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		/* Similar to the PMD case, NUMA hinting must take slow path */
87		if (pte_numa(pte)) {
88			pte_unmap(ptep);
89			return 0;
90		}
91
92		if ((pte_flags(pte) & (mask | _PAGE_SPECIAL)) != mask) {
93			pte_unmap(ptep);
94			return 0;
95		}
96		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
97		page = pte_page(pte);
98		get_page(page);
99		SetPageReferenced(page);
100		pages[*nr] = page;
101		(*nr)++;
102
103	} while (ptep++, addr += PAGE_SIZE, addr != end);
104	pte_unmap(ptep - 1);
105
106	return 1;
107}
108
109static inline void get_head_page_multiple(struct page *page, int nr)
110{
111	VM_BUG_ON_PAGE(page != compound_head(page), page);
112	VM_BUG_ON_PAGE(page_count(page) == 0, page);
113	atomic_add(nr, &page->_count);
114	SetPageReferenced(page);
115}
116
117static noinline int gup_huge_pmd(pmd_t pmd, unsigned long addr,
118		unsigned long end, int write, struct page **pages, int *nr)
119{
120	unsigned long mask;
121	pte_t pte = *(pte_t *)&pmd;
122	struct page *head, *page;
123	int refs;
124
125	mask = _PAGE_PRESENT|_PAGE_USER;
126	if (write)
127		mask |= _PAGE_RW;
128	if ((pte_flags(pte) & mask) != mask)
129		return 0;
130	/* hugepages are never "special" */
131	VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
132	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
133
134	refs = 0;
135	head = pte_page(pte);
136	page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
137	do {
138		VM_BUG_ON_PAGE(compound_head(page) != head, page);
139		pages[*nr] = page;
140		if (PageTail(page))
141			get_huge_page_tail(page);
142		(*nr)++;
143		page++;
144		refs++;
145	} while (addr += PAGE_SIZE, addr != end);
146	get_head_page_multiple(head, refs);
147
148	return 1;
149}
150
151static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
152		int write, struct page **pages, int *nr)
153{
154	unsigned long next;
155	pmd_t *pmdp;
156
157	pmdp = pmd_offset(&pud, addr);
158	do {
159		pmd_t pmd = *pmdp;
160
161		next = pmd_addr_end(addr, end);
162		/*
163		 * The pmd_trans_splitting() check below explains why
164		 * pmdp_splitting_flush has to flush the tlb, to stop
165		 * this gup-fast code from running while we set the
166		 * splitting bit in the pmd. Returning zero will take
167		 * the slow path that will call wait_split_huge_page()
168		 * if the pmd is still in splitting state. gup-fast
169		 * can't because it has irq disabled and
170		 * wait_split_huge_page() would never return as the
171		 * tlb flush IPI wouldn't run.
172		 */
173		if (pmd_none(pmd) || pmd_trans_splitting(pmd))
174			return 0;
175		if (unlikely(pmd_large(pmd))) {
176			/*
177			 * NUMA hinting faults need to be handled in the GUP
178			 * slowpath for accounting purposes and so that they
179			 * can be serialised against THP migration.
180			 */
181			if (pmd_numa(pmd))
182				return 0;
183			if (!gup_huge_pmd(pmd, addr, next, write, pages, nr))
184				return 0;
185		} else {
186			if (!gup_pte_range(pmd, addr, next, write, pages, nr))
187				return 0;
188		}
189	} while (pmdp++, addr = next, addr != end);
190
191	return 1;
192}
193
194static noinline int gup_huge_pud(pud_t pud, unsigned long addr,
195		unsigned long end, int write, struct page **pages, int *nr)
196{
197	unsigned long mask;
198	pte_t pte = *(pte_t *)&pud;
199	struct page *head, *page;
200	int refs;
201
202	mask = _PAGE_PRESENT|_PAGE_USER;
203	if (write)
204		mask |= _PAGE_RW;
205	if ((pte_flags(pte) & mask) != mask)
206		return 0;
207	/* hugepages are never "special" */
208	VM_BUG_ON(pte_flags(pte) & _PAGE_SPECIAL);
209	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
210
211	refs = 0;
212	head = pte_page(pte);
213	page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
214	do {
215		VM_BUG_ON_PAGE(compound_head(page) != head, page);
216		pages[*nr] = page;
217		if (PageTail(page))
218			get_huge_page_tail(page);
219		(*nr)++;
220		page++;
221		refs++;
222	} while (addr += PAGE_SIZE, addr != end);
223	get_head_page_multiple(head, refs);
224
225	return 1;
226}
227
228static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
229			int write, struct page **pages, int *nr)
230{
231	unsigned long next;
232	pud_t *pudp;
233
234	pudp = pud_offset(&pgd, addr);
235	do {
236		pud_t pud = *pudp;
237
238		next = pud_addr_end(addr, end);
239		if (pud_none(pud))
240			return 0;
241		if (unlikely(pud_large(pud))) {
242			if (!gup_huge_pud(pud, addr, next, write, pages, nr))
243				return 0;
244		} else {
245			if (!gup_pmd_range(pud, addr, next, write, pages, nr))
246				return 0;
247		}
248	} while (pudp++, addr = next, addr != end);
249
250	return 1;
251}
252
253/*
254 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
255 * back to the regular GUP.
256 */
257int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
258			  struct page **pages)
259{
260	struct mm_struct *mm = current->mm;
261	unsigned long addr, len, end;
262	unsigned long next;
263	unsigned long flags;
264	pgd_t *pgdp;
265	int nr = 0;
266
267	start &= PAGE_MASK;
268	addr = start;
269	len = (unsigned long) nr_pages << PAGE_SHIFT;
270	end = start + len;
271	if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
272					(void __user *)start, len)))
273		return 0;
274
275	/*
276	 * XXX: batch / limit 'nr', to avoid large irq off latency
277	 * needs some instrumenting to determine the common sizes used by
278	 * important workloads (eg. DB2), and whether limiting the batch size
279	 * will decrease performance.
280	 *
281	 * It seems like we're in the clear for the moment. Direct-IO is
282	 * the main guy that batches up lots of get_user_pages, and even
283	 * they are limited to 64-at-a-time which is not so many.
284	 */
285	/*
286	 * This doesn't prevent pagetable teardown, but does prevent
287	 * the pagetables and pages from being freed on x86.
288	 *
289	 * So long as we atomically load page table pointers versus teardown
290	 * (which we do on x86, with the above PAE exception), we can follow the
291	 * address down to the the page and take a ref on it.
292	 */
293	local_irq_save(flags);
294	pgdp = pgd_offset(mm, addr);
295	do {
296		pgd_t pgd = *pgdp;
297
298		next = pgd_addr_end(addr, end);
299		if (pgd_none(pgd))
300			break;
301		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
302			break;
303	} while (pgdp++, addr = next, addr != end);
304	local_irq_restore(flags);
305
306	return nr;
307}
308
309/**
310 * get_user_pages_fast() - pin user pages in memory
311 * @start:	starting user address
312 * @nr_pages:	number of pages from start to pin
313 * @write:	whether pages will be written to
314 * @pages:	array that receives pointers to the pages pinned.
315 * 		Should be at least nr_pages long.
316 *
317 * Attempt to pin user pages in memory without taking mm->mmap_sem.
318 * If not successful, it will fall back to taking the lock and
319 * calling get_user_pages().
320 *
321 * Returns number of pages pinned. This may be fewer than the number
322 * requested. If nr_pages is 0 or negative, returns 0. If no pages
323 * were pinned, returns -errno.
324 */
325int get_user_pages_fast(unsigned long start, int nr_pages, int write,
326			struct page **pages)
327{
328	struct mm_struct *mm = current->mm;
329	unsigned long addr, len, end;
330	unsigned long next;
331	pgd_t *pgdp;
332	int nr = 0;
333
334	start &= PAGE_MASK;
335	addr = start;
336	len = (unsigned long) nr_pages << PAGE_SHIFT;
337
338	end = start + len;
339	if (end < start)
340		goto slow_irqon;
341
342#ifdef CONFIG_X86_64
343	if (end >> __VIRTUAL_MASK_SHIFT)
344		goto slow_irqon;
345#endif
346
347	/*
348	 * XXX: batch / limit 'nr', to avoid large irq off latency
349	 * needs some instrumenting to determine the common sizes used by
350	 * important workloads (eg. DB2), and whether limiting the batch size
351	 * will decrease performance.
352	 *
353	 * It seems like we're in the clear for the moment. Direct-IO is
354	 * the main guy that batches up lots of get_user_pages, and even
355	 * they are limited to 64-at-a-time which is not so many.
356	 */
357	/*
358	 * This doesn't prevent pagetable teardown, but does prevent
359	 * the pagetables and pages from being freed on x86.
360	 *
361	 * So long as we atomically load page table pointers versus teardown
362	 * (which we do on x86, with the above PAE exception), we can follow the
363	 * address down to the the page and take a ref on it.
364	 */
365	local_irq_disable();
366	pgdp = pgd_offset(mm, addr);
367	do {
368		pgd_t pgd = *pgdp;
369
370		next = pgd_addr_end(addr, end);
371		if (pgd_none(pgd))
372			goto slow;
373		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
374			goto slow;
375	} while (pgdp++, addr = next, addr != end);
376	local_irq_enable();
377
378	VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
379	return nr;
380
381	{
382		int ret;
383
384slow:
385		local_irq_enable();
386slow_irqon:
387		/* Try to get the remaining pages with get_user_pages */
388		start += nr << PAGE_SHIFT;
389		pages += nr;
390
391		down_read(&mm->mmap_sem);
392		ret = get_user_pages(current, mm, start,
393			(end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
394		up_read(&mm->mmap_sem);
395
396		/* Have to be a bit careful with return values */
397		if (nr > 0) {
398			if (ret < 0)
399				ret = nr;
400			else
401				ret += nr;
402		}
403
404		return ret;
405	}
406}
407