2 * Lockless get_user_pages_fast for s390
4 * Copyright IBM Corp. 2010
7 #include <linux/sched.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>
16 * The performance critical leaf functions are made noinline otherwise gcc
17 * inlines everything into a single function which results in too much
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)
23 struct page *head, *page;
27 mask = (write ? _PAGE_PROTECT : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
29 ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
33 /* Similar to the PMD case, NUMA hinting must take slow path */
34 if (pte_protnone(pte))
36 if ((pte_val(pte) & mask) != 0)
38 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
40 head = compound_head(page);
41 if (!page_cache_get_speculative(head))
43 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
47 VM_BUG_ON_PAGE(compound_head(page) != head, page);
51 } while (ptep++, addr += PAGE_SIZE, addr != end);
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)
59 unsigned long mask, result;
60 struct page *head, *page;
63 result = write ? 0 : _SEGMENT_ENTRY_PROTECT;
64 mask = result | _SEGMENT_ENTRY_INVALID;
65 if ((pmd_val(pmd) & mask) != result)
67 VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
71 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
73 VM_BUG_ON(compound_head(page) != head);
78 } while (addr += PAGE_SIZE, addr != end);
80 if (!page_cache_add_speculative(head, refs)) {
85 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
96 static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
97 unsigned long end, int write, struct page **pages, int *nr)
102 pmdp = (pmd_t *) pudp;
103 if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
104 pmdp = (pmd_t *) pud_deref(pud);
105 pmdp += pmd_index(addr);
109 next = pmd_addr_end(addr, end);
112 if (unlikely(pmd_large(pmd))) {
114 * NUMA hinting faults need to be handled in the GUP
115 * slowpath for accounting purposes and so that they
116 * can be serialised against THP migration.
118 if (pmd_protnone(pmd))
120 if (!gup_huge_pmd(pmdp, pmd, addr, next,
123 } else if (!gup_pte_range(pmdp, pmd, addr, next,
126 } while (pmdp++, addr = next, addr != end);
131 static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr,
132 unsigned long end, int write, struct page **pages, int *nr)
134 struct page *head, *page;
138 mask = (write ? _REGION_ENTRY_PROTECT : 0) | _REGION_ENTRY_INVALID;
139 if ((pud_val(pud) & mask) != 0)
141 VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
144 head = pud_page(pud);
145 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
147 VM_BUG_ON_PAGE(compound_head(page) != head, page);
152 } while (addr += PAGE_SIZE, addr != end);
154 if (!page_cache_add_speculative(head, refs)) {
159 if (unlikely(pud_val(pud) != pud_val(*pudp))) {
169 static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
170 unsigned long end, int write, struct page **pages, int *nr)
175 pudp = (pud_t *) pgdp;
176 if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
177 pudp = (pud_t *) pgd_deref(pgd);
178 pudp += pud_index(addr);
182 next = pud_addr_end(addr, end);
185 if (unlikely(pud_large(pud))) {
186 if (!gup_huge_pud(pudp, pud, addr, next, write, pages,
189 } else if (!gup_pmd_range(pudp, pud, addr, next, write, pages,
192 } while (pudp++, addr = next, addr != end);
198 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
199 * back to the regular GUP.
201 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
204 struct mm_struct *mm = current->mm;
205 unsigned long addr, len, end;
206 unsigned long next, flags;
212 len = (unsigned long) nr_pages << PAGE_SHIFT;
214 if ((end <= start) || (end > TASK_SIZE))
217 * local_irq_save() doesn't prevent pagetable teardown, but does
218 * prevent the pagetables from being freed on s390.
220 * So long as we atomically load page table pointers versus teardown,
221 * we can follow the address down to the the page and take a ref on it.
223 local_irq_save(flags);
224 pgdp = pgd_offset(mm, addr);
228 next = pgd_addr_end(addr, end);
231 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
233 } while (pgdp++, addr = next, addr != end);
234 local_irq_restore(flags);
240 * get_user_pages_fast() - pin user pages in memory
241 * @start: starting user address
242 * @nr_pages: number of pages from start to pin
243 * @write: whether pages will be written to
244 * @pages: array that receives pointers to the pages pinned.
245 * Should be at least nr_pages long.
247 * Attempt to pin user pages in memory without taking mm->mmap_sem.
248 * If not successful, it will fall back to taking the lock and
249 * calling get_user_pages().
251 * Returns number of pages pinned. This may be fewer than the number
252 * requested. If nr_pages is 0 or negative, returns 0. If no pages
253 * were pinned, returns -errno.
255 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
262 nr = __get_user_pages_fast(start, nr_pages, write, pages);
266 /* Try to get the remaining pages with get_user_pages */
267 start += nr << PAGE_SHIFT;
269 ret = get_user_pages_unlocked(start, nr_pages - nr, write, 0, pages);
270 /* Have to be a bit careful with return values */
272 ret = (ret < 0) ? nr : ret + nr;