1 // SPDX-License-Identifier: GPL-2.0
3 * Lockless get_user_pages_fast for MIPS
5 * Copyright (C) 2008 Nick Piggin
6 * Copyright (C) 2008 Novell Inc.
7 * Copyright (C) 2011 Ralf Baechle
9 #include <linux/sched.h>
11 #include <linux/vmstat.h>
12 #include <linux/highmem.h>
13 #include <linux/swap.h>
14 #include <linux/hugetlb.h>
16 #include <asm/cpu-features.h>
17 #include <asm/pgtable.h>
19 static inline pte_t gup_get_pte(pte_t *ptep)
21 #if defined(CONFIG_PHYS_ADDR_T_64BIT) && defined(CONFIG_CPU_MIPS32)
25 pte.pte_low = ptep->pte_low;
27 pte.pte_high = ptep->pte_high;
29 if (unlikely(pte.pte_low != ptep->pte_low))
34 return READ_ONCE(*ptep);
38 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
39 int write, struct page **pages, int *nr)
41 pte_t *ptep = pte_offset_map(&pmd, addr);
43 pte_t pte = gup_get_pte(ptep);
46 if (!pte_present(pte) ||
47 pte_special(pte) || (write && !pte_write(pte))) {
51 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
54 SetPageReferenced(page);
58 } while (ptep++, addr += PAGE_SIZE, addr != end);
64 static inline void get_head_page_multiple(struct page *page, int nr)
66 VM_BUG_ON(page != compound_head(page));
67 VM_BUG_ON(page_count(page) == 0);
68 page_ref_add(page, nr);
69 SetPageReferenced(page);
72 static int gup_huge_pmd(pmd_t pmd, unsigned long addr, unsigned long end,
73 int write, struct page **pages, int *nr)
75 pte_t pte = *(pte_t *)&pmd;
76 struct page *head, *page;
79 if (write && !pte_write(pte))
81 /* hugepages are never "special" */
82 VM_BUG_ON(pte_special(pte));
83 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
87 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
89 VM_BUG_ON(compound_head(page) != head);
94 } while (addr += PAGE_SIZE, addr != end);
96 get_head_page_multiple(head, refs);
100 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
101 int write, struct page **pages, int *nr)
106 pmdp = pmd_offset(&pud, addr);
110 next = pmd_addr_end(addr, end);
113 if (unlikely(pmd_huge(pmd))) {
114 if (!gup_huge_pmd(pmd, addr, next, write, pages,nr))
117 if (!gup_pte_range(pmd, addr, next, write, pages,nr))
120 } while (pmdp++, addr = next, addr != end);
125 static int gup_huge_pud(pud_t pud, unsigned long addr, unsigned long end,
126 int write, struct page **pages, int *nr)
128 pte_t pte = *(pte_t *)&pud;
129 struct page *head, *page;
132 if (write && !pte_write(pte))
134 /* hugepages are never "special" */
135 VM_BUG_ON(pte_special(pte));
136 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
139 head = pte_page(pte);
140 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
142 VM_BUG_ON(compound_head(page) != head);
147 } while (addr += PAGE_SIZE, addr != end);
149 get_head_page_multiple(head, refs);
153 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
154 int write, struct page **pages, int *nr)
159 pudp = pud_offset(&pgd, addr);
163 next = pud_addr_end(addr, end);
166 if (unlikely(pud_huge(pud))) {
167 if (!gup_huge_pud(pud, addr, next, write, pages,nr))
170 if (!gup_pmd_range(pud, addr, next, write, pages,nr))
173 } while (pudp++, addr = next, addr != end);
179 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
180 * back to the regular GUP.
182 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
185 struct mm_struct *mm = current->mm;
186 unsigned long addr, len, end;
194 len = (unsigned long) nr_pages << PAGE_SHIFT;
196 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
197 (void __user *)start, len)))
201 * XXX: batch / limit 'nr', to avoid large irq off latency
202 * needs some instrumenting to determine the common sizes used by
203 * important workloads (eg. DB2), and whether limiting the batch
204 * size will decrease performance.
206 * It seems like we're in the clear for the moment. Direct-IO is
207 * the main guy that batches up lots of get_user_pages, and even
208 * they are limited to 64-at-a-time which is not so many.
211 * This doesn't prevent pagetable teardown, but does prevent
212 * the pagetables and pages from being freed.
214 * So long as we atomically load page table pointers versus teardown,
215 * we can follow the address down to the page and take a ref on it.
217 local_irq_save(flags);
218 pgdp = pgd_offset(mm, addr);
222 next = pgd_addr_end(addr, end);
225 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
227 } while (pgdp++, addr = next, addr != end);
228 local_irq_restore(flags);
234 * get_user_pages_fast() - pin user pages in memory
235 * @start: starting user address
236 * @nr_pages: number of pages from start to pin
237 * @write: whether pages will be written to
238 * @pages: array that receives pointers to the pages pinned.
239 * Should be at least nr_pages long.
241 * Attempt to pin user pages in memory without taking mm->mmap_sem.
242 * If not successful, it will fall back to taking the lock and
243 * calling get_user_pages().
245 * Returns number of pages pinned. This may be fewer than the number
246 * requested. If nr_pages is 0 or negative, returns 0. If no pages
247 * were pinned, returns -errno.
249 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
252 struct mm_struct *mm = current->mm;
253 unsigned long addr, len, end;
260 len = (unsigned long) nr_pages << PAGE_SHIFT;
263 if (end < start || cpu_has_dc_aliases)
266 /* XXX: batch / limit 'nr' */
268 pgdp = pgd_offset(mm, addr);
272 next = pgd_addr_end(addr, end);
276 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
277 * because get_user_pages() may need to cause an early COW in
278 * order to avoid confusing the normal COW routines. So only
279 * targets that are already writable are safe to do by just
280 * looking at the page tables.
282 if (!gup_pud_range(pgd, addr, next, 1, pages, &nr))
284 } while (pgdp++, addr = next, addr != end);
287 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
293 /* Try to get the remaining pages with get_user_pages */
294 start += nr << PAGE_SHIFT;
297 ret = get_user_pages_unlocked(start, (end - start) >> PAGE_SHIFT,
298 pages, write ? FOLL_WRITE : 0);
300 /* Have to be a bit careful with return values */