ptep, pte;
  27
  28         mask = (write ? _PAGE_PROTECT : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
  29
  30         ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
  31         do {
  32                 pte = *ptep;
  33                 barrier();
  34                 /* Similar to the PMD case, NUMA hinting must take slow path */
  35                 if (pte_protnone(pte))
  36                         return 0;
  37                 if ((pte_val(pte) & mask) != 0)
  38                         return 0;
  39                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
  40                 page = pte_page(pte);
  41                 head = compound_head(page);
  42                 if (unlikely(WARN_ON_ONCE(page_ref_count(head) < 0)
  43                     || !page_cache_get_speculative(head)))
  44                         return 0;
  45                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
  46                         put_page(head);
  47                         return 0;
  48                 }
  49                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
  50                 pages[*nr] = page;
  51                 (*nr)++;
  52
  53         } while (ptep++, addr += PAGE_SIZE, addr != end);
  54
  55         return 1;
  56 }
  57
  58 static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
  59                 unsigned long end, int write, struct page **pages, int *nr)
  60 {
  61         struct page *head, *page;
  62         unsigned long mask;
  63         int refs;
  64
  65         mask = (write ? _SEGMENT_ENTRY_PROTECT : 0) | _SEGMENT_ENTRY_INVALID;
  66         if ((pmd_val(pmd) & mask) != 0)
  67                 return 0;
  68         VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
  69
  70         refs = 0;
  71         head = pmd_page(pmd);
  72         page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
  73         do {
  74                 VM_BUG_ON(compound_head(page) != head);
  75                 pages[*nr] = page;
  76                 (*nr)++;
  77                 page++;
  78                 refs++;
  79         } while (addr += PAGE_SIZE, addr != end);
  80
  81         if (unlikely(WARN_ON_ONCE(page_ref_count(head) < 0)
  82             || !page_cache_add_speculative(head, refs))) {
  83                 *nr -= refs;
  84                 return 0;
  85         }
  86
  87         if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
  88                 *nr -= refs;
  89                 while (refs--)
  90                         put_page(head);
  91                 return 0;
  92         }
  93
  94         return 1;
  95 }
  96
  97
  98 static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
  99                 unsigned long end, int write, struct page **pages, int *nr)
 100 {
 101         unsigned long next;
 102         pmd_t *pmdp, pmd;
 103
 104         pmdp = (pmd_t *) pudp;
 105         if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
 106                 pmdp = (pmd_t *) pud_deref(pud);
 107         pmdp += pmd_index(addr);
 108         do {
 109                 pmd = *pmdp;
 110                 barrier();
 111                 next = pmd_addr_end(addr, end);
 112                 if (pmd_none(pmd))
 113                         return 0;
 114                 if (unlikely(pmd_large(pmd))) {
 115                         /*
 116                          * NUMA hinting faults need to be handled in the GUP
 117                          * slowpath for accounting purposes and so that they
 118                          * can be serialised against THP migration.
 119                          */
 120                         if (pmd_protnone(pmd))
 121                                 return 0;
 122                         if (!gup_huge_pmd(pmdp, pmd, addr, next,
 123                                           write, pages, nr))
 124                                 return 0;
 125                 } else if (!gup_pte_range(pmdp, pmd, addr, next,
 126                                           write, pages, nr))
 127                         return 0;
 128         } while (pmdp++, addr = next, addr != end);
 129
 130         return 1;
 131 }
 132
 133 static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr,
 134                 unsigned long end, int write, struct page **pages, int *nr)
 135 {
 136         struct page *head, *page;
 137         unsigned long mask;
 138         int refs;
 139
 140         mask = (write ? _REGION_ENTRY_PROTECT : 0) | _REGION_ENTRY_INVALID;
 141         if ((pud_val(pud) & mask) != 0)
 142                 return 0;
 143         VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
 144
 145         refs = 0;
 146         head = pud_page(pud);
 147         page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 148         do {
 149                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
 150                 pages[*nr] = page;
 151                 (*nr)++;
 152                 page++;
 153                 refs++;
 154         } while (addr += PAGE_SIZE, addr != end);
 155
 156         if (unlikely(WARN_ON_ONCE(page_ref_count(head) < 0)
 157             || !page_cache_add_speculative(head, refs))) {
 158                 *nr -= refs;
 159                 return 0;
 160         }
 161
 162         if (unlikely(pud_val(pud) != pud_val(*pudp))) {
 163                 *nr -= refs;
 164                 while (refs--)
 165                         put_page(head);
 166                 return 0;
 167         }
 168
 169         return 1;
 170 }
 171
 172 static inline int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr,
 173                 unsigned long end, int write, struct page **pages, int *nr)
 174 {
 175         unsigned long next;
 176         pud_t *pudp, pud;
 177
 178         pudp = (pud_t *) p4dp;
 179         if ((p4d_val(p4d) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
 180                 pudp = (pud_t *) p4d_deref(p4d);
 181         pudp += pud_index(addr);
 182         do {
 183                 pud = *pudp;
 184                 barrier();
 185                 next = pud_addr_end(addr, end);
 186                 if (pud_none(pud))
 187                         return 0;
 188                 if (unlikely(pud_large(pud))) {
 189                         if (!gup_huge_pud(pudp, pud, addr, next, write, pages,
 190                                           nr))
 191                                 return 0;
 192                 } else if (!gup_pmd_range(pudp, pud, addr, next, write, pages,
 193                                           nr))
 194                         return 0;
 195         } while (pudp++, addr = next, addr != end);
 196
 197         return 1;
 198 }
 199
 200 static inline int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
 201                 unsigned long end, int write, struct page **pages, int *nr)
 202 {
 203         unsigned long next;
 204         p4d_t *p4dp, p4d;
 205
 206         p4dp = (p4d_t *) pgdp;
 207         if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R1)
 208                 p4dp = (p4d_t *) pgd_deref(pgd);
 209         p4dp += p4d_index(addr);
 210         do {
 211                 p4d = *p4dp;
 212                 barrier();
 213                 next = p4d_addr_end(addr, end);
 214                 if (p4d_none(p4d))
 215                         return 0;
 216                 if (!gup_pud_range(p4dp, p4d, addr, next, write, pages, nr))
 217                         return 0;
 218         } while (p4dp++, addr = next, addr != end);
 219
 220         return 1;
 221 }
 222
 223 /*
 224  * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
 225  * back to the regular GUP.
 226  */
 227 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
 228                           struct page **pages)
 229 {
 230         struct mm_struct *mm = current->mm;
 231         unsigned long addr, len, end;
 232         unsigned long next, flags;
 233         pgd_t *pgdp, pgd;
 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 ((end <= start) || (end > mm->context.asce_limit))
 241                 return 0;
 242         /*
 243          * local_irq_save() doesn't prevent pagetable teardown, but does
 244          * prevent the pagetables from being freed on s390.
 245          *
 246          * So long as we atomically load page table pointers versus teardown,
 247          * we can follow the address down to the the page and take a ref on it.
 248          */
 249         local_irq_save(flags);
 250         pgdp = pgd_offset(mm, addr);
 251         do {
 252                 pgd = *pgdp;
 253                 barrier();
 254                 next = pgd_addr_end(addr, end);
 255                 if (pgd_none(pgd))
 256                         break;
 257                 if (!gup_p4d_range(pgdp, pgd, addr, next, write, pages, &nr))
 258                         break;
 259         } while (pgdp++, addr = next, addr != end);
 260         local_irq_restore(flags);
 261
 262         return nr;
 263 }
 264
 265 /**
 266  * get_user_pages_fast() - pin user pages in memory
 267  * @start:      starting user address
 268  * @nr_pages:   number of pages from start to pin
 269  * @write:      whether pages will be written to
 270  * @pages:      array that receives pointers to the pages pinned.
 271  *              Should be at least nr_pages long.
 272  *
 273  * Attempt to pin user pages in memory without taking mm->mmap_sem.
 274  * If not successful, it will fall back to taking the lock and
 275  * calling get_user_pages().
 276  *
 277  * Returns number of pages pinned. This may be fewer than the number
 278  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 279  * were pinned, returns -errno.
 280  */
 281 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
 282                         struct page **pages)
 283 {
 284         int nr, ret;
 285
 286         might_sleep();
 287         start &= PAGE_MASK;
 288         nr = __get_user_pages_fast(start, nr_pages, write, pages);
 289         if (nr == nr_pages)
 290                 return nr;
 291
 292         /* Try to get the remaining pages with get_user_pages */
 293         start += nr << PAGE_SHIFT;
 294         pages += nr;
 295         ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
 296                                       write ? FOLL_WRITE : 0);
 297         /* Have to be a bit careful with return values */
 298         if (nr > 0)
 299                 ret = (ret < 0) ? nr : ret + nr;
 300         return ret;
 301 }