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
 15static 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
 54retry:
 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 */
 71static 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
103static 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
111static inline void get_huge_page_tail(struct page *page)
112{
113	/*
114	 * __split_huge_page_refcount() cannot run
115	 * from under us.
116	 */
117	VM_BUG_ON(atomic_read(&page->_count) < 0);
118	atomic_inc(&page->_count);
119}
120
121static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
122		unsigned long end, int write, struct page **pages, int *nr)
123{
124	unsigned long mask;
125	pte_t pte = *(pte_t *)&pmd;
126	struct page *head, *page;
127	int refs;
128
129	mask = _PAGE_PRESENT|_PAGE_USER;
130	if (write)
131		mask |= _PAGE_RW;
132	if ((pte_flags(pte) & mask) != mask)
133		return 0;
134	/* hugepages are never "special" */
135	VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
136	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
137
138	refs = 0;
139	head = pte_page(pte);
140	page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
141	do {
142		VM_BUG_ON(compound_head(page) != head);
143		pages[*nr] = page;
144		if (PageTail(page))
145			get_huge_page_tail(page);
146		(*nr)++;
147		page++;
148		refs++;
149	} while (addr += PAGE_SIZE, addr != end);
150	get_head_page_multiple(head, refs);
151
152	return 1;
153}
154
155static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
156		int write, struct page **pages, int *nr)
157{
158	unsigned long next;
159	pmd_t *pmdp;
160
161	pmdp = pmd_offset(&pud, addr);
162	do {
163		pmd_t pmd = *pmdp;
164
165		next = pmd_addr_end(addr, end);
166		/*
167		 * The pmd_trans_splitting() check below explains why
168		 * pmdp_splitting_flush has to flush the tlb, to stop
169		 * this gup-fast code from running while we set the
170		 * splitting bit in the pmd. Returning zero will take
171		 * the slow path that will call wait_split_huge_page()
172		 * if the pmd is still in splitting state. gup-fast
173		 * can't because it has irq disabled and
174		 * wait_split_huge_page() would never return as the
175		 * tlb flush IPI wouldn't run.
176		 */
177		if (pmd_none(pmd) || pmd_trans_splitting(pmd))
178			return 0;
179		if (unlikely(pmd_large(pmd))) {
180			if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
181				return 0;
182		} else {
183			if (!gup_pte_range(pmd, addr, next, write, pages, nr))
184				return 0;
185		}
186	} while (pmdp++, addr = next, addr != end);
187
188	return 1;
189}
190
191static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
192		unsigned long end, int write, struct page **pages, int *nr)
193{
194	unsigned long mask;
195	pte_t pte = *(pte_t *)&pud;
196	struct page *head, *page;
197	int refs;
198
199	mask = _PAGE_PRESENT|_PAGE_USER;
200	if (write)
201		mask |= _PAGE_RW;
202	if ((pte_flags(pte) & mask) != mask)
203		return 0;
204	/* hugepages are never "special" */
205	VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
206	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
207
208	refs = 0;
209	head = pte_page(pte);
210	page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
211	do {
212		VM_BUG_ON(compound_head(page) != head);
213		pages[*nr] = page;
214		(*nr)++;
215		page++;
216		refs++;
217	} while (addr += PAGE_SIZE, addr != end);
218	get_head_page_multiple(head, refs);
219
220	return 1;
221}
222
223static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
224			int write, struct page **pages, int *nr)
225{
226	unsigned long next;
227	pud_t *pudp;
228
229	pudp = pud_offset(&pgd, addr);
230	do {
231		pud_t pud = *pudp;
232
233		next = pud_addr_end(addr, end);
234		if (pud_none(pud))
235			return 0;
236		if (unlikely(pud_large(pud))) {
237			if (!gup_huge_pud(pud, addr, next, write, pages, nr))
238				return 0;
239		} else {
240			if (!gup_pmd_range(pud, addr, next, write, pages, nr))
241				return 0;
242		}
243	} while (pudp++, addr = next, addr != end);
244
245	return 1;
246}
247
248/*
249 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
250 * back to the regular GUP.
251 */
252int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
253			  struct page **pages)
254{
255	struct mm_struct *mm = current->mm;
256	unsigned long addr, len, end;
257	unsigned long next;
258	unsigned long flags;
259	pgd_t *pgdp;
260	int nr = 0;
261
262	start &= PAGE_MASK;
263	addr = start;
264	len = (unsigned long) nr_pages << PAGE_SHIFT;
265	end = start + len;
266	if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
267					(void __user *)start, len)))
268		return 0;
269
270	/*
271	 * XXX: batch / limit 'nr', to avoid large irq off latency
272	 * needs some instrumenting to determine the common sizes used by
273	 * important workloads (eg. DB2), and whether limiting the batch size
274	 * will decrease performance.
275	 *
276	 * It seems like we're in the clear for the moment. Direct-IO is
277	 * the main guy that batches up lots of get_user_pages, and even
278	 * they are limited to 64-at-a-time which is not so many.
279	 */
280	/*
281	 * This doesn't prevent pagetable teardown, but does prevent
282	 * the pagetables and pages from being freed on x86.
283	 *
284	 * So long as we atomically load page table pointers versus teardown
285	 * (which we do on x86, with the above PAE exception), we can follow the
286	 * address down to the the page and take a ref on it.
287	 */
288	local_irq_save(flags);
289	pgdp = pgd_offset(mm, addr);
290	do {
291		pgd_t pgd = *pgdp;
292
293		next = pgd_addr_end(addr, end);
294		if (pgd_none(pgd))
295			break;
296		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
297			break;
298	} while (pgdp++, addr = next, addr != end);
299	local_irq_restore(flags);
300
301	return nr;
302}
303
304/**
305 * get_user_pages_fast() - pin user pages in memory
306 * @start:	starting user address
307 * @nr_pages:	number of pages from start to pin
308 * @write:	whether pages will be written to
309 * @pages:	array that receives pointers to the pages pinned.
310 * 		Should be at least nr_pages long.
311 *
312 * Attempt to pin user pages in memory without taking mm->mmap_sem.
313 * If not successful, it will fall back to taking the lock and
314 * calling get_user_pages().
315 *
316 * Returns number of pages pinned. This may be fewer than the number
317 * requested. If nr_pages is 0 or negative, returns 0. If no pages
318 * were pinned, returns -errno.
319 */
320int get_user_pages_fast(unsigned long start, int nr_pages, int write,
321			struct page **pages)
322{
323	struct mm_struct *mm = current->mm;
324	unsigned long addr, len, end;
325	unsigned long next;
326	pgd_t *pgdp;
327	int nr = 0;
328
329	start &= PAGE_MASK;
330	addr = start;
331	len = (unsigned long) nr_pages << PAGE_SHIFT;
332
333	end = start + len;
334	if (end < start)
335		goto slow_irqon;
336
337#ifdef CONFIG_X86_64
338	if (end >> __VIRTUAL_MASK_SHIFT)
339		goto slow_irqon;
340#endif
341
342	/*
343	 * XXX: batch / limit 'nr', to avoid large irq off latency
344	 * needs some instrumenting to determine the common sizes used by
345	 * important workloads (eg. DB2), and whether limiting the batch size
346	 * will decrease performance.
347	 *
348	 * It seems like we're in the clear for the moment. Direct-IO is
349	 * the main guy that batches up lots of get_user_pages, and even
350	 * they are limited to 64-at-a-time which is not so many.
351	 */
352	/*
353	 * This doesn't prevent pagetable teardown, but does prevent
354	 * the pagetables and pages from being freed on x86.
355	 *
356	 * So long as we atomically load page table pointers versus teardown
357	 * (which we do on x86, with the above PAE exception), we can follow the
358	 * address down to the the page and take a ref on it.
359	 */
360	local_irq_disable();
361	pgdp = pgd_offset(mm, addr);
362	do {
363		pgd_t pgd = *pgdp;
364
365		next = pgd_addr_end(addr, end);
366		if (pgd_none(pgd))
367			goto slow;
368		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
369			goto slow;
370	} while (pgdp++, addr = next, addr != end);
371	local_irq_enable();
372
373	VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
374	return nr;
375
376	{
377		int ret;
378
379slow:
380		local_irq_enable();
381slow_irqon:
382		/* Try to get the remaining pages with get_user_pages */
383		start += nr << PAGE_SHIFT;
384		pages += nr;
385
386		down_read(&mm->mmap_sem);
387		ret = get_user_pages(current, mm, start,
388			(end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
389		up_read(&mm->mmap_sem);
390
391		/* Have to be a bit careful with return values */
392		if (nr > 0) {
393			if (ret < 0)
394				ret = nr;
395			else
396				ret += nr;
397		}
398
399		return ret;
400	}
401}