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 MAX_SWAPFILES_SHIFT     5
  65 #define PM_SWAP_OFFSET(x)       (((x) & PM_PFRAME_MASK) >> MAX_SWAPFILES_SHIFT)
  66 #define PM_SOFT_DIRTY           (1ULL << 55)
  67 #define PM_MMAP_EXCLUSIVE       (1ULL << 56)
  68 #define PM_FILE                 (1ULL << 61)
  69 #define PM_SWAP                 (1ULL << 62)
  70 #define PM_PRESENT              (1ULL << 63)
  71
  72 /*
  73  * kernel page flags
  74  */
  75
  76 #define KPF_BYTES               8
  77 #define PROC_KPAGEFLAGS         "/proc/kpageflags"
  78 #define PROC_KPAGECGROUP        "/proc/kpagecgroup"
  79
  80 /* [32-] kernel hacking assistances */
  81 #define KPF_RESERVED            32
  82 #define KPF_MLOCKED             33
  83 #define KPF_MAPPEDTODISK        34
  84 #define KPF_PRIVATE             35
  85 #define KPF_PRIVATE_2           36
  86 #define KPF_OWNER_PRIVATE       37
  87 #define KPF_ARCH                38
  88 #define KPF_UNCACHED            39
  89 #define KPF_SOFTDIRTY           40
  90
  91 /* [48-] take some arbitrary free slots for expanding overloaded flags
  92  * not part of kernel API
  93  */
  94 #define KPF_READAHEAD           48
  95 #define KPF_SLOB_FREE           49
  96 #define KPF_SLUB_FROZEN         50
  97 #define KPF_SLUB_DEBUG          51
  98 #define KPF_FILE                61
  99 #define KPF_SWAP                62
 100 #define KPF_MMAP_EXCLUSIVE      63
 101
 102 #define KPF_ALL_BITS            ((uint64_t)~0ULL)
 103 #define KPF_HACKERS_BITS        (0xffffULL << 32)
 104 #define KPF_OVERLOADED_BITS     (0xffffULL << 48)
 105 #define BIT(name)               (1ULL << KPF_##name)
 106 #define BITS_COMPOUND           (BIT(COMPOUND_HEAD) | BIT(COMPOUND_TAIL))
 107
 108 static const char * const page_flag_names[] = {
 109         [KPF_LOCKED]            = "L:locked",
 110         [KPF_ERROR]             = "E:error",
 111         [KPF_REFERENCED]        = "R:referenced",
 112         [KPF_UPTODATE]          = "U:uptodate",
 113         [KPF_DIRTY]             = "D:dirty",
 114         [KPF_LRU]               = "l:lru",
 115         [KPF_ACTIVE]            = "A:active",
 116         [KPF_SLAB]              = "S:slab",
 117         [KPF_WRITEBACK]         = "W:writeback",
 118         [KPF_RECLAIM]           = "I:reclaim",
 119         [KPF_BUDDY]             = "B:buddy",
 120
 121         [KPF_MMAP]              = "M:mmap",
 122         [KPF_ANON]              = "a:anonymous",
 123         [KPF_SWAPCACHE]         = "s:swapcache",
 124         [KPF_SWAPBACKED]        = "b:swapbacked",
 125         [KPF_COMPOUND_HEAD]     = "H:compound_head",
 126         [KPF_COMPOUND_TAIL]     = "T:compound_tail",
 127         [KPF_HUGE]              = "G:huge",
 128         [KPF_UNEVICTABLE]       = "u:unevictable",
 129         [KPF_HWPOISON]          = "X:hwpoison",
 130         [KPF_NOPAGE]            = "n:nopage",
 131         [KPF_KSM]               = "x:ksm",
 132         [KPF_THP]               = "t:thp",
 133         [KPF_BALLOON]           = "o:balloon",
 134         [KPF_ZERO_PAGE]         = "z:zero_page",
 135         [KPF_IDLE]              = "i:idle_page",
 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         [KPF_FILE]              = "F:file",
 153         [KPF_SWAP]              = "w:swap",
 154         [KPF_MMAP_EXCLUSIVE]    = "1:mmap_exclusive",
 155 };
 156
 157
 158 static const char * const debugfs_known_mountpoints[] = {
 159         "/sys/kernel/debug",
 160         "/debug",
 161         0,
 162 };
 163
 164 /*
 165  * data structures
 166  */
 167
 168 static int              opt_raw;        /* for kernel developers */
 169 static int              opt_list;       /* list pages (in ranges) */
 170 static int              opt_no_summary; /* don't show summary */
 171 static pid_t            opt_pid;        /* process to walk */
 172 const char              *opt_file;      /* file or directory path */
 173 static uint64_t         opt_cgroup;     /* cgroup inode */
 174 static int              opt_list_cgroup;/* list page cgroup */
 175 static const char       *opt_kpageflags;/* kpageflags file to parse */
 176
 177 #define MAX_ADDR_RANGES 1024
 178 static int              nr_addr_ranges;
 179 static unsigned long    opt_offset[MAX_ADDR_RANGES];
 180 static unsigned long    opt_size[MAX_ADDR_RANGES];
 181
 182 #define MAX_VMAS        10240
 183 static int              nr_vmas;
 184 static unsigned long    pg_start[MAX_VMAS];
 185 static unsigned long    pg_end[MAX_VMAS];
 186
 187 #define MAX_BIT_FILTERS 64
 188 static int              nr_bit_filters;
 189 static uint64_t         opt_mask[MAX_BIT_FILTERS];
 190 static uint64_t         opt_bits[MAX_BIT_FILTERS];
 191
 192 static int              page_size;
 193
 194 static int              pagemap_fd;
 195 static int              kpageflags_fd;
 196 static int              kpagecgroup_fd = -1;
 197
 198 static int              opt_hwpoison;
 199 static int              opt_unpoison;
 200
 201 static const char       *hwpoison_debug_fs;
 202 static int              hwpoison_inject_fd;
 203 static int              hwpoison_forget_fd;
 204
 205 #define HASH_SHIFT      13
 206 #define HASH_SIZE       (1 << HASH_SHIFT)
 207 #define HASH_MASK       (HASH_SIZE - 1)
 208 #define HASH_KEY(flags) (flags & HASH_MASK)
 209
 210 static unsigned long    total_pages;
 211 static unsigned long    nr_pages[HASH_SIZE];
 212 static uint64_t         page_flags[HASH_SIZE];
 213
 214
 215 /*
 216  * helper functions
 217  */
 218
 219 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 220
 221 #define min_t(type, x, y) ({                    \
 222         type __min1 = (x);                      \
 223         type __min2 = (y);                      \
 224         __min1 < __min2 ? __min1 : __min2; })
 225
 226 #define max_t(type, x, y) ({                    \
 227         type __max1 = (x);                      \
 228         type __max2 = (y);                      \
 229         __max1 > __max2 ? __max1 : __max2; })
 230
 231 static unsigned long pages2mb(unsigned long pages)
 232 {
 233         return (pages * page_size) >> 20;
 234 }
 235
 236 static void fatal(const char *x, ...)
 237 {
 238         va_list ap;
 239
 240         va_start(ap, x);
 241         vfprintf(stderr, x, ap);
 242         va_end(ap);
 243         exit(EXIT_FAILURE);
 244 }
 245
 246 static int checked_open(const char *pathname, int flags)
 247 {
 248         int fd = open(pathname, flags);
 249
 250         if (fd < 0) {
 251                 perror(pathname);
 252                 exit(EXIT_FAILURE);
 253         }
 254
 255         return fd;
 256 }
 257
 258 /*
 259  * pagemap/kpageflags routines
 260  */
 261
 262 static unsigned long do_u64_read(int fd, const char *name,
 263                                  uint64_t *buf,
 264                                  unsigned long index,
 265                                  unsigned long count)
 266 {
 267         long bytes;
 268
 269         if (index > ULONG_MAX / 8)
 270                 fatal("index overflow: %lu\n", index);
 271
 272         bytes = pread(fd, buf, count * 8, (off_t)index * 8);
 273         if (bytes < 0) {
 274                 perror(name);
 275                 exit(EXIT_FAILURE);
 276         }
 277         if (bytes % 8)
 278                 fatal("partial read: %lu bytes\n", bytes);
 279
 280         return bytes / 8;
 281 }
 282
 283 static unsigned long kpageflags_read(uint64_t *buf,
 284                                      unsigned long index,
 285                                      unsigned long pages)
 286 {
 287         return do_u64_read(kpageflags_fd, opt_kpageflags, buf, index, pages);
 288 }
 289
 290 static unsigned long kpagecgroup_read(uint64_t *buf,
 291                                       unsigned long index,
 292                                       unsigned long pages)
 293 {
 294         if (kpagecgroup_fd < 0)
 295                 return pages;
 296
 297         return do_u64_read(kpagecgroup_fd, opt_kpageflags, buf, index, pages);
 298 }
 299
 300 static unsigned long pagemap_read(uint64_t *buf,
 301                                   unsigned long index,
 302                                   unsigned long pages)
 303 {
 304         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 305 }
 306
 307 static unsigned long pagemap_pfn(uint64_t val)
 308 {
 309         unsigned long pfn;
 310
 311         if (val & PM_PRESENT)
 312                 pfn = PM_PFRAME(val);
 313         else
 314                 pfn = 0;
 315
 316         return pfn;
 317 }
 318
 319 static unsigned long pagemap_swap_offset(uint64_t val)
 320 {
 321         return val & PM_SWAP ? PM_SWAP_OFFSET(val) : 0;
 322 }
 323
 324 /*
 325  * page flag names
 326  */
 327
 328 static char *page_flag_name(uint64_t flags)
 329 {
 330         static char buf[65];
 331         int present;
 332         size_t i, j;
 333
 334         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 335                 present = (flags >> i) & 1;
 336                 if (!page_flag_names[i]) {
 337                         if (present)
 338                                 fatal("unknown flag bit %d\n", i);
 339                         continue;
 340                 }
 341                 buf[j++] = present ? page_flag_names[i][0] : '_';
 342         }
 343
 344         return buf;
 345 }
 346
 347 static char *page_flag_longname(uint64_t flags)
 348 {
 349         static char buf[1024];
 350         size_t i, n;
 351
 352         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 353                 if (!page_flag_names[i])
 354                         continue;
 355                 if ((flags >> i) & 1)
 356                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 357                                         page_flag_names[i] + 2);
 358         }
 359         if (n)
 360                 n--;
 361         buf[n] = '\0';
 362
 363         return buf;
 364 }
 365
 366
 367 /*
 368  * page list and summary
 369  */
 370
 371 static void show_page_range(unsigned long voffset, unsigned long offset,
 372                             unsigned long size, uint64_t flags, uint64_t cgroup)
 373 {
 374         static uint64_t      flags0;
 375         static uint64_t      cgroup0;
 376         static unsigned long voff;
 377         static unsigned long index;
 378         static unsigned long count;
 379
 380         if (flags == flags0 && cgroup == cgroup0 && offset == index + count &&
 381             size && voffset == voff + count) {
 382                 count += size;
 383                 return;
 384         }
 385
 386         if (count) {
 387                 if (opt_pid)
 388                         printf("%lx\t", voff);
 389                 if (opt_file)
 390                         printf("%lu\t", voff);
 391                 if (opt_list_cgroup)
 392                         printf("@%llu\t", (unsigned long long)cgroup0);
 393                 printf("%lx\t%lx\t%s\n",
 394                                 index, count, page_flag_name(flags0));
 395         }
 396
 397         flags0 = flags;
 398         cgroup0= cgroup;
 399         index  = offset;
 400         voff   = voffset;
 401         count  = size;
 402 }
 403
 404 static void flush_page_range(void)
 405 {
 406         show_page_range(0, 0, 0, 0, 0);
 407 }
 408
 409 static void show_page(unsigned long voffset, unsigned long offset,
 410                       uint64_t flags, uint64_t cgroup)
 411 {
 412         if (opt_pid)
 413                 printf("%lx\t", voffset);
 414         if (opt_file)
 415                 printf("%lu\t", voffset);
 416         if (opt_list_cgroup)
 417                 printf("@%llu\t", (unsigned long long)cgroup);
 418         printf("%lx\t%s\n", offset, page_flag_name(flags));
 419 }
 420
 421 static void show_summary(void)
 422 {
 423         size_t i;
 424
 425         printf("             flags\tpage-count       MB"
 426                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 427
 428         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 429                 if (nr_pages[i])
 430                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 431                                 (unsigned long long)page_flags[i],
 432                                 nr_pages[i],
 433                                 pages2mb(nr_pages[i]),
 434                                 page_flag_name(page_flags[i]),
 435                                 page_flag_longname(page_flags[i]));
 436         }
 437
 438         printf("             total\t%10lu %8lu\n",
 439                         total_pages, pages2mb(total_pages));
 440 }
 441
 442
 443 /*
 444  * page flag filters
 445  */
 446
 447 static int bit_mask_ok(uint64_t flags)
 448 {
 449         int i;
 450
 451         for (i = 0; i < nr_bit_filters; i++) {
 452                 if (opt_bits[i] == KPF_ALL_BITS) {
 453                         if ((flags & opt_mask[i]) == 0)
 454                                 return 0;
 455                 } else {
 456                         if ((flags & opt_mask[i]) != opt_bits[i])
 457                                 return 0;
 458                 }
 459         }
 460
 461         return 1;
 462 }
 463
 464 static uint64_t expand_overloaded_flags(uint64_t flags, uint64_t pme)
 465 {
 466         /* SLOB/SLUB overload several page flags */
 467         if (flags & BIT(SLAB)) {
 468                 if (flags & BIT(PRIVATE))
 469                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 470                 if (flags & BIT(ACTIVE))
 471                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 472                 if (flags & BIT(ERROR))
 473                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 474         }
 475
 476         /* PG_reclaim is overloaded as PG_readahead in the read path */
 477         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 478                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 479
 480         if (pme & PM_SOFT_DIRTY)
 481                 flags |= BIT(SOFTDIRTY);
 482         if (pme & PM_FILE)
 483                 flags |= BIT(FILE);
 484         if (pme & PM_SWAP)
 485                 flags |= BIT(SWAP);
 486         if (pme & PM_MMAP_EXCLUSIVE)
 487                 flags |= BIT(MMAP_EXCLUSIVE);
 488
 489         return flags;
 490 }
 491
 492 static uint64_t well_known_flags(uint64_t flags)
 493 {
 494         /* hide flags intended only for kernel hacker */
 495         flags &= ~KPF_HACKERS_BITS;
 496
 497         /* hide non-hugeTLB compound pages */
 498         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 499                 flags &= ~BITS_COMPOUND;
 500
 501         return flags;
 502 }
 503
 504 static uint64_t kpageflags_flags(uint64_t flags, uint64_t pme)
 505 {
 506         if (opt_raw)
 507                 flags = expand_overloaded_flags(flags, pme);
 508         else
 509                 flags = well_known_flags(flags);
 510
 511         return flags;
 512 }
 513
 514 /*
 515  * page actions
 516  */
 517
 518 static void prepare_hwpoison_fd(void)
 519 {
 520         char buf[MAX_PATH + 1];
 521
 522         hwpoison_debug_fs = debugfs__mount();
 523         if (!hwpoison_debug_fs) {
 524                 perror("mount debugfs");
 525                 exit(EXIT_FAILURE);
 526         }
 527
 528         if (opt_hwpoison && !hwpoison_inject_fd) {
 529                 snprintf(buf, MAX_PATH, "%s/hwpoison/corrupt-pfn",
 530                         hwpoison_debug_fs);
 531                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 532         }
 533
 534         if (opt_unpoison && !hwpoison_forget_fd) {
 535                 snprintf(buf, MAX_PATH, "%s/hwpoison/unpoison-pfn",
 536                         hwpoison_debug_fs);
 537                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 538         }
 539 }
 540
 541 static int hwpoison_page(unsigned long offset)
 542 {
 543         char buf[100];
 544         int len;
 545
 546         len = sprintf(buf, "0x%lx\n", offset);
 547         len = write(hwpoison_inject_fd, buf, len);
 548         if (len < 0) {
 549                 perror("hwpoison inject");
 550                 return len;
 551         }
 552         return 0;
 553 }
 554
 555 static int unpoison_page(unsigned long offset)
 556 {
 557         char buf[100];
 558         int len;
 559
 560         len = sprintf(buf, "0x%lx\n", offset);
 561         len = write(hwpoison_forget_fd, buf, len);
 562         if (len < 0) {
 563                 perror("hwpoison forget");
 564                 return len;
 565         }
 566         return 0;
 567 }
 568
 569 /*
 570  * page frame walker
 571  */
 572
 573 static size_t hash_slot(uint64_t flags)
 574 {
 575         size_t k = HASH_KEY(flags);
 576         size_t i;
 577
 578         /* Explicitly reserve slot 0 for flags 0: the following logic
 579          * cannot distinguish an unoccupied slot from slot (flags==0).
 580          */
 581         if (flags == 0)
 582                 return 0;
 583
 584         /* search through the remaining (HASH_SIZE-1) slots */
 585         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 586                 if (!k || k >= ARRAY_SIZE(page_flags))
 587                         k = 1;
 588                 if (page_flags[k] == 0) {
 589                         page_flags[k] = flags;
 590                         return k;
 591                 }
 592                 if (page_flags[k] == flags)
 593                         return k;
 594         }
 595
 596         fatal("hash table full: bump up HASH_SHIFT?\n");
 597         exit(EXIT_FAILURE);
 598 }
 599
 600 static void add_page(unsigned long voffset, unsigned long offset,
 601                      uint64_t flags, uint64_t cgroup, uint64_t pme)
 602 {
 603         flags = kpageflags_flags(flags, pme);
 604
 605         if (!bit_mask_ok(flags))
 606                 return;
 607
 608         if (opt_cgroup && cgroup != (uint64_t)opt_cgroup)
 609                 return;
 610
 611         if (opt_hwpoison)
 612                 hwpoison_page(offset);
 613         if (opt_unpoison)
 614                 unpoison_page(offset);
 615
 616         if (opt_list == 1)
 617                 show_page_range(voffset, offset, 1, flags, cgroup);
 618         else if (opt_list == 2)
 619                 show_page(voffset, offset, flags, cgroup);
 620
 621         nr_pages[hash_slot(flags)]++;
 622         total_pages++;
 623 }
 624
 625 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 626 static void walk_pfn(unsigned long voffset,
 627                      unsigned long index,
 628                      unsigned long count,
 629                      uint64_t pme)
 630 {
 631         uint64_t buf[KPAGEFLAGS_BATCH];
 632         uint64_t cgi[KPAGEFLAGS_BATCH];
 633         unsigned long batch;
 634         unsigned long pages;
 635         unsigned long i;
 636
 637         /*
 638          * kpagecgroup_read() reads only if kpagecgroup were opened, but
 639          * /proc/kpagecgroup might even not exist, so it's better to fill
 640          * them with zeros here.
 641          */
 642         if (count == 1)
 643                 cgi[0] = 0;
 644         else
 645                 memset(cgi, 0, sizeof cgi);
 646
 647         while (count) {
 648                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 649                 pages = kpageflags_read(buf, index, batch);
 650                 if (pages == 0)
 651                         break;
 652
 653                 if (kpagecgroup_read(cgi, index, pages) != pages)
 654                         fatal("kpagecgroup returned fewer pages than expected");
 655
 656                 for (i = 0; i < pages; i++)
 657                         add_page(voffset + i, index + i, buf[i], cgi[i], pme);
 658
 659                 index += pages;
 660                 count -= pages;
 661         }
 662 }
 663
 664 static void walk_swap(unsigned long voffset, uint64_t pme)
 665 {
 666         uint64_t flags = kpageflags_flags(0, pme);
 667
 668         if (!bit_mask_ok(flags))
 669                 return;
 670
 671         if (opt_cgroup)
 672                 return;
 673
 674         if (opt_list == 1)
 675                 show_page_range(voffset, pagemap_swap_offset(pme), 1, flags, 0);
 676         else if (opt_list == 2)
 677                 show_page(voffset, pagemap_swap_offset(pme), flags, 0);
 678
 679         nr_pages[hash_slot(flags)]++;
 680         total_pages++;
 681 }
 682
 683 #define PAGEMAP_BATCH   (64 << 10)
 684 static void walk_vma(unsigned long index, unsigned long count)
 685 {
 686         uint64_t buf[PAGEMAP_BATCH];
 687         unsigned long batch;
 688         unsigned long pages;
 689         unsigned long pfn;
 690         unsigned long i;
 691
 692         while (count) {
 693                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 694                 pages = pagemap_read(buf, index, batch);
 695                 if (pages == 0)
 696                         break;
 697
 698                 for (i = 0; i < pages; i++) {
 699                         pfn = pagemap_pfn(buf[i]);
 700                         if (pfn)
 701                                 walk_pfn(index + i, pfn, 1, buf[i]);
 702                         if (buf[i] & PM_SWAP)
 703                                 walk_swap(index + i, buf[i]);
 704                 }
 705
 706                 index += pages;
 707                 count -= pages;
 708         }
 709 }
 710
 711 static void walk_task(unsigned long index, unsigned long count)
 712 {
 713         const unsigned long end = index + count;
 714         unsigned long start;
 715         int i = 0;
 716
 717         while (index < end) {
 718
 719                 while (pg_end[i] <= index)
 720                         if (++i >= nr_vmas)
 721                                 return;
 722                 if (pg_start[i] >= end)
 723                         return;
 724
 725                 start = max_t(unsigned long, pg_start[i], index);
 726                 index = min_t(unsigned long, pg_end[i], end);
 727
 728                 assert(start < index);
 729                 walk_vma(start, index - start);
 730         }
 731 }
 732
 733 static void add_addr_range(unsigned long offset, unsigned long size)
 734 {
 735         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 736                 fatal("too many addr ranges\n");
 737
 738         opt_offset[nr_addr_ranges] = offset;
 739         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 740         nr_addr_ranges++;
 741 }
 742
 743 static void walk_addr_ranges(void)
 744 {
 745         int i;
 746
 747         kpageflags_fd = checked_open(opt_kpageflags, O_RDONLY);
 748
 749         if (!nr_addr_ranges)
 750                 add_addr_range(0, ULONG_MAX);
 751
 752         for (i = 0; i < nr_addr_ranges; i++)
 753                 if (!opt_pid)
 754                         walk_pfn(opt_offset[i], opt_offset[i], opt_size[i], 0);
 755                 else
 756                         walk_task(opt_offset[i], opt_size[i]);
 757
 758         close(kpageflags_fd);
 759 }
 760
 761
 762 /*
 763  * user interface
 764  */
 765
 766 static const char *page_flag_type(uint64_t flag)
 767 {
 768         if (flag & KPF_HACKERS_BITS)
 769                 return "(r)";
 770         if (flag & KPF_OVERLOADED_BITS)
 771                 return "(o)";
 772         return "   ";
 773 }
 774
 775 static void usage(void)
 776 {
 777         size_t i, j;
 778
 779         printf(
 780 "page-types [options]\n"
 781 "            -r|--raw                   Raw mode, for kernel developers\n"
 782 "            -d|--describe flags        Describe flags\n"
 783 "            -a|--addr    addr-spec     Walk a range of pages\n"
 784 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 785 "            -c|--cgroup  path|@inode   Walk pages within memory cgroup\n"
 786 "            -p|--pid     pid           Walk process address space\n"
 787 "            -f|--file    filename      Walk file address space\n"
 788 "            -l|--list                  Show page details in ranges\n"
 789 "            -L|--list-each             Show page details one by one\n"
 790 "            -C|--list-cgroup           Show cgroup inode for pages\n"
 791 "            -N|--no-summary            Don't show summary info\n"
 792 "            -X|--hwpoison              hwpoison pages\n"
 793 "            -x|--unpoison              unpoison pages\n"
 794 "            -F|--kpageflags filename   kpageflags file to parse\n"
 795 "            -h|--help                  Show this usage message\n"
 796 "flags:\n"
 797 "            0x10                       bitfield format, e.g.\n"
 798 "            anon                       bit-name, e.g.\n"
 799 "            0x10,anon                  comma-separated list, e.g.\n"
 800 "addr-spec:\n"
 801 "            N                          one page at offset N (unit: pages)\n"
 802 "            N+M                        pages range from N to N+M-1\n"
 803 "            N,M                        pages range from N to M-1\n"
 804 "            N,                         pages range from N to end\n"
 805 "            ,M                         pages range from 0 to M-1\n"
 806 "bits-spec:\n"
 807 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 808 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 809 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 810 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 811 "bit-names:\n"
 812         );
 813
 814         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 815                 if (!page_flag_names[i])
 816                         continue;
 817                 printf("%16s%s", page_flag_names[i] + 2,
 818                                  page_flag_type(1ULL << i));
 819                 if (++j > 3) {
 820                         j = 0;
 821                         putchar('\n');
 822                 }
 823         }
 824         printf("\n                                   "
 825                 "(r) raw mode bits  (o) overloaded bits\n");
 826 }
 827
 828 static unsigned long long parse_number(const char *str)
 829 {
 830         unsigned long long n;
 831
 832         n = strtoll(str, NULL, 0);
 833
 834         if (n == 0 && str[0] != '0')
 835                 fatal("invalid name or number: %s\n", str);
 836
 837         return n;
 838 }
 839
 840 static void parse_pid(const char *str)
 841 {
 842         FILE *file;
 843         char buf[5000];
 844
 845         opt_pid = parse_number(str);
 846
 847         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 848         pagemap_fd = checked_open(buf, O_RDONLY);
 849
 850         sprintf(buf, "/proc/%d/maps", opt_pid);
 851         file = fopen(buf, "r");
 852         if (!file) {
 853                 perror(buf);
 854                 exit(EXIT_FAILURE);
 855         }
 856
 857         while (fgets(buf, sizeof(buf), file) != NULL) {
 858                 unsigned long vm_start;
 859                 unsigned long vm_end;
 860                 unsigned long long pgoff;
 861                 int major, minor;
 862                 char r, w, x, s;
 863                 unsigned long ino;
 864                 int n;
 865
 866                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 867                            &vm_start,
 868                            &vm_end,
 869                            &r, &w, &x, &s,
 870                            &pgoff,
 871                            &major, &minor,
 872                            &ino);
 873                 if (n < 10) {
 874                         fprintf(stderr, "unexpected line: %s\n", buf);
 875                         continue;
 876                 }
 877                 pg_start[nr_vmas] = vm_start / page_size;
 878                 pg_end[nr_vmas] = vm_end / page_size;
 879                 if (++nr_vmas >= MAX_VMAS) {
 880                         fprintf(stderr, "too many VMAs\n");
 881                         break;
 882                 }
 883         }
 884         fclose(file);
 885 }
 886
 887 static void show_file(const char *name, const struct stat *st)
 888 {
 889         unsigned long long size = st->st_size;
 890         char atime[64], mtime[64];
 891         long now = time(NULL);
 892
 893         printf("%s\tInode: %u\tSize: %llu (%llu pages)\n",
 894                         name, (unsigned)st->st_ino,
 895                         size, (size + page_size - 1) / page_size);
 896
 897         strftime(atime, sizeof(atime), "%c", localtime(&st->st_atime));
 898         strftime(mtime, sizeof(mtime), "%c", localtime(&st->st_mtime));
 899
 900         printf("Modify: %s (%ld seconds ago)\nAccess: %s (%ld seconds ago)\n",
 901                         mtime, now - st->st_mtime,
 902                         atime, now - st->st_atime);
 903 }
 904
 905 static sigjmp_buf sigbus_jmp;
 906
 907 static void * volatile sigbus_addr;
 908
 909 static void sigbus_handler(int sig, siginfo_t *info, void *ucontex)
 910 {
 911         (void)sig;
 912         (void)ucontex;
 913         sigbus_addr = info ? info->si_addr : NULL;
 914         siglongjmp(sigbus_jmp, 1);
 915 }
 916
 917 static struct sigaction sigbus_action = {
 918         .sa_sigaction = sigbus_handler,
 919         .sa_flags = SA_SIGINFO,
 920 };
 921
 922 static void walk_file(const char *name, const struct stat *st)
 923 {
 924         uint8_t vec[PAGEMAP_BATCH];
 925         uint64_t buf[PAGEMAP_BATCH], flags;
 926         uint64_t cgroup = 0;
 927         unsigned long nr_pages, pfn, i;
 928         off_t off, end = st->st_size;
 929         int fd;
 930         ssize_t len;
 931         void *ptr;
 932         int first = 1;
 933
 934         fd = checked_open(name, O_RDONLY|O_NOATIME|O_NOFOLLOW);
 935
 936         for (off = 0; off < end; off += len) {
 937                 nr_pages = (end - off + page_size - 1) / page_size;
 938                 if (nr_pages > PAGEMAP_BATCH)
 939                         nr_pages = PAGEMAP_BATCH;
 940                 len = nr_pages * page_size;
 941
 942                 ptr = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, off);
 943                 if (ptr == MAP_FAILED)
 944                         fatal("mmap failed: %s", name);
 945
 946                 /* determine cached pages */
 947                 if (mincore(ptr, len, vec))
 948                         fatal("mincore failed: %s", name);
 949
 950                 /* turn off readahead */
 951                 if (madvise(ptr, len, MADV_RANDOM))
 952                         fatal("madvice failed: %s", name);
 953
 954                 if (sigsetjmp(sigbus_jmp, 1)) {
 955                         end = off + sigbus_addr ? sigbus_addr - ptr : 0;
 956                         fprintf(stderr, "got sigbus at offset %lld: %s\n",
 957                                         (long long)end, name);
 958                         goto got_sigbus;
 959                 }
 960
 961                 /* populate ptes */
 962                 for (i = 0; i < nr_pages ; i++) {
 963                         if (vec[i] & 1)
 964                                 (void)*(volatile int *)(ptr + i * page_size);
 965                 }
 966 got_sigbus:
 967
 968                 /* turn off harvesting reference bits */
 969                 if (madvise(ptr, len, MADV_SEQUENTIAL))
 970                         fatal("madvice failed: %s", name);
 971
 972                 if (pagemap_read(buf, (unsigned long)ptr / page_size,
 973                                         nr_pages) != nr_pages)
 974                         fatal("cannot read pagemap");
 975
 976                 munmap(ptr, len);
 977
 978                 for (i = 0; i < nr_pages; i++) {
 979                         pfn = pagemap_pfn(buf[i]);
 980                         if (!pfn)
 981                                 continue;
 982                         if (!kpageflags_read(&flags, pfn, 1))
 983                                 continue;
 984                         if (!kpagecgroup_read(&cgroup, pfn, 1))
 985                                 fatal("kpagecgroup_read failed");
 986                         if (first && opt_list) {
 987                                 first = 0;
 988                                 flush_page_range();
 989                                 show_file(name, st);
 990                         }
 991                         add_page(off / page_size + i, pfn,
 992                                  flags, cgroup, buf[i]);
 993                 }
 994         }
 995
 996         close(fd);
 997 }
 998
 999 int walk_tree(const char *name, const struct stat *st, int type, struct FTW *f)
1000 {
1001         (void)f;
1002         switch (type) {
1003         case FTW_F:
1004                 if (S_ISREG(st->st_mode))
1005                         walk_file(name, st);
1006                 break;
1007         case FTW_DNR:
1008                 fprintf(stderr, "cannot read dir: %s\n", name);
1009                 break;
1010         }
1011         return 0;
1012 }
1013
1014 static void walk_page_cache(void)
1015 {
1016         struct stat st;
1017
1018         kpageflags_fd = checked_open(opt_kpageflags, O_RDONLY);
1019         pagemap_fd = checked_open("/proc/self/pagemap", O_RDONLY);
1020         sigaction(SIGBUS, &sigbus_action, NULL);
1021
1022         if (stat(opt_file, &st))
1023                 fatal("stat failed: %s\n", opt_file);
1024
1025         if (S_ISREG(st.st_mode)) {
1026                 walk_file(opt_file, &st);
1027         } else if (S_ISDIR(st.st_mode)) {
1028                 /* do not follow symlinks and mountpoints */
1029                 if (nftw(opt_file, walk_tree, 64, FTW_MOUNT | FTW_PHYS) < 0)
1030                         fatal("nftw failed: %s\n", opt_file);
1031         } else
1032                 fatal("unhandled file type: %s\n", opt_file);
1033
1034         close(kpageflags_fd);
1035         close(pagemap_fd);
1036         signal(SIGBUS, SIG_DFL);
1037 }
1038
1039 static void parse_file(const char *name)
1040 {
1041         opt_file = name;
1042 }
1043
1044 static void parse_cgroup(const char *path)
1045 {
1046         if (path[0] == '@') {
1047                 opt_cgroup = parse_number(path + 1);
1048                 return;
1049         }
1050
1051         struct stat st;
1052
1053         if (stat(path, &st))
1054                 fatal("stat failed: %s: %m\n", path);
1055
1056         if (!S_ISDIR(st.st_mode))
1057                 fatal("cgroup supposed to be a directory: %s\n", path);
1058
1059         opt_cgroup = st.st_ino;
1060 }
1061
1062 static void parse_addr_range(const char *optarg)
1063 {
1064         unsigned long offset;
1065         unsigned long size;
1066         char *p;
1067
1068         p = strchr(optarg, ',');
1069         if (!p)
1070                 p = strchr(optarg, '+');
1071
1072         if (p == optarg) {
1073                 offset = 0;
1074                 size   = parse_number(p + 1);
1075         } else if (p) {
1076                 offset = parse_number(optarg);
1077                 if (p[1] == '\0')
1078                         size = ULONG_MAX;
1079                 else {
1080                         size = parse_number(p + 1);
1081                         if (*p == ',') {
1082                                 if (size < offset)
1083                                         fatal("invalid range: %lu,%lu\n",
1084                                                         offset, size);
1085                                 size -= offset;
1086                         }
1087                 }
1088         } else {
1089                 offset = parse_number(optarg);
1090                 size   = 1;
1091         }
1092
1093         add_addr_range(offset, size);
1094 }
1095
1096 static void add_bits_filter(uint64_t mask, uint64_t bits)
1097 {
1098         if (nr_bit_filters >= MAX_BIT_FILTERS)
1099                 fatal("too much bit filters\n");
1100
1101         opt_mask[nr_bit_filters] = mask;
1102         opt_bits[nr_bit_filters] = bits;
1103         nr_bit_filters++;
1104 }
1105
1106 static uint64_t parse_flag_name(const char *str, int len)
1107 {
1108         size_t i;
1109
1110         if (!*str || !len)
1111                 return 0;
1112
1113         if (len <= 8 && !strncmp(str, "compound", len))
1114                 return BITS_COMPOUND;
1115
1116         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
1117                 if (!page_flag_names[i])
1118                         continue;
1119                 if (!strncmp(str, page_flag_names[i] + 2, len))
1120                         return 1ULL << i;
1121         }
1122
1123         return parse_number(str);
1124 }
1125
1126 static uint64_t parse_flag_names(const char *str, int all)
1127 {
1128         const char *p    = str;
1129         uint64_t   flags = 0;
1130
1131         while (1) {
1132                 if (*p == ',' || *p == '=' || *p == '\0') {
1133                         if ((*str != '~') || (*str == '~' && all && *++str))
1134                                 flags |= parse_flag_name(str, p - str);
1135                         if (*p != ',')
1136                                 break;
1137                         str = p + 1;
1138                 }
1139                 p++;
1140         }
1141
1142         return flags;
1143 }
1144
1145 static void parse_bits_mask(const char *optarg)
1146 {
1147         uint64_t mask;
1148         uint64_t bits;
1149         const char *p;
1150
1151         p = strchr(optarg, '=');
1152         if (p == optarg) {
1153                 mask = KPF_ALL_BITS;
1154                 bits = parse_flag_names(p + 1, 0);
1155         } else if (p) {
1156                 mask = parse_flag_names(optarg, 0);
1157                 bits = parse_flag_names(p + 1, 0);
1158         } else if (strchr(optarg, '~')) {
1159                 mask = parse_flag_names(optarg, 1);
1160                 bits = parse_flag_names(optarg, 0);
1161         } else {
1162                 mask = parse_flag_names(optarg, 0);
1163                 bits = KPF_ALL_BITS;
1164         }
1165
1166         add_bits_filter(mask, bits);
1167 }
1168
1169 static void parse_kpageflags(const char *name)
1170 {
1171         opt_kpageflags = name;
1172 }
1173
1174 static void describe_flags(const char *optarg)
1175 {
1176         uint64_t flags = parse_flag_names(optarg, 0);
1177
1178         printf("0x%016llx\t%s\t%s\n",
1179                 (unsigned long long)flags,
1180                 page_flag_name(flags),
1181                 page_flag_longname(flags));
1182 }
1183
1184 static const struct option opts[] = {
1185         { "raw"       , 0, NULL, 'r' },
1186         { "pid"       , 1, NULL, 'p' },
1187         { "file"      , 1, NULL, 'f' },
1188         { "addr"      , 1, NULL, 'a' },
1189         { "bits"      , 1, NULL, 'b' },
1190         { "cgroup"    , 1, NULL, 'c' },
1191         { "describe"  , 1, NULL, 'd' },
1192         { "list"      , 0, NULL, 'l' },
1193         { "list-each" , 0, NULL, 'L' },
1194         { "list-cgroup", 0, NULL, 'C' },
1195         { "no-summary", 0, NULL, 'N' },
1196         { "hwpoison"  , 0, NULL, 'X' },
1197         { "unpoison"  , 0, NULL, 'x' },
1198         { "kpageflags", 0, NULL, 'F' },
1199         { "help"      , 0, NULL, 'h' },
1200         { NULL        , 0, NULL, 0 }
1201 };
1202
1203 int main(int argc, char *argv[])
1204 {
1205         int c;
1206
1207         page_size = getpagesize();
1208
1209         while ((c = getopt_long(argc, argv,
1210                                 "rp:f:a:b:d:c:ClLNXxF:h", opts, NULL)) != -1) {
1211                 switch (c) {
1212                 case 'r':
1213                         opt_raw = 1;
1214                         break;
1215                 case 'p':
1216                         parse_pid(optarg);
1217                         break;
1218                 case 'f':
1219                         parse_file(optarg);
1220                         break;
1221                 case 'a':
1222                         parse_addr_range(optarg);
1223                         break;
1224                 case 'b':
1225                         parse_bits_mask(optarg);
1226                         break;
1227                 case 'c':
1228                         parse_cgroup(optarg);
1229                         break;
1230                 case 'C':
1231                         opt_list_cgroup = 1;
1232                         break;
1233                 case 'd':
1234                         describe_flags(optarg);
1235                         exit(0);
1236                 case 'l':
1237                         opt_list = 1;
1238                         break;
1239                 case 'L':
1240                         opt_list = 2;
1241                         break;
1242                 case 'N':
1243                         opt_no_summary = 1;
1244                         break;
1245                 case 'X':
1246                         opt_hwpoison = 1;
1247                         prepare_hwpoison_fd();
1248                         break;
1249                 case 'x':
1250                         opt_unpoison = 1;
1251                         prepare_hwpoison_fd();
1252                         break;
1253                 case 'F':
1254                         parse_kpageflags(optarg);
1255                         break;
1256                 case 'h':
1257                         usage();
1258                         exit(0);
1259                 default:
1260                         usage();
1261                         exit(1);
1262                 }
1263         }
1264
1265         if (!opt_kpageflags)
1266                 opt_kpageflags = PROC_KPAGEFLAGS;
1267
1268         if (opt_cgroup || opt_list_cgroup)
1269                 kpagecgroup_fd = checked_open(PROC_KPAGECGROUP, O_RDONLY);
1270
1271         if (opt_list && opt_pid)
1272                 printf("voffset\t");
1273         if (opt_list && opt_file)
1274                 printf("foffset\t");
1275         if (opt_list && opt_list_cgroup)
1276                 printf("cgroup\t");
1277         if (opt_list == 1)
1278                 printf("offset\tlen\tflags\n");
1279         if (opt_list == 2)
1280                 printf("offset\tflags\n");
1281
1282         if (opt_file)
1283                 walk_page_cache();
1284         else
1285                 walk_addr_ranges();
1286
1287         if (opt_list == 1)
1288                 flush_page_range();
1289
1290         if (opt_no_summary)
1291                 return 0;
1292
1293         if (opt_list)
1294                 printf("\n\n");
1295
1296         show_summary();
1297
1298         return 0;
1299 }