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