2 * Lockless get_user_pages_fast for s390
4 * Copyright IBM Corp. 2010
5 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
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)
27 mask = (write ? _PAGE_RO : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
29 ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
33 if ((pte_val(pte) & mask) != 0)
35 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
37 if (!page_cache_get_speculative(page))
39 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
46 } while (ptep++, addr += PAGE_SIZE, addr != end);
51 static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
52 unsigned long end, int write, struct page **pages, int *nr)
54 unsigned long mask, result;
55 struct page *head, *page, *tail;
58 result = write ? 0 : _SEGMENT_ENTRY_RO;
59 mask = result | _SEGMENT_ENTRY_INV;
60 if ((pmd_val(pmd) & mask) != result)
62 VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
66 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
69 VM_BUG_ON(compound_head(page) != head);
74 } while (addr += PAGE_SIZE, addr != end);
76 if (!page_cache_add_speculative(head, refs)) {
81 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
89 * Any tail page need their mapcount reference taken before we
94 get_huge_page_tail(tail);
102 static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
103 unsigned long end, int write, struct page **pages, int *nr)
108 pmdp = (pmd_t *) pudp;
110 if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
111 pmdp = (pmd_t *) pud_deref(pud);
112 pmdp += pmd_index(addr);
117 next = pmd_addr_end(addr, end);
120 if (unlikely(pmd_huge(pmd))) {
121 if (!gup_huge_pmd(pmdp, pmd, addr, next,
124 } else if (!gup_pte_range(pmdp, pmd, addr, next,
127 } while (pmdp++, addr = next, addr != end);
132 static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
133 unsigned long end, int write, struct page **pages, int *nr)
138 pudp = (pud_t *) pgdp;
140 if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
141 pudp = (pud_t *) pgd_deref(pgd);
142 pudp += pud_index(addr);
147 next = pud_addr_end(addr, end);
150 if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr))
152 } while (pudp++, addr = next, addr != end);
158 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
159 * back to the regular GUP.
161 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
164 struct mm_struct *mm = current->mm;
165 unsigned long addr, len, end;
166 unsigned long next, flags;
172 len = (unsigned long) nr_pages << PAGE_SHIFT;
174 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
175 (void __user *)start, len)))
178 local_irq_save(flags);
179 pgdp = pgd_offset(mm, addr);
183 next = pgd_addr_end(addr, end);
186 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
188 } while (pgdp++, addr = next, addr != end);
189 local_irq_restore(flags);
195 * get_user_pages_fast() - pin user pages in memory
196 * @start: starting user address
197 * @nr_pages: number of pages from start to pin
198 * @write: whether pages will be written to
199 * @pages: array that receives pointers to the pages pinned.
200 * Should be at least nr_pages long.
202 * Attempt to pin user pages in memory without taking mm->mmap_sem.
203 * If not successful, it will fall back to taking the lock and
204 * calling get_user_pages().
206 * Returns number of pages pinned. This may be fewer than the number
207 * requested. If nr_pages is 0 or negative, returns 0. If no pages
208 * were pinned, returns -errno.
210 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
213 struct mm_struct *mm = current->mm;
214 unsigned long addr, len, end;
221 len = (unsigned long) nr_pages << PAGE_SHIFT;
227 * local_irq_disable() doesn't prevent pagetable teardown, but does
228 * prevent the pagetables from being freed on s390.
230 * So long as we atomically load page table pointers versus teardown,
231 * we can follow the address down to the the page and take a ref on it.
234 pgdp = pgd_offset(mm, addr);
238 next = pgd_addr_end(addr, end);
241 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
243 } while (pgdp++, addr = next, addr != end);
246 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
254 /* Try to get the remaining pages with get_user_pages */
255 start += nr << PAGE_SHIFT;
258 down_read(&mm->mmap_sem);
259 ret = get_user_pages(current, mm, start,
260 (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
261 up_read(&mm->mmap_sem);
263 /* Have to be a bit careful with return values */