kernel/tools/vm/page-types.c

   1 /*
   2  * page-types: Tool for querying page flags
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License as published by the Free
   6  * Software Foundation; version 2.
   7  *
   8  * This program is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should find a copy of v2 of the GNU General Public License somewhere on
  14  * your Linux system; if not, write to the Free Software Foundation, Inc., 59
  15  * Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Copyright (C) 2009 Intel corporation
  18  *
  19  * Authors: Wu Fengguang <fengguang.wu@intel.com>
  20  */
  21
  22 #define _FILE_OFFSET_BITS 64
  23 #define _GNU_SOURCE
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <unistd.h>
  27 #include <stdint.h>
  28 #include <stdarg.h>
  29 #include <string.h>
  30 #include <getopt.h>
  31 #include <limits.h>
  32 #include <assert.h>
  33 #include <ftw.h>
  34 #include <time.h>
  35 #include <setjmp.h>
  36 #include <signal.h>
  37 #include <sys/types.h>
  38 #include <sys/errno.h>
  39 #include <sys/fcntl.h>
  40 #include <sys/mount.h>
  41 #include <sys/statfs.h>
  42 #include <sys/mman.h>
  43 #include "../../include/uapi/linux/magic.h"
  44 #include "../../include/uapi/linux/kernel-page-flags.h"
  45 #include <api/fs/fs.h>
  46
  47 #ifndef MAX_PATH
  48 # define MAX_PATH 256
  49 #endif
  50
  51 #ifndef STR
  52 # define _STR(x) #x
  53 # define STR(x) _STR(x)
  54 #endif
  55
  56 /*
  57  * pagemap kernel ABI bits
  58  */
  59
  60 #define PM_ENTRY_BYTES          8
  61 #define PM_PFRAME_BITS          55
  62 #define PM_PFRAME_MASK          ((1LL << PM_PFRAME_BITS) - 1)
  63 #define PM_PFRAME(x)            ((x) & PM_PFRAME_MASK)
  64 #define PM_SOFT_DIRTY           (1ULL << 55)
  65 #define PM_MMAP_EXCLUSIVE       (1ULL << 56)
  66 #define PM_FILE                 (1ULL << 61)
  67 #define PM_SWAP                 (1ULL << 62)
  68 #define PM_PRESENT              (1ULL << 63)
  69
  70 /*
  71  * kernel page flags
  72  */
  73
  74 #define KPF_BYTES               8
  75 #define PROC_KPAGEFLAGS         "/proc/kpageflags"
  76
  77 /* [32-] kernel hacking assistances */
  78 #define KPF_RESERVED            32
  79 #define KPF_MLOCKED             33
  80 #define KPF_MAPPEDTODISK        34
  81 #define KPF_PRIVATE             35
  82 #define KPF_PRIVATE_2           36
  83 #define KPF_OWNER_PRIVATE       37
  84 #define KPF_ARCH                38
  85 #define KPF_UNCACHED            39
  86 #define KPF_SOFTDIRTY           40
  87
  88 /* [48-] take some arbitrary free slots for expanding overloaded flags
  89  * not part of kernel API
  90  */
  91 #define KPF_READAHEAD           48
  92 #define KPF_SLOB_FREE           49
  93 #define KPF_SLUB_FROZEN         50
  94 #define KPF_SLUB_DEBUG          51
  95 #define KPF_FILE                62
  96 #define KPF_MMAP_EXCLUSIVE      63
  97
  98 #define KPF_ALL_BITS            ((uint64_t)~0ULL)
  99 #define KPF_HACKERS_BITS        (0xffffULL << 32)
 100 #define KPF_OVERLOADED_BITS     (0xffffULL << 48)
 101 #define BIT(name)               (1ULL << KPF_##name)
 102 #define BITS_COMPOUND           (BIT(COMPOUND_HEAD) | BIT(COMPOUND_TAIL))
 103
 104 static const char * const page_flag_names[] = {
 105         [KPF_LOCKED]            = "L:locked",
 106         [KPF_ERROR]             = "E:error",
 107         [KPF_REFERENCED]        = "R:referenced",
 108         [KPF_UPTODATE]          = "U:uptodate",
 109         [KPF_DIRTY]             = "D:dirty",
 110         [KPF_LRU]               = "l:lru",
 111         [KPF_ACTIVE]            = "A:active",
 112         [KPF_SLAB]              = "S:slab",
 113         [KPF_WRITEBACK]         = "W:writeback",
 114         [KPF_RECLAIM]           = "I:reclaim",
 115         [KPF_BUDDY]             = "B:buddy",
 116
 117         [KPF_MMAP]              = "M:mmap",
 118         [KPF_ANON]              = "a:anonymous",
 119         [KPF_SWAPCACHE]         = "s:swapcache",
 120         [KPF_SWAPBACKED]        = "b:swapbacked",
 121         [KPF_COMPOUND_HEAD]     = "H:compound_head",
 122         [KPF_COMPOUND_TAIL]     = "T:compound_tail",
 123         [KPF_HUGE]              = "G:huge",
 124         [KPF_UNEVICTABLE]       = "u:unevictable",
 125         [KPF_HWPOISON]          = "X:hwpoison",
 126         [KPF_NOPAGE]            = "n:nopage",
 127         [KPF_KSM]               = "x:ksm",
 128         [KPF_THP]               = "t:thp",
 129         [KPF_BALLOON]           = "o:balloon",
 130         [KPF_ZERO_PAGE]         = "z:zero_page",
 131         [KPF_IDLE]              = "i:idle_page",
 132
 133         [KPF_RESERVED]          = "r:reserved",
 134         [KPF_MLOCKED]           = "m:mlocked",
 135         [KPF_MAPPEDTODISK]      = "d:mappedtodisk",
 136         [KPF_PRIVATE]           = "P:private",
 137         [KPF_PRIVATE_2]         = "p:private_2",
 138         [KPF_OWNER_PRIVATE]     = "O:owner_private",
 139         [KPF_ARCH]              = "h:arch",
 140         [KPF_UNCACHED]          = "c:uncached",
 141         [KPF_SOFTDIRTY]         = "f:softdirty",
 142
 143         [KPF_READAHEAD]         = "I:readahead",
 144         [KPF_SLOB_FREE]         = "P:slob_free",
 145         [KPF_SLUB_FROZEN]       = "A:slub_frozen",
 146         [KPF_SLUB_DEBUG]        = "E:slub_debug",
 147
 148         [KPF_FILE]              = "F:file",
 149         [KPF_MMAP_EXCLUSIVE]    = "1:mmap_exclusive",
 150 };
 151
 152
 153 static const char * const debugfs_known_mountpoints[] = {
 154         "/sys/kernel/debug",
 155         "/debug",
 156         0,
 157 };
 158
 159 /*
 160  * data structures
 161  */
 162
 163 static int              opt_raw;        /* for kernel developers */
 164 static int              opt_list;       /* list pages (in ranges) */
 165 static int              opt_no_summary; /* don't show summary */
 166 static pid_t            opt_pid;        /* process to walk */
 167 const char *            opt_file;
 168
 169 #define MAX_ADDR_RANGES 1024
 170 static int              nr_addr_ranges;
 171 static unsigned long    opt_offset[MAX_ADDR_RANGES];
 172 static unsigned long    opt_size[MAX_ADDR_RANGES];
 173
 174 #define MAX_VMAS        10240
 175 static int              nr_vmas;
 176 static unsigned long    pg_start[MAX_VMAS];
 177 static unsigned long    pg_end[MAX_VMAS];
 178
 179 #define MAX_BIT_FILTERS 64
 180 static int              nr_bit_filters;
 181 static uint64_t         opt_mask[MAX_BIT_FILTERS];
 182 static uint64_t         opt_bits[MAX_BIT_FILTERS];
 183
 184 static int              page_size;
 185
 186 static int              pagemap_fd;
 187 static int              kpageflags_fd;
 188
 189 static int              opt_hwpoison;
 190 static int              opt_unpoison;
 191
 192 static const char       *hwpoison_debug_fs;
 193 static int              hwpoison_inject_fd;
 194 static int              hwpoison_forget_fd;
 195
 196 #define HASH_SHIFT      13
 197 #define HASH_SIZE       (1 << HASH_SHIFT)
 198 #define HASH_MASK       (HASH_SIZE - 1)
 199 #define HASH_KEY(flags) (flags & HASH_MASK)
 200
 201 static unsigned long    total_pages;
 202 static unsigned long    nr_pages[HASH_SIZE];
 203 static uint64_t         page_flags[HASH_SIZE];
 204
 205
 206 /*
 207  * helper functions
 208  */
 209
 210 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 211
 212 #define min_t(type, x, y) ({                    \
 213         type __min1 = (x);                      \
 214         type __min2 = (y);                      \
 215         __min1 < __min2 ? __min1 : __min2; })
 216
 217 #define max_t(type, x, y) ({                    \
 218         type __max1 = (x);                      \
 219         type __max2 = (y);                      \
 220         __max1 > __max2 ? __max1 : __max2; })
 221
 222 static unsigned long pages2mb(unsigned long pages)
 223 {
 224         return (pages * page_size) >> 20;
 225 }
 226
 227 static void fatal(const char *x, ...)
 228 {
 229         va_list ap;
 230
 231         va_start(ap, x);
 232         vfprintf(stderr, x, ap);
 233         va_end(ap);
 234         exit(EXIT_FAILURE);
 235 }
 236
 237 static int checked_open(const char *pathname, int flags)
 238 {
 239         int fd = open(pathname, flags);
 240
 241         if (fd < 0) {
 242                 perror(pathname);
 243                 exit(EXIT_FAILURE);
 244         }
 245
 246         return fd;
 247 }
 248
 249 /*
 250  * pagemap/kpageflags routines
 251  */
 252
 253 static unsigned long do_u64_read(int fd, char *name,
 254                                  uint64_t *buf,
 255                                  unsigned long index,
 256                                  unsigned long count)
 257 {
 258         long bytes;
 259
 260         if (index > ULONG_MAX / 8)
 261                 fatal("index overflow: %lu\n", index);
 262
 263         bytes = pread(fd, buf, count * 8, (off_t)index * 8);
 264         if (bytes < 0) {
 265                 perror(name);
 266                 exit(EXIT_FAILURE);
 267         }
 268         if (bytes % 8)
 269                 fatal("partial read: %lu bytes\n", bytes);
 270
 271         return bytes / 8;
 272 }
 273
 274 static unsigned long kpageflags_read(uint64_t *buf,
 275                                      unsigned long index,
 276                                      unsigned long pages)
 277 {
 278         return do_u64_read(kpageflags_fd, PROC_KPAGEFLAGS, buf, index, pages);
 279 }
 280
 281 static unsigned long pagemap_read(uint64_t *buf,
 282                                   unsigned long index,
 283                                   unsigned long pages)
 284 {
 285         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 286 }
 287
 288 static unsigned long pagemap_pfn(uint64_t val)
 289 {
 290         unsigned long pfn;
 291
 292         if (val & PM_PRESENT)
 293                 pfn = PM_PFRAME(val);
 294         else
 295                 pfn = 0;
 296
 297         return pfn;
 298 }
 299
 300
 301 /*
 302  * page flag names
 303  */
 304
 305 static char *page_flag_name(uint64_t flags)
 306 {
 307         static char buf[65];
 308         int present;
 309         size_t i, j;
 310
 311         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 312                 present = (flags >> i) & 1;
 313                 if (!page_flag_names[i]) {
 314                         if (present)
 315                                 fatal("unknown flag bit %d\n", i);
 316                         continue;
 317                 }
 318                 buf[j++] = present ? page_flag_names[i][0] : '_';
 319         }
 320
 321         return buf;
 322 }
 323
 324 static char *page_flag_longname(uint64_t flags)
 325 {
 326         static char buf[1024];
 327         size_t i, n;
 328
 329         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 330                 if (!page_flag_names[i])
 331                         continue;
 332                 if ((flags >> i) & 1)
 333                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 334                                         page_flag_names[i] + 2);
 335         }
 336         if (n)
 337                 n--;
 338         buf[n] = '\0';
 339
 340         return buf;
 341 }
 342
 343
 344 /*
 345  * page list and summary
 346  */
 347
 348 static void show_page_range(unsigned long voffset, unsigned long offset,
 349                             unsigned long size, uint64_t flags)
 350 {
 351         static uint64_t      flags0;
 352         static unsigned long voff;
 353         static unsigned long index;
 354         static unsigned long count;
 355
 356         if (flags == flags0 && offset == index + count &&
 357             size && voffset == voff + count) {
 358                 count += size;
 359                 return;
 360         }
 361
 362         if (count) {
 363                 if (opt_pid)
 364                         printf("%lx\t", voff);
 365                 if (opt_file)
 366                         printf("%lu\t", voff);
 367                 printf("%lx\t%lx\t%s\n",
 368                                 index, count, page_flag_name(flags0));
 369         }
 370
 371         flags0 = flags;
 372         index  = offset;
 373         voff   = voffset;
 374         count  = size;
 375 }
 376
 377 static void flush_page_range(void)
 378 {
 379         show_page_range(0, 0, 0, 0);
 380 }
 381
 382 static void show_page(unsigned long voffset,
 383                       unsigned long offset, uint64_t flags)
 384 {
 385         if (opt_pid)
 386                 printf("%lx\t", voffset);
 387         if (opt_file)
 388                 printf("%lu\t", voffset);
 389         printf("%lx\t%s\n", offset, page_flag_name(flags));
 390 }
 391
 392 static void show_summary(void)
 393 {
 394         size_t i;
 395
 396         printf("             flags\tpage-count       MB"
 397                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 398
 399         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 400                 if (nr_pages[i])
 401                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 402                                 (unsigned long long)page_flags[i],
 403                                 nr_pages[i],
 404                                 pages2mb(nr_pages[i]),
 405                                 page_flag_name(page_flags[i]),
 406                                 page_flag_longname(page_flags[i]));
 407         }
 408
 409         printf("             total\t%10lu %8lu\n",
 410                         total_pages, pages2mb(total_pages));
 411 }
 412
 413
 414 /*
 415  * page flag filters
 416  */
 417
 418 static int bit_mask_ok(uint64_t flags)
 419 {
 420         int i;
 421
 422         for (i = 0; i < nr_bit_filters; i++) {
 423                 if (opt_bits[i] == KPF_ALL_BITS) {
 424                         if ((flags & opt_mask[i]) == 0)
 425                                 return 0;
 426                 } else {
 427                         if ((flags & opt_mask[i]) != opt_bits[i])
 428                                 return 0;
 429                 }
 430         }
 431
 432         return 1;
 433 }
 434
 435 static uint64_t expand_overloaded_flags(uint64_t flags, uint64_t pme)
 436 {
 437         /* SLOB/SLUB overload several page flags */
 438         if (flags & BIT(SLAB)) {
 439                 if (flags & BIT(PRIVATE))
 440                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 441                 if (flags & BIT(ACTIVE))
 442                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 443                 if (flags & BIT(ERROR))
 444                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 445         }
 446
 447         /* PG_reclaim is overloaded as PG_readahead in the read path */
 448         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 449                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 450
 451         if (pme & PM_SOFT_DIRTY)
 452                 flags |= BIT(SOFTDIRTY);
 453         if (pme & PM_FILE)
 454                 flags |= BIT(FILE);
 455         if (pme & PM_MMAP_EXCLUSIVE)
 456                 flags |= BIT(MMAP_EXCLUSIVE);
 457
 458         return flags;
 459 }
 460
 461 static uint64_t well_known_flags(uint64_t flags)
 462 {
 463         /* hide flags intended only for kernel hacker */
 464         flags &= ~KPF_HACKERS_BITS;
 465
 466         /* hide non-hugeTLB compound pages */
 467         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 468                 flags &= ~BITS_COMPOUND;
 469
 470         return flags;
 471 }
 472
 473 static uint64_t kpageflags_flags(uint64_t flags, uint64_t pme)
 474 {
 475         if (opt_raw)
 476                 flags = expand_overloaded_flags(flags, pme);
 477         else
 478                 flags = well_known_flags(flags);
 479
 480         return flags;
 481 }
 482
 483 /*
 484  * page actions
 485  */
 486
 487 static void prepare_hwpoison_fd(void)
 488 {
 489         char buf[MAX_PATH + 1];
 490
 491         hwpoison_debug_fs = debugfs__mount();
 492         if (!hwpoison_debug_fs) {
 493                 perror("mount debugfs");
 494                 exit(EXIT_FAILURE);
 495         }
 496
 497         if (opt_hwpoison && !hwpoison_inject_fd) {
 498                 snprintf(buf, MAX_PATH, "%s/hwpoison/corrupt-pfn",
 499                         hwpoison_debug_fs);
 500                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 501         }
 502
 503         if (opt_unpoison && !hwpoison_forget_fd) {
 504                 snprintf(buf, MAX_PATH, "%s/hwpoison/unpoison-pfn",
 505                         hwpoison_debug_fs);
 506                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 507         }
 508 }
 509
 510 static int hwpoison_page(unsigned long offset)
 511 {
 512         char buf[100];
 513         int len;
 514
 515         len = sprintf(buf, "0x%lx\n", offset);
 516         len = write(hwpoison_inject_fd, buf, len);
 517         if (len < 0) {
 518                 perror("hwpoison inject");
 519                 return len;
 520         }
 521         return 0;
 522 }
 523
 524 static int unpoison_page(unsigned long offset)
 525 {
 526         char buf[100];
 527         int len;
 528
 529         len = sprintf(buf, "0x%lx\n", offset);
 530         len = write(hwpoison_forget_fd, buf, len);
 531         if (len < 0) {
 532                 perror("hwpoison forget");
 533                 return len;
 534         }
 535         return 0;
 536 }
 537
 538 /*
 539  * page frame walker
 540  */
 541
 542 static size_t hash_slot(uint64_t flags)
 543 {
 544         size_t k = HASH_KEY(flags);
 545         size_t i;
 546
 547         /* Explicitly reserve slot 0 for flags 0: the following logic
 548          * cannot distinguish an unoccupied slot from slot (flags==0).
 549          */
 550         if (flags == 0)
 551                 return 0;
 552
 553         /* search through the remaining (HASH_SIZE-1) slots */
 554         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 555                 if (!k || k >= ARRAY_SIZE(page_flags))
 556                         k = 1;
 557                 if (page_flags[k] == 0) {
 558                         page_flags[k] = flags;
 559                         return k;
 560                 }
 561                 if (page_flags[k] == flags)
 562                         return k;
 563         }
 564
 565         fatal("hash table full: bump up HASH_SHIFT?\n");
 566         exit(EXIT_FAILURE);
 567 }
 568
 569 static void add_page(unsigned long voffset,
 570                      unsigned long offset, uint64_t flags, uint64_t pme)
 571 {
 572         flags = kpageflags_flags(flags, pme);
 573
 574         if (!bit_mask_ok(flags))
 575                 return;
 576
 577         if (opt_hwpoison)
 578                 hwpoison_page(offset);
 579         if (opt_unpoison)
 580                 unpoison_page(offset);
 581
 582         if (opt_list == 1)
 583                 show_page_range(voffset, offset, 1, flags);
 584         else if (opt_list == 2)
 585                 show_page(voffset, offset, flags);
 586
 587         nr_pages[hash_slot(flags)]++;
 588         total_pages++;
 589 }
 590
 591 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 592 static void walk_pfn(unsigned long voffset,
 593                      unsigned long index,
 594                      unsigned long count,
 595                      uint64_t pme)
 596 {
 597         uint64_t buf[KPAGEFLAGS_BATCH];
 598         unsigned long batch;
 599         unsigned long pages;
 600         unsigned long i;
 601
 602         while (count) {
 603                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 604                 pages = kpageflags_read(buf, index, batch);
 605                 if (pages == 0)
 606                         break;
 607
 608                 for (i = 0; i < pages; i++)
 609                         add_page(voffset + i, index + i, buf[i], pme);
 610
 611                 index += pages;
 612                 count -= pages;
 613         }
 614 }
 615
 616 #define PAGEMAP_BATCH   (64 << 10)
 617 static void walk_vma(unsigned long index, unsigned long count)
 618 {
 619         uint64_t buf[PAGEMAP_BATCH];
 620         unsigned long batch;
 621         unsigned long pages;
 622         unsigned long pfn;
 623         unsigned long i;
 624
 625         while (count) {
 626                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 627                 pages = pagemap_read(buf, index, batch);
 628                 if (pages == 0)
 629                         break;
 630
 631                 for (i = 0; i < pages; i++) {
 632                         pfn = pagemap_pfn(buf[i]);
 633                         if (pfn)
 634                                 walk_pfn(index + i, pfn, 1, buf[i]);
 635                 }
 636
 637                 index += pages;
 638                 count -= pages;
 639         }
 640 }
 641
 642 static void walk_task(unsigned long index, unsigned long count)
 643 {
 644         const unsigned long end = index + count;
 645         unsigned long start;
 646         int i = 0;
 647
 648         while (index < end) {
 649
 650                 while (pg_end[i] <= index)
 651                         if (++i >= nr_vmas)
 652                                 return;
 653                 if (pg_start[i] >= end)
 654                         return;
 655
 656                 start = max_t(unsigned long, pg_start[i], index);
 657                 index = min_t(unsigned long, pg_end[i], end);
 658
 659                 assert(start < index);
 660                 walk_vma(start, index - start);
 661         }
 662 }
 663
 664 static void add_addr_range(unsigned long offset, unsigned long size)
 665 {
 666         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 667                 fatal("too many addr ranges\n");
 668
 669         opt_offset[nr_addr_ranges] = offset;
 670         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 671         nr_addr_ranges++;
 672 }
 673
 674 static void walk_addr_ranges(void)
 675 {
 676         int i;
 677
 678         kpageflags_fd = checked_open(PROC_KPAGEFLAGS, O_RDONLY);
 679
 680         if (!nr_addr_ranges)
 681                 add_addr_range(0, ULONG_MAX);
 682
 683         for (i = 0; i < nr_addr_ranges; i++)
 684                 if (!opt_pid)
 685                         walk_pfn(opt_offset[i], opt_offset[i], opt_size[i], 0);
 686                 else
 687                         walk_task(opt_offset[i], opt_size[i]);
 688
 689         close(kpageflags_fd);
 690 }
 691
 692
 693 /*
 694  * user interface
 695  */
 696
 697 static const char *page_flag_type(uint64_t flag)
 698 {
 699         if (flag & KPF_HACKERS_BITS)
 700                 return "(r)";
 701         if (flag & KPF_OVERLOADED_BITS)
 702                 return "(o)";
 703         return "   ";
 704 }
 705
 706 static void usage(void)
 707 {
 708         size_t i, j;
 709
 710         printf(
 711 "page-types [options]\n"
 712 "            -r|--raw                   Raw mode, for kernel developers\n"
 713 "            -d|--describe flags        Describe flags\n"
 714 "            -a|--addr    addr-spec     Walk a range of pages\n"
 715 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 716 "            -p|--pid     pid           Walk process address space\n"
 717 "            -f|--file    filename      Walk file address space\n"
 718 "            -l|--list                  Show page details in ranges\n"
 719 "            -L|--list-each             Show page details one by one\n"
 720 "            -N|--no-summary            Don't show summary info\n"
 721 "            -X|--hwpoison              hwpoison pages\n"
 722 "            -x|--unpoison              unpoison pages\n"
 723 "            -h|--help                  Show this usage message\n"
 724 "flags:\n"
 725 "            0x10                       bitfield format, e.g.\n"
 726 "            anon                       bit-name, e.g.\n"
 727 "            0x10,anon                  comma-separated list, e.g.\n"
 728 "addr-spec:\n"
 729 "            N                          one page at offset N (unit: pages)\n"
 730 "            N+M                        pages range from N to N+M-1\n"
 731 "            N,M                        pages range from N to M-1\n"
 732 "            N,                         pages range from N to end\n"
 733 "            ,M                         pages range from 0 to M-1\n"
 734 "bits-spec:\n"
 735 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 736 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 737 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 738 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 739 "bit-names:\n"
 740         );
 741
 742         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 743                 if (!page_flag_names[i])
 744                         continue;
 745                 printf("%16s%s", page_flag_names[i] + 2,
 746                                  page_flag_type(1ULL << i));
 747                 if (++j > 3) {
 748                         j = 0;
 749                         putchar('\n');
 750                 }
 751         }
 752         printf("\n                                   "
 753                 "(r) raw mode bits  (o) overloaded bits\n");
 754 }
 755
 756 static unsigned long long parse_number(const char *str)
 757 {
 758         unsigned long long n;
 759
 760         n = strtoll(str, NULL, 0);
 761
 762         if (n == 0 && str[0] != '0')
 763                 fatal("invalid name or number: %s\n", str);
 764
 765         return n;
 766 }
 767
 768 static void parse_pid(const char *str)
 769 {
 770         FILE *file;
 771         char buf[5000];
 772
 773         opt_pid = parse_number(str);
 774
 775         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 776         pagemap_fd = checked_open(buf, O_RDONLY);
 777
 778         sprintf(buf, "/proc/%d/maps", opt_pid);
 779         file = fopen(buf, "r");
 780         if (!file) {
 781                 perror(buf);
 782                 exit(EXIT_FAILURE);
 783         }
 784
 785         while (fgets(buf, sizeof(buf), file) != NULL) {
 786                 unsigned long vm_start;
 787                 unsigned long vm_end;
 788                 unsigned long long pgoff;
 789                 int major, minor;
 790                 char r, w, x, s;
 791                 unsigned long ino;
 792                 int n;
 793
 794                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 795                            &vm_start,
 796                            &vm_end,
 797                            &r, &w, &x, &s,
 798                            &pgoff,
 799                            &major, &minor,
 800                            &ino);
 801                 if (n < 10) {
 802                         fprintf(stderr, "unexpected line: %s\n", buf);
 803                         continue;
 804                 }
 805                 pg_start[nr_vmas] = vm_start / page_size;
 806                 pg_end[nr_vmas] = vm_end / page_size;
 807                 if (++nr_vmas >= MAX_VMAS) {
 808                         fprintf(stderr, "too many VMAs\n");
 809                         break;
 810                 }
 811         }
 812         fclose(file);
 813 }
 814
 815 static void show_file(const char *name, const struct stat *st)
 816 {
 817         unsigned long long size = st->st_size;
 818         char atime[64], mtime[64];
 819         long now = time(NULL);
 820
 821         printf("%s\tInode: %u\tSize: %llu (%llu pages)\n",
 822                         name, (unsigned)st->st_ino,
 823                         size, (size + page_size - 1) / page_size);
 824
 825         strftime(atime, sizeof(atime), "%c", localtime(&st->st_atime));
 826         strftime(mtime, sizeof(mtime), "%c", localtime(&st->st_mtime));
 827
 828         printf("Modify: %s (%ld seconds ago)\nAccess: %s (%ld seconds ago)\n",
 829                         mtime, now - st->st_mtime,
 830                         atime, now - st->st_atime);
 831 }
 832
 833 static sigjmp_buf sigbus_jmp;
 834
 835 static void * volatile sigbus_addr;
 836
 837 static void sigbus_handler(int sig, siginfo_t *info, void *ucontex)
 838 {
 839         (void)sig;
 840         (void)ucontex;
 841         sigbus_addr = info ? info->si_addr : NULL;
 842         siglongjmp(sigbus_jmp, 1);
 843 }
 844
 845 static struct sigaction sigbus_action = {
 846         .sa_sigaction = sigbus_handler,
 847         .sa_flags = SA_SIGINFO,
 848 };
 849
 850 static void walk_file(const char *name, const struct stat *st)
 851 {
 852         uint8_t vec[PAGEMAP_BATCH];
 853         uint64_t buf[PAGEMAP_BATCH], flags;
 854         unsigned long nr_pages, pfn, i;
 855         off_t off, end = st->st_size;
 856         int fd;
 857         ssize_t len;
 858         void *ptr;
 859         int first = 1;
 860
 861         fd = checked_open(name, O_RDONLY|O_NOATIME|O_NOFOLLOW);
 862
 863         for (off = 0; off < end; off += len) {
 864                 nr_pages = (end - off + page_size - 1) / page_size;
 865                 if (nr_pages > PAGEMAP_BATCH)
 866                         nr_pages = PAGEMAP_BATCH;
 867                 len = nr_pages * page_size;
 868
 869                 ptr = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, off);
 870                 if (ptr == MAP_FAILED)
 871                         fatal("mmap failed: %s", name);
 872
 873                 /* determine cached pages */
 874                 if (mincore(ptr, len, vec))
 875                         fatal("mincore failed: %s", name);
 876
 877                 /* turn off readahead */
 878                 if (madvise(ptr, len, MADV_RANDOM))
 879                         fatal("madvice failed: %s", name);
 880
 881                 if (sigsetjmp(sigbus_jmp, 1)) {
 882                         end = off + sigbus_addr ? sigbus_addr - ptr : 0;
 883                         fprintf(stderr, "got sigbus at offset %lld: %s\n",
 884                                         (long long)end, name);
 885                         goto got_sigbus;
 886                 }
 887
 888                 /* populate ptes */
 889                 for (i = 0; i < nr_pages ; i++) {
 890                         if (vec[i] & 1)
 891                                 (void)*(volatile int *)(ptr + i * page_size);
 892                 }
 893 got_sigbus:
 894
 895                 /* turn off harvesting reference bits */
 896                 if (madvise(ptr, len, MADV_SEQUENTIAL))
 897                         fatal("madvice failed: %s", name);
 898
 899                 if (pagemap_read(buf, (unsigned long)ptr / page_size,
 900                                         nr_pages) != nr_pages)
 901                         fatal("cannot read pagemap");
 902
 903                 munmap(ptr, len);
 904
 905                 for (i = 0; i < nr_pages; i++) {
 906                         pfn = pagemap_pfn(buf[i]);
 907                         if (!pfn)
 908                                 continue;
 909                         if (!kpageflags_read(&flags, pfn, 1))
 910                                 continue;
 911                         if (first && opt_list) {
 912                                 first = 0;
 913                                 flush_page_range();
 914                                 show_file(name, st);
 915                         }
 916                         add_page(off / page_size + i, pfn, flags, buf[i]);
 917                 }
 918         }
 919
 920         close(fd);
 921 }
 922
 923 int walk_tree(const char *name, const struct stat *st, int type, struct FTW *f)
 924 {
 925         (void)f;
 926         switch (type) {
 927         case FTW_F:
 928                 if (S_ISREG(st->st_mode))
 929                         walk_file(name, st);
 930                 break;
 931         case FTW_DNR:
 932                 fprintf(stderr, "cannot read dir: %s\n", name);
 933                 break;
 934         }
 935         return 0;
 936 }
 937
 938 static void walk_page_cache(void)
 939 {
 940         struct stat st;
 941
 942         kpageflags_fd = checked_open(PROC_KPAGEFLAGS, O_RDONLY);
 943         pagemap_fd = checked_open("/proc/self/pagemap", O_RDONLY);
 944         sigaction(SIGBUS, &sigbus_action, NULL);
 945
 946         if (stat(opt_file, &st))
 947                 fatal("stat failed: %s\n", opt_file);
 948
 949         if (S_ISREG(st.st_mode)) {
 950                 walk_file(opt_file, &st);
 951         } else if (S_ISDIR(st.st_mode)) {
 952                 /* do not follow symlinks and mountpoints */
 953                 if (nftw(opt_file, walk_tree, 64, FTW_MOUNT | FTW_PHYS) < 0)
 954                         fatal("nftw failed: %s\n", opt_file);
 955         } else
 956                 fatal("unhandled file type: %s\n", opt_file);
 957
 958         close(kpageflags_fd);
 959         close(pagemap_fd);
 960         signal(SIGBUS, SIG_DFL);
 961 }
 962
 963 static void parse_file(const char *name)
 964 {
 965         opt_file = name;
 966 }
 967
 968 static void parse_addr_range(const char *optarg)
 969 {
 970         unsigned long offset;
 971         unsigned long size;
 972         char *p;
 973
 974         p = strchr(optarg, ',');
 975         if (!p)
 976                 p = strchr(optarg, '+');
 977
 978         if (p == optarg) {
 979                 offset = 0;
 980                 size   = parse_number(p + 1);
 981         } else if (p) {
 982                 offset = parse_number(optarg);
 983                 if (p[1] == '\0')
 984                         size = ULONG_MAX;
 985                 else {
 986                         size = parse_number(p + 1);
 987                         if (*p == ',') {
 988                                 if (size < offset)
 989                                         fatal("invalid range: %lu,%lu\n",
 990                                                         offset, size);
 991                                 size -= offset;
 992                         }
 993                 }
 994         } else {
 995                 offset = parse_number(optarg);
 996                 size   = 1;
 997         }
 998
 999         add_addr_range(offset, size);
1000 }
1001
1002 static void add_bits_filter(uint64_t mask, uint64_t bits)
1003 {
1004         if (nr_bit_filters >= MAX_BIT_FILTERS)
1005                 fatal("too much bit filters\n");
1006
1007         opt_mask[nr_bit_filters] = mask;
1008         opt_bits[nr_bit_filters] = bits;
1009         nr_bit_filters++;
1010 }
1011
1012 static uint64_t parse_flag_name(const char *str, int len)
1013 {
1014         size_t i;
1015
1016         if (!*str || !len)
1017                 return 0;
1018
1019         if (len <= 8 && !strncmp(str, "compound", len))
1020                 return BITS_COMPOUND;
1021
1022         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
1023                 if (!page_flag_names[i])
1024                         continue;
1025                 if (!strncmp(str, page_flag_names[i] + 2, len))
1026                         return 1ULL << i;
1027         }
1028
1029         return parse_number(str);
1030 }
1031
1032 static uint64_t parse_flag_names(const char *str, int all)
1033 {
1034         const char *p    = str;
1035         uint64_t   flags = 0;
1036
1037         while (1) {
1038                 if (*p == ',' || *p == '=' || *p == '\0') {
1039                         if ((*str != '~') || (*str == '~' && all && *++str))
1040                                 flags |= parse_flag_name(str, p - str);
1041                         if (*p != ',')
1042                                 break;
1043                         str = p + 1;
1044                 }
1045                 p++;
1046         }
1047
1048         return flags;
1049 }
1050
1051 static void parse_bits_mask(const char *optarg)
1052 {
1053         uint64_t mask;
1054         uint64_t bits;
1055         const char *p;
1056
1057         p = strchr(optarg, '=');
1058         if (p == optarg) {
1059                 mask = KPF_ALL_BITS;
1060                 bits = parse_flag_names(p + 1, 0);
1061         } else if (p) {
1062                 mask = parse_flag_names(optarg, 0);
1063                 bits = parse_flag_names(p + 1, 0);
1064         } else if (strchr(optarg, '~')) {
1065                 mask = parse_flag_names(optarg, 1);
1066                 bits = parse_flag_names(optarg, 0);
1067         } else {
1068                 mask = parse_flag_names(optarg, 0);
1069                 bits = KPF_ALL_BITS;
1070         }
1071
1072         add_bits_filter(mask, bits);
1073 }
1074
1075 static void describe_flags(const char *optarg)
1076 {
1077         uint64_t flags = parse_flag_names(optarg, 0);
1078
1079         printf("0x%016llx\t%s\t%s\n",
1080                 (unsigned long long)flags,
1081                 page_flag_name(flags),
1082                 page_flag_longname(flags));
1083 }
1084
1085 static const struct option opts[] = {
1086         { "raw"       , 0, NULL, 'r' },
1087         { "pid"       , 1, NULL, 'p' },
1088         { "file"      , 1, NULL, 'f' },
1089         { "addr"      , 1, NULL, 'a' },
1090         { "bits"      , 1, NULL, 'b' },
1091         { "describe"  , 1, NULL, 'd' },
1092         { "list"      , 0, NULL, 'l' },
1093         { "list-each" , 0, NULL, 'L' },
1094         { "no-summary", 0, NULL, 'N' },
1095         { "hwpoison"  , 0, NULL, 'X' },
1096         { "unpoison"  , 0, NULL, 'x' },
1097         { "help"      , 0, NULL, 'h' },
1098         { NULL        , 0, NULL, 0 }
1099 };
1100
1101 int main(int argc, char *argv[])
1102 {
1103         int c;
1104
1105         page_size = getpagesize();
1106
1107         while ((c = getopt_long(argc, argv,
1108                                 "rp:f:a:b:d:lLNXxh", opts, NULL)) != -1) {
1109                 switch (c) {
1110                 case 'r':
1111                         opt_raw = 1;
1112                         break;
1113                 case 'p':
1114                         parse_pid(optarg);
1115                         break;
1116                 case 'f':
1117                         parse_file(optarg);
1118                         break;
1119                 case 'a':
1120                         parse_addr_range(optarg);
1121                         break;
1122                 case 'b':
1123                         parse_bits_mask(optarg);
1124                         break;
1125                 case 'd':
1126                         describe_flags(optarg);
1127                         exit(0);
1128                 case 'l':
1129                         opt_list = 1;
1130                         break;
1131                 case 'L':
1132                         opt_list = 2;
1133                         break;
1134                 case 'N':
1135                         opt_no_summary = 1;
1136                         break;
1137                 case 'X':
1138                         opt_hwpoison = 1;
1139                         prepare_hwpoison_fd();
1140                         break;
1141                 case 'x':
1142                         opt_unpoison = 1;
1143                         prepare_hwpoison_fd();
1144                         break;
1145                 case 'h':
1146                         usage();
1147                         exit(0);
1148                 default:
1149                         usage();
1150                         exit(1);
1151                 }
1152         }
1153
1154         if (opt_list && opt_pid)
1155                 printf("voffset\t");
1156         if (opt_list && opt_file)
1157                 printf("foffset\t");
1158         if (opt_list == 1)
1159                 printf("offset\tlen\tflags\n");
1160         if (opt_list == 2)
1161                 printf("offset\tflags\n");
1162
1163         if (opt_file)
1164                 walk_page_cache();
1165         else
1166                 walk_addr_ranges();
1167
1168         if (opt_list == 1)
1169                 flush_page_range();
1170
1171         if (opt_no_summary)
1172                 return 0;
1173
1174         if (opt_list)
1175                 printf("\n\n");
1176
1177         show_summary();
1178
1179         return 0;
1180 }