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 #include <linux/userfaultfd_k.h>
65 
66 #include <asm/io.h>
67 #include <asm/pgalloc.h>
68 #include <asm/uaccess.h>
69 #include <asm/tlb.h>
70 #include <asm/tlbflush.h>
71 #include <asm/pgtable.h>
72 
73 #include "internal.h"
74 
75 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
76 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
77 #endif
78 
79 #ifndef CONFIG_NEED_MULTIPLE_NODES
80 /* use the per-pgdat data instead for discontigmem - mbligh */
81 unsigned long max_mapnr;
82 struct page *mem_map;
83 
84 EXPORT_SYMBOL(max_mapnr);
85 EXPORT_SYMBOL(mem_map);
86 #endif
87 
88 /*
89  * A number of key systems in x86 including ioremap() rely on the assumption
90  * that high_memory defines the upper bound on direct map memory, then end
91  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
92  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
93  * and ZONE_HIGHMEM.
94  */
95 void * high_memory;
96 
97 EXPORT_SYMBOL(high_memory);
98 
99 /*
100  * Randomize the address space (stacks, mmaps, brk, etc.).
101  *
102  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
103  *   as ancient (libc5 based) binaries can segfault. )
104  */
105 int randomize_va_space __read_mostly =
106 #ifdef CONFIG_COMPAT_BRK
107 					1;
108 #else
109 					2;
110 #endif
111 
disable_randmaps(char * s)112 static int __init disable_randmaps(char *s)
113 {
114 	randomize_va_space = 0;
115 	return 1;
116 }
117 __setup("norandmaps", disable_randmaps);
118 
119 unsigned long zero_pfn __read_mostly;
120 unsigned long highest_memmap_pfn __read_mostly;
121 
122 EXPORT_SYMBOL(zero_pfn);
123 
124 /*
125  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
126  */
init_zero_pfn(void)127 static int __init init_zero_pfn(void)
128 {
129 	zero_pfn = page_to_pfn(ZERO_PAGE(0));
130 	return 0;
131 }
132 core_initcall(init_zero_pfn);
133 
134 
135 #if defined(SPLIT_RSS_COUNTING)
136 
sync_mm_rss(struct mm_struct * mm)137 void sync_mm_rss(struct mm_struct *mm)
138 {
139 	int i;
140 
141 	for (i = 0; i < NR_MM_COUNTERS; i++) {
142 		if (current->rss_stat.count[i]) {
143 			add_mm_counter(mm, i, current->rss_stat.count[i]);
144 			current->rss_stat.count[i] = 0;
145 		}
146 	}
147 	current->rss_stat.events = 0;
148 }
149 
add_mm_counter_fast(struct mm_struct * mm,int member,int val)150 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
151 {
152 	struct task_struct *task = current;
153 
154 	if (likely(task->mm == mm))
155 		task->rss_stat.count[member] += val;
156 	else
157 		add_mm_counter(mm, member, val);
158 }
159 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
160 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
161 
162 /* sync counter once per 64 page faults */
163 #define TASK_RSS_EVENTS_THRESH	(64)
check_sync_rss_stat(struct task_struct * task)164 static void check_sync_rss_stat(struct task_struct *task)
165 {
166 	if (unlikely(task != current))
167 		return;
168 	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
169 		sync_mm_rss(task->mm);
170 }
171 #else /* SPLIT_RSS_COUNTING */
172 
173 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
174 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
175 
check_sync_rss_stat(struct task_struct * task)176 static void check_sync_rss_stat(struct task_struct *task)
177 {
178 }
179 
180 #endif /* SPLIT_RSS_COUNTING */
181 
182 #ifdef HAVE_GENERIC_MMU_GATHER
183 
tlb_next_batch(struct mmu_gather * tlb)184 static bool tlb_next_batch(struct mmu_gather *tlb)
185 {
186 	struct mmu_gather_batch *batch;
187 
188 	batch = tlb->active;
189 	if (batch->next) {
190 		tlb->active = batch->next;
191 		return true;
192 	}
193 
194 	if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
195 		return false;
196 
197 	batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
198 	if (!batch)
199 		return false;
200 
201 	tlb->batch_count++;
202 	batch->next = NULL;
203 	batch->nr   = 0;
204 	batch->max  = MAX_GATHER_BATCH;
205 
206 	tlb->active->next = batch;
207 	tlb->active = batch;
208 
209 	return true;
210 }
211 
212 /* tlb_gather_mmu
213  *	Called to initialize an (on-stack) mmu_gather structure for page-table
214  *	tear-down from @mm. The @fullmm argument is used when @mm is without
215  *	users and we're going to destroy the full address space (exit/execve).
216  */
tlb_gather_mmu(struct mmu_gather * tlb,struct mm_struct * mm,unsigned long start,unsigned long end)217 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
218 {
219 	tlb->mm = mm;
220 
221 	/* Is it from 0 to ~0? */
222 	tlb->fullmm     = !(start | (end+1));
223 	tlb->need_flush_all = 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 	__tlb_reset_range(tlb);
235 }
236 
tlb_flush_mmu_tlbonly(struct mmu_gather * tlb)237 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
238 {
239 	if (!tlb->end)
240 		return;
241 
242 	tlb_flush(tlb);
243 	mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
244 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
245 	tlb_table_flush(tlb);
246 #endif
247 	__tlb_reset_range(tlb);
248 }
249 
tlb_flush_mmu_free(struct mmu_gather * tlb)250 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
251 {
252 	struct mmu_gather_batch *batch;
253 
254 	for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
255 		free_pages_and_swap_cache(batch->pages, batch->nr);
256 		batch->nr = 0;
257 	}
258 	tlb->active = &tlb->local;
259 }
260 
tlb_flush_mmu(struct mmu_gather * tlb)261 void tlb_flush_mmu(struct mmu_gather *tlb)
262 {
263 	tlb_flush_mmu_tlbonly(tlb);
264 	tlb_flush_mmu_free(tlb);
265 }
266 
267 /* tlb_finish_mmu
268  *	Called at the end of the shootdown operation to free up any resources
269  *	that were required.
270  */
tlb_finish_mmu(struct mmu_gather * tlb,unsigned long start,unsigned long end)271 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
272 {
273 	struct mmu_gather_batch *batch, *next;
274 
275 	tlb_flush_mmu(tlb);
276 
277 	/* keep the page table cache within bounds */
278 	check_pgt_cache();
279 
280 	for (batch = tlb->local.next; batch; batch = next) {
281 		next = batch->next;
282 		free_pages((unsigned long)batch, 0);
283 	}
284 	tlb->local.next = NULL;
285 }
286 
287 /* __tlb_remove_page
288  *	Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
289  *	handling the additional races in SMP caused by other CPUs caching valid
290  *	mappings in their TLBs. Returns the number of free page slots left.
291  *	When out of page slots we must call tlb_flush_mmu().
292  */
__tlb_remove_page(struct mmu_gather * tlb,struct page * page)293 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
294 {
295 	struct mmu_gather_batch *batch;
296 
297 	VM_BUG_ON(!tlb->end);
298 
299 	batch = tlb->active;
300 	batch->pages[batch->nr++] = page;
301 	if (batch->nr == batch->max) {
302 		if (!tlb_next_batch(tlb))
303 			return 0;
304 		batch = tlb->active;
305 	}
306 	VM_BUG_ON_PAGE(batch->nr > batch->max, page);
307 
308 	return batch->max - batch->nr;
309 }
310 
311 #endif /* HAVE_GENERIC_MMU_GATHER */
312 
313 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
314 
315 /*
316  * See the comment near struct mmu_table_batch.
317  */
318 
tlb_remove_table_smp_sync(void * arg)319 static void tlb_remove_table_smp_sync(void *arg)
320 {
321 	/* Simply deliver the interrupt */
322 }
323 
tlb_remove_table_one(void * table)324 static void tlb_remove_table_one(void *table)
325 {
326 	/*
327 	 * This isn't an RCU grace period and hence the page-tables cannot be
328 	 * assumed to be actually RCU-freed.
329 	 *
330 	 * It is however sufficient for software page-table walkers that rely on
331 	 * IRQ disabling. See the comment near struct mmu_table_batch.
332 	 */
333 	smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
334 	__tlb_remove_table(table);
335 }
336 
tlb_remove_table_rcu(struct rcu_head * head)337 static void tlb_remove_table_rcu(struct rcu_head *head)
338 {
339 	struct mmu_table_batch *batch;
340 	int i;
341 
342 	batch = container_of(head, struct mmu_table_batch, rcu);
343 
344 	for (i = 0; i < batch->nr; i++)
345 		__tlb_remove_table(batch->tables[i]);
346 
347 	free_page((unsigned long)batch);
348 }
349 
tlb_table_flush(struct mmu_gather * tlb)350 void tlb_table_flush(struct mmu_gather *tlb)
351 {
352 	struct mmu_table_batch **batch = &tlb->batch;
353 
354 	if (*batch) {
355 		call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
356 		*batch = NULL;
357 	}
358 }
359 
tlb_remove_table(struct mmu_gather * tlb,void * table)360 void tlb_remove_table(struct mmu_gather *tlb, void *table)
361 {
362 	struct mmu_table_batch **batch = &tlb->batch;
363 
364 	/*
365 	 * When there's less then two users of this mm there cannot be a
366 	 * concurrent page-table walk.
367 	 */
368 	if (atomic_read(&tlb->mm->mm_users) < 2) {
369 		__tlb_remove_table(table);
370 		return;
371 	}
372 
373 	if (*batch == NULL) {
374 		*batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
375 		if (*batch == NULL) {
376 			tlb_remove_table_one(table);
377 			return;
378 		}
379 		(*batch)->nr = 0;
380 	}
381 	(*batch)->tables[(*batch)->nr++] = table;
382 	if ((*batch)->nr == MAX_TABLE_BATCH)
383 		tlb_table_flush(tlb);
384 }
385 
386 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
387 
388 /*
389  * Note: this doesn't free the actual pages themselves. That
390  * has been handled earlier when unmapping all the memory regions.
391  */
free_pte_range(struct mmu_gather * tlb,pmd_t * pmd,unsigned long addr)392 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
393 			   unsigned long addr)
394 {
395 	pgtable_t token = pmd_pgtable(*pmd);
396 	pmd_clear(pmd);
397 	pte_free_tlb(tlb, token, addr);
398 	atomic_long_dec(&tlb->mm->nr_ptes);
399 }
400 
free_pmd_range(struct mmu_gather * tlb,pud_t * pud,unsigned long addr,unsigned long end,unsigned long floor,unsigned long ceiling)401 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
402 				unsigned long addr, unsigned long end,
403 				unsigned long floor, unsigned long ceiling)
404 {
405 	pmd_t *pmd;
406 	unsigned long next;
407 	unsigned long start;
408 
409 	start = addr;
410 	pmd = pmd_offset(pud, addr);
411 	do {
412 		next = pmd_addr_end(addr, end);
413 		if (pmd_none_or_clear_bad(pmd))
414 			continue;
415 		free_pte_range(tlb, pmd, addr);
416 	} while (pmd++, addr = next, addr != end);
417 
418 	start &= PUD_MASK;
419 	if (start < floor)
420 		return;
421 	if (ceiling) {
422 		ceiling &= PUD_MASK;
423 		if (!ceiling)
424 			return;
425 	}
426 	if (end - 1 > ceiling - 1)
427 		return;
428 
429 	pmd = pmd_offset(pud, start);
430 	pud_clear(pud);
431 	pmd_free_tlb(tlb, pmd, start);
432 	mm_dec_nr_pmds(tlb->mm);
433 }
434 
free_pud_range(struct mmu_gather * tlb,pgd_t * pgd,unsigned long addr,unsigned long end,unsigned long floor,unsigned long ceiling)435 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
436 				unsigned long addr, unsigned long end,
437 				unsigned long floor, unsigned long ceiling)
438 {
439 	pud_t *pud;
440 	unsigned long next;
441 	unsigned long start;
442 
443 	start = addr;
444 	pud = pud_offset(pgd, addr);
445 	do {
446 		next = pud_addr_end(addr, end);
447 		if (pud_none_or_clear_bad(pud))
448 			continue;
449 		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
450 	} while (pud++, addr = next, addr != end);
451 
452 	start &= PGDIR_MASK;
453 	if (start < floor)
454 		return;
455 	if (ceiling) {
456 		ceiling &= PGDIR_MASK;
457 		if (!ceiling)
458 			return;
459 	}
460 	if (end - 1 > ceiling - 1)
461 		return;
462 
463 	pud = pud_offset(pgd, start);
464 	pgd_clear(pgd);
465 	pud_free_tlb(tlb, pud, start);
466 }
467 
468 /*
469  * This function frees user-level page tables of a process.
470  */
free_pgd_range(struct mmu_gather * tlb,unsigned long addr,unsigned long end,unsigned long floor,unsigned long ceiling)471 void free_pgd_range(struct mmu_gather *tlb,
472 			unsigned long addr, unsigned long end,
473 			unsigned long floor, unsigned long ceiling)
474 {
475 	pgd_t *pgd;
476 	unsigned long next;
477 
478 	/*
479 	 * The next few lines have given us lots of grief...
480 	 *
481 	 * Why are we testing PMD* at this top level?  Because often
482 	 * there will be no work to do at all, and we'd prefer not to
483 	 * go all the way down to the bottom just to discover that.
484 	 *
485 	 * Why all these "- 1"s?  Because 0 represents both the bottom
486 	 * of the address space and the top of it (using -1 for the
487 	 * top wouldn't help much: the masks would do the wrong thing).
488 	 * The rule is that addr 0 and floor 0 refer to the bottom of
489 	 * the address space, but end 0 and ceiling 0 refer to the top
490 	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
491 	 * that end 0 case should be mythical).
492 	 *
493 	 * Wherever addr is brought up or ceiling brought down, we must
494 	 * be careful to reject "the opposite 0" before it confuses the
495 	 * subsequent tests.  But what about where end is brought down
496 	 * by PMD_SIZE below? no, end can't go down to 0 there.
497 	 *
498 	 * Whereas we round start (addr) and ceiling down, by different
499 	 * masks at different levels, in order to test whether a table
500 	 * now has no other vmas using it, so can be freed, we don't
501 	 * bother to round floor or end up - the tests don't need that.
502 	 */
503 
504 	addr &= PMD_MASK;
505 	if (addr < floor) {
506 		addr += PMD_SIZE;
507 		if (!addr)
508 			return;
509 	}
510 	if (ceiling) {
511 		ceiling &= PMD_MASK;
512 		if (!ceiling)
513 			return;
514 	}
515 	if (end - 1 > ceiling - 1)
516 		end -= PMD_SIZE;
517 	if (addr > end - 1)
518 		return;
519 
520 	pgd = pgd_offset(tlb->mm, addr);
521 	do {
522 		next = pgd_addr_end(addr, end);
523 		if (pgd_none_or_clear_bad(pgd))
524 			continue;
525 		free_pud_range(tlb, pgd, addr, next, floor, ceiling);
526 	} while (pgd++, addr = next, addr != end);
527 }
528 
free_pgtables(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long floor,unsigned long ceiling)529 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
530 		unsigned long floor, unsigned long ceiling)
531 {
532 	while (vma) {
533 		struct vm_area_struct *next = vma->vm_next;
534 		unsigned long addr = vma->vm_start;
535 
536 		/*
537 		 * Hide vma from rmap and truncate_pagecache before freeing
538 		 * pgtables
539 		 */
540 		unlink_anon_vmas(vma);
541 		unlink_file_vma(vma);
542 
543 		if (is_vm_hugetlb_page(vma)) {
544 			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
545 				floor, next? next->vm_start: ceiling);
546 		} else {
547 			/*
548 			 * Optimization: gather nearby vmas into one call down
549 			 */
550 			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
551 			       && !is_vm_hugetlb_page(next)) {
552 				vma = next;
553 				next = vma->vm_next;
554 				unlink_anon_vmas(vma);
555 				unlink_file_vma(vma);
556 			}
557 			free_pgd_range(tlb, addr, vma->vm_end,
558 				floor, next? next->vm_start: ceiling);
559 		}
560 		vma = next;
561 	}
562 }
563 
__pte_alloc(struct mm_struct * mm,struct vm_area_struct * vma,pmd_t * pmd,unsigned long address)564 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
565 		pmd_t *pmd, unsigned long address)
566 {
567 	spinlock_t *ptl;
568 	pgtable_t new = pte_alloc_one(mm, address);
569 	int wait_split_huge_page;
570 	if (!new)
571 		return -ENOMEM;
572 
573 	/*
574 	 * Ensure all pte setup (eg. pte page lock and page clearing) are
575 	 * visible before the pte is made visible to other CPUs by being
576 	 * put into page tables.
577 	 *
578 	 * The other side of the story is the pointer chasing in the page
579 	 * table walking code (when walking the page table without locking;
580 	 * ie. most of the time). Fortunately, these data accesses consist
581 	 * of a chain of data-dependent loads, meaning most CPUs (alpha
582 	 * being the notable exception) will already guarantee loads are
583 	 * seen in-order. See the alpha page table accessors for the
584 	 * smp_read_barrier_depends() barriers in page table walking code.
585 	 */
586 	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
587 
588 	ptl = pmd_lock(mm, pmd);
589 	wait_split_huge_page = 0;
590 	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
591 		atomic_long_inc(&mm->nr_ptes);
592 		pmd_populate(mm, pmd, new);
593 		new = NULL;
594 	} else if (unlikely(pmd_trans_splitting(*pmd)))
595 		wait_split_huge_page = 1;
596 	spin_unlock(ptl);
597 	if (new)
598 		pte_free(mm, new);
599 	if (wait_split_huge_page)
600 		wait_split_huge_page(vma->anon_vma, pmd);
601 	return 0;
602 }
603 
__pte_alloc_kernel(pmd_t * pmd,unsigned long address)604 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
605 {
606 	pte_t *new = pte_alloc_one_kernel(&init_mm, address);
607 	if (!new)
608 		return -ENOMEM;
609 
610 	smp_wmb(); /* See comment in __pte_alloc */
611 
612 	spin_lock(&init_mm.page_table_lock);
613 	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
614 		pmd_populate_kernel(&init_mm, pmd, new);
615 		new = NULL;
616 	} else
617 		VM_BUG_ON(pmd_trans_splitting(*pmd));
618 	spin_unlock(&init_mm.page_table_lock);
619 	if (new)
620 		pte_free_kernel(&init_mm, new);
621 	return 0;
622 }
623 
init_rss_vec(int * rss)624 static inline void init_rss_vec(int *rss)
625 {
626 	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
627 }
628 
add_mm_rss_vec(struct mm_struct * mm,int * rss)629 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
630 {
631 	int i;
632 
633 	if (current->mm == mm)
634 		sync_mm_rss(mm);
635 	for (i = 0; i < NR_MM_COUNTERS; i++)
636 		if (rss[i])
637 			add_mm_counter(mm, i, rss[i]);
638 }
639 
640 /*
641  * This function is called to print an error when a bad pte
642  * is found. For example, we might have a PFN-mapped pte in
643  * a region that doesn't allow it.
644  *
645  * The calling function must still handle the error.
646  */
print_bad_pte(struct vm_area_struct * vma,unsigned long addr,pte_t pte,struct page * page)647 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
648 			  pte_t pte, struct page *page)
649 {
650 	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
651 	pud_t *pud = pud_offset(pgd, addr);
652 	pmd_t *pmd = pmd_offset(pud, addr);
653 	struct address_space *mapping;
654 	pgoff_t index;
655 	static unsigned long resume;
656 	static unsigned long nr_shown;
657 	static unsigned long nr_unshown;
658 
659 	/*
660 	 * Allow a burst of 60 reports, then keep quiet for that minute;
661 	 * or allow a steady drip of one report per second.
662 	 */
663 	if (nr_shown == 60) {
664 		if (time_before(jiffies, resume)) {
665 			nr_unshown++;
666 			return;
667 		}
668 		if (nr_unshown) {
669 			printk(KERN_ALERT
670 				"BUG: Bad page map: %lu messages suppressed\n",
671 				nr_unshown);
672 			nr_unshown = 0;
673 		}
674 		nr_shown = 0;
675 	}
676 	if (nr_shown++ == 0)
677 		resume = jiffies + 60 * HZ;
678 
679 	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
680 	index = linear_page_index(vma, addr);
681 
682 	printk(KERN_ALERT
683 		"BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
684 		current->comm,
685 		(long long)pte_val(pte), (long long)pmd_val(*pmd));
686 	if (page)
687 		dump_page(page, "bad pte");
688 	printk(KERN_ALERT
689 		"addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
690 		(void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
691 	/*
692 	 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
693 	 */
694 	pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
695 		 vma->vm_file,
696 		 vma->vm_ops ? vma->vm_ops->fault : NULL,
697 		 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
698 		 mapping ? mapping->a_ops->readpage : NULL);
699 	dump_stack();
700 	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
701 }
702 
703 /*
704  * vm_normal_page -- This function gets the "struct page" associated with a pte.
705  *
706  * "Special" mappings do not wish to be associated with a "struct page" (either
707  * it doesn't exist, or it exists but they don't want to touch it). In this
708  * case, NULL is returned here. "Normal" mappings do have a struct page.
709  *
710  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
711  * pte bit, in which case this function is trivial. Secondly, an architecture
712  * may not have a spare pte bit, which requires a more complicated scheme,
713  * described below.
714  *
715  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
716  * special mapping (even if there are underlying and valid "struct pages").
717  * COWed pages of a VM_PFNMAP are always normal.
718  *
719  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
720  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
721  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
722  * mapping will always honor the rule
723  *
724  *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
725  *
726  * And for normal mappings this is false.
727  *
728  * This restricts such mappings to be a linear translation from virtual address
729  * to pfn. To get around this restriction, we allow arbitrary mappings so long
730  * as the vma is not a COW mapping; in that case, we know that all ptes are
731  * special (because none can have been COWed).
732  *
733  *
734  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
735  *
736  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
737  * page" backing, however the difference is that _all_ pages with a struct
738  * page (that is, those where pfn_valid is true) are refcounted and considered
739  * normal pages by the VM. The disadvantage is that pages are refcounted
740  * (which can be slower and simply not an option for some PFNMAP users). The
741  * advantage is that we don't have to follow the strict linearity rule of
742  * PFNMAP mappings in order to support COWable mappings.
743  *
744  */
745 #ifdef __HAVE_ARCH_PTE_SPECIAL
746 # define HAVE_PTE_SPECIAL 1
747 #else
748 # define HAVE_PTE_SPECIAL 0
749 #endif
vm_normal_page(struct vm_area_struct * vma,unsigned long addr,pte_t pte)750 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
751 				pte_t pte)
752 {
753 	unsigned long pfn = pte_pfn(pte);
754 
755 	if (HAVE_PTE_SPECIAL) {
756 		if (likely(!pte_special(pte)))
757 			goto check_pfn;
758 		if (vma->vm_ops && vma->vm_ops->find_special_page)
759 			return vma->vm_ops->find_special_page(vma, addr);
760 		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
761 			return NULL;
762 		if (!is_zero_pfn(pfn))
763 			print_bad_pte(vma, addr, pte, NULL);
764 		return NULL;
765 	}
766 
767 	/* !HAVE_PTE_SPECIAL case follows: */
768 
769 	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
770 		if (vma->vm_flags & VM_MIXEDMAP) {
771 			if (!pfn_valid(pfn))
772 				return NULL;
773 			goto out;
774 		} else {
775 			unsigned long off;
776 			off = (addr - vma->vm_start) >> PAGE_SHIFT;
777 			if (pfn == vma->vm_pgoff + off)
778 				return NULL;
779 			if (!is_cow_mapping(vma->vm_flags))
780 				return NULL;
781 		}
782 	}
783 
784 	if (is_zero_pfn(pfn))
785 		return NULL;
786 check_pfn:
787 	if (unlikely(pfn > highest_memmap_pfn)) {
788 		print_bad_pte(vma, addr, pte, NULL);
789 		return NULL;
790 	}
791 
792 	/*
793 	 * NOTE! We still have PageReserved() pages in the page tables.
794 	 * eg. VDSO mappings can cause them to exist.
795 	 */
796 out:
797 	return pfn_to_page(pfn);
798 }
799 
800 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
vm_normal_page_pmd(struct vm_area_struct * vma,unsigned long addr,pmd_t pmd)801 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
802 				pmd_t pmd)
803 {
804 	unsigned long pfn = pmd_pfn(pmd);
805 
806 	/*
807 	 * There is no pmd_special() but there may be special pmds, e.g.
808 	 * in a direct-access (dax) mapping, so let's just replicate the
809 	 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
810 	 */
811 	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
812 		if (vma->vm_flags & VM_MIXEDMAP) {
813 			if (!pfn_valid(pfn))
814 				return NULL;
815 			goto out;
816 		} else {
817 			unsigned long off;
818 			off = (addr - vma->vm_start) >> PAGE_SHIFT;
819 			if (pfn == vma->vm_pgoff + off)
820 				return NULL;
821 			if (!is_cow_mapping(vma->vm_flags))
822 				return NULL;
823 		}
824 	}
825 
826 	if (is_zero_pfn(pfn))
827 		return NULL;
828 	if (unlikely(pfn > highest_memmap_pfn))
829 		return NULL;
830 
831 	/*
832 	 * NOTE! We still have PageReserved() pages in the page tables.
833 	 * eg. VDSO mappings can cause them to exist.
834 	 */
835 out:
836 	return pfn_to_page(pfn);
837 }
838 #endif
839 
840 /*
841  * copy one vm_area from one task to the other. Assumes the page tables
842  * already present in the new task to be cleared in the whole range
843  * covered by this vma.
844  */
845 
846 static inline unsigned long
copy_one_pte(struct mm_struct * dst_mm,struct mm_struct * src_mm,pte_t * dst_pte,pte_t * src_pte,struct vm_area_struct * vma,unsigned long addr,int * rss)847 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
848 		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
849 		unsigned long addr, int *rss)
850 {
851 	unsigned long vm_flags = vma->vm_flags;
852 	pte_t pte = *src_pte;
853 	struct page *page;
854 
855 	/* pte contains position in swap or file, so copy. */
856 	if (unlikely(!pte_present(pte))) {
857 		swp_entry_t entry = pte_to_swp_entry(pte);
858 
859 		if (likely(!non_swap_entry(entry))) {
860 			if (swap_duplicate(entry) < 0)
861 				return entry.val;
862 
863 			/* make sure dst_mm is on swapoff's mmlist. */
864 			if (unlikely(list_empty(&dst_mm->mmlist))) {
865 				spin_lock(&mmlist_lock);
866 				if (list_empty(&dst_mm->mmlist))
867 					list_add(&dst_mm->mmlist,
868 							&src_mm->mmlist);
869 				spin_unlock(&mmlist_lock);
870 			}
871 			rss[MM_SWAPENTS]++;
872 		} else if (is_migration_entry(entry)) {
873 			page = migration_entry_to_page(entry);
874 
875 			if (PageAnon(page))
876 				rss[MM_ANONPAGES]++;
877 			else
878 				rss[MM_FILEPAGES]++;
879 
880 			if (is_write_migration_entry(entry) &&
881 					is_cow_mapping(vm_flags)) {
882 				/*
883 				 * COW mappings require pages in both
884 				 * parent and child to be set to read.
885 				 */
886 				make_migration_entry_read(&entry);
887 				pte = swp_entry_to_pte(entry);
888 				if (pte_swp_soft_dirty(*src_pte))
889 					pte = pte_swp_mksoft_dirty(pte);
890 				set_pte_at(src_mm, addr, src_pte, pte);
891 			}
892 		}
893 		goto out_set_pte;
894 	}
895 
896 	/*
897 	 * If it's a COW mapping, write protect it both
898 	 * in the parent and the child
899 	 */
900 	if (is_cow_mapping(vm_flags)) {
901 		ptep_set_wrprotect(src_mm, addr, src_pte);
902 		pte = pte_wrprotect(pte);
903 	}
904 
905 	/*
906 	 * If it's a shared mapping, mark it clean in
907 	 * the child
908 	 */
909 	if (vm_flags & VM_SHARED)
910 		pte = pte_mkclean(pte);
911 	pte = pte_mkold(pte);
912 
913 	page = vm_normal_page(vma, addr, pte);
914 	if (page) {
915 		get_page(page);
916 		page_dup_rmap(page);
917 		if (PageAnon(page))
918 			rss[MM_ANONPAGES]++;
919 		else
920 			rss[MM_FILEPAGES]++;
921 	}
922 
923 out_set_pte:
924 	set_pte_at(dst_mm, addr, dst_pte, pte);
925 	return 0;
926 }
927 
copy_pte_range(struct mm_struct * dst_mm,struct mm_struct * src_mm,pmd_t * dst_pmd,pmd_t * src_pmd,struct vm_area_struct * vma,unsigned long addr,unsigned long end)928 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
929 		   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
930 		   unsigned long addr, unsigned long end)
931 {
932 	pte_t *orig_src_pte, *orig_dst_pte;
933 	pte_t *src_pte, *dst_pte;
934 	spinlock_t *src_ptl, *dst_ptl;
935 	int progress = 0;
936 	int rss[NR_MM_COUNTERS];
937 	swp_entry_t entry = (swp_entry_t){0};
938 
939 again:
940 	init_rss_vec(rss);
941 
942 	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
943 	if (!dst_pte)
944 		return -ENOMEM;
945 	src_pte = pte_offset_map(src_pmd, addr);
946 	src_ptl = pte_lockptr(src_mm, src_pmd);
947 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
948 	orig_src_pte = src_pte;
949 	orig_dst_pte = dst_pte;
950 	arch_enter_lazy_mmu_mode();
951 
952 	do {
953 		/*
954 		 * We are holding two locks at this point - either of them
955 		 * could generate latencies in another task on another CPU.
956 		 */
957 		if (progress >= 32) {
958 			progress = 0;
959 			if (need_resched() ||
960 			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
961 				break;
962 		}
963 		if (pte_none(*src_pte)) {
964 			progress++;
965 			continue;
966 		}
967 		entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
968 							vma, addr, rss);
969 		if (entry.val)
970 			break;
971 		progress += 8;
972 	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
973 
974 	arch_leave_lazy_mmu_mode();
975 	spin_unlock(src_ptl);
976 	pte_unmap(orig_src_pte);
977 	add_mm_rss_vec(dst_mm, rss);
978 	pte_unmap_unlock(orig_dst_pte, dst_ptl);
979 	cond_resched();
980 
981 	if (entry.val) {
982 		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
983 			return -ENOMEM;
984 		progress = 0;
985 	}
986 	if (addr != end)
987 		goto again;
988 	return 0;
989 }
990 
copy_pmd_range(struct mm_struct * dst_mm,struct mm_struct * src_mm,pud_t * dst_pud,pud_t * src_pud,struct vm_area_struct * vma,unsigned long addr,unsigned long end)991 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
992 		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
993 		unsigned long addr, unsigned long end)
994 {
995 	pmd_t *src_pmd, *dst_pmd;
996 	unsigned long next;
997 
998 	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
999 	if (!dst_pmd)
1000 		return -ENOMEM;
1001 	src_pmd = pmd_offset(src_pud, addr);
1002 	do {
1003 		next = pmd_addr_end(addr, end);
1004 		if (pmd_trans_huge(*src_pmd)) {
1005 			int err;
1006 			VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1007 			err = copy_huge_pmd(dst_mm, src_mm,
1008 					    dst_pmd, src_pmd, addr, vma);
1009 			if (err == -ENOMEM)
1010 				return -ENOMEM;
1011 			if (!err)
1012 				continue;
1013 			/* fall through */
1014 		}
1015 		if (pmd_none_or_clear_bad(src_pmd))
1016 			continue;
1017 		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1018 						vma, addr, next))
1019 			return -ENOMEM;
1020 	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
1021 	return 0;
1022 }
1023 
copy_pud_range(struct mm_struct * dst_mm,struct mm_struct * src_mm,pgd_t * dst_pgd,pgd_t * src_pgd,struct vm_area_struct * vma,unsigned long addr,unsigned long end)1024 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1025 		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1026 		unsigned long addr, unsigned long end)
1027 {
1028 	pud_t *src_pud, *dst_pud;
1029 	unsigned long next;
1030 
1031 	dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1032 	if (!dst_pud)
1033 		return -ENOMEM;
1034 	src_pud = pud_offset(src_pgd, addr);
1035 	do {
1036 		next = pud_addr_end(addr, end);
1037 		if (pud_none_or_clear_bad(src_pud))
1038 			continue;
1039 		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1040 						vma, addr, next))
1041 			return -ENOMEM;
1042 	} while (dst_pud++, src_pud++, addr = next, addr != end);
1043 	return 0;
1044 }
1045 
copy_page_range(struct mm_struct * dst_mm,struct mm_struct * src_mm,struct vm_area_struct * vma)1046 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1047 		struct vm_area_struct *vma)
1048 {
1049 	pgd_t *src_pgd, *dst_pgd;
1050 	unsigned long next;
1051 	unsigned long addr = vma->vm_start;
1052 	unsigned long end = vma->vm_end;
1053 	unsigned long mmun_start;	/* For mmu_notifiers */
1054 	unsigned long mmun_end;		/* For mmu_notifiers */
1055 	bool is_cow;
1056 	int ret;
1057 
1058 	/*
1059 	 * Don't copy ptes where a page fault will fill them correctly.
1060 	 * Fork becomes much lighter when there are big shared or private
1061 	 * readonly mappings. The tradeoff is that copy_page_range is more
1062 	 * efficient than faulting.
1063 	 */
1064 	if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1065 			!vma->anon_vma)
1066 		return 0;
1067 
1068 	if (is_vm_hugetlb_page(vma))
1069 		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1070 
1071 	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1072 		/*
1073 		 * We do not free on error cases below as remove_vma
1074 		 * gets called on error from higher level routine
1075 		 */
1076 		ret = track_pfn_copy(vma);
1077 		if (ret)
1078 			return ret;
1079 	}
1080 
1081 	/*
1082 	 * We need to invalidate the secondary MMU mappings only when
1083 	 * there could be a permission downgrade on the ptes of the
1084 	 * parent mm. And a permission downgrade will only happen if
1085 	 * is_cow_mapping() returns true.
1086 	 */
1087 	is_cow = is_cow_mapping(vma->vm_flags);
1088 	mmun_start = addr;
1089 	mmun_end   = end;
1090 	if (is_cow)
1091 		mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1092 						    mmun_end);
1093 
1094 	ret = 0;
1095 	dst_pgd = pgd_offset(dst_mm, addr);
1096 	src_pgd = pgd_offset(src_mm, addr);
1097 	do {
1098 		next = pgd_addr_end(addr, end);
1099 		if (pgd_none_or_clear_bad(src_pgd))
1100 			continue;
1101 		if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1102 					    vma, addr, next))) {
1103 			ret = -ENOMEM;
1104 			break;
1105 		}
1106 	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1107 
1108 	if (is_cow)
1109 		mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1110 	return ret;
1111 }
1112 
zap_pte_range(struct mmu_gather * tlb,struct vm_area_struct * vma,pmd_t * pmd,unsigned long addr,unsigned long end,struct zap_details * details)1113 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1114 				struct vm_area_struct *vma, pmd_t *pmd,
1115 				unsigned long addr, unsigned long end,
1116 				struct zap_details *details)
1117 {
1118 	struct mm_struct *mm = tlb->mm;
1119 	int force_flush = 0;
1120 	int rss[NR_MM_COUNTERS];
1121 	spinlock_t *ptl;
1122 	pte_t *start_pte;
1123 	pte_t *pte;
1124 	swp_entry_t entry;
1125 
1126 again:
1127 	init_rss_vec(rss);
1128 	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1129 	pte = start_pte;
1130 	arch_enter_lazy_mmu_mode();
1131 	do {
1132 		pte_t ptent = *pte;
1133 		if (pte_none(ptent)) {
1134 			continue;
1135 		}
1136 
1137 		if (pte_present(ptent)) {
1138 			struct page *page;
1139 
1140 			page = vm_normal_page(vma, addr, ptent);
1141 			if (unlikely(details) && page) {
1142 				/*
1143 				 * unmap_shared_mapping_pages() wants to
1144 				 * invalidate cache without truncating:
1145 				 * unmap shared but keep private pages.
1146 				 */
1147 				if (details->check_mapping &&
1148 				    details->check_mapping != page->mapping)
1149 					continue;
1150 			}
1151 			ptent = ptep_get_and_clear_full(mm, addr, pte,
1152 							tlb->fullmm);
1153 			tlb_remove_tlb_entry(tlb, pte, addr);
1154 			if (unlikely(!page))
1155 				continue;
1156 			if (PageAnon(page))
1157 				rss[MM_ANONPAGES]--;
1158 			else {
1159 				if (pte_dirty(ptent)) {
1160 					force_flush = 1;
1161 					set_page_dirty(page);
1162 				}
1163 				if (pte_young(ptent) &&
1164 				    likely(!(vma->vm_flags & VM_SEQ_READ)))
1165 					mark_page_accessed(page);
1166 				rss[MM_FILEPAGES]--;
1167 			}
1168 			page_remove_rmap(page);
1169 			if (unlikely(page_mapcount(page) < 0))
1170 				print_bad_pte(vma, addr, ptent, page);
1171 			if (unlikely(!__tlb_remove_page(tlb, page))) {
1172 				force_flush = 1;
1173 				addr += PAGE_SIZE;
1174 				break;
1175 			}
1176 			continue;
1177 		}
1178 		/* If details->check_mapping, we leave swap entries. */
1179 		if (unlikely(details))
1180 			continue;
1181 
1182 		entry = pte_to_swp_entry(ptent);
1183 		if (!non_swap_entry(entry))
1184 			rss[MM_SWAPENTS]--;
1185 		else if (is_migration_entry(entry)) {
1186 			struct page *page;
1187 
1188 			page = migration_entry_to_page(entry);
1189 
1190 			if (PageAnon(page))
1191 				rss[MM_ANONPAGES]--;
1192 			else
1193 				rss[MM_FILEPAGES]--;
1194 		}
1195 		if (unlikely(!free_swap_and_cache(entry)))
1196 			print_bad_pte(vma, addr, ptent, NULL);
1197 		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198 	} while (pte++, addr += PAGE_SIZE, addr != end);
1199 
1200 	add_mm_rss_vec(mm, rss);
1201 	arch_leave_lazy_mmu_mode();
1202 
1203 	/* Do the actual TLB flush before dropping ptl */
1204 	if (force_flush)
1205 		tlb_flush_mmu_tlbonly(tlb);
1206 	pte_unmap_unlock(start_pte, ptl);
1207 
1208 	/*
1209 	 * If we forced a TLB flush (either due to running out of
1210 	 * batch buffers or because we needed to flush dirty TLB
1211 	 * entries before releasing the ptl), free the batched
1212 	 * memory too. Restart if we didn't do everything.
1213 	 */
1214 	if (force_flush) {
1215 		force_flush = 0;
1216 		tlb_flush_mmu_free(tlb);
1217 
1218 		if (addr != end)
1219 			goto again;
1220 	}
1221 
1222 	return addr;
1223 }
1224 
zap_pmd_range(struct mmu_gather * tlb,struct vm_area_struct * vma,pud_t * pud,unsigned long addr,unsigned long end,struct zap_details * details)1225 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1226 				struct vm_area_struct *vma, pud_t *pud,
1227 				unsigned long addr, unsigned long end,
1228 				struct zap_details *details)
1229 {
1230 	pmd_t *pmd;
1231 	unsigned long next;
1232 
1233 	pmd = pmd_offset(pud, addr);
1234 	do {
1235 		next = pmd_addr_end(addr, end);
1236 		if (pmd_trans_huge(*pmd)) {
1237 			if (next - addr != HPAGE_PMD_SIZE) {
1238 #ifdef CONFIG_DEBUG_VM
1239 				if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1240 					pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1241 						__func__, addr, end,
1242 						vma->vm_start,
1243 						vma->vm_end);
1244 					BUG();
1245 				}
1246 #endif
1247 				split_huge_page_pmd(vma, addr, pmd);
1248 			} else if (zap_huge_pmd(tlb, vma, pmd, addr))
1249 				goto next;
1250 			/* fall through */
1251 		}
1252 		/*
1253 		 * Here there can be other concurrent MADV_DONTNEED or
1254 		 * trans huge page faults running, and if the pmd is
1255 		 * none or trans huge it can change under us. This is
1256 		 * because MADV_DONTNEED holds the mmap_sem in read
1257 		 * mode.
1258 		 */
1259 		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1260 			goto next;
1261 		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1262 next:
1263 		cond_resched();
1264 	} while (pmd++, addr = next, addr != end);
1265 
1266 	return addr;
1267 }
1268 
zap_pud_range(struct mmu_gather * tlb,struct vm_area_struct * vma,pgd_t * pgd,unsigned long addr,unsigned long end,struct zap_details * details)1269 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1270 				struct vm_area_struct *vma, pgd_t *pgd,
1271 				unsigned long addr, unsigned long end,
1272 				struct zap_details *details)
1273 {
1274 	pud_t *pud;
1275 	unsigned long next;
1276 
1277 	pud = pud_offset(pgd, addr);
1278 	do {
1279 		next = pud_addr_end(addr, end);
1280 		if (pud_none_or_clear_bad(pud))
1281 			continue;
1282 		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1283 	} while (pud++, addr = next, addr != end);
1284 
1285 	return addr;
1286 }
1287 
unmap_page_range(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long addr,unsigned long end,struct zap_details * details)1288 static void unmap_page_range(struct mmu_gather *tlb,
1289 			     struct vm_area_struct *vma,
1290 			     unsigned long addr, unsigned long end,
1291 			     struct zap_details *details)
1292 {
1293 	pgd_t *pgd;
1294 	unsigned long next;
1295 
1296 	if (details && !details->check_mapping)
1297 		details = NULL;
1298 
1299 	BUG_ON(addr >= end);
1300 	tlb_start_vma(tlb, vma);
1301 	pgd = pgd_offset(vma->vm_mm, addr);
1302 	do {
1303 		next = pgd_addr_end(addr, end);
1304 		if (pgd_none_or_clear_bad(pgd))
1305 			continue;
1306 		next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 	} while (pgd++, addr = next, addr != end);
1308 	tlb_end_vma(tlb, vma);
1309 }
1310 
1311 
unmap_single_vma(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long start_addr,unsigned long end_addr,struct zap_details * details)1312 static void unmap_single_vma(struct mmu_gather *tlb,
1313 		struct vm_area_struct *vma, unsigned long start_addr,
1314 		unsigned long end_addr,
1315 		struct zap_details *details)
1316 {
1317 	unsigned long start = max(vma->vm_start, start_addr);
1318 	unsigned long end;
1319 
1320 	if (start >= vma->vm_end)
1321 		return;
1322 	end = min(vma->vm_end, end_addr);
1323 	if (end <= vma->vm_start)
1324 		return;
1325 
1326 	if (vma->vm_file)
1327 		uprobe_munmap(vma, start, end);
1328 
1329 	if (unlikely(vma->vm_flags & VM_PFNMAP))
1330 		untrack_pfn(vma, 0, 0);
1331 
1332 	if (start != end) {
1333 		if (unlikely(is_vm_hugetlb_page(vma))) {
1334 			/*
1335 			 * It is undesirable to test vma->vm_file as it
1336 			 * should be non-null for valid hugetlb area.
1337 			 * However, vm_file will be NULL in the error
1338 			 * cleanup path of mmap_region. When
1339 			 * hugetlbfs ->mmap method fails,
1340 			 * mmap_region() nullifies vma->vm_file
1341 			 * before calling this function to clean up.
1342 			 * Since no pte has actually been setup, it is
1343 			 * safe to do nothing in this case.
1344 			 */
1345 			if (vma->vm_file) {
1346 				i_mmap_lock_write(vma->vm_file->f_mapping);
1347 				__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1348 				i_mmap_unlock_write(vma->vm_file->f_mapping);
1349 			}
1350 		} else
1351 			unmap_page_range(tlb, vma, start, end, details);
1352 	}
1353 }
1354 
1355 /**
1356  * unmap_vmas - unmap a range of memory covered by a list of vma's
1357  * @tlb: address of the caller's struct mmu_gather
1358  * @vma: the starting vma
1359  * @start_addr: virtual address at which to start unmapping
1360  * @end_addr: virtual address at which to end unmapping
1361  *
1362  * Unmap all pages in the vma list.
1363  *
1364  * Only addresses between `start' and `end' will be unmapped.
1365  *
1366  * The VMA list must be sorted in ascending virtual address order.
1367  *
1368  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1369  * range after unmap_vmas() returns.  So the only responsibility here is to
1370  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1371  * drops the lock and schedules.
1372  */
unmap_vmas(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long start_addr,unsigned long end_addr)1373 void unmap_vmas(struct mmu_gather *tlb,
1374 		struct vm_area_struct *vma, unsigned long start_addr,
1375 		unsigned long end_addr)
1376 {
1377 	struct mm_struct *mm = vma->vm_mm;
1378 
1379 	mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1380 	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1381 		unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1382 	mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1383 }
1384 
1385 /**
1386  * zap_page_range - remove user pages in a given range
1387  * @vma: vm_area_struct holding the applicable pages
1388  * @start: starting address of pages to zap
1389  * @size: number of bytes to zap
1390  * @details: details of shared cache invalidation
1391  *
1392  * Caller must protect the VMA list
1393  */
zap_page_range(struct vm_area_struct * vma,unsigned long start,unsigned long size,struct zap_details * details)1394 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1395 		unsigned long size, struct zap_details *details)
1396 {
1397 	struct mm_struct *mm = vma->vm_mm;
1398 	struct mmu_gather tlb;
1399 	unsigned long end = start + size;
1400 
1401 	lru_add_drain();
1402 	tlb_gather_mmu(&tlb, mm, start, end);
1403 	update_hiwater_rss(mm);
1404 	mmu_notifier_invalidate_range_start(mm, start, end);
1405 	for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1406 		unmap_single_vma(&tlb, vma, start, end, details);
1407 	mmu_notifier_invalidate_range_end(mm, start, end);
1408 	tlb_finish_mmu(&tlb, start, end);
1409 }
1410 
1411 /**
1412  * zap_page_range_single - remove user pages in a given range
1413  * @vma: vm_area_struct holding the applicable pages
1414  * @address: starting address of pages to zap
1415  * @size: number of bytes to zap
1416  * @details: details of shared cache invalidation
1417  *
1418  * The range must fit into one VMA.
1419  */
zap_page_range_single(struct vm_area_struct * vma,unsigned long address,unsigned long size,struct zap_details * details)1420 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1421 		unsigned long size, struct zap_details *details)
1422 {
1423 	struct mm_struct *mm = vma->vm_mm;
1424 	struct mmu_gather tlb;
1425 	unsigned long end = address + size;
1426 
1427 	lru_add_drain();
1428 	tlb_gather_mmu(&tlb, mm, address, end);
1429 	update_hiwater_rss(mm);
1430 	mmu_notifier_invalidate_range_start(mm, address, end);
1431 	unmap_single_vma(&tlb, vma, address, end, details);
1432 	mmu_notifier_invalidate_range_end(mm, address, end);
1433 	tlb_finish_mmu(&tlb, address, end);
1434 }
1435 
1436 /**
1437  * zap_vma_ptes - remove ptes mapping the vma
1438  * @vma: vm_area_struct holding ptes to be zapped
1439  * @address: starting address of pages to zap
1440  * @size: number of bytes to zap
1441  *
1442  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1443  *
1444  * The entire address range must be fully contained within the vma.
1445  *
1446  * Returns 0 if successful.
1447  */
zap_vma_ptes(struct vm_area_struct * vma,unsigned long address,unsigned long size)1448 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1449 		unsigned long size)
1450 {
1451 	if (address < vma->vm_start || address + size > vma->vm_end ||
1452 	    		!(vma->vm_flags & VM_PFNMAP))
1453 		return -1;
1454 	zap_page_range_single(vma, address, size, NULL);
1455 	return 0;
1456 }
1457 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1458 
__get_locked_pte(struct mm_struct * mm,unsigned long addr,spinlock_t ** ptl)1459 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1460 			spinlock_t **ptl)
1461 {
1462 	pgd_t * pgd = pgd_offset(mm, addr);
1463 	pud_t * pud = pud_alloc(mm, pgd, addr);
1464 	if (pud) {
1465 		pmd_t * pmd = pmd_alloc(mm, pud, addr);
1466 		if (pmd) {
1467 			VM_BUG_ON(pmd_trans_huge(*pmd));
1468 			return pte_alloc_map_lock(mm, pmd, addr, ptl);
1469 		}
1470 	}
1471 	return NULL;
1472 }
1473 
1474 /*
1475  * This is the old fallback for page remapping.
1476  *
1477  * For historical reasons, it only allows reserved pages. Only
1478  * old drivers should use this, and they needed to mark their
1479  * pages reserved for the old functions anyway.
1480  */
insert_page(struct vm_area_struct * vma,unsigned long addr,struct page * page,pgprot_t prot)1481 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1482 			struct page *page, pgprot_t prot)
1483 {
1484 	struct mm_struct *mm = vma->vm_mm;
1485 	int retval;
1486 	pte_t *pte;
1487 	spinlock_t *ptl;
1488 
1489 	retval = -EINVAL;
1490 	if (PageAnon(page))
1491 		goto out;
1492 	retval = -ENOMEM;
1493 	flush_dcache_page(page);
1494 	pte = get_locked_pte(mm, addr, &ptl);
1495 	if (!pte)
1496 		goto out;
1497 	retval = -EBUSY;
1498 	if (!pte_none(*pte))
1499 		goto out_unlock;
1500 
1501 	/* Ok, finally just insert the thing.. */
1502 	get_page(page);
1503 	inc_mm_counter_fast(mm, MM_FILEPAGES);
1504 	page_add_file_rmap(page);
1505 	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1506 
1507 	retval = 0;
1508 	pte_unmap_unlock(pte, ptl);
1509 	return retval;
1510 out_unlock:
1511 	pte_unmap_unlock(pte, ptl);
1512 out:
1513 	return retval;
1514 }
1515 
1516 /**
1517  * vm_insert_page - insert single page into user vma
1518  * @vma: user vma to map to
1519  * @addr: target user address of this page
1520  * @page: source kernel page
1521  *
1522  * This allows drivers to insert individual pages they've allocated
1523  * into a user vma.
1524  *
1525  * The page has to be a nice clean _individual_ kernel allocation.
1526  * If you allocate a compound page, you need to have marked it as
1527  * such (__GFP_COMP), or manually just split the page up yourself
1528  * (see split_page()).
1529  *
1530  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1531  * took an arbitrary page protection parameter. This doesn't allow
1532  * that. Your vma protection will have to be set up correctly, which
1533  * means that if you want a shared writable mapping, you'd better
1534  * ask for a shared writable mapping!
1535  *
1536  * The page does not need to be reserved.
1537  *
1538  * Usually this function is called from f_op->mmap() handler
1539  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1540  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1541  * function from other places, for example from page-fault handler.
1542  */
vm_insert_page(struct vm_area_struct * vma,unsigned long addr,struct page * page)1543 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1544 			struct page *page)
1545 {
1546 	if (addr < vma->vm_start || addr >= vma->vm_end)
1547 		return -EFAULT;
1548 	if (!page_count(page))
1549 		return -EINVAL;
1550 	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1551 		BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1552 		BUG_ON(vma->vm_flags & VM_PFNMAP);
1553 		vma->vm_flags |= VM_MIXEDMAP;
1554 	}
1555 	return insert_page(vma, addr, page, vma->vm_page_prot);
1556 }
1557 EXPORT_SYMBOL(vm_insert_page);
1558 
insert_pfn(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn,pgprot_t prot)1559 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1560 			unsigned long pfn, pgprot_t prot)
1561 {
1562 	struct mm_struct *mm = vma->vm_mm;
1563 	int retval;
1564 	pte_t *pte, entry;
1565 	spinlock_t *ptl;
1566 
1567 	retval = -ENOMEM;
1568 	pte = get_locked_pte(mm, addr, &ptl);
1569 	if (!pte)
1570 		goto out;
1571 	retval = -EBUSY;
1572 	if (!pte_none(*pte))
1573 		goto out_unlock;
1574 
1575 	/* Ok, finally just insert the thing.. */
1576 	entry = pte_mkspecial(pfn_pte(pfn, prot));
1577 	set_pte_at(mm, addr, pte, entry);
1578 	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1579 
1580 	retval = 0;
1581 out_unlock:
1582 	pte_unmap_unlock(pte, ptl);
1583 out:
1584 	return retval;
1585 }
1586 
1587 /**
1588  * vm_insert_pfn - insert single pfn into user vma
1589  * @vma: user vma to map to
1590  * @addr: target user address of this page
1591  * @pfn: source kernel pfn
1592  *
1593  * Similar to vm_insert_page, this allows drivers to insert individual pages
1594  * they've allocated into a user vma. Same comments apply.
1595  *
1596  * This function should only be called from a vm_ops->fault handler, and
1597  * in that case the handler should return NULL.
1598  *
1599  * vma cannot be a COW mapping.
1600  *
1601  * As this is called only for pages that do not currently exist, we
1602  * do not need to flush old virtual caches or the TLB.
1603  */
vm_insert_pfn(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn)1604 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1605 			unsigned long pfn)
1606 {
1607 	int ret;
1608 	pgprot_t pgprot = vma->vm_page_prot;
1609 	/*
1610 	 * Technically, architectures with pte_special can avoid all these
1611 	 * restrictions (same for remap_pfn_range).  However we would like
1612 	 * consistency in testing and feature parity among all, so we should
1613 	 * try to keep these invariants in place for everybody.
1614 	 */
1615 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1616 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1617 						(VM_PFNMAP|VM_MIXEDMAP));
1618 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1619 	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1620 
1621 	if (addr < vma->vm_start || addr >= vma->vm_end)
1622 		return -EFAULT;
1623 	if (track_pfn_insert(vma, &pgprot, pfn))
1624 		return -EINVAL;
1625 
1626 	ret = insert_pfn(vma, addr, pfn, pgprot);
1627 
1628 	return ret;
1629 }
1630 EXPORT_SYMBOL(vm_insert_pfn);
1631 
vm_insert_mixed(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn)1632 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1633 			unsigned long pfn)
1634 {
1635 	BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1636 
1637 	if (addr < vma->vm_start || addr >= vma->vm_end)
1638 		return -EFAULT;
1639 
1640 	/*
1641 	 * If we don't have pte special, then we have to use the pfn_valid()
1642 	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1643 	 * refcount the page if pfn_valid is true (hence insert_page rather
1644 	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1645 	 * without pte special, it would there be refcounted as a normal page.
1646 	 */
1647 	if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1648 		struct page *page;
1649 
1650 		page = pfn_to_page(pfn);
1651 		return insert_page(vma, addr, page, vma->vm_page_prot);
1652 	}
1653 	return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1654 }
1655 EXPORT_SYMBOL(vm_insert_mixed);
1656 
1657 /*
1658  * maps a range of physical memory into the requested pages. the old
1659  * mappings are removed. any references to nonexistent pages results
1660  * in null mappings (currently treated as "copy-on-access")
1661  */
remap_pte_range(struct mm_struct * mm,pmd_t * pmd,unsigned long addr,unsigned long end,unsigned long pfn,pgprot_t prot)1662 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1663 			unsigned long addr, unsigned long end,
1664 			unsigned long pfn, pgprot_t prot)
1665 {
1666 	pte_t *pte;
1667 	spinlock_t *ptl;
1668 
1669 	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1670 	if (!pte)
1671 		return -ENOMEM;
1672 	arch_enter_lazy_mmu_mode();
1673 	do {
1674 		BUG_ON(!pte_none(*pte));
1675 		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1676 		pfn++;
1677 	} while (pte++, addr += PAGE_SIZE, addr != end);
1678 	arch_leave_lazy_mmu_mode();
1679 	pte_unmap_unlock(pte - 1, ptl);
1680 	return 0;
1681 }
1682 
remap_pmd_range(struct mm_struct * mm,pud_t * pud,unsigned long addr,unsigned long end,unsigned long pfn,pgprot_t prot)1683 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1684 			unsigned long addr, unsigned long end,
1685 			unsigned long pfn, pgprot_t prot)
1686 {
1687 	pmd_t *pmd;
1688 	unsigned long next;
1689 
1690 	pfn -= addr >> PAGE_SHIFT;
1691 	pmd = pmd_alloc(mm, pud, addr);
1692 	if (!pmd)
1693 		return -ENOMEM;
1694 	VM_BUG_ON(pmd_trans_huge(*pmd));
1695 	do {
1696 		next = pmd_addr_end(addr, end);
1697 		if (remap_pte_range(mm, pmd, addr, next,
1698 				pfn + (addr >> PAGE_SHIFT), prot))
1699 			return -ENOMEM;
1700 	} while (pmd++, addr = next, addr != end);
1701 	return 0;
1702 }
1703 
remap_pud_range(struct mm_struct * mm,pgd_t * pgd,unsigned long addr,unsigned long end,unsigned long pfn,pgprot_t prot)1704 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1705 			unsigned long addr, unsigned long end,
1706 			unsigned long pfn, pgprot_t prot)
1707 {
1708 	pud_t *pud;
1709 	unsigned long next;
1710 
1711 	pfn -= addr >> PAGE_SHIFT;
1712 	pud = pud_alloc(mm, pgd, addr);
1713 	if (!pud)
1714 		return -ENOMEM;
1715 	do {
1716 		next = pud_addr_end(addr, end);
1717 		if (remap_pmd_range(mm, pud, addr, next,
1718 				pfn + (addr >> PAGE_SHIFT), prot))
1719 			return -ENOMEM;
1720 	} while (pud++, addr = next, addr != end);
1721 	return 0;
1722 }
1723 
1724 /**
1725  * remap_pfn_range - remap kernel memory to userspace
1726  * @vma: user vma to map to
1727  * @addr: target user address to start at
1728  * @pfn: physical address of kernel memory
1729  * @size: size of map area
1730  * @prot: page protection flags for this mapping
1731  *
1732  *  Note: this is only safe if the mm semaphore is held when called.
1733  */
remap_pfn_range(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn,unsigned long size,pgprot_t prot)1734 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1735 		    unsigned long pfn, unsigned long size, pgprot_t prot)
1736 {
1737 	pgd_t *pgd;
1738 	unsigned long next;
1739 	unsigned long end = addr + PAGE_ALIGN(size);
1740 	struct mm_struct *mm = vma->vm_mm;
1741 	int err;
1742 
1743 	/*
1744 	 * Physically remapped pages are special. Tell the
1745 	 * rest of the world about it:
1746 	 *   VM_IO tells people not to look at these pages
1747 	 *	(accesses can have side effects).
1748 	 *   VM_PFNMAP tells the core MM that the base pages are just
1749 	 *	raw PFN mappings, and do not have a "struct page" associated
1750 	 *	with them.
1751 	 *   VM_DONTEXPAND
1752 	 *      Disable vma merging and expanding with mremap().
1753 	 *   VM_DONTDUMP
1754 	 *      Omit vma from core dump, even when VM_IO turned off.
1755 	 *
1756 	 * There's a horrible special case to handle copy-on-write
1757 	 * behaviour that some programs depend on. We mark the "original"
1758 	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1759 	 * See vm_normal_page() for details.
1760 	 */
1761 	if (is_cow_mapping(vma->vm_flags)) {
1762 		if (addr != vma->vm_start || end != vma->vm_end)
1763 			return -EINVAL;
1764 		vma->vm_pgoff = pfn;
1765 	}
1766 
1767 	err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1768 	if (err)
1769 		return -EINVAL;
1770 
1771 	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1772 
1773 	BUG_ON(addr >= end);
1774 	pfn -= addr >> PAGE_SHIFT;
1775 	pgd = pgd_offset(mm, addr);
1776 	flush_cache_range(vma, addr, end);
1777 	do {
1778 		next = pgd_addr_end(addr, end);
1779 		err = remap_pud_range(mm, pgd, addr, next,
1780 				pfn + (addr >> PAGE_SHIFT), prot);
1781 		if (err)
1782 			break;
1783 	} while (pgd++, addr = next, addr != end);
1784 
1785 	if (err)
1786 		untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1787 
1788 	return err;
1789 }
1790 EXPORT_SYMBOL(remap_pfn_range);
1791 
1792 /**
1793  * vm_iomap_memory - remap memory to userspace
1794  * @vma: user vma to map to
1795  * @start: start of area
1796  * @len: size of area
1797  *
1798  * This is a simplified io_remap_pfn_range() for common driver use. The
1799  * driver just needs to give us the physical memory range to be mapped,
1800  * we'll figure out the rest from the vma information.
1801  *
1802  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1803  * whatever write-combining details or similar.
1804  */
vm_iomap_memory(struct vm_area_struct * vma,phys_addr_t start,unsigned long len)1805 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1806 {
1807 	unsigned long vm_len, pfn, pages;
1808 
1809 	/* Check that the physical memory area passed in looks valid */
1810 	if (start + len < start)
1811 		return -EINVAL;
1812 	/*
1813 	 * You *really* shouldn't map things that aren't page-aligned,
1814 	 * but we've historically allowed it because IO memory might
1815 	 * just have smaller alignment.
1816 	 */
1817 	len += start & ~PAGE_MASK;
1818 	pfn = start >> PAGE_SHIFT;
1819 	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1820 	if (pfn + pages < pfn)
1821 		return -EINVAL;
1822 
1823 	/* We start the mapping 'vm_pgoff' pages into the area */
1824 	if (vma->vm_pgoff > pages)
1825 		return -EINVAL;
1826 	pfn += vma->vm_pgoff;
1827 	pages -= vma->vm_pgoff;
1828 
1829 	/* Can we fit all of the mapping? */
1830 	vm_len = vma->vm_end - vma->vm_start;
1831 	if (vm_len >> PAGE_SHIFT > pages)
1832 		return -EINVAL;
1833 
1834 	/* Ok, let it rip */
1835 	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1836 }
1837 EXPORT_SYMBOL(vm_iomap_memory);
1838 
apply_to_pte_range(struct mm_struct * mm,pmd_t * pmd,unsigned long addr,unsigned long end,pte_fn_t fn,void * data)1839 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1840 				     unsigned long addr, unsigned long end,
1841 				     pte_fn_t fn, void *data)
1842 {
1843 	pte_t *pte;
1844 	int err;
1845 	pgtable_t token;
1846 	spinlock_t *uninitialized_var(ptl);
1847 
1848 	pte = (mm == &init_mm) ?
1849 		pte_alloc_kernel(pmd, addr) :
1850 		pte_alloc_map_lock(mm, pmd, addr, &ptl);
1851 	if (!pte)
1852 		return -ENOMEM;
1853 
1854 	BUG_ON(pmd_huge(*pmd));
1855 
1856 	arch_enter_lazy_mmu_mode();
1857 
1858 	token = pmd_pgtable(*pmd);
1859 
1860 	do {
1861 		err = fn(pte++, token, addr, data);
1862 		if (err)
1863 			break;
1864 	} while (addr += PAGE_SIZE, addr != end);
1865 
1866 	arch_leave_lazy_mmu_mode();
1867 
1868 	if (mm != &init_mm)
1869 		pte_unmap_unlock(pte-1, ptl);
1870 	return err;
1871 }
1872 
apply_to_pmd_range(struct mm_struct * mm,pud_t * pud,unsigned long addr,unsigned long end,pte_fn_t fn,void * data)1873 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1874 				     unsigned long addr, unsigned long end,
1875 				     pte_fn_t fn, void *data)
1876 {
1877 	pmd_t *pmd;
1878 	unsigned long next;
1879 	int err;
1880 
1881 	BUG_ON(pud_huge(*pud));
1882 
1883 	pmd = pmd_alloc(mm, pud, addr);
1884 	if (!pmd)
1885 		return -ENOMEM;
1886 	do {
1887 		next = pmd_addr_end(addr, end);
1888 		err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1889 		if (err)
1890 			break;
1891 	} while (pmd++, addr = next, addr != end);
1892 	return err;
1893 }
1894 
apply_to_pud_range(struct mm_struct * mm,pgd_t * pgd,unsigned long addr,unsigned long end,pte_fn_t fn,void * data)1895 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1896 				     unsigned long addr, unsigned long end,
1897 				     pte_fn_t fn, void *data)
1898 {
1899 	pud_t *pud;
1900 	unsigned long next;
1901 	int err;
1902 
1903 	pud = pud_alloc(mm, pgd, addr);
1904 	if (!pud)
1905 		return -ENOMEM;
1906 	do {
1907 		next = pud_addr_end(addr, end);
1908 		err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1909 		if (err)
1910 			break;
1911 	} while (pud++, addr = next, addr != end);
1912 	return err;
1913 }
1914 
1915 /*
1916  * Scan a region of virtual memory, filling in page tables as necessary
1917  * and calling a provided function on each leaf page table.
1918  */
apply_to_page_range(struct mm_struct * mm,unsigned long addr,unsigned long size,pte_fn_t fn,void * data)1919 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1920 			unsigned long size, pte_fn_t fn, void *data)
1921 {
1922 	pgd_t *pgd;
1923 	unsigned long next;
1924 	unsigned long end = addr + size;
1925 	int err;
1926 
1927 	BUG_ON(addr >= end);
1928 	pgd = pgd_offset(mm, addr);
1929 	do {
1930 		next = pgd_addr_end(addr, end);
1931 		err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1932 		if (err)
1933 			break;
1934 	} while (pgd++, addr = next, addr != end);
1935 
1936 	return err;
1937 }
1938 EXPORT_SYMBOL_GPL(apply_to_page_range);
1939 
1940 /*
1941  * handle_pte_fault chooses page fault handler according to an entry which was
1942  * read non-atomically.  Before making any commitment, on those architectures
1943  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1944  * parts, do_swap_page must check under lock before unmapping the pte and
1945  * proceeding (but do_wp_page is only called after already making such a check;
1946  * and do_anonymous_page can safely check later on).
1947  */
pte_unmap_same(struct mm_struct * mm,pmd_t * pmd,pte_t * page_table,pte_t orig_pte)1948 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1949 				pte_t *page_table, pte_t orig_pte)
1950 {
1951 	int same = 1;
1952 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1953 	if (sizeof(pte_t) > sizeof(unsigned long)) {
1954 		spinlock_t *ptl = pte_lockptr(mm, pmd);
1955 		spin_lock(ptl);
1956 		same = pte_same(*page_table, orig_pte);
1957 		spin_unlock(ptl);
1958 	}
1959 #endif
1960 	pte_unmap(page_table);
1961 	return same;
1962 }
1963 
cow_user_page(struct page * dst,struct page * src,unsigned long va,struct vm_area_struct * vma)1964 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1965 {
1966 	debug_dma_assert_idle(src);
1967 
1968 	/*
1969 	 * If the source page was a PFN mapping, we don't have
1970 	 * a "struct page" for it. We do a best-effort copy by
1971 	 * just copying from the original user address. If that
1972 	 * fails, we just zero-fill it. Live with it.
1973 	 */
1974 	if (unlikely(!src)) {
1975 		void *kaddr = kmap_atomic(dst);
1976 		void __user *uaddr = (void __user *)(va & PAGE_MASK);
1977 
1978 		/*
1979 		 * This really shouldn't fail, because the page is there
1980 		 * in the page tables. But it might just be unreadable,
1981 		 * in which case we just give up and fill the result with
1982 		 * zeroes.
1983 		 */
1984 		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1985 			clear_page(kaddr);
1986 		kunmap_atomic(kaddr);
1987 		flush_dcache_page(dst);
1988 	} else
1989 		copy_user_highpage(dst, src, va, vma);
1990 }
1991 
1992 /*
1993  * Notify the address space that the page is about to become writable so that
1994  * it can prohibit this or wait for the page to get into an appropriate state.
1995  *
1996  * We do this without the lock held, so that it can sleep if it needs to.
1997  */
do_page_mkwrite(struct vm_area_struct * vma,struct page * page,unsigned long address)1998 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1999 	       unsigned long address)
2000 {
2001 	struct vm_fault vmf;
2002 	int ret;
2003 
2004 	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2005 	vmf.pgoff = page->index;
2006 	vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2007 	vmf.page = page;
2008 	vmf.cow_page = NULL;
2009 
2010 	ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2011 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2012 		return ret;
2013 	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2014 		lock_page(page);
2015 		if (!page->mapping) {
2016 			unlock_page(page);
2017 			return 0; /* retry */
2018 		}
2019 		ret |= VM_FAULT_LOCKED;
2020 	} else
2021 		VM_BUG_ON_PAGE(!PageLocked(page), page);
2022 	return ret;
2023 }
2024 
2025 /*
2026  * Handle write page faults for pages that can be reused in the current vma
2027  *
2028  * This can happen either due to the mapping being with the VM_SHARED flag,
2029  * or due to us being the last reference standing to the page. In either
2030  * case, all we need to do here is to mark the page as writable and update
2031  * any related book-keeping.
2032  */
wp_page_reuse(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * page_table,spinlock_t * ptl,pte_t orig_pte,struct page * page,int page_mkwrite,int dirty_shared)2033 static inline int wp_page_reuse(struct mm_struct *mm,
2034 			struct vm_area_struct *vma, unsigned long address,
2035 			pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2036 			struct page *page, int page_mkwrite,
2037 			int dirty_shared)
2038 	__releases(ptl)
2039 {
2040 	pte_t entry;
2041 	/*
2042 	 * Clear the pages cpupid information as the existing
2043 	 * information potentially belongs to a now completely
2044 	 * unrelated process.
2045 	 */
2046 	if (page)
2047 		page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2048 
2049 	flush_cache_page(vma, address, pte_pfn(orig_pte));
2050 	entry = pte_mkyoung(orig_pte);
2051 	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2052 	if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2053 		update_mmu_cache(vma, address, page_table);
2054 	pte_unmap_unlock(page_table, ptl);
2055 
2056 	if (dirty_shared) {
2057 		struct address_space *mapping;
2058 		int dirtied;
2059 
2060 		if (!page_mkwrite)
2061 			lock_page(page);
2062 
2063 		dirtied = set_page_dirty(page);
2064 		VM_BUG_ON_PAGE(PageAnon(page), page);
2065 		mapping = page->mapping;
2066 		unlock_page(page);
2067 		page_cache_release(page);
2068 
2069 		if ((dirtied || page_mkwrite) && mapping) {
2070 			/*
2071 			 * Some device drivers do not set page.mapping
2072 			 * but still dirty their pages
2073 			 */
2074 			balance_dirty_pages_ratelimited(mapping);
2075 		}
2076 
2077 		if (!page_mkwrite)
2078 			file_update_time(vma->vm_file);
2079 	}
2080 
2081 	return VM_FAULT_WRITE;
2082 }
2083 
2084 /*
2085  * Handle the case of a page which we actually need to copy to a new page.
2086  *
2087  * Called with mmap_sem locked and the old page referenced, but
2088  * without the ptl held.
2089  *
2090  * High level logic flow:
2091  *
2092  * - Allocate a page, copy the content of the old page to the new one.
2093  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2094  * - Take the PTL. If the pte changed, bail out and release the allocated page
2095  * - If the pte is still the way we remember it, update the page table and all
2096  *   relevant references. This includes dropping the reference the page-table
2097  *   held to the old page, as well as updating the rmap.
2098  * - In any case, unlock the PTL and drop the reference we took to the old page.
2099  */
wp_page_copy(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * page_table,pmd_t * pmd,pte_t orig_pte,struct page * old_page)2100 static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2101 			unsigned long address, pte_t *page_table, pmd_t *pmd,
2102 			pte_t orig_pte, struct page *old_page)
2103 {
2104 	struct page *new_page = NULL;
2105 	spinlock_t *ptl = NULL;
2106 	pte_t entry;
2107 	int page_copied = 0;
2108 	const unsigned long mmun_start = address & PAGE_MASK;	/* For mmu_notifiers */
2109 	const unsigned long mmun_end = mmun_start + PAGE_SIZE;	/* For mmu_notifiers */
2110 	struct mem_cgroup *memcg;
2111 
2112 	if (unlikely(anon_vma_prepare(vma)))
2113 		goto oom;
2114 
2115 	if (is_zero_pfn(pte_pfn(orig_pte))) {
2116 		new_page = alloc_zeroed_user_highpage_movable(vma, address);
2117 		if (!new_page)
2118 			goto oom;
2119 	} else {
2120 		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2121 		if (!new_page)
2122 			goto oom;
2123 		cow_user_page(new_page, old_page, address, vma);
2124 	}
2125 
2126 	if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
2127 		goto oom_free_new;
2128 
2129 	__SetPageUptodate(new_page);
2130 
2131 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2132 
2133 	/*
2134 	 * Re-check the pte - we dropped the lock
2135 	 */
2136 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2137 	if (likely(pte_same(*page_table, orig_pte))) {
2138 		if (old_page) {
2139 			if (!PageAnon(old_page)) {
2140 				dec_mm_counter_fast(mm, MM_FILEPAGES);
2141 				inc_mm_counter_fast(mm, MM_ANONPAGES);
2142 			}
2143 		} else {
2144 			inc_mm_counter_fast(mm, MM_ANONPAGES);
2145 		}
2146 		flush_cache_page(vma, address, pte_pfn(orig_pte));
2147 		entry = mk_pte(new_page, vma->vm_page_prot);
2148 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2149 		/*
2150 		 * Clear the pte entry and flush it first, before updating the
2151 		 * pte with the new entry. This will avoid a race condition
2152 		 * seen in the presence of one thread doing SMC and another
2153 		 * thread doing COW.
2154 		 */
2155 		ptep_clear_flush_notify(vma, address, page_table);
2156 		page_add_new_anon_rmap(new_page, vma, address);
2157 		mem_cgroup_commit_charge(new_page, memcg, false);
2158 		lru_cache_add_active_or_unevictable(new_page, vma);
2159 		/*
2160 		 * We call the notify macro here because, when using secondary
2161 		 * mmu page tables (such as kvm shadow page tables), we want the
2162 		 * new page to be mapped directly into the secondary page table.
2163 		 */
2164 		set_pte_at_notify(mm, address, page_table, entry);
2165 		update_mmu_cache(vma, address, page_table);
2166 		if (old_page) {
2167 			/*
2168 			 * Only after switching the pte to the new page may
2169 			 * we remove the mapcount here. Otherwise another
2170 			 * process may come and find the rmap count decremented
2171 			 * before the pte is switched to the new page, and
2172 			 * "reuse" the old page writing into it while our pte
2173 			 * here still points into it and can be read by other
2174 			 * threads.
2175 			 *
2176 			 * The critical issue is to order this
2177 			 * page_remove_rmap with the ptp_clear_flush above.
2178 			 * Those stores are ordered by (if nothing else,)
2179 			 * the barrier present in the atomic_add_negative
2180 			 * in page_remove_rmap.
2181 			 *
2182 			 * Then the TLB flush in ptep_clear_flush ensures that
2183 			 * no process can access the old page before the
2184 			 * decremented mapcount is visible. And the old page
2185 			 * cannot be reused until after the decremented
2186 			 * mapcount is visible. So transitively, TLBs to
2187 			 * old page will be flushed before it can be reused.
2188 			 */
2189 			page_remove_rmap(old_page);
2190 		}
2191 
2192 		/* Free the old page.. */
2193 		new_page = old_page;
2194 		page_copied = 1;
2195 	} else {
2196 		mem_cgroup_cancel_charge(new_page, memcg);
2197 	}
2198 
2199 	if (new_page)
2200 		page_cache_release(new_page);
2201 
2202 	pte_unmap_unlock(page_table, ptl);
2203 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2204 	if (old_page) {
2205 		/*
2206 		 * Don't let another task, with possibly unlocked vma,
2207 		 * keep the mlocked page.
2208 		 */
2209 		if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2210 			lock_page(old_page);	/* LRU manipulation */
2211 			munlock_vma_page(old_page);
2212 			unlock_page(old_page);
2213 		}
2214 		page_cache_release(old_page);
2215 	}
2216 	return page_copied ? VM_FAULT_WRITE : 0;
2217 oom_free_new:
2218 	page_cache_release(new_page);
2219 oom:
2220 	if (old_page)
2221 		page_cache_release(old_page);
2222 	return VM_FAULT_OOM;
2223 }
2224 
2225 /*
2226  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2227  * mapping
2228  */
wp_pfn_shared(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * page_table,spinlock_t * ptl,pte_t orig_pte,pmd_t * pmd)2229 static int wp_pfn_shared(struct mm_struct *mm,
2230 			struct vm_area_struct *vma, unsigned long address,
2231 			pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2232 			pmd_t *pmd)
2233 {
2234 	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2235 		struct vm_fault vmf = {
2236 			.page = NULL,
2237 			.pgoff = linear_page_index(vma, address),
2238 			.virtual_address = (void __user *)(address & PAGE_MASK),
2239 			.flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2240 		};
2241 		int ret;
2242 
2243 		pte_unmap_unlock(page_table, ptl);
2244 		ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2245 		if (ret & VM_FAULT_ERROR)
2246 			return ret;
2247 		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2248 		/*
2249 		 * We might have raced with another page fault while we
2250 		 * released the pte_offset_map_lock.
2251 		 */
2252 		if (!pte_same(*page_table, orig_pte)) {
2253 			pte_unmap_unlock(page_table, ptl);
2254 			return 0;
2255 		}
2256 	}
2257 	return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2258 			     NULL, 0, 0);
2259 }
2260 
wp_page_shared(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * page_table,pmd_t * pmd,spinlock_t * ptl,pte_t orig_pte,struct page * old_page)2261 static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2262 			  unsigned long address, pte_t *page_table,
2263 			  pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2264 			  struct page *old_page)
2265 	__releases(ptl)
2266 {
2267 	int page_mkwrite = 0;
2268 
2269 	page_cache_get(old_page);
2270 
2271 	/*
2272 	 * Only catch write-faults on shared writable pages,
2273 	 * read-only shared pages can get COWed by
2274 	 * get_user_pages(.write=1, .force=1).
2275 	 */
2276 	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2277 		int tmp;
2278 
2279 		pte_unmap_unlock(page_table, ptl);
2280 		tmp = do_page_mkwrite(vma, old_page, address);
2281 		if (unlikely(!tmp || (tmp &
2282 				      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2283 			page_cache_release(old_page);
2284 			return tmp;
2285 		}
2286 		/*
2287 		 * Since we dropped the lock we need to revalidate
2288 		 * the PTE as someone else may have changed it.  If
2289 		 * they did, we just return, as we can count on the
2290 		 * MMU to tell us if they didn't also make it writable.
2291 		 */
2292 		page_table = pte_offset_map_lock(mm, pmd, address,
2293 						 &ptl);
2294 		if (!pte_same(*page_table, orig_pte)) {
2295 			unlock_page(old_page);
2296 			pte_unmap_unlock(page_table, ptl);
2297 			page_cache_release(old_page);
2298 			return 0;
2299 		}
2300 		page_mkwrite = 1;
2301 	}
2302 
2303 	return wp_page_reuse(mm, vma, address, page_table, ptl,
2304 			     orig_pte, old_page, page_mkwrite, 1);
2305 }
2306 
2307 /*
2308  * This routine handles present pages, when users try to write
2309  * to a shared page. It is done by copying the page to a new address
2310  * and decrementing the shared-page counter for the old page.
2311  *
2312  * Note that this routine assumes that the protection checks have been
2313  * done by the caller (the low-level page fault routine in most cases).
2314  * Thus we can safely just mark it writable once we've done any necessary
2315  * COW.
2316  *
2317  * We also mark the page dirty at this point even though the page will
2318  * change only once the write actually happens. This avoids a few races,
2319  * and potentially makes it more efficient.
2320  *
2321  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2322  * but allow concurrent faults), with pte both mapped and locked.
2323  * We return with mmap_sem still held, but pte unmapped and unlocked.
2324  */
do_wp_page(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * page_table,pmd_t * pmd,spinlock_t * ptl,pte_t orig_pte)2325 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2326 		unsigned long address, pte_t *page_table, pmd_t *pmd,
2327 		spinlock_t *ptl, pte_t orig_pte)
2328 	__releases(ptl)
2329 {
2330 	struct page *old_page;
2331 
2332 	old_page = vm_normal_page(vma, address, orig_pte);
2333 	if (!old_page) {
2334 		/*
2335 		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2336 		 * VM_PFNMAP VMA.
2337 		 *
2338 		 * We should not cow pages in a shared writeable mapping.
2339 		 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2340 		 */
2341 		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2342 				     (VM_WRITE|VM_SHARED))
2343 			return wp_pfn_shared(mm, vma, address, page_table, ptl,
2344 					     orig_pte, pmd);
2345 
2346 		pte_unmap_unlock(page_table, ptl);
2347 		return wp_page_copy(mm, vma, address, page_table, pmd,
2348 				    orig_pte, old_page);
2349 	}
2350 
2351 	/*
2352 	 * Take out anonymous pages first, anonymous shared vmas are
2353 	 * not dirty accountable.
2354 	 */
2355 	if (PageAnon(old_page) && !PageKsm(old_page)) {
2356 		if (!trylock_page(old_page)) {
2357 			page_cache_get(old_page);
2358 			pte_unmap_unlock(page_table, ptl);
2359 			lock_page(old_page);
2360 			page_table = pte_offset_map_lock(mm, pmd, address,
2361 							 &ptl);
2362 			if (!pte_same(*page_table, orig_pte)) {
2363 				unlock_page(old_page);
2364 				pte_unmap_unlock(page_table, ptl);
2365 				page_cache_release(old_page);
2366 				return 0;
2367 			}
2368 			page_cache_release(old_page);
2369 		}
2370 		if (reuse_swap_page(old_page)) {
2371 			/*
2372 			 * The page is all ours.  Move it to our anon_vma so
2373 			 * the rmap code will not search our parent or siblings.
2374 			 * Protected against the rmap code by the page lock.
2375 			 */
2376 			page_move_anon_rmap(old_page, vma, address);
2377 			unlock_page(old_page);
2378 			return wp_page_reuse(mm, vma, address, page_table, ptl,
2379 					     orig_pte, old_page, 0, 0);
2380 		}
2381 		unlock_page(old_page);
2382 	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2383 					(VM_WRITE|VM_SHARED))) {
2384 		return wp_page_shared(mm, vma, address, page_table, pmd,
2385 				      ptl, orig_pte, old_page);
2386 	}
2387 
2388 	/*
2389 	 * Ok, we need to copy. Oh, well..
2390 	 */
2391 	page_cache_get(old_page);
2392 
2393 	pte_unmap_unlock(page_table, ptl);
2394 	return wp_page_copy(mm, vma, address, page_table, pmd,
2395 			    orig_pte, old_page);
2396 }
2397 
unmap_mapping_range_vma(struct vm_area_struct * vma,unsigned long start_addr,unsigned long end_addr,struct zap_details * details)2398 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2399 		unsigned long start_addr, unsigned long end_addr,
2400 		struct zap_details *details)
2401 {
2402 	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2403 }
2404 
unmap_mapping_range_tree(struct rb_root * root,struct zap_details * details)2405 static inline void unmap_mapping_range_tree(struct rb_root *root,
2406 					    struct zap_details *details)
2407 {
2408 	struct vm_area_struct *vma;
2409 	pgoff_t vba, vea, zba, zea;
2410 
2411 	vma_interval_tree_foreach(vma, root,
2412 			details->first_index, details->last_index) {
2413 
2414 		vba = vma->vm_pgoff;
2415 		vea = vba + vma_pages(vma) - 1;
2416 		/* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2417 		zba = details->first_index;
2418 		if (zba < vba)
2419 			zba = vba;
2420 		zea = details->last_index;
2421 		if (zea > vea)
2422 			zea = vea;
2423 
2424 		unmap_mapping_range_vma(vma,
2425 			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2426 			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2427 				details);
2428 	}
2429 }
2430 
2431 /**
2432  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2433  * address_space corresponding to the specified page range in the underlying
2434  * file.
2435  *
2436  * @mapping: the address space containing mmaps to be unmapped.
2437  * @holebegin: byte in first page to unmap, relative to the start of
2438  * the underlying file.  This will be rounded down to a PAGE_SIZE
2439  * boundary.  Note that this is different from truncate_pagecache(), which
2440  * must keep the partial page.  In contrast, we must get rid of
2441  * partial pages.
2442  * @holelen: size of prospective hole in bytes.  This will be rounded
2443  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2444  * end of the file.
2445  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2446  * but 0 when invalidating pagecache, don't throw away private data.
2447  */
unmap_mapping_range(struct address_space * mapping,loff_t const holebegin,loff_t const holelen,int even_cows)2448 void unmap_mapping_range(struct address_space *mapping,
2449 		loff_t const holebegin, loff_t const holelen, int even_cows)
2450 {
2451 	struct zap_details details;
2452 	pgoff_t hba = holebegin >> PAGE_SHIFT;
2453 	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2454 
2455 	/* Check for overflow. */
2456 	if (sizeof(holelen) > sizeof(hlen)) {
2457 		long long holeend =
2458 			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2459 		if (holeend & ~(long long)ULONG_MAX)
2460 			hlen = ULONG_MAX - hba + 1;
2461 	}
2462 
2463 	details.check_mapping = even_cows? NULL: mapping;
2464 	details.first_index = hba;
2465 	details.last_index = hba + hlen - 1;
2466 	if (details.last_index < details.first_index)
2467 		details.last_index = ULONG_MAX;
2468 
2469 
2470 	/* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2471 	i_mmap_lock_write(mapping);
2472 	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2473 		unmap_mapping_range_tree(&mapping->i_mmap, &details);
2474 	i_mmap_unlock_write(mapping);
2475 }
2476 EXPORT_SYMBOL(unmap_mapping_range);
2477 
2478 /*
2479  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2480  * but allow concurrent faults), and pte mapped but not yet locked.
2481  * We return with pte unmapped and unlocked.
2482  *
2483  * We return with the mmap_sem locked or unlocked in the same cases
2484  * as does filemap_fault().
2485  */
do_swap_page(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * page_table,pmd_t * pmd,unsigned int flags,pte_t orig_pte)2486 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2487 		unsigned long address, pte_t *page_table, pmd_t *pmd,
2488 		unsigned int flags, pte_t orig_pte)
2489 {
2490 	spinlock_t *ptl;
2491 	struct page *page, *swapcache;
2492 	struct mem_cgroup *memcg;
2493 	swp_entry_t entry;
2494 	pte_t pte;
2495 	int locked;
2496 	int exclusive = 0;
2497 	int ret = 0;
2498 
2499 	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2500 		goto out;
2501 
2502 	entry = pte_to_swp_entry(orig_pte);
2503 	if (unlikely(non_swap_entry(entry))) {
2504 		if (is_migration_entry(entry)) {
2505 			migration_entry_wait(mm, pmd, address);
2506 		} else if (is_hwpoison_entry(entry)) {
2507 			ret = VM_FAULT_HWPOISON;
2508 		} else {
2509 			print_bad_pte(vma, address, orig_pte, NULL);
2510 			ret = VM_FAULT_SIGBUS;
2511 		}
2512 		goto out;
2513 	}
2514 	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2515 	page = lookup_swap_cache(entry);
2516 	if (!page) {
2517 		page = swapin_readahead(entry,
2518 					GFP_HIGHUSER_MOVABLE, vma, address);
2519 		if (!page) {
2520 			/*
2521 			 * Back out if somebody else faulted in this pte
2522 			 * while we released the pte lock.
2523 			 */
2524 			page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2525 			if (likely(pte_same(*page_table, orig_pte)))
2526 				ret = VM_FAULT_OOM;
2527 			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2528 			goto unlock;
2529 		}
2530 
2531 		/* Had to read the page from swap area: Major fault */
2532 		ret = VM_FAULT_MAJOR;
2533 		count_vm_event(PGMAJFAULT);
2534 		mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2535 	} else if (PageHWPoison(page)) {
2536 		/*
2537 		 * hwpoisoned dirty swapcache pages are kept for killing
2538 		 * owner processes (which may be unknown at hwpoison time)
2539 		 */
2540 		ret = VM_FAULT_HWPOISON;
2541 		delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2542 		swapcache = page;
2543 		goto out_release;
2544 	}
2545 
2546 	swapcache = page;
2547 	locked = lock_page_or_retry(page, mm, flags);
2548 
2549 	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2550 	if (!locked) {
2551 		ret |= VM_FAULT_RETRY;
2552 		goto out_release;
2553 	}
2554 
2555 	/*
2556 	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2557 	 * release the swapcache from under us.  The page pin, and pte_same
2558 	 * test below, are not enough to exclude that.  Even if it is still
2559 	 * swapcache, we need to check that the page's swap has not changed.
2560 	 */
2561 	if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2562 		goto out_page;
2563 
2564 	page = ksm_might_need_to_copy(page, vma, address);
2565 	if (unlikely(!page)) {
2566 		ret = VM_FAULT_OOM;
2567 		page = swapcache;
2568 		goto out_page;
2569 	}
2570 
2571 	if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
2572 		ret = VM_FAULT_OOM;
2573 		goto out_page;
2574 	}
2575 
2576 	/*
2577 	 * Back out if somebody else already faulted in this pte.
2578 	 */
2579 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2580 	if (unlikely(!pte_same(*page_table, orig_pte)))
2581 		goto out_nomap;
2582 
2583 	if (unlikely(!PageUptodate(page))) {
2584 		ret = VM_FAULT_SIGBUS;
2585 		goto out_nomap;
2586 	}
2587 
2588 	/*
2589 	 * The page isn't present yet, go ahead with the fault.
2590 	 *
2591 	 * Be careful about the sequence of operations here.
2592 	 * To get its accounting right, reuse_swap_page() must be called
2593 	 * while the page is counted on swap but not yet in mapcount i.e.
2594 	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2595 	 * must be called after the swap_free(), or it will never succeed.
2596 	 */
2597 
2598 	inc_mm_counter_fast(mm, MM_ANONPAGES);
2599 	dec_mm_counter_fast(mm, MM_SWAPENTS);
2600 	pte = mk_pte(page, vma->vm_page_prot);
2601 	if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2602 		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2603 		flags &= ~FAULT_FLAG_WRITE;
2604 		ret |= VM_FAULT_WRITE;
2605 		exclusive = 1;
2606 	}
2607 	flush_icache_page(vma, page);
2608 	if (pte_swp_soft_dirty(orig_pte))
2609 		pte = pte_mksoft_dirty(pte);
2610 	set_pte_at(mm, address, page_table, pte);
2611 	if (page == swapcache) {
2612 		do_page_add_anon_rmap(page, vma, address, exclusive);
2613 		mem_cgroup_commit_charge(page, memcg, true);
2614 	} else { /* ksm created a completely new copy */
2615 		page_add_new_anon_rmap(page, vma, address);
2616 		mem_cgroup_commit_charge(page, memcg, false);
2617 		lru_cache_add_active_or_unevictable(page, vma);
2618 	}
2619 
2620 	swap_free(entry);
2621 	if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2622 		try_to_free_swap(page);
2623 	unlock_page(page);
2624 	if (page != swapcache) {
2625 		/*
2626 		 * Hold the lock to avoid the swap entry to be reused
2627 		 * until we take the PT lock for the pte_same() check
2628 		 * (to avoid false positives from pte_same). For
2629 		 * further safety release the lock after the swap_free
2630 		 * so that the swap count won't change under a
2631 		 * parallel locked swapcache.
2632 		 */
2633 		unlock_page(swapcache);
2634 		page_cache_release(swapcache);
2635 	}
2636 
2637 	if (flags & FAULT_FLAG_WRITE) {
2638 		ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2639 		if (ret & VM_FAULT_ERROR)
2640 			ret &= VM_FAULT_ERROR;
2641 		goto out;
2642 	}
2643 
2644 	/* No need to invalidate - it was non-present before */
2645 	update_mmu_cache(vma, address, page_table);
2646 unlock:
2647 	pte_unmap_unlock(page_table, ptl);
2648 out:
2649 	return ret;
2650 out_nomap:
2651 	mem_cgroup_cancel_charge(page, memcg);
2652 	pte_unmap_unlock(page_table, ptl);
2653 out_page:
2654 	unlock_page(page);
2655 out_release:
2656 	page_cache_release(page);
2657 	if (page != swapcache) {
2658 		unlock_page(swapcache);
2659 		page_cache_release(swapcache);
2660 	}
2661 	return ret;
2662 }
2663 
2664 /*
2665  * This is like a special single-page "expand_{down|up}wards()",
2666  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2667  * doesn't hit another vma.
2668  */
check_stack_guard_page(struct vm_area_struct * vma,unsigned long address)2669 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2670 {
2671 	address &= PAGE_MASK;
2672 	if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2673 		struct vm_area_struct *prev = vma->vm_prev;
2674 
2675 		/*
2676 		 * Is there a mapping abutting this one below?
2677 		 *
2678 		 * That's only ok if it's the same stack mapping
2679 		 * that has gotten split..
2680 		 */
2681 		if (prev && prev->vm_end == address)
2682 			return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2683 
2684 		return expand_downwards(vma, address - PAGE_SIZE);
2685 	}
2686 	if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2687 		struct vm_area_struct *next = vma->vm_next;
2688 
2689 		/* As VM_GROWSDOWN but s/below/above/ */
2690 		if (next && next->vm_start == address + PAGE_SIZE)
2691 			return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2692 
2693 		return expand_upwards(vma, address + PAGE_SIZE);
2694 	}
2695 	return 0;
2696 }
2697 
2698 /*
2699  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2700  * but allow concurrent faults), and pte mapped but not yet locked.
2701  * We return with mmap_sem still held, but pte unmapped and unlocked.
2702  */
do_anonymous_page(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * page_table,pmd_t * pmd,unsigned int flags)2703 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2704 		unsigned long address, pte_t *page_table, pmd_t *pmd,
2705 		unsigned int flags)
2706 {
2707 	struct mem_cgroup *memcg;
2708 	struct page *page;
2709 	spinlock_t *ptl;
2710 	pte_t entry;
2711 
2712 	pte_unmap(page_table);
2713 
2714 	/* File mapping without ->vm_ops ? */
2715 	if (vma->vm_flags & VM_SHARED)
2716 		return VM_FAULT_SIGBUS;
2717 
2718 	/* Check if we need to add a guard page to the stack */
2719 	if (check_stack_guard_page(vma, address) < 0)
2720 		return VM_FAULT_SIGSEGV;
2721 
2722 	/* Use the zero-page for reads */
2723 	if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2724 		entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2725 						vma->vm_page_prot));
2726 		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2727 		if (!pte_none(*page_table))
2728 			goto unlock;
2729 		/* Deliver the page fault to userland, check inside PT lock */
2730 		if (userfaultfd_missing(vma)) {
2731 			pte_unmap_unlock(page_table, ptl);
2732 			return handle_userfault(vma, address, flags,
2733 						VM_UFFD_MISSING);
2734 		}
2735 		goto setpte;
2736 	}
2737 
2738 	/* Allocate our own private page. */
2739 	if (unlikely(anon_vma_prepare(vma)))
2740 		goto oom;
2741 	page = alloc_zeroed_user_highpage_movable(vma, address);
2742 	if (!page)
2743 		goto oom;
2744 
2745 	if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
2746 		goto oom_free_page;
2747 
2748 	/*
2749 	 * The memory barrier inside __SetPageUptodate makes sure that
2750 	 * preceeding stores to the page contents become visible before
2751 	 * the set_pte_at() write.
2752 	 */
2753 	__SetPageUptodate(page);
2754 
2755 	entry = mk_pte(page, vma->vm_page_prot);
2756 	if (vma->vm_flags & VM_WRITE)
2757 		entry = pte_mkwrite(pte_mkdirty(entry));
2758 
2759 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2760 	if (!pte_none(*page_table))
2761 		goto release;
2762 
2763 	/* Deliver the page fault to userland, check inside PT lock */
2764 	if (userfaultfd_missing(vma)) {
2765 		pte_unmap_unlock(page_table, ptl);
2766 		mem_cgroup_cancel_charge(page, memcg);
2767 		page_cache_release(page);
2768 		return handle_userfault(vma, address, flags,
2769 					VM_UFFD_MISSING);
2770 	}
2771 
2772 	inc_mm_counter_fast(mm, MM_ANONPAGES);
2773 	page_add_new_anon_rmap(page, vma, address);
2774 	mem_cgroup_commit_charge(page, memcg, false);
2775 	lru_cache_add_active_or_unevictable(page, vma);
2776 setpte:
2777 	set_pte_at(mm, address, page_table, entry);
2778 
2779 	/* No need to invalidate - it was non-present before */
2780 	update_mmu_cache(vma, address, page_table);
2781 unlock:
2782 	pte_unmap_unlock(page_table, ptl);
2783 	return 0;
2784 release:
2785 	mem_cgroup_cancel_charge(page, memcg);
2786 	page_cache_release(page);
2787 	goto unlock;
2788 oom_free_page:
2789 	page_cache_release(page);
2790 oom:
2791 	return VM_FAULT_OOM;
2792 }
2793 
2794 /*
2795  * The mmap_sem must have been held on entry, and may have been
2796  * released depending on flags and vma->vm_ops->fault() return value.
2797  * See filemap_fault() and __lock_page_retry().
2798  */
__do_fault(struct vm_area_struct * vma,unsigned long address,pgoff_t pgoff,unsigned int flags,struct page * cow_page,struct page ** page)2799 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2800 			pgoff_t pgoff, unsigned int flags,
2801 			struct page *cow_page, struct page **page)
2802 {
2803 	struct vm_fault vmf;
2804 	int ret;
2805 
2806 	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2807 	vmf.pgoff = pgoff;
2808 	vmf.flags = flags;
2809 	vmf.page = NULL;
2810 	vmf.cow_page = cow_page;
2811 
2812 	ret = vma->vm_ops->fault(vma, &vmf);
2813 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2814 		return ret;
2815 	if (!vmf.page)
2816 		goto out;
2817 
2818 	if (unlikely(PageHWPoison(vmf.page))) {
2819 		if (ret & VM_FAULT_LOCKED)
2820 			unlock_page(vmf.page);
2821 		page_cache_release(vmf.page);
2822 		return VM_FAULT_HWPOISON;
2823 	}
2824 
2825 	if (unlikely(!(ret & VM_FAULT_LOCKED)))
2826 		lock_page(vmf.page);
2827 	else
2828 		VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2829 
2830  out:
2831 	*page = vmf.page;
2832 	return ret;
2833 }
2834 
2835 /**
2836  * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2837  *
2838  * @vma: virtual memory area
2839  * @address: user virtual address
2840  * @page: page to map
2841  * @pte: pointer to target page table entry
2842  * @write: true, if new entry is writable
2843  * @anon: true, if it's anonymous page
2844  *
2845  * Caller must hold page table lock relevant for @pte.
2846  *
2847  * Target users are page handler itself and implementations of
2848  * vm_ops->map_pages.
2849  */
do_set_pte(struct vm_area_struct * vma,unsigned long address,struct page * page,pte_t * pte,bool write,bool anon)2850 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2851 		struct page *page, pte_t *pte, bool write, bool anon)
2852 {
2853 	pte_t entry;
2854 
2855 	flush_icache_page(vma, page);
2856 	entry = mk_pte(page, vma->vm_page_prot);
2857 	if (write)
2858 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2859 	if (anon) {
2860 		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2861 		page_add_new_anon_rmap(page, vma, address);
2862 	} else {
2863 		inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2864 		page_add_file_rmap(page);
2865 	}
2866 	set_pte_at(vma->vm_mm, address, pte, entry);
2867 
2868 	/* no need to invalidate: a not-present page won't be cached */
2869 	update_mmu_cache(vma, address, pte);
2870 }
2871 
2872 static unsigned long fault_around_bytes __read_mostly =
2873 	rounddown_pow_of_two(65536);
2874 
2875 #ifdef CONFIG_DEBUG_FS
fault_around_bytes_get(void * data,u64 * val)2876 static int fault_around_bytes_get(void *data, u64 *val)
2877 {
2878 	*val = fault_around_bytes;
2879 	return 0;
2880 }
2881 
2882 /*
2883  * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2884  * rounded down to nearest page order. It's what do_fault_around() expects to
2885  * see.
2886  */
fault_around_bytes_set(void * data,u64 val)2887 static int fault_around_bytes_set(void *data, u64 val)
2888 {
2889 	if (val / PAGE_SIZE > PTRS_PER_PTE)
2890 		return -EINVAL;
2891 	if (val > PAGE_SIZE)
2892 		fault_around_bytes = rounddown_pow_of_two(val);
2893 	else
2894 		fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2895 	return 0;
2896 }
2897 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2898 		fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2899 
fault_around_debugfs(void)2900 static int __init fault_around_debugfs(void)
2901 {
2902 	void *ret;
2903 
2904 	ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2905 			&fault_around_bytes_fops);
2906 	if (!ret)
2907 		pr_warn("Failed to create fault_around_bytes in debugfs");
2908 	return 0;
2909 }
2910 late_initcall(fault_around_debugfs);
2911 #endif
2912 
2913 /*
2914  * do_fault_around() tries to map few pages around the fault address. The hope
2915  * is that the pages will be needed soon and this will lower the number of
2916  * faults to handle.
2917  *
2918  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2919  * not ready to be mapped: not up-to-date, locked, etc.
2920  *
2921  * This function is called with the page table lock taken. In the split ptlock
2922  * case the page table lock only protects only those entries which belong to
2923  * the page table corresponding to the fault address.
2924  *
2925  * This function doesn't cross the VMA boundaries, in order to call map_pages()
2926  * only once.
2927  *
2928  * fault_around_pages() defines how many pages we'll try to map.
2929  * do_fault_around() expects it to return a power of two less than or equal to
2930  * PTRS_PER_PTE.
2931  *
2932  * The virtual address of the area that we map is naturally aligned to the
2933  * fault_around_pages() value (and therefore to page order).  This way it's
2934  * easier to guarantee that we don't cross page table boundaries.
2935  */
do_fault_around(struct vm_area_struct * vma,unsigned long address,pte_t * pte,pgoff_t pgoff,unsigned int flags)2936 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2937 		pte_t *pte, pgoff_t pgoff, unsigned int flags)
2938 {
2939 	unsigned long start_addr, nr_pages, mask;
2940 	pgoff_t max_pgoff;
2941 	struct vm_fault vmf;
2942 	int off;
2943 
2944 	nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2945 	mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2946 
2947 	start_addr = max(address & mask, vma->vm_start);
2948 	off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2949 	pte -= off;
2950 	pgoff -= off;
2951 
2952 	/*
2953 	 *  max_pgoff is either end of page table or end of vma
2954 	 *  or fault_around_pages() from pgoff, depending what is nearest.
2955 	 */
2956 	max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2957 		PTRS_PER_PTE - 1;
2958 	max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2959 			pgoff + nr_pages - 1);
2960 
2961 	/* Check if it makes any sense to call ->map_pages */
2962 	while (!pte_none(*pte)) {
2963 		if (++pgoff > max_pgoff)
2964 			return;
2965 		start_addr += PAGE_SIZE;
2966 		if (start_addr >= vma->vm_end)
2967 			return;
2968 		pte++;
2969 	}
2970 
2971 	vmf.virtual_address = (void __user *) start_addr;
2972 	vmf.pte = pte;
2973 	vmf.pgoff = pgoff;
2974 	vmf.max_pgoff = max_pgoff;
2975 	vmf.flags = flags;
2976 	vma->vm_ops->map_pages(vma, &vmf);
2977 }
2978 
do_read_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,pgoff_t pgoff,unsigned int flags,pte_t orig_pte)2979 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2980 		unsigned long address, pmd_t *pmd,
2981 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2982 {
2983 	struct page *fault_page;
2984 	spinlock_t *ptl;
2985 	pte_t *pte;
2986 	int ret = 0;
2987 
2988 	/*
2989 	 * Let's call ->map_pages() first and use ->fault() as fallback
2990 	 * if page by the offset is not ready to be mapped (cold cache or
2991 	 * something).
2992 	 */
2993 	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
2994 		pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2995 		do_fault_around(vma, address, pte, pgoff, flags);
2996 		if (!pte_same(*pte, orig_pte))
2997 			goto unlock_out;
2998 		pte_unmap_unlock(pte, ptl);
2999 	}
3000 
3001 	ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3002 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3003 		return ret;
3004 
3005 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3006 	if (unlikely(!pte_same(*pte, orig_pte))) {
3007 		pte_unmap_unlock(pte, ptl);
3008 		unlock_page(fault_page);
3009 		page_cache_release(fault_page);
3010 		return ret;
3011 	}
3012 	do_set_pte(vma, address, fault_page, pte, false, false);
3013 	unlock_page(fault_page);
3014 unlock_out:
3015 	pte_unmap_unlock(pte, ptl);
3016 	return ret;
3017 }
3018 
do_cow_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,pgoff_t pgoff,unsigned int flags,pte_t orig_pte)3019 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3020 		unsigned long address, pmd_t *pmd,
3021 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3022 {
3023 	struct page *fault_page, *new_page;
3024 	struct mem_cgroup *memcg;
3025 	spinlock_t *ptl;
3026 	pte_t *pte;
3027 	int ret;
3028 
3029 	if (unlikely(anon_vma_prepare(vma)))
3030 		return VM_FAULT_OOM;
3031 
3032 	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3033 	if (!new_page)
3034 		return VM_FAULT_OOM;
3035 
3036 	if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
3037 		page_cache_release(new_page);
3038 		return VM_FAULT_OOM;
3039 	}
3040 
3041 	ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
3042 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3043 		goto uncharge_out;
3044 
3045 	if (fault_page)
3046 		copy_user_highpage(new_page, fault_page, address, vma);
3047 	__SetPageUptodate(new_page);
3048 
3049 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3050 	if (unlikely(!pte_same(*pte, orig_pte))) {
3051 		pte_unmap_unlock(pte, ptl);
3052 		if (fault_page) {
3053 			unlock_page(fault_page);
3054 			page_cache_release(fault_page);
3055 		} else {
3056 			/*
3057 			 * The fault handler has no page to lock, so it holds
3058 			 * i_mmap_lock for read to protect against truncate.
3059 			 */
3060 			i_mmap_unlock_read(vma->vm_file->f_mapping);
3061 		}
3062 		goto uncharge_out;
3063 	}
3064 	do_set_pte(vma, address, new_page, pte, true, true);
3065 	mem_cgroup_commit_charge(new_page, memcg, false);
3066 	lru_cache_add_active_or_unevictable(new_page, vma);
3067 	pte_unmap_unlock(pte, ptl);
3068 	if (fault_page) {
3069 		unlock_page(fault_page);
3070 		page_cache_release(fault_page);
3071 	} else {
3072 		/*
3073 		 * The fault handler has no page to lock, so it holds
3074 		 * i_mmap_lock for read to protect against truncate.
3075 		 */
3076 		i_mmap_unlock_read(vma->vm_file->f_mapping);
3077 	}
3078 	return ret;
3079 uncharge_out:
3080 	mem_cgroup_cancel_charge(new_page, memcg);
3081 	page_cache_release(new_page);
3082 	return ret;
3083 }
3084 
do_shared_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,pgoff_t pgoff,unsigned int flags,pte_t orig_pte)3085 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3086 		unsigned long address, pmd_t *pmd,
3087 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3088 {
3089 	struct page *fault_page;
3090 	struct address_space *mapping;
3091 	spinlock_t *ptl;
3092 	pte_t *pte;
3093 	int dirtied = 0;
3094 	int ret, tmp;
3095 
3096 	ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3097 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3098 		return ret;
3099 
3100 	/*
3101 	 * Check if the backing address space wants to know that the page is
3102 	 * about to become writable
3103 	 */
3104 	if (vma->vm_ops->page_mkwrite) {
3105 		unlock_page(fault_page);
3106 		tmp = do_page_mkwrite(vma, fault_page, address);
3107 		if (unlikely(!tmp ||
3108 				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3109 			page_cache_release(fault_page);
3110 			return tmp;
3111 		}
3112 	}
3113 
3114 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3115 	if (unlikely(!pte_same(*pte, orig_pte))) {
3116 		pte_unmap_unlock(pte, ptl);
3117 		unlock_page(fault_page);
3118 		page_cache_release(fault_page);
3119 		return ret;
3120 	}
3121 	do_set_pte(vma, address, fault_page, pte, true, false);
3122 	pte_unmap_unlock(pte, ptl);
3123 
3124 	if (set_page_dirty(fault_page))
3125 		dirtied = 1;
3126 	/*
3127 	 * Take a local copy of the address_space - page.mapping may be zeroed
3128 	 * by truncate after unlock_page().   The address_space itself remains
3129 	 * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
3130 	 * release semantics to prevent the compiler from undoing this copying.
3131 	 */
3132 	mapping = fault_page->mapping;
3133 	unlock_page(fault_page);
3134 	if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3135 		/*
3136 		 * Some device drivers do not set page.mapping but still
3137 		 * dirty their pages
3138 		 */
3139 		balance_dirty_pages_ratelimited(mapping);
3140 	}
3141 
3142 	if (!vma->vm_ops->page_mkwrite)
3143 		file_update_time(vma->vm_file);
3144 
3145 	return ret;
3146 }
3147 
3148 /*
3149  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3150  * but allow concurrent faults).
3151  * The mmap_sem may have been released depending on flags and our
3152  * return value.  See filemap_fault() and __lock_page_or_retry().
3153  */
do_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * page_table,pmd_t * pmd,unsigned int flags,pte_t orig_pte)3154 static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3155 		unsigned long address, pte_t *page_table, pmd_t *pmd,
3156 		unsigned int flags, pte_t orig_pte)
3157 {
3158 	pgoff_t pgoff = (((address & PAGE_MASK)
3159 			- vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3160 
3161 	pte_unmap(page_table);
3162 	/* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3163 	if (!vma->vm_ops->fault)
3164 		return VM_FAULT_SIGBUS;
3165 	if (!(flags & FAULT_FLAG_WRITE))
3166 		return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3167 				orig_pte);
3168 	if (!(vma->vm_flags & VM_SHARED))
3169 		return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3170 				orig_pte);
3171 	return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3172 }
3173 
numa_migrate_prep(struct page * page,struct vm_area_struct * vma,unsigned long addr,int page_nid,int * flags)3174 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3175 				unsigned long addr, int page_nid,
3176 				int *flags)
3177 {
3178 	get_page(page);
3179 
3180 	count_vm_numa_event(NUMA_HINT_FAULTS);
3181 	if (page_nid == numa_node_id()) {
3182 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3183 		*flags |= TNF_FAULT_LOCAL;
3184 	}
3185 
3186 	return mpol_misplaced(page, vma, addr);
3187 }
3188 
do_numa_page(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pte_t pte,pte_t * ptep,pmd_t * pmd)3189 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3190 		   unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3191 {
3192 	struct page *page = NULL;
3193 	spinlock_t *ptl;
3194 	int page_nid = -1;
3195 	int last_cpupid;
3196 	int target_nid;
3197 	bool migrated = false;
3198 	bool was_writable = pte_write(pte);
3199 	int flags = 0;
3200 
3201 	/* A PROT_NONE fault should not end up here */
3202 	BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3203 
3204 	/*
3205 	* The "pte" at this point cannot be used safely without
3206 	* validation through pte_unmap_same(). It's of NUMA type but
3207 	* the pfn may be screwed if the read is non atomic.
3208 	*
3209 	* We can safely just do a "set_pte_at()", because the old
3210 	* page table entry is not accessible, so there would be no
3211 	* concurrent hardware modifications to the PTE.
3212 	*/
3213 	ptl = pte_lockptr(mm, pmd);
3214 	spin_lock(ptl);
3215 	if (unlikely(!pte_same(*ptep, pte))) {
3216 		pte_unmap_unlock(ptep, ptl);
3217 		goto out;
3218 	}
3219 
3220 	/* Make it present again */
3221 	pte = pte_modify(pte, vma->vm_page_prot);
3222 	pte = pte_mkyoung(pte);
3223 	if (was_writable)
3224 		pte = pte_mkwrite(pte);
3225 	set_pte_at(mm, addr, ptep, pte);
3226 	update_mmu_cache(vma, addr, ptep);
3227 
3228 	page = vm_normal_page(vma, addr, pte);
3229 	if (!page) {
3230 		pte_unmap_unlock(ptep, ptl);
3231 		return 0;
3232 	}
3233 
3234 	/*
3235 	 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3236 	 * much anyway since they can be in shared cache state. This misses
3237 	 * the case where a mapping is writable but the process never writes
3238 	 * to it but pte_write gets cleared during protection updates and
3239 	 * pte_dirty has unpredictable behaviour between PTE scan updates,
3240 	 * background writeback, dirty balancing and application behaviour.
3241 	 */
3242 	if (!(vma->vm_flags & VM_WRITE))
3243 		flags |= TNF_NO_GROUP;
3244 
3245 	/*
3246 	 * Flag if the page is shared between multiple address spaces. This
3247 	 * is later used when determining whether to group tasks together
3248 	 */
3249 	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3250 		flags |= TNF_SHARED;
3251 
3252 	last_cpupid = page_cpupid_last(page);
3253 	page_nid = page_to_nid(page);
3254 	target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3255 	pte_unmap_unlock(ptep, ptl);
3256 	if (target_nid == -1) {
3257 		put_page(page);
3258 		goto out;
3259 	}
3260 
3261 	/* Migrate to the requested node */
3262 	migrated = migrate_misplaced_page(page, vma, target_nid);
3263 	if (migrated) {
3264 		page_nid = target_nid;
3265 		flags |= TNF_MIGRATED;
3266 	} else
3267 		flags |= TNF_MIGRATE_FAIL;
3268 
3269 out:
3270 	if (page_nid != -1)
3271 		task_numa_fault(last_cpupid, page_nid, 1, flags);
3272 	return 0;
3273 }
3274 
create_huge_pmd(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,unsigned int flags)3275 static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3276 			unsigned long address, pmd_t *pmd, unsigned int flags)
3277 {
3278 	if (vma_is_anonymous(vma))
3279 		return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3280 	if (vma->vm_ops->pmd_fault)
3281 		return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3282 	return VM_FAULT_FALLBACK;
3283 }
3284 
wp_huge_pmd(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,pmd_t orig_pmd,unsigned int flags)3285 static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3286 			unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3287 			unsigned int flags)
3288 {
3289 	if (vma_is_anonymous(vma))
3290 		return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3291 	if (vma->vm_ops->pmd_fault)
3292 		return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3293 	return VM_FAULT_FALLBACK;
3294 }
3295 
3296 /*
3297  * These routines also need to handle stuff like marking pages dirty
3298  * and/or accessed for architectures that don't do it in hardware (most
3299  * RISC architectures).  The early dirtying is also good on the i386.
3300  *
3301  * There is also a hook called "update_mmu_cache()" that architectures
3302  * with external mmu caches can use to update those (ie the Sparc or
3303  * PowerPC hashed page tables that act as extended TLBs).
3304  *
3305  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3306  * but allow concurrent faults), and pte mapped but not yet locked.
3307  * We return with pte unmapped and unlocked.
3308  *
3309  * The mmap_sem may have been released depending on flags and our
3310  * return value.  See filemap_fault() and __lock_page_or_retry().
3311  */
handle_pte_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * pte,pmd_t * pmd,unsigned int flags)3312 static int handle_pte_fault(struct mm_struct *mm,
3313 		     struct vm_area_struct *vma, unsigned long address,
3314 		     pte_t *pte, pmd_t *pmd, unsigned int flags)
3315 {
3316 	pte_t entry;
3317 	spinlock_t *ptl;
3318 
3319 	/*
3320 	 * some architectures can have larger ptes than wordsize,
3321 	 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3322 	 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3323 	 * The code below just needs a consistent view for the ifs and
3324 	 * we later double check anyway with the ptl lock held. So here
3325 	 * a barrier will do.
3326 	 */
3327 	entry = *pte;
3328 	barrier();
3329 	if (!pte_present(entry)) {
3330 		if (pte_none(entry)) {
3331 			if (vma_is_anonymous(vma))
3332 				return do_anonymous_page(mm, vma, address,
3333 							 pte, pmd, flags);
3334 			else
3335 				return do_fault(mm, vma, address, pte, pmd,
3336 						flags, entry);
3337 		}
3338 		return do_swap_page(mm, vma, address,
3339 					pte, pmd, flags, entry);
3340 	}
3341 
3342 	if (pte_protnone(entry))
3343 		return do_numa_page(mm, vma, address, entry, pte, pmd);
3344 
3345 	ptl = pte_lockptr(mm, pmd);
3346 	spin_lock(ptl);
3347 	if (unlikely(!pte_same(*pte, entry)))
3348 		goto unlock;
3349 	if (flags & FAULT_FLAG_WRITE) {
3350 		if (!pte_write(entry))
3351 			return do_wp_page(mm, vma, address,
3352 					pte, pmd, ptl, entry);
3353 		entry = pte_mkdirty(entry);
3354 	}
3355 	entry = pte_mkyoung(entry);
3356 	if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3357 		update_mmu_cache(vma, address, pte);
3358 	} else {
3359 		/*
3360 		 * This is needed only for protection faults but the arch code
3361 		 * is not yet telling us if this is a protection fault or not.
3362 		 * This still avoids useless tlb flushes for .text page faults
3363 		 * with threads.
3364 		 */
3365 		if (flags & FAULT_FLAG_WRITE)
3366 			flush_tlb_fix_spurious_fault(vma, address);
3367 	}
3368 unlock:
3369 	pte_unmap_unlock(pte, ptl);
3370 	return 0;
3371 }
3372 
3373 /*
3374  * By the time we get here, we already hold the mm semaphore
3375  *
3376  * The mmap_sem may have been released depending on flags and our
3377  * return value.  See filemap_fault() and __lock_page_or_retry().
3378  */
__handle_mm_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,unsigned int flags)3379 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3380 			     unsigned long address, unsigned int flags)
3381 {
3382 	pgd_t *pgd;
3383 	pud_t *pud;
3384 	pmd_t *pmd;
3385 	pte_t *pte;
3386 
3387 	if (unlikely(is_vm_hugetlb_page(vma)))
3388 		return hugetlb_fault(mm, vma, address, flags);
3389 
3390 	pgd = pgd_offset(mm, address);
3391 	pud = pud_alloc(mm, pgd, address);
3392 	if (!pud)
3393 		return VM_FAULT_OOM;
3394 	pmd = pmd_alloc(mm, pud, address);
3395 	if (!pmd)
3396 		return VM_FAULT_OOM;
3397 	if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3398 		int ret = create_huge_pmd(mm, vma, address, pmd, flags);
3399 		if (!(ret & VM_FAULT_FALLBACK))
3400 			return ret;
3401 	} else {
3402 		pmd_t orig_pmd = *pmd;
3403 		int ret;
3404 
3405 		barrier();
3406 		if (pmd_trans_huge(orig_pmd)) {
3407 			unsigned int dirty = flags & FAULT_FLAG_WRITE;
3408 
3409 			/*
3410 			 * If the pmd is splitting, return and retry the
3411 			 * the fault.  Alternative: wait until the split
3412 			 * is done, and goto retry.
3413 			 */
3414 			if (pmd_trans_splitting(orig_pmd))
3415 				return 0;
3416 
3417 			if (pmd_protnone(orig_pmd))
3418 				return do_huge_pmd_numa_page(mm, vma, address,
3419 							     orig_pmd, pmd);
3420 
3421 			if (dirty && !pmd_write(orig_pmd)) {
3422 				ret = wp_huge_pmd(mm, vma, address, pmd,
3423 							orig_pmd, flags);
3424 				if (!(ret & VM_FAULT_FALLBACK))
3425 					return ret;
3426 			} else {
3427 				huge_pmd_set_accessed(mm, vma, address, pmd,
3428 						      orig_pmd, dirty);
3429 				return 0;
3430 			}
3431 		}
3432 	}
3433 
3434 	/*
3435 	 * Use __pte_alloc instead of pte_alloc_map, because we can't
3436 	 * run pte_offset_map on the pmd, if an huge pmd could
3437 	 * materialize from under us from a different thread.
3438 	 */
3439 	if (unlikely(pmd_none(*pmd)) &&
3440 	    unlikely(__pte_alloc(mm, vma, pmd, address)))
3441 		return VM_FAULT_OOM;
3442 	/*
3443 	 * If a huge pmd materialized under us just retry later.  Use
3444 	 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3445 	 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3446 	 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3447 	 * in a different thread of this mm, in turn leading to a misleading
3448 	 * pmd_trans_huge() retval.  All we have to ensure is that it is a
3449 	 * regular pmd that we can walk with pte_offset_map() and we can do that
3450 	 * through an atomic read in C, which is what pmd_trans_unstable()
3451 	 * provides.
3452 	 */
3453 	if (unlikely(pmd_trans_unstable(pmd)))
3454 		return 0;
3455 	/*
3456 	 * A regular pmd is established and it can't morph into a huge pmd
3457 	 * from under us anymore at this point because we hold the mmap_sem
3458 	 * read mode and khugepaged takes it in write mode. So now it's
3459 	 * safe to run pte_offset_map().
3460 	 */
3461 	pte = pte_offset_map(pmd, address);
3462 
3463 	return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3464 }
3465 
3466 /*
3467  * By the time we get here, we already hold the mm semaphore
3468  *
3469  * The mmap_sem may have been released depending on flags and our
3470  * return value.  See filemap_fault() and __lock_page_or_retry().
3471  */
handle_mm_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,unsigned int flags)3472 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3473 		    unsigned long address, unsigned int flags)
3474 {
3475 	int ret;
3476 
3477 	__set_current_state(TASK_RUNNING);
3478 
3479 	count_vm_event(PGFAULT);
3480 	mem_cgroup_count_vm_event(mm, PGFAULT);
3481 
3482 	/* do counter updates before entering really critical section. */
3483 	check_sync_rss_stat(current);
3484 
3485 	/*
3486 	 * Enable the memcg OOM handling for faults triggered in user
3487 	 * space.  Kernel faults are handled more gracefully.
3488 	 */
3489 	if (flags & FAULT_FLAG_USER)
3490 		mem_cgroup_oom_enable();
3491 
3492 	ret = __handle_mm_fault(mm, vma, address, flags);
3493 
3494 	if (flags & FAULT_FLAG_USER) {
3495 		mem_cgroup_oom_disable();
3496                 /*
3497                  * The task may have entered a memcg OOM situation but
3498                  * if the allocation error was handled gracefully (no
3499                  * VM_FAULT_OOM), there is no need to kill anything.
3500                  * Just clean up the OOM state peacefully.
3501                  */
3502                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3503                         mem_cgroup_oom_synchronize(false);
3504 	}
3505 
3506 	return ret;
3507 }
3508 EXPORT_SYMBOL_GPL(handle_mm_fault);
3509 
3510 #ifndef __PAGETABLE_PUD_FOLDED
3511 /*
3512  * Allocate page upper directory.
3513  * We've already handled the fast-path in-line.
3514  */
__pud_alloc(struct mm_struct * mm,pgd_t * pgd,unsigned long address)3515 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3516 {
3517 	pud_t *new = pud_alloc_one(mm, address);
3518 	if (!new)
3519 		return -ENOMEM;
3520 
3521 	smp_wmb(); /* See comment in __pte_alloc */
3522 
3523 	spin_lock(&mm->page_table_lock);
3524 	if (pgd_present(*pgd))		/* Another has populated it */
3525 		pud_free(mm, new);
3526 	else
3527 		pgd_populate(mm, pgd, new);
3528 	spin_unlock(&mm->page_table_lock);
3529 	return 0;
3530 }
3531 #endif /* __PAGETABLE_PUD_FOLDED */
3532 
3533 #ifndef __PAGETABLE_PMD_FOLDED
3534 /*
3535  * Allocate page middle directory.
3536  * We've already handled the fast-path in-line.
3537  */
__pmd_alloc(struct mm_struct * mm,pud_t * pud,unsigned long address)3538 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3539 {
3540 	pmd_t *new = pmd_alloc_one(mm, address);
3541 	if (!new)
3542 		return -ENOMEM;
3543 
3544 	smp_wmb(); /* See comment in __pte_alloc */
3545 
3546 	spin_lock(&mm->page_table_lock);
3547 #ifndef __ARCH_HAS_4LEVEL_HACK
3548 	if (!pud_present(*pud)) {
3549 		mm_inc_nr_pmds(mm);
3550 		pud_populate(mm, pud, new);
3551 	} else	/* Another has populated it */
3552 		pmd_free(mm, new);
3553 #else
3554 	if (!pgd_present(*pud)) {
3555 		mm_inc_nr_pmds(mm);
3556 		pgd_populate(mm, pud, new);
3557 	} else /* Another has populated it */
3558 		pmd_free(mm, new);
3559 #endif /* __ARCH_HAS_4LEVEL_HACK */
3560 	spin_unlock(&mm->page_table_lock);
3561 	return 0;
3562 }
3563 #endif /* __PAGETABLE_PMD_FOLDED */
3564 
__follow_pte(struct mm_struct * mm,unsigned long address,pte_t ** ptepp,spinlock_t ** ptlp)3565 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3566 		pte_t **ptepp, spinlock_t **ptlp)
3567 {
3568 	pgd_t *pgd;
3569 	pud_t *pud;
3570 	pmd_t *pmd;
3571 	pte_t *ptep;
3572 
3573 	pgd = pgd_offset(mm, address);
3574 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3575 		goto out;
3576 
3577 	pud = pud_offset(pgd, address);
3578 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3579 		goto out;
3580 
3581 	pmd = pmd_offset(pud, address);
3582 	VM_BUG_ON(pmd_trans_huge(*pmd));
3583 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3584 		goto out;
3585 
3586 	/* We cannot handle huge page PFN maps. Luckily they don't exist. */
3587 	if (pmd_huge(*pmd))
3588 		goto out;
3589 
3590 	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3591 	if (!ptep)
3592 		goto out;
3593 	if (!pte_present(*ptep))
3594 		goto unlock;
3595 	*ptepp = ptep;
3596 	return 0;
3597 unlock:
3598 	pte_unmap_unlock(ptep, *ptlp);
3599 out:
3600 	return -EINVAL;
3601 }
3602 
follow_pte(struct mm_struct * mm,unsigned long address,pte_t ** ptepp,spinlock_t ** ptlp)3603 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3604 			     pte_t **ptepp, spinlock_t **ptlp)
3605 {
3606 	int res;
3607 
3608 	/* (void) is needed to make gcc happy */
3609 	(void) __cond_lock(*ptlp,
3610 			   !(res = __follow_pte(mm, address, ptepp, ptlp)));
3611 	return res;
3612 }
3613 
3614 /**
3615  * follow_pfn - look up PFN at a user virtual address
3616  * @vma: memory mapping
3617  * @address: user virtual address
3618  * @pfn: location to store found PFN
3619  *
3620  * Only IO mappings and raw PFN mappings are allowed.
3621  *
3622  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3623  */
follow_pfn(struct vm_area_struct * vma,unsigned long address,unsigned long * pfn)3624 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3625 	unsigned long *pfn)
3626 {
3627 	int ret = -EINVAL;
3628 	spinlock_t *ptl;
3629 	pte_t *ptep;
3630 
3631 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3632 		return ret;
3633 
3634 	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3635 	if (ret)
3636 		return ret;
3637 	*pfn = pte_pfn(*ptep);
3638 	pte_unmap_unlock(ptep, ptl);
3639 	return 0;
3640 }
3641 EXPORT_SYMBOL(follow_pfn);
3642 
3643 #ifdef CONFIG_HAVE_IOREMAP_PROT
follow_phys(struct vm_area_struct * vma,unsigned long address,unsigned int flags,unsigned long * prot,resource_size_t * phys)3644 int follow_phys(struct vm_area_struct *vma,
3645 		unsigned long address, unsigned int flags,
3646 		unsigned long *prot, resource_size_t *phys)
3647 {
3648 	int ret = -EINVAL;
3649 	pte_t *ptep, pte;
3650 	spinlock_t *ptl;
3651 
3652 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3653 		goto out;
3654 
3655 	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3656 		goto out;
3657 	pte = *ptep;
3658 
3659 	if ((flags & FOLL_WRITE) && !pte_write(pte))
3660 		goto unlock;
3661 
3662 	*prot = pgprot_val(pte_pgprot(pte));
3663 	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3664 
3665 	ret = 0;
3666 unlock:
3667 	pte_unmap_unlock(ptep, ptl);
3668 out:
3669 	return ret;
3670 }
3671 
generic_access_phys(struct vm_area_struct * vma,unsigned long addr,void * buf,int len,int write)3672 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3673 			void *buf, int len, int write)
3674 {
3675 	resource_size_t phys_addr;
3676 	unsigned long prot = 0;
3677 	void __iomem *maddr;
3678 	int offset = addr & (PAGE_SIZE-1);
3679 
3680 	if (follow_phys(vma, addr, write, &prot, &phys_addr))
3681 		return -EINVAL;
3682 
3683 	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3684 	if (write)
3685 		memcpy_toio(maddr + offset, buf, len);
3686 	else
3687 		memcpy_fromio(buf, maddr + offset, len);
3688 	iounmap(maddr);
3689 
3690 	return len;
3691 }
3692 EXPORT_SYMBOL_GPL(generic_access_phys);
3693 #endif
3694 
3695 /*
3696  * Access another process' address space as given in mm.  If non-NULL, use the
3697  * given task for page fault accounting.
3698  */
__access_remote_vm(struct task_struct * tsk,struct mm_struct * mm,unsigned long addr,void * buf,int len,int write)3699 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3700 		unsigned long addr, void *buf, int len, int write)
3701 {
3702 	struct vm_area_struct *vma;
3703 	void *old_buf = buf;
3704 
3705 	down_read(&mm->mmap_sem);
3706 	/* ignore errors, just check how much was successfully transferred */
3707 	while (len) {
3708 		int bytes, ret, offset;
3709 		void *maddr;
3710 		struct page *page = NULL;
3711 
3712 		ret = get_user_pages(tsk, mm, addr, 1,
3713 				write, 1, &page, &vma);
3714 		if (ret <= 0) {
3715 #ifndef CONFIG_HAVE_IOREMAP_PROT
3716 			break;
3717 #else
3718 			/*
3719 			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3720 			 * we can access using slightly different code.
3721 			 */
3722 			vma = find_vma(mm, addr);
3723 			if (!vma || vma->vm_start > addr)
3724 				break;
3725 			if (vma->vm_ops && vma->vm_ops->access)
3726 				ret = vma->vm_ops->access(vma, addr, buf,
3727 							  len, write);
3728 			if (ret <= 0)
3729 				break;
3730 			bytes = ret;
3731 #endif
3732 		} else {
3733 			bytes = len;
3734 			offset = addr & (PAGE_SIZE-1);
3735 			if (bytes > PAGE_SIZE-offset)
3736 				bytes = PAGE_SIZE-offset;
3737 
3738 			maddr = kmap(page);
3739 			if (write) {
3740 				copy_to_user_page(vma, page, addr,
3741 						  maddr + offset, buf, bytes);
3742 				set_page_dirty_lock(page);
3743 			} else {
3744 				copy_from_user_page(vma, page, addr,
3745 						    buf, maddr + offset, bytes);
3746 			}
3747 			kunmap(page);
3748 			page_cache_release(page);
3749 		}
3750 		len -= bytes;
3751 		buf += bytes;
3752 		addr += bytes;
3753 	}
3754 	up_read(&mm->mmap_sem);
3755 
3756 	return buf - old_buf;
3757 }
3758 
3759 /**
3760  * access_remote_vm - access another process' address space
3761  * @mm:		the mm_struct of the target address space
3762  * @addr:	start address to access
3763  * @buf:	source or destination buffer
3764  * @len:	number of bytes to transfer
3765  * @write:	whether the access is a write
3766  *
3767  * The caller must hold a reference on @mm.
3768  */
access_remote_vm(struct mm_struct * mm,unsigned long addr,void * buf,int len,int write)3769 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3770 		void *buf, int len, int write)
3771 {
3772 	return __access_remote_vm(NULL, mm, addr, buf, len, write);
3773 }
3774 
3775 /*
3776  * Access another process' address space.
3777  * Source/target buffer must be kernel space,
3778  * Do not walk the page table directly, use get_user_pages
3779  */
access_process_vm(struct task_struct * tsk,unsigned long addr,void * buf,int len,int write)3780 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3781 		void *buf, int len, int write)
3782 {
3783 	struct mm_struct *mm;
3784 	int ret;
3785 
3786 	mm = get_task_mm(tsk);
3787 	if (!mm)
3788 		return 0;
3789 
3790 	ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3791 	mmput(mm);
3792 
3793 	return ret;
3794 }
3795 
3796 /*
3797  * Print the name of a VMA.
3798  */
print_vma_addr(char * prefix,unsigned long ip)3799 void print_vma_addr(char *prefix, unsigned long ip)
3800 {
3801 	struct mm_struct *mm = current->mm;
3802 	struct vm_area_struct *vma;
3803 
3804 	/*
3805 	 * Do not print if we are in atomic
3806 	 * contexts (in exception stacks, etc.):
3807 	 */
3808 	if (preempt_count())
3809 		return;
3810 
3811 	down_read(&mm->mmap_sem);
3812 	vma = find_vma(mm, ip);
3813 	if (vma && vma->vm_file) {
3814 		struct file *f = vma->vm_file;
3815 		char *buf = (char *)__get_free_page(GFP_KERNEL);
3816 		if (buf) {
3817 			char *p;
3818 
3819 			p = file_path(f, buf, PAGE_SIZE);
3820 			if (IS_ERR(p))
3821 				p = "?";
3822 			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3823 					vma->vm_start,
3824 					vma->vm_end - vma->vm_start);
3825 			free_page((unsigned long)buf);
3826 		}
3827 	}
3828 	up_read(&mm->mmap_sem);
3829 }
3830 
3831 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
__might_fault(const char * file,int line)3832 void __might_fault(const char *file, int line)
3833 {
3834 	/*
3835 	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3836 	 * holding the mmap_sem, this is safe because kernel memory doesn't
3837 	 * get paged out, therefore we'll never actually fault, and the
3838 	 * below annotations will generate false positives.
3839 	 */
3840 	if (segment_eq(get_fs(), KERNEL_DS))
3841 		return;
3842 	if (pagefault_disabled())
3843 		return;
3844 	__might_sleep(file, line, 0);
3845 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3846 	if (current->mm)
3847 		might_lock_read(&current->mm->mmap_sem);
3848 #endif
3849 }
3850 EXPORT_SYMBOL(__might_fault);
3851 #endif
3852 
3853 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
clear_gigantic_page(struct page * page,unsigned long addr,unsigned int pages_per_huge_page)3854 static void clear_gigantic_page(struct page *page,
3855 				unsigned long addr,
3856 				unsigned int pages_per_huge_page)
3857 {
3858 	int i;
3859 	struct page *p = page;
3860 
3861 	might_sleep();
3862 	for (i = 0; i < pages_per_huge_page;
3863 	     i++, p = mem_map_next(p, page, i)) {
3864 		cond_resched();
3865 		clear_user_highpage(p, addr + i * PAGE_SIZE);
3866 	}
3867 }
clear_huge_page(struct page * page,unsigned long addr,unsigned int pages_per_huge_page)3868 void clear_huge_page(struct page *page,
3869 		     unsigned long addr, unsigned int pages_per_huge_page)
3870 {
3871 	int i;
3872 
3873 	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3874 		clear_gigantic_page(page, addr, pages_per_huge_page);
3875 		return;
3876 	}
3877 
3878 	might_sleep();
3879 	for (i = 0; i < pages_per_huge_page; i++) {
3880 		cond_resched();
3881 		clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3882 	}
3883 }
3884 
copy_user_gigantic_page(struct page * dst,struct page * src,unsigned long addr,struct vm_area_struct * vma,unsigned int pages_per_huge_page)3885 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3886 				    unsigned long addr,
3887 				    struct vm_area_struct *vma,
3888 				    unsigned int pages_per_huge_page)
3889 {
3890 	int i;
3891 	struct page *dst_base = dst;
3892 	struct page *src_base = src;
3893 
3894 	for (i = 0; i < pages_per_huge_page; ) {
3895 		cond_resched();
3896 		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3897 
3898 		i++;
3899 		dst = mem_map_next(dst, dst_base, i);
3900 		src = mem_map_next(src, src_base, i);
3901 	}
3902 }
3903 
copy_user_huge_page(struct page * dst,struct page * src,unsigned long addr,struct vm_area_struct * vma,unsigned int pages_per_huge_page)3904 void copy_user_huge_page(struct page *dst, struct page *src,
3905 			 unsigned long addr, struct vm_area_struct *vma,
3906 			 unsigned int pages_per_huge_page)
3907 {
3908 	int i;
3909 
3910 	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3911 		copy_user_gigantic_page(dst, src, addr, vma,
3912 					pages_per_huge_page);
3913 		return;
3914 	}
3915 
3916 	might_sleep();
3917 	for (i = 0; i < pages_per_huge_page; i++) {
3918 		cond_resched();
3919 		copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3920 	}
3921 }
3922 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3923 
3924 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3925 
3926 static struct kmem_cache *page_ptl_cachep;
3927 
ptlock_cache_init(void)3928 void __init ptlock_cache_init(void)
3929 {
3930 	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3931 			SLAB_PANIC, NULL);
3932 }
3933 
ptlock_alloc(struct page * page)3934 bool ptlock_alloc(struct page *page)
3935 {
3936 	spinlock_t *ptl;
3937 
3938 	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3939 	if (!ptl)
3940 		return false;
3941 	page->ptl = ptl;
3942 	return true;
3943 }
3944 
ptlock_free(struct page * page)3945 void ptlock_free(struct page *page)
3946 {
3947 	kmem_cache_free(page_ptl_cachep, page->ptl);
3948 }
3949 #endif
3950