tools/vm/page-types.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * page-types: Tool for querying page flags
   4  *
   5  * Copyright (C) 2009 Intel corporation
   6  *
   7  * Authors: Wu Fengguang <fengguang.wu@intel.com>
   8  */
   9
  10 #define _FILE_OFFSET_BITS 64
  11 #define _GNU_SOURCE
  12 #include <stdio.h>
  13 #include <stdlib.h>
  14 #include <unistd.h>
  15 #include <stdint.h>
  16 #include <stdarg.h>
  17 #include <string.h>
  18 #include <getopt.h>
  19 #include <limits.h>
  20 #include <assert.h>
  21 #include <ftw.h>
  22 #include <time.h>
  23 #include <setjmp.h>
  24 #include <signal.h>
  25 #include <sys/types.h>
  26 #include <sys/errno.h>
  27 #include <sys/fcntl.h>
  28 #include <sys/mount.h>
  29 #include <sys/statfs.h>
  30 #include <sys/mman.h>
  31 #include "../../include/uapi/linux/magic.h"
  32 #include "../../include/uapi/linux/kernel-page-flags.h"
  33 #include <api/fs/fs.h>
  34
  35 #ifndef MAX_PATH
  36 # define MAX_PATH 256
  37 #endif
  38
  39 #ifndef STR
  40 # define _STR(x) #x
  41 # define STR(x) _STR(x)
  42 #endif
  43
  44 /*
  45  * pagemap kernel ABI bits
  46  */
  47
  48 #define PM_ENTRY_BYTES          8
  49 #define PM_PFRAME_BITS          55
  50 #define PM_PFRAME_MASK          ((1LL << PM_PFRAME_BITS) - 1)
  51 #define PM_PFRAME(x)            ((x) & PM_PFRAME_MASK)
  52 #define MAX_SWAPFILES_SHIFT     5
  53 #define PM_SWAP_OFFSET(x)       (((x) & PM_PFRAME_MASK) >> MAX_SWAPFILES_SHIFT)
  54 #define PM_SOFT_DIRTY           (1ULL << 55)
  55 #define PM_MMAP_EXCLUSIVE       (1ULL << 56)
  56 #define PM_FILE                 (1ULL << 61)
  57 #define PM_SWAP                 (1ULL << 62)
  58 #define PM_PRESENT              (1ULL << 63)
  59
  60 /*
  61  * kernel page flags
  62  */
  63
  64 #define KPF_BYTES               8
  65 #define PROC_KPAGEFLAGS         "/proc/kpageflags"
  66 #define PROC_KPAGECOUNT         "/proc/kpagecount"
  67 #define PROC_KPAGECGROUP        "/proc/kpagecgroup"
  68
  69 #define SYS_KERNEL_MM_PAGE_IDLE "/sys/kernel/mm/page_idle/bitmap"
  70
  71 /* [32-] kernel hacking assistances */
  72 #define KPF_RESERVED            32
  73 #define KPF_MLOCKED             33
  74 #define KPF_MAPPEDTODISK        34
  75 #define KPF_PRIVATE             35
  76 #define KPF_PRIVATE_2           36
  77 #define KPF_OWNER_PRIVATE       37
  78 #define KPF_ARCH                38
  79 #define KPF_UNCACHED            39
  80 #define KPF_SOFTDIRTY           40
  81 #define KPF_ARCH_2              41
  82
  83 /* [48-] take some arbitrary free slots for expanding overloaded flags
  84  * not part of kernel API
  85  */
  86 #define KPF_READAHEAD           48
  87 #define KPF_SLOB_FREE           49
  88 #define KPF_SLUB_FROZEN         50
  89 #define KPF_SLUB_DEBUG          51
  90 #define KPF_FILE                61
  91 #define KPF_SWAP                62
  92 #define KPF_MMAP_EXCLUSIVE      63
  93
  94 #define KPF_ALL_BITS            ((uint64_t)~0ULL)
  95 #define KPF_HACKERS_BITS        (0xffffULL << 32)
  96 #define KPF_OVERLOADED_BITS     (0xffffULL << 48)
  97 #define BIT(name)               (1ULL << KPF_##name)
  98 #define BITS_COMPOUND           (BIT(COMPOUND_HEAD) | BIT(COMPOUND_TAIL))
  99
 100 static const char * const page_flag_names[] = {
 101         [KPF_LOCKED]            = "L:locked",
 102         [KPF_ERROR]             = "E:error",
 103         [KPF_REFERENCED]        = "R:referenced",
 104         [KPF_UPTODATE]          = "U:uptodate",
 105         [KPF_DIRTY]             = "D:dirty",
 106         [KPF_LRU]               = "l:lru",
 107         [KPF_ACTIVE]            = "A:active",
 108         [KPF_SLAB]              = "S:slab",
 109         [KPF_WRITEBACK]         = "W:writeback",
 110         [KPF_RECLAIM]           = "I:reclaim",
 111         [KPF_BUDDY]             = "B:buddy",
 112
 113         [KPF_MMAP]              = "M:mmap",
 114         [KPF_ANON]              = "a:anonymous",
 115         [KPF_SWAPCACHE]         = "s:swapcache",
 116         [KPF_SWAPBACKED]        = "b:swapbacked",
 117         [KPF_COMPOUND_HEAD]     = "H:compound_head",
 118         [KPF_COMPOUND_TAIL]     = "T:compound_tail",
 119         [KPF_HUGE]              = "G:huge",
 120         [KPF_UNEVICTABLE]       = "u:unevictable",
 121         [KPF_HWPOISON]          = "X:hwpoison",
 122         [KPF_NOPAGE]            = "n:nopage",
 123         [KPF_KSM]               = "x:ksm",
 124         [KPF_THP]               = "t:thp",
 125         [KPF_OFFLINE]           = "o:offline",
 126         [KPF_PGTABLE]           = "g:pgtable",
 127         [KPF_ZERO_PAGE]         = "z:zero_page",
 128         [KPF_IDLE]              = "i:idle_page",
 129
 130         [KPF_RESERVED]          = "r:reserved",
 131         [KPF_MLOCKED]           = "m:mlocked",
 132         [KPF_MAPPEDTODISK]      = "d:mappedtodisk",
 133         [KPF_PRIVATE]           = "P:private",
 134         [KPF_PRIVATE_2]         = "p:private_2",
 135         [KPF_OWNER_PRIVATE]     = "O:owner_private",
 136         [KPF_ARCH]              = "h:arch",
 137         [KPF_UNCACHED]          = "c:uncached",
 138         [KPF_SOFTDIRTY]         = "f:softdirty",
 139         [KPF_ARCH_2]            = "H:arch_2",
 140
 141         [KPF_READAHEAD]         = "I:readahead",
 142         [KPF_SLOB_FREE]         = "P:slob_free",
 143         [KPF_SLUB_FROZEN]       = "A:slub_frozen",
 144         [KPF_SLUB_DEBUG]        = "E:slub_debug",
 145
 146         [KPF_FILE]              = "F:file",
 147         [KPF_SWAP]              = "w:swap",
 148         [KPF_MMAP_EXCLUSIVE]    = "1:mmap_exclusive",
 149 };
 150
 151
 152 /*
 153  * data structures
 154  */
 155
 156 static int              opt_raw;        /* for kernel developers */
 157 static int              opt_list;       /* list pages (in ranges) */
 158 static int              opt_mark_idle;  /* set accessed bit */
 159 static int              opt_no_summary; /* don't show summary */
 160 static pid_t            opt_pid;        /* process to walk */
 161 const char              *opt_file;      /* file or directory path */
 162 static uint64_t         opt_cgroup;     /* cgroup inode */
 163 static int              opt_list_cgroup;/* list page cgroup */
 164 static int              opt_list_mapcnt;/* list page map count */
 165 static const char       *opt_kpageflags;/* kpageflags file to parse */
 166
 167 #define MAX_ADDR_RANGES 1024
 168 static int              nr_addr_ranges;
 169 static unsigned long    opt_offset[MAX_ADDR_RANGES];
 170 static unsigned long    opt_size[MAX_ADDR_RANGES];
 171
 172 #define MAX_VMAS        10240
 173 static int              nr_vmas;
 174 static unsigned long    pg_start[MAX_VMAS];
 175 static unsigned long    pg_end[MAX_VMAS];
 176
 177 #define MAX_BIT_FILTERS 64
 178 static int              nr_bit_filters;
 179 static uint64_t         opt_mask[MAX_BIT_FILTERS];
 180 static uint64_t         opt_bits[MAX_BIT_FILTERS];
 181
 182 static int              page_size;
 183
 184 static int              pagemap_fd;
 185 static int              kpageflags_fd;
 186 static int              kpagecount_fd = -1;
 187 static int              kpagecgroup_fd = -1;
 188 static int              page_idle_fd = -1;
 189
 190 static int              opt_hwpoison;
 191 static int              opt_unpoison;
 192
 193 static const 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, const 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, opt_kpageflags, buf, index, pages);
 280 }
 281
 282 static unsigned long kpagecgroup_read(uint64_t *buf,
 283                                       unsigned long index,
 284                                       unsigned long pages)
 285 {
 286         if (kpagecgroup_fd < 0)
 287                 return pages;
 288
 289         return do_u64_read(kpagecgroup_fd, opt_kpageflags, buf, index, pages);
 290 }
 291
 292 static unsigned long kpagecount_read(uint64_t *buf,
 293                                      unsigned long index,
 294                                      unsigned long pages)
 295 {
 296         return kpagecount_fd < 0 ? pages :
 297                 do_u64_read(kpagecount_fd, PROC_KPAGECOUNT,
 298                             buf, index, pages);
 299 }
 300
 301 static unsigned long pagemap_read(uint64_t *buf,
 302                                   unsigned long index,
 303                                   unsigned long pages)
 304 {
 305         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 306 }
 307
 308 static unsigned long pagemap_pfn(uint64_t val)
 309 {
 310         unsigned long pfn;
 311
 312         if (val & PM_PRESENT)
 313                 pfn = PM_PFRAME(val);
 314         else
 315                 pfn = 0;
 316
 317         return pfn;
 318 }
 319
 320 static unsigned long pagemap_swap_offset(uint64_t val)
 321 {
 322         return val & PM_SWAP ? PM_SWAP_OFFSET(val) : 0;
 323 }
 324
 325 /*
 326  * page flag names
 327  */
 328
 329 static char *page_flag_name(uint64_t flags)
 330 {
 331         static char buf[65];
 332         int present;
 333         size_t i, j;
 334
 335         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 336                 present = (flags >> i) & 1;
 337                 if (!page_flag_names[i]) {
 338                         if (present)
 339                                 fatal("unknown flag bit %d\n", i);
 340                         continue;
 341                 }
 342                 buf[j++] = present ? page_flag_names[i][0] : '_';
 343         }
 344
 345         return buf;
 346 }
 347
 348 static char *page_flag_longname(uint64_t flags)
 349 {
 350         static char buf[1024];
 351         size_t i, n;
 352
 353         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 354                 if (!page_flag_names[i])
 355                         continue;
 356                 if ((flags >> i) & 1)
 357                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 358                                         page_flag_names[i] + 2);
 359         }
 360         if (n)
 361                 n--;
 362         buf[n] = '\0';
 363
 364         return buf;
 365 }
 366
 367
 368 /*
 369  * page list and summary
 370  */
 371
 372 static void show_page_range(unsigned long voffset, unsigned long offset,
 373                             unsigned long size, uint64_t flags,
 374                             uint64_t cgroup, uint64_t mapcnt)
 375 {
 376         static uint64_t      flags0;
 377         static uint64_t      cgroup0;
 378         static uint64_t      mapcnt0;
 379         static unsigned long voff;
 380         static unsigned long index;
 381         static unsigned long count;
 382
 383         if (flags == flags0 && cgroup == cgroup0 && mapcnt == mapcnt0 &&
 384             offset == index + count && size && voffset == voff + count) {
 385                 count += size;
 386                 return;
 387         }
 388
 389         if (count) {
 390                 if (opt_pid)
 391                         printf("%lx\t", voff);
 392                 if (opt_file)
 393                         printf("%lu\t", voff);
 394                 if (opt_list_cgroup)
 395                         printf("@%llu\t", (unsigned long long)cgroup0);
 396                 if (opt_list_mapcnt)
 397                         printf("%lu\t", mapcnt0);
 398                 printf("%lx\t%lx\t%s\n",
 399                                 index, count, page_flag_name(flags0));
 400         }
 401
 402         flags0 = flags;
 403         cgroup0 = cgroup;
 404         mapcnt0 = mapcnt;
 405         index  = offset;
 406         voff   = voffset;
 407         count  = size;
 408 }
 409
 410 static void flush_page_range(void)
 411 {
 412         show_page_range(0, 0, 0, 0, 0, 0);
 413 }
 414
 415 static void show_page(unsigned long voffset, unsigned long offset,
 416                       uint64_t flags, uint64_t cgroup, uint64_t mapcnt)
 417 {
 418         if (opt_pid)
 419                 printf("%lx\t", voffset);
 420         if (opt_file)
 421                 printf("%lu\t", voffset);
 422         if (opt_list_cgroup)
 423                 printf("@%llu\t", (unsigned long long)cgroup);
 424         if (opt_list_mapcnt)
 425                 printf("%lu\t", mapcnt);
 426
 427         printf("%lx\t%s\n", offset, page_flag_name(flags));
 428 }
 429
 430 static void show_summary(void)
 431 {
 432         size_t i;
 433
 434         printf("             flags\tpage-count       MB"
 435                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 436
 437         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 438                 if (nr_pages[i])
 439                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 440                                 (unsigned long long)page_flags[i],
 441                                 nr_pages[i],
 442                                 pages2mb(nr_pages[i]),
 443                                 page_flag_name(page_flags[i]),
 444                                 page_flag_longname(page_flags[i]));
 445         }
 446
 447         printf("             total\t%10lu %8lu\n",
 448                         total_pages, pages2mb(total_pages));
 449 }
 450
 451
 452 /*
 453  * page flag filters
 454  */
 455
 456 static int bit_mask_ok(uint64_t flags)
 457 {
 458         int i;
 459
 460         for (i = 0; i < nr_bit_filters; i++) {
 461                 if (opt_bits[i] == KPF_ALL_BITS) {
 462                         if ((flags & opt_mask[i]) == 0)
 463                                 return 0;
 464                 } else {
 465                         if ((flags & opt_mask[i]) != opt_bits[i])
 466                                 return 0;
 467                 }
 468         }
 469
 470         return 1;
 471 }
 472
 473 static uint64_t expand_overloaded_flags(uint64_t flags, uint64_t pme)
 474 {
 475         /* SLOB/SLUB overload several page flags */
 476         if (flags & BIT(SLAB)) {
 477                 if (flags & BIT(PRIVATE))
 478                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 479                 if (flags & BIT(ACTIVE))
 480                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 481                 if (flags & BIT(ERROR))
 482                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 483         }
 484
 485         /* PG_reclaim is overloaded as PG_readahead in the read path */
 486         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 487                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 488
 489         if (pme & PM_SOFT_DIRTY)
 490                 flags |= BIT(SOFTDIRTY);
 491         if (pme & PM_FILE)
 492                 flags |= BIT(FILE);
 493         if (pme & PM_SWAP)
 494                 flags |= BIT(SWAP);
 495         if (pme & PM_MMAP_EXCLUSIVE)
 496                 flags |= BIT(MMAP_EXCLUSIVE);
 497
 498         return flags;
 499 }
 500
 501 static uint64_t well_known_flags(uint64_t flags)
 502 {
 503         /* hide flags intended only for kernel hacker */
 504         flags &= ~KPF_HACKERS_BITS;
 505
 506         /* hide non-hugeTLB compound pages */
 507         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 508                 flags &= ~BITS_COMPOUND;
 509
 510         return flags;
 511 }
 512
 513 static uint64_t kpageflags_flags(uint64_t flags, uint64_t pme)
 514 {
 515         if (opt_raw)
 516                 flags = expand_overloaded_flags(flags, pme);
 517         else
 518                 flags = well_known_flags(flags);
 519
 520         return flags;
 521 }
 522
 523 /*
 524  * page actions
 525  */
 526
 527 static void prepare_hwpoison_fd(void)
 528 {
 529         char buf[MAX_PATH + 1];
 530
 531         hwpoison_debug_fs = debugfs__mount();
 532         if (!hwpoison_debug_fs) {
 533                 perror("mount debugfs");
 534                 exit(EXIT_FAILURE);
 535         }
 536
 537         if (opt_hwpoison && !hwpoison_inject_fd) {
 538                 snprintf(buf, MAX_PATH, "%s/hwpoison/corrupt-pfn",
 539                         hwpoison_debug_fs);
 540                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 541         }
 542
 543         if (opt_unpoison && !hwpoison_forget_fd) {
 544                 snprintf(buf, MAX_PATH, "%s/hwpoison/unpoison-pfn",
 545                         hwpoison_debug_fs);
 546                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 547         }
 548 }
 549
 550 static int hwpoison_page(unsigned long offset)
 551 {
 552         char buf[100];
 553         int len;
 554
 555         len = sprintf(buf, "0x%lx\n", offset);
 556         len = write(hwpoison_inject_fd, buf, len);
 557         if (len < 0) {
 558                 perror("hwpoison inject");
 559                 return len;
 560         }
 561         return 0;
 562 }
 563
 564 static int unpoison_page(unsigned long offset)
 565 {
 566         char buf[100];
 567         int len;
 568
 569         len = sprintf(buf, "0x%lx\n", offset);
 570         len = write(hwpoison_forget_fd, buf, len);
 571         if (len < 0) {
 572                 perror("hwpoison forget");
 573                 return len;
 574         }
 575         return 0;
 576 }
 577
 578 static int mark_page_idle(unsigned long offset)
 579 {
 580         static unsigned long off;
 581         static uint64_t buf;
 582         int len;
 583
 584         if ((offset / 64 == off / 64) || buf == 0) {
 585                 buf |= 1UL << (offset % 64);
 586                 off = offset;
 587                 return 0;
 588         }
 589
 590         len = pwrite(page_idle_fd, &buf, 8, 8 * (off / 64));
 591         if (len < 0) {
 592                 perror("mark page idle");
 593                 return len;
 594         }
 595
 596         buf = 1UL << (offset % 64);
 597         off = offset;
 598
 599         return 0;
 600 }
 601
 602 /*
 603  * page frame walker
 604  */
 605
 606 static size_t hash_slot(uint64_t flags)
 607 {
 608         size_t k = HASH_KEY(flags);
 609         size_t i;
 610
 611         /* Explicitly reserve slot 0 for flags 0: the following logic
 612          * cannot distinguish an unoccupied slot from slot (flags==0).
 613          */
 614         if (flags == 0)
 615                 return 0;
 616
 617         /* search through the remaining (HASH_SIZE-1) slots */
 618         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 619                 if (!k || k >= ARRAY_SIZE(page_flags))
 620                         k = 1;
 621                 if (page_flags[k] == 0) {
 622                         page_flags[k] = flags;
 623                         return k;
 624                 }
 625                 if (page_flags[k] == flags)
 626                         return k;
 627         }
 628
 629         fatal("hash table full: bump up HASH_SHIFT?\n");
 630         exit(EXIT_FAILURE);
 631 }
 632
 633 static void add_page(unsigned long voffset, unsigned long offset,
 634                      uint64_t flags, uint64_t cgroup, uint64_t mapcnt,
 635                      uint64_t pme)
 636 {
 637         flags = kpageflags_flags(flags, pme);
 638
 639         if (!bit_mask_ok(flags))
 640                 return;
 641
 642         if (opt_cgroup && cgroup != (uint64_t)opt_cgroup)
 643                 return;
 644
 645         if (opt_hwpoison)
 646                 hwpoison_page(offset);
 647         if (opt_unpoison)
 648                 unpoison_page(offset);
 649
 650         if (opt_mark_idle)
 651                 mark_page_idle(offset);
 652
 653         if (opt_list == 1)
 654                 show_page_range(voffset, offset, 1, flags, cgroup, mapcnt);
 655         else if (opt_list == 2)
 656                 show_page(voffset, offset, flags, cgroup, mapcnt);
 657
 658         nr_pages[hash_slot(flags)]++;
 659         total_pages++;
 660 }
 661
 662 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 663 static void walk_pfn(unsigned long voffset,
 664                      unsigned long index,
 665                      unsigned long count,
 666                      uint64_t pme)
 667 {
 668         uint64_t buf[KPAGEFLAGS_BATCH];
 669         uint64_t cgi[KPAGEFLAGS_BATCH];
 670         uint64_t cnt[KPAGEFLAGS_BATCH];
 671         unsigned long batch;
 672         unsigned long pages;
 673         unsigned long i;
 674
 675         /*
 676          * kpagecgroup_read() reads only if kpagecgroup were opened, but
 677          * /proc/kpagecgroup might even not exist, so it's better to fill
 678          * them with zeros here.
 679          */
 680         if (count == 1)
 681                 cgi[0] = 0;
 682         else
 683                 memset(cgi, 0, sizeof cgi);
 684
 685         while (count) {
 686                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 687                 pages = kpageflags_read(buf, index, batch);
 688                 if (pages == 0)
 689                         break;
 690
 691                 if (kpagecgroup_read(cgi, index, pages) != pages)
 692                         fatal("kpagecgroup returned fewer pages than expected");
 693
 694                 if (kpagecount_read(cnt, index, pages) != pages)
 695                         fatal("kpagecount returned fewer pages than expected");
 696
 697                 for (i = 0; i < pages; i++)
 698                         add_page(voffset + i, index + i,
 699                                  buf[i], cgi[i], cnt[i], pme);
 700
 701                 index += pages;
 702                 count -= pages;
 703         }
 704 }
 705
 706 static void walk_swap(unsigned long voffset, uint64_t pme)
 707 {
 708         uint64_t flags = kpageflags_flags(0, pme);
 709
 710         if (!bit_mask_ok(flags))
 711                 return;
 712
 713         if (opt_cgroup)
 714                 return;
 715
 716         if (opt_list == 1)
 717                 show_page_range(voffset, pagemap_swap_offset(pme),
 718                                 1, flags, 0, 0);
 719         else if (opt_list == 2)
 720                 show_page(voffset, pagemap_swap_offset(pme), flags, 0, 0);
 721
 722         nr_pages[hash_slot(flags)]++;
 723         total_pages++;
 724 }
 725
 726 #define PAGEMAP_BATCH   (64 << 10)
 727 static void walk_vma(unsigned long index, unsigned long count)
 728 {
 729         uint64_t buf[PAGEMAP_BATCH];
 730         unsigned long batch;
 731         unsigned long pages;
 732         unsigned long pfn;
 733         unsigned long i;
 734
 735         while (count) {
 736                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 737                 pages = pagemap_read(buf, index, batch);
 738                 if (pages == 0)
 739                         break;
 740
 741                 for (i = 0; i < pages; i++) {
 742                         pfn = pagemap_pfn(buf[i]);
 743                         if (pfn)
 744                                 walk_pfn(index + i, pfn, 1, buf[i]);
 745                         if (buf[i] & PM_SWAP)
 746                                 walk_swap(index + i, buf[i]);
 747                 }
 748
 749                 index += pages;
 750                 count -= pages;
 751         }
 752 }
 753
 754 static void walk_task(unsigned long index, unsigned long count)
 755 {
 756         const unsigned long end = index + count;
 757         unsigned long start;
 758         int i = 0;
 759
 760         while (index < end) {
 761
 762                 while (pg_end[i] <= index)
 763                         if (++i >= nr_vmas)
 764                                 return;
 765                 if (pg_start[i] >= end)
 766                         return;
 767
 768                 start = max_t(unsigned long, pg_start[i], index);
 769                 index = min_t(unsigned long, pg_end[i], end);
 770
 771                 assert(start < index);
 772                 walk_vma(start, index - start);
 773         }
 774 }
 775
 776 static void add_addr_range(unsigned long offset, unsigned long size)
 777 {
 778         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 779                 fatal("too many addr ranges\n");
 780
 781         opt_offset[nr_addr_ranges] = offset;
 782         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 783         nr_addr_ranges++;
 784 }
 785
 786 static void walk_addr_ranges(void)
 787 {
 788         int i;
 789
 790         kpageflags_fd = checked_open(opt_kpageflags, O_RDONLY);
 791
 792         if (!nr_addr_ranges)
 793                 add_addr_range(0, ULONG_MAX);
 794
 795         for (i = 0; i < nr_addr_ranges; i++)
 796                 if (!opt_pid)
 797                         walk_pfn(opt_offset[i], opt_offset[i], opt_size[i], 0);
 798                 else
 799                         walk_task(opt_offset[i], opt_size[i]);
 800
 801         if (opt_mark_idle)
 802                 mark_page_idle(0);
 803
 804         close(kpageflags_fd);
 805 }
 806
 807
 808 /*
 809  * user interface
 810  */
 811
 812 static const char *page_flag_type(uint64_t flag)
 813 {
 814         if (flag & KPF_HACKERS_BITS)
 815                 return "(r)";
 816         if (flag & KPF_OVERLOADED_BITS)
 817                 return "(o)";
 818         return "   ";
 819 }
 820
 821 static void usage(void)
 822 {
 823         size_t i, j;
 824
 825         printf(
 826 "page-types [options]\n"
 827 "            -r|--raw                   Raw mode, for kernel developers\n"
 828 "            -d|--describe flags        Describe flags\n"
 829 "            -a|--addr    addr-spec     Walk a range of pages\n"
 830 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 831 "            -c|--cgroup  path|@inode   Walk pages within memory cgroup\n"
 832 "            -p|--pid     pid           Walk process address space\n"
 833 "            -f|--file    filename      Walk file address space\n"
 834 "            -i|--mark-idle             Mark pages idle\n"
 835 "            -l|--list                  Show page details in ranges\n"
 836 "            -L|--list-each             Show page details one by one\n"
 837 "            -C|--list-cgroup           Show cgroup inode for pages\n"
 838 "            -M|--list-mapcnt           Show page map count\n"
 839 "            -N|--no-summary            Don't show summary info\n"
 840 "            -X|--hwpoison              hwpoison pages\n"
 841 "            -x|--unpoison              unpoison pages\n"
 842 "            -F|--kpageflags filename   kpageflags file to parse\n"
 843 "            -h|--help                  Show this usage message\n"
 844 "flags:\n"
 845 "            0x10                       bitfield format, e.g.\n"
 846 "            anon                       bit-name, e.g.\n"
 847 "            0x10,anon                  comma-separated list, e.g.\n"
 848 "addr-spec:\n"
 849 "            N                          one page at offset N (unit: pages)\n"
 850 "            N+M                        pages range from N to N+M-1\n"
 851 "            N,M                        pages range from N to M-1\n"
 852 "            N,                         pages range from N to end\n"
 853 "            ,M                         pages range from 0 to M-1\n"
 854 "bits-spec:\n"
 855 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 856 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 857 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 858 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 859 "bit-names:\n"
 860         );
 861
 862         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 863                 if (!page_flag_names[i])
 864                         continue;
 865                 printf("%16s%s", page_flag_names[i] + 2,
 866                                  page_flag_type(1ULL << i));
 867                 if (++j > 3) {
 868                         j = 0;
 869                         putchar('\n');
 870                 }
 871         }
 872         printf("\n                                   "
 873                 "(r) raw mode bits  (o) overloaded bits\n");
 874 }
 875
 876 static unsigned long long parse_number(const char *str)
 877 {
 878         unsigned long long n;
 879
 880         n = strtoll(str, NULL, 0);
 881
 882         if (n == 0 && str[0] != '0')
 883                 fatal("invalid name or number: %s\n", str);
 884
 885         return n;
 886 }
 887
 888 static void parse_pid(const char *str)
 889 {
 890         FILE *file;
 891         char buf[5000];
 892
 893         opt_pid = parse_number(str);
 894
 895         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 896         pagemap_fd = checked_open(buf, O_RDONLY);
 897
 898         sprintf(buf, "/proc/%d/maps", opt_pid);
 899         file = fopen(buf, "r");
 900         if (!file) {
 901                 perror(buf);
 902                 exit(EXIT_FAILURE);
 903         }
 904
 905         while (fgets(buf, sizeof(buf), file) != NULL) {
 906                 unsigned long vm_start;
 907                 unsigned long vm_end;
 908                 unsigned long long pgoff;
 909                 int major, minor;
 910                 char r, w, x, s;
 911                 unsigned long ino;
 912                 int n;
 913
 914                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 915                            &vm_start,
 916                            &vm_end,
 917                            &r, &w, &x, &s,
 918                            &pgoff,
 919                            &major, &minor,
 920                            &ino);
 921                 if (n < 10) {
 922                         fprintf(stderr, "unexpected line: %s\n", buf);
 923                         continue;
 924                 }
 925                 pg_start[nr_vmas] = vm_start / page_size;
 926                 pg_end[nr_vmas] = vm_end / page_size;
 927                 if (++nr_vmas >= MAX_VMAS) {
 928                         fprintf(stderr, "too many VMAs\n");
 929                         break;
 930                 }
 931         }
 932         fclose(file);
 933 }
 934
 935 static void show_file(const char *name, const struct stat *st)
 936 {
 937         unsigned long long size = st->st_size;
 938         char atime[64], mtime[64];
 939         long now = time(NULL);
 940
 941         printf("%s\tInode: %u\tSize: %llu (%llu pages)\n",
 942                         name, (unsigned)st->st_ino,
 943                         size, (size + page_size - 1) / page_size);
 944
 945         strftime(atime, sizeof(atime), "%c", localtime(&st->st_atime));
 946         strftime(mtime, sizeof(mtime), "%c", localtime(&st->st_mtime));
 947
 948         printf("Modify: %s (%ld seconds ago)\nAccess: %s (%ld seconds ago)\n",
 949                         mtime, now - st->st_mtime,
 950                         atime, now - st->st_atime);
 951 }
 952
 953 static sigjmp_buf sigbus_jmp;
 954
 955 static void * volatile sigbus_addr;
 956
 957 static void sigbus_handler(int sig, siginfo_t *info, void *ucontex)
 958 {
 959         (void)sig;
 960         (void)ucontex;
 961         sigbus_addr = info ? info->si_addr : NULL;
 962         siglongjmp(sigbus_jmp, 1);
 963 }
 964
 965 static struct sigaction sigbus_action = {
 966         .sa_sigaction = sigbus_handler,
 967         .sa_flags = SA_SIGINFO,
 968 };
 969
 970 static void walk_file(const char *name, const struct stat *st)
 971 {
 972         uint8_t vec[PAGEMAP_BATCH];
 973         uint64_t buf[PAGEMAP_BATCH], flags;
 974         uint64_t cgroup = 0;
 975         uint64_t mapcnt = 0;
 976         unsigned long nr_pages, pfn, i;
 977         off_t off, end = st->st_size;
 978         int fd;
 979         ssize_t len;
 980         void *ptr;
 981         int first = 1;
 982
 983         fd = checked_open(name, O_RDONLY|O_NOATIME|O_NOFOLLOW);
 984
 985         for (off = 0; off < end; off += len) {
 986                 nr_pages = (end - off + page_size - 1) / page_size;
 987                 if (nr_pages > PAGEMAP_BATCH)
 988                         nr_pages = PAGEMAP_BATCH;
 989                 len = nr_pages * page_size;
 990
 991                 ptr = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, off);
 992                 if (ptr == MAP_FAILED)
 993                         fatal("mmap failed: %s", name);
 994
 995                 /* determine cached pages */
 996                 if (mincore(ptr, len, vec))
 997                         fatal("mincore failed: %s", name);
 998
 999                 /* turn off readahead */
1000                 if (madvise(ptr, len, MADV_RANDOM))
1001                         fatal("madvice failed: %s", name);
1002
1003                 if (sigsetjmp(sigbus_jmp, 1)) {
1004                         end = off + sigbus_addr ? sigbus_addr - ptr : 0;
1005                         fprintf(stderr, "got sigbus at offset %lld: %s\n",
1006                                         (long long)end, name);
1007                         goto got_sigbus;
1008                 }
1009
1010                 /* populate ptes */
1011                 for (i = 0; i < nr_pages ; i++) {
1012                         if (vec[i] & 1)
1013                                 (void)*(volatile int *)(ptr + i * page_size);
1014                 }
1015 got_sigbus:
1016
1017                 /* turn off harvesting reference bits */
1018                 if (madvise(ptr, len, MADV_SEQUENTIAL))
1019                         fatal("madvice failed: %s", name);
1020
1021                 if (pagemap_read(buf, (unsigned long)ptr / page_size,
1022                                         nr_pages) != nr_pages)
1023                         fatal("cannot read pagemap");
1024
1025                 munmap(ptr, len);
1026
1027                 for (i = 0; i < nr_pages; i++) {
1028                         pfn = pagemap_pfn(buf[i]);
1029                         if (!pfn)
1030                                 continue;
1031                         if (!kpageflags_read(&flags, pfn, 1))
1032                                 continue;
1033                         if (!kpagecgroup_read(&cgroup, pfn, 1))
1034                                 fatal("kpagecgroup_read failed");
1035                         if (!kpagecount_read(&mapcnt, pfn, 1))
1036                                 fatal("kpagecount_read failed");
1037                         if (first && opt_list) {
1038                                 first = 0;
1039                                 flush_page_range();
1040                                 show_file(name, st);
1041                         }
1042                         add_page(off / page_size + i, pfn,
1043                                  flags, cgroup, mapcnt, buf[i]);
1044                 }
1045         }
1046
1047         close(fd);
1048 }
1049
1050 int walk_tree(const char *name, const struct stat *st, int type, struct FTW *f)
1051 {
1052         (void)f;
1053         switch (type) {
1054         case FTW_F:
1055                 if (S_ISREG(st->st_mode))
1056                         walk_file(name, st);
1057                 break;
1058         case FTW_DNR:
1059                 fprintf(stderr, "cannot read dir: %s\n", name);
1060                 break;
1061         }
1062         return 0;
1063 }
1064
1065 static void walk_page_cache(void)
1066 {
1067         struct stat st;
1068
1069         kpageflags_fd = checked_open(opt_kpageflags, O_RDONLY);
1070         pagemap_fd = checked_open("/proc/self/pagemap", O_RDONLY);
1071         sigaction(SIGBUS, &sigbus_action, NULL);
1072
1073         if (stat(opt_file, &st))
1074                 fatal("stat failed: %s\n", opt_file);
1075
1076         if (S_ISREG(st.st_mode)) {
1077                 walk_file(opt_file, &st);
1078         } else if (S_ISDIR(st.st_mode)) {
1079                 /* do not follow symlinks and mountpoints */
1080                 if (nftw(opt_file, walk_tree, 64, FTW_MOUNT | FTW_PHYS) < 0)
1081                         fatal("nftw failed: %s\n", opt_file);
1082         } else
1083                 fatal("unhandled file type: %s\n", opt_file);
1084
1085         close(kpageflags_fd);
1086         close(pagemap_fd);
1087         signal(SIGBUS, SIG_DFL);
1088 }
1089
1090 static void parse_file(const char *name)
1091 {
1092         opt_file = name;
1093 }
1094
1095 static void parse_cgroup(const char *path)
1096 {
1097         if (path[0] == '@') {
1098                 opt_cgroup = parse_number(path + 1);
1099                 return;
1100         }
1101
1102         struct stat st;
1103
1104         if (stat(path, &st))
1105                 fatal("stat failed: %s: %m\n", path);
1106
1107         if (!S_ISDIR(st.st_mode))
1108                 fatal("cgroup supposed to be a directory: %s\n", path);
1109
1110         opt_cgroup = st.st_ino;
1111 }
1112
1113 static void parse_addr_range(const char *optarg)
1114 {
1115         unsigned long offset;
1116         unsigned long size;
1117         char *p;
1118
1119         p = strchr(optarg, ',');
1120         if (!p)
1121                 p = strchr(optarg, '+');
1122
1123         if (p == optarg) {
1124                 offset = 0;
1125                 size   = parse_number(p + 1);
1126         } else if (p) {
1127                 offset = parse_number(optarg);
1128                 if (p[1] == '\0')
1129                         size = ULONG_MAX;
1130                 else {
1131                         size = parse_number(p + 1);
1132                         if (*p == ',') {
1133                                 if (size < offset)
1134                                         fatal("invalid range: %lu,%lu\n",
1135                                                         offset, size);
1136                                 size -= offset;
1137                         }
1138                 }
1139         } else {
1140                 offset = parse_number(optarg);
1141                 size   = 1;
1142         }
1143
1144         add_addr_range(offset, size);
1145 }
1146
1147 static void add_bits_filter(uint64_t mask, uint64_t bits)
1148 {
1149         if (nr_bit_filters >= MAX_BIT_FILTERS)
1150                 fatal("too much bit filters\n");
1151
1152         opt_mask[nr_bit_filters] = mask;
1153         opt_bits[nr_bit_filters] = bits;
1154         nr_bit_filters++;
1155 }
1156
1157 static uint64_t parse_flag_name(const char *str, int len)
1158 {
1159         size_t i;
1160
1161         if (!*str || !len)
1162                 return 0;
1163
1164         if (len <= 8 && !strncmp(str, "compound", len))
1165                 return BITS_COMPOUND;
1166
1167         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
1168                 if (!page_flag_names[i])
1169                         continue;
1170                 if (!strncmp(str, page_flag_names[i] + 2, len))
1171                         return 1ULL << i;
1172         }
1173
1174         return parse_number(str);
1175 }
1176
1177 static uint64_t parse_flag_names(const char *str, int all)
1178 {
1179         const char *p    = str;
1180         uint64_t   flags = 0;
1181
1182         while (1) {
1183                 if (*p == ',' || *p == '=' || *p == '\0') {
1184                         if ((*str != '~') || (*str == '~' && all && *++str))
1185                                 flags |= parse_flag_name(str, p - str);
1186                         if (*p != ',')
1187                                 break;
1188                         str = p + 1;
1189                 }
1190                 p++;
1191         }
1192
1193         return flags;
1194 }
1195
1196 static void parse_bits_mask(const char *optarg)
1197 {
1198         uint64_t mask;
1199         uint64_t bits;
1200         const char *p;
1201
1202         p = strchr(optarg, '=');
1203         if (p == optarg) {
1204                 mask = KPF_ALL_BITS;
1205                 bits = parse_flag_names(p + 1, 0);
1206         } else if (p) {
1207                 mask = parse_flag_names(optarg, 0);
1208                 bits = parse_flag_names(p + 1, 0);
1209         } else if (strchr(optarg, '~')) {
1210                 mask = parse_flag_names(optarg, 1);
1211                 bits = parse_flag_names(optarg, 0);
1212         } else {
1213                 mask = parse_flag_names(optarg, 0);
1214                 bits = KPF_ALL_BITS;
1215         }
1216
1217         add_bits_filter(mask, bits);
1218 }
1219
1220 static void parse_kpageflags(const char *name)
1221 {
1222         opt_kpageflags = name;
1223 }
1224
1225 static void describe_flags(const char *optarg)
1226 {
1227         uint64_t flags = parse_flag_names(optarg, 0);
1228
1229         printf("0x%016llx\t%s\t%s\n",
1230                 (unsigned long long)flags,
1231                 page_flag_name(flags),
1232                 page_flag_longname(flags));
1233 }
1234
1235 static const struct option opts[] = {
1236         { "raw"       , 0, NULL, 'r' },
1237         { "pid"       , 1, NULL, 'p' },
1238         { "file"      , 1, NULL, 'f' },
1239         { "addr"      , 1, NULL, 'a' },
1240         { "bits"      , 1, NULL, 'b' },
1241         { "cgroup"    , 1, NULL, 'c' },
1242         { "describe"  , 1, NULL, 'd' },
1243         { "mark-idle" , 0, NULL, 'i' },
1244         { "list"      , 0, NULL, 'l' },
1245         { "list-each" , 0, NULL, 'L' },
1246         { "list-cgroup", 0, NULL, 'C' },
1247         { "list-mapcnt", 0, NULL, 'M' },
1248         { "no-summary", 0, NULL, 'N' },
1249         { "hwpoison"  , 0, NULL, 'X' },
1250         { "unpoison"  , 0, NULL, 'x' },
1251         { "kpageflags", 0, NULL, 'F' },
1252         { "help"      , 0, NULL, 'h' },
1253         { NULL        , 0, NULL, 0 }
1254 };
1255
1256 int main(int argc, char *argv[])
1257 {
1258         int c;
1259
1260         page_size = getpagesize();
1261
1262         while ((c = getopt_long(argc, argv,
1263                                 "rp:f:a:b:d:c:CilLMNXxF:h",
1264                                 opts, NULL)) != -1) {
1265                 switch (c) {
1266                 case 'r':
1267                         opt_raw = 1;
1268                         break;
1269                 case 'p':
1270                         parse_pid(optarg);
1271                         break;
1272                 case 'f':
1273                         parse_file(optarg);
1274                         break;
1275                 case 'a':
1276                         parse_addr_range(optarg);
1277                         break;
1278                 case 'b':
1279                         parse_bits_mask(optarg);
1280                         break;
1281                 case 'c':
1282                         parse_cgroup(optarg);
1283                         break;
1284                 case 'C':
1285                         opt_list_cgroup = 1;
1286                         break;
1287                 case 'd':
1288                         describe_flags(optarg);
1289                         exit(0);
1290                 case 'i':
1291                         opt_mark_idle = 1;
1292                         break;
1293                 case 'l':
1294                         opt_list = 1;
1295                         break;
1296                 case 'L':
1297                         opt_list = 2;
1298                         break;
1299                 case 'M':
1300                         opt_list_mapcnt = 1;
1301                         break;
1302                 case 'N':
1303                         opt_no_summary = 1;
1304                         break;
1305                 case 'X':
1306                         opt_hwpoison = 1;
1307                         prepare_hwpoison_fd();
1308                         break;
1309                 case 'x':
1310                         opt_unpoison = 1;
1311                         prepare_hwpoison_fd();
1312                         break;
1313                 case 'F':
1314                         parse_kpageflags(optarg);
1315                         break;
1316                 case 'h':
1317                         usage();
1318                         exit(0);
1319                 default:
1320                         usage();
1321                         exit(1);
1322                 }
1323         }
1324
1325         if (!opt_kpageflags)
1326                 opt_kpageflags = PROC_KPAGEFLAGS;
1327
1328         if (opt_cgroup || opt_list_cgroup)
1329                 kpagecgroup_fd = checked_open(PROC_KPAGECGROUP, O_RDONLY);
1330
1331         if (opt_list && opt_list_mapcnt)
1332                 kpagecount_fd = checked_open(PROC_KPAGECOUNT, O_RDONLY);
1333
1334         if (opt_mark_idle && opt_file)
1335                 page_idle_fd = checked_open(SYS_KERNEL_MM_PAGE_IDLE, O_RDWR);
1336
1337         if (opt_list && opt_pid)
1338                 printf("voffset\t");
1339         if (opt_list && opt_file)
1340                 printf("foffset\t");
1341         if (opt_list && opt_list_cgroup)
1342                 printf("cgroup\t");
1343         if (opt_list && opt_list_mapcnt)
1344                 printf("map-cnt\t");
1345
1346         if (opt_list == 1)
1347                 printf("offset\tlen\tflags\n");
1348         if (opt_list == 2)
1349                 printf("offset\tflags\n");
1350
1351         if (opt_file)
1352                 walk_page_cache();
1353         else
1354                 walk_addr_ranges();
1355
1356         if (opt_list == 1)
1357                 flush_page_range();
1358
1359         if (opt_no_summary)
1360                 return 0;
1361
1362         if (opt_list)
1363                 printf("\n\n");
1364
1365         show_summary();
1366
1367         if (opt_list_mapcnt)
1368                 close(kpagecount_fd);
1369
1370         if (page_idle_fd >= 0)
1371                 close(page_idle_fd);
1372
1373         return 0;
1374 }