xref: /openbmc/linux/mm/memory.c (revision b34e08d5)
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/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
62 #include <linux/dma-debug.h>
63 #include <linux/debugfs.h>
64 
65 #include <asm/io.h>
66 #include <asm/pgalloc.h>
67 #include <asm/uaccess.h>
68 #include <asm/tlb.h>
69 #include <asm/tlbflush.h>
70 #include <asm/pgtable.h>
71 
72 #include "internal.h"
73 
74 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76 #endif
77 
78 #ifndef CONFIG_NEED_MULTIPLE_NODES
79 /* use the per-pgdat data instead for discontigmem - mbligh */
80 unsigned long max_mapnr;
81 struct page *mem_map;
82 
83 EXPORT_SYMBOL(max_mapnr);
84 EXPORT_SYMBOL(mem_map);
85 #endif
86 
87 /*
88  * A number of key systems in x86 including ioremap() rely on the assumption
89  * that high_memory defines the upper bound on direct map memory, then end
90  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
91  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92  * and ZONE_HIGHMEM.
93  */
94 void * high_memory;
95 
96 EXPORT_SYMBOL(high_memory);
97 
98 /*
99  * Randomize the address space (stacks, mmaps, brk, etc.).
100  *
101  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102  *   as ancient (libc5 based) binaries can segfault. )
103  */
104 int randomize_va_space __read_mostly =
105 #ifdef CONFIG_COMPAT_BRK
106 					1;
107 #else
108 					2;
109 #endif
110 
111 static int __init disable_randmaps(char *s)
112 {
113 	randomize_va_space = 0;
114 	return 1;
115 }
116 __setup("norandmaps", disable_randmaps);
117 
118 unsigned long zero_pfn __read_mostly;
119 unsigned long highest_memmap_pfn __read_mostly;
120 
121 /*
122  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123  */
124 static int __init init_zero_pfn(void)
125 {
126 	zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 	return 0;
128 }
129 core_initcall(init_zero_pfn);
130 
131 
132 #if defined(SPLIT_RSS_COUNTING)
133 
134 void sync_mm_rss(struct mm_struct *mm)
135 {
136 	int i;
137 
138 	for (i = 0; i < NR_MM_COUNTERS; i++) {
139 		if (current->rss_stat.count[i]) {
140 			add_mm_counter(mm, i, current->rss_stat.count[i]);
141 			current->rss_stat.count[i] = 0;
142 		}
143 	}
144 	current->rss_stat.events = 0;
145 }
146 
147 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148 {
149 	struct task_struct *task = current;
150 
151 	if (likely(task->mm == mm))
152 		task->rss_stat.count[member] += val;
153 	else
154 		add_mm_counter(mm, member, val);
155 }
156 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158 
159 /* sync counter once per 64 page faults */
160 #define TASK_RSS_EVENTS_THRESH	(64)
161 static void check_sync_rss_stat(struct task_struct *task)
162 {
163 	if (unlikely(task != current))
164 		return;
165 	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166 		sync_mm_rss(task->mm);
167 }
168 #else /* SPLIT_RSS_COUNTING */
169 
170 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172 
173 static void check_sync_rss_stat(struct task_struct *task)
174 {
175 }
176 
177 #endif /* SPLIT_RSS_COUNTING */
178 
179 #ifdef HAVE_GENERIC_MMU_GATHER
180 
181 static int tlb_next_batch(struct mmu_gather *tlb)
182 {
183 	struct mmu_gather_batch *batch;
184 
185 	batch = tlb->active;
186 	if (batch->next) {
187 		tlb->active = batch->next;
188 		return 1;
189 	}
190 
191 	if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 		return 0;
193 
194 	batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195 	if (!batch)
196 		return 0;
197 
198 	tlb->batch_count++;
199 	batch->next = NULL;
200 	batch->nr   = 0;
201 	batch->max  = MAX_GATHER_BATCH;
202 
203 	tlb->active->next = batch;
204 	tlb->active = batch;
205 
206 	return 1;
207 }
208 
209 /* tlb_gather_mmu
210  *	Called to initialize an (on-stack) mmu_gather structure for page-table
211  *	tear-down from @mm. The @fullmm argument is used when @mm is without
212  *	users and we're going to destroy the full address space (exit/execve).
213  */
214 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215 {
216 	tlb->mm = mm;
217 
218 	/* Is it from 0 to ~0? */
219 	tlb->fullmm     = !(start | (end+1));
220 	tlb->need_flush_all = 0;
221 	tlb->start	= start;
222 	tlb->end	= end;
223 	tlb->need_flush = 0;
224 	tlb->local.next = NULL;
225 	tlb->local.nr   = 0;
226 	tlb->local.max  = ARRAY_SIZE(tlb->__pages);
227 	tlb->active     = &tlb->local;
228 	tlb->batch_count = 0;
229 
230 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 	tlb->batch = NULL;
232 #endif
233 }
234 
235 void tlb_flush_mmu(struct mmu_gather *tlb)
236 {
237 	struct mmu_gather_batch *batch;
238 
239 	if (!tlb->need_flush)
240 		return;
241 	tlb->need_flush = 0;
242 	tlb_flush(tlb);
243 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
244 	tlb_table_flush(tlb);
245 #endif
246 
247 	for (batch = &tlb->local; batch; batch = batch->next) {
248 		free_pages_and_swap_cache(batch->pages, batch->nr);
249 		batch->nr = 0;
250 	}
251 	tlb->active = &tlb->local;
252 }
253 
254 /* tlb_finish_mmu
255  *	Called at the end of the shootdown operation to free up any resources
256  *	that were required.
257  */
258 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
259 {
260 	struct mmu_gather_batch *batch, *next;
261 
262 	tlb_flush_mmu(tlb);
263 
264 	/* keep the page table cache within bounds */
265 	check_pgt_cache();
266 
267 	for (batch = tlb->local.next; batch; batch = next) {
268 		next = batch->next;
269 		free_pages((unsigned long)batch, 0);
270 	}
271 	tlb->local.next = NULL;
272 }
273 
274 /* __tlb_remove_page
275  *	Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
276  *	handling the additional races in SMP caused by other CPUs caching valid
277  *	mappings in their TLBs. Returns the number of free page slots left.
278  *	When out of page slots we must call tlb_flush_mmu().
279  */
280 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
281 {
282 	struct mmu_gather_batch *batch;
283 
284 	VM_BUG_ON(!tlb->need_flush);
285 
286 	batch = tlb->active;
287 	batch->pages[batch->nr++] = page;
288 	if (batch->nr == batch->max) {
289 		if (!tlb_next_batch(tlb))
290 			return 0;
291 		batch = tlb->active;
292 	}
293 	VM_BUG_ON_PAGE(batch->nr > batch->max, page);
294 
295 	return batch->max - batch->nr;
296 }
297 
298 #endif /* HAVE_GENERIC_MMU_GATHER */
299 
300 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
301 
302 /*
303  * See the comment near struct mmu_table_batch.
304  */
305 
306 static void tlb_remove_table_smp_sync(void *arg)
307 {
308 	/* Simply deliver the interrupt */
309 }
310 
311 static void tlb_remove_table_one(void *table)
312 {
313 	/*
314 	 * This isn't an RCU grace period and hence the page-tables cannot be
315 	 * assumed to be actually RCU-freed.
316 	 *
317 	 * It is however sufficient for software page-table walkers that rely on
318 	 * IRQ disabling. See the comment near struct mmu_table_batch.
319 	 */
320 	smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
321 	__tlb_remove_table(table);
322 }
323 
324 static void tlb_remove_table_rcu(struct rcu_head *head)
325 {
326 	struct mmu_table_batch *batch;
327 	int i;
328 
329 	batch = container_of(head, struct mmu_table_batch, rcu);
330 
331 	for (i = 0; i < batch->nr; i++)
332 		__tlb_remove_table(batch->tables[i]);
333 
334 	free_page((unsigned long)batch);
335 }
336 
337 void tlb_table_flush(struct mmu_gather *tlb)
338 {
339 	struct mmu_table_batch **batch = &tlb->batch;
340 
341 	if (*batch) {
342 		call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
343 		*batch = NULL;
344 	}
345 }
346 
347 void tlb_remove_table(struct mmu_gather *tlb, void *table)
348 {
349 	struct mmu_table_batch **batch = &tlb->batch;
350 
351 	tlb->need_flush = 1;
352 
353 	/*
354 	 * When there's less then two users of this mm there cannot be a
355 	 * concurrent page-table walk.
356 	 */
357 	if (atomic_read(&tlb->mm->mm_users) < 2) {
358 		__tlb_remove_table(table);
359 		return;
360 	}
361 
362 	if (*batch == NULL) {
363 		*batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
364 		if (*batch == NULL) {
365 			tlb_remove_table_one(table);
366 			return;
367 		}
368 		(*batch)->nr = 0;
369 	}
370 	(*batch)->tables[(*batch)->nr++] = table;
371 	if ((*batch)->nr == MAX_TABLE_BATCH)
372 		tlb_table_flush(tlb);
373 }
374 
375 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
376 
377 /*
378  * Note: this doesn't free the actual pages themselves. That
379  * has been handled earlier when unmapping all the memory regions.
380  */
381 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
382 			   unsigned long addr)
383 {
384 	pgtable_t token = pmd_pgtable(*pmd);
385 	pmd_clear(pmd);
386 	pte_free_tlb(tlb, token, addr);
387 	atomic_long_dec(&tlb->mm->nr_ptes);
388 }
389 
390 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
391 				unsigned long addr, unsigned long end,
392 				unsigned long floor, unsigned long ceiling)
393 {
394 	pmd_t *pmd;
395 	unsigned long next;
396 	unsigned long start;
397 
398 	start = addr;
399 	pmd = pmd_offset(pud, addr);
400 	do {
401 		next = pmd_addr_end(addr, end);
402 		if (pmd_none_or_clear_bad(pmd))
403 			continue;
404 		free_pte_range(tlb, pmd, addr);
405 	} while (pmd++, addr = next, addr != end);
406 
407 	start &= PUD_MASK;
408 	if (start < floor)
409 		return;
410 	if (ceiling) {
411 		ceiling &= PUD_MASK;
412 		if (!ceiling)
413 			return;
414 	}
415 	if (end - 1 > ceiling - 1)
416 		return;
417 
418 	pmd = pmd_offset(pud, start);
419 	pud_clear(pud);
420 	pmd_free_tlb(tlb, pmd, start);
421 }
422 
423 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
424 				unsigned long addr, unsigned long end,
425 				unsigned long floor, unsigned long ceiling)
426 {
427 	pud_t *pud;
428 	unsigned long next;
429 	unsigned long start;
430 
431 	start = addr;
432 	pud = pud_offset(pgd, addr);
433 	do {
434 		next = pud_addr_end(addr, end);
435 		if (pud_none_or_clear_bad(pud))
436 			continue;
437 		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
438 	} while (pud++, addr = next, addr != end);
439 
440 	start &= PGDIR_MASK;
441 	if (start < floor)
442 		return;
443 	if (ceiling) {
444 		ceiling &= PGDIR_MASK;
445 		if (!ceiling)
446 			return;
447 	}
448 	if (end - 1 > ceiling - 1)
449 		return;
450 
451 	pud = pud_offset(pgd, start);
452 	pgd_clear(pgd);
453 	pud_free_tlb(tlb, pud, start);
454 }
455 
456 /*
457  * This function frees user-level page tables of a process.
458  */
459 void free_pgd_range(struct mmu_gather *tlb,
460 			unsigned long addr, unsigned long end,
461 			unsigned long floor, unsigned long ceiling)
462 {
463 	pgd_t *pgd;
464 	unsigned long next;
465 
466 	/*
467 	 * The next few lines have given us lots of grief...
468 	 *
469 	 * Why are we testing PMD* at this top level?  Because often
470 	 * there will be no work to do at all, and we'd prefer not to
471 	 * go all the way down to the bottom just to discover that.
472 	 *
473 	 * Why all these "- 1"s?  Because 0 represents both the bottom
474 	 * of the address space and the top of it (using -1 for the
475 	 * top wouldn't help much: the masks would do the wrong thing).
476 	 * The rule is that addr 0 and floor 0 refer to the bottom of
477 	 * the address space, but end 0 and ceiling 0 refer to the top
478 	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
479 	 * that end 0 case should be mythical).
480 	 *
481 	 * Wherever addr is brought up or ceiling brought down, we must
482 	 * be careful to reject "the opposite 0" before it confuses the
483 	 * subsequent tests.  But what about where end is brought down
484 	 * by PMD_SIZE below? no, end can't go down to 0 there.
485 	 *
486 	 * Whereas we round start (addr) and ceiling down, by different
487 	 * masks at different levels, in order to test whether a table
488 	 * now has no other vmas using it, so can be freed, we don't
489 	 * bother to round floor or end up - the tests don't need that.
490 	 */
491 
492 	addr &= PMD_MASK;
493 	if (addr < floor) {
494 		addr += PMD_SIZE;
495 		if (!addr)
496 			return;
497 	}
498 	if (ceiling) {
499 		ceiling &= PMD_MASK;
500 		if (!ceiling)
501 			return;
502 	}
503 	if (end - 1 > ceiling - 1)
504 		end -= PMD_SIZE;
505 	if (addr > end - 1)
506 		return;
507 
508 	pgd = pgd_offset(tlb->mm, addr);
509 	do {
510 		next = pgd_addr_end(addr, end);
511 		if (pgd_none_or_clear_bad(pgd))
512 			continue;
513 		free_pud_range(tlb, pgd, addr, next, floor, ceiling);
514 	} while (pgd++, addr = next, addr != end);
515 }
516 
517 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
518 		unsigned long floor, unsigned long ceiling)
519 {
520 	while (vma) {
521 		struct vm_area_struct *next = vma->vm_next;
522 		unsigned long addr = vma->vm_start;
523 
524 		/*
525 		 * Hide vma from rmap and truncate_pagecache before freeing
526 		 * pgtables
527 		 */
528 		unlink_anon_vmas(vma);
529 		unlink_file_vma(vma);
530 
531 		if (is_vm_hugetlb_page(vma)) {
532 			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
533 				floor, next? next->vm_start: ceiling);
534 		} else {
535 			/*
536 			 * Optimization: gather nearby vmas into one call down
537 			 */
538 			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
539 			       && !is_vm_hugetlb_page(next)) {
540 				vma = next;
541 				next = vma->vm_next;
542 				unlink_anon_vmas(vma);
543 				unlink_file_vma(vma);
544 			}
545 			free_pgd_range(tlb, addr, vma->vm_end,
546 				floor, next? next->vm_start: ceiling);
547 		}
548 		vma = next;
549 	}
550 }
551 
552 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
553 		pmd_t *pmd, unsigned long address)
554 {
555 	spinlock_t *ptl;
556 	pgtable_t new = pte_alloc_one(mm, address);
557 	int wait_split_huge_page;
558 	if (!new)
559 		return -ENOMEM;
560 
561 	/*
562 	 * Ensure all pte setup (eg. pte page lock and page clearing) are
563 	 * visible before the pte is made visible to other CPUs by being
564 	 * put into page tables.
565 	 *
566 	 * The other side of the story is the pointer chasing in the page
567 	 * table walking code (when walking the page table without locking;
568 	 * ie. most of the time). Fortunately, these data accesses consist
569 	 * of a chain of data-dependent loads, meaning most CPUs (alpha
570 	 * being the notable exception) will already guarantee loads are
571 	 * seen in-order. See the alpha page table accessors for the
572 	 * smp_read_barrier_depends() barriers in page table walking code.
573 	 */
574 	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
575 
576 	ptl = pmd_lock(mm, pmd);
577 	wait_split_huge_page = 0;
578 	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
579 		atomic_long_inc(&mm->nr_ptes);
580 		pmd_populate(mm, pmd, new);
581 		new = NULL;
582 	} else if (unlikely(pmd_trans_splitting(*pmd)))
583 		wait_split_huge_page = 1;
584 	spin_unlock(ptl);
585 	if (new)
586 		pte_free(mm, new);
587 	if (wait_split_huge_page)
588 		wait_split_huge_page(vma->anon_vma, pmd);
589 	return 0;
590 }
591 
592 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
593 {
594 	pte_t *new = pte_alloc_one_kernel(&init_mm, address);
595 	if (!new)
596 		return -ENOMEM;
597 
598 	smp_wmb(); /* See comment in __pte_alloc */
599 
600 	spin_lock(&init_mm.page_table_lock);
601 	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
602 		pmd_populate_kernel(&init_mm, pmd, new);
603 		new = NULL;
604 	} else
605 		VM_BUG_ON(pmd_trans_splitting(*pmd));
606 	spin_unlock(&init_mm.page_table_lock);
607 	if (new)
608 		pte_free_kernel(&init_mm, new);
609 	return 0;
610 }
611 
612 static inline void init_rss_vec(int *rss)
613 {
614 	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
615 }
616 
617 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
618 {
619 	int i;
620 
621 	if (current->mm == mm)
622 		sync_mm_rss(mm);
623 	for (i = 0; i < NR_MM_COUNTERS; i++)
624 		if (rss[i])
625 			add_mm_counter(mm, i, rss[i]);
626 }
627 
628 /*
629  * This function is called to print an error when a bad pte
630  * is found. For example, we might have a PFN-mapped pte in
631  * a region that doesn't allow it.
632  *
633  * The calling function must still handle the error.
634  */
635 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
636 			  pte_t pte, struct page *page)
637 {
638 	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
639 	pud_t *pud = pud_offset(pgd, addr);
640 	pmd_t *pmd = pmd_offset(pud, addr);
641 	struct address_space *mapping;
642 	pgoff_t index;
643 	static unsigned long resume;
644 	static unsigned long nr_shown;
645 	static unsigned long nr_unshown;
646 
647 	/*
648 	 * Allow a burst of 60 reports, then keep quiet for that minute;
649 	 * or allow a steady drip of one report per second.
650 	 */
651 	if (nr_shown == 60) {
652 		if (time_before(jiffies, resume)) {
653 			nr_unshown++;
654 			return;
655 		}
656 		if (nr_unshown) {
657 			printk(KERN_ALERT
658 				"BUG: Bad page map: %lu messages suppressed\n",
659 				nr_unshown);
660 			nr_unshown = 0;
661 		}
662 		nr_shown = 0;
663 	}
664 	if (nr_shown++ == 0)
665 		resume = jiffies + 60 * HZ;
666 
667 	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
668 	index = linear_page_index(vma, addr);
669 
670 	printk(KERN_ALERT
671 		"BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
672 		current->comm,
673 		(long long)pte_val(pte), (long long)pmd_val(*pmd));
674 	if (page)
675 		dump_page(page, "bad pte");
676 	printk(KERN_ALERT
677 		"addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
678 		(void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
679 	/*
680 	 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
681 	 */
682 	if (vma->vm_ops)
683 		printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
684 		       vma->vm_ops->fault);
685 	if (vma->vm_file)
686 		printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
687 		       vma->vm_file->f_op->mmap);
688 	dump_stack();
689 	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
690 }
691 
692 static inline bool is_cow_mapping(vm_flags_t flags)
693 {
694 	return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
695 }
696 
697 /*
698  * vm_normal_page -- This function gets the "struct page" associated with a pte.
699  *
700  * "Special" mappings do not wish to be associated with a "struct page" (either
701  * it doesn't exist, or it exists but they don't want to touch it). In this
702  * case, NULL is returned here. "Normal" mappings do have a struct page.
703  *
704  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
705  * pte bit, in which case this function is trivial. Secondly, an architecture
706  * may not have a spare pte bit, which requires a more complicated scheme,
707  * described below.
708  *
709  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
710  * special mapping (even if there are underlying and valid "struct pages").
711  * COWed pages of a VM_PFNMAP are always normal.
712  *
713  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
714  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
715  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
716  * mapping will always honor the rule
717  *
718  *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
719  *
720  * And for normal mappings this is false.
721  *
722  * This restricts such mappings to be a linear translation from virtual address
723  * to pfn. To get around this restriction, we allow arbitrary mappings so long
724  * as the vma is not a COW mapping; in that case, we know that all ptes are
725  * special (because none can have been COWed).
726  *
727  *
728  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
729  *
730  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
731  * page" backing, however the difference is that _all_ pages with a struct
732  * page (that is, those where pfn_valid is true) are refcounted and considered
733  * normal pages by the VM. The disadvantage is that pages are refcounted
734  * (which can be slower and simply not an option for some PFNMAP users). The
735  * advantage is that we don't have to follow the strict linearity rule of
736  * PFNMAP mappings in order to support COWable mappings.
737  *
738  */
739 #ifdef __HAVE_ARCH_PTE_SPECIAL
740 # define HAVE_PTE_SPECIAL 1
741 #else
742 # define HAVE_PTE_SPECIAL 0
743 #endif
744 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
745 				pte_t pte)
746 {
747 	unsigned long pfn = pte_pfn(pte);
748 
749 	if (HAVE_PTE_SPECIAL) {
750 		if (likely(!pte_special(pte)))
751 			goto check_pfn;
752 		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
753 			return NULL;
754 		if (!is_zero_pfn(pfn))
755 			print_bad_pte(vma, addr, pte, NULL);
756 		return NULL;
757 	}
758 
759 	/* !HAVE_PTE_SPECIAL case follows: */
760 
761 	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
762 		if (vma->vm_flags & VM_MIXEDMAP) {
763 			if (!pfn_valid(pfn))
764 				return NULL;
765 			goto out;
766 		} else {
767 			unsigned long off;
768 			off = (addr - vma->vm_start) >> PAGE_SHIFT;
769 			if (pfn == vma->vm_pgoff + off)
770 				return NULL;
771 			if (!is_cow_mapping(vma->vm_flags))
772 				return NULL;
773 		}
774 	}
775 
776 	if (is_zero_pfn(pfn))
777 		return NULL;
778 check_pfn:
779 	if (unlikely(pfn > highest_memmap_pfn)) {
780 		print_bad_pte(vma, addr, pte, NULL);
781 		return NULL;
782 	}
783 
784 	/*
785 	 * NOTE! We still have PageReserved() pages in the page tables.
786 	 * eg. VDSO mappings can cause them to exist.
787 	 */
788 out:
789 	return pfn_to_page(pfn);
790 }
791 
792 /*
793  * copy one vm_area from one task to the other. Assumes the page tables
794  * already present in the new task to be cleared in the whole range
795  * covered by this vma.
796  */
797 
798 static inline unsigned long
799 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
800 		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
801 		unsigned long addr, int *rss)
802 {
803 	unsigned long vm_flags = vma->vm_flags;
804 	pte_t pte = *src_pte;
805 	struct page *page;
806 
807 	/* pte contains position in swap or file, so copy. */
808 	if (unlikely(!pte_present(pte))) {
809 		if (!pte_file(pte)) {
810 			swp_entry_t entry = pte_to_swp_entry(pte);
811 
812 			if (swap_duplicate(entry) < 0)
813 				return entry.val;
814 
815 			/* make sure dst_mm is on swapoff's mmlist. */
816 			if (unlikely(list_empty(&dst_mm->mmlist))) {
817 				spin_lock(&mmlist_lock);
818 				if (list_empty(&dst_mm->mmlist))
819 					list_add(&dst_mm->mmlist,
820 						 &src_mm->mmlist);
821 				spin_unlock(&mmlist_lock);
822 			}
823 			if (likely(!non_swap_entry(entry)))
824 				rss[MM_SWAPENTS]++;
825 			else if (is_migration_entry(entry)) {
826 				page = migration_entry_to_page(entry);
827 
828 				if (PageAnon(page))
829 					rss[MM_ANONPAGES]++;
830 				else
831 					rss[MM_FILEPAGES]++;
832 
833 				if (is_write_migration_entry(entry) &&
834 				    is_cow_mapping(vm_flags)) {
835 					/*
836 					 * COW mappings require pages in both
837 					 * parent and child to be set to read.
838 					 */
839 					make_migration_entry_read(&entry);
840 					pte = swp_entry_to_pte(entry);
841 					if (pte_swp_soft_dirty(*src_pte))
842 						pte = pte_swp_mksoft_dirty(pte);
843 					set_pte_at(src_mm, addr, src_pte, pte);
844 				}
845 			}
846 		}
847 		goto out_set_pte;
848 	}
849 
850 	/*
851 	 * If it's a COW mapping, write protect it both
852 	 * in the parent and the child
853 	 */
854 	if (is_cow_mapping(vm_flags)) {
855 		ptep_set_wrprotect(src_mm, addr, src_pte);
856 		pte = pte_wrprotect(pte);
857 	}
858 
859 	/*
860 	 * If it's a shared mapping, mark it clean in
861 	 * the child
862 	 */
863 	if (vm_flags & VM_SHARED)
864 		pte = pte_mkclean(pte);
865 	pte = pte_mkold(pte);
866 
867 	page = vm_normal_page(vma, addr, pte);
868 	if (page) {
869 		get_page(page);
870 		page_dup_rmap(page);
871 		if (PageAnon(page))
872 			rss[MM_ANONPAGES]++;
873 		else
874 			rss[MM_FILEPAGES]++;
875 	}
876 
877 out_set_pte:
878 	set_pte_at(dst_mm, addr, dst_pte, pte);
879 	return 0;
880 }
881 
882 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
883 		   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
884 		   unsigned long addr, unsigned long end)
885 {
886 	pte_t *orig_src_pte, *orig_dst_pte;
887 	pte_t *src_pte, *dst_pte;
888 	spinlock_t *src_ptl, *dst_ptl;
889 	int progress = 0;
890 	int rss[NR_MM_COUNTERS];
891 	swp_entry_t entry = (swp_entry_t){0};
892 
893 again:
894 	init_rss_vec(rss);
895 
896 	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
897 	if (!dst_pte)
898 		return -ENOMEM;
899 	src_pte = pte_offset_map(src_pmd, addr);
900 	src_ptl = pte_lockptr(src_mm, src_pmd);
901 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
902 	orig_src_pte = src_pte;
903 	orig_dst_pte = dst_pte;
904 	arch_enter_lazy_mmu_mode();
905 
906 	do {
907 		/*
908 		 * We are holding two locks at this point - either of them
909 		 * could generate latencies in another task on another CPU.
910 		 */
911 		if (progress >= 32) {
912 			progress = 0;
913 			if (need_resched() ||
914 			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
915 				break;
916 		}
917 		if (pte_none(*src_pte)) {
918 			progress++;
919 			continue;
920 		}
921 		entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
922 							vma, addr, rss);
923 		if (entry.val)
924 			break;
925 		progress += 8;
926 	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
927 
928 	arch_leave_lazy_mmu_mode();
929 	spin_unlock(src_ptl);
930 	pte_unmap(orig_src_pte);
931 	add_mm_rss_vec(dst_mm, rss);
932 	pte_unmap_unlock(orig_dst_pte, dst_ptl);
933 	cond_resched();
934 
935 	if (entry.val) {
936 		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
937 			return -ENOMEM;
938 		progress = 0;
939 	}
940 	if (addr != end)
941 		goto again;
942 	return 0;
943 }
944 
945 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
946 		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
947 		unsigned long addr, unsigned long end)
948 {
949 	pmd_t *src_pmd, *dst_pmd;
950 	unsigned long next;
951 
952 	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
953 	if (!dst_pmd)
954 		return -ENOMEM;
955 	src_pmd = pmd_offset(src_pud, addr);
956 	do {
957 		next = pmd_addr_end(addr, end);
958 		if (pmd_trans_huge(*src_pmd)) {
959 			int err;
960 			VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
961 			err = copy_huge_pmd(dst_mm, src_mm,
962 					    dst_pmd, src_pmd, addr, vma);
963 			if (err == -ENOMEM)
964 				return -ENOMEM;
965 			if (!err)
966 				continue;
967 			/* fall through */
968 		}
969 		if (pmd_none_or_clear_bad(src_pmd))
970 			continue;
971 		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
972 						vma, addr, next))
973 			return -ENOMEM;
974 	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
975 	return 0;
976 }
977 
978 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
979 		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
980 		unsigned long addr, unsigned long end)
981 {
982 	pud_t *src_pud, *dst_pud;
983 	unsigned long next;
984 
985 	dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
986 	if (!dst_pud)
987 		return -ENOMEM;
988 	src_pud = pud_offset(src_pgd, addr);
989 	do {
990 		next = pud_addr_end(addr, end);
991 		if (pud_none_or_clear_bad(src_pud))
992 			continue;
993 		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
994 						vma, addr, next))
995 			return -ENOMEM;
996 	} while (dst_pud++, src_pud++, addr = next, addr != end);
997 	return 0;
998 }
999 
1000 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1001 		struct vm_area_struct *vma)
1002 {
1003 	pgd_t *src_pgd, *dst_pgd;
1004 	unsigned long next;
1005 	unsigned long addr = vma->vm_start;
1006 	unsigned long end = vma->vm_end;
1007 	unsigned long mmun_start;	/* For mmu_notifiers */
1008 	unsigned long mmun_end;		/* For mmu_notifiers */
1009 	bool is_cow;
1010 	int ret;
1011 
1012 	/*
1013 	 * Don't copy ptes where a page fault will fill them correctly.
1014 	 * Fork becomes much lighter when there are big shared or private
1015 	 * readonly mappings. The tradeoff is that copy_page_range is more
1016 	 * efficient than faulting.
1017 	 */
1018 	if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1019 			       VM_PFNMAP | VM_MIXEDMAP))) {
1020 		if (!vma->anon_vma)
1021 			return 0;
1022 	}
1023 
1024 	if (is_vm_hugetlb_page(vma))
1025 		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1026 
1027 	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1028 		/*
1029 		 * We do not free on error cases below as remove_vma
1030 		 * gets called on error from higher level routine
1031 		 */
1032 		ret = track_pfn_copy(vma);
1033 		if (ret)
1034 			return ret;
1035 	}
1036 
1037 	/*
1038 	 * We need to invalidate the secondary MMU mappings only when
1039 	 * there could be a permission downgrade on the ptes of the
1040 	 * parent mm. And a permission downgrade will only happen if
1041 	 * is_cow_mapping() returns true.
1042 	 */
1043 	is_cow = is_cow_mapping(vma->vm_flags);
1044 	mmun_start = addr;
1045 	mmun_end   = end;
1046 	if (is_cow)
1047 		mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1048 						    mmun_end);
1049 
1050 	ret = 0;
1051 	dst_pgd = pgd_offset(dst_mm, addr);
1052 	src_pgd = pgd_offset(src_mm, addr);
1053 	do {
1054 		next = pgd_addr_end(addr, end);
1055 		if (pgd_none_or_clear_bad(src_pgd))
1056 			continue;
1057 		if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1058 					    vma, addr, next))) {
1059 			ret = -ENOMEM;
1060 			break;
1061 		}
1062 	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1063 
1064 	if (is_cow)
1065 		mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1066 	return ret;
1067 }
1068 
1069 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1070 				struct vm_area_struct *vma, pmd_t *pmd,
1071 				unsigned long addr, unsigned long end,
1072 				struct zap_details *details)
1073 {
1074 	struct mm_struct *mm = tlb->mm;
1075 	int force_flush = 0;
1076 	int rss[NR_MM_COUNTERS];
1077 	spinlock_t *ptl;
1078 	pte_t *start_pte;
1079 	pte_t *pte;
1080 
1081 again:
1082 	init_rss_vec(rss);
1083 	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1084 	pte = start_pte;
1085 	arch_enter_lazy_mmu_mode();
1086 	do {
1087 		pte_t ptent = *pte;
1088 		if (pte_none(ptent)) {
1089 			continue;
1090 		}
1091 
1092 		if (pte_present(ptent)) {
1093 			struct page *page;
1094 
1095 			page = vm_normal_page(vma, addr, ptent);
1096 			if (unlikely(details) && page) {
1097 				/*
1098 				 * unmap_shared_mapping_pages() wants to
1099 				 * invalidate cache without truncating:
1100 				 * unmap shared but keep private pages.
1101 				 */
1102 				if (details->check_mapping &&
1103 				    details->check_mapping != page->mapping)
1104 					continue;
1105 				/*
1106 				 * Each page->index must be checked when
1107 				 * invalidating or truncating nonlinear.
1108 				 */
1109 				if (details->nonlinear_vma &&
1110 				    (page->index < details->first_index ||
1111 				     page->index > details->last_index))
1112 					continue;
1113 			}
1114 			ptent = ptep_get_and_clear_full(mm, addr, pte,
1115 							tlb->fullmm);
1116 			tlb_remove_tlb_entry(tlb, pte, addr);
1117 			if (unlikely(!page))
1118 				continue;
1119 			if (unlikely(details) && details->nonlinear_vma
1120 			    && linear_page_index(details->nonlinear_vma,
1121 						addr) != page->index) {
1122 				pte_t ptfile = pgoff_to_pte(page->index);
1123 				if (pte_soft_dirty(ptent))
1124 					pte_file_mksoft_dirty(ptfile);
1125 				set_pte_at(mm, addr, pte, ptfile);
1126 			}
1127 			if (PageAnon(page))
1128 				rss[MM_ANONPAGES]--;
1129 			else {
1130 				if (pte_dirty(ptent))
1131 					set_page_dirty(page);
1132 				if (pte_young(ptent) &&
1133 				    likely(!(vma->vm_flags & VM_SEQ_READ)))
1134 					mark_page_accessed(page);
1135 				rss[MM_FILEPAGES]--;
1136 			}
1137 			page_remove_rmap(page);
1138 			if (unlikely(page_mapcount(page) < 0))
1139 				print_bad_pte(vma, addr, ptent, page);
1140 			force_flush = !__tlb_remove_page(tlb, page);
1141 			if (force_flush)
1142 				break;
1143 			continue;
1144 		}
1145 		/*
1146 		 * If details->check_mapping, we leave swap entries;
1147 		 * if details->nonlinear_vma, we leave file entries.
1148 		 */
1149 		if (unlikely(details))
1150 			continue;
1151 		if (pte_file(ptent)) {
1152 			if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1153 				print_bad_pte(vma, addr, ptent, NULL);
1154 		} else {
1155 			swp_entry_t entry = pte_to_swp_entry(ptent);
1156 
1157 			if (!non_swap_entry(entry))
1158 				rss[MM_SWAPENTS]--;
1159 			else if (is_migration_entry(entry)) {
1160 				struct page *page;
1161 
1162 				page = migration_entry_to_page(entry);
1163 
1164 				if (PageAnon(page))
1165 					rss[MM_ANONPAGES]--;
1166 				else
1167 					rss[MM_FILEPAGES]--;
1168 			}
1169 			if (unlikely(!free_swap_and_cache(entry)))
1170 				print_bad_pte(vma, addr, ptent, NULL);
1171 		}
1172 		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1173 	} while (pte++, addr += PAGE_SIZE, addr != end);
1174 
1175 	add_mm_rss_vec(mm, rss);
1176 	arch_leave_lazy_mmu_mode();
1177 	pte_unmap_unlock(start_pte, ptl);
1178 
1179 	/*
1180 	 * mmu_gather ran out of room to batch pages, we break out of
1181 	 * the PTE lock to avoid doing the potential expensive TLB invalidate
1182 	 * and page-free while holding it.
1183 	 */
1184 	if (force_flush) {
1185 		unsigned long old_end;
1186 
1187 		force_flush = 0;
1188 
1189 		/*
1190 		 * Flush the TLB just for the previous segment,
1191 		 * then update the range to be the remaining
1192 		 * TLB range.
1193 		 */
1194 		old_end = tlb->end;
1195 		tlb->end = addr;
1196 
1197 		tlb_flush_mmu(tlb);
1198 
1199 		tlb->start = addr;
1200 		tlb->end = old_end;
1201 
1202 		if (addr != end)
1203 			goto again;
1204 	}
1205 
1206 	return addr;
1207 }
1208 
1209 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1210 				struct vm_area_struct *vma, pud_t *pud,
1211 				unsigned long addr, unsigned long end,
1212 				struct zap_details *details)
1213 {
1214 	pmd_t *pmd;
1215 	unsigned long next;
1216 
1217 	pmd = pmd_offset(pud, addr);
1218 	do {
1219 		next = pmd_addr_end(addr, end);
1220 		if (pmd_trans_huge(*pmd)) {
1221 			if (next - addr != HPAGE_PMD_SIZE) {
1222 #ifdef CONFIG_DEBUG_VM
1223 				if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1224 					pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1225 						__func__, addr, end,
1226 						vma->vm_start,
1227 						vma->vm_end);
1228 					BUG();
1229 				}
1230 #endif
1231 				split_huge_page_pmd(vma, addr, pmd);
1232 			} else if (zap_huge_pmd(tlb, vma, pmd, addr))
1233 				goto next;
1234 			/* fall through */
1235 		}
1236 		/*
1237 		 * Here there can be other concurrent MADV_DONTNEED or
1238 		 * trans huge page faults running, and if the pmd is
1239 		 * none or trans huge it can change under us. This is
1240 		 * because MADV_DONTNEED holds the mmap_sem in read
1241 		 * mode.
1242 		 */
1243 		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1244 			goto next;
1245 		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1246 next:
1247 		cond_resched();
1248 	} while (pmd++, addr = next, addr != end);
1249 
1250 	return addr;
1251 }
1252 
1253 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1254 				struct vm_area_struct *vma, pgd_t *pgd,
1255 				unsigned long addr, unsigned long end,
1256 				struct zap_details *details)
1257 {
1258 	pud_t *pud;
1259 	unsigned long next;
1260 
1261 	pud = pud_offset(pgd, addr);
1262 	do {
1263 		next = pud_addr_end(addr, end);
1264 		if (pud_none_or_clear_bad(pud))
1265 			continue;
1266 		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1267 	} while (pud++, addr = next, addr != end);
1268 
1269 	return addr;
1270 }
1271 
1272 static void unmap_page_range(struct mmu_gather *tlb,
1273 			     struct vm_area_struct *vma,
1274 			     unsigned long addr, unsigned long end,
1275 			     struct zap_details *details)
1276 {
1277 	pgd_t *pgd;
1278 	unsigned long next;
1279 
1280 	if (details && !details->check_mapping && !details->nonlinear_vma)
1281 		details = NULL;
1282 
1283 	BUG_ON(addr >= end);
1284 	mem_cgroup_uncharge_start();
1285 	tlb_start_vma(tlb, vma);
1286 	pgd = pgd_offset(vma->vm_mm, addr);
1287 	do {
1288 		next = pgd_addr_end(addr, end);
1289 		if (pgd_none_or_clear_bad(pgd))
1290 			continue;
1291 		next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1292 	} while (pgd++, addr = next, addr != end);
1293 	tlb_end_vma(tlb, vma);
1294 	mem_cgroup_uncharge_end();
1295 }
1296 
1297 
1298 static void unmap_single_vma(struct mmu_gather *tlb,
1299 		struct vm_area_struct *vma, unsigned long start_addr,
1300 		unsigned long end_addr,
1301 		struct zap_details *details)
1302 {
1303 	unsigned long start = max(vma->vm_start, start_addr);
1304 	unsigned long end;
1305 
1306 	if (start >= vma->vm_end)
1307 		return;
1308 	end = min(vma->vm_end, end_addr);
1309 	if (end <= vma->vm_start)
1310 		return;
1311 
1312 	if (vma->vm_file)
1313 		uprobe_munmap(vma, start, end);
1314 
1315 	if (unlikely(vma->vm_flags & VM_PFNMAP))
1316 		untrack_pfn(vma, 0, 0);
1317 
1318 	if (start != end) {
1319 		if (unlikely(is_vm_hugetlb_page(vma))) {
1320 			/*
1321 			 * It is undesirable to test vma->vm_file as it
1322 			 * should be non-null for valid hugetlb area.
1323 			 * However, vm_file will be NULL in the error
1324 			 * cleanup path of mmap_region. When
1325 			 * hugetlbfs ->mmap method fails,
1326 			 * mmap_region() nullifies vma->vm_file
1327 			 * before calling this function to clean up.
1328 			 * Since no pte has actually been setup, it is
1329 			 * safe to do nothing in this case.
1330 			 */
1331 			if (vma->vm_file) {
1332 				mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1333 				__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1334 				mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1335 			}
1336 		} else
1337 			unmap_page_range(tlb, vma, start, end, details);
1338 	}
1339 }
1340 
1341 /**
1342  * unmap_vmas - unmap a range of memory covered by a list of vma's
1343  * @tlb: address of the caller's struct mmu_gather
1344  * @vma: the starting vma
1345  * @start_addr: virtual address at which to start unmapping
1346  * @end_addr: virtual address at which to end unmapping
1347  *
1348  * Unmap all pages in the vma list.
1349  *
1350  * Only addresses between `start' and `end' will be unmapped.
1351  *
1352  * The VMA list must be sorted in ascending virtual address order.
1353  *
1354  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1355  * range after unmap_vmas() returns.  So the only responsibility here is to
1356  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1357  * drops the lock and schedules.
1358  */
1359 void unmap_vmas(struct mmu_gather *tlb,
1360 		struct vm_area_struct *vma, unsigned long start_addr,
1361 		unsigned long end_addr)
1362 {
1363 	struct mm_struct *mm = vma->vm_mm;
1364 
1365 	mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1366 	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1367 		unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1368 	mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1369 }
1370 
1371 /**
1372  * zap_page_range - remove user pages in a given range
1373  * @vma: vm_area_struct holding the applicable pages
1374  * @start: starting address of pages to zap
1375  * @size: number of bytes to zap
1376  * @details: details of nonlinear truncation or shared cache invalidation
1377  *
1378  * Caller must protect the VMA list
1379  */
1380 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1381 		unsigned long size, struct zap_details *details)
1382 {
1383 	struct mm_struct *mm = vma->vm_mm;
1384 	struct mmu_gather tlb;
1385 	unsigned long end = start + size;
1386 
1387 	lru_add_drain();
1388 	tlb_gather_mmu(&tlb, mm, start, end);
1389 	update_hiwater_rss(mm);
1390 	mmu_notifier_invalidate_range_start(mm, start, end);
1391 	for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1392 		unmap_single_vma(&tlb, vma, start, end, details);
1393 	mmu_notifier_invalidate_range_end(mm, start, end);
1394 	tlb_finish_mmu(&tlb, start, end);
1395 }
1396 
1397 /**
1398  * zap_page_range_single - remove user pages in a given range
1399  * @vma: vm_area_struct holding the applicable pages
1400  * @address: starting address of pages to zap
1401  * @size: number of bytes to zap
1402  * @details: details of nonlinear truncation or shared cache invalidation
1403  *
1404  * The range must fit into one VMA.
1405  */
1406 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1407 		unsigned long size, struct zap_details *details)
1408 {
1409 	struct mm_struct *mm = vma->vm_mm;
1410 	struct mmu_gather tlb;
1411 	unsigned long end = address + size;
1412 
1413 	lru_add_drain();
1414 	tlb_gather_mmu(&tlb, mm, address, end);
1415 	update_hiwater_rss(mm);
1416 	mmu_notifier_invalidate_range_start(mm, address, end);
1417 	unmap_single_vma(&tlb, vma, address, end, details);
1418 	mmu_notifier_invalidate_range_end(mm, address, end);
1419 	tlb_finish_mmu(&tlb, address, end);
1420 }
1421 
1422 /**
1423  * zap_vma_ptes - remove ptes mapping the vma
1424  * @vma: vm_area_struct holding ptes to be zapped
1425  * @address: starting address of pages to zap
1426  * @size: number of bytes to zap
1427  *
1428  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1429  *
1430  * The entire address range must be fully contained within the vma.
1431  *
1432  * Returns 0 if successful.
1433  */
1434 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1435 		unsigned long size)
1436 {
1437 	if (address < vma->vm_start || address + size > vma->vm_end ||
1438 	    		!(vma->vm_flags & VM_PFNMAP))
1439 		return -1;
1440 	zap_page_range_single(vma, address, size, NULL);
1441 	return 0;
1442 }
1443 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1444 
1445 /**
1446  * follow_page_mask - look up a page descriptor from a user-virtual address
1447  * @vma: vm_area_struct mapping @address
1448  * @address: virtual address to look up
1449  * @flags: flags modifying lookup behaviour
1450  * @page_mask: on output, *page_mask is set according to the size of the page
1451  *
1452  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1453  *
1454  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1455  * an error pointer if there is a mapping to something not represented
1456  * by a page descriptor (see also vm_normal_page()).
1457  */
1458 struct page *follow_page_mask(struct vm_area_struct *vma,
1459 			      unsigned long address, unsigned int flags,
1460 			      unsigned int *page_mask)
1461 {
1462 	pgd_t *pgd;
1463 	pud_t *pud;
1464 	pmd_t *pmd;
1465 	pte_t *ptep, pte;
1466 	spinlock_t *ptl;
1467 	struct page *page;
1468 	struct mm_struct *mm = vma->vm_mm;
1469 
1470 	*page_mask = 0;
1471 
1472 	page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1473 	if (!IS_ERR(page)) {
1474 		BUG_ON(flags & FOLL_GET);
1475 		goto out;
1476 	}
1477 
1478 	page = NULL;
1479 	pgd = pgd_offset(mm, address);
1480 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1481 		goto no_page_table;
1482 
1483 	pud = pud_offset(pgd, address);
1484 	if (pud_none(*pud))
1485 		goto no_page_table;
1486 	if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1487 		if (flags & FOLL_GET)
1488 			goto out;
1489 		page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1490 		goto out;
1491 	}
1492 	if (unlikely(pud_bad(*pud)))
1493 		goto no_page_table;
1494 
1495 	pmd = pmd_offset(pud, address);
1496 	if (pmd_none(*pmd))
1497 		goto no_page_table;
1498 	if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1499 		page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1500 		if (flags & FOLL_GET) {
1501 			/*
1502 			 * Refcount on tail pages are not well-defined and
1503 			 * shouldn't be taken. The caller should handle a NULL
1504 			 * return when trying to follow tail pages.
1505 			 */
1506 			if (PageHead(page))
1507 				get_page(page);
1508 			else {
1509 				page = NULL;
1510 				goto out;
1511 			}
1512 		}
1513 		goto out;
1514 	}
1515 	if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1516 		goto no_page_table;
1517 	if (pmd_trans_huge(*pmd)) {
1518 		if (flags & FOLL_SPLIT) {
1519 			split_huge_page_pmd(vma, address, pmd);
1520 			goto split_fallthrough;
1521 		}
1522 		ptl = pmd_lock(mm, pmd);
1523 		if (likely(pmd_trans_huge(*pmd))) {
1524 			if (unlikely(pmd_trans_splitting(*pmd))) {
1525 				spin_unlock(ptl);
1526 				wait_split_huge_page(vma->anon_vma, pmd);
1527 			} else {
1528 				page = follow_trans_huge_pmd(vma, address,
1529 							     pmd, flags);
1530 				spin_unlock(ptl);
1531 				*page_mask = HPAGE_PMD_NR - 1;
1532 				goto out;
1533 			}
1534 		} else
1535 			spin_unlock(ptl);
1536 		/* fall through */
1537 	}
1538 split_fallthrough:
1539 	if (unlikely(pmd_bad(*pmd)))
1540 		goto no_page_table;
1541 
1542 	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1543 
1544 	pte = *ptep;
1545 	if (!pte_present(pte)) {
1546 		swp_entry_t entry;
1547 		/*
1548 		 * KSM's break_ksm() relies upon recognizing a ksm page
1549 		 * even while it is being migrated, so for that case we
1550 		 * need migration_entry_wait().
1551 		 */
1552 		if (likely(!(flags & FOLL_MIGRATION)))
1553 			goto no_page;
1554 		if (pte_none(pte) || pte_file(pte))
1555 			goto no_page;
1556 		entry = pte_to_swp_entry(pte);
1557 		if (!is_migration_entry(entry))
1558 			goto no_page;
1559 		pte_unmap_unlock(ptep, ptl);
1560 		migration_entry_wait(mm, pmd, address);
1561 		goto split_fallthrough;
1562 	}
1563 	if ((flags & FOLL_NUMA) && pte_numa(pte))
1564 		goto no_page;
1565 	if ((flags & FOLL_WRITE) && !pte_write(pte))
1566 		goto unlock;
1567 
1568 	page = vm_normal_page(vma, address, pte);
1569 	if (unlikely(!page)) {
1570 		if ((flags & FOLL_DUMP) ||
1571 		    !is_zero_pfn(pte_pfn(pte)))
1572 			goto bad_page;
1573 		page = pte_page(pte);
1574 	}
1575 
1576 	if (flags & FOLL_GET)
1577 		get_page_foll(page);
1578 	if (flags & FOLL_TOUCH) {
1579 		if ((flags & FOLL_WRITE) &&
1580 		    !pte_dirty(pte) && !PageDirty(page))
1581 			set_page_dirty(page);
1582 		/*
1583 		 * pte_mkyoung() would be more correct here, but atomic care
1584 		 * is needed to avoid losing the dirty bit: it is easier to use
1585 		 * mark_page_accessed().
1586 		 */
1587 		mark_page_accessed(page);
1588 	}
1589 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1590 		/*
1591 		 * The preliminary mapping check is mainly to avoid the
1592 		 * pointless overhead of lock_page on the ZERO_PAGE
1593 		 * which might bounce very badly if there is contention.
1594 		 *
1595 		 * If the page is already locked, we don't need to
1596 		 * handle it now - vmscan will handle it later if and
1597 		 * when it attempts to reclaim the page.
1598 		 */
1599 		if (page->mapping && trylock_page(page)) {
1600 			lru_add_drain();  /* push cached pages to LRU */
1601 			/*
1602 			 * Because we lock page here, and migration is
1603 			 * blocked by the pte's page reference, and we
1604 			 * know the page is still mapped, we don't even
1605 			 * need to check for file-cache page truncation.
1606 			 */
1607 			mlock_vma_page(page);
1608 			unlock_page(page);
1609 		}
1610 	}
1611 unlock:
1612 	pte_unmap_unlock(ptep, ptl);
1613 out:
1614 	return page;
1615 
1616 bad_page:
1617 	pte_unmap_unlock(ptep, ptl);
1618 	return ERR_PTR(-EFAULT);
1619 
1620 no_page:
1621 	pte_unmap_unlock(ptep, ptl);
1622 	if (!pte_none(pte))
1623 		return page;
1624 
1625 no_page_table:
1626 	/*
1627 	 * When core dumping an enormous anonymous area that nobody
1628 	 * has touched so far, we don't want to allocate unnecessary pages or
1629 	 * page tables.  Return error instead of NULL to skip handle_mm_fault,
1630 	 * then get_dump_page() will return NULL to leave a hole in the dump.
1631 	 * But we can only make this optimization where a hole would surely
1632 	 * be zero-filled if handle_mm_fault() actually did handle it.
1633 	 */
1634 	if ((flags & FOLL_DUMP) &&
1635 	    (!vma->vm_ops || !vma->vm_ops->fault))
1636 		return ERR_PTR(-EFAULT);
1637 	return page;
1638 }
1639 
1640 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1641 {
1642 	return stack_guard_page_start(vma, addr) ||
1643 	       stack_guard_page_end(vma, addr+PAGE_SIZE);
1644 }
1645 
1646 /**
1647  * __get_user_pages() - pin user pages in memory
1648  * @tsk:	task_struct of target task
1649  * @mm:		mm_struct of target mm
1650  * @start:	starting user address
1651  * @nr_pages:	number of pages from start to pin
1652  * @gup_flags:	flags modifying pin behaviour
1653  * @pages:	array that receives pointers to the pages pinned.
1654  *		Should be at least nr_pages long. Or NULL, if caller
1655  *		only intends to ensure the pages are faulted in.
1656  * @vmas:	array of pointers to vmas corresponding to each page.
1657  *		Or NULL if the caller does not require them.
1658  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1659  *
1660  * Returns number of pages pinned. This may be fewer than the number
1661  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1662  * were pinned, returns -errno. Each page returned must be released
1663  * with a put_page() call when it is finished with. vmas will only
1664  * remain valid while mmap_sem is held.
1665  *
1666  * Must be called with mmap_sem held for read or write.
1667  *
1668  * __get_user_pages walks a process's page tables and takes a reference to
1669  * each struct page that each user address corresponds to at a given
1670  * instant. That is, it takes the page that would be accessed if a user
1671  * thread accesses the given user virtual address at that instant.
1672  *
1673  * This does not guarantee that the page exists in the user mappings when
1674  * __get_user_pages returns, and there may even be a completely different
1675  * page there in some cases (eg. if mmapped pagecache has been invalidated
1676  * and subsequently re faulted). However it does guarantee that the page
1677  * won't be freed completely. And mostly callers simply care that the page
1678  * contains data that was valid *at some point in time*. Typically, an IO
1679  * or similar operation cannot guarantee anything stronger anyway because
1680  * locks can't be held over the syscall boundary.
1681  *
1682  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1683  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1684  * appropriate) must be called after the page is finished with, and
1685  * before put_page is called.
1686  *
1687  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1688  * or mmap_sem contention, and if waiting is needed to pin all pages,
1689  * *@nonblocking will be set to 0.
1690  *
1691  * In most cases, get_user_pages or get_user_pages_fast should be used
1692  * instead of __get_user_pages. __get_user_pages should be used only if
1693  * you need some special @gup_flags.
1694  */
1695 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1696 		unsigned long start, unsigned long nr_pages,
1697 		unsigned int gup_flags, struct page **pages,
1698 		struct vm_area_struct **vmas, int *nonblocking)
1699 {
1700 	long i;
1701 	unsigned long vm_flags;
1702 	unsigned int page_mask;
1703 
1704 	if (!nr_pages)
1705 		return 0;
1706 
1707 	VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1708 
1709 	/*
1710 	 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1711 	 * would be called on PROT_NONE ranges. We must never invoke
1712 	 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1713 	 * page faults would unprotect the PROT_NONE ranges if
1714 	 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1715 	 * bitflag. So to avoid that, don't set FOLL_NUMA if
1716 	 * FOLL_FORCE is set.
1717 	 */
1718 	if (!(gup_flags & FOLL_FORCE))
1719 		gup_flags |= FOLL_NUMA;
1720 
1721 	i = 0;
1722 
1723 	do {
1724 		struct vm_area_struct *vma;
1725 
1726 		vma = find_extend_vma(mm, start);
1727 		if (!vma && in_gate_area(mm, start)) {
1728 			unsigned long pg = start & PAGE_MASK;
1729 			pgd_t *pgd;
1730 			pud_t *pud;
1731 			pmd_t *pmd;
1732 			pte_t *pte;
1733 
1734 			/* user gate pages are read-only */
1735 			if (gup_flags & FOLL_WRITE)
1736 				goto efault;
1737 			if (pg > TASK_SIZE)
1738 				pgd = pgd_offset_k(pg);
1739 			else
1740 				pgd = pgd_offset_gate(mm, pg);
1741 			BUG_ON(pgd_none(*pgd));
1742 			pud = pud_offset(pgd, pg);
1743 			BUG_ON(pud_none(*pud));
1744 			pmd = pmd_offset(pud, pg);
1745 			if (pmd_none(*pmd))
1746 				goto efault;
1747 			VM_BUG_ON(pmd_trans_huge(*pmd));
1748 			pte = pte_offset_map(pmd, pg);
1749 			if (pte_none(*pte)) {
1750 				pte_unmap(pte);
1751 				goto efault;
1752 			}
1753 			vma = get_gate_vma(mm);
1754 			if (pages) {
1755 				struct page *page;
1756 
1757 				page = vm_normal_page(vma, start, *pte);
1758 				if (!page) {
1759 					if (!(gup_flags & FOLL_DUMP) &&
1760 					     is_zero_pfn(pte_pfn(*pte)))
1761 						page = pte_page(*pte);
1762 					else {
1763 						pte_unmap(pte);
1764 						goto efault;
1765 					}
1766 				}
1767 				pages[i] = page;
1768 				get_page(page);
1769 			}
1770 			pte_unmap(pte);
1771 			page_mask = 0;
1772 			goto next_page;
1773 		}
1774 
1775 		if (!vma)
1776 			goto efault;
1777 		vm_flags = vma->vm_flags;
1778 		if (vm_flags & (VM_IO | VM_PFNMAP))
1779 			goto efault;
1780 
1781 		if (gup_flags & FOLL_WRITE) {
1782 			if (!(vm_flags & VM_WRITE)) {
1783 				if (!(gup_flags & FOLL_FORCE))
1784 					goto efault;
1785 				/*
1786 				 * We used to let the write,force case do COW
1787 				 * in a VM_MAYWRITE VM_SHARED !VM_WRITE vma, so
1788 				 * ptrace could set a breakpoint in a read-only
1789 				 * mapping of an executable, without corrupting
1790 				 * the file (yet only when that file had been
1791 				 * opened for writing!).  Anon pages in shared
1792 				 * mappings are surprising: now just reject it.
1793 				 */
1794 				if (!is_cow_mapping(vm_flags)) {
1795 					WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
1796 					goto efault;
1797 				}
1798 			}
1799 		} else {
1800 			if (!(vm_flags & VM_READ)) {
1801 				if (!(gup_flags & FOLL_FORCE))
1802 					goto efault;
1803 				/*
1804 				 * Is there actually any vma we can reach here
1805 				 * which does not have VM_MAYREAD set?
1806 				 */
1807 				if (!(vm_flags & VM_MAYREAD))
1808 					goto efault;
1809 			}
1810 		}
1811 
1812 		if (is_vm_hugetlb_page(vma)) {
1813 			i = follow_hugetlb_page(mm, vma, pages, vmas,
1814 					&start, &nr_pages, i, gup_flags);
1815 			continue;
1816 		}
1817 
1818 		do {
1819 			struct page *page;
1820 			unsigned int foll_flags = gup_flags;
1821 			unsigned int page_increm;
1822 
1823 			/*
1824 			 * If we have a pending SIGKILL, don't keep faulting
1825 			 * pages and potentially allocating memory.
1826 			 */
1827 			if (unlikely(fatal_signal_pending(current)))
1828 				return i ? i : -ERESTARTSYS;
1829 
1830 			cond_resched();
1831 			while (!(page = follow_page_mask(vma, start,
1832 						foll_flags, &page_mask))) {
1833 				int ret;
1834 				unsigned int fault_flags = 0;
1835 
1836 				/* For mlock, just skip the stack guard page. */
1837 				if (foll_flags & FOLL_MLOCK) {
1838 					if (stack_guard_page(vma, start))
1839 						goto next_page;
1840 				}
1841 				if (foll_flags & FOLL_WRITE)
1842 					fault_flags |= FAULT_FLAG_WRITE;
1843 				if (nonblocking)
1844 					fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1845 				if (foll_flags & FOLL_NOWAIT)
1846 					fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1847 
1848 				ret = handle_mm_fault(mm, vma, start,
1849 							fault_flags);
1850 
1851 				if (ret & VM_FAULT_ERROR) {
1852 					if (ret & VM_FAULT_OOM)
1853 						return i ? i : -ENOMEM;
1854 					if (ret & (VM_FAULT_HWPOISON |
1855 						   VM_FAULT_HWPOISON_LARGE)) {
1856 						if (i)
1857 							return i;
1858 						else if (gup_flags & FOLL_HWPOISON)
1859 							return -EHWPOISON;
1860 						else
1861 							return -EFAULT;
1862 					}
1863 					if (ret & VM_FAULT_SIGBUS)
1864 						goto efault;
1865 					BUG();
1866 				}
1867 
1868 				if (tsk) {
1869 					if (ret & VM_FAULT_MAJOR)
1870 						tsk->maj_flt++;
1871 					else
1872 						tsk->min_flt++;
1873 				}
1874 
1875 				if (ret & VM_FAULT_RETRY) {
1876 					if (nonblocking)
1877 						*nonblocking = 0;
1878 					return i;
1879 				}
1880 
1881 				/*
1882 				 * The VM_FAULT_WRITE bit tells us that
1883 				 * do_wp_page has broken COW when necessary,
1884 				 * even if maybe_mkwrite decided not to set
1885 				 * pte_write. We can thus safely do subsequent
1886 				 * page lookups as if they were reads. But only
1887 				 * do so when looping for pte_write is futile:
1888 				 * in some cases userspace may also be wanting
1889 				 * to write to the gotten user page, which a
1890 				 * read fault here might prevent (a readonly
1891 				 * page might get reCOWed by userspace write).
1892 				 */
1893 				if ((ret & VM_FAULT_WRITE) &&
1894 				    !(vma->vm_flags & VM_WRITE))
1895 					foll_flags &= ~FOLL_WRITE;
1896 
1897 				cond_resched();
1898 			}
1899 			if (IS_ERR(page))
1900 				return i ? i : PTR_ERR(page);
1901 			if (pages) {
1902 				pages[i] = page;
1903 
1904 				flush_anon_page(vma, page, start);
1905 				flush_dcache_page(page);
1906 				page_mask = 0;
1907 			}
1908 next_page:
1909 			if (vmas) {
1910 				vmas[i] = vma;
1911 				page_mask = 0;
1912 			}
1913 			page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1914 			if (page_increm > nr_pages)
1915 				page_increm = nr_pages;
1916 			i += page_increm;
1917 			start += page_increm * PAGE_SIZE;
1918 			nr_pages -= page_increm;
1919 		} while (nr_pages && start < vma->vm_end);
1920 	} while (nr_pages);
1921 	return i;
1922 efault:
1923 	return i ? : -EFAULT;
1924 }
1925 EXPORT_SYMBOL(__get_user_pages);
1926 
1927 /*
1928  * fixup_user_fault() - manually resolve a user page fault
1929  * @tsk:	the task_struct to use for page fault accounting, or
1930  *		NULL if faults are not to be recorded.
1931  * @mm:		mm_struct of target mm
1932  * @address:	user address
1933  * @fault_flags:flags to pass down to handle_mm_fault()
1934  *
1935  * This is meant to be called in the specific scenario where for locking reasons
1936  * we try to access user memory in atomic context (within a pagefault_disable()
1937  * section), this returns -EFAULT, and we want to resolve the user fault before
1938  * trying again.
1939  *
1940  * Typically this is meant to be used by the futex code.
1941  *
1942  * The main difference with get_user_pages() is that this function will
1943  * unconditionally call handle_mm_fault() which will in turn perform all the
1944  * necessary SW fixup of the dirty and young bits in the PTE, while
1945  * handle_mm_fault() only guarantees to update these in the struct page.
1946  *
1947  * This is important for some architectures where those bits also gate the
1948  * access permission to the page because they are maintained in software.  On
1949  * such architectures, gup() will not be enough to make a subsequent access
1950  * succeed.
1951  *
1952  * This should be called with the mm_sem held for read.
1953  */
1954 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1955 		     unsigned long address, unsigned int fault_flags)
1956 {
1957 	struct vm_area_struct *vma;
1958 	int ret;
1959 
1960 	vma = find_extend_vma(mm, address);
1961 	if (!vma || address < vma->vm_start)
1962 		return -EFAULT;
1963 
1964 	ret = handle_mm_fault(mm, vma, address, fault_flags);
1965 	if (ret & VM_FAULT_ERROR) {
1966 		if (ret & VM_FAULT_OOM)
1967 			return -ENOMEM;
1968 		if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1969 			return -EHWPOISON;
1970 		if (ret & VM_FAULT_SIGBUS)
1971 			return -EFAULT;
1972 		BUG();
1973 	}
1974 	if (tsk) {
1975 		if (ret & VM_FAULT_MAJOR)
1976 			tsk->maj_flt++;
1977 		else
1978 			tsk->min_flt++;
1979 	}
1980 	return 0;
1981 }
1982 
1983 /*
1984  * get_user_pages() - pin user pages in memory
1985  * @tsk:	the task_struct to use for page fault accounting, or
1986  *		NULL if faults are not to be recorded.
1987  * @mm:		mm_struct of target mm
1988  * @start:	starting user address
1989  * @nr_pages:	number of pages from start to pin
1990  * @write:	whether pages will be written to by the caller
1991  * @force:	whether to force access even when user mapping is currently
1992  *		protected (but never forces write access to shared mapping).
1993  * @pages:	array that receives pointers to the pages pinned.
1994  *		Should be at least nr_pages long. Or NULL, if caller
1995  *		only intends to ensure the pages are faulted in.
1996  * @vmas:	array of pointers to vmas corresponding to each page.
1997  *		Or NULL if the caller does not require them.
1998  *
1999  * Returns number of pages pinned. This may be fewer than the number
2000  * requested. If nr_pages is 0 or negative, returns 0. If no pages
2001  * were pinned, returns -errno. Each page returned must be released
2002  * with a put_page() call when it is finished with. vmas will only
2003  * remain valid while mmap_sem is held.
2004  *
2005  * Must be called with mmap_sem held for read or write.
2006  *
2007  * get_user_pages walks a process's page tables and takes a reference to
2008  * each struct page that each user address corresponds to at a given
2009  * instant. That is, it takes the page that would be accessed if a user
2010  * thread accesses the given user virtual address at that instant.
2011  *
2012  * This does not guarantee that the page exists in the user mappings when
2013  * get_user_pages returns, and there may even be a completely different
2014  * page there in some cases (eg. if mmapped pagecache has been invalidated
2015  * and subsequently re faulted). However it does guarantee that the page
2016  * won't be freed completely. And mostly callers simply care that the page
2017  * contains data that was valid *at some point in time*. Typically, an IO
2018  * or similar operation cannot guarantee anything stronger anyway because
2019  * locks can't be held over the syscall boundary.
2020  *
2021  * If write=0, the page must not be written to. If the page is written to,
2022  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2023  * after the page is finished with, and before put_page is called.
2024  *
2025  * get_user_pages is typically used for fewer-copy IO operations, to get a
2026  * handle on the memory by some means other than accesses via the user virtual
2027  * addresses. The pages may be submitted for DMA to devices or accessed via
2028  * their kernel linear mapping (via the kmap APIs). Care should be taken to
2029  * use the correct cache flushing APIs.
2030  *
2031  * See also get_user_pages_fast, for performance critical applications.
2032  */
2033 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2034 		unsigned long start, unsigned long nr_pages, int write,
2035 		int force, struct page **pages, struct vm_area_struct **vmas)
2036 {
2037 	int flags = FOLL_TOUCH;
2038 
2039 	if (pages)
2040 		flags |= FOLL_GET;
2041 	if (write)
2042 		flags |= FOLL_WRITE;
2043 	if (force)
2044 		flags |= FOLL_FORCE;
2045 
2046 	return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2047 				NULL);
2048 }
2049 EXPORT_SYMBOL(get_user_pages);
2050 
2051 /**
2052  * get_dump_page() - pin user page in memory while writing it to core dump
2053  * @addr: user address
2054  *
2055  * Returns struct page pointer of user page pinned for dump,
2056  * to be freed afterwards by page_cache_release() or put_page().
2057  *
2058  * Returns NULL on any kind of failure - a hole must then be inserted into
2059  * the corefile, to preserve alignment with its headers; and also returns
2060  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2061  * allowing a hole to be left in the corefile to save diskspace.
2062  *
2063  * Called without mmap_sem, but after all other threads have been killed.
2064  */
2065 #ifdef CONFIG_ELF_CORE
2066 struct page *get_dump_page(unsigned long addr)
2067 {
2068 	struct vm_area_struct *vma;
2069 	struct page *page;
2070 
2071 	if (__get_user_pages(current, current->mm, addr, 1,
2072 			     FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2073 			     NULL) < 1)
2074 		return NULL;
2075 	flush_cache_page(vma, addr, page_to_pfn(page));
2076 	return page;
2077 }
2078 #endif /* CONFIG_ELF_CORE */
2079 
2080 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2081 			spinlock_t **ptl)
2082 {
2083 	pgd_t * pgd = pgd_offset(mm, addr);
2084 	pud_t * pud = pud_alloc(mm, pgd, addr);
2085 	if (pud) {
2086 		pmd_t * pmd = pmd_alloc(mm, pud, addr);
2087 		if (pmd) {
2088 			VM_BUG_ON(pmd_trans_huge(*pmd));
2089 			return pte_alloc_map_lock(mm, pmd, addr, ptl);
2090 		}
2091 	}
2092 	return NULL;
2093 }
2094 
2095 /*
2096  * This is the old fallback for page remapping.
2097  *
2098  * For historical reasons, it only allows reserved pages. Only
2099  * old drivers should use this, and they needed to mark their
2100  * pages reserved for the old functions anyway.
2101  */
2102 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2103 			struct page *page, pgprot_t prot)
2104 {
2105 	struct mm_struct *mm = vma->vm_mm;
2106 	int retval;
2107 	pte_t *pte;
2108 	spinlock_t *ptl;
2109 
2110 	retval = -EINVAL;
2111 	if (PageAnon(page))
2112 		goto out;
2113 	retval = -ENOMEM;
2114 	flush_dcache_page(page);
2115 	pte = get_locked_pte(mm, addr, &ptl);
2116 	if (!pte)
2117 		goto out;
2118 	retval = -EBUSY;
2119 	if (!pte_none(*pte))
2120 		goto out_unlock;
2121 
2122 	/* Ok, finally just insert the thing.. */
2123 	get_page(page);
2124 	inc_mm_counter_fast(mm, MM_FILEPAGES);
2125 	page_add_file_rmap(page);
2126 	set_pte_at(mm, addr, pte, mk_pte(page, prot));
2127 
2128 	retval = 0;
2129 	pte_unmap_unlock(pte, ptl);
2130 	return retval;
2131 out_unlock:
2132 	pte_unmap_unlock(pte, ptl);
2133 out:
2134 	return retval;
2135 }
2136 
2137 /**
2138  * vm_insert_page - insert single page into user vma
2139  * @vma: user vma to map to
2140  * @addr: target user address of this page
2141  * @page: source kernel page
2142  *
2143  * This allows drivers to insert individual pages they've allocated
2144  * into a user vma.
2145  *
2146  * The page has to be a nice clean _individual_ kernel allocation.
2147  * If you allocate a compound page, you need to have marked it as
2148  * such (__GFP_COMP), or manually just split the page up yourself
2149  * (see split_page()).
2150  *
2151  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2152  * took an arbitrary page protection parameter. This doesn't allow
2153  * that. Your vma protection will have to be set up correctly, which
2154  * means that if you want a shared writable mapping, you'd better
2155  * ask for a shared writable mapping!
2156  *
2157  * The page does not need to be reserved.
2158  *
2159  * Usually this function is called from f_op->mmap() handler
2160  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2161  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2162  * function from other places, for example from page-fault handler.
2163  */
2164 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2165 			struct page *page)
2166 {
2167 	if (addr < vma->vm_start || addr >= vma->vm_end)
2168 		return -EFAULT;
2169 	if (!page_count(page))
2170 		return -EINVAL;
2171 	if (!(vma->vm_flags & VM_MIXEDMAP)) {
2172 		BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2173 		BUG_ON(vma->vm_flags & VM_PFNMAP);
2174 		vma->vm_flags |= VM_MIXEDMAP;
2175 	}
2176 	return insert_page(vma, addr, page, vma->vm_page_prot);
2177 }
2178 EXPORT_SYMBOL(vm_insert_page);
2179 
2180 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2181 			unsigned long pfn, pgprot_t prot)
2182 {
2183 	struct mm_struct *mm = vma->vm_mm;
2184 	int retval;
2185 	pte_t *pte, entry;
2186 	spinlock_t *ptl;
2187 
2188 	retval = -ENOMEM;
2189 	pte = get_locked_pte(mm, addr, &ptl);
2190 	if (!pte)
2191 		goto out;
2192 	retval = -EBUSY;
2193 	if (!pte_none(*pte))
2194 		goto out_unlock;
2195 
2196 	/* Ok, finally just insert the thing.. */
2197 	entry = pte_mkspecial(pfn_pte(pfn, prot));
2198 	set_pte_at(mm, addr, pte, entry);
2199 	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2200 
2201 	retval = 0;
2202 out_unlock:
2203 	pte_unmap_unlock(pte, ptl);
2204 out:
2205 	return retval;
2206 }
2207 
2208 /**
2209  * vm_insert_pfn - insert single pfn into user vma
2210  * @vma: user vma to map to
2211  * @addr: target user address of this page
2212  * @pfn: source kernel pfn
2213  *
2214  * Similar to vm_insert_page, this allows drivers to insert individual pages
2215  * they've allocated into a user vma. Same comments apply.
2216  *
2217  * This function should only be called from a vm_ops->fault handler, and
2218  * in that case the handler should return NULL.
2219  *
2220  * vma cannot be a COW mapping.
2221  *
2222  * As this is called only for pages that do not currently exist, we
2223  * do not need to flush old virtual caches or the TLB.
2224  */
2225 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2226 			unsigned long pfn)
2227 {
2228 	int ret;
2229 	pgprot_t pgprot = vma->vm_page_prot;
2230 	/*
2231 	 * Technically, architectures with pte_special can avoid all these
2232 	 * restrictions (same for remap_pfn_range).  However we would like
2233 	 * consistency in testing and feature parity among all, so we should
2234 	 * try to keep these invariants in place for everybody.
2235 	 */
2236 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2237 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2238 						(VM_PFNMAP|VM_MIXEDMAP));
2239 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2240 	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2241 
2242 	if (addr < vma->vm_start || addr >= vma->vm_end)
2243 		return -EFAULT;
2244 	if (track_pfn_insert(vma, &pgprot, pfn))
2245 		return -EINVAL;
2246 
2247 	ret = insert_pfn(vma, addr, pfn, pgprot);
2248 
2249 	return ret;
2250 }
2251 EXPORT_SYMBOL(vm_insert_pfn);
2252 
2253 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2254 			unsigned long pfn)
2255 {
2256 	BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2257 
2258 	if (addr < vma->vm_start || addr >= vma->vm_end)
2259 		return -EFAULT;
2260 
2261 	/*
2262 	 * If we don't have pte special, then we have to use the pfn_valid()
2263 	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2264 	 * refcount the page if pfn_valid is true (hence insert_page rather
2265 	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2266 	 * without pte special, it would there be refcounted as a normal page.
2267 	 */
2268 	if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2269 		struct page *page;
2270 
2271 		page = pfn_to_page(pfn);
2272 		return insert_page(vma, addr, page, vma->vm_page_prot);
2273 	}
2274 	return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2275 }
2276 EXPORT_SYMBOL(vm_insert_mixed);
2277 
2278 /*
2279  * maps a range of physical memory into the requested pages. the old
2280  * mappings are removed. any references to nonexistent pages results
2281  * in null mappings (currently treated as "copy-on-access")
2282  */
2283 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2284 			unsigned long addr, unsigned long end,
2285 			unsigned long pfn, pgprot_t prot)
2286 {
2287 	pte_t *pte;
2288 	spinlock_t *ptl;
2289 
2290 	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2291 	if (!pte)
2292 		return -ENOMEM;
2293 	arch_enter_lazy_mmu_mode();
2294 	do {
2295 		BUG_ON(!pte_none(*pte));
2296 		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2297 		pfn++;
2298 	} while (pte++, addr += PAGE_SIZE, addr != end);
2299 	arch_leave_lazy_mmu_mode();
2300 	pte_unmap_unlock(pte - 1, ptl);
2301 	return 0;
2302 }
2303 
2304 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2305 			unsigned long addr, unsigned long end,
2306 			unsigned long pfn, pgprot_t prot)
2307 {
2308 	pmd_t *pmd;
2309 	unsigned long next;
2310 
2311 	pfn -= addr >> PAGE_SHIFT;
2312 	pmd = pmd_alloc(mm, pud, addr);
2313 	if (!pmd)
2314 		return -ENOMEM;
2315 	VM_BUG_ON(pmd_trans_huge(*pmd));
2316 	do {
2317 		next = pmd_addr_end(addr, end);
2318 		if (remap_pte_range(mm, pmd, addr, next,
2319 				pfn + (addr >> PAGE_SHIFT), prot))
2320 			return -ENOMEM;
2321 	} while (pmd++, addr = next, addr != end);
2322 	return 0;
2323 }
2324 
2325 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2326 			unsigned long addr, unsigned long end,
2327 			unsigned long pfn, pgprot_t prot)
2328 {
2329 	pud_t *pud;
2330 	unsigned long next;
2331 
2332 	pfn -= addr >> PAGE_SHIFT;
2333 	pud = pud_alloc(mm, pgd, addr);
2334 	if (!pud)
2335 		return -ENOMEM;
2336 	do {
2337 		next = pud_addr_end(addr, end);
2338 		if (remap_pmd_range(mm, pud, addr, next,
2339 				pfn + (addr >> PAGE_SHIFT), prot))
2340 			return -ENOMEM;
2341 	} while (pud++, addr = next, addr != end);
2342 	return 0;
2343 }
2344 
2345 /**
2346  * remap_pfn_range - remap kernel memory to userspace
2347  * @vma: user vma to map to
2348  * @addr: target user address to start at
2349  * @pfn: physical address of kernel memory
2350  * @size: size of map area
2351  * @prot: page protection flags for this mapping
2352  *
2353  *  Note: this is only safe if the mm semaphore is held when called.
2354  */
2355 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2356 		    unsigned long pfn, unsigned long size, pgprot_t prot)
2357 {
2358 	pgd_t *pgd;
2359 	unsigned long next;
2360 	unsigned long end = addr + PAGE_ALIGN(size);
2361 	struct mm_struct *mm = vma->vm_mm;
2362 	int err;
2363 
2364 	/*
2365 	 * Physically remapped pages are special. Tell the
2366 	 * rest of the world about it:
2367 	 *   VM_IO tells people not to look at these pages
2368 	 *	(accesses can have side effects).
2369 	 *   VM_PFNMAP tells the core MM that the base pages are just
2370 	 *	raw PFN mappings, and do not have a "struct page" associated
2371 	 *	with them.
2372 	 *   VM_DONTEXPAND
2373 	 *      Disable vma merging and expanding with mremap().
2374 	 *   VM_DONTDUMP
2375 	 *      Omit vma from core dump, even when VM_IO turned off.
2376 	 *
2377 	 * There's a horrible special case to handle copy-on-write
2378 	 * behaviour that some programs depend on. We mark the "original"
2379 	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2380 	 * See vm_normal_page() for details.
2381 	 */
2382 	if (is_cow_mapping(vma->vm_flags)) {
2383 		if (addr != vma->vm_start || end != vma->vm_end)
2384 			return -EINVAL;
2385 		vma->vm_pgoff = pfn;
2386 	}
2387 
2388 	err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2389 	if (err)
2390 		return -EINVAL;
2391 
2392 	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2393 
2394 	BUG_ON(addr >= end);
2395 	pfn -= addr >> PAGE_SHIFT;
2396 	pgd = pgd_offset(mm, addr);
2397 	flush_cache_range(vma, addr, end);
2398 	do {
2399 		next = pgd_addr_end(addr, end);
2400 		err = remap_pud_range(mm, pgd, addr, next,
2401 				pfn + (addr >> PAGE_SHIFT), prot);
2402 		if (err)
2403 			break;
2404 	} while (pgd++, addr = next, addr != end);
2405 
2406 	if (err)
2407 		untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2408 
2409 	return err;
2410 }
2411 EXPORT_SYMBOL(remap_pfn_range);
2412 
2413 /**
2414  * vm_iomap_memory - remap memory to userspace
2415  * @vma: user vma to map to
2416  * @start: start of area
2417  * @len: size of area
2418  *
2419  * This is a simplified io_remap_pfn_range() for common driver use. The
2420  * driver just needs to give us the physical memory range to be mapped,
2421  * we'll figure out the rest from the vma information.
2422  *
2423  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2424  * whatever write-combining details or similar.
2425  */
2426 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2427 {
2428 	unsigned long vm_len, pfn, pages;
2429 
2430 	/* Check that the physical memory area passed in looks valid */
2431 	if (start + len < start)
2432 		return -EINVAL;
2433 	/*
2434 	 * You *really* shouldn't map things that aren't page-aligned,
2435 	 * but we've historically allowed it because IO memory might
2436 	 * just have smaller alignment.
2437 	 */
2438 	len += start & ~PAGE_MASK;
2439 	pfn = start >> PAGE_SHIFT;
2440 	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2441 	if (pfn + pages < pfn)
2442 		return -EINVAL;
2443 
2444 	/* We start the mapping 'vm_pgoff' pages into the area */
2445 	if (vma->vm_pgoff > pages)
2446 		return -EINVAL;
2447 	pfn += vma->vm_pgoff;
2448 	pages -= vma->vm_pgoff;
2449 
2450 	/* Can we fit all of the mapping? */
2451 	vm_len = vma->vm_end - vma->vm_start;
2452 	if (vm_len >> PAGE_SHIFT > pages)
2453 		return -EINVAL;
2454 
2455 	/* Ok, let it rip */
2456 	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2457 }
2458 EXPORT_SYMBOL(vm_iomap_memory);
2459 
2460 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2461 				     unsigned long addr, unsigned long end,
2462 				     pte_fn_t fn, void *data)
2463 {
2464 	pte_t *pte;
2465 	int err;
2466 	pgtable_t token;
2467 	spinlock_t *uninitialized_var(ptl);
2468 
2469 	pte = (mm == &init_mm) ?
2470 		pte_alloc_kernel(pmd, addr) :
2471 		pte_alloc_map_lock(mm, pmd, addr, &ptl);
2472 	if (!pte)
2473 		return -ENOMEM;
2474 
2475 	BUG_ON(pmd_huge(*pmd));
2476 
2477 	arch_enter_lazy_mmu_mode();
2478 
2479 	token = pmd_pgtable(*pmd);
2480 
2481 	do {
2482 		err = fn(pte++, token, addr, data);
2483 		if (err)
2484 			break;
2485 	} while (addr += PAGE_SIZE, addr != end);
2486 
2487 	arch_leave_lazy_mmu_mode();
2488 
2489 	if (mm != &init_mm)
2490 		pte_unmap_unlock(pte-1, ptl);
2491 	return err;
2492 }
2493 
2494 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2495 				     unsigned long addr, unsigned long end,
2496 				     pte_fn_t fn, void *data)
2497 {
2498 	pmd_t *pmd;
2499 	unsigned long next;
2500 	int err;
2501 
2502 	BUG_ON(pud_huge(*pud));
2503 
2504 	pmd = pmd_alloc(mm, pud, addr);
2505 	if (!pmd)
2506 		return -ENOMEM;
2507 	do {
2508 		next = pmd_addr_end(addr, end);
2509 		err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2510 		if (err)
2511 			break;
2512 	} while (pmd++, addr = next, addr != end);
2513 	return err;
2514 }
2515 
2516 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2517 				     unsigned long addr, unsigned long end,
2518 				     pte_fn_t fn, void *data)
2519 {
2520 	pud_t *pud;
2521 	unsigned long next;
2522 	int err;
2523 
2524 	pud = pud_alloc(mm, pgd, addr);
2525 	if (!pud)
2526 		return -ENOMEM;
2527 	do {
2528 		next = pud_addr_end(addr, end);
2529 		err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2530 		if (err)
2531 			break;
2532 	} while (pud++, addr = next, addr != end);
2533 	return err;
2534 }
2535 
2536 /*
2537  * Scan a region of virtual memory, filling in page tables as necessary
2538  * and calling a provided function on each leaf page table.
2539  */
2540 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2541 			unsigned long size, pte_fn_t fn, void *data)
2542 {
2543 	pgd_t *pgd;
2544 	unsigned long next;
2545 	unsigned long end = addr + size;
2546 	int err;
2547 
2548 	BUG_ON(addr >= end);
2549 	pgd = pgd_offset(mm, addr);
2550 	do {
2551 		next = pgd_addr_end(addr, end);
2552 		err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2553 		if (err)
2554 			break;
2555 	} while (pgd++, addr = next, addr != end);
2556 
2557 	return err;
2558 }
2559 EXPORT_SYMBOL_GPL(apply_to_page_range);
2560 
2561 /*
2562  * handle_pte_fault chooses page fault handler according to an entry
2563  * which was read non-atomically.  Before making any commitment, on
2564  * those architectures or configurations (e.g. i386 with PAE) which
2565  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2566  * must check under lock before unmapping the pte and proceeding
2567  * (but do_wp_page is only called after already making such a check;
2568  * and do_anonymous_page can safely check later on).
2569  */
2570 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2571 				pte_t *page_table, pte_t orig_pte)
2572 {
2573 	int same = 1;
2574 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2575 	if (sizeof(pte_t) > sizeof(unsigned long)) {
2576 		spinlock_t *ptl = pte_lockptr(mm, pmd);
2577 		spin_lock(ptl);
2578 		same = pte_same(*page_table, orig_pte);
2579 		spin_unlock(ptl);
2580 	}
2581 #endif
2582 	pte_unmap(page_table);
2583 	return same;
2584 }
2585 
2586 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2587 {
2588 	debug_dma_assert_idle(src);
2589 
2590 	/*
2591 	 * If the source page was a PFN mapping, we don't have
2592 	 * a "struct page" for it. We do a best-effort copy by
2593 	 * just copying from the original user address. If that
2594 	 * fails, we just zero-fill it. Live with it.
2595 	 */
2596 	if (unlikely(!src)) {
2597 		void *kaddr = kmap_atomic(dst);
2598 		void __user *uaddr = (void __user *)(va & PAGE_MASK);
2599 
2600 		/*
2601 		 * This really shouldn't fail, because the page is there
2602 		 * in the page tables. But it might just be unreadable,
2603 		 * in which case we just give up and fill the result with
2604 		 * zeroes.
2605 		 */
2606 		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2607 			clear_page(kaddr);
2608 		kunmap_atomic(kaddr);
2609 		flush_dcache_page(dst);
2610 	} else
2611 		copy_user_highpage(dst, src, va, vma);
2612 }
2613 
2614 /*
2615  * Notify the address space that the page is about to become writable so that
2616  * it can prohibit this or wait for the page to get into an appropriate state.
2617  *
2618  * We do this without the lock held, so that it can sleep if it needs to.
2619  */
2620 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2621 	       unsigned long address)
2622 {
2623 	struct vm_fault vmf;
2624 	int ret;
2625 
2626 	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2627 	vmf.pgoff = page->index;
2628 	vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2629 	vmf.page = page;
2630 
2631 	ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2632 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2633 		return ret;
2634 	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2635 		lock_page(page);
2636 		if (!page->mapping) {
2637 			unlock_page(page);
2638 			return 0; /* retry */
2639 		}
2640 		ret |= VM_FAULT_LOCKED;
2641 	} else
2642 		VM_BUG_ON_PAGE(!PageLocked(page), page);
2643 	return ret;
2644 }
2645 
2646 /*
2647  * This routine handles present pages, when users try to write
2648  * to a shared page. It is done by copying the page to a new address
2649  * and decrementing the shared-page counter for the old page.
2650  *
2651  * Note that this routine assumes that the protection checks have been
2652  * done by the caller (the low-level page fault routine in most cases).
2653  * Thus we can safely just mark it writable once we've done any necessary
2654  * COW.
2655  *
2656  * We also mark the page dirty at this point even though the page will
2657  * change only once the write actually happens. This avoids a few races,
2658  * and potentially makes it more efficient.
2659  *
2660  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2661  * but allow concurrent faults), with pte both mapped and locked.
2662  * We return with mmap_sem still held, but pte unmapped and unlocked.
2663  */
2664 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2665 		unsigned long address, pte_t *page_table, pmd_t *pmd,
2666 		spinlock_t *ptl, pte_t orig_pte)
2667 	__releases(ptl)
2668 {
2669 	struct page *old_page, *new_page = NULL;
2670 	pte_t entry;
2671 	int ret = 0;
2672 	int page_mkwrite = 0;
2673 	struct page *dirty_page = NULL;
2674 	unsigned long mmun_start = 0;	/* For mmu_notifiers */
2675 	unsigned long mmun_end = 0;	/* For mmu_notifiers */
2676 
2677 	old_page = vm_normal_page(vma, address, orig_pte);
2678 	if (!old_page) {
2679 		/*
2680 		 * VM_MIXEDMAP !pfn_valid() case
2681 		 *
2682 		 * We should not cow pages in a shared writeable mapping.
2683 		 * Just mark the pages writable as we can't do any dirty
2684 		 * accounting on raw pfn maps.
2685 		 */
2686 		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2687 				     (VM_WRITE|VM_SHARED))
2688 			goto reuse;
2689 		goto gotten;
2690 	}
2691 
2692 	/*
2693 	 * Take out anonymous pages first, anonymous shared vmas are
2694 	 * not dirty accountable.
2695 	 */
2696 	if (PageAnon(old_page) && !PageKsm(old_page)) {
2697 		if (!trylock_page(old_page)) {
2698 			page_cache_get(old_page);
2699 			pte_unmap_unlock(page_table, ptl);
2700 			lock_page(old_page);
2701 			page_table = pte_offset_map_lock(mm, pmd, address,
2702 							 &ptl);
2703 			if (!pte_same(*page_table, orig_pte)) {
2704 				unlock_page(old_page);
2705 				goto unlock;
2706 			}
2707 			page_cache_release(old_page);
2708 		}
2709 		if (reuse_swap_page(old_page)) {
2710 			/*
2711 			 * The page is all ours.  Move it to our anon_vma so
2712 			 * the rmap code will not search our parent or siblings.
2713 			 * Protected against the rmap code by the page lock.
2714 			 */
2715 			page_move_anon_rmap(old_page, vma, address);
2716 			unlock_page(old_page);
2717 			goto reuse;
2718 		}
2719 		unlock_page(old_page);
2720 	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2721 					(VM_WRITE|VM_SHARED))) {
2722 		/*
2723 		 * Only catch write-faults on shared writable pages,
2724 		 * read-only shared pages can get COWed by
2725 		 * get_user_pages(.write=1, .force=1).
2726 		 */
2727 		if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2728 			int tmp;
2729 			page_cache_get(old_page);
2730 			pte_unmap_unlock(page_table, ptl);
2731 			tmp = do_page_mkwrite(vma, old_page, address);
2732 			if (unlikely(!tmp || (tmp &
2733 					(VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2734 				page_cache_release(old_page);
2735 				return tmp;
2736 			}
2737 			/*
2738 			 * Since we dropped the lock we need to revalidate
2739 			 * the PTE as someone else may have changed it.  If
2740 			 * they did, we just return, as we can count on the
2741 			 * MMU to tell us if they didn't also make it writable.
2742 			 */
2743 			page_table = pte_offset_map_lock(mm, pmd, address,
2744 							 &ptl);
2745 			if (!pte_same(*page_table, orig_pte)) {
2746 				unlock_page(old_page);
2747 				goto unlock;
2748 			}
2749 
2750 			page_mkwrite = 1;
2751 		}
2752 		dirty_page = old_page;
2753 		get_page(dirty_page);
2754 
2755 reuse:
2756 		/*
2757 		 * Clear the pages cpupid information as the existing
2758 		 * information potentially belongs to a now completely
2759 		 * unrelated process.
2760 		 */
2761 		if (old_page)
2762 			page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2763 
2764 		flush_cache_page(vma, address, pte_pfn(orig_pte));
2765 		entry = pte_mkyoung(orig_pte);
2766 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2767 		if (ptep_set_access_flags(vma, address, page_table, entry,1))
2768 			update_mmu_cache(vma, address, page_table);
2769 		pte_unmap_unlock(page_table, ptl);
2770 		ret |= VM_FAULT_WRITE;
2771 
2772 		if (!dirty_page)
2773 			return ret;
2774 
2775 		/*
2776 		 * Yes, Virginia, this is actually required to prevent a race
2777 		 * with clear_page_dirty_for_io() from clearing the page dirty
2778 		 * bit after it clear all dirty ptes, but before a racing
2779 		 * do_wp_page installs a dirty pte.
2780 		 *
2781 		 * do_shared_fault is protected similarly.
2782 		 */
2783 		if (!page_mkwrite) {
2784 			wait_on_page_locked(dirty_page);
2785 			set_page_dirty_balance(dirty_page);
2786 			/* file_update_time outside page_lock */
2787 			if (vma->vm_file)
2788 				file_update_time(vma->vm_file);
2789 		}
2790 		put_page(dirty_page);
2791 		if (page_mkwrite) {
2792 			struct address_space *mapping = dirty_page->mapping;
2793 
2794 			set_page_dirty(dirty_page);
2795 			unlock_page(dirty_page);
2796 			page_cache_release(dirty_page);
2797 			if (mapping)	{
2798 				/*
2799 				 * Some device drivers do not set page.mapping
2800 				 * but still dirty their pages
2801 				 */
2802 				balance_dirty_pages_ratelimited(mapping);
2803 			}
2804 		}
2805 
2806 		return ret;
2807 	}
2808 
2809 	/*
2810 	 * Ok, we need to copy. Oh, well..
2811 	 */
2812 	page_cache_get(old_page);
2813 gotten:
2814 	pte_unmap_unlock(page_table, ptl);
2815 
2816 	if (unlikely(anon_vma_prepare(vma)))
2817 		goto oom;
2818 
2819 	if (is_zero_pfn(pte_pfn(orig_pte))) {
2820 		new_page = alloc_zeroed_user_highpage_movable(vma, address);
2821 		if (!new_page)
2822 			goto oom;
2823 	} else {
2824 		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2825 		if (!new_page)
2826 			goto oom;
2827 		cow_user_page(new_page, old_page, address, vma);
2828 	}
2829 	__SetPageUptodate(new_page);
2830 
2831 	if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))
2832 		goto oom_free_new;
2833 
2834 	mmun_start  = address & PAGE_MASK;
2835 	mmun_end    = mmun_start + PAGE_SIZE;
2836 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2837 
2838 	/*
2839 	 * Re-check the pte - we dropped the lock
2840 	 */
2841 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2842 	if (likely(pte_same(*page_table, orig_pte))) {
2843 		if (old_page) {
2844 			if (!PageAnon(old_page)) {
2845 				dec_mm_counter_fast(mm, MM_FILEPAGES);
2846 				inc_mm_counter_fast(mm, MM_ANONPAGES);
2847 			}
2848 		} else
2849 			inc_mm_counter_fast(mm, MM_ANONPAGES);
2850 		flush_cache_page(vma, address, pte_pfn(orig_pte));
2851 		entry = mk_pte(new_page, vma->vm_page_prot);
2852 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2853 		/*
2854 		 * Clear the pte entry and flush it first, before updating the
2855 		 * pte with the new entry. This will avoid a race condition
2856 		 * seen in the presence of one thread doing SMC and another
2857 		 * thread doing COW.
2858 		 */
2859 		ptep_clear_flush(vma, address, page_table);
2860 		page_add_new_anon_rmap(new_page, vma, address);
2861 		/*
2862 		 * We call the notify macro here because, when using secondary
2863 		 * mmu page tables (such as kvm shadow page tables), we want the
2864 		 * new page to be mapped directly into the secondary page table.
2865 		 */
2866 		set_pte_at_notify(mm, address, page_table, entry);
2867 		update_mmu_cache(vma, address, page_table);
2868 		if (old_page) {
2869 			/*
2870 			 * Only after switching the pte to the new page may
2871 			 * we remove the mapcount here. Otherwise another
2872 			 * process may come and find the rmap count decremented
2873 			 * before the pte is switched to the new page, and
2874 			 * "reuse" the old page writing into it while our pte
2875 			 * here still points into it and can be read by other
2876 			 * threads.
2877 			 *
2878 			 * The critical issue is to order this
2879 			 * page_remove_rmap with the ptp_clear_flush above.
2880 			 * Those stores are ordered by (if nothing else,)
2881 			 * the barrier present in the atomic_add_negative
2882 			 * in page_remove_rmap.
2883 			 *
2884 			 * Then the TLB flush in ptep_clear_flush ensures that
2885 			 * no process can access the old page before the
2886 			 * decremented mapcount is visible. And the old page
2887 			 * cannot be reused until after the decremented
2888 			 * mapcount is visible. So transitively, TLBs to
2889 			 * old page will be flushed before it can be reused.
2890 			 */
2891 			page_remove_rmap(old_page);
2892 		}
2893 
2894 		/* Free the old page.. */
2895 		new_page = old_page;
2896 		ret |= VM_FAULT_WRITE;
2897 	} else
2898 		mem_cgroup_uncharge_page(new_page);
2899 
2900 	if (new_page)
2901 		page_cache_release(new_page);
2902 unlock:
2903 	pte_unmap_unlock(page_table, ptl);
2904 	if (mmun_end > mmun_start)
2905 		mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2906 	if (old_page) {
2907 		/*
2908 		 * Don't let another task, with possibly unlocked vma,
2909 		 * keep the mlocked page.
2910 		 */
2911 		if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2912 			lock_page(old_page);	/* LRU manipulation */
2913 			munlock_vma_page(old_page);
2914 			unlock_page(old_page);
2915 		}
2916 		page_cache_release(old_page);
2917 	}
2918 	return ret;
2919 oom_free_new:
2920 	page_cache_release(new_page);
2921 oom:
2922 	if (old_page)
2923 		page_cache_release(old_page);
2924 	return VM_FAULT_OOM;
2925 }
2926 
2927 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2928 		unsigned long start_addr, unsigned long end_addr,
2929 		struct zap_details *details)
2930 {
2931 	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2932 }
2933 
2934 static inline void unmap_mapping_range_tree(struct rb_root *root,
2935 					    struct zap_details *details)
2936 {
2937 	struct vm_area_struct *vma;
2938 	pgoff_t vba, vea, zba, zea;
2939 
2940 	vma_interval_tree_foreach(vma, root,
2941 			details->first_index, details->last_index) {
2942 
2943 		vba = vma->vm_pgoff;
2944 		vea = vba + vma_pages(vma) - 1;
2945 		/* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2946 		zba = details->first_index;
2947 		if (zba < vba)
2948 			zba = vba;
2949 		zea = details->last_index;
2950 		if (zea > vea)
2951 			zea = vea;
2952 
2953 		unmap_mapping_range_vma(vma,
2954 			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2955 			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2956 				details);
2957 	}
2958 }
2959 
2960 static inline void unmap_mapping_range_list(struct list_head *head,
2961 					    struct zap_details *details)
2962 {
2963 	struct vm_area_struct *vma;
2964 
2965 	/*
2966 	 * In nonlinear VMAs there is no correspondence between virtual address
2967 	 * offset and file offset.  So we must perform an exhaustive search
2968 	 * across *all* the pages in each nonlinear VMA, not just the pages
2969 	 * whose virtual address lies outside the file truncation point.
2970 	 */
2971 	list_for_each_entry(vma, head, shared.nonlinear) {
2972 		details->nonlinear_vma = vma;
2973 		unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2974 	}
2975 }
2976 
2977 /**
2978  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2979  * @mapping: the address space containing mmaps to be unmapped.
2980  * @holebegin: byte in first page to unmap, relative to the start of
2981  * the underlying file.  This will be rounded down to a PAGE_SIZE
2982  * boundary.  Note that this is different from truncate_pagecache(), which
2983  * must keep the partial page.  In contrast, we must get rid of
2984  * partial pages.
2985  * @holelen: size of prospective hole in bytes.  This will be rounded
2986  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2987  * end of the file.
2988  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2989  * but 0 when invalidating pagecache, don't throw away private data.
2990  */
2991 void unmap_mapping_range(struct address_space *mapping,
2992 		loff_t const holebegin, loff_t const holelen, int even_cows)
2993 {
2994 	struct zap_details details;
2995 	pgoff_t hba = holebegin >> PAGE_SHIFT;
2996 	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2997 
2998 	/* Check for overflow. */
2999 	if (sizeof(holelen) > sizeof(hlen)) {
3000 		long long holeend =
3001 			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3002 		if (holeend & ~(long long)ULONG_MAX)
3003 			hlen = ULONG_MAX - hba + 1;
3004 	}
3005 
3006 	details.check_mapping = even_cows? NULL: mapping;
3007 	details.nonlinear_vma = NULL;
3008 	details.first_index = hba;
3009 	details.last_index = hba + hlen - 1;
3010 	if (details.last_index < details.first_index)
3011 		details.last_index = ULONG_MAX;
3012 
3013 
3014 	mutex_lock(&mapping->i_mmap_mutex);
3015 	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
3016 		unmap_mapping_range_tree(&mapping->i_mmap, &details);
3017 	if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
3018 		unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3019 	mutex_unlock(&mapping->i_mmap_mutex);
3020 }
3021 EXPORT_SYMBOL(unmap_mapping_range);
3022 
3023 /*
3024  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3025  * but allow concurrent faults), and pte mapped but not yet locked.
3026  * We return with mmap_sem still held, but pte unmapped and unlocked.
3027  */
3028 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3029 		unsigned long address, pte_t *page_table, pmd_t *pmd,
3030 		unsigned int flags, pte_t orig_pte)
3031 {
3032 	spinlock_t *ptl;
3033 	struct page *page, *swapcache;
3034 	swp_entry_t entry;
3035 	pte_t pte;
3036 	int locked;
3037 	struct mem_cgroup *ptr;
3038 	int exclusive = 0;
3039 	int ret = 0;
3040 
3041 	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3042 		goto out;
3043 
3044 	entry = pte_to_swp_entry(orig_pte);
3045 	if (unlikely(non_swap_entry(entry))) {
3046 		if (is_migration_entry(entry)) {
3047 			migration_entry_wait(mm, pmd, address);
3048 		} else if (is_hwpoison_entry(entry)) {
3049 			ret = VM_FAULT_HWPOISON;
3050 		} else {
3051 			print_bad_pte(vma, address, orig_pte, NULL);
3052 			ret = VM_FAULT_SIGBUS;
3053 		}
3054 		goto out;
3055 	}
3056 	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3057 	page = lookup_swap_cache(entry);
3058 	if (!page) {
3059 		page = swapin_readahead(entry,
3060 					GFP_HIGHUSER_MOVABLE, vma, address);
3061 		if (!page) {
3062 			/*
3063 			 * Back out if somebody else faulted in this pte
3064 			 * while we released the pte lock.
3065 			 */
3066 			page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3067 			if (likely(pte_same(*page_table, orig_pte)))
3068 				ret = VM_FAULT_OOM;
3069 			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3070 			goto unlock;
3071 		}
3072 
3073 		/* Had to read the page from swap area: Major fault */
3074 		ret = VM_FAULT_MAJOR;
3075 		count_vm_event(PGMAJFAULT);
3076 		mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3077 	} else if (PageHWPoison(page)) {
3078 		/*
3079 		 * hwpoisoned dirty swapcache pages are kept for killing
3080 		 * owner processes (which may be unknown at hwpoison time)
3081 		 */
3082 		ret = VM_FAULT_HWPOISON;
3083 		delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3084 		swapcache = page;
3085 		goto out_release;
3086 	}
3087 
3088 	swapcache = page;
3089 	locked = lock_page_or_retry(page, mm, flags);
3090 
3091 	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3092 	if (!locked) {
3093 		ret |= VM_FAULT_RETRY;
3094 		goto out_release;
3095 	}
3096 
3097 	/*
3098 	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3099 	 * release the swapcache from under us.  The page pin, and pte_same
3100 	 * test below, are not enough to exclude that.  Even if it is still
3101 	 * swapcache, we need to check that the page's swap has not changed.
3102 	 */
3103 	if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3104 		goto out_page;
3105 
3106 	page = ksm_might_need_to_copy(page, vma, address);
3107 	if (unlikely(!page)) {
3108 		ret = VM_FAULT_OOM;
3109 		page = swapcache;
3110 		goto out_page;
3111 	}
3112 
3113 	if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3114 		ret = VM_FAULT_OOM;
3115 		goto out_page;
3116 	}
3117 
3118 	/*
3119 	 * Back out if somebody else already faulted in this pte.
3120 	 */
3121 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3122 	if (unlikely(!pte_same(*page_table, orig_pte)))
3123 		goto out_nomap;
3124 
3125 	if (unlikely(!PageUptodate(page))) {
3126 		ret = VM_FAULT_SIGBUS;
3127 		goto out_nomap;
3128 	}
3129 
3130 	/*
3131 	 * The page isn't present yet, go ahead with the fault.
3132 	 *
3133 	 * Be careful about the sequence of operations here.
3134 	 * To get its accounting right, reuse_swap_page() must be called
3135 	 * while the page is counted on swap but not yet in mapcount i.e.
3136 	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3137 	 * must be called after the swap_free(), or it will never succeed.
3138 	 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3139 	 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3140 	 * in page->private. In this case, a record in swap_cgroup  is silently
3141 	 * discarded at swap_free().
3142 	 */
3143 
3144 	inc_mm_counter_fast(mm, MM_ANONPAGES);
3145 	dec_mm_counter_fast(mm, MM_SWAPENTS);
3146 	pte = mk_pte(page, vma->vm_page_prot);
3147 	if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3148 		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3149 		flags &= ~FAULT_FLAG_WRITE;
3150 		ret |= VM_FAULT_WRITE;
3151 		exclusive = 1;
3152 	}
3153 	flush_icache_page(vma, page);
3154 	if (pte_swp_soft_dirty(orig_pte))
3155 		pte = pte_mksoft_dirty(pte);
3156 	set_pte_at(mm, address, page_table, pte);
3157 	if (page == swapcache)
3158 		do_page_add_anon_rmap(page, vma, address, exclusive);
3159 	else /* ksm created a completely new copy */
3160 		page_add_new_anon_rmap(page, vma, address);
3161 	/* It's better to call commit-charge after rmap is established */
3162 	mem_cgroup_commit_charge_swapin(page, ptr);
3163 
3164 	swap_free(entry);
3165 	if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3166 		try_to_free_swap(page);
3167 	unlock_page(page);
3168 	if (page != swapcache) {
3169 		/*
3170 		 * Hold the lock to avoid the swap entry to be reused
3171 		 * until we take the PT lock for the pte_same() check
3172 		 * (to avoid false positives from pte_same). For
3173 		 * further safety release the lock after the swap_free
3174 		 * so that the swap count won't change under a
3175 		 * parallel locked swapcache.
3176 		 */
3177 		unlock_page(swapcache);
3178 		page_cache_release(swapcache);
3179 	}
3180 
3181 	if (flags & FAULT_FLAG_WRITE) {
3182 		ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3183 		if (ret & VM_FAULT_ERROR)
3184 			ret &= VM_FAULT_ERROR;
3185 		goto out;
3186 	}
3187 
3188 	/* No need to invalidate - it was non-present before */
3189 	update_mmu_cache(vma, address, page_table);
3190 unlock:
3191 	pte_unmap_unlock(page_table, ptl);
3192 out:
3193 	return ret;
3194 out_nomap:
3195 	mem_cgroup_cancel_charge_swapin(ptr);
3196 	pte_unmap_unlock(page_table, ptl);
3197 out_page:
3198 	unlock_page(page);
3199 out_release:
3200 	page_cache_release(page);
3201 	if (page != swapcache) {
3202 		unlock_page(swapcache);
3203 		page_cache_release(swapcache);
3204 	}
3205 	return ret;
3206 }
3207 
3208 /*
3209  * This is like a special single-page "expand_{down|up}wards()",
3210  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3211  * doesn't hit another vma.
3212  */
3213 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3214 {
3215 	address &= PAGE_MASK;
3216 	if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3217 		struct vm_area_struct *prev = vma->vm_prev;
3218 
3219 		/*
3220 		 * Is there a mapping abutting this one below?
3221 		 *
3222 		 * That's only ok if it's the same stack mapping
3223 		 * that has gotten split..
3224 		 */
3225 		if (prev && prev->vm_end == address)
3226 			return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3227 
3228 		expand_downwards(vma, address - PAGE_SIZE);
3229 	}
3230 	if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3231 		struct vm_area_struct *next = vma->vm_next;
3232 
3233 		/* As VM_GROWSDOWN but s/below/above/ */
3234 		if (next && next->vm_start == address + PAGE_SIZE)
3235 			return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3236 
3237 		expand_upwards(vma, address + PAGE_SIZE);
3238 	}
3239 	return 0;
3240 }
3241 
3242 /*
3243  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3244  * but allow concurrent faults), and pte mapped but not yet locked.
3245  * We return with mmap_sem still held, but pte unmapped and unlocked.
3246  */
3247 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3248 		unsigned long address, pte_t *page_table, pmd_t *pmd,
3249 		unsigned int flags)
3250 {
3251 	struct page *page;
3252 	spinlock_t *ptl;
3253 	pte_t entry;
3254 
3255 	pte_unmap(page_table);
3256 
3257 	/* Check if we need to add a guard page to the stack */
3258 	if (check_stack_guard_page(vma, address) < 0)
3259 		return VM_FAULT_SIGBUS;
3260 
3261 	/* Use the zero-page for reads */
3262 	if (!(flags & FAULT_FLAG_WRITE)) {
3263 		entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3264 						vma->vm_page_prot));
3265 		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3266 		if (!pte_none(*page_table))
3267 			goto unlock;
3268 		goto setpte;
3269 	}
3270 
3271 	/* Allocate our own private page. */
3272 	if (unlikely(anon_vma_prepare(vma)))
3273 		goto oom;
3274 	page = alloc_zeroed_user_highpage_movable(vma, address);
3275 	if (!page)
3276 		goto oom;
3277 	/*
3278 	 * The memory barrier inside __SetPageUptodate makes sure that
3279 	 * preceeding stores to the page contents become visible before
3280 	 * the set_pte_at() write.
3281 	 */
3282 	__SetPageUptodate(page);
3283 
3284 	if (mem_cgroup_charge_anon(page, mm, GFP_KERNEL))
3285 		goto oom_free_page;
3286 
3287 	entry = mk_pte(page, vma->vm_page_prot);
3288 	if (vma->vm_flags & VM_WRITE)
3289 		entry = pte_mkwrite(pte_mkdirty(entry));
3290 
3291 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3292 	if (!pte_none(*page_table))
3293 		goto release;
3294 
3295 	inc_mm_counter_fast(mm, MM_ANONPAGES);
3296 	page_add_new_anon_rmap(page, vma, address);
3297 setpte:
3298 	set_pte_at(mm, address, page_table, entry);
3299 
3300 	/* No need to invalidate - it was non-present before */
3301 	update_mmu_cache(vma, address, page_table);
3302 unlock:
3303 	pte_unmap_unlock(page_table, ptl);
3304 	return 0;
3305 release:
3306 	mem_cgroup_uncharge_page(page);
3307 	page_cache_release(page);
3308 	goto unlock;
3309 oom_free_page:
3310 	page_cache_release(page);
3311 oom:
3312 	return VM_FAULT_OOM;
3313 }
3314 
3315 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
3316 		pgoff_t pgoff, unsigned int flags, struct page **page)
3317 {
3318 	struct vm_fault vmf;
3319 	int ret;
3320 
3321 	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3322 	vmf.pgoff = pgoff;
3323 	vmf.flags = flags;
3324 	vmf.page = NULL;
3325 
3326 	ret = vma->vm_ops->fault(vma, &vmf);
3327 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3328 		return ret;
3329 
3330 	if (unlikely(PageHWPoison(vmf.page))) {
3331 		if (ret & VM_FAULT_LOCKED)
3332 			unlock_page(vmf.page);
3333 		page_cache_release(vmf.page);
3334 		return VM_FAULT_HWPOISON;
3335 	}
3336 
3337 	if (unlikely(!(ret & VM_FAULT_LOCKED)))
3338 		lock_page(vmf.page);
3339 	else
3340 		VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
3341 
3342 	*page = vmf.page;
3343 	return ret;
3344 }
3345 
3346 /**
3347  * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
3348  *
3349  * @vma: virtual memory area
3350  * @address: user virtual address
3351  * @page: page to map
3352  * @pte: pointer to target page table entry
3353  * @write: true, if new entry is writable
3354  * @anon: true, if it's anonymous page
3355  *
3356  * Caller must hold page table lock relevant for @pte.
3357  *
3358  * Target users are page handler itself and implementations of
3359  * vm_ops->map_pages.
3360  */
3361 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
3362 		struct page *page, pte_t *pte, bool write, bool anon)
3363 {
3364 	pte_t entry;
3365 
3366 	flush_icache_page(vma, page);
3367 	entry = mk_pte(page, vma->vm_page_prot);
3368 	if (write)
3369 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3370 	else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
3371 		pte_mksoft_dirty(entry);
3372 	if (anon) {
3373 		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3374 		page_add_new_anon_rmap(page, vma, address);
3375 	} else {
3376 		inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
3377 		page_add_file_rmap(page);
3378 	}
3379 	set_pte_at(vma->vm_mm, address, pte, entry);
3380 
3381 	/* no need to invalidate: a not-present page won't be cached */
3382 	update_mmu_cache(vma, address, pte);
3383 }
3384 
3385 #define FAULT_AROUND_ORDER 4
3386 
3387 #ifdef CONFIG_DEBUG_FS
3388 static unsigned int fault_around_order = FAULT_AROUND_ORDER;
3389 
3390 static int fault_around_order_get(void *data, u64 *val)
3391 {
3392 	*val = fault_around_order;
3393 	return 0;
3394 }
3395 
3396 static int fault_around_order_set(void *data, u64 val)
3397 {
3398 	BUILD_BUG_ON((1UL << FAULT_AROUND_ORDER) > PTRS_PER_PTE);
3399 	if (1UL << val > PTRS_PER_PTE)
3400 		return -EINVAL;
3401 	fault_around_order = val;
3402 	return 0;
3403 }
3404 DEFINE_SIMPLE_ATTRIBUTE(fault_around_order_fops,
3405 		fault_around_order_get, fault_around_order_set, "%llu\n");
3406 
3407 static int __init fault_around_debugfs(void)
3408 {
3409 	void *ret;
3410 
3411 	ret = debugfs_create_file("fault_around_order",	0644, NULL, NULL,
3412 			&fault_around_order_fops);
3413 	if (!ret)
3414 		pr_warn("Failed to create fault_around_order in debugfs");
3415 	return 0;
3416 }
3417 late_initcall(fault_around_debugfs);
3418 
3419 static inline unsigned long fault_around_pages(void)
3420 {
3421 	return 1UL << fault_around_order;
3422 }
3423 
3424 static inline unsigned long fault_around_mask(void)
3425 {
3426 	return ~((1UL << (PAGE_SHIFT + fault_around_order)) - 1);
3427 }
3428 #else
3429 static inline unsigned long fault_around_pages(void)
3430 {
3431 	unsigned long nr_pages;
3432 
3433 	nr_pages = 1UL << FAULT_AROUND_ORDER;
3434 	BUILD_BUG_ON(nr_pages > PTRS_PER_PTE);
3435 	return nr_pages;
3436 }
3437 
3438 static inline unsigned long fault_around_mask(void)
3439 {
3440 	return ~((1UL << (PAGE_SHIFT + FAULT_AROUND_ORDER)) - 1);
3441 }
3442 #endif
3443 
3444 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
3445 		pte_t *pte, pgoff_t pgoff, unsigned int flags)
3446 {
3447 	unsigned long start_addr;
3448 	pgoff_t max_pgoff;
3449 	struct vm_fault vmf;
3450 	int off;
3451 
3452 	start_addr = max(address & fault_around_mask(), vma->vm_start);
3453 	off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3454 	pte -= off;
3455 	pgoff -= off;
3456 
3457 	/*
3458 	 *  max_pgoff is either end of page table or end of vma
3459 	 *  or fault_around_pages() from pgoff, depending what is neast.
3460 	 */
3461 	max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3462 		PTRS_PER_PTE - 1;
3463 	max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
3464 			pgoff + fault_around_pages() - 1);
3465 
3466 	/* Check if it makes any sense to call ->map_pages */
3467 	while (!pte_none(*pte)) {
3468 		if (++pgoff > max_pgoff)
3469 			return;
3470 		start_addr += PAGE_SIZE;
3471 		if (start_addr >= vma->vm_end)
3472 			return;
3473 		pte++;
3474 	}
3475 
3476 	vmf.virtual_address = (void __user *) start_addr;
3477 	vmf.pte = pte;
3478 	vmf.pgoff = pgoff;
3479 	vmf.max_pgoff = max_pgoff;
3480 	vmf.flags = flags;
3481 	vma->vm_ops->map_pages(vma, &vmf);
3482 }
3483 
3484 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3485 		unsigned long address, pmd_t *pmd,
3486 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3487 {
3488 	struct page *fault_page;
3489 	spinlock_t *ptl;
3490 	pte_t *pte;
3491 	int ret = 0;
3492 
3493 	/*
3494 	 * Let's call ->map_pages() first and use ->fault() as fallback
3495 	 * if page by the offset is not ready to be mapped (cold cache or
3496 	 * something).
3497 	 */
3498 	if (vma->vm_ops->map_pages) {
3499 		pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3500 		do_fault_around(vma, address, pte, pgoff, flags);
3501 		if (!pte_same(*pte, orig_pte))
3502 			goto unlock_out;
3503 		pte_unmap_unlock(pte, ptl);
3504 	}
3505 
3506 	ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3507 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3508 		return ret;
3509 
3510 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3511 	if (unlikely(!pte_same(*pte, orig_pte))) {
3512 		pte_unmap_unlock(pte, ptl);
3513 		unlock_page(fault_page);
3514 		page_cache_release(fault_page);
3515 		return ret;
3516 	}
3517 	do_set_pte(vma, address, fault_page, pte, false, false);
3518 	unlock_page(fault_page);
3519 unlock_out:
3520 	pte_unmap_unlock(pte, ptl);
3521 	return ret;
3522 }
3523 
3524 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3525 		unsigned long address, pmd_t *pmd,
3526 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3527 {
3528 	struct page *fault_page, *new_page;
3529 	spinlock_t *ptl;
3530 	pte_t *pte;
3531 	int ret;
3532 
3533 	if (unlikely(anon_vma_prepare(vma)))
3534 		return VM_FAULT_OOM;
3535 
3536 	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3537 	if (!new_page)
3538 		return VM_FAULT_OOM;
3539 
3540 	if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)) {
3541 		page_cache_release(new_page);
3542 		return VM_FAULT_OOM;
3543 	}
3544 
3545 	ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3546 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3547 		goto uncharge_out;
3548 
3549 	copy_user_highpage(new_page, fault_page, address, vma);
3550 	__SetPageUptodate(new_page);
3551 
3552 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3553 	if (unlikely(!pte_same(*pte, orig_pte))) {
3554 		pte_unmap_unlock(pte, ptl);
3555 		unlock_page(fault_page);
3556 		page_cache_release(fault_page);
3557 		goto uncharge_out;
3558 	}
3559 	do_set_pte(vma, address, new_page, pte, true, true);
3560 	pte_unmap_unlock(pte, ptl);
3561 	unlock_page(fault_page);
3562 	page_cache_release(fault_page);
3563 	return ret;
3564 uncharge_out:
3565 	mem_cgroup_uncharge_page(new_page);
3566 	page_cache_release(new_page);
3567 	return ret;
3568 }
3569 
3570 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3571 		unsigned long address, pmd_t *pmd,
3572 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3573 {
3574 	struct page *fault_page;
3575 	struct address_space *mapping;
3576 	spinlock_t *ptl;
3577 	pte_t *pte;
3578 	int dirtied = 0;
3579 	int ret, tmp;
3580 
3581 	ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3582 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3583 		return ret;
3584 
3585 	/*
3586 	 * Check if the backing address space wants to know that the page is
3587 	 * about to become writable
3588 	 */
3589 	if (vma->vm_ops->page_mkwrite) {
3590 		unlock_page(fault_page);
3591 		tmp = do_page_mkwrite(vma, fault_page, address);
3592 		if (unlikely(!tmp ||
3593 				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3594 			page_cache_release(fault_page);
3595 			return tmp;
3596 		}
3597 	}
3598 
3599 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3600 	if (unlikely(!pte_same(*pte, orig_pte))) {
3601 		pte_unmap_unlock(pte, ptl);
3602 		unlock_page(fault_page);
3603 		page_cache_release(fault_page);
3604 		return ret;
3605 	}
3606 	do_set_pte(vma, address, fault_page, pte, true, false);
3607 	pte_unmap_unlock(pte, ptl);
3608 
3609 	if (set_page_dirty(fault_page))
3610 		dirtied = 1;
3611 	mapping = fault_page->mapping;
3612 	unlock_page(fault_page);
3613 	if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3614 		/*
3615 		 * Some device drivers do not set page.mapping but still
3616 		 * dirty their pages
3617 		 */
3618 		balance_dirty_pages_ratelimited(mapping);
3619 	}
3620 
3621 	/* file_update_time outside page_lock */
3622 	if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3623 		file_update_time(vma->vm_file);
3624 
3625 	return ret;
3626 }
3627 
3628 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3629 		unsigned long address, pte_t *page_table, pmd_t *pmd,
3630 		unsigned int flags, pte_t orig_pte)
3631 {
3632 	pgoff_t pgoff = (((address & PAGE_MASK)
3633 			- vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3634 
3635 	pte_unmap(page_table);
3636 	if (!(flags & FAULT_FLAG_WRITE))
3637 		return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3638 				orig_pte);
3639 	if (!(vma->vm_flags & VM_SHARED))
3640 		return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3641 				orig_pte);
3642 	return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3643 }
3644 
3645 /*
3646  * Fault of a previously existing named mapping. Repopulate the pte
3647  * from the encoded file_pte if possible. This enables swappable
3648  * nonlinear vmas.
3649  *
3650  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3651  * but allow concurrent faults), and pte mapped but not yet locked.
3652  * We return with mmap_sem still held, but pte unmapped and unlocked.
3653  */
3654 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3655 		unsigned long address, pte_t *page_table, pmd_t *pmd,
3656 		unsigned int flags, pte_t orig_pte)
3657 {
3658 	pgoff_t pgoff;
3659 
3660 	flags |= FAULT_FLAG_NONLINEAR;
3661 
3662 	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3663 		return 0;
3664 
3665 	if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3666 		/*
3667 		 * Page table corrupted: show pte and kill process.
3668 		 */
3669 		print_bad_pte(vma, address, orig_pte, NULL);
3670 		return VM_FAULT_SIGBUS;
3671 	}
3672 
3673 	pgoff = pte_to_pgoff(orig_pte);
3674 	if (!(flags & FAULT_FLAG_WRITE))
3675 		return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3676 				orig_pte);
3677 	if (!(vma->vm_flags & VM_SHARED))
3678 		return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3679 				orig_pte);
3680 	return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3681 }
3682 
3683 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3684 				unsigned long addr, int page_nid,
3685 				int *flags)
3686 {
3687 	get_page(page);
3688 
3689 	count_vm_numa_event(NUMA_HINT_FAULTS);
3690 	if (page_nid == numa_node_id()) {
3691 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3692 		*flags |= TNF_FAULT_LOCAL;
3693 	}
3694 
3695 	return mpol_misplaced(page, vma, addr);
3696 }
3697 
3698 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3699 		   unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3700 {
3701 	struct page *page = NULL;
3702 	spinlock_t *ptl;
3703 	int page_nid = -1;
3704 	int last_cpupid;
3705 	int target_nid;
3706 	bool migrated = false;
3707 	int flags = 0;
3708 
3709 	/*
3710 	* The "pte" at this point cannot be used safely without
3711 	* validation through pte_unmap_same(). It's of NUMA type but
3712 	* the pfn may be screwed if the read is non atomic.
3713 	*
3714 	* ptep_modify_prot_start is not called as this is clearing
3715 	* the _PAGE_NUMA bit and it is not really expected that there
3716 	* would be concurrent hardware modifications to the PTE.
3717 	*/
3718 	ptl = pte_lockptr(mm, pmd);
3719 	spin_lock(ptl);
3720 	if (unlikely(!pte_same(*ptep, pte))) {
3721 		pte_unmap_unlock(ptep, ptl);
3722 		goto out;
3723 	}
3724 
3725 	pte = pte_mknonnuma(pte);
3726 	set_pte_at(mm, addr, ptep, pte);
3727 	update_mmu_cache(vma, addr, ptep);
3728 
3729 	page = vm_normal_page(vma, addr, pte);
3730 	if (!page) {
3731 		pte_unmap_unlock(ptep, ptl);
3732 		return 0;
3733 	}
3734 	BUG_ON(is_zero_pfn(page_to_pfn(page)));
3735 
3736 	/*
3737 	 * Avoid grouping on DSO/COW pages in specific and RO pages
3738 	 * in general, RO pages shouldn't hurt as much anyway since
3739 	 * they can be in shared cache state.
3740 	 */
3741 	if (!pte_write(pte))
3742 		flags |= TNF_NO_GROUP;
3743 
3744 	/*
3745 	 * Flag if the page is shared between multiple address spaces. This
3746 	 * is later used when determining whether to group tasks together
3747 	 */
3748 	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3749 		flags |= TNF_SHARED;
3750 
3751 	last_cpupid = page_cpupid_last(page);
3752 	page_nid = page_to_nid(page);
3753 	target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3754 	pte_unmap_unlock(ptep, ptl);
3755 	if (target_nid == -1) {
3756 		put_page(page);
3757 		goto out;
3758 	}
3759 
3760 	/* Migrate to the requested node */
3761 	migrated = migrate_misplaced_page(page, vma, target_nid);
3762 	if (migrated) {
3763 		page_nid = target_nid;
3764 		flags |= TNF_MIGRATED;
3765 	}
3766 
3767 out:
3768 	if (page_nid != -1)
3769 		task_numa_fault(last_cpupid, page_nid, 1, flags);
3770 	return 0;
3771 }
3772 
3773 /*
3774  * These routines also need to handle stuff like marking pages dirty
3775  * and/or accessed for architectures that don't do it in hardware (most
3776  * RISC architectures).  The early dirtying is also good on the i386.
3777  *
3778  * There is also a hook called "update_mmu_cache()" that architectures
3779  * with external mmu caches can use to update those (ie the Sparc or
3780  * PowerPC hashed page tables that act as extended TLBs).
3781  *
3782  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3783  * but allow concurrent faults), and pte mapped but not yet locked.
3784  * We return with mmap_sem still held, but pte unmapped and unlocked.
3785  */
3786 static int handle_pte_fault(struct mm_struct *mm,
3787 		     struct vm_area_struct *vma, unsigned long address,
3788 		     pte_t *pte, pmd_t *pmd, unsigned int flags)
3789 {
3790 	pte_t entry;
3791 	spinlock_t *ptl;
3792 
3793 	entry = *pte;
3794 	if (!pte_present(entry)) {
3795 		if (pte_none(entry)) {
3796 			if (vma->vm_ops) {
3797 				if (likely(vma->vm_ops->fault))
3798 					return do_linear_fault(mm, vma, address,
3799 						pte, pmd, flags, entry);
3800 			}
3801 			return do_anonymous_page(mm, vma, address,
3802 						 pte, pmd, flags);
3803 		}
3804 		if (pte_file(entry))
3805 			return do_nonlinear_fault(mm, vma, address,
3806 					pte, pmd, flags, entry);
3807 		return do_swap_page(mm, vma, address,
3808 					pte, pmd, flags, entry);
3809 	}
3810 
3811 	if (pte_numa(entry))
3812 		return do_numa_page(mm, vma, address, entry, pte, pmd);
3813 
3814 	ptl = pte_lockptr(mm, pmd);
3815 	spin_lock(ptl);
3816 	if (unlikely(!pte_same(*pte, entry)))
3817 		goto unlock;
3818 	if (flags & FAULT_FLAG_WRITE) {
3819 		if (!pte_write(entry))
3820 			return do_wp_page(mm, vma, address,
3821 					pte, pmd, ptl, entry);
3822 		entry = pte_mkdirty(entry);
3823 	}
3824 	entry = pte_mkyoung(entry);
3825 	if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3826 		update_mmu_cache(vma, address, pte);
3827 	} else {
3828 		/*
3829 		 * This is needed only for protection faults but the arch code
3830 		 * is not yet telling us if this is a protection fault or not.
3831 		 * This still avoids useless tlb flushes for .text page faults
3832 		 * with threads.
3833 		 */
3834 		if (flags & FAULT_FLAG_WRITE)
3835 			flush_tlb_fix_spurious_fault(vma, address);
3836 	}
3837 unlock:
3838 	pte_unmap_unlock(pte, ptl);
3839 	return 0;
3840 }
3841 
3842 /*
3843  * By the time we get here, we already hold the mm semaphore
3844  */
3845 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3846 			     unsigned long address, unsigned int flags)
3847 {
3848 	pgd_t *pgd;
3849 	pud_t *pud;
3850 	pmd_t *pmd;
3851 	pte_t *pte;
3852 
3853 	if (unlikely(is_vm_hugetlb_page(vma)))
3854 		return hugetlb_fault(mm, vma, address, flags);
3855 
3856 	pgd = pgd_offset(mm, address);
3857 	pud = pud_alloc(mm, pgd, address);
3858 	if (!pud)
3859 		return VM_FAULT_OOM;
3860 	pmd = pmd_alloc(mm, pud, address);
3861 	if (!pmd)
3862 		return VM_FAULT_OOM;
3863 	if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3864 		int ret = VM_FAULT_FALLBACK;
3865 		if (!vma->vm_ops)
3866 			ret = do_huge_pmd_anonymous_page(mm, vma, address,
3867 					pmd, flags);
3868 		if (!(ret & VM_FAULT_FALLBACK))
3869 			return ret;
3870 	} else {
3871 		pmd_t orig_pmd = *pmd;
3872 		int ret;
3873 
3874 		barrier();
3875 		if (pmd_trans_huge(orig_pmd)) {
3876 			unsigned int dirty = flags & FAULT_FLAG_WRITE;
3877 
3878 			/*
3879 			 * If the pmd is splitting, return and retry the
3880 			 * the fault.  Alternative: wait until the split
3881 			 * is done, and goto retry.
3882 			 */
3883 			if (pmd_trans_splitting(orig_pmd))
3884 				return 0;
3885 
3886 			if (pmd_numa(orig_pmd))
3887 				return do_huge_pmd_numa_page(mm, vma, address,
3888 							     orig_pmd, pmd);
3889 
3890 			if (dirty && !pmd_write(orig_pmd)) {
3891 				ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3892 							  orig_pmd);
3893 				if (!(ret & VM_FAULT_FALLBACK))
3894 					return ret;
3895 			} else {
3896 				huge_pmd_set_accessed(mm, vma, address, pmd,
3897 						      orig_pmd, dirty);
3898 				return 0;
3899 			}
3900 		}
3901 	}
3902 
3903 	/* THP should already have been handled */
3904 	BUG_ON(pmd_numa(*pmd));
3905 
3906 	/*
3907 	 * Use __pte_alloc instead of pte_alloc_map, because we can't
3908 	 * run pte_offset_map on the pmd, if an huge pmd could
3909 	 * materialize from under us from a different thread.
3910 	 */
3911 	if (unlikely(pmd_none(*pmd)) &&
3912 	    unlikely(__pte_alloc(mm, vma, pmd, address)))
3913 		return VM_FAULT_OOM;
3914 	/* if an huge pmd materialized from under us just retry later */
3915 	if (unlikely(pmd_trans_huge(*pmd)))
3916 		return 0;
3917 	/*
3918 	 * A regular pmd is established and it can't morph into a huge pmd
3919 	 * from under us anymore at this point because we hold the mmap_sem
3920 	 * read mode and khugepaged takes it in write mode. So now it's
3921 	 * safe to run pte_offset_map().
3922 	 */
3923 	pte = pte_offset_map(pmd, address);
3924 
3925 	return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3926 }
3927 
3928 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3929 		    unsigned long address, unsigned int flags)
3930 {
3931 	int ret;
3932 
3933 	__set_current_state(TASK_RUNNING);
3934 
3935 	count_vm_event(PGFAULT);
3936 	mem_cgroup_count_vm_event(mm, PGFAULT);
3937 
3938 	/* do counter updates before entering really critical section. */
3939 	check_sync_rss_stat(current);
3940 
3941 	/*
3942 	 * Enable the memcg OOM handling for faults triggered in user
3943 	 * space.  Kernel faults are handled more gracefully.
3944 	 */
3945 	if (flags & FAULT_FLAG_USER)
3946 		mem_cgroup_oom_enable();
3947 
3948 	ret = __handle_mm_fault(mm, vma, address, flags);
3949 
3950 	if (flags & FAULT_FLAG_USER) {
3951 		mem_cgroup_oom_disable();
3952                 /*
3953                  * The task may have entered a memcg OOM situation but
3954                  * if the allocation error was handled gracefully (no
3955                  * VM_FAULT_OOM), there is no need to kill anything.
3956                  * Just clean up the OOM state peacefully.
3957                  */
3958                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3959                         mem_cgroup_oom_synchronize(false);
3960 	}
3961 
3962 	return ret;
3963 }
3964 
3965 #ifndef __PAGETABLE_PUD_FOLDED
3966 /*
3967  * Allocate page upper directory.
3968  * We've already handled the fast-path in-line.
3969  */
3970 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3971 {
3972 	pud_t *new = pud_alloc_one(mm, address);
3973 	if (!new)
3974 		return -ENOMEM;
3975 
3976 	smp_wmb(); /* See comment in __pte_alloc */
3977 
3978 	spin_lock(&mm->page_table_lock);
3979 	if (pgd_present(*pgd))		/* Another has populated it */
3980 		pud_free(mm, new);
3981 	else
3982 		pgd_populate(mm, pgd, new);
3983 	spin_unlock(&mm->page_table_lock);
3984 	return 0;
3985 }
3986 #endif /* __PAGETABLE_PUD_FOLDED */
3987 
3988 #ifndef __PAGETABLE_PMD_FOLDED
3989 /*
3990  * Allocate page middle directory.
3991  * We've already handled the fast-path in-line.
3992  */
3993 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3994 {
3995 	pmd_t *new = pmd_alloc_one(mm, address);
3996 	if (!new)
3997 		return -ENOMEM;
3998 
3999 	smp_wmb(); /* See comment in __pte_alloc */
4000 
4001 	spin_lock(&mm->page_table_lock);
4002 #ifndef __ARCH_HAS_4LEVEL_HACK
4003 	if (pud_present(*pud))		/* Another has populated it */
4004 		pmd_free(mm, new);
4005 	else
4006 		pud_populate(mm, pud, new);
4007 #else
4008 	if (pgd_present(*pud))		/* Another has populated it */
4009 		pmd_free(mm, new);
4010 	else
4011 		pgd_populate(mm, pud, new);
4012 #endif /* __ARCH_HAS_4LEVEL_HACK */
4013 	spin_unlock(&mm->page_table_lock);
4014 	return 0;
4015 }
4016 #endif /* __PAGETABLE_PMD_FOLDED */
4017 
4018 #if !defined(__HAVE_ARCH_GATE_AREA)
4019 
4020 #if defined(AT_SYSINFO_EHDR)
4021 static struct vm_area_struct gate_vma;
4022 
4023 static int __init gate_vma_init(void)
4024 {
4025 	gate_vma.vm_mm = NULL;
4026 	gate_vma.vm_start = FIXADDR_USER_START;
4027 	gate_vma.vm_end = FIXADDR_USER_END;
4028 	gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
4029 	gate_vma.vm_page_prot = __P101;
4030 
4031 	return 0;
4032 }
4033 __initcall(gate_vma_init);
4034 #endif
4035 
4036 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
4037 {
4038 #ifdef AT_SYSINFO_EHDR
4039 	return &gate_vma;
4040 #else
4041 	return NULL;
4042 #endif
4043 }
4044 
4045 int in_gate_area_no_mm(unsigned long addr)
4046 {
4047 #ifdef AT_SYSINFO_EHDR
4048 	if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
4049 		return 1;
4050 #endif
4051 	return 0;
4052 }
4053 
4054 #endif	/* __HAVE_ARCH_GATE_AREA */
4055 
4056 static int __follow_pte(struct mm_struct *mm, unsigned long address,
4057 		pte_t **ptepp, spinlock_t **ptlp)
4058 {
4059 	pgd_t *pgd;
4060 	pud_t *pud;
4061 	pmd_t *pmd;
4062 	pte_t *ptep;
4063 
4064 	pgd = pgd_offset(mm, address);
4065 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4066 		goto out;
4067 
4068 	pud = pud_offset(pgd, address);
4069 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4070 		goto out;
4071 
4072 	pmd = pmd_offset(pud, address);
4073 	VM_BUG_ON(pmd_trans_huge(*pmd));
4074 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4075 		goto out;
4076 
4077 	/* We cannot handle huge page PFN maps. Luckily they don't exist. */
4078 	if (pmd_huge(*pmd))
4079 		goto out;
4080 
4081 	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4082 	if (!ptep)
4083 		goto out;
4084 	if (!pte_present(*ptep))
4085 		goto unlock;
4086 	*ptepp = ptep;
4087 	return 0;
4088 unlock:
4089 	pte_unmap_unlock(ptep, *ptlp);
4090 out:
4091 	return -EINVAL;
4092 }
4093 
4094 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4095 			     pte_t **ptepp, spinlock_t **ptlp)
4096 {
4097 	int res;
4098 
4099 	/* (void) is needed to make gcc happy */
4100 	(void) __cond_lock(*ptlp,
4101 			   !(res = __follow_pte(mm, address, ptepp, ptlp)));
4102 	return res;
4103 }
4104 
4105 /**
4106  * follow_pfn - look up PFN at a user virtual address
4107  * @vma: memory mapping
4108  * @address: user virtual address
4109  * @pfn: location to store found PFN
4110  *
4111  * Only IO mappings and raw PFN mappings are allowed.
4112  *
4113  * Returns zero and the pfn at @pfn on success, -ve otherwise.
4114  */
4115 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4116 	unsigned long *pfn)
4117 {
4118 	int ret = -EINVAL;
4119 	spinlock_t *ptl;
4120 	pte_t *ptep;
4121 
4122 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4123 		return ret;
4124 
4125 	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4126 	if (ret)
4127 		return ret;
4128 	*pfn = pte_pfn(*ptep);
4129 	pte_unmap_unlock(ptep, ptl);
4130 	return 0;
4131 }
4132 EXPORT_SYMBOL(follow_pfn);
4133 
4134 #ifdef CONFIG_HAVE_IOREMAP_PROT
4135 int follow_phys(struct vm_area_struct *vma,
4136 		unsigned long address, unsigned int flags,
4137 		unsigned long *prot, resource_size_t *phys)
4138 {
4139 	int ret = -EINVAL;
4140 	pte_t *ptep, pte;
4141 	spinlock_t *ptl;
4142 
4143 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4144 		goto out;
4145 
4146 	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4147 		goto out;
4148 	pte = *ptep;
4149 
4150 	if ((flags & FOLL_WRITE) && !pte_write(pte))
4151 		goto unlock;
4152 
4153 	*prot = pgprot_val(pte_pgprot(pte));
4154 	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4155 
4156 	ret = 0;
4157 unlock:
4158 	pte_unmap_unlock(ptep, ptl);
4159 out:
4160 	return ret;
4161 }
4162 
4163 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4164 			void *buf, int len, int write)
4165 {
4166 	resource_size_t phys_addr;
4167 	unsigned long prot = 0;
4168 	void __iomem *maddr;
4169 	int offset = addr & (PAGE_SIZE-1);
4170 
4171 	if (follow_phys(vma, addr, write, &prot, &phys_addr))
4172 		return -EINVAL;
4173 
4174 	maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4175 	if (write)
4176 		memcpy_toio(maddr + offset, buf, len);
4177 	else
4178 		memcpy_fromio(buf, maddr + offset, len);
4179 	iounmap(maddr);
4180 
4181 	return len;
4182 }
4183 EXPORT_SYMBOL_GPL(generic_access_phys);
4184 #endif
4185 
4186 /*
4187  * Access another process' address space as given in mm.  If non-NULL, use the
4188  * given task for page fault accounting.
4189  */
4190 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4191 		unsigned long addr, void *buf, int len, int write)
4192 {
4193 	struct vm_area_struct *vma;
4194 	void *old_buf = buf;
4195 
4196 	down_read(&mm->mmap_sem);
4197 	/* ignore errors, just check how much was successfully transferred */
4198 	while (len) {
4199 		int bytes, ret, offset;
4200 		void *maddr;
4201 		struct page *page = NULL;
4202 
4203 		ret = get_user_pages(tsk, mm, addr, 1,
4204 				write, 1, &page, &vma);
4205 		if (ret <= 0) {
4206 			/*
4207 			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4208 			 * we can access using slightly different code.
4209 			 */
4210 #ifdef CONFIG_HAVE_IOREMAP_PROT
4211 			vma = find_vma(mm, addr);
4212 			if (!vma || vma->vm_start > addr)
4213 				break;
4214 			if (vma->vm_ops && vma->vm_ops->access)
4215 				ret = vma->vm_ops->access(vma, addr, buf,
4216 							  len, write);
4217 			if (ret <= 0)
4218 #endif
4219 				break;
4220 			bytes = ret;
4221 		} else {
4222 			bytes = len;
4223 			offset = addr & (PAGE_SIZE-1);
4224 			if (bytes > PAGE_SIZE-offset)
4225 				bytes = PAGE_SIZE-offset;
4226 
4227 			maddr = kmap(page);
4228 			if (write) {
4229 				copy_to_user_page(vma, page, addr,
4230 						  maddr + offset, buf, bytes);
4231 				set_page_dirty_lock(page);
4232 			} else {
4233 				copy_from_user_page(vma, page, addr,
4234 						    buf, maddr + offset, bytes);
4235 			}
4236 			kunmap(page);
4237 			page_cache_release(page);
4238 		}
4239 		len -= bytes;
4240 		buf += bytes;
4241 		addr += bytes;
4242 	}
4243 	up_read(&mm->mmap_sem);
4244 
4245 	return buf - old_buf;
4246 }
4247 
4248 /**
4249  * access_remote_vm - access another process' address space
4250  * @mm:		the mm_struct of the target address space
4251  * @addr:	start address to access
4252  * @buf:	source or destination buffer
4253  * @len:	number of bytes to transfer
4254  * @write:	whether the access is a write
4255  *
4256  * The caller must hold a reference on @mm.
4257  */
4258 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4259 		void *buf, int len, int write)
4260 {
4261 	return __access_remote_vm(NULL, mm, addr, buf, len, write);
4262 }
4263 
4264 /*
4265  * Access another process' address space.
4266  * Source/target buffer must be kernel space,
4267  * Do not walk the page table directly, use get_user_pages
4268  */
4269 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4270 		void *buf, int len, int write)
4271 {
4272 	struct mm_struct *mm;
4273 	int ret;
4274 
4275 	mm = get_task_mm(tsk);
4276 	if (!mm)
4277 		return 0;
4278 
4279 	ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4280 	mmput(mm);
4281 
4282 	return ret;
4283 }
4284 
4285 /*
4286  * Print the name of a VMA.
4287  */
4288 void print_vma_addr(char *prefix, unsigned long ip)
4289 {
4290 	struct mm_struct *mm = current->mm;
4291 	struct vm_area_struct *vma;
4292 
4293 	/*
4294 	 * Do not print if we are in atomic
4295 	 * contexts (in exception stacks, etc.):
4296 	 */
4297 	if (preempt_count())
4298 		return;
4299 
4300 	down_read(&mm->mmap_sem);
4301 	vma = find_vma(mm, ip);
4302 	if (vma && vma->vm_file) {
4303 		struct file *f = vma->vm_file;
4304 		char *buf = (char *)__get_free_page(GFP_KERNEL);
4305 		if (buf) {
4306 			char *p;
4307 
4308 			p = d_path(&f->f_path, buf, PAGE_SIZE);
4309 			if (IS_ERR(p))
4310 				p = "?";
4311 			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4312 					vma->vm_start,
4313 					vma->vm_end - vma->vm_start);
4314 			free_page((unsigned long)buf);
4315 		}
4316 	}
4317 	up_read(&mm->mmap_sem);
4318 }
4319 
4320 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4321 void might_fault(void)
4322 {
4323 	/*
4324 	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4325 	 * holding the mmap_sem, this is safe because kernel memory doesn't
4326 	 * get paged out, therefore we'll never actually fault, and the
4327 	 * below annotations will generate false positives.
4328 	 */
4329 	if (segment_eq(get_fs(), KERNEL_DS))
4330 		return;
4331 
4332 	/*
4333 	 * it would be nicer only to annotate paths which are not under
4334 	 * pagefault_disable, however that requires a larger audit and
4335 	 * providing helpers like get_user_atomic.
4336 	 */
4337 	if (in_atomic())
4338 		return;
4339 
4340 	__might_sleep(__FILE__, __LINE__, 0);
4341 
4342 	if (current->mm)
4343 		might_lock_read(&current->mm->mmap_sem);
4344 }
4345 EXPORT_SYMBOL(might_fault);
4346 #endif
4347 
4348 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4349 static void clear_gigantic_page(struct page *page,
4350 				unsigned long addr,
4351 				unsigned int pages_per_huge_page)
4352 {
4353 	int i;
4354 	struct page *p = page;
4355 
4356 	might_sleep();
4357 	for (i = 0; i < pages_per_huge_page;
4358 	     i++, p = mem_map_next(p, page, i)) {
4359 		cond_resched();
4360 		clear_user_highpage(p, addr + i * PAGE_SIZE);
4361 	}
4362 }
4363 void clear_huge_page(struct page *page,
4364 		     unsigned long addr, unsigned int pages_per_huge_page)
4365 {
4366 	int i;
4367 
4368 	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4369 		clear_gigantic_page(page, addr, pages_per_huge_page);
4370 		return;
4371 	}
4372 
4373 	might_sleep();
4374 	for (i = 0; i < pages_per_huge_page; i++) {
4375 		cond_resched();
4376 		clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4377 	}
4378 }
4379 
4380 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4381 				    unsigned long addr,
4382 				    struct vm_area_struct *vma,
4383 				    unsigned int pages_per_huge_page)
4384 {
4385 	int i;
4386 	struct page *dst_base = dst;
4387 	struct page *src_base = src;
4388 
4389 	for (i = 0; i < pages_per_huge_page; ) {
4390 		cond_resched();
4391 		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4392 
4393 		i++;
4394 		dst = mem_map_next(dst, dst_base, i);
4395 		src = mem_map_next(src, src_base, i);
4396 	}
4397 }
4398 
4399 void copy_user_huge_page(struct page *dst, struct page *src,
4400 			 unsigned long addr, struct vm_area_struct *vma,
4401 			 unsigned int pages_per_huge_page)
4402 {
4403 	int i;
4404 
4405 	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4406 		copy_user_gigantic_page(dst, src, addr, vma,
4407 					pages_per_huge_page);
4408 		return;
4409 	}
4410 
4411 	might_sleep();
4412 	for (i = 0; i < pages_per_huge_page; i++) {
4413 		cond_resched();
4414 		copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4415 	}
4416 }
4417 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4418 
4419 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4420 
4421 static struct kmem_cache *page_ptl_cachep;
4422 
4423 void __init ptlock_cache_init(void)
4424 {
4425 	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4426 			SLAB_PANIC, NULL);
4427 }
4428 
4429 bool ptlock_alloc(struct page *page)
4430 {
4431 	spinlock_t *ptl;
4432 
4433 	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4434 	if (!ptl)
4435 		return false;
4436 	page->ptl = ptl;
4437 	return true;
4438 }
4439 
4440 void ptlock_free(struct page *page)
4441 {
4442 	kmem_cache_free(page_ptl_cachep, page->ptl);
4443 }
4444 #endif
4445