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_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 if (WARN_ON_ONCE(page_ref_count(page) < 0)
41 || !page_cache_get_speculative(page))
43 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
50 } while (ptep++, addr += PAGE_SIZE, addr != end);
55 static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
56 unsigned long end, int write, struct page **pages, int *nr)
58 unsigned long mask, result;
59 struct page *head, *page, *tail;
62 result = write ? 0 : _SEGMENT_ENTRY_PROTECT;
63 mask = result | _SEGMENT_ENTRY_INVALID;
64 if ((pmd_val(pmd) & mask) != result)
66 VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
70 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
73 VM_BUG_ON(compound_head(page) != head);
78 } while (addr += PAGE_SIZE, addr != end);
80 if (WARN_ON_ONCE(page_ref_count(head) < 0)
81 || !page_cache_add_speculative(head, refs)) {
86 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
94 * Any tail page need their mapcount reference taken before we
99 get_huge_page_tail(tail);
107 static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
108 unsigned long end, int write, struct page **pages, int *nr)
113 pmdp = (pmd_t *) pudp;
114 if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
115 pmdp = (pmd_t *) pud_deref(pud);
116 pmdp += pmd_index(addr);
120 next = pmd_addr_end(addr, end);
122 * The pmd_trans_splitting() check below explains why
123 * pmdp_splitting_flush() has to serialize with
124 * smp_call_function() against our disabled IRQs, to stop
125 * this gup-fast code from running while we set the
126 * splitting bit in the pmd. Returning zero will take
127 * the slow path that will call wait_split_huge_page()
128 * if the pmd is still in splitting state.
130 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
132 if (unlikely(pmd_large(pmd))) {
134 * NUMA hinting faults need to be handled in the GUP
135 * slowpath for accounting purposes and so that they
136 * can be serialised against THP migration.
138 if (pmd_protnone(pmd))
140 if (!gup_huge_pmd(pmdp, pmd, addr, next,
143 } else if (!gup_pte_range(pmdp, pmd, addr, next,
146 } while (pmdp++, addr = next, addr != end);
151 static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
152 unsigned long end, int write, struct page **pages, int *nr)
157 pudp = (pud_t *) pgdp;
158 if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
159 pudp = (pud_t *) pgd_deref(pgd);
160 pudp += pud_index(addr);
164 next = pud_addr_end(addr, end);
167 if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr))
169 } while (pudp++, addr = next, addr != end);
175 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
176 * back to the regular GUP.
178 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
181 struct mm_struct *mm = current->mm;
182 unsigned long addr, len, end;
183 unsigned long next, flags;
189 len = (unsigned long) nr_pages << PAGE_SHIFT;
191 if ((end <= start) || (end > TASK_SIZE))
194 * local_irq_save() doesn't prevent pagetable teardown, but does
195 * prevent the pagetables from being freed on s390.
197 * So long as we atomically load page table pointers versus teardown,
198 * we can follow the address down to the the page and take a ref on it.
200 local_irq_save(flags);
201 pgdp = pgd_offset(mm, addr);
205 next = pgd_addr_end(addr, end);
208 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
210 } while (pgdp++, addr = next, addr != end);
211 local_irq_restore(flags);
217 * get_user_pages_fast() - pin user pages in memory
218 * @start: starting user address
219 * @nr_pages: number of pages from start to pin
220 * @write: whether pages will be written to
221 * @pages: array that receives pointers to the pages pinned.
222 * Should be at least nr_pages long.
224 * Attempt to pin user pages in memory without taking mm->mmap_sem.
225 * If not successful, it will fall back to taking the lock and
226 * calling get_user_pages().
228 * Returns number of pages pinned. This may be fewer than the number
229 * requested. If nr_pages is 0 or negative, returns 0. If no pages
230 * were pinned, returns -errno.
232 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
235 struct mm_struct *mm = current->mm;
239 nr = __get_user_pages_fast(start, nr_pages, write, pages);
243 /* Try to get the remaining pages with get_user_pages */
244 start += nr << PAGE_SHIFT;
246 ret = get_user_pages_unlocked(current, mm, start,
247 nr_pages - nr, pages, write ? FOLL_WRITE : 0);
248 /* Have to be a bit careful with return values */
250 ret = (ret < 0) ? nr : ret + nr;