memory.c revision 5d2a2dbbc1025dbf7998b9289574d9592b8f21cc
1/*
2 *  linux/mm/memory.c
3 *
4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27 *		Found it. Everything seems to work now.
28 * 20.12.91  -  Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94  -  Multi-page memory management added for v1.1.
33 * 		Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 *		(Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/hugetlb.h>
44#include <linux/mman.h>
45#include <linux/swap.h>
46#include <linux/highmem.h>
47#include <linux/pagemap.h>
48#include <linux/rmap.h>
49#include <linux/module.h>
50#include <linux/init.h>
51
52#include <asm/pgalloc.h>
53#include <asm/uaccess.h>
54#include <asm/tlb.h>
55#include <asm/tlbflush.h>
56#include <asm/pgtable.h>
57
58#include <linux/swapops.h>
59#include <linux/elf.h>
60
61#ifndef CONFIG_NEED_MULTIPLE_NODES
62/* use the per-pgdat data instead for discontigmem - mbligh */
63unsigned long max_mapnr;
64struct page *mem_map;
65
66EXPORT_SYMBOL(max_mapnr);
67EXPORT_SYMBOL(mem_map);
68#endif
69
70unsigned long num_physpages;
71/*
72 * A number of key systems in x86 including ioremap() rely on the assumption
73 * that high_memory defines the upper bound on direct map memory, then end
74 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76 * and ZONE_HIGHMEM.
77 */
78void * high_memory;
79unsigned long vmalloc_earlyreserve;
80
81EXPORT_SYMBOL(num_physpages);
82EXPORT_SYMBOL(high_memory);
83EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85/*
86 * If a p?d_bad entry is found while walking page tables, report
87 * the error, before resetting entry to p?d_none.  Usually (but
88 * very seldom) called out from the p?d_none_or_clear_bad macros.
89 */
90
91void pgd_clear_bad(pgd_t *pgd)
92{
93	pgd_ERROR(*pgd);
94	pgd_clear(pgd);
95}
96
97void pud_clear_bad(pud_t *pud)
98{
99	pud_ERROR(*pud);
100	pud_clear(pud);
101}
102
103void pmd_clear_bad(pmd_t *pmd)
104{
105	pmd_ERROR(*pmd);
106	pmd_clear(pmd);
107}
108
109/*
110 * Note: this doesn't free the actual pages themselves. That
111 * has been handled earlier when unmapping all the memory regions.
112 */
113static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
114{
115	struct page *page = pmd_page(*pmd);
116	pmd_clear(pmd);
117	pte_lock_deinit(page);
118	pte_free_tlb(tlb, page);
119	dec_page_state(nr_page_table_pages);
120	tlb->mm->nr_ptes--;
121}
122
123static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
124				unsigned long addr, unsigned long end,
125				unsigned long floor, unsigned long ceiling)
126{
127	pmd_t *pmd;
128	unsigned long next;
129	unsigned long start;
130
131	start = addr;
132	pmd = pmd_offset(pud, addr);
133	do {
134		next = pmd_addr_end(addr, end);
135		if (pmd_none_or_clear_bad(pmd))
136			continue;
137		free_pte_range(tlb, pmd);
138	} while (pmd++, addr = next, addr != end);
139
140	start &= PUD_MASK;
141	if (start < floor)
142		return;
143	if (ceiling) {
144		ceiling &= PUD_MASK;
145		if (!ceiling)
146			return;
147	}
148	if (end - 1 > ceiling - 1)
149		return;
150
151	pmd = pmd_offset(pud, start);
152	pud_clear(pud);
153	pmd_free_tlb(tlb, pmd);
154}
155
156static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
157				unsigned long addr, unsigned long end,
158				unsigned long floor, unsigned long ceiling)
159{
160	pud_t *pud;
161	unsigned long next;
162	unsigned long start;
163
164	start = addr;
165	pud = pud_offset(pgd, addr);
166	do {
167		next = pud_addr_end(addr, end);
168		if (pud_none_or_clear_bad(pud))
169			continue;
170		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
171	} while (pud++, addr = next, addr != end);
172
173	start &= PGDIR_MASK;
174	if (start < floor)
175		return;
176	if (ceiling) {
177		ceiling &= PGDIR_MASK;
178		if (!ceiling)
179			return;
180	}
181	if (end - 1 > ceiling - 1)
182		return;
183
184	pud = pud_offset(pgd, start);
185	pgd_clear(pgd);
186	pud_free_tlb(tlb, pud);
187}
188
189/*
190 * This function frees user-level page tables of a process.
191 *
192 * Must be called with pagetable lock held.
193 */
194void free_pgd_range(struct mmu_gather **tlb,
195			unsigned long addr, unsigned long end,
196			unsigned long floor, unsigned long ceiling)
197{
198	pgd_t *pgd;
199	unsigned long next;
200	unsigned long start;
201
202	/*
203	 * The next few lines have given us lots of grief...
204	 *
205	 * Why are we testing PMD* at this top level?  Because often
206	 * there will be no work to do at all, and we'd prefer not to
207	 * go all the way down to the bottom just to discover that.
208	 *
209	 * Why all these "- 1"s?  Because 0 represents both the bottom
210	 * of the address space and the top of it (using -1 for the
211	 * top wouldn't help much: the masks would do the wrong thing).
212	 * The rule is that addr 0 and floor 0 refer to the bottom of
213	 * the address space, but end 0 and ceiling 0 refer to the top
214	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
215	 * that end 0 case should be mythical).
216	 *
217	 * Wherever addr is brought up or ceiling brought down, we must
218	 * be careful to reject "the opposite 0" before it confuses the
219	 * subsequent tests.  But what about where end is brought down
220	 * by PMD_SIZE below? no, end can't go down to 0 there.
221	 *
222	 * Whereas we round start (addr) and ceiling down, by different
223	 * masks at different levels, in order to test whether a table
224	 * now has no other vmas using it, so can be freed, we don't
225	 * bother to round floor or end up - the tests don't need that.
226	 */
227
228	addr &= PMD_MASK;
229	if (addr < floor) {
230		addr += PMD_SIZE;
231		if (!addr)
232			return;
233	}
234	if (ceiling) {
235		ceiling &= PMD_MASK;
236		if (!ceiling)
237			return;
238	}
239	if (end - 1 > ceiling - 1)
240		end -= PMD_SIZE;
241	if (addr > end - 1)
242		return;
243
244	start = addr;
245	pgd = pgd_offset((*tlb)->mm, addr);
246	do {
247		next = pgd_addr_end(addr, end);
248		if (pgd_none_or_clear_bad(pgd))
249			continue;
250		free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
251	} while (pgd++, addr = next, addr != end);
252
253	if (!(*tlb)->fullmm)
254		flush_tlb_pgtables((*tlb)->mm, start, end);
255}
256
257void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
258		unsigned long floor, unsigned long ceiling)
259{
260	while (vma) {
261		struct vm_area_struct *next = vma->vm_next;
262		unsigned long addr = vma->vm_start;
263
264		/*
265		 * Hide vma from rmap and vmtruncate before freeing pgtables
266		 */
267		anon_vma_unlink(vma);
268		unlink_file_vma(vma);
269
270		if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
271			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
272				floor, next? next->vm_start: ceiling);
273		} else {
274			/*
275			 * Optimization: gather nearby vmas into one call down
276			 */
277			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
278			  && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
279							HPAGE_SIZE)) {
280				vma = next;
281				next = vma->vm_next;
282				anon_vma_unlink(vma);
283				unlink_file_vma(vma);
284			}
285			free_pgd_range(tlb, addr, vma->vm_end,
286				floor, next? next->vm_start: ceiling);
287		}
288		vma = next;
289	}
290}
291
292int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
293{
294	struct page *new = pte_alloc_one(mm, address);
295	if (!new)
296		return -ENOMEM;
297
298	pte_lock_init(new);
299	spin_lock(&mm->page_table_lock);
300	if (pmd_present(*pmd)) {	/* Another has populated it */
301		pte_lock_deinit(new);
302		pte_free(new);
303	} else {
304		mm->nr_ptes++;
305		inc_page_state(nr_page_table_pages);
306		pmd_populate(mm, pmd, new);
307	}
308	spin_unlock(&mm->page_table_lock);
309	return 0;
310}
311
312int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
313{
314	pte_t *new = pte_alloc_one_kernel(&init_mm, address);
315	if (!new)
316		return -ENOMEM;
317
318	spin_lock(&init_mm.page_table_lock);
319	if (pmd_present(*pmd))		/* Another has populated it */
320		pte_free_kernel(new);
321	else
322		pmd_populate_kernel(&init_mm, pmd, new);
323	spin_unlock(&init_mm.page_table_lock);
324	return 0;
325}
326
327static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
328{
329	if (file_rss)
330		add_mm_counter(mm, file_rss, file_rss);
331	if (anon_rss)
332		add_mm_counter(mm, anon_rss, anon_rss);
333}
334
335/*
336 * This function is called to print an error when a bad pte
337 * is found. For example, we might have a PFN-mapped pte in
338 * a region that doesn't allow it.
339 *
340 * The calling function must still handle the error.
341 */
342void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
343{
344	printk(KERN_ERR "Bad pte = %08llx, process = %s, "
345			"vm_flags = %lx, vaddr = %lx\n",
346		(long long)pte_val(pte),
347		(vma->vm_mm == current->mm ? current->comm : "???"),
348		vma->vm_flags, vaddr);
349	dump_stack();
350}
351
352/*
353 * This function gets the "struct page" associated with a pte.
354 *
355 * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
356 * will have each page table entry just pointing to a raw page frame
357 * number, and as far as the VM layer is concerned, those do not have
358 * pages associated with them - even if the PFN might point to memory
359 * that otherwise is perfectly fine and has a "struct page".
360 *
361 * The way we recognize those mappings is through the rules set up
362 * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
363 * and the vm_pgoff will point to the first PFN mapped: thus every
364 * page that is a raw mapping will always honor the rule
365 *
366 *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
367 *
368 * and if that isn't true, the page has been COW'ed (in which case it
369 * _does_ have a "struct page" associated with it even if it is in a
370 * VM_PFNMAP range).
371 */
372struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
373{
374	unsigned long pfn = pte_pfn(pte);
375
376	if (vma->vm_flags & VM_PFNMAP) {
377		unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
378		if (pfn == vma->vm_pgoff + off)
379			return NULL;
380	}
381
382	/*
383	 * Add some anal sanity checks for now. Eventually,
384	 * we should just do "return pfn_to_page(pfn)", but
385	 * in the meantime we check that we get a valid pfn,
386	 * and that the resulting page looks ok.
387	 *
388	 * Remove this test eventually!
389	 */
390	if (unlikely(!pfn_valid(pfn))) {
391		print_bad_pte(vma, pte, addr);
392		return NULL;
393	}
394
395	/*
396	 * NOTE! We still have PageReserved() pages in the page
397	 * tables.
398	 *
399	 * The PAGE_ZERO() pages and various VDSO mappings can
400	 * cause them to exist.
401	 */
402	return pfn_to_page(pfn);
403}
404
405/*
406 * copy one vm_area from one task to the other. Assumes the page tables
407 * already present in the new task to be cleared in the whole range
408 * covered by this vma.
409 */
410
411static inline void
412copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
413		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
414		unsigned long addr, int *rss)
415{
416	unsigned long vm_flags = vma->vm_flags;
417	pte_t pte = *src_pte;
418	struct page *page;
419
420	/* pte contains position in swap or file, so copy. */
421	if (unlikely(!pte_present(pte))) {
422		if (!pte_file(pte)) {
423			swap_duplicate(pte_to_swp_entry(pte));
424			/* make sure dst_mm is on swapoff's mmlist. */
425			if (unlikely(list_empty(&dst_mm->mmlist))) {
426				spin_lock(&mmlist_lock);
427				if (list_empty(&dst_mm->mmlist))
428					list_add(&dst_mm->mmlist,
429						 &src_mm->mmlist);
430				spin_unlock(&mmlist_lock);
431			}
432		}
433		goto out_set_pte;
434	}
435
436	/*
437	 * If it's a COW mapping, write protect it both
438	 * in the parent and the child
439	 */
440	if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
441		ptep_set_wrprotect(src_mm, addr, src_pte);
442		pte = *src_pte;
443	}
444
445	/*
446	 * If it's a shared mapping, mark it clean in
447	 * the child
448	 */
449	if (vm_flags & VM_SHARED)
450		pte = pte_mkclean(pte);
451	pte = pte_mkold(pte);
452
453	page = vm_normal_page(vma, addr, pte);
454	if (page) {
455		get_page(page);
456		page_dup_rmap(page);
457		rss[!!PageAnon(page)]++;
458	}
459
460out_set_pte:
461	set_pte_at(dst_mm, addr, dst_pte, pte);
462}
463
464static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
465		pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
466		unsigned long addr, unsigned long end)
467{
468	pte_t *src_pte, *dst_pte;
469	spinlock_t *src_ptl, *dst_ptl;
470	int progress = 0;
471	int rss[2];
472
473again:
474	rss[1] = rss[0] = 0;
475	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
476	if (!dst_pte)
477		return -ENOMEM;
478	src_pte = pte_offset_map_nested(src_pmd, addr);
479	src_ptl = pte_lockptr(src_mm, src_pmd);
480	spin_lock(src_ptl);
481
482	do {
483		/*
484		 * We are holding two locks at this point - either of them
485		 * could generate latencies in another task on another CPU.
486		 */
487		if (progress >= 32) {
488			progress = 0;
489			if (need_resched() ||
490			    need_lockbreak(src_ptl) ||
491			    need_lockbreak(dst_ptl))
492				break;
493		}
494		if (pte_none(*src_pte)) {
495			progress++;
496			continue;
497		}
498		copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
499		progress += 8;
500	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
501
502	spin_unlock(src_ptl);
503	pte_unmap_nested(src_pte - 1);
504	add_mm_rss(dst_mm, rss[0], rss[1]);
505	pte_unmap_unlock(dst_pte - 1, dst_ptl);
506	cond_resched();
507	if (addr != end)
508		goto again;
509	return 0;
510}
511
512static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
513		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
514		unsigned long addr, unsigned long end)
515{
516	pmd_t *src_pmd, *dst_pmd;
517	unsigned long next;
518
519	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
520	if (!dst_pmd)
521		return -ENOMEM;
522	src_pmd = pmd_offset(src_pud, addr);
523	do {
524		next = pmd_addr_end(addr, end);
525		if (pmd_none_or_clear_bad(src_pmd))
526			continue;
527		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
528						vma, addr, next))
529			return -ENOMEM;
530	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
531	return 0;
532}
533
534static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
535		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
536		unsigned long addr, unsigned long end)
537{
538	pud_t *src_pud, *dst_pud;
539	unsigned long next;
540
541	dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
542	if (!dst_pud)
543		return -ENOMEM;
544	src_pud = pud_offset(src_pgd, addr);
545	do {
546		next = pud_addr_end(addr, end);
547		if (pud_none_or_clear_bad(src_pud))
548			continue;
549		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
550						vma, addr, next))
551			return -ENOMEM;
552	} while (dst_pud++, src_pud++, addr = next, addr != end);
553	return 0;
554}
555
556int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
557		struct vm_area_struct *vma)
558{
559	pgd_t *src_pgd, *dst_pgd;
560	unsigned long next;
561	unsigned long addr = vma->vm_start;
562	unsigned long end = vma->vm_end;
563
564	/*
565	 * Don't copy ptes where a page fault will fill them correctly.
566	 * Fork becomes much lighter when there are big shared or private
567	 * readonly mappings. The tradeoff is that copy_page_range is more
568	 * efficient than faulting.
569	 */
570	if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP))) {
571		if (!vma->anon_vma)
572			return 0;
573	}
574
575	if (is_vm_hugetlb_page(vma))
576		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
577
578	dst_pgd = pgd_offset(dst_mm, addr);
579	src_pgd = pgd_offset(src_mm, addr);
580	do {
581		next = pgd_addr_end(addr, end);
582		if (pgd_none_or_clear_bad(src_pgd))
583			continue;
584		if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
585						vma, addr, next))
586			return -ENOMEM;
587	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
588	return 0;
589}
590
591static unsigned long zap_pte_range(struct mmu_gather *tlb,
592				struct vm_area_struct *vma, pmd_t *pmd,
593				unsigned long addr, unsigned long end,
594				long *zap_work, struct zap_details *details)
595{
596	struct mm_struct *mm = tlb->mm;
597	pte_t *pte;
598	spinlock_t *ptl;
599	int file_rss = 0;
600	int anon_rss = 0;
601
602	pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
603	do {
604		pte_t ptent = *pte;
605		if (pte_none(ptent)) {
606			(*zap_work)--;
607			continue;
608		}
609		if (pte_present(ptent)) {
610			struct page *page;
611
612			(*zap_work) -= PAGE_SIZE;
613
614			page = vm_normal_page(vma, addr, ptent);
615			if (unlikely(details) && page) {
616				/*
617				 * unmap_shared_mapping_pages() wants to
618				 * invalidate cache without truncating:
619				 * unmap shared but keep private pages.
620				 */
621				if (details->check_mapping &&
622				    details->check_mapping != page->mapping)
623					continue;
624				/*
625				 * Each page->index must be checked when
626				 * invalidating or truncating nonlinear.
627				 */
628				if (details->nonlinear_vma &&
629				    (page->index < details->first_index ||
630				     page->index > details->last_index))
631					continue;
632			}
633			ptent = ptep_get_and_clear_full(mm, addr, pte,
634							tlb->fullmm);
635			tlb_remove_tlb_entry(tlb, pte, addr);
636			if (unlikely(!page))
637				continue;
638			if (unlikely(details) && details->nonlinear_vma
639			    && linear_page_index(details->nonlinear_vma,
640						addr) != page->index)
641				set_pte_at(mm, addr, pte,
642					   pgoff_to_pte(page->index));
643			if (PageAnon(page))
644				anon_rss--;
645			else {
646				if (pte_dirty(ptent))
647					set_page_dirty(page);
648				if (pte_young(ptent))
649					mark_page_accessed(page);
650				file_rss--;
651			}
652			page_remove_rmap(page);
653			tlb_remove_page(tlb, page);
654			continue;
655		}
656		/*
657		 * If details->check_mapping, we leave swap entries;
658		 * if details->nonlinear_vma, we leave file entries.
659		 */
660		if (unlikely(details))
661			continue;
662		if (!pte_file(ptent))
663			free_swap_and_cache(pte_to_swp_entry(ptent));
664		pte_clear_full(mm, addr, pte, tlb->fullmm);
665	} while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
666
667	add_mm_rss(mm, file_rss, anon_rss);
668	pte_unmap_unlock(pte - 1, ptl);
669
670	return addr;
671}
672
673static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
674				struct vm_area_struct *vma, pud_t *pud,
675				unsigned long addr, unsigned long end,
676				long *zap_work, struct zap_details *details)
677{
678	pmd_t *pmd;
679	unsigned long next;
680
681	pmd = pmd_offset(pud, addr);
682	do {
683		next = pmd_addr_end(addr, end);
684		if (pmd_none_or_clear_bad(pmd)) {
685			(*zap_work)--;
686			continue;
687		}
688		next = zap_pte_range(tlb, vma, pmd, addr, next,
689						zap_work, details);
690	} while (pmd++, addr = next, (addr != end && *zap_work > 0));
691
692	return addr;
693}
694
695static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
696				struct vm_area_struct *vma, pgd_t *pgd,
697				unsigned long addr, unsigned long end,
698				long *zap_work, struct zap_details *details)
699{
700	pud_t *pud;
701	unsigned long next;
702
703	pud = pud_offset(pgd, addr);
704	do {
705		next = pud_addr_end(addr, end);
706		if (pud_none_or_clear_bad(pud)) {
707			(*zap_work)--;
708			continue;
709		}
710		next = zap_pmd_range(tlb, vma, pud, addr, next,
711						zap_work, details);
712	} while (pud++, addr = next, (addr != end && *zap_work > 0));
713
714	return addr;
715}
716
717static unsigned long unmap_page_range(struct mmu_gather *tlb,
718				struct vm_area_struct *vma,
719				unsigned long addr, unsigned long end,
720				long *zap_work, struct zap_details *details)
721{
722	pgd_t *pgd;
723	unsigned long next;
724
725	if (details && !details->check_mapping && !details->nonlinear_vma)
726		details = NULL;
727
728	BUG_ON(addr >= end);
729	tlb_start_vma(tlb, vma);
730	pgd = pgd_offset(vma->vm_mm, addr);
731	do {
732		next = pgd_addr_end(addr, end);
733		if (pgd_none_or_clear_bad(pgd)) {
734			(*zap_work)--;
735			continue;
736		}
737		next = zap_pud_range(tlb, vma, pgd, addr, next,
738						zap_work, details);
739	} while (pgd++, addr = next, (addr != end && *zap_work > 0));
740	tlb_end_vma(tlb, vma);
741
742	return addr;
743}
744
745#ifdef CONFIG_PREEMPT
746# define ZAP_BLOCK_SIZE	(8 * PAGE_SIZE)
747#else
748/* No preempt: go for improved straight-line efficiency */
749# define ZAP_BLOCK_SIZE	(1024 * PAGE_SIZE)
750#endif
751
752/**
753 * unmap_vmas - unmap a range of memory covered by a list of vma's
754 * @tlbp: address of the caller's struct mmu_gather
755 * @vma: the starting vma
756 * @start_addr: virtual address at which to start unmapping
757 * @end_addr: virtual address at which to end unmapping
758 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
759 * @details: details of nonlinear truncation or shared cache invalidation
760 *
761 * Returns the end address of the unmapping (restart addr if interrupted).
762 *
763 * Unmap all pages in the vma list.
764 *
765 * We aim to not hold locks for too long (for scheduling latency reasons).
766 * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
767 * return the ending mmu_gather to the caller.
768 *
769 * Only addresses between `start' and `end' will be unmapped.
770 *
771 * The VMA list must be sorted in ascending virtual address order.
772 *
773 * unmap_vmas() assumes that the caller will flush the whole unmapped address
774 * range after unmap_vmas() returns.  So the only responsibility here is to
775 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
776 * drops the lock and schedules.
777 */
778unsigned long unmap_vmas(struct mmu_gather **tlbp,
779		struct vm_area_struct *vma, unsigned long start_addr,
780		unsigned long end_addr, unsigned long *nr_accounted,
781		struct zap_details *details)
782{
783	long zap_work = ZAP_BLOCK_SIZE;
784	unsigned long tlb_start = 0;	/* For tlb_finish_mmu */
785	int tlb_start_valid = 0;
786	unsigned long start = start_addr;
787	spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
788	int fullmm = (*tlbp)->fullmm;
789
790	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
791		unsigned long end;
792
793		start = max(vma->vm_start, start_addr);
794		if (start >= vma->vm_end)
795			continue;
796		end = min(vma->vm_end, end_addr);
797		if (end <= vma->vm_start)
798			continue;
799
800		if (vma->vm_flags & VM_ACCOUNT)
801			*nr_accounted += (end - start) >> PAGE_SHIFT;
802
803		while (start != end) {
804			if (!tlb_start_valid) {
805				tlb_start = start;
806				tlb_start_valid = 1;
807			}
808
809			if (unlikely(is_vm_hugetlb_page(vma))) {
810				unmap_hugepage_range(vma, start, end);
811				zap_work -= (end - start) /
812						(HPAGE_SIZE / PAGE_SIZE);
813				start = end;
814			} else
815				start = unmap_page_range(*tlbp, vma,
816						start, end, &zap_work, details);
817
818			if (zap_work > 0) {
819				BUG_ON(start != end);
820				break;
821			}
822
823			tlb_finish_mmu(*tlbp, tlb_start, start);
824
825			if (need_resched() ||
826				(i_mmap_lock && need_lockbreak(i_mmap_lock))) {
827				if (i_mmap_lock) {
828					*tlbp = NULL;
829					goto out;
830				}
831				cond_resched();
832			}
833
834			*tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
835			tlb_start_valid = 0;
836			zap_work = ZAP_BLOCK_SIZE;
837		}
838	}
839out:
840	return start;	/* which is now the end (or restart) address */
841}
842
843/**
844 * zap_page_range - remove user pages in a given range
845 * @vma: vm_area_struct holding the applicable pages
846 * @address: starting address of pages to zap
847 * @size: number of bytes to zap
848 * @details: details of nonlinear truncation or shared cache invalidation
849 */
850unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
851		unsigned long size, struct zap_details *details)
852{
853	struct mm_struct *mm = vma->vm_mm;
854	struct mmu_gather *tlb;
855	unsigned long end = address + size;
856	unsigned long nr_accounted = 0;
857
858	lru_add_drain();
859	tlb = tlb_gather_mmu(mm, 0);
860	update_hiwater_rss(mm);
861	end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
862	if (tlb)
863		tlb_finish_mmu(tlb, address, end);
864	return end;
865}
866
867/*
868 * Do a quick page-table lookup for a single page.
869 */
870struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
871			unsigned int flags)
872{
873	pgd_t *pgd;
874	pud_t *pud;
875	pmd_t *pmd;
876	pte_t *ptep, pte;
877	spinlock_t *ptl;
878	struct page *page;
879	struct mm_struct *mm = vma->vm_mm;
880
881	page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
882	if (!IS_ERR(page)) {
883		BUG_ON(flags & FOLL_GET);
884		goto out;
885	}
886
887	page = NULL;
888	pgd = pgd_offset(mm, address);
889	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
890		goto no_page_table;
891
892	pud = pud_offset(pgd, address);
893	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
894		goto no_page_table;
895
896	pmd = pmd_offset(pud, address);
897	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
898		goto no_page_table;
899
900	if (pmd_huge(*pmd)) {
901		BUG_ON(flags & FOLL_GET);
902		page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
903		goto out;
904	}
905
906	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
907	if (!ptep)
908		goto out;
909
910	pte = *ptep;
911	if (!pte_present(pte))
912		goto unlock;
913	if ((flags & FOLL_WRITE) && !pte_write(pte))
914		goto unlock;
915	page = vm_normal_page(vma, address, pte);
916	if (unlikely(!page))
917		goto unlock;
918
919	if (flags & FOLL_GET)
920		get_page(page);
921	if (flags & FOLL_TOUCH) {
922		if ((flags & FOLL_WRITE) &&
923		    !pte_dirty(pte) && !PageDirty(page))
924			set_page_dirty(page);
925		mark_page_accessed(page);
926	}
927unlock:
928	pte_unmap_unlock(ptep, ptl);
929out:
930	return page;
931
932no_page_table:
933	/*
934	 * When core dumping an enormous anonymous area that nobody
935	 * has touched so far, we don't want to allocate page tables.
936	 */
937	if (flags & FOLL_ANON) {
938		page = ZERO_PAGE(address);
939		if (flags & FOLL_GET)
940			get_page(page);
941		BUG_ON(flags & FOLL_WRITE);
942	}
943	return page;
944}
945
946int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
947		unsigned long start, int len, int write, int force,
948		struct page **pages, struct vm_area_struct **vmas)
949{
950	int i;
951	unsigned int vm_flags;
952
953	/*
954	 * Require read or write permissions.
955	 * If 'force' is set, we only require the "MAY" flags.
956	 */
957	vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
958	vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
959	i = 0;
960
961	do {
962		struct vm_area_struct *vma;
963		unsigned int foll_flags;
964
965		vma = find_extend_vma(mm, start);
966		if (!vma && in_gate_area(tsk, start)) {
967			unsigned long pg = start & PAGE_MASK;
968			struct vm_area_struct *gate_vma = get_gate_vma(tsk);
969			pgd_t *pgd;
970			pud_t *pud;
971			pmd_t *pmd;
972			pte_t *pte;
973			if (write) /* user gate pages are read-only */
974				return i ? : -EFAULT;
975			if (pg > TASK_SIZE)
976				pgd = pgd_offset_k(pg);
977			else
978				pgd = pgd_offset_gate(mm, pg);
979			BUG_ON(pgd_none(*pgd));
980			pud = pud_offset(pgd, pg);
981			BUG_ON(pud_none(*pud));
982			pmd = pmd_offset(pud, pg);
983			if (pmd_none(*pmd))
984				return i ? : -EFAULT;
985			pte = pte_offset_map(pmd, pg);
986			if (pte_none(*pte)) {
987				pte_unmap(pte);
988				return i ? : -EFAULT;
989			}
990			if (pages) {
991				struct page *page = vm_normal_page(gate_vma, start, *pte);
992				pages[i] = page;
993				if (page)
994					get_page(page);
995			}
996			pte_unmap(pte);
997			if (vmas)
998				vmas[i] = gate_vma;
999			i++;
1000			start += PAGE_SIZE;
1001			len--;
1002			continue;
1003		}
1004
1005		if (!vma || (vma->vm_flags & VM_IO)
1006				|| !(vm_flags & vma->vm_flags))
1007			return i ? : -EFAULT;
1008
1009		if (is_vm_hugetlb_page(vma)) {
1010			i = follow_hugetlb_page(mm, vma, pages, vmas,
1011						&start, &len, i);
1012			continue;
1013		}
1014
1015		foll_flags = FOLL_TOUCH;
1016		if (pages)
1017			foll_flags |= FOLL_GET;
1018		if (!write && !(vma->vm_flags & VM_LOCKED) &&
1019		    (!vma->vm_ops || !vma->vm_ops->nopage))
1020			foll_flags |= FOLL_ANON;
1021
1022		do {
1023			struct page *page;
1024
1025			if (write)
1026				foll_flags |= FOLL_WRITE;
1027
1028			cond_resched();
1029			while (!(page = follow_page(vma, start, foll_flags))) {
1030				int ret;
1031				ret = __handle_mm_fault(mm, vma, start,
1032						foll_flags & FOLL_WRITE);
1033				/*
1034				 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1035				 * broken COW when necessary, even if maybe_mkwrite
1036				 * decided not to set pte_write. We can thus safely do
1037				 * subsequent page lookups as if they were reads.
1038				 */
1039				if (ret & VM_FAULT_WRITE)
1040					foll_flags &= ~FOLL_WRITE;
1041
1042				switch (ret & ~VM_FAULT_WRITE) {
1043				case VM_FAULT_MINOR:
1044					tsk->min_flt++;
1045					break;
1046				case VM_FAULT_MAJOR:
1047					tsk->maj_flt++;
1048					break;
1049				case VM_FAULT_SIGBUS:
1050					return i ? i : -EFAULT;
1051				case VM_FAULT_OOM:
1052					return i ? i : -ENOMEM;
1053				default:
1054					BUG();
1055				}
1056			}
1057			if (pages) {
1058				pages[i] = page;
1059				flush_dcache_page(page);
1060			}
1061			if (vmas)
1062				vmas[i] = vma;
1063			i++;
1064			start += PAGE_SIZE;
1065			len--;
1066		} while (len && start < vma->vm_end);
1067	} while (len);
1068	return i;
1069}
1070EXPORT_SYMBOL(get_user_pages);
1071
1072static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1073			unsigned long addr, unsigned long end, pgprot_t prot)
1074{
1075	pte_t *pte;
1076	spinlock_t *ptl;
1077
1078	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1079	if (!pte)
1080		return -ENOMEM;
1081	do {
1082		struct page *page = ZERO_PAGE(addr);
1083		pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1084		page_cache_get(page);
1085		page_add_file_rmap(page);
1086		inc_mm_counter(mm, file_rss);
1087		BUG_ON(!pte_none(*pte));
1088		set_pte_at(mm, addr, pte, zero_pte);
1089	} while (pte++, addr += PAGE_SIZE, addr != end);
1090	pte_unmap_unlock(pte - 1, ptl);
1091	return 0;
1092}
1093
1094static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1095			unsigned long addr, unsigned long end, pgprot_t prot)
1096{
1097	pmd_t *pmd;
1098	unsigned long next;
1099
1100	pmd = pmd_alloc(mm, pud, addr);
1101	if (!pmd)
1102		return -ENOMEM;
1103	do {
1104		next = pmd_addr_end(addr, end);
1105		if (zeromap_pte_range(mm, pmd, addr, next, prot))
1106			return -ENOMEM;
1107	} while (pmd++, addr = next, addr != end);
1108	return 0;
1109}
1110
1111static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1112			unsigned long addr, unsigned long end, pgprot_t prot)
1113{
1114	pud_t *pud;
1115	unsigned long next;
1116
1117	pud = pud_alloc(mm, pgd, addr);
1118	if (!pud)
1119		return -ENOMEM;
1120	do {
1121		next = pud_addr_end(addr, end);
1122		if (zeromap_pmd_range(mm, pud, addr, next, prot))
1123			return -ENOMEM;
1124	} while (pud++, addr = next, addr != end);
1125	return 0;
1126}
1127
1128int zeromap_page_range(struct vm_area_struct *vma,
1129			unsigned long addr, unsigned long size, pgprot_t prot)
1130{
1131	pgd_t *pgd;
1132	unsigned long next;
1133	unsigned long end = addr + size;
1134	struct mm_struct *mm = vma->vm_mm;
1135	int err;
1136
1137	BUG_ON(addr >= end);
1138	pgd = pgd_offset(mm, addr);
1139	flush_cache_range(vma, addr, end);
1140	do {
1141		next = pgd_addr_end(addr, end);
1142		err = zeromap_pud_range(mm, pgd, addr, next, prot);
1143		if (err)
1144			break;
1145	} while (pgd++, addr = next, addr != end);
1146	return err;
1147}
1148
1149pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1150{
1151	pgd_t * pgd = pgd_offset(mm, addr);
1152	pud_t * pud = pud_alloc(mm, pgd, addr);
1153	if (pud) {
1154		pmd_t * pmd = pmd_alloc(mm, pgd, addr);
1155		if (pmd)
1156			return pte_alloc_map_lock(mm, pmd, addr, ptl);
1157	}
1158	return NULL;
1159}
1160
1161/*
1162 * This is the old fallback for page remapping.
1163 *
1164 * For historical reasons, it only allows reserved pages. Only
1165 * old drivers should use this, and they needed to mark their
1166 * pages reserved for the old functions anyway.
1167 */
1168static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1169{
1170	int retval;
1171	pte_t *pte;
1172	spinlock_t *ptl;
1173
1174	retval = -EINVAL;
1175	if (PageAnon(page) || !PageReserved(page))
1176		goto out;
1177	retval = -ENOMEM;
1178	flush_dcache_page(page);
1179	pte = get_locked_pte(mm, addr, &ptl);
1180	if (!pte)
1181		goto out;
1182	retval = -EBUSY;
1183	if (!pte_none(*pte))
1184		goto out_unlock;
1185
1186	/* Ok, finally just insert the thing.. */
1187	get_page(page);
1188	inc_mm_counter(mm, file_rss);
1189	page_add_file_rmap(page);
1190	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1191
1192	retval = 0;
1193out_unlock:
1194	pte_unmap_unlock(pte, ptl);
1195out:
1196	return retval;
1197}
1198
1199/*
1200 * Somebody does a pfn remapping that doesn't actually work as a vma.
1201 *
1202 * Do it as individual pages instead, and warn about it. It's bad form,
1203 * and very inefficient.
1204 */
1205static int incomplete_pfn_remap(struct vm_area_struct *vma,
1206		unsigned long start, unsigned long end,
1207		unsigned long pfn, pgprot_t prot)
1208{
1209	static int warn = 10;
1210	struct page *page;
1211	int retval;
1212
1213	if (!(vma->vm_flags & VM_INCOMPLETE)) {
1214		if (warn) {
1215			warn--;
1216			printk("%s does an incomplete pfn remapping", current->comm);
1217			dump_stack();
1218		}
1219	}
1220	vma->vm_flags |= VM_INCOMPLETE | VM_IO | VM_RESERVED;
1221
1222	if (start < vma->vm_start || end > vma->vm_end)
1223		return -EINVAL;
1224
1225	if (!pfn_valid(pfn))
1226		return -EINVAL;
1227
1228	retval = 0;
1229	page = pfn_to_page(pfn);
1230	while (start < end) {
1231		retval = insert_page(vma->vm_mm, start, page, prot);
1232		if (retval < 0)
1233			break;
1234		start += PAGE_SIZE;
1235		page++;
1236	}
1237	return retval;
1238}
1239
1240/*
1241 * maps a range of physical memory into the requested pages. the old
1242 * mappings are removed. any references to nonexistent pages results
1243 * in null mappings (currently treated as "copy-on-access")
1244 */
1245static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1246			unsigned long addr, unsigned long end,
1247			unsigned long pfn, pgprot_t prot)
1248{
1249	pte_t *pte;
1250	spinlock_t *ptl;
1251
1252	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1253	if (!pte)
1254		return -ENOMEM;
1255	do {
1256		BUG_ON(!pte_none(*pte));
1257		set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1258		pfn++;
1259	} while (pte++, addr += PAGE_SIZE, addr != end);
1260	pte_unmap_unlock(pte - 1, ptl);
1261	return 0;
1262}
1263
1264static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1265			unsigned long addr, unsigned long end,
1266			unsigned long pfn, pgprot_t prot)
1267{
1268	pmd_t *pmd;
1269	unsigned long next;
1270
1271	pfn -= addr >> PAGE_SHIFT;
1272	pmd = pmd_alloc(mm, pud, addr);
1273	if (!pmd)
1274		return -ENOMEM;
1275	do {
1276		next = pmd_addr_end(addr, end);
1277		if (remap_pte_range(mm, pmd, addr, next,
1278				pfn + (addr >> PAGE_SHIFT), prot))
1279			return -ENOMEM;
1280	} while (pmd++, addr = next, addr != end);
1281	return 0;
1282}
1283
1284static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1285			unsigned long addr, unsigned long end,
1286			unsigned long pfn, pgprot_t prot)
1287{
1288	pud_t *pud;
1289	unsigned long next;
1290
1291	pfn -= addr >> PAGE_SHIFT;
1292	pud = pud_alloc(mm, pgd, addr);
1293	if (!pud)
1294		return -ENOMEM;
1295	do {
1296		next = pud_addr_end(addr, end);
1297		if (remap_pmd_range(mm, pud, addr, next,
1298				pfn + (addr >> PAGE_SHIFT), prot))
1299			return -ENOMEM;
1300	} while (pud++, addr = next, addr != end);
1301	return 0;
1302}
1303
1304/*  Note: this is only safe if the mm semaphore is held when called. */
1305int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1306		    unsigned long pfn, unsigned long size, pgprot_t prot)
1307{
1308	pgd_t *pgd;
1309	unsigned long next;
1310	unsigned long end = addr + PAGE_ALIGN(size);
1311	struct mm_struct *mm = vma->vm_mm;
1312	int err;
1313
1314	if (addr != vma->vm_start || end != vma->vm_end)
1315		return incomplete_pfn_remap(vma, addr, end, pfn, prot);
1316
1317	/*
1318	 * Physically remapped pages are special. Tell the
1319	 * rest of the world about it:
1320	 *   VM_IO tells people not to look at these pages
1321	 *	(accesses can have side effects).
1322	 *   VM_RESERVED is specified all over the place, because
1323	 *	in 2.4 it kept swapout's vma scan off this vma; but
1324	 *	in 2.6 the LRU scan won't even find its pages, so this
1325	 *	flag means no more than count its pages in reserved_vm,
1326	 * 	and omit it from core dump, even when VM_IO turned off.
1327	 *   VM_PFNMAP tells the core MM that the base pages are just
1328	 *	raw PFN mappings, and do not have a "struct page" associated
1329	 *	with them.
1330	 */
1331	vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1332	vma->vm_pgoff = pfn;
1333
1334	BUG_ON(addr >= end);
1335	pfn -= addr >> PAGE_SHIFT;
1336	pgd = pgd_offset(mm, addr);
1337	flush_cache_range(vma, addr, end);
1338	do {
1339		next = pgd_addr_end(addr, end);
1340		err = remap_pud_range(mm, pgd, addr, next,
1341				pfn + (addr >> PAGE_SHIFT), prot);
1342		if (err)
1343			break;
1344	} while (pgd++, addr = next, addr != end);
1345	return err;
1346}
1347EXPORT_SYMBOL(remap_pfn_range);
1348
1349/*
1350 * handle_pte_fault chooses page fault handler according to an entry
1351 * which was read non-atomically.  Before making any commitment, on
1352 * those architectures or configurations (e.g. i386 with PAE) which
1353 * might give a mix of unmatched parts, do_swap_page and do_file_page
1354 * must check under lock before unmapping the pte and proceeding
1355 * (but do_wp_page is only called after already making such a check;
1356 * and do_anonymous_page and do_no_page can safely check later on).
1357 */
1358static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1359				pte_t *page_table, pte_t orig_pte)
1360{
1361	int same = 1;
1362#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1363	if (sizeof(pte_t) > sizeof(unsigned long)) {
1364		spinlock_t *ptl = pte_lockptr(mm, pmd);
1365		spin_lock(ptl);
1366		same = pte_same(*page_table, orig_pte);
1367		spin_unlock(ptl);
1368	}
1369#endif
1370	pte_unmap(page_table);
1371	return same;
1372}
1373
1374/*
1375 * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1376 * servicing faults for write access.  In the normal case, do always want
1377 * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1378 * that do not have writing enabled, when used by access_process_vm.
1379 */
1380static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1381{
1382	if (likely(vma->vm_flags & VM_WRITE))
1383		pte = pte_mkwrite(pte);
1384	return pte;
1385}
1386
1387static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1388{
1389	/*
1390	 * If the source page was a PFN mapping, we don't have
1391	 * a "struct page" for it. We do a best-effort copy by
1392	 * just copying from the original user address. If that
1393	 * fails, we just zero-fill it. Live with it.
1394	 */
1395	if (unlikely(!src)) {
1396		void *kaddr = kmap_atomic(dst, KM_USER0);
1397		void __user *uaddr = (void __user *)(va & PAGE_MASK);
1398
1399		/*
1400		 * This really shouldn't fail, because the page is there
1401		 * in the page tables. But it might just be unreadable,
1402		 * in which case we just give up and fill the result with
1403		 * zeroes.
1404		 */
1405		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1406			memset(kaddr, 0, PAGE_SIZE);
1407		kunmap_atomic(kaddr, KM_USER0);
1408		return;
1409
1410	}
1411	copy_user_highpage(dst, src, va);
1412}
1413
1414/*
1415 * This routine handles present pages, when users try to write
1416 * to a shared page. It is done by copying the page to a new address
1417 * and decrementing the shared-page counter for the old page.
1418 *
1419 * Note that this routine assumes that the protection checks have been
1420 * done by the caller (the low-level page fault routine in most cases).
1421 * Thus we can safely just mark it writable once we've done any necessary
1422 * COW.
1423 *
1424 * We also mark the page dirty at this point even though the page will
1425 * change only once the write actually happens. This avoids a few races,
1426 * and potentially makes it more efficient.
1427 *
1428 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1429 * but allow concurrent faults), with pte both mapped and locked.
1430 * We return with mmap_sem still held, but pte unmapped and unlocked.
1431 */
1432static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1433		unsigned long address, pte_t *page_table, pmd_t *pmd,
1434		spinlock_t *ptl, pte_t orig_pte)
1435{
1436	struct page *old_page, *src_page, *new_page;
1437	pte_t entry;
1438	int ret = VM_FAULT_MINOR;
1439
1440	old_page = vm_normal_page(vma, address, orig_pte);
1441	src_page = old_page;
1442	if (!old_page)
1443		goto gotten;
1444
1445	if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1446		int reuse = can_share_swap_page(old_page);
1447		unlock_page(old_page);
1448		if (reuse) {
1449			flush_cache_page(vma, address, pte_pfn(orig_pte));
1450			entry = pte_mkyoung(orig_pte);
1451			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1452			ptep_set_access_flags(vma, address, page_table, entry, 1);
1453			update_mmu_cache(vma, address, entry);
1454			lazy_mmu_prot_update(entry);
1455			ret |= VM_FAULT_WRITE;
1456			goto unlock;
1457		}
1458	}
1459
1460	/*
1461	 * Ok, we need to copy. Oh, well..
1462	 */
1463	page_cache_get(old_page);
1464gotten:
1465	pte_unmap_unlock(page_table, ptl);
1466
1467	if (unlikely(anon_vma_prepare(vma)))
1468		goto oom;
1469	if (src_page == ZERO_PAGE(address)) {
1470		new_page = alloc_zeroed_user_highpage(vma, address);
1471		if (!new_page)
1472			goto oom;
1473	} else {
1474		new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1475		if (!new_page)
1476			goto oom;
1477		cow_user_page(new_page, src_page, address);
1478	}
1479
1480	/*
1481	 * Re-check the pte - we dropped the lock
1482	 */
1483	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1484	if (likely(pte_same(*page_table, orig_pte))) {
1485		if (old_page) {
1486			page_remove_rmap(old_page);
1487			if (!PageAnon(old_page)) {
1488				dec_mm_counter(mm, file_rss);
1489				inc_mm_counter(mm, anon_rss);
1490			}
1491		} else
1492			inc_mm_counter(mm, anon_rss);
1493		flush_cache_page(vma, address, pte_pfn(orig_pte));
1494		entry = mk_pte(new_page, vma->vm_page_prot);
1495		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1496		ptep_establish(vma, address, page_table, entry);
1497		update_mmu_cache(vma, address, entry);
1498		lazy_mmu_prot_update(entry);
1499		lru_cache_add_active(new_page);
1500		page_add_anon_rmap(new_page, vma, address);
1501
1502		/* Free the old page.. */
1503		new_page = old_page;
1504		ret |= VM_FAULT_WRITE;
1505	}
1506	if (new_page)
1507		page_cache_release(new_page);
1508	if (old_page)
1509		page_cache_release(old_page);
1510unlock:
1511	pte_unmap_unlock(page_table, ptl);
1512	return ret;
1513oom:
1514	if (old_page)
1515		page_cache_release(old_page);
1516	return VM_FAULT_OOM;
1517}
1518
1519/*
1520 * Helper functions for unmap_mapping_range().
1521 *
1522 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1523 *
1524 * We have to restart searching the prio_tree whenever we drop the lock,
1525 * since the iterator is only valid while the lock is held, and anyway
1526 * a later vma might be split and reinserted earlier while lock dropped.
1527 *
1528 * The list of nonlinear vmas could be handled more efficiently, using
1529 * a placeholder, but handle it in the same way until a need is shown.
1530 * It is important to search the prio_tree before nonlinear list: a vma
1531 * may become nonlinear and be shifted from prio_tree to nonlinear list
1532 * while the lock is dropped; but never shifted from list to prio_tree.
1533 *
1534 * In order to make forward progress despite restarting the search,
1535 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1536 * quickly skip it next time around.  Since the prio_tree search only
1537 * shows us those vmas affected by unmapping the range in question, we
1538 * can't efficiently keep all vmas in step with mapping->truncate_count:
1539 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1540 * mapping->truncate_count and vma->vm_truncate_count are protected by
1541 * i_mmap_lock.
1542 *
1543 * In order to make forward progress despite repeatedly restarting some
1544 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1545 * and restart from that address when we reach that vma again.  It might
1546 * have been split or merged, shrunk or extended, but never shifted: so
1547 * restart_addr remains valid so long as it remains in the vma's range.
1548 * unmap_mapping_range forces truncate_count to leap over page-aligned
1549 * values so we can save vma's restart_addr in its truncate_count field.
1550 */
1551#define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1552
1553static void reset_vma_truncate_counts(struct address_space *mapping)
1554{
1555	struct vm_area_struct *vma;
1556	struct prio_tree_iter iter;
1557
1558	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1559		vma->vm_truncate_count = 0;
1560	list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1561		vma->vm_truncate_count = 0;
1562}
1563
1564static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1565		unsigned long start_addr, unsigned long end_addr,
1566		struct zap_details *details)
1567{
1568	unsigned long restart_addr;
1569	int need_break;
1570
1571again:
1572	restart_addr = vma->vm_truncate_count;
1573	if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1574		start_addr = restart_addr;
1575		if (start_addr >= end_addr) {
1576			/* Top of vma has been split off since last time */
1577			vma->vm_truncate_count = details->truncate_count;
1578			return 0;
1579		}
1580	}
1581
1582	restart_addr = zap_page_range(vma, start_addr,
1583					end_addr - start_addr, details);
1584	need_break = need_resched() ||
1585			need_lockbreak(details->i_mmap_lock);
1586
1587	if (restart_addr >= end_addr) {
1588		/* We have now completed this vma: mark it so */
1589		vma->vm_truncate_count = details->truncate_count;
1590		if (!need_break)
1591			return 0;
1592	} else {
1593		/* Note restart_addr in vma's truncate_count field */
1594		vma->vm_truncate_count = restart_addr;
1595		if (!need_break)
1596			goto again;
1597	}
1598
1599	spin_unlock(details->i_mmap_lock);
1600	cond_resched();
1601	spin_lock(details->i_mmap_lock);
1602	return -EINTR;
1603}
1604
1605static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1606					    struct zap_details *details)
1607{
1608	struct vm_area_struct *vma;
1609	struct prio_tree_iter iter;
1610	pgoff_t vba, vea, zba, zea;
1611
1612restart:
1613	vma_prio_tree_foreach(vma, &iter, root,
1614			details->first_index, details->last_index) {
1615		/* Skip quickly over those we have already dealt with */
1616		if (vma->vm_truncate_count == details->truncate_count)
1617			continue;
1618
1619		vba = vma->vm_pgoff;
1620		vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1621		/* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1622		zba = details->first_index;
1623		if (zba < vba)
1624			zba = vba;
1625		zea = details->last_index;
1626		if (zea > vea)
1627			zea = vea;
1628
1629		if (unmap_mapping_range_vma(vma,
1630			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1631			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1632				details) < 0)
1633			goto restart;
1634	}
1635}
1636
1637static inline void unmap_mapping_range_list(struct list_head *head,
1638					    struct zap_details *details)
1639{
1640	struct vm_area_struct *vma;
1641
1642	/*
1643	 * In nonlinear VMAs there is no correspondence between virtual address
1644	 * offset and file offset.  So we must perform an exhaustive search
1645	 * across *all* the pages in each nonlinear VMA, not just the pages
1646	 * whose virtual address lies outside the file truncation point.
1647	 */
1648restart:
1649	list_for_each_entry(vma, head, shared.vm_set.list) {
1650		/* Skip quickly over those we have already dealt with */
1651		if (vma->vm_truncate_count == details->truncate_count)
1652			continue;
1653		details->nonlinear_vma = vma;
1654		if (unmap_mapping_range_vma(vma, vma->vm_start,
1655					vma->vm_end, details) < 0)
1656			goto restart;
1657	}
1658}
1659
1660/**
1661 * unmap_mapping_range - unmap the portion of all mmaps
1662 * in the specified address_space corresponding to the specified
1663 * page range in the underlying file.
1664 * @mapping: the address space containing mmaps to be unmapped.
1665 * @holebegin: byte in first page to unmap, relative to the start of
1666 * the underlying file.  This will be rounded down to a PAGE_SIZE
1667 * boundary.  Note that this is different from vmtruncate(), which
1668 * must keep the partial page.  In contrast, we must get rid of
1669 * partial pages.
1670 * @holelen: size of prospective hole in bytes.  This will be rounded
1671 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1672 * end of the file.
1673 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1674 * but 0 when invalidating pagecache, don't throw away private data.
1675 */
1676void unmap_mapping_range(struct address_space *mapping,
1677		loff_t const holebegin, loff_t const holelen, int even_cows)
1678{
1679	struct zap_details details;
1680	pgoff_t hba = holebegin >> PAGE_SHIFT;
1681	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1682
1683	/* Check for overflow. */
1684	if (sizeof(holelen) > sizeof(hlen)) {
1685		long long holeend =
1686			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1687		if (holeend & ~(long long)ULONG_MAX)
1688			hlen = ULONG_MAX - hba + 1;
1689	}
1690
1691	details.check_mapping = even_cows? NULL: mapping;
1692	details.nonlinear_vma = NULL;
1693	details.first_index = hba;
1694	details.last_index = hba + hlen - 1;
1695	if (details.last_index < details.first_index)
1696		details.last_index = ULONG_MAX;
1697	details.i_mmap_lock = &mapping->i_mmap_lock;
1698
1699	spin_lock(&mapping->i_mmap_lock);
1700
1701	/* serialize i_size write against truncate_count write */
1702	smp_wmb();
1703	/* Protect against page faults, and endless unmapping loops */
1704	mapping->truncate_count++;
1705	/*
1706	 * For archs where spin_lock has inclusive semantics like ia64
1707	 * this smp_mb() will prevent to read pagetable contents
1708	 * before the truncate_count increment is visible to
1709	 * other cpus.
1710	 */
1711	smp_mb();
1712	if (unlikely(is_restart_addr(mapping->truncate_count))) {
1713		if (mapping->truncate_count == 0)
1714			reset_vma_truncate_counts(mapping);
1715		mapping->truncate_count++;
1716	}
1717	details.truncate_count = mapping->truncate_count;
1718
1719	if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1720		unmap_mapping_range_tree(&mapping->i_mmap, &details);
1721	if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1722		unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1723	spin_unlock(&mapping->i_mmap_lock);
1724}
1725EXPORT_SYMBOL(unmap_mapping_range);
1726
1727/*
1728 * Handle all mappings that got truncated by a "truncate()"
1729 * system call.
1730 *
1731 * NOTE! We have to be ready to update the memory sharing
1732 * between the file and the memory map for a potential last
1733 * incomplete page.  Ugly, but necessary.
1734 */
1735int vmtruncate(struct inode * inode, loff_t offset)
1736{
1737	struct address_space *mapping = inode->i_mapping;
1738	unsigned long limit;
1739
1740	if (inode->i_size < offset)
1741		goto do_expand;
1742	/*
1743	 * truncation of in-use swapfiles is disallowed - it would cause
1744	 * subsequent swapout to scribble on the now-freed blocks.
1745	 */
1746	if (IS_SWAPFILE(inode))
1747		goto out_busy;
1748	i_size_write(inode, offset);
1749	unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1750	truncate_inode_pages(mapping, offset);
1751	goto out_truncate;
1752
1753do_expand:
1754	limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1755	if (limit != RLIM_INFINITY && offset > limit)
1756		goto out_sig;
1757	if (offset > inode->i_sb->s_maxbytes)
1758		goto out_big;
1759	i_size_write(inode, offset);
1760
1761out_truncate:
1762	if (inode->i_op && inode->i_op->truncate)
1763		inode->i_op->truncate(inode);
1764	return 0;
1765out_sig:
1766	send_sig(SIGXFSZ, current, 0);
1767out_big:
1768	return -EFBIG;
1769out_busy:
1770	return -ETXTBSY;
1771}
1772
1773EXPORT_SYMBOL(vmtruncate);
1774
1775/*
1776 * Primitive swap readahead code. We simply read an aligned block of
1777 * (1 << page_cluster) entries in the swap area. This method is chosen
1778 * because it doesn't cost us any seek time.  We also make sure to queue
1779 * the 'original' request together with the readahead ones...
1780 *
1781 * This has been extended to use the NUMA policies from the mm triggering
1782 * the readahead.
1783 *
1784 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1785 */
1786void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1787{
1788#ifdef CONFIG_NUMA
1789	struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1790#endif
1791	int i, num;
1792	struct page *new_page;
1793	unsigned long offset;
1794
1795	/*
1796	 * Get the number of handles we should do readahead io to.
1797	 */
1798	num = valid_swaphandles(entry, &offset);
1799	for (i = 0; i < num; offset++, i++) {
1800		/* Ok, do the async read-ahead now */
1801		new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1802							   offset), vma, addr);
1803		if (!new_page)
1804			break;
1805		page_cache_release(new_page);
1806#ifdef CONFIG_NUMA
1807		/*
1808		 * Find the next applicable VMA for the NUMA policy.
1809		 */
1810		addr += PAGE_SIZE;
1811		if (addr == 0)
1812			vma = NULL;
1813		if (vma) {
1814			if (addr >= vma->vm_end) {
1815				vma = next_vma;
1816				next_vma = vma ? vma->vm_next : NULL;
1817			}
1818			if (vma && addr < vma->vm_start)
1819				vma = NULL;
1820		} else {
1821			if (next_vma && addr >= next_vma->vm_start) {
1822				vma = next_vma;
1823				next_vma = vma->vm_next;
1824			}
1825		}
1826#endif
1827	}
1828	lru_add_drain();	/* Push any new pages onto the LRU now */
1829}
1830
1831/*
1832 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1833 * but allow concurrent faults), and pte mapped but not yet locked.
1834 * We return with mmap_sem still held, but pte unmapped and unlocked.
1835 */
1836static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1837		unsigned long address, pte_t *page_table, pmd_t *pmd,
1838		int write_access, pte_t orig_pte)
1839{
1840	spinlock_t *ptl;
1841	struct page *page;
1842	swp_entry_t entry;
1843	pte_t pte;
1844	int ret = VM_FAULT_MINOR;
1845
1846	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1847		goto out;
1848
1849	entry = pte_to_swp_entry(orig_pte);
1850	page = lookup_swap_cache(entry);
1851	if (!page) {
1852 		swapin_readahead(entry, address, vma);
1853 		page = read_swap_cache_async(entry, vma, address);
1854		if (!page) {
1855			/*
1856			 * Back out if somebody else faulted in this pte
1857			 * while we released the pte lock.
1858			 */
1859			page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1860			if (likely(pte_same(*page_table, orig_pte)))
1861				ret = VM_FAULT_OOM;
1862			goto unlock;
1863		}
1864
1865		/* Had to read the page from swap area: Major fault */
1866		ret = VM_FAULT_MAJOR;
1867		inc_page_state(pgmajfault);
1868		grab_swap_token();
1869	}
1870
1871	mark_page_accessed(page);
1872	lock_page(page);
1873
1874	/*
1875	 * Back out if somebody else already faulted in this pte.
1876	 */
1877	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1878	if (unlikely(!pte_same(*page_table, orig_pte)))
1879		goto out_nomap;
1880
1881	if (unlikely(!PageUptodate(page))) {
1882		ret = VM_FAULT_SIGBUS;
1883		goto out_nomap;
1884	}
1885
1886	/* The page isn't present yet, go ahead with the fault. */
1887
1888	inc_mm_counter(mm, anon_rss);
1889	pte = mk_pte(page, vma->vm_page_prot);
1890	if (write_access && can_share_swap_page(page)) {
1891		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1892		write_access = 0;
1893	}
1894
1895	flush_icache_page(vma, page);
1896	set_pte_at(mm, address, page_table, pte);
1897	page_add_anon_rmap(page, vma, address);
1898
1899	swap_free(entry);
1900	if (vm_swap_full())
1901		remove_exclusive_swap_page(page);
1902	unlock_page(page);
1903
1904	if (write_access) {
1905		if (do_wp_page(mm, vma, address,
1906				page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1907			ret = VM_FAULT_OOM;
1908		goto out;
1909	}
1910
1911	/* No need to invalidate - it was non-present before */
1912	update_mmu_cache(vma, address, pte);
1913	lazy_mmu_prot_update(pte);
1914unlock:
1915	pte_unmap_unlock(page_table, ptl);
1916out:
1917	return ret;
1918out_nomap:
1919	pte_unmap_unlock(page_table, ptl);
1920	unlock_page(page);
1921	page_cache_release(page);
1922	return ret;
1923}
1924
1925/*
1926 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1927 * but allow concurrent faults), and pte mapped but not yet locked.
1928 * We return with mmap_sem still held, but pte unmapped and unlocked.
1929 */
1930static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1931		unsigned long address, pte_t *page_table, pmd_t *pmd,
1932		int write_access)
1933{
1934	struct page *page;
1935	spinlock_t *ptl;
1936	pte_t entry;
1937
1938	if (write_access) {
1939		/* Allocate our own private page. */
1940		pte_unmap(page_table);
1941
1942		if (unlikely(anon_vma_prepare(vma)))
1943			goto oom;
1944		page = alloc_zeroed_user_highpage(vma, address);
1945		if (!page)
1946			goto oom;
1947
1948		entry = mk_pte(page, vma->vm_page_prot);
1949		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1950
1951		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1952		if (!pte_none(*page_table))
1953			goto release;
1954		inc_mm_counter(mm, anon_rss);
1955		lru_cache_add_active(page);
1956		SetPageReferenced(page);
1957		page_add_anon_rmap(page, vma, address);
1958	} else {
1959		/* Map the ZERO_PAGE - vm_page_prot is readonly */
1960		page = ZERO_PAGE(address);
1961		page_cache_get(page);
1962		entry = mk_pte(page, vma->vm_page_prot);
1963
1964		ptl = pte_lockptr(mm, pmd);
1965		spin_lock(ptl);
1966		if (!pte_none(*page_table))
1967			goto release;
1968		inc_mm_counter(mm, file_rss);
1969		page_add_file_rmap(page);
1970	}
1971
1972	set_pte_at(mm, address, page_table, entry);
1973
1974	/* No need to invalidate - it was non-present before */
1975	update_mmu_cache(vma, address, entry);
1976	lazy_mmu_prot_update(entry);
1977unlock:
1978	pte_unmap_unlock(page_table, ptl);
1979	return VM_FAULT_MINOR;
1980release:
1981	page_cache_release(page);
1982	goto unlock;
1983oom:
1984	return VM_FAULT_OOM;
1985}
1986
1987/*
1988 * do_no_page() tries to create a new page mapping. It aggressively
1989 * tries to share with existing pages, but makes a separate copy if
1990 * the "write_access" parameter is true in order to avoid the next
1991 * page fault.
1992 *
1993 * As this is called only for pages that do not currently exist, we
1994 * do not need to flush old virtual caches or the TLB.
1995 *
1996 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1997 * but allow concurrent faults), and pte mapped but not yet locked.
1998 * We return with mmap_sem still held, but pte unmapped and unlocked.
1999 */
2000static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2001		unsigned long address, pte_t *page_table, pmd_t *pmd,
2002		int write_access)
2003{
2004	spinlock_t *ptl;
2005	struct page *new_page;
2006	struct address_space *mapping = NULL;
2007	pte_t entry;
2008	unsigned int sequence = 0;
2009	int ret = VM_FAULT_MINOR;
2010	int anon = 0;
2011
2012	pte_unmap(page_table);
2013	if (vma->vm_file) {
2014		mapping = vma->vm_file->f_mapping;
2015		sequence = mapping->truncate_count;
2016		smp_rmb(); /* serializes i_size against truncate_count */
2017	}
2018retry:
2019	new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2020	/*
2021	 * No smp_rmb is needed here as long as there's a full
2022	 * spin_lock/unlock sequence inside the ->nopage callback
2023	 * (for the pagecache lookup) that acts as an implicit
2024	 * smp_mb() and prevents the i_size read to happen
2025	 * after the next truncate_count read.
2026	 */
2027
2028	/* no page was available -- either SIGBUS or OOM */
2029	if (new_page == NOPAGE_SIGBUS)
2030		return VM_FAULT_SIGBUS;
2031	if (new_page == NOPAGE_OOM)
2032		return VM_FAULT_OOM;
2033
2034	/*
2035	 * Should we do an early C-O-W break?
2036	 */
2037	if (write_access && !(vma->vm_flags & VM_SHARED)) {
2038		struct page *page;
2039
2040		if (unlikely(anon_vma_prepare(vma)))
2041			goto oom;
2042		page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2043		if (!page)
2044			goto oom;
2045		cow_user_page(page, new_page, address);
2046		page_cache_release(new_page);
2047		new_page = page;
2048		anon = 1;
2049	}
2050
2051	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2052	/*
2053	 * For a file-backed vma, someone could have truncated or otherwise
2054	 * invalidated this page.  If unmap_mapping_range got called,
2055	 * retry getting the page.
2056	 */
2057	if (mapping && unlikely(sequence != mapping->truncate_count)) {
2058		pte_unmap_unlock(page_table, ptl);
2059		page_cache_release(new_page);
2060		cond_resched();
2061		sequence = mapping->truncate_count;
2062		smp_rmb();
2063		goto retry;
2064	}
2065
2066	/*
2067	 * This silly early PAGE_DIRTY setting removes a race
2068	 * due to the bad i386 page protection. But it's valid
2069	 * for other architectures too.
2070	 *
2071	 * Note that if write_access is true, we either now have
2072	 * an exclusive copy of the page, or this is a shared mapping,
2073	 * so we can make it writable and dirty to avoid having to
2074	 * handle that later.
2075	 */
2076	/* Only go through if we didn't race with anybody else... */
2077	if (pte_none(*page_table)) {
2078		flush_icache_page(vma, new_page);
2079		entry = mk_pte(new_page, vma->vm_page_prot);
2080		if (write_access)
2081			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2082		set_pte_at(mm, address, page_table, entry);
2083		if (anon) {
2084			inc_mm_counter(mm, anon_rss);
2085			lru_cache_add_active(new_page);
2086			page_add_anon_rmap(new_page, vma, address);
2087		} else {
2088			inc_mm_counter(mm, file_rss);
2089			page_add_file_rmap(new_page);
2090		}
2091	} else {
2092		/* One of our sibling threads was faster, back out. */
2093		page_cache_release(new_page);
2094		goto unlock;
2095	}
2096
2097	/* no need to invalidate: a not-present page shouldn't be cached */
2098	update_mmu_cache(vma, address, entry);
2099	lazy_mmu_prot_update(entry);
2100unlock:
2101	pte_unmap_unlock(page_table, ptl);
2102	return ret;
2103oom:
2104	page_cache_release(new_page);
2105	return VM_FAULT_OOM;
2106}
2107
2108/*
2109 * Fault of a previously existing named mapping. Repopulate the pte
2110 * from the encoded file_pte if possible. This enables swappable
2111 * nonlinear vmas.
2112 *
2113 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2114 * but allow concurrent faults), and pte mapped but not yet locked.
2115 * We return with mmap_sem still held, but pte unmapped and unlocked.
2116 */
2117static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2118		unsigned long address, pte_t *page_table, pmd_t *pmd,
2119		int write_access, pte_t orig_pte)
2120{
2121	pgoff_t pgoff;
2122	int err;
2123
2124	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2125		return VM_FAULT_MINOR;
2126
2127	if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2128		/*
2129		 * Page table corrupted: show pte and kill process.
2130		 */
2131		print_bad_pte(vma, orig_pte, address);
2132		return VM_FAULT_OOM;
2133	}
2134	/* We can then assume vm->vm_ops && vma->vm_ops->populate */
2135
2136	pgoff = pte_to_pgoff(orig_pte);
2137	err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2138					vma->vm_page_prot, pgoff, 0);
2139	if (err == -ENOMEM)
2140		return VM_FAULT_OOM;
2141	if (err)
2142		return VM_FAULT_SIGBUS;
2143	return VM_FAULT_MAJOR;
2144}
2145
2146/*
2147 * These routines also need to handle stuff like marking pages dirty
2148 * and/or accessed for architectures that don't do it in hardware (most
2149 * RISC architectures).  The early dirtying is also good on the i386.
2150 *
2151 * There is also a hook called "update_mmu_cache()" that architectures
2152 * with external mmu caches can use to update those (ie the Sparc or
2153 * PowerPC hashed page tables that act as extended TLBs).
2154 *
2155 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2156 * but allow concurrent faults), and pte mapped but not yet locked.
2157 * We return with mmap_sem still held, but pte unmapped and unlocked.
2158 */
2159static inline int handle_pte_fault(struct mm_struct *mm,
2160		struct vm_area_struct *vma, unsigned long address,
2161		pte_t *pte, pmd_t *pmd, int write_access)
2162{
2163	pte_t entry;
2164	pte_t old_entry;
2165	spinlock_t *ptl;
2166
2167	old_entry = entry = *pte;
2168	if (!pte_present(entry)) {
2169		if (pte_none(entry)) {
2170			if (!vma->vm_ops || !vma->vm_ops->nopage)
2171				return do_anonymous_page(mm, vma, address,
2172					pte, pmd, write_access);
2173			return do_no_page(mm, vma, address,
2174					pte, pmd, write_access);
2175		}
2176		if (pte_file(entry))
2177			return do_file_page(mm, vma, address,
2178					pte, pmd, write_access, entry);
2179		return do_swap_page(mm, vma, address,
2180					pte, pmd, write_access, entry);
2181	}
2182
2183	ptl = pte_lockptr(mm, pmd);
2184	spin_lock(ptl);
2185	if (unlikely(!pte_same(*pte, entry)))
2186		goto unlock;
2187	if (write_access) {
2188		if (!pte_write(entry))
2189			return do_wp_page(mm, vma, address,
2190					pte, pmd, ptl, entry);
2191		entry = pte_mkdirty(entry);
2192	}
2193	entry = pte_mkyoung(entry);
2194	if (!pte_same(old_entry, entry)) {
2195		ptep_set_access_flags(vma, address, pte, entry, write_access);
2196		update_mmu_cache(vma, address, entry);
2197		lazy_mmu_prot_update(entry);
2198	} else {
2199		/*
2200		 * This is needed only for protection faults but the arch code
2201		 * is not yet telling us if this is a protection fault or not.
2202		 * This still avoids useless tlb flushes for .text page faults
2203		 * with threads.
2204		 */
2205		if (write_access)
2206			flush_tlb_page(vma, address);
2207	}
2208unlock:
2209	pte_unmap_unlock(pte, ptl);
2210	return VM_FAULT_MINOR;
2211}
2212
2213/*
2214 * By the time we get here, we already hold the mm semaphore
2215 */
2216int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2217		unsigned long address, int write_access)
2218{
2219	pgd_t *pgd;
2220	pud_t *pud;
2221	pmd_t *pmd;
2222	pte_t *pte;
2223
2224	__set_current_state(TASK_RUNNING);
2225
2226	inc_page_state(pgfault);
2227
2228	if (unlikely(is_vm_hugetlb_page(vma)))
2229		return hugetlb_fault(mm, vma, address, write_access);
2230
2231	pgd = pgd_offset(mm, address);
2232	pud = pud_alloc(mm, pgd, address);
2233	if (!pud)
2234		return VM_FAULT_OOM;
2235	pmd = pmd_alloc(mm, pud, address);
2236	if (!pmd)
2237		return VM_FAULT_OOM;
2238	pte = pte_alloc_map(mm, pmd, address);
2239	if (!pte)
2240		return VM_FAULT_OOM;
2241
2242	return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2243}
2244
2245#ifndef __PAGETABLE_PUD_FOLDED
2246/*
2247 * Allocate page upper directory.
2248 * We've already handled the fast-path in-line.
2249 */
2250int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2251{
2252	pud_t *new = pud_alloc_one(mm, address);
2253	if (!new)
2254		return -ENOMEM;
2255
2256	spin_lock(&mm->page_table_lock);
2257	if (pgd_present(*pgd))		/* Another has populated it */
2258		pud_free(new);
2259	else
2260		pgd_populate(mm, pgd, new);
2261	spin_unlock(&mm->page_table_lock);
2262	return 0;
2263}
2264#else
2265/* Workaround for gcc 2.96 */
2266int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2267{
2268	return 0;
2269}
2270#endif /* __PAGETABLE_PUD_FOLDED */
2271
2272#ifndef __PAGETABLE_PMD_FOLDED
2273/*
2274 * Allocate page middle directory.
2275 * We've already handled the fast-path in-line.
2276 */
2277int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2278{
2279	pmd_t *new = pmd_alloc_one(mm, address);
2280	if (!new)
2281		return -ENOMEM;
2282
2283	spin_lock(&mm->page_table_lock);
2284#ifndef __ARCH_HAS_4LEVEL_HACK
2285	if (pud_present(*pud))		/* Another has populated it */
2286		pmd_free(new);
2287	else
2288		pud_populate(mm, pud, new);
2289#else
2290	if (pgd_present(*pud))		/* Another has populated it */
2291		pmd_free(new);
2292	else
2293		pgd_populate(mm, pud, new);
2294#endif /* __ARCH_HAS_4LEVEL_HACK */
2295	spin_unlock(&mm->page_table_lock);
2296	return 0;
2297}
2298#else
2299/* Workaround for gcc 2.96 */
2300int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2301{
2302	return 0;
2303}
2304#endif /* __PAGETABLE_PMD_FOLDED */
2305
2306int make_pages_present(unsigned long addr, unsigned long end)
2307{
2308	int ret, len, write;
2309	struct vm_area_struct * vma;
2310
2311	vma = find_vma(current->mm, addr);
2312	if (!vma)
2313		return -1;
2314	write = (vma->vm_flags & VM_WRITE) != 0;
2315	if (addr >= end)
2316		BUG();
2317	if (end > vma->vm_end)
2318		BUG();
2319	len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2320	ret = get_user_pages(current, current->mm, addr,
2321			len, write, 0, NULL, NULL);
2322	if (ret < 0)
2323		return ret;
2324	return ret == len ? 0 : -1;
2325}
2326
2327/*
2328 * Map a vmalloc()-space virtual address to the physical page.
2329 */
2330struct page * vmalloc_to_page(void * vmalloc_addr)
2331{
2332	unsigned long addr = (unsigned long) vmalloc_addr;
2333	struct page *page = NULL;
2334	pgd_t *pgd = pgd_offset_k(addr);
2335	pud_t *pud;
2336	pmd_t *pmd;
2337	pte_t *ptep, pte;
2338
2339	if (!pgd_none(*pgd)) {
2340		pud = pud_offset(pgd, addr);
2341		if (!pud_none(*pud)) {
2342			pmd = pmd_offset(pud, addr);
2343			if (!pmd_none(*pmd)) {
2344				ptep = pte_offset_map(pmd, addr);
2345				pte = *ptep;
2346				if (pte_present(pte))
2347					page = pte_page(pte);
2348				pte_unmap(ptep);
2349			}
2350		}
2351	}
2352	return page;
2353}
2354
2355EXPORT_SYMBOL(vmalloc_to_page);
2356
2357/*
2358 * Map a vmalloc()-space virtual address to the physical page frame number.
2359 */
2360unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2361{
2362	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2363}
2364
2365EXPORT_SYMBOL(vmalloc_to_pfn);
2366
2367#if !defined(__HAVE_ARCH_GATE_AREA)
2368
2369#if defined(AT_SYSINFO_EHDR)
2370static struct vm_area_struct gate_vma;
2371
2372static int __init gate_vma_init(void)
2373{
2374	gate_vma.vm_mm = NULL;
2375	gate_vma.vm_start = FIXADDR_USER_START;
2376	gate_vma.vm_end = FIXADDR_USER_END;
2377	gate_vma.vm_page_prot = PAGE_READONLY;
2378	gate_vma.vm_flags = 0;
2379	return 0;
2380}
2381__initcall(gate_vma_init);
2382#endif
2383
2384struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2385{
2386#ifdef AT_SYSINFO_EHDR
2387	return &gate_vma;
2388#else
2389	return NULL;
2390#endif
2391}
2392
2393int in_gate_area_no_task(unsigned long addr)
2394{
2395#ifdef AT_SYSINFO_EHDR
2396	if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2397		return 1;
2398#endif
2399	return 0;
2400}
2401
2402#endif	/* __HAVE_ARCH_GATE_AREA */
2403