kernel/tools/vm/page-types.c

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