Pileus Git - ~andy/linux/blob - 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 _LARGEFILE64_SOURCE
  23 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <unistd.h>
  26 #include <stdint.h>
  27 #include <stdarg.h>
  28 #include <string.h>
  29 #include <getopt.h>
  30 #include <limits.h>
  31 #include <assert.h>
  32 #include <sys/types.h>
  33 #include <sys/errno.h>
  34 #include <sys/fcntl.h>
  35 #include <sys/mount.h>
  36 #include <sys/statfs.h>
  37 #include "../../include/uapi/linux/magic.h"
  38 #include "../../include/uapi/linux/kernel-page-flags.h"
  39 #include <api/fs/debugfs.h>
  40
  41 #ifndef MAX_PATH
  42 # define MAX_PATH 256
  43 #endif
  44
  45 #ifndef STR
  46 # define _STR(x) #x
  47 # define STR(x) _STR(x)
  48 #endif
  49
  50 /*
  51  * pagemap kernel ABI bits
  52  */
  53
  54 #define PM_ENTRY_BYTES      sizeof(uint64_t)
  55 #define PM_STATUS_BITS      3
  56 #define PM_STATUS_OFFSET    (64 - PM_STATUS_BITS)
  57 #define PM_STATUS_MASK      (((1LL << PM_STATUS_BITS) - 1) << PM_STATUS_OFFSET)
  58 #define PM_STATUS(nr)       (((nr) << PM_STATUS_OFFSET) & PM_STATUS_MASK)
  59 #define PM_PSHIFT_BITS      6
  60 #define PM_PSHIFT_OFFSET    (PM_STATUS_OFFSET - PM_PSHIFT_BITS)
  61 #define PM_PSHIFT_MASK      (((1LL << PM_PSHIFT_BITS) - 1) << PM_PSHIFT_OFFSET)
  62 #define __PM_PSHIFT(x)      (((uint64_t) (x) << PM_PSHIFT_OFFSET) & PM_PSHIFT_MASK)
  63 #define PM_PFRAME_MASK      ((1LL << PM_PSHIFT_OFFSET) - 1)
  64 #define PM_PFRAME(x)        ((x) & PM_PFRAME_MASK)
  65
  66 #define __PM_SOFT_DIRTY      (1LL)
  67 #define PM_PRESENT          PM_STATUS(4LL)
  68 #define PM_SWAP             PM_STATUS(2LL)
  69 #define PM_SOFT_DIRTY       __PM_PSHIFT(__PM_SOFT_DIRTY)
  70
  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
  98 #define KPF_ALL_BITS            ((uint64_t)~0ULL)
  99 #define KPF_HACKERS_BITS        (0xffffULL << 32)
 100 #define KPF_OVERLOADED_BITS     (0xffffULL << 48)
 101 #define BIT(name)               (1ULL << KPF_##name)
 102 #define BITS_COMPOUND           (BIT(COMPOUND_HEAD) | BIT(COMPOUND_TAIL))
 103
 104 static const char * const page_flag_names[] = {
 105         [KPF_LOCKED]            = "L:locked",
 106         [KPF_ERROR]             = "E:error",
 107         [KPF_REFERENCED]        = "R:referenced",
 108         [KPF_UPTODATE]          = "U:uptodate",
 109         [KPF_DIRTY]             = "D:dirty",
 110         [KPF_LRU]               = "l:lru",
 111         [KPF_ACTIVE]            = "A:active",
 112         [KPF_SLAB]              = "S:slab",
 113         [KPF_WRITEBACK]         = "W:writeback",
 114         [KPF_RECLAIM]           = "I:reclaim",
 115         [KPF_BUDDY]             = "B:buddy",
 116
 117         [KPF_MMAP]              = "M:mmap",
 118         [KPF_ANON]              = "a:anonymous",
 119         [KPF_SWAPCACHE]         = "s:swapcache",
 120         [KPF_SWAPBACKED]        = "b:swapbacked",
 121         [KPF_COMPOUND_HEAD]     = "H:compound_head",
 122         [KPF_COMPOUND_TAIL]     = "T:compound_tail",
 123         [KPF_HUGE]              = "G:huge",
 124         [KPF_UNEVICTABLE]       = "u:unevictable",
 125         [KPF_HWPOISON]          = "X:hwpoison",
 126         [KPF_NOPAGE]            = "n:nopage",
 127         [KPF_KSM]               = "x:ksm",
 128         [KPF_THP]               = "t:thp",
 129
 130         [KPF_RESERVED]          = "r:reserved",
 131         [KPF_MLOCKED]           = "m:mlocked",
 132         [KPF_MAPPEDTODISK]      = "d:mappedtodisk",
 133         [KPF_PRIVATE]           = "P:private",
 134         [KPF_PRIVATE_2]         = "p:private_2",
 135         [KPF_OWNER_PRIVATE]     = "O:owner_private",
 136         [KPF_ARCH]              = "h:arch",
 137         [KPF_UNCACHED]          = "c:uncached",
 138         [KPF_SOFTDIRTY]         = "f:softdirty",
 139
 140         [KPF_READAHEAD]         = "I:readahead",
 141         [KPF_SLOB_FREE]         = "P:slob_free",
 142         [KPF_SLUB_FROZEN]       = "A:slub_frozen",
 143         [KPF_SLUB_DEBUG]        = "E:slub_debug",
 144 };
 145
 146
 147 static const char * const debugfs_known_mountpoints[] = {
 148         "/sys/kernel/debug",
 149         "/debug",
 150         0,
 151 };
 152
 153 /*
 154  * data structures
 155  */
 156
 157 static int              opt_raw;        /* for kernel developers */
 158 static int              opt_list;       /* list pages (in ranges) */
 159 static int              opt_no_summary; /* don't show summary */
 160 static pid_t            opt_pid;        /* process to walk */
 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         if (lseek(fd, index * 8, SEEK_SET) < 0) {
 257                 perror(name);
 258                 exit(EXIT_FAILURE);
 259         }
 260
 261         bytes = read(fd, buf, count * 8);
 262         if (bytes < 0) {
 263                 perror(name);
 264                 exit(EXIT_FAILURE);
 265         }
 266         if (bytes % 8)
 267                 fatal("partial read: %lu bytes\n", bytes);
 268
 269         return bytes / 8;
 270 }
 271
 272 static unsigned long kpageflags_read(uint64_t *buf,
 273                                      unsigned long index,
 274                                      unsigned long pages)
 275 {
 276         return do_u64_read(kpageflags_fd, PROC_KPAGEFLAGS, buf, index, pages);
 277 }
 278
 279 static unsigned long pagemap_read(uint64_t *buf,
 280                                   unsigned long index,
 281                                   unsigned long pages)
 282 {
 283         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 284 }
 285
 286 static unsigned long pagemap_pfn(uint64_t val)
 287 {
 288         unsigned long pfn;
 289
 290         if (val & PM_PRESENT)
 291                 pfn = PM_PFRAME(val);
 292         else
 293                 pfn = 0;
 294
 295         return pfn;
 296 }
 297
 298
 299 /*
 300  * page flag names
 301  */
 302
 303 static char *page_flag_name(uint64_t flags)
 304 {
 305         static char buf[65];
 306         int present;
 307         size_t i, j;
 308
 309         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 310                 present = (flags >> i) & 1;
 311                 if (!page_flag_names[i]) {
 312                         if (present)
 313                                 fatal("unknown flag bit %d\n", i);
 314                         continue;
 315                 }
 316                 buf[j++] = present ? page_flag_names[i][0] : '_';
 317         }
 318
 319         return buf;
 320 }
 321
 322 static char *page_flag_longname(uint64_t flags)
 323 {
 324         static char buf[1024];
 325         size_t i, n;
 326
 327         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 328                 if (!page_flag_names[i])
 329                         continue;
 330                 if ((flags >> i) & 1)
 331                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 332                                         page_flag_names[i] + 2);
 333         }
 334         if (n)
 335                 n--;
 336         buf[n] = '\0';
 337
 338         return buf;
 339 }
 340
 341
 342 /*
 343  * page list and summary
 344  */
 345
 346 static void show_page_range(unsigned long voffset,
 347                             unsigned long offset, uint64_t flags)
 348 {
 349         static uint64_t      flags0;
 350         static unsigned long voff;
 351         static unsigned long index;
 352         static unsigned long count;
 353
 354         if (flags == flags0 && offset == index + count &&
 355             (!opt_pid || voffset == voff + count)) {
 356                 count++;
 357                 return;
 358         }
 359
 360         if (count) {
 361                 if (opt_pid)
 362                         printf("%lx\t", voff);
 363                 printf("%lx\t%lx\t%s\n",
 364                                 index, count, page_flag_name(flags0));
 365         }
 366
 367         flags0 = flags;
 368         index  = offset;
 369         voff   = voffset;
 370         count  = 1;
 371 }
 372
 373 static void show_page(unsigned long voffset,
 374                       unsigned long offset, uint64_t flags)
 375 {
 376         if (opt_pid)
 377                 printf("%lx\t", voffset);
 378         printf("%lx\t%s\n", offset, page_flag_name(flags));
 379 }
 380
 381 static void show_summary(void)
 382 {
 383         size_t i;
 384
 385         printf("             flags\tpage-count       MB"
 386                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 387
 388         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 389                 if (nr_pages[i])
 390                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 391                                 (unsigned long long)page_flags[i],
 392                                 nr_pages[i],
 393                                 pages2mb(nr_pages[i]),
 394                                 page_flag_name(page_flags[i]),
 395                                 page_flag_longname(page_flags[i]));
 396         }
 397
 398         printf("             total\t%10lu %8lu\n",
 399                         total_pages, pages2mb(total_pages));
 400 }
 401
 402
 403 /*
 404  * page flag filters
 405  */
 406
 407 static int bit_mask_ok(uint64_t flags)
 408 {
 409         int i;
 410
 411         for (i = 0; i < nr_bit_filters; i++) {
 412                 if (opt_bits[i] == KPF_ALL_BITS) {
 413                         if ((flags & opt_mask[i]) == 0)
 414                                 return 0;
 415                 } else {
 416                         if ((flags & opt_mask[i]) != opt_bits[i])
 417                                 return 0;
 418                 }
 419         }
 420
 421         return 1;
 422 }
 423
 424 static uint64_t expand_overloaded_flags(uint64_t flags, uint64_t pme)
 425 {
 426         /* SLOB/SLUB overload several page flags */
 427         if (flags & BIT(SLAB)) {
 428                 if (flags & BIT(PRIVATE))
 429                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 430                 if (flags & BIT(ACTIVE))
 431                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 432                 if (flags & BIT(ERROR))
 433                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 434         }
 435
 436         /* PG_reclaim is overloaded as PG_readahead in the read path */
 437         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 438                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 439
 440         if (pme & PM_SOFT_DIRTY)
 441                 flags |= BIT(SOFTDIRTY);
 442
 443         return flags;
 444 }
 445
 446 static uint64_t well_known_flags(uint64_t flags)
 447 {
 448         /* hide flags intended only for kernel hacker */
 449         flags &= ~KPF_HACKERS_BITS;
 450
 451         /* hide non-hugeTLB compound pages */
 452         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 453                 flags &= ~BITS_COMPOUND;
 454
 455         return flags;
 456 }
 457
 458 static uint64_t kpageflags_flags(uint64_t flags, uint64_t pme)
 459 {
 460         if (opt_raw)
 461                 flags = expand_overloaded_flags(flags, pme);
 462         else
 463                 flags = well_known_flags(flags);
 464
 465         return flags;
 466 }
 467
 468 /*
 469  * page actions
 470  */
 471
 472 static void prepare_hwpoison_fd(void)
 473 {
 474         char buf[MAX_PATH + 1];
 475
 476         hwpoison_debug_fs = debugfs_mount(NULL);
 477         if (!hwpoison_debug_fs) {
 478                 perror("mount debugfs");
 479                 exit(EXIT_FAILURE);
 480         }
 481
 482         if (opt_hwpoison && !hwpoison_inject_fd) {
 483                 snprintf(buf, MAX_PATH, "%s/hwpoison/corrupt-pfn",
 484                         hwpoison_debug_fs);
 485                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 486         }
 487
 488         if (opt_unpoison && !hwpoison_forget_fd) {
 489                 snprintf(buf, MAX_PATH, "%s/hwpoison/unpoison-pfn",
 490                         hwpoison_debug_fs);
 491                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 492         }
 493 }
 494
 495 static int hwpoison_page(unsigned long offset)
 496 {
 497         char buf[100];
 498         int len;
 499
 500         len = sprintf(buf, "0x%lx\n", offset);
 501         len = write(hwpoison_inject_fd, buf, len);
 502         if (len < 0) {
 503                 perror("hwpoison inject");
 504                 return len;
 505         }
 506         return 0;
 507 }
 508
 509 static int unpoison_page(unsigned long offset)
 510 {
 511         char buf[100];
 512         int len;
 513
 514         len = sprintf(buf, "0x%lx\n", offset);
 515         len = write(hwpoison_forget_fd, buf, len);
 516         if (len < 0) {
 517                 perror("hwpoison forget");
 518                 return len;
 519         }
 520         return 0;
 521 }
 522
 523 /*
 524  * page frame walker
 525  */
 526
 527 static size_t hash_slot(uint64_t flags)
 528 {
 529         size_t k = HASH_KEY(flags);
 530         size_t i;
 531
 532         /* Explicitly reserve slot 0 for flags 0: the following logic
 533          * cannot distinguish an unoccupied slot from slot (flags==0).
 534          */
 535         if (flags == 0)
 536                 return 0;
 537
 538         /* search through the remaining (HASH_SIZE-1) slots */
 539         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 540                 if (!k || k >= ARRAY_SIZE(page_flags))
 541                         k = 1;
 542                 if (page_flags[k] == 0) {
 543                         page_flags[k] = flags;
 544                         return k;
 545                 }
 546                 if (page_flags[k] == flags)
 547                         return k;
 548         }
 549
 550         fatal("hash table full: bump up HASH_SHIFT?\n");
 551         exit(EXIT_FAILURE);
 552 }
 553
 554 static void add_page(unsigned long voffset,
 555                      unsigned long offset, uint64_t flags, uint64_t pme)
 556 {
 557         flags = kpageflags_flags(flags, pme);
 558
 559         if (!bit_mask_ok(flags))
 560                 return;
 561
 562         if (opt_hwpoison)
 563                 hwpoison_page(offset);
 564         if (opt_unpoison)
 565                 unpoison_page(offset);
 566
 567         if (opt_list == 1)
 568                 show_page_range(voffset, offset, flags);
 569         else if (opt_list == 2)
 570                 show_page(voffset, offset, flags);
 571
 572         nr_pages[hash_slot(flags)]++;
 573         total_pages++;
 574 }
 575
 576 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 577 static void walk_pfn(unsigned long voffset,
 578                      unsigned long index,
 579                      unsigned long count,
 580                      uint64_t pme)
 581 {
 582         uint64_t buf[KPAGEFLAGS_BATCH];
 583         unsigned long batch;
 584         unsigned long pages;
 585         unsigned long i;
 586
 587         while (count) {
 588                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 589                 pages = kpageflags_read(buf, index, batch);
 590                 if (pages == 0)
 591                         break;
 592
 593                 for (i = 0; i < pages; i++)
 594                         add_page(voffset + i, index + i, buf[i], pme);
 595
 596                 index += pages;
 597                 count -= pages;
 598         }
 599 }
 600
 601 #define PAGEMAP_BATCH   (64 << 10)
 602 static void walk_vma(unsigned long index, unsigned long count)
 603 {
 604         uint64_t buf[PAGEMAP_BATCH];
 605         unsigned long batch;
 606         unsigned long pages;
 607         unsigned long pfn;
 608         unsigned long i;
 609
 610         while (count) {
 611                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 612                 pages = pagemap_read(buf, index, batch);
 613                 if (pages == 0)
 614                         break;
 615
 616                 for (i = 0; i < pages; i++) {
 617                         pfn = pagemap_pfn(buf[i]);
 618                         if (pfn)
 619                                 walk_pfn(index + i, pfn, 1, buf[i]);
 620                 }
 621
 622                 index += pages;
 623                 count -= pages;
 624         }
 625 }
 626
 627 static void walk_task(unsigned long index, unsigned long count)
 628 {
 629         const unsigned long end = index + count;
 630         unsigned long start;
 631         int i = 0;
 632
 633         while (index < end) {
 634
 635                 while (pg_end[i] <= index)
 636                         if (++i >= nr_vmas)
 637                                 return;
 638                 if (pg_start[i] >= end)
 639                         return;
 640
 641                 start = max_t(unsigned long, pg_start[i], index);
 642                 index = min_t(unsigned long, pg_end[i], end);
 643
 644                 assert(start < index);
 645                 walk_vma(start, index - start);
 646         }
 647 }
 648
 649 static void add_addr_range(unsigned long offset, unsigned long size)
 650 {
 651         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 652                 fatal("too many addr ranges\n");
 653
 654         opt_offset[nr_addr_ranges] = offset;
 655         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 656         nr_addr_ranges++;
 657 }
 658
 659 static void walk_addr_ranges(void)
 660 {
 661         int i;
 662
 663         kpageflags_fd = checked_open(PROC_KPAGEFLAGS, O_RDONLY);
 664
 665         if (!nr_addr_ranges)
 666                 add_addr_range(0, ULONG_MAX);
 667
 668         for (i = 0; i < nr_addr_ranges; i++)
 669                 if (!opt_pid)
 670                         walk_pfn(0, opt_offset[i], opt_size[i], 0);
 671                 else
 672                         walk_task(opt_offset[i], opt_size[i]);
 673
 674         close(kpageflags_fd);
 675 }
 676
 677
 678 /*
 679  * user interface
 680  */
 681
 682 static const char *page_flag_type(uint64_t flag)
 683 {
 684         if (flag & KPF_HACKERS_BITS)
 685                 return "(r)";
 686         if (flag & KPF_OVERLOADED_BITS)
 687                 return "(o)";
 688         return "   ";
 689 }
 690
 691 static void usage(void)
 692 {
 693         size_t i, j;
 694
 695         printf(
 696 "page-types [options]\n"
 697 "            -r|--raw                   Raw mode, for kernel developers\n"
 698 "            -d|--describe flags        Describe flags\n"
 699 "            -a|--addr    addr-spec     Walk a range of pages\n"
 700 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 701 "            -p|--pid     pid           Walk process address space\n"
 702 #if 0 /* planned features */
 703 "            -f|--file    filename      Walk file address space\n"
 704 #endif
 705 "            -l|--list                  Show page details in ranges\n"
 706 "            -L|--list-each             Show page details one by one\n"
 707 "            -N|--no-summary            Don't show summary info\n"
 708 "            -X|--hwpoison              hwpoison pages\n"
 709 "            -x|--unpoison              unpoison pages\n"
 710 "            -h|--help                  Show this usage message\n"
 711 "flags:\n"
 712 "            0x10                       bitfield format, e.g.\n"
 713 "            anon                       bit-name, e.g.\n"
 714 "            0x10,anon                  comma-separated list, e.g.\n"
 715 "addr-spec:\n"
 716 "            N                          one page at offset N (unit: pages)\n"
 717 "            N+M                        pages range from N to N+M-1\n"
 718 "            N,M                        pages range from N to M-1\n"
 719 "            N,                         pages range from N to end\n"
 720 "            ,M                         pages range from 0 to M-1\n"
 721 "bits-spec:\n"
 722 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 723 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 724 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 725 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 726 "bit-names:\n"
 727         );
 728
 729         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 730                 if (!page_flag_names[i])
 731                         continue;
 732                 printf("%16s%s", page_flag_names[i] + 2,
 733                                  page_flag_type(1ULL << i));
 734                 if (++j > 3) {
 735                         j = 0;
 736                         putchar('\n');
 737                 }
 738         }
 739         printf("\n                                   "
 740                 "(r) raw mode bits  (o) overloaded bits\n");
 741 }
 742
 743 static unsigned long long parse_number(const char *str)
 744 {
 745         unsigned long long n;
 746
 747         n = strtoll(str, NULL, 0);
 748
 749         if (n == 0 && str[0] != '0')
 750                 fatal("invalid name or number: %s\n", str);
 751
 752         return n;
 753 }
 754
 755 static void parse_pid(const char *str)
 756 {
 757         FILE *file;
 758         char buf[5000];
 759
 760         opt_pid = parse_number(str);
 761
 762         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 763         pagemap_fd = checked_open(buf, O_RDONLY);
 764
 765         sprintf(buf, "/proc/%d/maps", opt_pid);
 766         file = fopen(buf, "r");
 767         if (!file) {
 768                 perror(buf);
 769                 exit(EXIT_FAILURE);
 770         }
 771
 772         while (fgets(buf, sizeof(buf), file) != NULL) {
 773                 unsigned long vm_start;
 774                 unsigned long vm_end;
 775                 unsigned long long pgoff;
 776                 int major, minor;
 777                 char r, w, x, s;
 778                 unsigned long ino;
 779                 int n;
 780
 781                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 782                            &vm_start,
 783                            &vm_end,
 784                            &r, &w, &x, &s,
 785                            &pgoff,
 786                            &major, &minor,
 787                            &ino);
 788                 if (n < 10) {
 789                         fprintf(stderr, "unexpected line: %s\n", buf);
 790                         continue;
 791                 }
 792                 pg_start[nr_vmas] = vm_start / page_size;
 793                 pg_end[nr_vmas] = vm_end / page_size;
 794                 if (++nr_vmas >= MAX_VMAS) {
 795                         fprintf(stderr, "too many VMAs\n");
 796                         break;
 797                 }
 798         }
 799         fclose(file);
 800 }
 801
 802 static void parse_file(const char *name)
 803 {
 804 }
 805
 806 static void parse_addr_range(const char *optarg)
 807 {
 808         unsigned long offset;
 809         unsigned long size;
 810         char *p;
 811
 812         p = strchr(optarg, ',');
 813         if (!p)
 814                 p = strchr(optarg, '+');
 815
 816         if (p == optarg) {
 817                 offset = 0;
 818                 size   = parse_number(p + 1);
 819         } else if (p) {
 820                 offset = parse_number(optarg);
 821                 if (p[1] == '\0')
 822                         size = ULONG_MAX;
 823                 else {
 824                         size = parse_number(p + 1);
 825                         if (*p == ',') {
 826                                 if (size < offset)
 827                                         fatal("invalid range: %lu,%lu\n",
 828                                                         offset, size);
 829                                 size -= offset;
 830                         }
 831                 }
 832         } else {
 833                 offset = parse_number(optarg);
 834                 size   = 1;
 835         }
 836
 837         add_addr_range(offset, size);
 838 }
 839
 840 static void add_bits_filter(uint64_t mask, uint64_t bits)
 841 {
 842         if (nr_bit_filters >= MAX_BIT_FILTERS)
 843                 fatal("too much bit filters\n");
 844
 845         opt_mask[nr_bit_filters] = mask;
 846         opt_bits[nr_bit_filters] = bits;
 847         nr_bit_filters++;
 848 }
 849
 850 static uint64_t parse_flag_name(const char *str, int len)
 851 {
 852         size_t i;
 853
 854         if (!*str || !len)
 855                 return 0;
 856
 857         if (len <= 8 && !strncmp(str, "compound", len))
 858                 return BITS_COMPOUND;
 859
 860         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 861                 if (!page_flag_names[i])
 862                         continue;
 863                 if (!strncmp(str, page_flag_names[i] + 2, len))
 864                         return 1ULL << i;
 865         }
 866
 867         return parse_number(str);
 868 }
 869
 870 static uint64_t parse_flag_names(const char *str, int all)
 871 {
 872         const char *p    = str;
 873         uint64_t   flags = 0;
 874
 875         while (1) {
 876                 if (*p == ',' || *p == '=' || *p == '\0') {
 877                         if ((*str != '~') || (*str == '~' && all && *++str))
 878                                 flags |= parse_flag_name(str, p - str);
 879                         if (*p != ',')
 880                                 break;
 881                         str = p + 1;
 882                 }
 883                 p++;
 884         }
 885
 886         return flags;
 887 }
 888
 889 static void parse_bits_mask(const char *optarg)
 890 {
 891         uint64_t mask;
 892         uint64_t bits;
 893         const char *p;
 894
 895         p = strchr(optarg, '=');
 896         if (p == optarg) {
 897                 mask = KPF_ALL_BITS;
 898                 bits = parse_flag_names(p + 1, 0);
 899         } else if (p) {
 900                 mask = parse_flag_names(optarg, 0);
 901                 bits = parse_flag_names(p + 1, 0);
 902         } else if (strchr(optarg, '~')) {
 903                 mask = parse_flag_names(optarg, 1);
 904                 bits = parse_flag_names(optarg, 0);
 905         } else {
 906                 mask = parse_flag_names(optarg, 0);
 907                 bits = KPF_ALL_BITS;
 908         }
 909
 910         add_bits_filter(mask, bits);
 911 }
 912
 913 static void describe_flags(const char *optarg)
 914 {
 915         uint64_t flags = parse_flag_names(optarg, 0);
 916
 917         printf("0x%016llx\t%s\t%s\n",
 918                 (unsigned long long)flags,
 919                 page_flag_name(flags),
 920                 page_flag_longname(flags));
 921 }
 922
 923 static const struct option opts[] = {
 924         { "raw"       , 0, NULL, 'r' },
 925         { "pid"       , 1, NULL, 'p' },
 926         { "file"      , 1, NULL, 'f' },
 927         { "addr"      , 1, NULL, 'a' },
 928         { "bits"      , 1, NULL, 'b' },
 929         { "describe"  , 1, NULL, 'd' },
 930         { "list"      , 0, NULL, 'l' },
 931         { "list-each" , 0, NULL, 'L' },
 932         { "no-summary", 0, NULL, 'N' },
 933         { "hwpoison"  , 0, NULL, 'X' },
 934         { "unpoison"  , 0, NULL, 'x' },
 935         { "help"      , 0, NULL, 'h' },
 936         { NULL        , 0, NULL, 0 }
 937 };
 938
 939 int main(int argc, char *argv[])
 940 {
 941         int c;
 942
 943         page_size = getpagesize();
 944
 945         while ((c = getopt_long(argc, argv,
 946                                 "rp:f:a:b:d:lLNXxh", opts, NULL)) != -1) {
 947                 switch (c) {
 948                 case 'r':
 949                         opt_raw = 1;
 950                         break;
 951                 case 'p':
 952                         parse_pid(optarg);
 953                         break;
 954                 case 'f':
 955                         parse_file(optarg);
 956                         break;
 957                 case 'a':
 958                         parse_addr_range(optarg);
 959                         break;
 960                 case 'b':
 961                         parse_bits_mask(optarg);
 962                         break;
 963                 case 'd':
 964                         describe_flags(optarg);
 965                         exit(0);
 966                 case 'l':
 967                         opt_list = 1;
 968                         break;
 969                 case 'L':
 970                         opt_list = 2;
 971                         break;
 972                 case 'N':
 973                         opt_no_summary = 1;
 974                         break;
 975                 case 'X':
 976                         opt_hwpoison = 1;
 977                         prepare_hwpoison_fd();
 978                         break;
 979                 case 'x':
 980                         opt_unpoison = 1;
 981                         prepare_hwpoison_fd();
 982                         break;
 983                 case 'h':
 984                         usage();
 985                         exit(0);
 986                 default:
 987                         usage();
 988                         exit(1);
 989                 }
 990         }
 991
 992         if (opt_list && opt_pid)
 993                 printf("voffset\t");
 994         if (opt_list == 1)
 995                 printf("offset\tlen\tflags\n");
 996         if (opt_list == 2)
 997                 printf("offset\tflags\n");
 998
 999         walk_addr_ranges();
1000
1001         if (opt_list == 1)
1002                 show_page_range(0, 0, 0);  /* drain the buffer */
1003
1004         if (opt_no_summary)
1005                 return 0;
1006
1007         if (opt_list)
1008                 printf("\n\n");
1009
1010         show_summary();
1011
1012         return 0;
1013 }