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