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
  79 /* [32-] kernel hacking assistances */
  80 #define KPF_RESERVED            32
  81 #define KPF_MLOCKED             33
  82 #define KPF_MAPPEDTODISK        34
  83 #define KPF_PRIVATE             35
  84 #define KPF_PRIVATE_2           36
  85 #define KPF_OWNER_PRIVATE       37
  86 #define KPF_ARCH                38
  87 #define KPF_UNCACHED            39
  88 #define KPF_SOFTDIRTY           40
  89
  90 /* [48-] take some arbitrary free slots for expanding overloaded flags
  91  * not part of kernel API
  92  */
  93 #define KPF_READAHEAD           48
  94 #define KPF_SLOB_FREE           49
  95 #define KPF_SLUB_FROZEN         50
  96 #define KPF_SLUB_DEBUG          51
  97 #define KPF_FILE                61
  98 #define KPF_SWAP                62
  99 #define KPF_MMAP_EXCLUSIVE      63
 100
 101 #define KPF_ALL_BITS            ((uint64_t)~0ULL)
 102 #define KPF_HACKERS_BITS        (0xffffULL << 32)
 103 #define KPF_OVERLOADED_BITS     (0xffffULL << 48)
 104 #define BIT(name)               (1ULL << KPF_##name)
 105 #define BITS_COMPOUND           (BIT(COMPOUND_HEAD) | BIT(COMPOUND_TAIL))
 106
 107 static const char * const page_flag_names[] = {
 108         [KPF_LOCKED]            = "L:locked",
 109         [KPF_ERROR]             = "E:error",
 110         [KPF_REFERENCED]        = "R:referenced",
 111         [KPF_UPTODATE]          = "U:uptodate",
 112         [KPF_DIRTY]             = "D:dirty",
 113         [KPF_LRU]               = "l:lru",
 114         [KPF_ACTIVE]            = "A:active",
 115         [KPF_SLAB]              = "S:slab",
 116         [KPF_WRITEBACK]         = "W:writeback",
 117         [KPF_RECLAIM]           = "I:reclaim",
 118         [KPF_BUDDY]             = "B:buddy",
 119
 120         [KPF_MMAP]              = "M:mmap",
 121         [KPF_ANON]              = "a:anonymous",
 122         [KPF_SWAPCACHE]         = "s:swapcache",
 123         [KPF_SWAPBACKED]        = "b:swapbacked",
 124         [KPF_COMPOUND_HEAD]     = "H:compound_head",
 125         [KPF_COMPOUND_TAIL]     = "T:compound_tail",
 126         [KPF_HUGE]              = "G:huge",
 127         [KPF_UNEVICTABLE]       = "u:unevictable",
 128         [KPF_HWPOISON]          = "X:hwpoison",
 129         [KPF_NOPAGE]            = "n:nopage",
 130         [KPF_KSM]               = "x:ksm",
 131         [KPF_THP]               = "t:thp",
 132         [KPF_BALLOON]           = "o:balloon",
 133         [KPF_ZERO_PAGE]         = "z:zero_page",
 134         [KPF_IDLE]              = "i:idle_page",
 135
 136         [KPF_RESERVED]          = "r:reserved",
 137         [KPF_MLOCKED]           = "m:mlocked",
 138         [KPF_MAPPEDTODISK]      = "d:mappedtodisk",
 139         [KPF_PRIVATE]           = "P:private",
 140         [KPF_PRIVATE_2]         = "p:private_2",
 141         [KPF_OWNER_PRIVATE]     = "O:owner_private",
 142         [KPF_ARCH]              = "h:arch",
 143         [KPF_UNCACHED]          = "c:uncached",
 144         [KPF_SOFTDIRTY]         = "f:softdirty",
 145
 146         [KPF_READAHEAD]         = "I:readahead",
 147         [KPF_SLOB_FREE]         = "P:slob_free",
 148         [KPF_SLUB_FROZEN]       = "A:slub_frozen",
 149         [KPF_SLUB_DEBUG]        = "E:slub_debug",
 150
 151         [KPF_FILE]              = "F:file",
 152         [KPF_SWAP]              = "w:swap",
 153         [KPF_MMAP_EXCLUSIVE]    = "1:mmap_exclusive",
 154 };
 155
 156
 157 static const char * const debugfs_known_mountpoints[] = {
 158         "/sys/kernel/debug",
 159         "/debug",
 160         0,
 161 };
 162
 163 /*
 164  * data structures
 165  */
 166
 167 static int              opt_raw;        /* for kernel developers */
 168 static int              opt_list;       /* list pages (in ranges) */
 169 static int              opt_no_summary; /* don't show summary */
 170 static pid_t            opt_pid;        /* process to walk */
 171 const char *            opt_file;
 172
 173 #define MAX_ADDR_RANGES 1024
 174 static int              nr_addr_ranges;
 175 static unsigned long    opt_offset[MAX_ADDR_RANGES];
 176 static unsigned long    opt_size[MAX_ADDR_RANGES];
 177
 178 #define MAX_VMAS        10240
 179 static int              nr_vmas;
 180 static unsigned long    pg_start[MAX_VMAS];
 181 static unsigned long    pg_end[MAX_VMAS];
 182
 183 #define MAX_BIT_FILTERS 64
 184 static int              nr_bit_filters;
 185 static uint64_t         opt_mask[MAX_BIT_FILTERS];
 186 static uint64_t         opt_bits[MAX_BIT_FILTERS];
 187
 188 static int              page_size;
 189
 190 static int              pagemap_fd;
 191 static int              kpageflags_fd;
 192
 193 static int              opt_hwpoison;
 194 static int              opt_unpoison;
 195
 196 static const char       *hwpoison_debug_fs;
 197 static int              hwpoison_inject_fd;
 198 static int              hwpoison_forget_fd;
 199
 200 #define HASH_SHIFT      13
 201 #define HASH_SIZE       (1 << HASH_SHIFT)
 202 #define HASH_MASK       (HASH_SIZE - 1)
 203 #define HASH_KEY(flags) (flags & HASH_MASK)
 204
 205 static unsigned long    total_pages;
 206 static unsigned long    nr_pages[HASH_SIZE];
 207 static uint64_t         page_flags[HASH_SIZE];
 208
 209
 210 /*
 211  * helper functions
 212  */
 213
 214 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 215
 216 #define min_t(type, x, y) ({                    \
 217         type __min1 = (x);                      \
 218         type __min2 = (y);                      \
 219         __min1 < __min2 ? __min1 : __min2; })
 220
 221 #define max_t(type, x, y) ({                    \
 222         type __max1 = (x);                      \
 223         type __max2 = (y);                      \
 224         __max1 > __max2 ? __max1 : __max2; })
 225
 226 static unsigned long pages2mb(unsigned long pages)
 227 {
 228         return (pages * page_size) >> 20;
 229 }
 230
 231 static void fatal(const char *x, ...)
 232 {
 233         va_list ap;
 234
 235         va_start(ap, x);
 236         vfprintf(stderr, x, ap);
 237         va_end(ap);
 238         exit(EXIT_FAILURE);
 239 }
 240
 241 static int checked_open(const char *pathname, int flags)
 242 {
 243         int fd = open(pathname, flags);
 244
 245         if (fd < 0) {
 246                 perror(pathname);
 247                 exit(EXIT_FAILURE);
 248         }
 249
 250         return fd;
 251 }
 252
 253 /*
 254  * pagemap/kpageflags routines
 255  */
 256
 257 static unsigned long do_u64_read(int fd, char *name,
 258                                  uint64_t *buf,
 259                                  unsigned long index,
 260                                  unsigned long count)
 261 {
 262         long bytes;
 263
 264         if (index > ULONG_MAX / 8)
 265                 fatal("index overflow: %lu\n", index);
 266
 267         bytes = pread(fd, buf, count * 8, (off_t)index * 8);
 268         if (bytes < 0) {
 269                 perror(name);
 270                 exit(EXIT_FAILURE);
 271         }
 272         if (bytes % 8)
 273                 fatal("partial read: %lu bytes\n", bytes);
 274
 275         return bytes / 8;
 276 }
 277
 278 static unsigned long kpageflags_read(uint64_t *buf,
 279                                      unsigned long index,
 280                                      unsigned long pages)
 281 {
 282         return do_u64_read(kpageflags_fd, PROC_KPAGEFLAGS, buf, index, pages);
 283 }
 284
 285 static unsigned long pagemap_read(uint64_t *buf,
 286                                   unsigned long index,
 287                                   unsigned long pages)
 288 {
 289         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 290 }
 291
 292 static unsigned long pagemap_pfn(uint64_t val)
 293 {
 294         unsigned long pfn;
 295
 296         if (val & PM_PRESENT)
 297                 pfn = PM_PFRAME(val);
 298         else
 299                 pfn = 0;
 300
 301         return pfn;
 302 }
 303
 304 static unsigned long pagemap_swap_offset(uint64_t val)
 305 {
 306         return val & PM_SWAP ? PM_SWAP_OFFSET(val) : 0;
 307 }
 308
 309 /*
 310  * page flag names
 311  */
 312
 313 static char *page_flag_name(uint64_t flags)
 314 {
 315         static char buf[65];
 316         int present;
 317         size_t i, j;
 318
 319         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 320                 present = (flags >> i) & 1;
 321                 if (!page_flag_names[i]) {
 322                         if (present)
 323                                 fatal("unknown flag bit %d\n", i);
 324                         continue;
 325                 }
 326                 buf[j++] = present ? page_flag_names[i][0] : '_';
 327         }
 328
 329         return buf;
 330 }
 331
 332 static char *page_flag_longname(uint64_t flags)
 333 {
 334         static char buf[1024];
 335         size_t i, n;
 336
 337         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 338                 if (!page_flag_names[i])
 339                         continue;
 340                 if ((flags >> i) & 1)
 341                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 342                                         page_flag_names[i] + 2);
 343         }
 344         if (n)
 345                 n--;
 346         buf[n] = '\0';
 347
 348         return buf;
 349 }
 350
 351
 352 /*
 353  * page list and summary
 354  */
 355
 356 static void show_page_range(unsigned long voffset, unsigned long offset,
 357                             unsigned long size, uint64_t flags)
 358 {
 359         static uint64_t      flags0;
 360         static unsigned long voff;
 361         static unsigned long index;
 362         static unsigned long count;
 363
 364         if (flags == flags0 && offset == index + count &&
 365             size && voffset == voff + count) {
 366                 count += size;
 367                 return;
 368         }
 369
 370         if (count) {
 371                 if (opt_pid)
 372                         printf("%lx\t", voff);
 373                 if (opt_file)
 374                         printf("%lu\t", voff);
 375                 printf("%lx\t%lx\t%s\n",
 376                                 index, count, page_flag_name(flags0));
 377         }
 378
 379         flags0 = flags;
 380         index  = offset;
 381         voff   = voffset;
 382         count  = size;
 383 }
 384
 385 static void flush_page_range(void)
 386 {
 387         show_page_range(0, 0, 0, 0);
 388 }
 389
 390 static void show_page(unsigned long voffset,
 391                       unsigned long offset, uint64_t flags)
 392 {
 393         if (opt_pid)
 394                 printf("%lx\t", voffset);
 395         if (opt_file)
 396                 printf("%lu\t", voffset);
 397         printf("%lx\t%s\n", offset, page_flag_name(flags));
 398 }
 399
 400 static void show_summary(void)
 401 {
 402         size_t i;
 403
 404         printf("             flags\tpage-count       MB"
 405                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 406
 407         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 408                 if (nr_pages[i])
 409                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 410                                 (unsigned long long)page_flags[i],
 411                                 nr_pages[i],
 412                                 pages2mb(nr_pages[i]),
 413                                 page_flag_name(page_flags[i]),
 414                                 page_flag_longname(page_flags[i]));
 415         }
 416
 417         printf("             total\t%10lu %8lu\n",
 418                         total_pages, pages2mb(total_pages));
 419 }
 420
 421
 422 /*
 423  * page flag filters
 424  */
 425
 426 static int bit_mask_ok(uint64_t flags)
 427 {
 428         int i;
 429
 430         for (i = 0; i < nr_bit_filters; i++) {
 431                 if (opt_bits[i] == KPF_ALL_BITS) {
 432                         if ((flags & opt_mask[i]) == 0)
 433                                 return 0;
 434                 } else {
 435                         if ((flags & opt_mask[i]) != opt_bits[i])
 436                                 return 0;
 437                 }
 438         }
 439
 440         return 1;
 441 }
 442
 443 static uint64_t expand_overloaded_flags(uint64_t flags, uint64_t pme)
 444 {
 445         /* SLOB/SLUB overload several page flags */
 446         if (flags & BIT(SLAB)) {
 447                 if (flags & BIT(PRIVATE))
 448                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 449                 if (flags & BIT(ACTIVE))
 450                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 451                 if (flags & BIT(ERROR))
 452                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 453         }
 454
 455         /* PG_reclaim is overloaded as PG_readahead in the read path */
 456         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 457                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 458
 459         if (pme & PM_SOFT_DIRTY)
 460                 flags |= BIT(SOFTDIRTY);
 461         if (pme & PM_FILE)
 462                 flags |= BIT(FILE);
 463         if (pme & PM_SWAP)
 464                 flags |= BIT(SWAP);
 465         if (pme & PM_MMAP_EXCLUSIVE)
 466                 flags |= BIT(MMAP_EXCLUSIVE);
 467
 468         return flags;
 469 }
 470
 471 static uint64_t well_known_flags(uint64_t flags)
 472 {
 473         /* hide flags intended only for kernel hacker */
 474         flags &= ~KPF_HACKERS_BITS;
 475
 476         /* hide non-hugeTLB compound pages */
 477         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 478                 flags &= ~BITS_COMPOUND;
 479
 480         return flags;
 481 }
 482
 483 static uint64_t kpageflags_flags(uint64_t flags, uint64_t pme)
 484 {
 485         if (opt_raw)
 486                 flags = expand_overloaded_flags(flags, pme);
 487         else
 488                 flags = well_known_flags(flags);
 489
 490         return flags;
 491 }
 492
 493 /*
 494  * page actions
 495  */
 496
 497 static void prepare_hwpoison_fd(void)
 498 {
 499         char buf[MAX_PATH + 1];
 500
 501         hwpoison_debug_fs = debugfs__mount();
 502         if (!hwpoison_debug_fs) {
 503                 perror("mount debugfs");
 504                 exit(EXIT_FAILURE);
 505         }
 506
 507         if (opt_hwpoison && !hwpoison_inject_fd) {
 508                 snprintf(buf, MAX_PATH, "%s/hwpoison/corrupt-pfn",
 509                         hwpoison_debug_fs);
 510                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 511         }
 512
 513         if (opt_unpoison && !hwpoison_forget_fd) {
 514                 snprintf(buf, MAX_PATH, "%s/hwpoison/unpoison-pfn",
 515                         hwpoison_debug_fs);
 516                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 517         }
 518 }
 519
 520 static int hwpoison_page(unsigned long offset)
 521 {
 522         char buf[100];
 523         int len;
 524
 525         len = sprintf(buf, "0x%lx\n", offset);
 526         len = write(hwpoison_inject_fd, buf, len);
 527         if (len < 0) {
 528                 perror("hwpoison inject");
 529                 return len;
 530         }
 531         return 0;
 532 }
 533
 534 static int unpoison_page(unsigned long offset)
 535 {
 536         char buf[100];
 537         int len;
 538
 539         len = sprintf(buf, "0x%lx\n", offset);
 540         len = write(hwpoison_forget_fd, buf, len);
 541         if (len < 0) {
 542                 perror("hwpoison forget");
 543                 return len;
 544         }
 545         return 0;
 546 }
 547
 548 /*
 549  * page frame walker
 550  */
 551
 552 static size_t hash_slot(uint64_t flags)
 553 {
 554         size_t k = HASH_KEY(flags);
 555         size_t i;
 556
 557         /* Explicitly reserve slot 0 for flags 0: the following logic
 558          * cannot distinguish an unoccupied slot from slot (flags==0).
 559          */
 560         if (flags == 0)
 561                 return 0;
 562
 563         /* search through the remaining (HASH_SIZE-1) slots */
 564         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 565                 if (!k || k >= ARRAY_SIZE(page_flags))
 566                         k = 1;
 567                 if (page_flags[k] == 0) {
 568                         page_flags[k] = flags;
 569                         return k;
 570                 }
 571                 if (page_flags[k] == flags)
 572                         return k;
 573         }
 574
 575         fatal("hash table full: bump up HASH_SHIFT?\n");
 576         exit(EXIT_FAILURE);
 577 }
 578
 579 static void add_page(unsigned long voffset,
 580                      unsigned long offset, uint64_t flags, uint64_t pme)
 581 {
 582         flags = kpageflags_flags(flags, pme);
 583
 584         if (!bit_mask_ok(flags))
 585                 return;
 586
 587         if (opt_hwpoison)
 588                 hwpoison_page(offset);
 589         if (opt_unpoison)
 590                 unpoison_page(offset);
 591
 592         if (opt_list == 1)
 593                 show_page_range(voffset, offset, 1, flags);
 594         else if (opt_list == 2)
 595                 show_page(voffset, offset, flags);
 596
 597         nr_pages[hash_slot(flags)]++;
 598         total_pages++;
 599 }
 600
 601 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 602 static void walk_pfn(unsigned long voffset,
 603                      unsigned long index,
 604                      unsigned long count,
 605                      uint64_t pme)
 606 {
 607         uint64_t buf[KPAGEFLAGS_BATCH];
 608         unsigned long batch;
 609         unsigned long pages;
 610         unsigned long i;
 611
 612         while (count) {
 613                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 614                 pages = kpageflags_read(buf, index, batch);
 615                 if (pages == 0)
 616                         break;
 617
 618                 for (i = 0; i < pages; i++)
 619                         add_page(voffset + i, index + i, buf[i], pme);
 620
 621                 index += pages;
 622                 count -= pages;
 623         }
 624 }
 625
 626 static void walk_swap(unsigned long voffset, uint64_t pme)
 627 {
 628         uint64_t flags = kpageflags_flags(0, pme);
 629
 630         if (!bit_mask_ok(flags))
 631                 return;
 632
 633         if (opt_list == 1)
 634                 show_page_range(voffset, pagemap_swap_offset(pme), 1, flags);
 635         else if (opt_list == 2)
 636                 show_page(voffset, pagemap_swap_offset(pme), flags);
 637
 638         nr_pages[hash_slot(flags)]++;
 639         total_pages++;
 640 }
 641
 642 #define PAGEMAP_BATCH   (64 << 10)
 643 static void walk_vma(unsigned long index, unsigned long count)
 644 {
 645         uint64_t buf[PAGEMAP_BATCH];
 646         unsigned long batch;
 647         unsigned long pages;
 648         unsigned long pfn;
 649         unsigned long i;
 650
 651         while (count) {
 652                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 653                 pages = pagemap_read(buf, index, batch);
 654                 if (pages == 0)
 655                         break;
 656
 657                 for (i = 0; i < pages; i++) {
 658                         pfn = pagemap_pfn(buf[i]);
 659                         if (pfn)
 660                                 walk_pfn(index + i, pfn, 1, buf[i]);
 661                         if (buf[i] & PM_SWAP)
 662                                 walk_swap(index + i, buf[i]);
 663                 }
 664
 665                 index += pages;
 666                 count -= pages;
 667         }
 668 }
 669
 670 static void walk_task(unsigned long index, unsigned long count)
 671 {
 672         const unsigned long end = index + count;
 673         unsigned long start;
 674         int i = 0;
 675
 676         while (index < end) {
 677
 678                 while (pg_end[i] <= index)
 679                         if (++i >= nr_vmas)
 680                                 return;
 681                 if (pg_start[i] >= end)
 682                         return;
 683
 684                 start = max_t(unsigned long, pg_start[i], index);
 685                 index = min_t(unsigned long, pg_end[i], end);
 686
 687                 assert(start < index);
 688                 walk_vma(start, index - start);
 689         }
 690 }
 691
 692 static void add_addr_range(unsigned long offset, unsigned long size)
 693 {
 694         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 695                 fatal("too many addr ranges\n");
 696
 697         opt_offset[nr_addr_ranges] = offset;
 698         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 699         nr_addr_ranges++;
 700 }
 701
 702 static void walk_addr_ranges(void)
 703 {
 704         int i;
 705
 706         kpageflags_fd = checked_open(PROC_KPAGEFLAGS, O_RDONLY);
 707
 708         if (!nr_addr_ranges)
 709                 add_addr_range(0, ULONG_MAX);
 710
 711         for (i = 0; i < nr_addr_ranges; i++)
 712                 if (!opt_pid)
 713                         walk_pfn(opt_offset[i], opt_offset[i], opt_size[i], 0);
 714                 else
 715                         walk_task(opt_offset[i], opt_size[i]);
 716
 717         close(kpageflags_fd);
 718 }
 719
 720
 721 /*
 722  * user interface
 723  */
 724
 725 static const char *page_flag_type(uint64_t flag)
 726 {
 727         if (flag & KPF_HACKERS_BITS)
 728                 return "(r)";
 729         if (flag & KPF_OVERLOADED_BITS)
 730                 return "(o)";
 731         return "   ";
 732 }
 733
 734 static void usage(void)
 735 {
 736         size_t i, j;
 737
 738         printf(
 739 "page-types [options]\n"
 740 "            -r|--raw                   Raw mode, for kernel developers\n"
 741 "            -d|--describe flags        Describe flags\n"
 742 "            -a|--addr    addr-spec     Walk a range of pages\n"
 743 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 744 "            -p|--pid     pid           Walk process address space\n"
 745 "            -f|--file    filename      Walk file address space\n"
 746 "            -l|--list                  Show page details in ranges\n"
 747 "            -L|--list-each             Show page details one by one\n"
 748 "            -N|--no-summary            Don't show summary info\n"
 749 "            -X|--hwpoison              hwpoison pages\n"
 750 "            -x|--unpoison              unpoison pages\n"
 751 "            -h|--help                  Show this usage message\n"
 752 "flags:\n"
 753 "            0x10                       bitfield format, e.g.\n"
 754 "            anon                       bit-name, e.g.\n"
 755 "            0x10,anon                  comma-separated list, e.g.\n"
 756 "addr-spec:\n"
 757 "            N                          one page at offset N (unit: pages)\n"
 758 "            N+M                        pages range from N to N+M-1\n"
 759 "            N,M                        pages range from N to M-1\n"
 760 "            N,                         pages range from N to end\n"
 761 "            ,M                         pages range from 0 to M-1\n"
 762 "bits-spec:\n"
 763 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 764 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 765 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 766 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 767 "bit-names:\n"
 768         );
 769
 770         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 771                 if (!page_flag_names[i])
 772                         continue;
 773                 printf("%16s%s", page_flag_names[i] + 2,
 774                                  page_flag_type(1ULL << i));
 775                 if (++j > 3) {
 776                         j = 0;
 777                         putchar('\n');
 778                 }
 779         }
 780         printf("\n                                   "
 781                 "(r) raw mode bits  (o) overloaded bits\n");
 782 }
 783
 784 static unsigned long long parse_number(const char *str)
 785 {
 786         unsigned long long n;
 787
 788         n = strtoll(str, NULL, 0);
 789
 790         if (n == 0 && str[0] != '0')
 791                 fatal("invalid name or number: %s\n", str);
 792
 793         return n;
 794 }
 795
 796 static void parse_pid(const char *str)
 797 {
 798         FILE *file;
 799         char buf[5000];
 800
 801         opt_pid = parse_number(str);
 802
 803         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 804         pagemap_fd = checked_open(buf, O_RDONLY);
 805
 806         sprintf(buf, "/proc/%d/maps", opt_pid);
 807         file = fopen(buf, "r");
 808         if (!file) {
 809                 perror(buf);
 810                 exit(EXIT_FAILURE);
 811         }
 812
 813         while (fgets(buf, sizeof(buf), file) != NULL) {
 814                 unsigned long vm_start;
 815                 unsigned long vm_end;
 816                 unsigned long long pgoff;
 817                 int major, minor;
 818                 char r, w, x, s;
 819                 unsigned long ino;
 820                 int n;
 821
 822                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 823                            &vm_start,
 824                            &vm_end,
 825                            &r, &w, &x, &s,
 826                            &pgoff,
 827                            &major, &minor,
 828                            &ino);
 829                 if (n < 10) {
 830                         fprintf(stderr, "unexpected line: %s\n", buf);
 831                         continue;
 832                 }
 833                 pg_start[nr_vmas] = vm_start / page_size;
 834                 pg_end[nr_vmas] = vm_end / page_size;
 835                 if (++nr_vmas >= MAX_VMAS) {
 836                         fprintf(stderr, "too many VMAs\n");
 837                         break;
 838                 }
 839         }
 840         fclose(file);
 841 }
 842
 843 static void show_file(const char *name, const struct stat *st)
 844 {
 845         unsigned long long size = st->st_size;
 846         char atime[64], mtime[64];
 847         long now = time(NULL);
 848
 849         printf("%s\tInode: %u\tSize: %llu (%llu pages)\n",
 850                         name, (unsigned)st->st_ino,
 851                         size, (size + page_size - 1) / page_size);
 852
 853         strftime(atime, sizeof(atime), "%c", localtime(&st->st_atime));
 854         strftime(mtime, sizeof(mtime), "%c", localtime(&st->st_mtime));
 855
 856         printf("Modify: %s (%ld seconds ago)\nAccess: %s (%ld seconds ago)\n",
 857                         mtime, now - st->st_mtime,
 858                         atime, now - st->st_atime);
 859 }
 860
 861 static sigjmp_buf sigbus_jmp;
 862
 863 static void * volatile sigbus_addr;
 864
 865 static void sigbus_handler(int sig, siginfo_t *info, void *ucontex)
 866 {
 867         (void)sig;
 868         (void)ucontex;
 869         sigbus_addr = info ? info->si_addr : NULL;
 870         siglongjmp(sigbus_jmp, 1);
 871 }
 872
 873 static struct sigaction sigbus_action = {
 874         .sa_sigaction = sigbus_handler,
 875         .sa_flags = SA_SIGINFO,
 876 };
 877
 878 static void walk_file(const char *name, const struct stat *st)
 879 {
 880         uint8_t vec[PAGEMAP_BATCH];
 881         uint64_t buf[PAGEMAP_BATCH], flags;
 882         unsigned long nr_pages, pfn, i;
 883         off_t off, end = st->st_size;
 884         int fd;
 885         ssize_t len;
 886         void *ptr;
 887         int first = 1;
 888
 889         fd = checked_open(name, O_RDONLY|O_NOATIME|O_NOFOLLOW);
 890
 891         for (off = 0; off < end; off += len) {
 892                 nr_pages = (end - off + page_size - 1) / page_size;
 893                 if (nr_pages > PAGEMAP_BATCH)
 894                         nr_pages = PAGEMAP_BATCH;
 895                 len = nr_pages * page_size;
 896
 897                 ptr = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, off);
 898                 if (ptr == MAP_FAILED)
 899                         fatal("mmap failed: %s", name);
 900
 901                 /* determine cached pages */
 902                 if (mincore(ptr, len, vec))
 903                         fatal("mincore failed: %s", name);
 904
 905                 /* turn off readahead */
 906                 if (madvise(ptr, len, MADV_RANDOM))
 907                         fatal("madvice failed: %s", name);
 908
 909                 if (sigsetjmp(sigbus_jmp, 1)) {
 910                         end = off + sigbus_addr ? sigbus_addr - ptr : 0;
 911                         fprintf(stderr, "got sigbus at offset %lld: %s\n",
 912                                         (long long)end, name);
 913                         goto got_sigbus;
 914                 }
 915
 916                 /* populate ptes */
 917                 for (i = 0; i < nr_pages ; i++) {
 918                         if (vec[i] & 1)
 919                                 (void)*(volatile int *)(ptr + i * page_size);
 920                 }
 921 got_sigbus:
 922
 923                 /* turn off harvesting reference bits */
 924                 if (madvise(ptr, len, MADV_SEQUENTIAL))
 925                         fatal("madvice failed: %s", name);
 926
 927                 if (pagemap_read(buf, (unsigned long)ptr / page_size,
 928                                         nr_pages) != nr_pages)
 929                         fatal("cannot read pagemap");
 930
 931                 munmap(ptr, len);
 932
 933                 for (i = 0; i < nr_pages; i++) {
 934                         pfn = pagemap_pfn(buf[i]);
 935                         if (!pfn)
 936                                 continue;
 937                         if (!kpageflags_read(&flags, pfn, 1))
 938                                 continue;
 939                         if (first && opt_list) {
 940                                 first = 0;
 941                                 flush_page_range();
 942                                 show_file(name, st);
 943                         }
 944                         add_page(off / page_size + i, pfn, flags, buf[i]);
 945                 }
 946         }
 947
 948         close(fd);
 949 }
 950
 951 int walk_tree(const char *name, const struct stat *st, int type, struct FTW *f)
 952 {
 953         (void)f;
 954         switch (type) {
 955         case FTW_F:
 956                 if (S_ISREG(st->st_mode))
 957                         walk_file(name, st);
 958                 break;
 959         case FTW_DNR:
 960                 fprintf(stderr, "cannot read dir: %s\n", name);
 961                 break;
 962         }
 963         return 0;
 964 }
 965
 966 static void walk_page_cache(void)
 967 {
 968         struct stat st;
 969
 970         kpageflags_fd = checked_open(PROC_KPAGEFLAGS, O_RDONLY);
 971         pagemap_fd = checked_open("/proc/self/pagemap", O_RDONLY);
 972         sigaction(SIGBUS, &sigbus_action, NULL);
 973
 974         if (stat(opt_file, &st))
 975                 fatal("stat failed: %s\n", opt_file);
 976
 977         if (S_ISREG(st.st_mode)) {
 978                 walk_file(opt_file, &st);
 979         } else if (S_ISDIR(st.st_mode)) {
 980                 /* do not follow symlinks and mountpoints */
 981                 if (nftw(opt_file, walk_tree, 64, FTW_MOUNT | FTW_PHYS) < 0)
 982                         fatal("nftw failed: %s\n", opt_file);
 983         } else
 984                 fatal("unhandled file type: %s\n", opt_file);
 985
 986         close(kpageflags_fd);
 987         close(pagemap_fd);
 988         signal(SIGBUS, SIG_DFL);
 989 }
 990
 991 static void parse_file(const char *name)
 992 {
 993         opt_file = name;
 994 }
 995
 996 static void parse_addr_range(const char *optarg)
 997 {
 998         unsigned long offset;
 999         unsigned long size;
1000         char *p;
1001
1002         p = strchr(optarg, ',');
1003         if (!p)
1004                 p = strchr(optarg, '+');
1005
1006         if (p == optarg) {
1007                 offset = 0;
1008                 size   = parse_number(p + 1);
1009         } else if (p) {
1010                 offset = parse_number(optarg);
1011                 if (p[1] == '\0')
1012                         size = ULONG_MAX;
1013                 else {
1014                         size = parse_number(p + 1);
1015                         if (*p == ',') {
1016                                 if (size < offset)
1017                                         fatal("invalid range: %lu,%lu\n",
1018                                                         offset, size);
1019                                 size -= offset;
1020                         }
1021                 }
1022         } else {
1023                 offset = parse_number(optarg);
1024                 size   = 1;
1025         }
1026
1027         add_addr_range(offset, size);
1028 }
1029
1030 static void add_bits_filter(uint64_t mask, uint64_t bits)
1031 {
1032         if (nr_bit_filters >= MAX_BIT_FILTERS)
1033                 fatal("too much bit filters\n");
1034
1035         opt_mask[nr_bit_filters] = mask;
1036         opt_bits[nr_bit_filters] = bits;
1037         nr_bit_filters++;
1038 }
1039
1040 static uint64_t parse_flag_name(const char *str, int len)
1041 {
1042         size_t i;
1043
1044         if (!*str || !len)
1045                 return 0;
1046
1047         if (len <= 8 && !strncmp(str, "compound", len))
1048                 return BITS_COMPOUND;
1049
1050         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
1051                 if (!page_flag_names[i])
1052                         continue;
1053                 if (!strncmp(str, page_flag_names[i] + 2, len))
1054                         return 1ULL << i;
1055         }
1056
1057         return parse_number(str);
1058 }
1059
1060 static uint64_t parse_flag_names(const char *str, int all)
1061 {
1062         const char *p    = str;
1063         uint64_t   flags = 0;
1064
1065         while (1) {
1066                 if (*p == ',' || *p == '=' || *p == '\0') {
1067                         if ((*str != '~') || (*str == '~' && all && *++str))
1068                                 flags |= parse_flag_name(str, p - str);
1069                         if (*p != ',')
1070                                 break;
1071                         str = p + 1;
1072                 }
1073                 p++;
1074         }
1075
1076         return flags;
1077 }
1078
1079 static void parse_bits_mask(const char *optarg)
1080 {
1081         uint64_t mask;
1082         uint64_t bits;
1083         const char *p;
1084
1085         p = strchr(optarg, '=');
1086         if (p == optarg) {
1087                 mask = KPF_ALL_BITS;
1088                 bits = parse_flag_names(p + 1, 0);
1089         } else if (p) {
1090                 mask = parse_flag_names(optarg, 0);
1091                 bits = parse_flag_names(p + 1, 0);
1092         } else if (strchr(optarg, '~')) {
1093                 mask = parse_flag_names(optarg, 1);
1094                 bits = parse_flag_names(optarg, 0);
1095         } else {
1096                 mask = parse_flag_names(optarg, 0);
1097                 bits = KPF_ALL_BITS;
1098         }
1099
1100         add_bits_filter(mask, bits);
1101 }
1102
1103 static void describe_flags(const char *optarg)
1104 {
1105         uint64_t flags = parse_flag_names(optarg, 0);
1106
1107         printf("0x%016llx\t%s\t%s\n",
1108                 (unsigned long long)flags,
1109                 page_flag_name(flags),
1110                 page_flag_longname(flags));
1111 }
1112
1113 static const struct option opts[] = {
1114         { "raw"       , 0, NULL, 'r' },
1115         { "pid"       , 1, NULL, 'p' },
1116         { "file"      , 1, NULL, 'f' },
1117         { "addr"      , 1, NULL, 'a' },
1118         { "bits"      , 1, NULL, 'b' },
1119         { "describe"  , 1, NULL, 'd' },
1120         { "list"      , 0, NULL, 'l' },
1121         { "list-each" , 0, NULL, 'L' },
1122         { "no-summary", 0, NULL, 'N' },
1123         { "hwpoison"  , 0, NULL, 'X' },
1124         { "unpoison"  , 0, NULL, 'x' },
1125         { "help"      , 0, NULL, 'h' },
1126         { NULL        , 0, NULL, 0 }
1127 };
1128
1129 int main(int argc, char *argv[])
1130 {
1131         int c;
1132
1133         page_size = getpagesize();
1134
1135         while ((c = getopt_long(argc, argv,
1136                                 "rp:f:a:b:d:lLNXxh", opts, NULL)) != -1) {
1137                 switch (c) {
1138                 case 'r':
1139                         opt_raw = 1;
1140                         break;
1141                 case 'p':
1142                         parse_pid(optarg);
1143                         break;
1144                 case 'f':
1145                         parse_file(optarg);
1146                         break;
1147                 case 'a':
1148                         parse_addr_range(optarg);
1149                         break;
1150                 case 'b':
1151                         parse_bits_mask(optarg);
1152                         break;
1153                 case 'd':
1154                         describe_flags(optarg);
1155                         exit(0);
1156                 case 'l':
1157                         opt_list = 1;
1158                         break;
1159                 case 'L':
1160                         opt_list = 2;
1161                         break;
1162                 case 'N':
1163                         opt_no_summary = 1;
1164                         break;
1165                 case 'X':
1166                         opt_hwpoison = 1;
1167                         prepare_hwpoison_fd();
1168                         break;
1169                 case 'x':
1170                         opt_unpoison = 1;
1171                         prepare_hwpoison_fd();
1172                         break;
1173                 case 'h':
1174                         usage();
1175                         exit(0);
1176                 default:
1177                         usage();
1178                         exit(1);
1179                 }
1180         }
1181
1182         if (opt_list && opt_pid)
1183                 printf("voffset\t");
1184         if (opt_list && opt_file)
1185                 printf("foffset\t");
1186         if (opt_list == 1)
1187                 printf("offset\tlen\tflags\n");
1188         if (opt_list == 2)
1189                 printf("offset\tflags\n");
1190
1191         if (opt_file)
1192                 walk_page_cache();
1193         else
1194                 walk_addr_ranges();
1195
1196         if (opt_list == 1)
1197                 flush_page_range();
1198
1199         if (opt_no_summary)
1200                 return 0;
1201
1202         if (opt_list)
1203                 printf("\n\n");
1204
1205         show_summary();
1206
1207         return 0;
1208 }