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