xref: /linux/fs/proc/task_mmu.c (revision 26b0d14106954ae46d2f4f7eec3481828a210f7d)
1 #include <linux/mm.h>
2 #include <linux/hugetlb.h>
3 #include <linux/huge_mm.h>
4 #include <linux/mount.h>
5 #include <linux/seq_file.h>
6 #include <linux/highmem.h>
7 #include <linux/ptrace.h>
8 #include <linux/slab.h>
9 #include <linux/pagemap.h>
10 #include <linux/mempolicy.h>
11 #include <linux/rmap.h>
12 #include <linux/swap.h>
13 #include <linux/swapops.h>
14 
15 #include <asm/elf.h>
16 #include <asm/uaccess.h>
17 #include <asm/tlbflush.h>
18 #include "internal.h"
19 
20 void task_mem(struct seq_file *m, struct mm_struct *mm)
21 {
22 	unsigned long data, text, lib, swap;
23 	unsigned long hiwater_vm, total_vm, hiwater_rss, total_rss;
24 
25 	/*
26 	 * Note: to minimize their overhead, mm maintains hiwater_vm and
27 	 * hiwater_rss only when about to *lower* total_vm or rss.  Any
28 	 * collector of these hiwater stats must therefore get total_vm
29 	 * and rss too, which will usually be the higher.  Barriers? not
30 	 * worth the effort, such snapshots can always be inconsistent.
31 	 */
32 	hiwater_vm = total_vm = mm->total_vm;
33 	if (hiwater_vm < mm->hiwater_vm)
34 		hiwater_vm = mm->hiwater_vm;
35 	hiwater_rss = total_rss = get_mm_rss(mm);
36 	if (hiwater_rss < mm->hiwater_rss)
37 		hiwater_rss = mm->hiwater_rss;
38 
39 	data = mm->total_vm - mm->shared_vm - mm->stack_vm;
40 	text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK)) >> 10;
41 	lib = (mm->exec_vm << (PAGE_SHIFT-10)) - text;
42 	swap = get_mm_counter(mm, MM_SWAPENTS);
43 	seq_printf(m,
44 		"VmPeak:\t%8lu kB\n"
45 		"VmSize:\t%8lu kB\n"
46 		"VmLck:\t%8lu kB\n"
47 		"VmPin:\t%8lu kB\n"
48 		"VmHWM:\t%8lu kB\n"
49 		"VmRSS:\t%8lu kB\n"
50 		"VmData:\t%8lu kB\n"
51 		"VmStk:\t%8lu kB\n"
52 		"VmExe:\t%8lu kB\n"
53 		"VmLib:\t%8lu kB\n"
54 		"VmPTE:\t%8lu kB\n"
55 		"VmSwap:\t%8lu kB\n",
56 		hiwater_vm << (PAGE_SHIFT-10),
57 		(total_vm - mm->reserved_vm) << (PAGE_SHIFT-10),
58 		mm->locked_vm << (PAGE_SHIFT-10),
59 		mm->pinned_vm << (PAGE_SHIFT-10),
60 		hiwater_rss << (PAGE_SHIFT-10),
61 		total_rss << (PAGE_SHIFT-10),
62 		data << (PAGE_SHIFT-10),
63 		mm->stack_vm << (PAGE_SHIFT-10), text, lib,
64 		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10,
65 		swap << (PAGE_SHIFT-10));
66 }
67 
68 unsigned long task_vsize(struct mm_struct *mm)
69 {
70 	return PAGE_SIZE * mm->total_vm;
71 }
72 
73 unsigned long task_statm(struct mm_struct *mm,
74 			 unsigned long *shared, unsigned long *text,
75 			 unsigned long *data, unsigned long *resident)
76 {
77 	*shared = get_mm_counter(mm, MM_FILEPAGES);
78 	*text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK))
79 								>> PAGE_SHIFT;
80 	*data = mm->total_vm - mm->shared_vm;
81 	*resident = *shared + get_mm_counter(mm, MM_ANONPAGES);
82 	return mm->total_vm;
83 }
84 
85 static void pad_len_spaces(struct seq_file *m, int len)
86 {
87 	len = 25 + sizeof(void*) * 6 - len;
88 	if (len < 1)
89 		len = 1;
90 	seq_printf(m, "%*c", len, ' ');
91 }
92 
93 static void vma_stop(struct proc_maps_private *priv, struct vm_area_struct *vma)
94 {
95 	if (vma && vma != priv->tail_vma) {
96 		struct mm_struct *mm = vma->vm_mm;
97 		up_read(&mm->mmap_sem);
98 		mmput(mm);
99 	}
100 }
101 
102 static void *m_start(struct seq_file *m, loff_t *pos)
103 {
104 	struct proc_maps_private *priv = m->private;
105 	unsigned long last_addr = m->version;
106 	struct mm_struct *mm;
107 	struct vm_area_struct *vma, *tail_vma = NULL;
108 	loff_t l = *pos;
109 
110 	/* Clear the per syscall fields in priv */
111 	priv->task = NULL;
112 	priv->tail_vma = NULL;
113 
114 	/*
115 	 * We remember last_addr rather than next_addr to hit with
116 	 * mmap_cache most of the time. We have zero last_addr at
117 	 * the beginning and also after lseek. We will have -1 last_addr
118 	 * after the end of the vmas.
119 	 */
120 
121 	if (last_addr == -1UL)
122 		return NULL;
123 
124 	priv->task = get_pid_task(priv->pid, PIDTYPE_PID);
125 	if (!priv->task)
126 		return ERR_PTR(-ESRCH);
127 
128 	mm = mm_access(priv->task, PTRACE_MODE_READ);
129 	if (!mm || IS_ERR(mm))
130 		return mm;
131 	down_read(&mm->mmap_sem);
132 
133 	tail_vma = get_gate_vma(priv->task->mm);
134 	priv->tail_vma = tail_vma;
135 
136 	/* Start with last addr hint */
137 	vma = find_vma(mm, last_addr);
138 	if (last_addr && vma) {
139 		vma = vma->vm_next;
140 		goto out;
141 	}
142 
143 	/*
144 	 * Check the vma index is within the range and do
145 	 * sequential scan until m_index.
146 	 */
147 	vma = NULL;
148 	if ((unsigned long)l < mm->map_count) {
149 		vma = mm->mmap;
150 		while (l-- && vma)
151 			vma = vma->vm_next;
152 		goto out;
153 	}
154 
155 	if (l != mm->map_count)
156 		tail_vma = NULL; /* After gate vma */
157 
158 out:
159 	if (vma)
160 		return vma;
161 
162 	/* End of vmas has been reached */
163 	m->version = (tail_vma != NULL)? 0: -1UL;
164 	up_read(&mm->mmap_sem);
165 	mmput(mm);
166 	return tail_vma;
167 }
168 
169 static void *m_next(struct seq_file *m, void *v, loff_t *pos)
170 {
171 	struct proc_maps_private *priv = m->private;
172 	struct vm_area_struct *vma = v;
173 	struct vm_area_struct *tail_vma = priv->tail_vma;
174 
175 	(*pos)++;
176 	if (vma && (vma != tail_vma) && vma->vm_next)
177 		return vma->vm_next;
178 	vma_stop(priv, vma);
179 	return (vma != tail_vma)? tail_vma: NULL;
180 }
181 
182 static void m_stop(struct seq_file *m, void *v)
183 {
184 	struct proc_maps_private *priv = m->private;
185 	struct vm_area_struct *vma = v;
186 
187 	if (!IS_ERR(vma))
188 		vma_stop(priv, vma);
189 	if (priv->task)
190 		put_task_struct(priv->task);
191 }
192 
193 static int do_maps_open(struct inode *inode, struct file *file,
194 			const struct seq_operations *ops)
195 {
196 	struct proc_maps_private *priv;
197 	int ret = -ENOMEM;
198 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
199 	if (priv) {
200 		priv->pid = proc_pid(inode);
201 		ret = seq_open(file, ops);
202 		if (!ret) {
203 			struct seq_file *m = file->private_data;
204 			m->private = priv;
205 		} else {
206 			kfree(priv);
207 		}
208 	}
209 	return ret;
210 }
211 
212 static void
213 show_map_vma(struct seq_file *m, struct vm_area_struct *vma, int is_pid)
214 {
215 	struct mm_struct *mm = vma->vm_mm;
216 	struct file *file = vma->vm_file;
217 	struct proc_maps_private *priv = m->private;
218 	struct task_struct *task = priv->task;
219 	vm_flags_t flags = vma->vm_flags;
220 	unsigned long ino = 0;
221 	unsigned long long pgoff = 0;
222 	unsigned long start, end;
223 	dev_t dev = 0;
224 	int len;
225 	const char *name = NULL;
226 
227 	if (file) {
228 		struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
229 		dev = inode->i_sb->s_dev;
230 		ino = inode->i_ino;
231 		pgoff = ((loff_t)vma->vm_pgoff) << PAGE_SHIFT;
232 	}
233 
234 	/* We don't show the stack guard page in /proc/maps */
235 	start = vma->vm_start;
236 	if (stack_guard_page_start(vma, start))
237 		start += PAGE_SIZE;
238 	end = vma->vm_end;
239 	if (stack_guard_page_end(vma, end))
240 		end -= PAGE_SIZE;
241 
242 	seq_printf(m, "%08lx-%08lx %c%c%c%c %08llx %02x:%02x %lu %n",
243 			start,
244 			end,
245 			flags & VM_READ ? 'r' : '-',
246 			flags & VM_WRITE ? 'w' : '-',
247 			flags & VM_EXEC ? 'x' : '-',
248 			flags & VM_MAYSHARE ? 's' : 'p',
249 			pgoff,
250 			MAJOR(dev), MINOR(dev), ino, &len);
251 
252 	/*
253 	 * Print the dentry name for named mappings, and a
254 	 * special [heap] marker for the heap:
255 	 */
256 	if (file) {
257 		pad_len_spaces(m, len);
258 		seq_path(m, &file->f_path, "\n");
259 		goto done;
260 	}
261 
262 	name = arch_vma_name(vma);
263 	if (!name) {
264 		pid_t tid;
265 
266 		if (!mm) {
267 			name = "[vdso]";
268 			goto done;
269 		}
270 
271 		if (vma->vm_start <= mm->brk &&
272 		    vma->vm_end >= mm->start_brk) {
273 			name = "[heap]";
274 			goto done;
275 		}
276 
277 		tid = vm_is_stack(task, vma, is_pid);
278 
279 		if (tid != 0) {
280 			/*
281 			 * Thread stack in /proc/PID/task/TID/maps or
282 			 * the main process stack.
283 			 */
284 			if (!is_pid || (vma->vm_start <= mm->start_stack &&
285 			    vma->vm_end >= mm->start_stack)) {
286 				name = "[stack]";
287 			} else {
288 				/* Thread stack in /proc/PID/maps */
289 				pad_len_spaces(m, len);
290 				seq_printf(m, "[stack:%d]", tid);
291 			}
292 		}
293 	}
294 
295 done:
296 	if (name) {
297 		pad_len_spaces(m, len);
298 		seq_puts(m, name);
299 	}
300 	seq_putc(m, '\n');
301 }
302 
303 static int show_map(struct seq_file *m, void *v, int is_pid)
304 {
305 	struct vm_area_struct *vma = v;
306 	struct proc_maps_private *priv = m->private;
307 	struct task_struct *task = priv->task;
308 
309 	show_map_vma(m, vma, is_pid);
310 
311 	if (m->count < m->size)  /* vma is copied successfully */
312 		m->version = (vma != get_gate_vma(task->mm))
313 			? vma->vm_start : 0;
314 	return 0;
315 }
316 
317 static int show_pid_map(struct seq_file *m, void *v)
318 {
319 	return show_map(m, v, 1);
320 }
321 
322 static int show_tid_map(struct seq_file *m, void *v)
323 {
324 	return show_map(m, v, 0);
325 }
326 
327 static const struct seq_operations proc_pid_maps_op = {
328 	.start	= m_start,
329 	.next	= m_next,
330 	.stop	= m_stop,
331 	.show	= show_pid_map
332 };
333 
334 static const struct seq_operations proc_tid_maps_op = {
335 	.start	= m_start,
336 	.next	= m_next,
337 	.stop	= m_stop,
338 	.show	= show_tid_map
339 };
340 
341 static int pid_maps_open(struct inode *inode, struct file *file)
342 {
343 	return do_maps_open(inode, file, &proc_pid_maps_op);
344 }
345 
346 static int tid_maps_open(struct inode *inode, struct file *file)
347 {
348 	return do_maps_open(inode, file, &proc_tid_maps_op);
349 }
350 
351 const struct file_operations proc_pid_maps_operations = {
352 	.open		= pid_maps_open,
353 	.read		= seq_read,
354 	.llseek		= seq_lseek,
355 	.release	= seq_release_private,
356 };
357 
358 const struct file_operations proc_tid_maps_operations = {
359 	.open		= tid_maps_open,
360 	.read		= seq_read,
361 	.llseek		= seq_lseek,
362 	.release	= seq_release_private,
363 };
364 
365 /*
366  * Proportional Set Size(PSS): my share of RSS.
367  *
368  * PSS of a process is the count of pages it has in memory, where each
369  * page is divided by the number of processes sharing it.  So if a
370  * process has 1000 pages all to itself, and 1000 shared with one other
371  * process, its PSS will be 1500.
372  *
373  * To keep (accumulated) division errors low, we adopt a 64bit
374  * fixed-point pss counter to minimize division errors. So (pss >>
375  * PSS_SHIFT) would be the real byte count.
376  *
377  * A shift of 12 before division means (assuming 4K page size):
378  * 	- 1M 3-user-pages add up to 8KB errors;
379  * 	- supports mapcount up to 2^24, or 16M;
380  * 	- supports PSS up to 2^52 bytes, or 4PB.
381  */
382 #define PSS_SHIFT 12
383 
384 #ifdef CONFIG_PROC_PAGE_MONITOR
385 struct mem_size_stats {
386 	struct vm_area_struct *vma;
387 	unsigned long resident;
388 	unsigned long shared_clean;
389 	unsigned long shared_dirty;
390 	unsigned long private_clean;
391 	unsigned long private_dirty;
392 	unsigned long referenced;
393 	unsigned long anonymous;
394 	unsigned long anonymous_thp;
395 	unsigned long swap;
396 	unsigned long nonlinear;
397 	u64 pss;
398 };
399 
400 
401 static void smaps_pte_entry(pte_t ptent, unsigned long addr,
402 		unsigned long ptent_size, struct mm_walk *walk)
403 {
404 	struct mem_size_stats *mss = walk->private;
405 	struct vm_area_struct *vma = mss->vma;
406 	pgoff_t pgoff = linear_page_index(vma, addr);
407 	struct page *page = NULL;
408 	int mapcount;
409 
410 	if (pte_present(ptent)) {
411 		page = vm_normal_page(vma, addr, ptent);
412 	} else if (is_swap_pte(ptent)) {
413 		swp_entry_t swpent = pte_to_swp_entry(ptent);
414 
415 		if (!non_swap_entry(swpent))
416 			mss->swap += ptent_size;
417 		else if (is_migration_entry(swpent))
418 			page = migration_entry_to_page(swpent);
419 	} else if (pte_file(ptent)) {
420 		if (pte_to_pgoff(ptent) != pgoff)
421 			mss->nonlinear += ptent_size;
422 	}
423 
424 	if (!page)
425 		return;
426 
427 	if (PageAnon(page))
428 		mss->anonymous += ptent_size;
429 
430 	if (page->index != pgoff)
431 		mss->nonlinear += ptent_size;
432 
433 	mss->resident += ptent_size;
434 	/* Accumulate the size in pages that have been accessed. */
435 	if (pte_young(ptent) || PageReferenced(page))
436 		mss->referenced += ptent_size;
437 	mapcount = page_mapcount(page);
438 	if (mapcount >= 2) {
439 		if (pte_dirty(ptent) || PageDirty(page))
440 			mss->shared_dirty += ptent_size;
441 		else
442 			mss->shared_clean += ptent_size;
443 		mss->pss += (ptent_size << PSS_SHIFT) / mapcount;
444 	} else {
445 		if (pte_dirty(ptent) || PageDirty(page))
446 			mss->private_dirty += ptent_size;
447 		else
448 			mss->private_clean += ptent_size;
449 		mss->pss += (ptent_size << PSS_SHIFT);
450 	}
451 }
452 
453 static int smaps_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
454 			   struct mm_walk *walk)
455 {
456 	struct mem_size_stats *mss = walk->private;
457 	struct vm_area_struct *vma = mss->vma;
458 	pte_t *pte;
459 	spinlock_t *ptl;
460 
461 	if (pmd_trans_huge_lock(pmd, vma) == 1) {
462 		smaps_pte_entry(*(pte_t *)pmd, addr, HPAGE_PMD_SIZE, walk);
463 		spin_unlock(&walk->mm->page_table_lock);
464 		mss->anonymous_thp += HPAGE_PMD_SIZE;
465 		return 0;
466 	}
467 
468 	if (pmd_trans_unstable(pmd))
469 		return 0;
470 	/*
471 	 * The mmap_sem held all the way back in m_start() is what
472 	 * keeps khugepaged out of here and from collapsing things
473 	 * in here.
474 	 */
475 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
476 	for (; addr != end; pte++, addr += PAGE_SIZE)
477 		smaps_pte_entry(*pte, addr, PAGE_SIZE, walk);
478 	pte_unmap_unlock(pte - 1, ptl);
479 	cond_resched();
480 	return 0;
481 }
482 
483 static int show_smap(struct seq_file *m, void *v, int is_pid)
484 {
485 	struct proc_maps_private *priv = m->private;
486 	struct task_struct *task = priv->task;
487 	struct vm_area_struct *vma = v;
488 	struct mem_size_stats mss;
489 	struct mm_walk smaps_walk = {
490 		.pmd_entry = smaps_pte_range,
491 		.mm = vma->vm_mm,
492 		.private = &mss,
493 	};
494 
495 	memset(&mss, 0, sizeof mss);
496 	mss.vma = vma;
497 	/* mmap_sem is held in m_start */
498 	if (vma->vm_mm && !is_vm_hugetlb_page(vma))
499 		walk_page_range(vma->vm_start, vma->vm_end, &smaps_walk);
500 
501 	show_map_vma(m, vma, is_pid);
502 
503 	seq_printf(m,
504 		   "Size:           %8lu kB\n"
505 		   "Rss:            %8lu kB\n"
506 		   "Pss:            %8lu kB\n"
507 		   "Shared_Clean:   %8lu kB\n"
508 		   "Shared_Dirty:   %8lu kB\n"
509 		   "Private_Clean:  %8lu kB\n"
510 		   "Private_Dirty:  %8lu kB\n"
511 		   "Referenced:     %8lu kB\n"
512 		   "Anonymous:      %8lu kB\n"
513 		   "AnonHugePages:  %8lu kB\n"
514 		   "Swap:           %8lu kB\n"
515 		   "KernelPageSize: %8lu kB\n"
516 		   "MMUPageSize:    %8lu kB\n"
517 		   "Locked:         %8lu kB\n",
518 		   (vma->vm_end - vma->vm_start) >> 10,
519 		   mss.resident >> 10,
520 		   (unsigned long)(mss.pss >> (10 + PSS_SHIFT)),
521 		   mss.shared_clean  >> 10,
522 		   mss.shared_dirty  >> 10,
523 		   mss.private_clean >> 10,
524 		   mss.private_dirty >> 10,
525 		   mss.referenced >> 10,
526 		   mss.anonymous >> 10,
527 		   mss.anonymous_thp >> 10,
528 		   mss.swap >> 10,
529 		   vma_kernel_pagesize(vma) >> 10,
530 		   vma_mmu_pagesize(vma) >> 10,
531 		   (vma->vm_flags & VM_LOCKED) ?
532 			(unsigned long)(mss.pss >> (10 + PSS_SHIFT)) : 0);
533 
534 	if (vma->vm_flags & VM_NONLINEAR)
535 		seq_printf(m, "Nonlinear:      %8lu kB\n",
536 				mss.nonlinear >> 10);
537 
538 	if (m->count < m->size)  /* vma is copied successfully */
539 		m->version = (vma != get_gate_vma(task->mm))
540 			? vma->vm_start : 0;
541 	return 0;
542 }
543 
544 static int show_pid_smap(struct seq_file *m, void *v)
545 {
546 	return show_smap(m, v, 1);
547 }
548 
549 static int show_tid_smap(struct seq_file *m, void *v)
550 {
551 	return show_smap(m, v, 0);
552 }
553 
554 static const struct seq_operations proc_pid_smaps_op = {
555 	.start	= m_start,
556 	.next	= m_next,
557 	.stop	= m_stop,
558 	.show	= show_pid_smap
559 };
560 
561 static const struct seq_operations proc_tid_smaps_op = {
562 	.start	= m_start,
563 	.next	= m_next,
564 	.stop	= m_stop,
565 	.show	= show_tid_smap
566 };
567 
568 static int pid_smaps_open(struct inode *inode, struct file *file)
569 {
570 	return do_maps_open(inode, file, &proc_pid_smaps_op);
571 }
572 
573 static int tid_smaps_open(struct inode *inode, struct file *file)
574 {
575 	return do_maps_open(inode, file, &proc_tid_smaps_op);
576 }
577 
578 const struct file_operations proc_pid_smaps_operations = {
579 	.open		= pid_smaps_open,
580 	.read		= seq_read,
581 	.llseek		= seq_lseek,
582 	.release	= seq_release_private,
583 };
584 
585 const struct file_operations proc_tid_smaps_operations = {
586 	.open		= tid_smaps_open,
587 	.read		= seq_read,
588 	.llseek		= seq_lseek,
589 	.release	= seq_release_private,
590 };
591 
592 static int clear_refs_pte_range(pmd_t *pmd, unsigned long addr,
593 				unsigned long end, struct mm_walk *walk)
594 {
595 	struct vm_area_struct *vma = walk->private;
596 	pte_t *pte, ptent;
597 	spinlock_t *ptl;
598 	struct page *page;
599 
600 	split_huge_page_pmd(walk->mm, pmd);
601 	if (pmd_trans_unstable(pmd))
602 		return 0;
603 
604 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
605 	for (; addr != end; pte++, addr += PAGE_SIZE) {
606 		ptent = *pte;
607 		if (!pte_present(ptent))
608 			continue;
609 
610 		page = vm_normal_page(vma, addr, ptent);
611 		if (!page)
612 			continue;
613 
614 		/* Clear accessed and referenced bits. */
615 		ptep_test_and_clear_young(vma, addr, pte);
616 		ClearPageReferenced(page);
617 	}
618 	pte_unmap_unlock(pte - 1, ptl);
619 	cond_resched();
620 	return 0;
621 }
622 
623 #define CLEAR_REFS_ALL 1
624 #define CLEAR_REFS_ANON 2
625 #define CLEAR_REFS_MAPPED 3
626 
627 static ssize_t clear_refs_write(struct file *file, const char __user *buf,
628 				size_t count, loff_t *ppos)
629 {
630 	struct task_struct *task;
631 	char buffer[PROC_NUMBUF];
632 	struct mm_struct *mm;
633 	struct vm_area_struct *vma;
634 	int type;
635 	int rv;
636 
637 	memset(buffer, 0, sizeof(buffer));
638 	if (count > sizeof(buffer) - 1)
639 		count = sizeof(buffer) - 1;
640 	if (copy_from_user(buffer, buf, count))
641 		return -EFAULT;
642 	rv = kstrtoint(strstrip(buffer), 10, &type);
643 	if (rv < 0)
644 		return rv;
645 	if (type < CLEAR_REFS_ALL || type > CLEAR_REFS_MAPPED)
646 		return -EINVAL;
647 	task = get_proc_task(file->f_path.dentry->d_inode);
648 	if (!task)
649 		return -ESRCH;
650 	mm = get_task_mm(task);
651 	if (mm) {
652 		struct mm_walk clear_refs_walk = {
653 			.pmd_entry = clear_refs_pte_range,
654 			.mm = mm,
655 		};
656 		down_read(&mm->mmap_sem);
657 		for (vma = mm->mmap; vma; vma = vma->vm_next) {
658 			clear_refs_walk.private = vma;
659 			if (is_vm_hugetlb_page(vma))
660 				continue;
661 			/*
662 			 * Writing 1 to /proc/pid/clear_refs affects all pages.
663 			 *
664 			 * Writing 2 to /proc/pid/clear_refs only affects
665 			 * Anonymous pages.
666 			 *
667 			 * Writing 3 to /proc/pid/clear_refs only affects file
668 			 * mapped pages.
669 			 */
670 			if (type == CLEAR_REFS_ANON && vma->vm_file)
671 				continue;
672 			if (type == CLEAR_REFS_MAPPED && !vma->vm_file)
673 				continue;
674 			walk_page_range(vma->vm_start, vma->vm_end,
675 					&clear_refs_walk);
676 		}
677 		flush_tlb_mm(mm);
678 		up_read(&mm->mmap_sem);
679 		mmput(mm);
680 	}
681 	put_task_struct(task);
682 
683 	return count;
684 }
685 
686 const struct file_operations proc_clear_refs_operations = {
687 	.write		= clear_refs_write,
688 	.llseek		= noop_llseek,
689 };
690 
691 typedef struct {
692 	u64 pme;
693 } pagemap_entry_t;
694 
695 struct pagemapread {
696 	int pos, len;
697 	pagemap_entry_t *buffer;
698 };
699 
700 #define PAGEMAP_WALK_SIZE	(PMD_SIZE)
701 #define PAGEMAP_WALK_MASK	(PMD_MASK)
702 
703 #define PM_ENTRY_BYTES      sizeof(u64)
704 #define PM_STATUS_BITS      3
705 #define PM_STATUS_OFFSET    (64 - PM_STATUS_BITS)
706 #define PM_STATUS_MASK      (((1LL << PM_STATUS_BITS) - 1) << PM_STATUS_OFFSET)
707 #define PM_STATUS(nr)       (((nr) << PM_STATUS_OFFSET) & PM_STATUS_MASK)
708 #define PM_PSHIFT_BITS      6
709 #define PM_PSHIFT_OFFSET    (PM_STATUS_OFFSET - PM_PSHIFT_BITS)
710 #define PM_PSHIFT_MASK      (((1LL << PM_PSHIFT_BITS) - 1) << PM_PSHIFT_OFFSET)
711 #define PM_PSHIFT(x)        (((u64) (x) << PM_PSHIFT_OFFSET) & PM_PSHIFT_MASK)
712 #define PM_PFRAME_MASK      ((1LL << PM_PSHIFT_OFFSET) - 1)
713 #define PM_PFRAME(x)        ((x) & PM_PFRAME_MASK)
714 
715 #define PM_PRESENT          PM_STATUS(4LL)
716 #define PM_SWAP             PM_STATUS(2LL)
717 #define PM_FILE             PM_STATUS(1LL)
718 #define PM_NOT_PRESENT      PM_PSHIFT(PAGE_SHIFT)
719 #define PM_END_OF_BUFFER    1
720 
721 static inline pagemap_entry_t make_pme(u64 val)
722 {
723 	return (pagemap_entry_t) { .pme = val };
724 }
725 
726 static int add_to_pagemap(unsigned long addr, pagemap_entry_t *pme,
727 			  struct pagemapread *pm)
728 {
729 	pm->buffer[pm->pos++] = *pme;
730 	if (pm->pos >= pm->len)
731 		return PM_END_OF_BUFFER;
732 	return 0;
733 }
734 
735 static int pagemap_pte_hole(unsigned long start, unsigned long end,
736 				struct mm_walk *walk)
737 {
738 	struct pagemapread *pm = walk->private;
739 	unsigned long addr;
740 	int err = 0;
741 	pagemap_entry_t pme = make_pme(PM_NOT_PRESENT);
742 
743 	for (addr = start; addr < end; addr += PAGE_SIZE) {
744 		err = add_to_pagemap(addr, &pme, pm);
745 		if (err)
746 			break;
747 	}
748 	return err;
749 }
750 
751 static void pte_to_pagemap_entry(pagemap_entry_t *pme,
752 		struct vm_area_struct *vma, unsigned long addr, pte_t pte)
753 {
754 	u64 frame, flags;
755 	struct page *page = NULL;
756 
757 	if (pte_present(pte)) {
758 		frame = pte_pfn(pte);
759 		flags = PM_PRESENT;
760 		page = vm_normal_page(vma, addr, pte);
761 	} else if (is_swap_pte(pte)) {
762 		swp_entry_t entry = pte_to_swp_entry(pte);
763 
764 		frame = swp_type(entry) |
765 			(swp_offset(entry) << MAX_SWAPFILES_SHIFT);
766 		flags = PM_SWAP;
767 		if (is_migration_entry(entry))
768 			page = migration_entry_to_page(entry);
769 	} else {
770 		*pme = make_pme(PM_NOT_PRESENT);
771 		return;
772 	}
773 
774 	if (page && !PageAnon(page))
775 		flags |= PM_FILE;
776 
777 	*pme = make_pme(PM_PFRAME(frame) | PM_PSHIFT(PAGE_SHIFT) | flags);
778 }
779 
780 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
781 static void thp_pmd_to_pagemap_entry(pagemap_entry_t *pme,
782 					pmd_t pmd, int offset)
783 {
784 	/*
785 	 * Currently pmd for thp is always present because thp can not be
786 	 * swapped-out, migrated, or HWPOISONed (split in such cases instead.)
787 	 * This if-check is just to prepare for future implementation.
788 	 */
789 	if (pmd_present(pmd))
790 		*pme = make_pme(PM_PFRAME(pmd_pfn(pmd) + offset)
791 				| PM_PSHIFT(PAGE_SHIFT) | PM_PRESENT);
792 	else
793 		*pme = make_pme(PM_NOT_PRESENT);
794 }
795 #else
796 static inline void thp_pmd_to_pagemap_entry(pagemap_entry_t *pme,
797 						pmd_t pmd, int offset)
798 {
799 }
800 #endif
801 
802 static int pagemap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
803 			     struct mm_walk *walk)
804 {
805 	struct vm_area_struct *vma;
806 	struct pagemapread *pm = walk->private;
807 	pte_t *pte;
808 	int err = 0;
809 	pagemap_entry_t pme = make_pme(PM_NOT_PRESENT);
810 
811 	/* find the first VMA at or above 'addr' */
812 	vma = find_vma(walk->mm, addr);
813 	if (vma && pmd_trans_huge_lock(pmd, vma) == 1) {
814 		for (; addr != end; addr += PAGE_SIZE) {
815 			unsigned long offset;
816 
817 			offset = (addr & ~PAGEMAP_WALK_MASK) >>
818 					PAGE_SHIFT;
819 			thp_pmd_to_pagemap_entry(&pme, *pmd, offset);
820 			err = add_to_pagemap(addr, &pme, pm);
821 			if (err)
822 				break;
823 		}
824 		spin_unlock(&walk->mm->page_table_lock);
825 		return err;
826 	}
827 
828 	if (pmd_trans_unstable(pmd))
829 		return 0;
830 	for (; addr != end; addr += PAGE_SIZE) {
831 
832 		/* check to see if we've left 'vma' behind
833 		 * and need a new, higher one */
834 		if (vma && (addr >= vma->vm_end)) {
835 			vma = find_vma(walk->mm, addr);
836 			pme = make_pme(PM_NOT_PRESENT);
837 		}
838 
839 		/* check that 'vma' actually covers this address,
840 		 * and that it isn't a huge page vma */
841 		if (vma && (vma->vm_start <= addr) &&
842 		    !is_vm_hugetlb_page(vma)) {
843 			pte = pte_offset_map(pmd, addr);
844 			pte_to_pagemap_entry(&pme, vma, addr, *pte);
845 			/* unmap before userspace copy */
846 			pte_unmap(pte);
847 		}
848 		err = add_to_pagemap(addr, &pme, pm);
849 		if (err)
850 			return err;
851 	}
852 
853 	cond_resched();
854 
855 	return err;
856 }
857 
858 #ifdef CONFIG_HUGETLB_PAGE
859 static void huge_pte_to_pagemap_entry(pagemap_entry_t *pme,
860 					pte_t pte, int offset)
861 {
862 	if (pte_present(pte))
863 		*pme = make_pme(PM_PFRAME(pte_pfn(pte) + offset)
864 				| PM_PSHIFT(PAGE_SHIFT) | PM_PRESENT);
865 	else
866 		*pme = make_pme(PM_NOT_PRESENT);
867 }
868 
869 /* This function walks within one hugetlb entry in the single call */
870 static int pagemap_hugetlb_range(pte_t *pte, unsigned long hmask,
871 				 unsigned long addr, unsigned long end,
872 				 struct mm_walk *walk)
873 {
874 	struct pagemapread *pm = walk->private;
875 	int err = 0;
876 	pagemap_entry_t pme;
877 
878 	for (; addr != end; addr += PAGE_SIZE) {
879 		int offset = (addr & ~hmask) >> PAGE_SHIFT;
880 		huge_pte_to_pagemap_entry(&pme, *pte, offset);
881 		err = add_to_pagemap(addr, &pme, pm);
882 		if (err)
883 			return err;
884 	}
885 
886 	cond_resched();
887 
888 	return err;
889 }
890 #endif /* HUGETLB_PAGE */
891 
892 /*
893  * /proc/pid/pagemap - an array mapping virtual pages to pfns
894  *
895  * For each page in the address space, this file contains one 64-bit entry
896  * consisting of the following:
897  *
898  * Bits 0-54  page frame number (PFN) if present
899  * Bits 0-4   swap type if swapped
900  * Bits 5-54  swap offset if swapped
901  * Bits 55-60 page shift (page size = 1<<page shift)
902  * Bit  61    page is file-page or shared-anon
903  * Bit  62    page swapped
904  * Bit  63    page present
905  *
906  * If the page is not present but in swap, then the PFN contains an
907  * encoding of the swap file number and the page's offset into the
908  * swap. Unmapped pages return a null PFN. This allows determining
909  * precisely which pages are mapped (or in swap) and comparing mapped
910  * pages between processes.
911  *
912  * Efficient users of this interface will use /proc/pid/maps to
913  * determine which areas of memory are actually mapped and llseek to
914  * skip over unmapped regions.
915  */
916 static ssize_t pagemap_read(struct file *file, char __user *buf,
917 			    size_t count, loff_t *ppos)
918 {
919 	struct task_struct *task = get_proc_task(file->f_path.dentry->d_inode);
920 	struct mm_struct *mm;
921 	struct pagemapread pm;
922 	int ret = -ESRCH;
923 	struct mm_walk pagemap_walk = {};
924 	unsigned long src;
925 	unsigned long svpfn;
926 	unsigned long start_vaddr;
927 	unsigned long end_vaddr;
928 	int copied = 0;
929 
930 	if (!task)
931 		goto out;
932 
933 	ret = -EINVAL;
934 	/* file position must be aligned */
935 	if ((*ppos % PM_ENTRY_BYTES) || (count % PM_ENTRY_BYTES))
936 		goto out_task;
937 
938 	ret = 0;
939 	if (!count)
940 		goto out_task;
941 
942 	pm.len = PM_ENTRY_BYTES * (PAGEMAP_WALK_SIZE >> PAGE_SHIFT);
943 	pm.buffer = kmalloc(pm.len, GFP_TEMPORARY);
944 	ret = -ENOMEM;
945 	if (!pm.buffer)
946 		goto out_task;
947 
948 	mm = mm_access(task, PTRACE_MODE_READ);
949 	ret = PTR_ERR(mm);
950 	if (!mm || IS_ERR(mm))
951 		goto out_free;
952 
953 	pagemap_walk.pmd_entry = pagemap_pte_range;
954 	pagemap_walk.pte_hole = pagemap_pte_hole;
955 #ifdef CONFIG_HUGETLB_PAGE
956 	pagemap_walk.hugetlb_entry = pagemap_hugetlb_range;
957 #endif
958 	pagemap_walk.mm = mm;
959 	pagemap_walk.private = &pm;
960 
961 	src = *ppos;
962 	svpfn = src / PM_ENTRY_BYTES;
963 	start_vaddr = svpfn << PAGE_SHIFT;
964 	end_vaddr = TASK_SIZE_OF(task);
965 
966 	/* watch out for wraparound */
967 	if (svpfn > TASK_SIZE_OF(task) >> PAGE_SHIFT)
968 		start_vaddr = end_vaddr;
969 
970 	/*
971 	 * The odds are that this will stop walking way
972 	 * before end_vaddr, because the length of the
973 	 * user buffer is tracked in "pm", and the walk
974 	 * will stop when we hit the end of the buffer.
975 	 */
976 	ret = 0;
977 	while (count && (start_vaddr < end_vaddr)) {
978 		int len;
979 		unsigned long end;
980 
981 		pm.pos = 0;
982 		end = (start_vaddr + PAGEMAP_WALK_SIZE) & PAGEMAP_WALK_MASK;
983 		/* overflow ? */
984 		if (end < start_vaddr || end > end_vaddr)
985 			end = end_vaddr;
986 		down_read(&mm->mmap_sem);
987 		ret = walk_page_range(start_vaddr, end, &pagemap_walk);
988 		up_read(&mm->mmap_sem);
989 		start_vaddr = end;
990 
991 		len = min(count, PM_ENTRY_BYTES * pm.pos);
992 		if (copy_to_user(buf, pm.buffer, len)) {
993 			ret = -EFAULT;
994 			goto out_mm;
995 		}
996 		copied += len;
997 		buf += len;
998 		count -= len;
999 	}
1000 	*ppos += copied;
1001 	if (!ret || ret == PM_END_OF_BUFFER)
1002 		ret = copied;
1003 
1004 out_mm:
1005 	mmput(mm);
1006 out_free:
1007 	kfree(pm.buffer);
1008 out_task:
1009 	put_task_struct(task);
1010 out:
1011 	return ret;
1012 }
1013 
1014 const struct file_operations proc_pagemap_operations = {
1015 	.llseek		= mem_lseek, /* borrow this */
1016 	.read		= pagemap_read,
1017 };
1018 #endif /* CONFIG_PROC_PAGE_MONITOR */
1019 
1020 #ifdef CONFIG_NUMA
1021 
1022 struct numa_maps {
1023 	struct vm_area_struct *vma;
1024 	unsigned long pages;
1025 	unsigned long anon;
1026 	unsigned long active;
1027 	unsigned long writeback;
1028 	unsigned long mapcount_max;
1029 	unsigned long dirty;
1030 	unsigned long swapcache;
1031 	unsigned long node[MAX_NUMNODES];
1032 };
1033 
1034 struct numa_maps_private {
1035 	struct proc_maps_private proc_maps;
1036 	struct numa_maps md;
1037 };
1038 
1039 static void gather_stats(struct page *page, struct numa_maps *md, int pte_dirty,
1040 			unsigned long nr_pages)
1041 {
1042 	int count = page_mapcount(page);
1043 
1044 	md->pages += nr_pages;
1045 	if (pte_dirty || PageDirty(page))
1046 		md->dirty += nr_pages;
1047 
1048 	if (PageSwapCache(page))
1049 		md->swapcache += nr_pages;
1050 
1051 	if (PageActive(page) || PageUnevictable(page))
1052 		md->active += nr_pages;
1053 
1054 	if (PageWriteback(page))
1055 		md->writeback += nr_pages;
1056 
1057 	if (PageAnon(page))
1058 		md->anon += nr_pages;
1059 
1060 	if (count > md->mapcount_max)
1061 		md->mapcount_max = count;
1062 
1063 	md->node[page_to_nid(page)] += nr_pages;
1064 }
1065 
1066 static struct page *can_gather_numa_stats(pte_t pte, struct vm_area_struct *vma,
1067 		unsigned long addr)
1068 {
1069 	struct page *page;
1070 	int nid;
1071 
1072 	if (!pte_present(pte))
1073 		return NULL;
1074 
1075 	page = vm_normal_page(vma, addr, pte);
1076 	if (!page)
1077 		return NULL;
1078 
1079 	if (PageReserved(page))
1080 		return NULL;
1081 
1082 	nid = page_to_nid(page);
1083 	if (!node_isset(nid, node_states[N_HIGH_MEMORY]))
1084 		return NULL;
1085 
1086 	return page;
1087 }
1088 
1089 static int gather_pte_stats(pmd_t *pmd, unsigned long addr,
1090 		unsigned long end, struct mm_walk *walk)
1091 {
1092 	struct numa_maps *md;
1093 	spinlock_t *ptl;
1094 	pte_t *orig_pte;
1095 	pte_t *pte;
1096 
1097 	md = walk->private;
1098 
1099 	if (pmd_trans_huge_lock(pmd, md->vma) == 1) {
1100 		pte_t huge_pte = *(pte_t *)pmd;
1101 		struct page *page;
1102 
1103 		page = can_gather_numa_stats(huge_pte, md->vma, addr);
1104 		if (page)
1105 			gather_stats(page, md, pte_dirty(huge_pte),
1106 				     HPAGE_PMD_SIZE/PAGE_SIZE);
1107 		spin_unlock(&walk->mm->page_table_lock);
1108 		return 0;
1109 	}
1110 
1111 	if (pmd_trans_unstable(pmd))
1112 		return 0;
1113 	orig_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
1114 	do {
1115 		struct page *page = can_gather_numa_stats(*pte, md->vma, addr);
1116 		if (!page)
1117 			continue;
1118 		gather_stats(page, md, pte_dirty(*pte), 1);
1119 
1120 	} while (pte++, addr += PAGE_SIZE, addr != end);
1121 	pte_unmap_unlock(orig_pte, ptl);
1122 	return 0;
1123 }
1124 #ifdef CONFIG_HUGETLB_PAGE
1125 static int gather_hugetbl_stats(pte_t *pte, unsigned long hmask,
1126 		unsigned long addr, unsigned long end, struct mm_walk *walk)
1127 {
1128 	struct numa_maps *md;
1129 	struct page *page;
1130 
1131 	if (pte_none(*pte))
1132 		return 0;
1133 
1134 	page = pte_page(*pte);
1135 	if (!page)
1136 		return 0;
1137 
1138 	md = walk->private;
1139 	gather_stats(page, md, pte_dirty(*pte), 1);
1140 	return 0;
1141 }
1142 
1143 #else
1144 static int gather_hugetbl_stats(pte_t *pte, unsigned long hmask,
1145 		unsigned long addr, unsigned long end, struct mm_walk *walk)
1146 {
1147 	return 0;
1148 }
1149 #endif
1150 
1151 /*
1152  * Display pages allocated per node and memory policy via /proc.
1153  */
1154 static int show_numa_map(struct seq_file *m, void *v, int is_pid)
1155 {
1156 	struct numa_maps_private *numa_priv = m->private;
1157 	struct proc_maps_private *proc_priv = &numa_priv->proc_maps;
1158 	struct vm_area_struct *vma = v;
1159 	struct numa_maps *md = &numa_priv->md;
1160 	struct file *file = vma->vm_file;
1161 	struct mm_struct *mm = vma->vm_mm;
1162 	struct mm_walk walk = {};
1163 	struct mempolicy *pol;
1164 	int n;
1165 	char buffer[50];
1166 
1167 	if (!mm)
1168 		return 0;
1169 
1170 	/* Ensure we start with an empty set of numa_maps statistics. */
1171 	memset(md, 0, sizeof(*md));
1172 
1173 	md->vma = vma;
1174 
1175 	walk.hugetlb_entry = gather_hugetbl_stats;
1176 	walk.pmd_entry = gather_pte_stats;
1177 	walk.private = md;
1178 	walk.mm = mm;
1179 
1180 	pol = get_vma_policy(proc_priv->task, vma, vma->vm_start);
1181 	mpol_to_str(buffer, sizeof(buffer), pol, 0);
1182 	mpol_cond_put(pol);
1183 
1184 	seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1185 
1186 	if (file) {
1187 		seq_printf(m, " file=");
1188 		seq_path(m, &file->f_path, "\n\t= ");
1189 	} else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1190 		seq_printf(m, " heap");
1191 	} else {
1192 		pid_t tid = vm_is_stack(proc_priv->task, vma, is_pid);
1193 		if (tid != 0) {
1194 			/*
1195 			 * Thread stack in /proc/PID/task/TID/maps or
1196 			 * the main process stack.
1197 			 */
1198 			if (!is_pid || (vma->vm_start <= mm->start_stack &&
1199 			    vma->vm_end >= mm->start_stack))
1200 				seq_printf(m, " stack");
1201 			else
1202 				seq_printf(m, " stack:%d", tid);
1203 		}
1204 	}
1205 
1206 	if (is_vm_hugetlb_page(vma))
1207 		seq_printf(m, " huge");
1208 
1209 	walk_page_range(vma->vm_start, vma->vm_end, &walk);
1210 
1211 	if (!md->pages)
1212 		goto out;
1213 
1214 	if (md->anon)
1215 		seq_printf(m, " anon=%lu", md->anon);
1216 
1217 	if (md->dirty)
1218 		seq_printf(m, " dirty=%lu", md->dirty);
1219 
1220 	if (md->pages != md->anon && md->pages != md->dirty)
1221 		seq_printf(m, " mapped=%lu", md->pages);
1222 
1223 	if (md->mapcount_max > 1)
1224 		seq_printf(m, " mapmax=%lu", md->mapcount_max);
1225 
1226 	if (md->swapcache)
1227 		seq_printf(m, " swapcache=%lu", md->swapcache);
1228 
1229 	if (md->active < md->pages && !is_vm_hugetlb_page(vma))
1230 		seq_printf(m, " active=%lu", md->active);
1231 
1232 	if (md->writeback)
1233 		seq_printf(m, " writeback=%lu", md->writeback);
1234 
1235 	for_each_node_state(n, N_HIGH_MEMORY)
1236 		if (md->node[n])
1237 			seq_printf(m, " N%d=%lu", n, md->node[n]);
1238 out:
1239 	seq_putc(m, '\n');
1240 
1241 	if (m->count < m->size)
1242 		m->version = (vma != proc_priv->tail_vma) ? vma->vm_start : 0;
1243 	return 0;
1244 }
1245 
1246 static int show_pid_numa_map(struct seq_file *m, void *v)
1247 {
1248 	return show_numa_map(m, v, 1);
1249 }
1250 
1251 static int show_tid_numa_map(struct seq_file *m, void *v)
1252 {
1253 	return show_numa_map(m, v, 0);
1254 }
1255 
1256 static const struct seq_operations proc_pid_numa_maps_op = {
1257 	.start  = m_start,
1258 	.next   = m_next,
1259 	.stop   = m_stop,
1260 	.show   = show_pid_numa_map,
1261 };
1262 
1263 static const struct seq_operations proc_tid_numa_maps_op = {
1264 	.start  = m_start,
1265 	.next   = m_next,
1266 	.stop   = m_stop,
1267 	.show   = show_tid_numa_map,
1268 };
1269 
1270 static int numa_maps_open(struct inode *inode, struct file *file,
1271 			  const struct seq_operations *ops)
1272 {
1273 	struct numa_maps_private *priv;
1274 	int ret = -ENOMEM;
1275 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
1276 	if (priv) {
1277 		priv->proc_maps.pid = proc_pid(inode);
1278 		ret = seq_open(file, ops);
1279 		if (!ret) {
1280 			struct seq_file *m = file->private_data;
1281 			m->private = priv;
1282 		} else {
1283 			kfree(priv);
1284 		}
1285 	}
1286 	return ret;
1287 }
1288 
1289 static int pid_numa_maps_open(struct inode *inode, struct file *file)
1290 {
1291 	return numa_maps_open(inode, file, &proc_pid_numa_maps_op);
1292 }
1293 
1294 static int tid_numa_maps_open(struct inode *inode, struct file *file)
1295 {
1296 	return numa_maps_open(inode, file, &proc_tid_numa_maps_op);
1297 }
1298 
1299 const struct file_operations proc_pid_numa_maps_operations = {
1300 	.open		= pid_numa_maps_open,
1301 	.read		= seq_read,
1302 	.llseek		= seq_lseek,
1303 	.release	= seq_release_private,
1304 };
1305 
1306 const struct file_operations proc_tid_numa_maps_operations = {
1307 	.open		= tid_numa_maps_open,
1308 	.read		= seq_read,
1309 	.llseek		= seq_lseek,
1310 	.release	= seq_release_private,
1311 };
1312 #endif /* CONFIG_NUMA */
1313