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)
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);
119 * The pmd_trans_splitting() check below explains why
120 * pmdp_splitting_flush() has to serialize with
121 * smp_call_function() against our disabled IRQs, to stop
122 * this gup-fast code from running while we set the
123 * splitting bit in the pmd. Returning zero will take
124 * the slow path that will call wait_split_huge_page()
125 * if the pmd is still in splitting state.
127 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
129 if (unlikely(pmd_large(pmd))) {
130 if (!gup_huge_pmd(pmdp, pmd, addr, next,
133 } else if (!gup_pte_range(pmdp, pmd, addr, next,
136 } while (pmdp++, addr = next, addr != end);
141 static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
142 unsigned long end, int write, struct page **pages, int *nr)
147 pudp = (pud_t *) pgdp;
149 if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
150 pudp = (pud_t *) pgd_deref(pgd);
151 pudp += pud_index(addr);
156 next = pud_addr_end(addr, end);
159 if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr))
161 } while (pudp++, addr = next, addr != end);
167 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
168 * back to the regular GUP.
170 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
173 struct mm_struct *mm = current->mm;
174 unsigned long addr, len, end;
175 unsigned long next, flags;
181 len = (unsigned long) nr_pages << PAGE_SHIFT;
183 if ((end < start) || (end > TASK_SIZE))
186 local_irq_save(flags);
187 pgdp = pgd_offset(mm, addr);
191 next = pgd_addr_end(addr, end);
194 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
196 } while (pgdp++, addr = next, addr != end);
197 local_irq_restore(flags);
203 * get_user_pages_fast() - pin user pages in memory
204 * @start: starting user address
205 * @nr_pages: number of pages from start to pin
206 * @write: whether pages will be written to
207 * @pages: array that receives pointers to the pages pinned.
208 * Should be at least nr_pages long.
210 * Attempt to pin user pages in memory without taking mm->mmap_sem.
211 * If not successful, it will fall back to taking the lock and
212 * calling get_user_pages().
214 * Returns number of pages pinned. This may be fewer than the number
215 * requested. If nr_pages is 0 or negative, returns 0. If no pages
216 * were pinned, returns -errno.
218 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
221 struct mm_struct *mm = current->mm;
222 unsigned long addr, len, end;
229 len = (unsigned long) nr_pages << PAGE_SHIFT;
231 if ((end < start) || (end > TASK_SIZE))
235 * local_irq_disable() doesn't prevent pagetable teardown, but does
236 * prevent the pagetables from being freed on s390.
238 * So long as we atomically load page table pointers versus teardown,
239 * we can follow the address down to the the page and take a ref on it.
242 pgdp = pgd_offset(mm, addr);
246 next = pgd_addr_end(addr, end);
249 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
251 } while (pgdp++, addr = next, addr != end);
254 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
262 /* Try to get the remaining pages with get_user_pages */
263 start += nr << PAGE_SHIFT;
266 down_read(&mm->mmap_sem);
267 ret = get_user_pages(current, mm, start,
268 (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
269 up_read(&mm->mmap_sem);
271 /* Have to be a bit careful with return values */