1 // SPDX-License-Identifier: GPL-2.0
3 * Lockless get_user_pages_fast for sparc, cribbed from powerpc
5 * Copyright (C) 2008 Nick Piggin
6 * Copyright (C) 2008 Novell Inc.
9 #include <linux/sched.h>
11 #include <linux/vmstat.h>
12 #include <linux/pagemap.h>
13 #include <linux/rwsem.h>
14 #include <asm/pgtable.h>
18 * The performance critical leaf functions are made noinline otherwise gcc
19 * inlines everything into a single function which results in too much
22 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
23 unsigned long end, int write, struct page **pages, int *nr)
25 unsigned long mask, result;
28 if (tlb_type == hypervisor) {
29 result = _PAGE_PRESENT_4V|_PAGE_P_4V;
31 result |= _PAGE_WRITE_4V;
33 result = _PAGE_PRESENT_4U|_PAGE_P_4U;
35 result |= _PAGE_WRITE_4U;
37 mask = result | _PAGE_SPECIAL;
39 ptep = pte_offset_kernel(&pmd, addr);
41 struct page *page, *head;
44 if ((pte_val(pte) & mask) != result)
46 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
48 /* The hugepage case is simplified on sparc64 because
49 * we encode the sub-page pfn offsets into the
50 * hugepage PTEs. We could optimize this in the future
51 * use page_cache_add_speculative() for the hugepage case.
54 head = compound_head(page);
55 if (!page_cache_get_speculative(head))
57 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
64 } while (ptep++, addr += PAGE_SIZE, addr != end);
69 static int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
70 unsigned long end, int write, struct page **pages,
73 struct page *head, *page;
76 if (!(pmd_val(pmd) & _PAGE_VALID))
79 if (write && !pmd_write(pmd))
83 page = pmd_page(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
84 head = compound_head(page);
86 VM_BUG_ON(compound_head(page) != head);
91 } while (addr += PAGE_SIZE, addr != end);
93 if (!page_cache_add_speculative(head, refs)) {
98 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
108 static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr,
109 unsigned long end, int write, struct page **pages,
112 struct page *head, *page;
115 if (!(pud_val(pud) & _PAGE_VALID))
118 if (write && !pud_write(pud))
122 page = pud_page(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
123 head = compound_head(page);
125 VM_BUG_ON(compound_head(page) != head);
130 } while (addr += PAGE_SIZE, addr != end);
132 if (!page_cache_add_speculative(head, refs)) {
137 if (unlikely(pud_val(pud) != pud_val(*pudp))) {
147 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
148 int write, struct page **pages, int *nr)
153 pmdp = pmd_offset(&pud, addr);
157 next = pmd_addr_end(addr, end);
160 if (unlikely(pmd_large(pmd))) {
161 if (!gup_huge_pmd(pmdp, pmd, addr, next,
164 } else if (!gup_pte_range(pmd, addr, next, write,
167 } while (pmdp++, addr = next, addr != end);
172 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
173 int write, struct page **pages, int *nr)
178 pudp = pud_offset(&pgd, addr);
182 next = pud_addr_end(addr, end);
185 if (unlikely(pud_large(pud))) {
186 if (!gup_huge_pud(pudp, pud, addr, next,
189 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
191 } while (pudp++, addr = next, addr != end);
197 * Note a difference with get_user_pages_fast: this always returns the
198 * number of pages pinned, 0 if no pages were pinned.
200 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
203 struct mm_struct *mm = current->mm;
204 unsigned long addr, len, end;
205 unsigned long next, flags;
209 #ifdef CONFIG_SPARC64
213 /* If userspace has passed a versioned address, kernel
214 * will not find it in the VMAs since it does not store
215 * the version tags in the list of VMAs. Storing version
216 * tags in list of VMAs is impractical since they can be
217 * changed any time from userspace without dropping into
218 * kernel. Any address search in VMAs will be done with
219 * non-versioned addresses. Ensure the ADI version bits
220 * are dropped here by sign extending the last bit before
221 * ADI bits. IOMMU does not implement version tags.
223 addr = (addr << (long)adi_nbits()) >> (long)adi_nbits();
229 len = (unsigned long) nr_pages << PAGE_SHIFT;
232 local_irq_save(flags);
233 pgdp = pgd_offset(mm, addr);
237 next = pgd_addr_end(addr, end);
240 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
242 } while (pgdp++, addr = next, addr != end);
243 local_irq_restore(flags);
248 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
251 struct mm_struct *mm = current->mm;
252 unsigned long addr, len, end;
257 #ifdef CONFIG_SPARC64
261 /* If userspace has passed a versioned address, kernel
262 * will not find it in the VMAs since it does not store
263 * the version tags in the list of VMAs. Storing version
264 * tags in list of VMAs is impractical since they can be
265 * changed any time from userspace without dropping into
266 * kernel. Any address search in VMAs will be done with
267 * non-versioned addresses. Ensure the ADI version bits
268 * are dropped here by sign extending the last bit before
269 * ADI bits. IOMMU does not implements version tags,
271 addr = (addr << (long)adi_nbits()) >> (long)adi_nbits();
277 len = (unsigned long) nr_pages << PAGE_SHIFT;
281 * XXX: batch / limit 'nr', to avoid large irq off latency
282 * needs some instrumenting to determine the common sizes used by
283 * important workloads (eg. DB2), and whether limiting the batch size
284 * will decrease performance.
286 * It seems like we're in the clear for the moment. Direct-IO is
287 * the main guy that batches up lots of get_user_pages, and even
288 * they are limited to 64-at-a-time which is not so many.
291 * This doesn't prevent pagetable teardown, but does prevent
292 * the pagetables from being freed on sparc.
294 * So long as we atomically load page table pointers versus teardown,
295 * we can follow the address down to the the page and take a ref on it.
299 pgdp = pgd_offset(mm, addr);
303 next = pgd_addr_end(addr, end);
307 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
308 * because get_user_pages() may need to cause an early COW in
309 * order to avoid confusing the normal COW routines. So only
310 * targets that are already writable are safe to do by just
311 * looking at the page tables.
313 if (!gup_pud_range(pgd, addr, next, 1, pages, &nr))
315 } while (pgdp++, addr = next, addr != end);
319 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
328 /* Try to get the remaining pages with get_user_pages */
329 start += nr << PAGE_SHIFT;
332 ret = get_user_pages_unlocked(start,
333 (end - start) >> PAGE_SHIFT, pages,
334 write ? FOLL_WRITE : 0);
336 /* Have to be a bit careful with return values */