1 /*
2 * Lockless get_user_pages_fast for x86
3 *
4 * Copyright (C) 2008 Nick Piggin
5 * Copyright (C) 2008 Novell Inc.
6 */
7 #include <linux/sched.h>
8 #include <linux/mm.h>
9 #include <linux/vmstat.h>
10 #include <linux/highmem.h>
11 #include <linux/swap.h>
12
13 #include <asm/pgtable.h>
14
gup_get_pte(pte_t * ptep)15 static inline pte_t gup_get_pte(pte_t *ptep)
16 {
17 #ifndef CONFIG_X86_PAE
18 return ACCESS_ONCE(*ptep);
19 #else
20 /*
21 * With get_user_pages_fast, we walk down the pagetables without taking
22 * any locks. For this we would like to load the pointers atomically,
23 * but that is not possible (without expensive cmpxchg8b) on PAE. What
24 * we do have is the guarantee that a pte will only either go from not
25 * present to present, or present to not present or both -- it will not
26 * switch to a completely different present page without a TLB flush in
27 * between; something that we are blocking by holding interrupts off.
28 *
29 * Setting ptes from not present to present goes:
30 * ptep->pte_high = h;
31 * smp_wmb();
32 * ptep->pte_low = l;
33 *
34 * And present to not present goes:
35 * ptep->pte_low = 0;
36 * smp_wmb();
37 * ptep->pte_high = 0;
38 *
39 * We must ensure here that the load of pte_low sees l iff pte_high
40 * sees h. We load pte_high *after* loading pte_low, which ensures we
41 * don't see an older value of pte_high. *Then* we recheck pte_low,
42 * which ensures that we haven't picked up a changed pte high. We might
43 * have got rubbish values from pte_low and pte_high, but we are
44 * guaranteed that pte_low will not have the present bit set *unless*
45 * it is 'l'. And get_user_pages_fast only operates on present ptes, so
46 * we're safe.
47 *
48 * gup_get_pte should not be used or copied outside gup.c without being
49 * very careful -- it does not atomically load the pte or anything that
50 * is likely to be useful for you.
51 */
52 pte_t pte;
53
54 retry:
55 pte.pte_low = ptep->pte_low;
56 smp_rmb();
57 pte.pte_high = ptep->pte_high;
58 smp_rmb();
59 if (unlikely(pte.pte_low != ptep->pte_low))
60 goto retry;
61
62 return pte;
63 #endif
64 }
65
66 /*
67 * The performance critical leaf functions are made noinline otherwise gcc
68 * inlines everything into a single function which results in too much
69 * register pressure.
70 */
gup_pte_range(pmd_t pmd,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)71 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
72 unsigned long end, int write, struct page **pages, int *nr)
73 {
74 unsigned long mask;
75 pte_t *ptep;
76
77 mask = _PAGE_PRESENT|_PAGE_USER;
78 if (write)
79 mask |= _PAGE_RW;
80
81 ptep = pte_offset_map(&pmd, addr);
82 do {
83 pte_t pte = gup_get_pte(ptep);
84 struct page *page;
85
86 if ((pte_flags(pte) & (mask | _PAGE_SPECIAL)) != mask) {
87 pte_unmap(ptep);
88 return 0;
89 }
90 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
91 page = pte_page(pte);
92 get_page(page);
93 SetPageReferenced(page);
94 pages[*nr] = page;
95 (*nr)++;
96
97 } while (ptep++, addr += PAGE_SIZE, addr != end);
98 pte_unmap(ptep - 1);
99
100 return 1;
101 }
102
get_head_page_multiple(struct page * page,int nr)103 static inline void get_head_page_multiple(struct page *page, int nr)
104 {
105 VM_BUG_ON(page != compound_head(page));
106 VM_BUG_ON(page_count(page) == 0);
107 atomic_add(nr, &page->_count);
108 SetPageReferenced(page);
109 }
110
gup_huge_pmd(pmd_t pmd,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)111 static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
112 unsigned long end, int write, struct page **pages, int *nr)
113 {
114 unsigned long mask;
115 pte_t pte = *(pte_t *)&pmd;
116 struct page *head, *page;
117 int refs;
118
119 mask = _PAGE_PRESENT|_PAGE_USER;
120 if (write)
121 mask |= _PAGE_RW;
122 if ((pte_flags(pte) & mask) != mask)
123 return 0;
124 /* hugepages are never "special" */
125 VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
126 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
127
128 refs = 0;
129 head = pte_page(pte);
130 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
131 do {
132 VM_BUG_ON(compound_head(page) != head);
133 pages[*nr] = page;
134 if (PageTail(page))
135 get_huge_page_tail(page);
136 (*nr)++;
137 page++;
138 refs++;
139 } while (addr += PAGE_SIZE, addr != end);
140 get_head_page_multiple(head, refs);
141
142 return 1;
143 }
144
gup_pmd_range(pud_t pud,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)145 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
146 int write, struct page **pages, int *nr)
147 {
148 unsigned long next;
149 pmd_t *pmdp;
150
151 pmdp = pmd_offset(&pud, addr);
152 do {
153 pmd_t pmd = *pmdp;
154
155 next = pmd_addr_end(addr, end);
156 /*
157 * The pmd_trans_splitting() check below explains why
158 * pmdp_splitting_flush has to flush the tlb, to stop
159 * this gup-fast code from running while we set the
160 * splitting bit in the pmd. Returning zero will take
161 * the slow path that will call wait_split_huge_page()
162 * if the pmd is still in splitting state. gup-fast
163 * can't because it has irq disabled and
164 * wait_split_huge_page() would never return as the
165 * tlb flush IPI wouldn't run.
166 */
167 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
168 return 0;
169 if (unlikely(pmd_large(pmd))) {
170 if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
171 return 0;
172 } else {
173 if (!gup_pte_range(pmd, addr, next, write, pages, nr))
174 return 0;
175 }
176 } while (pmdp++, addr = next, addr != end);
177
178 return 1;
179 }
180
gup_huge_pud(pud_t pud,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)181 static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
182 unsigned long end, int write, struct page **pages, int *nr)
183 {
184 unsigned long mask;
185 pte_t pte = *(pte_t *)&pud;
186 struct page *head, *page;
187 int refs;
188
189 mask = _PAGE_PRESENT|_PAGE_USER;
190 if (write)
191 mask |= _PAGE_RW;
192 if ((pte_flags(pte) & mask) != mask)
193 return 0;
194 /* hugepages are never "special" */
195 VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
196 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
197
198 refs = 0;
199 head = pte_page(pte);
200 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
201 do {
202 VM_BUG_ON(compound_head(page) != head);
203 pages[*nr] = page;
204 if (PageTail(page))
205 get_huge_page_tail(page);
206 (*nr)++;
207 page++;
208 refs++;
209 } while (addr += PAGE_SIZE, addr != end);
210 get_head_page_multiple(head, refs);
211
212 return 1;
213 }
214
gup_pud_range(pgd_t pgd,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)215 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
216 int write, struct page **pages, int *nr)
217 {
218 unsigned long next;
219 pud_t *pudp;
220
221 pudp = pud_offset(&pgd, addr);
222 do {
223 pud_t pud = *pudp;
224
225 next = pud_addr_end(addr, end);
226 if (pud_none(pud))
227 return 0;
228 if (unlikely(pud_large(pud))) {
229 if (!gup_huge_pud(pud, addr, next, write, pages, nr))
230 return 0;
231 } else {
232 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
233 return 0;
234 }
235 } while (pudp++, addr = next, addr != end);
236
237 return 1;
238 }
239
240 /*
241 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
242 * back to the regular GUP.
243 */
__get_user_pages_fast(unsigned long start,int nr_pages,int write,struct page ** pages)244 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
245 struct page **pages)
246 {
247 struct mm_struct *mm = current->mm;
248 unsigned long addr, len, end;
249 unsigned long next;
250 unsigned long flags;
251 pgd_t *pgdp;
252 int nr = 0;
253
254 start &= PAGE_MASK;
255 addr = start;
256 len = (unsigned long) nr_pages << PAGE_SHIFT;
257 end = start + len;
258 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
259 (void __user *)start, len)))
260 return 0;
261
262 /*
263 * XXX: batch / limit 'nr', to avoid large irq off latency
264 * needs some instrumenting to determine the common sizes used by
265 * important workloads (eg. DB2), and whether limiting the batch size
266 * will decrease performance.
267 *
268 * It seems like we're in the clear for the moment. Direct-IO is
269 * the main guy that batches up lots of get_user_pages, and even
270 * they are limited to 64-at-a-time which is not so many.
271 */
272 /*
273 * This doesn't prevent pagetable teardown, but does prevent
274 * the pagetables and pages from being freed on x86.
275 *
276 * So long as we atomically load page table pointers versus teardown
277 * (which we do on x86, with the above PAE exception), we can follow the
278 * address down to the the page and take a ref on it.
279 */
280 local_irq_save(flags);
281 pgdp = pgd_offset(mm, addr);
282 do {
283 pgd_t pgd = *pgdp;
284
285 next = pgd_addr_end(addr, end);
286 if (pgd_none(pgd))
287 break;
288 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
289 break;
290 } while (pgdp++, addr = next, addr != end);
291 local_irq_restore(flags);
292
293 return nr;
294 }
295
296 /**
297 * get_user_pages_fast() - pin user pages in memory
298 * @start: starting user address
299 * @nr_pages: number of pages from start to pin
300 * @write: whether pages will be written to
301 * @pages: array that receives pointers to the pages pinned.
302 * Should be at least nr_pages long.
303 *
304 * Attempt to pin user pages in memory without taking mm->mmap_sem.
305 * If not successful, it will fall back to taking the lock and
306 * calling get_user_pages().
307 *
308 * Returns number of pages pinned. This may be fewer than the number
309 * requested. If nr_pages is 0 or negative, returns 0. If no pages
310 * were pinned, returns -errno.
311 */
get_user_pages_fast(unsigned long start,int nr_pages,int write,struct page ** pages)312 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
313 struct page **pages)
314 {
315 struct mm_struct *mm = current->mm;
316 unsigned long addr, len, end;
317 unsigned long next;
318 pgd_t *pgdp;
319 int nr = 0;
320
321 start &= PAGE_MASK;
322 addr = start;
323 len = (unsigned long) nr_pages << PAGE_SHIFT;
324
325 end = start + len;
326 if (end < start)
327 goto slow_irqon;
328
329 #ifdef CONFIG_X86_64
330 if (end >> __VIRTUAL_MASK_SHIFT)
331 goto slow_irqon;
332 #endif
333
334 /*
335 * XXX: batch / limit 'nr', to avoid large irq off latency
336 * needs some instrumenting to determine the common sizes used by
337 * important workloads (eg. DB2), and whether limiting the batch size
338 * will decrease performance.
339 *
340 * It seems like we're in the clear for the moment. Direct-IO is
341 * the main guy that batches up lots of get_user_pages, and even
342 * they are limited to 64-at-a-time which is not so many.
343 */
344 /*
345 * This doesn't prevent pagetable teardown, but does prevent
346 * the pagetables and pages from being freed on x86.
347 *
348 * So long as we atomically load page table pointers versus teardown
349 * (which we do on x86, with the above PAE exception), we can follow the
350 * address down to the the page and take a ref on it.
351 */
352 local_irq_disable();
353 pgdp = pgd_offset(mm, addr);
354 do {
355 pgd_t pgd = *pgdp;
356
357 next = pgd_addr_end(addr, end);
358 if (pgd_none(pgd))
359 goto slow;
360 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
361 goto slow;
362 } while (pgdp++, addr = next, addr != end);
363 local_irq_enable();
364
365 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
366 return nr;
367
368 {
369 int ret;
370
371 slow:
372 local_irq_enable();
373 slow_irqon:
374 /* Try to get the remaining pages with get_user_pages */
375 start += nr << PAGE_SHIFT;
376 pages += nr;
377
378 down_read(&mm->mmap_sem);
379 ret = get_user_pages(current, mm, start,
380 (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
381 up_read(&mm->mmap_sem);
382
383 /* Have to be a bit careful with return values */
384 if (nr > 0) {
385 if (ret < 0)
386 ret = nr;
387 else
388 ret += nr;
389 }
390
391 return ret;
392 }
393 }
394