?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
pagemap: avoid splitting thp when reading /proc/pid/pagemap
[linux/fpc-iii.git] / fs / proc / task_mmu.c
blob95264c0ef30888172f9851c47bba238fe00cb56c
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_for_maps(priv->task);
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 show_map_vma(struct seq_file *m, struct vm_area_struct *vma)
213 {
214 struct mm_struct *mm = vma->vm_mm;
215 struct file *file = vma->vm_file;
216 vm_flags_t flags = vma->vm_flags;
217 unsigned long ino = 0;
218 unsigned long long pgoff = 0;
219 unsigned long start, end;
220 dev_t dev = 0;
221 int len;
222
223 if (file) {
224 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
225 dev = inode->i_sb->s_dev;
226 ino = inode->i_ino;
227 pgoff = ((loff_t)vma->vm_pgoff) << PAGE_SHIFT;
228 }
229
230 /* We don't show the stack guard page in /proc/maps */
231 start = vma->vm_start;
232 if (stack_guard_page_start(vma, start))
233 start += PAGE_SIZE;
234 end = vma->vm_end;
235 if (stack_guard_page_end(vma, end))
236 end -= PAGE_SIZE;
237
238 seq_printf(m, "%08lx-%08lx %c%c%c%c %08llx %02x:%02x %lu %n",
239 start,
240 end,
241 flags & VM_READ ? 'r' : '-',
242 flags & VM_WRITE ? 'w' : '-',
243 flags & VM_EXEC ? 'x' : '-',
244 flags & VM_MAYSHARE ? 's' : 'p',
245 pgoff,
246 MAJOR(dev), MINOR(dev), ino, &len);
247
248 /*
249 * Print the dentry name for named mappings, and a
250 * special [heap] marker for the heap:
251 */
252 if (file) {
253 pad_len_spaces(m, len);
254 seq_path(m, &file->f_path, "\n");
255 } else {
256 const char *name = arch_vma_name(vma);
257 if (!name) {
258 if (mm) {
259 if (vma->vm_start <= mm->brk &&
260 vma->vm_end >= mm->start_brk) {
261 name = "[heap]";
262 } else if (vma->vm_start <= mm->start_stack &&
263 vma->vm_end >= mm->start_stack) {
264 name = "[stack]";
265 }
266 } else {
267 name = "[vdso]";
268 }
269 }
270 if (name) {
271 pad_len_spaces(m, len);
272 seq_puts(m, name);
273 }
274 }
275 seq_putc(m, '\n');
276 }
277
278 static int show_map(struct seq_file *m, void *v)
279 {
280 struct vm_area_struct *vma = v;
281 struct proc_maps_private *priv = m->private;
282 struct task_struct *task = priv->task;
283
284 show_map_vma(m, vma);
285
286 if (m->count < m->size) /* vma is copied successfully */
287 m->version = (vma != get_gate_vma(task->mm))
288 ? vma->vm_start : 0;
289 return 0;
290 }
291
292 static const struct seq_operations proc_pid_maps_op = {
293 .start = m_start,
294 .next = m_next,
295 .stop = m_stop,
296 .show = show_map
297 };
298
299 static int maps_open(struct inode *inode, struct file *file)
300 {
301 return do_maps_open(inode, file, &proc_pid_maps_op);
302 }
303
304 const struct file_operations proc_maps_operations = {
305 .open = maps_open,
306 .read = seq_read,
307 .llseek = seq_lseek,
308 .release = seq_release_private,
309 };
310
311 /*
312 * Proportional Set Size(PSS): my share of RSS.
313 *
314 * PSS of a process is the count of pages it has in memory, where each
315 * page is divided by the number of processes sharing it. So if a
316 * process has 1000 pages all to itself, and 1000 shared with one other
317 * process, its PSS will be 1500.
318 *
319 * To keep (accumulated) division errors low, we adopt a 64bit
320 * fixed-point pss counter to minimize division errors. So (pss >>
321 * PSS_SHIFT) would be the real byte count.
322 *
323 * A shift of 12 before division means (assuming 4K page size):
324 * - 1M 3-user-pages add up to 8KB errors;
325 * - supports mapcount up to 2^24, or 16M;
326 * - supports PSS up to 2^52 bytes, or 4PB.
327 */
328 #define PSS_SHIFT 12
329
330 #ifdef CONFIG_PROC_PAGE_MONITOR
331 struct mem_size_stats {
332 struct vm_area_struct *vma;
333 unsigned long resident;
334 unsigned long shared_clean;
335 unsigned long shared_dirty;
336 unsigned long private_clean;
337 unsigned long private_dirty;
338 unsigned long referenced;
339 unsigned long anonymous;
340 unsigned long anonymous_thp;
341 unsigned long swap;
342 u64 pss;
343 };
344
345
346 static void smaps_pte_entry(pte_t ptent, unsigned long addr,
347 unsigned long ptent_size, struct mm_walk *walk)
348 {
349 struct mem_size_stats *mss = walk->private;
350 struct vm_area_struct *vma = mss->vma;
351 struct page *page;
352 int mapcount;
353
354 if (is_swap_pte(ptent)) {
355 mss->swap += ptent_size;
356 return;
357 }
358
359 if (!pte_present(ptent))
360 return;
361
362 page = vm_normal_page(vma, addr, ptent);
363 if (!page)
364 return;
365
366 if (PageAnon(page))
367 mss->anonymous += ptent_size;
368
369 mss->resident += ptent_size;
370 /* Accumulate the size in pages that have been accessed. */
371 if (pte_young(ptent) || PageReferenced(page))
372 mss->referenced += ptent_size;
373 mapcount = page_mapcount(page);
374 if (mapcount >= 2) {
375 if (pte_dirty(ptent) || PageDirty(page))
376 mss->shared_dirty += ptent_size;
377 else
378 mss->shared_clean += ptent_size;
379 mss->pss += (ptent_size << PSS_SHIFT) / mapcount;
380 } else {
381 if (pte_dirty(ptent) || PageDirty(page))
382 mss->private_dirty += ptent_size;
383 else
384 mss->private_clean += ptent_size;
385 mss->pss += (ptent_size << PSS_SHIFT);
386 }
387 }
388
389 static int smaps_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
390 struct mm_walk *walk)
391 {
392 struct mem_size_stats *mss = walk->private;
393 struct vm_area_struct *vma = mss->vma;
394 pte_t *pte;
395 spinlock_t *ptl;
396
397 spin_lock(&walk->mm->page_table_lock);
398 if (pmd_trans_huge(*pmd)) {
399 if (pmd_trans_splitting(*pmd)) {
400 spin_unlock(&walk->mm->page_table_lock);
401 wait_split_huge_page(vma->anon_vma, pmd);
402 } else {
403 smaps_pte_entry(*(pte_t *)pmd, addr,
404 HPAGE_PMD_SIZE, walk);
405 spin_unlock(&walk->mm->page_table_lock);
406 mss->anonymous_thp += HPAGE_PMD_SIZE;
407 return 0;
408 }
409 } else {
410 spin_unlock(&walk->mm->page_table_lock);
411 }
412
413 if (pmd_trans_unstable(pmd))
414 return 0;
415 /*
416 * The mmap_sem held all the way back in m_start() is what
417 * keeps khugepaged out of here and from collapsing things
418 * in here.
419 */
420 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
421 for (; addr != end; pte++, addr += PAGE_SIZE)
422 smaps_pte_entry(*pte, addr, PAGE_SIZE, walk);
423 pte_unmap_unlock(pte - 1, ptl);
424 cond_resched();
425 return 0;
426 }
427
428 static int show_smap(struct seq_file *m, void *v)
429 {
430 struct proc_maps_private *priv = m->private;
431 struct task_struct *task = priv->task;
432 struct vm_area_struct *vma = v;
433 struct mem_size_stats mss;
434 struct mm_walk smaps_walk = {
435 .pmd_entry = smaps_pte_range,
436 .mm = vma->vm_mm,
437 .private = &mss,
438 };
439
440 memset(&mss, 0, sizeof mss);
441 mss.vma = vma;
442 /* mmap_sem is held in m_start */
443 if (vma->vm_mm && !is_vm_hugetlb_page(vma))
444 walk_page_range(vma->vm_start, vma->vm_end, &smaps_walk);
445
446 show_map_vma(m, vma);
447
448 seq_printf(m,
449 "Size: %8lu kB\n"
450 "Rss: %8lu kB\n"
451 "Pss: %8lu kB\n"
452 "Shared_Clean: %8lu kB\n"
453 "Shared_Dirty: %8lu kB\n"
454 "Private_Clean: %8lu kB\n"
455 "Private_Dirty: %8lu kB\n"
456 "Referenced: %8lu kB\n"
457 "Anonymous: %8lu kB\n"
458 "AnonHugePages: %8lu kB\n"
459 "Swap: %8lu kB\n"
460 "KernelPageSize: %8lu kB\n"
461 "MMUPageSize: %8lu kB\n"
462 "Locked: %8lu kB\n",
463 (vma->vm_end - vma->vm_start) >> 10,
464 mss.resident >> 10,
465 (unsigned long)(mss.pss >> (10 + PSS_SHIFT)),
466 mss.shared_clean >> 10,
467 mss.shared_dirty >> 10,
468 mss.private_clean >> 10,
469 mss.private_dirty >> 10,
470 mss.referenced >> 10,
471 mss.anonymous >> 10,
472 mss.anonymous_thp >> 10,
473 mss.swap >> 10,
474 vma_kernel_pagesize(vma) >> 10,
475 vma_mmu_pagesize(vma) >> 10,
476 (vma->vm_flags & VM_LOCKED) ?
477 (unsigned long)(mss.pss >> (10 + PSS_SHIFT)) : 0);
478
479 if (m->count < m->size) /* vma is copied successfully */
480 m->version = (vma != get_gate_vma(task->mm))
481 ? vma->vm_start : 0;
482 return 0;
483 }
484
485 static const struct seq_operations proc_pid_smaps_op = {
486 .start = m_start,
487 .next = m_next,
488 .stop = m_stop,
489 .show = show_smap
490 };
491
492 static int smaps_open(struct inode *inode, struct file *file)
493 {
494 return do_maps_open(inode, file, &proc_pid_smaps_op);
495 }
496
497 const struct file_operations proc_smaps_operations = {
498 .open = smaps_open,
499 .read = seq_read,
500 .llseek = seq_lseek,
501 .release = seq_release_private,
502 };
503
504 static int clear_refs_pte_range(pmd_t *pmd, unsigned long addr,
505 unsigned long end, struct mm_walk *walk)
506 {
507 struct vm_area_struct *vma = walk->private;
508 pte_t *pte, ptent;
509 spinlock_t *ptl;
510 struct page *page;
511
512 split_huge_page_pmd(walk->mm, pmd);
513 if (pmd_trans_unstable(pmd))
514 return 0;
515
516 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
517 for (; addr != end; pte++, addr += PAGE_SIZE) {
518 ptent = *pte;
519 if (!pte_present(ptent))
520 continue;
521
522 page = vm_normal_page(vma, addr, ptent);
523 if (!page)
524 continue;
525
526 if (PageReserved(page))
527 continue;
528
529 /* Clear accessed and referenced bits. */
530 ptep_test_and_clear_young(vma, addr, pte);
531 ClearPageReferenced(page);
532 }
533 pte_unmap_unlock(pte - 1, ptl);
534 cond_resched();
535 return 0;
536 }
537
538 #define CLEAR_REFS_ALL 1
539 #define CLEAR_REFS_ANON 2
540 #define CLEAR_REFS_MAPPED 3
541
542 static ssize_t clear_refs_write(struct file *file, const char __user *buf,
543 size_t count, loff_t *ppos)
544 {
545 struct task_struct *task;
546 char buffer[PROC_NUMBUF];
547 struct mm_struct *mm;
548 struct vm_area_struct *vma;
549 int type;
550 int rv;
551
552 memset(buffer, 0, sizeof(buffer));
553 if (count > sizeof(buffer) - 1)
554 count = sizeof(buffer) - 1;
555 if (copy_from_user(buffer, buf, count))
556 return -EFAULT;
557 rv = kstrtoint(strstrip(buffer), 10, &type);
558 if (rv < 0)
559 return rv;
560 if (type < CLEAR_REFS_ALL || type > CLEAR_REFS_MAPPED)
561 return -EINVAL;
562 task = get_proc_task(file->f_path.dentry->d_inode);
563 if (!task)
564 return -ESRCH;
565 mm = get_task_mm(task);
566 if (mm) {
567 struct mm_walk clear_refs_walk = {
568 .pmd_entry = clear_refs_pte_range,
569 .mm = mm,
570 };
571 down_read(&mm->mmap_sem);
572 for (vma = mm->mmap; vma; vma = vma->vm_next) {
573 clear_refs_walk.private = vma;
574 if (is_vm_hugetlb_page(vma))
575 continue;
576 /*
577 * Writing 1 to /proc/pid/clear_refs affects all pages.
578 *
579 * Writing 2 to /proc/pid/clear_refs only affects
580 * Anonymous pages.
581 *
582 * Writing 3 to /proc/pid/clear_refs only affects file
583 * mapped pages.
584 */
585 if (type == CLEAR_REFS_ANON && vma->vm_file)
586 continue;
587 if (type == CLEAR_REFS_MAPPED && !vma->vm_file)
588 continue;
589 walk_page_range(vma->vm_start, vma->vm_end,
590 &clear_refs_walk);
591 }
592 flush_tlb_mm(mm);
593 up_read(&mm->mmap_sem);
594 mmput(mm);
595 }
596 put_task_struct(task);
597
598 return count;
599 }
600
601 const struct file_operations proc_clear_refs_operations = {
602 .write = clear_refs_write,
603 .llseek = noop_llseek,
604 };
605
606 struct pagemapread {
607 int pos, len;
608 u64 *buffer;
609 };
610
611 #define PAGEMAP_WALK_SIZE (PMD_SIZE)
612 #define PAGEMAP_WALK_MASK (PMD_MASK)
613
614 #define PM_ENTRY_BYTES sizeof(u64)
615 #define PM_STATUS_BITS 3
616 #define PM_STATUS_OFFSET (64 - PM_STATUS_BITS)
617 #define PM_STATUS_MASK (((1LL << PM_STATUS_BITS) - 1) << PM_STATUS_OFFSET)
618 #define PM_STATUS(nr) (((nr) << PM_STATUS_OFFSET) & PM_STATUS_MASK)
619 #define PM_PSHIFT_BITS 6
620 #define PM_PSHIFT_OFFSET (PM_STATUS_OFFSET - PM_PSHIFT_BITS)
621 #define PM_PSHIFT_MASK (((1LL << PM_PSHIFT_BITS) - 1) << PM_PSHIFT_OFFSET)
622 #define PM_PSHIFT(x) (((u64) (x) << PM_PSHIFT_OFFSET) & PM_PSHIFT_MASK)
623 #define PM_PFRAME_MASK ((1LL << PM_PSHIFT_OFFSET) - 1)
624 #define PM_PFRAME(x) ((x) & PM_PFRAME_MASK)
625
626 #define PM_PRESENT PM_STATUS(4LL)
627 #define PM_SWAP PM_STATUS(2LL)
628 #define PM_NOT_PRESENT PM_PSHIFT(PAGE_SHIFT)
629 #define PM_END_OF_BUFFER 1
630
631 static int add_to_pagemap(unsigned long addr, u64 pfn,
632 struct pagemapread *pm)
633 {
634 pm->buffer[pm->pos++] = pfn;
635 if (pm->pos >= pm->len)
636 return PM_END_OF_BUFFER;
637 return 0;
638 }
639
640 static int pagemap_pte_hole(unsigned long start, unsigned long end,
641 struct mm_walk *walk)
642 {
643 struct pagemapread *pm = walk->private;
644 unsigned long addr;
645 int err = 0;
646 for (addr = start; addr < end; addr += PAGE_SIZE) {
647 err = add_to_pagemap(addr, PM_NOT_PRESENT, pm);
648 if (err)
649 break;
650 }
651 return err;
652 }
653
654 static u64 swap_pte_to_pagemap_entry(pte_t pte)
655 {
656 swp_entry_t e = pte_to_swp_entry(pte);
657 return swp_type(e) | (swp_offset(e) << MAX_SWAPFILES_SHIFT);
658 }
659
660 static u64 pte_to_pagemap_entry(pte_t pte)
661 {
662 u64 pme = 0;
663 if (is_swap_pte(pte))
664 pme = PM_PFRAME(swap_pte_to_pagemap_entry(pte))
665 | PM_PSHIFT(PAGE_SHIFT) | PM_SWAP;
666 else if (pte_present(pte))
667 pme = PM_PFRAME(pte_pfn(pte))
668 | PM_PSHIFT(PAGE_SHIFT) | PM_PRESENT;
669 return pme;
670 }
671
672 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
673 static u64 thp_pmd_to_pagemap_entry(pmd_t pmd, int offset)
674 {
675 u64 pme = 0;
676 /*
677 * Currently pmd for thp is always present because thp can not be
678 * swapped-out, migrated, or HWPOISONed (split in such cases instead.)
679 * This if-check is just to prepare for future implementation.
680 */
681 if (pmd_present(pmd))
682 pme = PM_PFRAME(pmd_pfn(pmd) + offset)
683 | PM_PSHIFT(PAGE_SHIFT) | PM_PRESENT;
684 return pme;
685 }
686 #else
687 static inline u64 thp_pmd_to_pagemap_entry(pmd_t pmd, int offset)
688 {
689 return 0;
690 }
691 #endif
692
693 static int pagemap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
694 struct mm_walk *walk)
695 {
696 struct vm_area_struct *vma;
697 struct pagemapread *pm = walk->private;
698 pte_t *pte;
699 int err = 0;
700 u64 pfn = PM_NOT_PRESENT;
701
702 if (pmd_trans_unstable(pmd))
703 return 0;
704
705 /* find the first VMA at or above 'addr' */
706 vma = find_vma(walk->mm, addr);
707 spin_lock(&walk->mm->page_table_lock);
708 if (pmd_trans_huge(*pmd)) {
709 if (pmd_trans_splitting(*pmd)) {
710 spin_unlock(&walk->mm->page_table_lock);
711 wait_split_huge_page(vma->anon_vma, pmd);
712 } else {
713 for (; addr != end; addr += PAGE_SIZE) {
714 unsigned long offset;
715
716 offset = (addr & ~PAGEMAP_WALK_MASK) >>
717 PAGE_SHIFT;
718 pfn = thp_pmd_to_pagemap_entry(*pmd, offset);
719 err = add_to_pagemap(addr, pfn, pm);
720 if (err)
721 break;
722 }
723 spin_unlock(&walk->mm->page_table_lock);
724 return err;
725 }
726 } else {
727 spin_unlock(&walk->mm->page_table_lock);
728 }
729
730 for (; addr != end; addr += PAGE_SIZE) {
731
732 /* check to see if we've left 'vma' behind
733 * and need a new, higher one */
734 if (vma && (addr >= vma->vm_end))
735 vma = find_vma(walk->mm, addr);
736
737 /* check that 'vma' actually covers this address,
738 * and that it isn't a huge page vma */
739 if (vma && (vma->vm_start <= addr) &&
740 !is_vm_hugetlb_page(vma)) {
741 pte = pte_offset_map(pmd, addr);
742 pfn = pte_to_pagemap_entry(*pte);
743 /* unmap before userspace copy */
744 pte_unmap(pte);
745 }
746 err = add_to_pagemap(addr, pfn, pm);
747 if (err)
748 return err;
749 }
750
751 cond_resched();
752
753 return err;
754 }
755
756 #ifdef CONFIG_HUGETLB_PAGE
757 static u64 huge_pte_to_pagemap_entry(pte_t pte, int offset)
758 {
759 u64 pme = 0;
760 if (pte_present(pte))
761 pme = PM_PFRAME(pte_pfn(pte) + offset)
762 | PM_PSHIFT(PAGE_SHIFT) | PM_PRESENT;
763 return pme;
764 }
765
766 /* This function walks within one hugetlb entry in the single call */
767 static int pagemap_hugetlb_range(pte_t *pte, unsigned long hmask,
768 unsigned long addr, unsigned long end,
769 struct mm_walk *walk)
770 {
771 struct pagemapread *pm = walk->private;
772 int err = 0;
773 u64 pfn;
774
775 for (; addr != end; addr += PAGE_SIZE) {
776 int offset = (addr & ~hmask) >> PAGE_SHIFT;
777 pfn = huge_pte_to_pagemap_entry(*pte, offset);
778 err = add_to_pagemap(addr, pfn, pm);
779 if (err)
780 return err;
781 }
782
783 cond_resched();
784
785 return err;
786 }
787 #endif /* HUGETLB_PAGE */
788
789 /*
790 * /proc/pid/pagemap - an array mapping virtual pages to pfns
791 *
792 * For each page in the address space, this file contains one 64-bit entry
793 * consisting of the following:
794 *
795 * Bits 0-55 page frame number (PFN) if present
796 * Bits 0-4 swap type if swapped
797 * Bits 5-55 swap offset if swapped
798 * Bits 55-60 page shift (page size = 1<<page shift)
799 * Bit 61 reserved for future use
800 * Bit 62 page swapped
801 * Bit 63 page present
802 *
803 * If the page is not present but in swap, then the PFN contains an
804 * encoding of the swap file number and the page's offset into the
805 * swap. Unmapped pages return a null PFN. This allows determining
806 * precisely which pages are mapped (or in swap) and comparing mapped
807 * pages between processes.
808 *
809 * Efficient users of this interface will use /proc/pid/maps to
810 * determine which areas of memory are actually mapped and llseek to
811 * skip over unmapped regions.
812 */
813 static ssize_t pagemap_read(struct file *file, char __user *buf,
814 size_t count, loff_t *ppos)
815 {
816 struct task_struct *task = get_proc_task(file->f_path.dentry->d_inode);
817 struct mm_struct *mm;
818 struct pagemapread pm;
819 int ret = -ESRCH;
820 struct mm_walk pagemap_walk = {};
821 unsigned long src;
822 unsigned long svpfn;
823 unsigned long start_vaddr;
824 unsigned long end_vaddr;
825 int copied = 0;
826
827 if (!task)
828 goto out;
829
830 ret = -EINVAL;
831 /* file position must be aligned */
832 if ((*ppos % PM_ENTRY_BYTES) || (count % PM_ENTRY_BYTES))
833 goto out_task;
834
835 ret = 0;
836 if (!count)
837 goto out_task;
838
839 pm.len = PM_ENTRY_BYTES * (PAGEMAP_WALK_SIZE >> PAGE_SHIFT);
840 pm.buffer = kmalloc(pm.len, GFP_TEMPORARY);
841 ret = -ENOMEM;
842 if (!pm.buffer)
843 goto out_task;
844
845 mm = mm_for_maps(task);
846 ret = PTR_ERR(mm);
847 if (!mm || IS_ERR(mm))
848 goto out_free;
849
850 pagemap_walk.pmd_entry = pagemap_pte_range;
851 pagemap_walk.pte_hole = pagemap_pte_hole;
852 #ifdef CONFIG_HUGETLB_PAGE
853 pagemap_walk.hugetlb_entry = pagemap_hugetlb_range;
854 #endif
855 pagemap_walk.mm = mm;
856 pagemap_walk.private = &pm;
857
858 src = *ppos;
859 svpfn = src / PM_ENTRY_BYTES;
860 start_vaddr = svpfn << PAGE_SHIFT;
861 end_vaddr = TASK_SIZE_OF(task);
862
863 /* watch out for wraparound */
864 if (svpfn > TASK_SIZE_OF(task) >> PAGE_SHIFT)
865 start_vaddr = end_vaddr;
866
867 /*
868 * The odds are that this will stop walking way
869 * before end_vaddr, because the length of the
870 * user buffer is tracked in "pm", and the walk
871 * will stop when we hit the end of the buffer.
872 */
873 ret = 0;
874 while (count && (start_vaddr < end_vaddr)) {
875 int len;
876 unsigned long end;
877
878 pm.pos = 0;
879 end = (start_vaddr + PAGEMAP_WALK_SIZE) & PAGEMAP_WALK_MASK;
880 /* overflow ? */
881 if (end < start_vaddr || end > end_vaddr)
882 end = end_vaddr;
883 down_read(&mm->mmap_sem);
884 ret = walk_page_range(start_vaddr, end, &pagemap_walk);
885 up_read(&mm->mmap_sem);
886 start_vaddr = end;
887
888 len = min(count, PM_ENTRY_BYTES * pm.pos);
889 if (copy_to_user(buf, pm.buffer, len)) {
890 ret = -EFAULT;
891 goto out_mm;
892 }
893 copied += len;
894 buf += len;
895 count -= len;
896 }
897 *ppos += copied;
898 if (!ret || ret == PM_END_OF_BUFFER)
899 ret = copied;
900
901 out_mm:
902 mmput(mm);
903 out_free:
904 kfree(pm.buffer);
905 out_task:
906 put_task_struct(task);
907 out:
908 return ret;
909 }
910
911 const struct file_operations proc_pagemap_operations = {
912 .llseek = mem_lseek, /* borrow this */
913 .read = pagemap_read,
914 };
915 #endif /* CONFIG_PROC_PAGE_MONITOR */
916
917 #ifdef CONFIG_NUMA
918
919 struct numa_maps {
920 struct vm_area_struct *vma;
921 unsigned long pages;
922 unsigned long anon;
923 unsigned long active;
924 unsigned long writeback;
925 unsigned long mapcount_max;
926 unsigned long dirty;
927 unsigned long swapcache;
928 unsigned long node[MAX_NUMNODES];
929 };
930
931 struct numa_maps_private {
932 struct proc_maps_private proc_maps;
933 struct numa_maps md;
934 };
935
936 static void gather_stats(struct page *page, struct numa_maps *md, int pte_dirty,
937 unsigned long nr_pages)
938 {
939 int count = page_mapcount(page);
940
941 md->pages += nr_pages;
942 if (pte_dirty || PageDirty(page))
943 md->dirty += nr_pages;
944
945 if (PageSwapCache(page))
946 md->swapcache += nr_pages;
947
948 if (PageActive(page) || PageUnevictable(page))
949 md->active += nr_pages;
950
951 if (PageWriteback(page))
952 md->writeback += nr_pages;
953
954 if (PageAnon(page))
955 md->anon += nr_pages;
956
957 if (count > md->mapcount_max)
958 md->mapcount_max = count;
959
960 md->node[page_to_nid(page)] += nr_pages;
961 }
962
963 static struct page *can_gather_numa_stats(pte_t pte, struct vm_area_struct *vma,
964 unsigned long addr)
965 {
966 struct page *page;
967 int nid;
968
969 if (!pte_present(pte))
970 return NULL;
971
972 page = vm_normal_page(vma, addr, pte);
973 if (!page)
974 return NULL;
975
976 if (PageReserved(page))
977 return NULL;
978
979 nid = page_to_nid(page);
980 if (!node_isset(nid, node_states[N_HIGH_MEMORY]))
981 return NULL;
982
983 return page;
984 }
985
986 static int gather_pte_stats(pmd_t *pmd, unsigned long addr,
987 unsigned long end, struct mm_walk *walk)
988 {
989 struct numa_maps *md;
990 spinlock_t *ptl;
991 pte_t *orig_pte;
992 pte_t *pte;
993
994 md = walk->private;
995 spin_lock(&walk->mm->page_table_lock);
996 if (pmd_trans_huge(*pmd)) {
997 if (pmd_trans_splitting(*pmd)) {
998 spin_unlock(&walk->mm->page_table_lock);
999 wait_split_huge_page(md->vma->anon_vma, pmd);
1000 } else {
1001 pte_t huge_pte = *(pte_t *)pmd;
1002 struct page *page;
1003
1004 page = can_gather_numa_stats(huge_pte, md->vma, addr);
1005 if (page)
1006 gather_stats(page, md, pte_dirty(huge_pte),
1007 HPAGE_PMD_SIZE/PAGE_SIZE);
1008 spin_unlock(&walk->mm->page_table_lock);
1009 return 0;
1010 }
1011 } else {
1012 spin_unlock(&walk->mm->page_table_lock);
1013 }
1014
1015 if (pmd_trans_unstable(pmd))
1016 return 0;
1017 orig_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
1018 do {
1019 struct page *page = can_gather_numa_stats(*pte, md->vma, addr);
1020 if (!page)
1021 continue;
1022 gather_stats(page, md, pte_dirty(*pte), 1);
1023
1024 } while (pte++, addr += PAGE_SIZE, addr != end);
1025 pte_unmap_unlock(orig_pte, ptl);
1026 return 0;
1027 }
1028 #ifdef CONFIG_HUGETLB_PAGE
1029 static int gather_hugetbl_stats(pte_t *pte, unsigned long hmask,
1030 unsigned long addr, unsigned long end, struct mm_walk *walk)
1031 {
1032 struct numa_maps *md;
1033 struct page *page;
1034
1035 if (pte_none(*pte))
1036 return 0;
1037
1038 page = pte_page(*pte);
1039 if (!page)
1040 return 0;
1041
1042 md = walk->private;
1043 gather_stats(page, md, pte_dirty(*pte), 1);
1044 return 0;
1045 }
1046
1047 #else
1048 static int gather_hugetbl_stats(pte_t *pte, unsigned long hmask,
1049 unsigned long addr, unsigned long end, struct mm_walk *walk)
1050 {
1051 return 0;
1052 }
1053 #endif
1054
1055 /*
1056 * Display pages allocated per node and memory policy via /proc.
1057 */
1058 static int show_numa_map(struct seq_file *m, void *v)
1059 {
1060 struct numa_maps_private *numa_priv = m->private;
1061 struct proc_maps_private *proc_priv = &numa_priv->proc_maps;
1062 struct vm_area_struct *vma = v;
1063 struct numa_maps *md = &numa_priv->md;
1064 struct file *file = vma->vm_file;
1065 struct mm_struct *mm = vma->vm_mm;
1066 struct mm_walk walk = {};
1067 struct mempolicy *pol;
1068 int n;
1069 char buffer[50];
1070
1071 if (!mm)
1072 return 0;
1073
1074 /* Ensure we start with an empty set of numa_maps statistics. */
1075 memset(md, 0, sizeof(*md));
1076
1077 md->vma = vma;
1078
1079 walk.hugetlb_entry = gather_hugetbl_stats;
1080 walk.pmd_entry = gather_pte_stats;
1081 walk.private = md;
1082 walk.mm = mm;
1083
1084 pol = get_vma_policy(proc_priv->task, vma, vma->vm_start);
1085 mpol_to_str(buffer, sizeof(buffer), pol, 0);
1086 mpol_cond_put(pol);
1087
1088 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1089
1090 if (file) {
1091 seq_printf(m, " file=");
1092 seq_path(m, &file->f_path, "\n\t= ");
1093 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1094 seq_printf(m, " heap");
1095 } else if (vma->vm_start <= mm->start_stack &&
1096 vma->vm_end >= mm->start_stack) {
1097 seq_printf(m, " stack");
1098 }
1099
1100 if (is_vm_hugetlb_page(vma))
1101 seq_printf(m, " huge");
1102
1103 walk_page_range(vma->vm_start, vma->vm_end, &walk);
1104
1105 if (!md->pages)
1106 goto out;
1107
1108 if (md->anon)
1109 seq_printf(m, " anon=%lu", md->anon);
1110
1111 if (md->dirty)
1112 seq_printf(m, " dirty=%lu", md->dirty);
1113
1114 if (md->pages != md->anon && md->pages != md->dirty)
1115 seq_printf(m, " mapped=%lu", md->pages);
1116
1117 if (md->mapcount_max > 1)
1118 seq_printf(m, " mapmax=%lu", md->mapcount_max);
1119
1120 if (md->swapcache)
1121 seq_printf(m, " swapcache=%lu", md->swapcache);
1122
1123 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
1124 seq_printf(m, " active=%lu", md->active);
1125
1126 if (md->writeback)
1127 seq_printf(m, " writeback=%lu", md->writeback);
1128
1129 for_each_node_state(n, N_HIGH_MEMORY)
1130 if (md->node[n])
1131 seq_printf(m, " N%d=%lu", n, md->node[n]);
1132 out:
1133 seq_putc(m, '\n');
1134
1135 if (m->count < m->size)
1136 m->version = (vma != proc_priv->tail_vma) ? vma->vm_start : 0;
1137 return 0;
1138 }
1139
1140 static const struct seq_operations proc_pid_numa_maps_op = {
1141 .start = m_start,
1142 .next = m_next,
1143 .stop = m_stop,
1144 .show = show_numa_map,
1145 };
1146
1147 static int numa_maps_open(struct inode *inode, struct file *file)
1148 {
1149 struct numa_maps_private *priv;
1150 int ret = -ENOMEM;
1151 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
1152 if (priv) {
1153 priv->proc_maps.pid = proc_pid(inode);
1154 ret = seq_open(file, &proc_pid_numa_maps_op);
1155 if (!ret) {
1156 struct seq_file *m = file->private_data;
1157 m->private = priv;
1158 } else {
1159 kfree(priv);
1160 }
1161 }
1162 return ret;
1163 }
1164
1165 const struct file_operations proc_numa_maps_operations = {
1166 .open = numa_maps_open,
1167 .read = seq_read,
1168 .llseek = seq_lseek,
1169 .release = seq_release_private,
1170 };
1171 #endif /* CONFIG_NUMA */
Linux with added FPC-III support