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/linux/magic.h"
  38 #include "../../include/linux/kernel-page-flags.h"
  39
  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)        (((u64) (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_PRESENT          PM_STATUS(4LL)
  67 #define PM_SWAP             PM_STATUS(2LL)
  68
  69
  70 /*
  71  * kernel page flags
  72  */
  73
  74 #define KPF_BYTES               8
  75 #define PROC_KPAGEFLAGS         "/proc/kpageflags"
  76
  77 /* [32-] kernel hacking assistances */
  78 #define KPF_RESERVED            32
  79 #define KPF_MLOCKED             33
  80 #define KPF_MAPPEDTODISK        34
  81 #define KPF_PRIVATE             35
  82 #define KPF_PRIVATE_2           36
  83 #define KPF_OWNER_PRIVATE       37
  84 #define KPF_ARCH                38
  85 #define KPF_UNCACHED            39
  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
 127         [KPF_RESERVED]          = "r:reserved",
 128         [KPF_MLOCKED]           = "m:mlocked",
 129         [KPF_MAPPEDTODISK]      = "d:mappedtodisk",
 130         [KPF_PRIVATE]           = "P:private",
 131         [KPF_PRIVATE_2]         = "p:private_2",
 132         [KPF_OWNER_PRIVATE]     = "O:owner_private",
 133         [KPF_ARCH]              = "h:arch",
 134         [KPF_UNCACHED]          = "c:uncached",
 135
 136         [KPF_READAHEAD]         = "I:readahead",
 137         [KPF_SLOB_FREE]         = "P:slob_free",
 138         [KPF_SLUB_FROZEN]       = "A:slub_frozen",
 139         [KPF_SLUB_DEBUG]        = "E:slub_debug",
 140 };
 141
 142
 143 static const char * const debugfs_known_mountpoints[] = {
 144         "/sys/kernel/debug",
 145         "/debug",
 146         0,
 147 };
 148
 149 /*
 150  * data structures
 151  */
 152
 153 static int              opt_raw;        /* for kernel developers */
 154 static int              opt_list;       /* list pages (in ranges) */
 155 static int              opt_no_summary; /* don't show summary */
 156 static pid_t            opt_pid;        /* process to walk */
 157
 158 #define MAX_ADDR_RANGES 1024
 159 static int              nr_addr_ranges;
 160 static unsigned long    opt_offset[MAX_ADDR_RANGES];
 161 static unsigned long    opt_size[MAX_ADDR_RANGES];
 162
 163 #define MAX_VMAS        10240
 164 static int              nr_vmas;
 165 static unsigned long    pg_start[MAX_VMAS];
 166 static unsigned long    pg_end[MAX_VMAS];
 167
 168 #define MAX_BIT_FILTERS 64
 169 static int              nr_bit_filters;
 170 static uint64_t         opt_mask[MAX_BIT_FILTERS];
 171 static uint64_t         opt_bits[MAX_BIT_FILTERS];
 172
 173 static int              page_size;
 174
 175 static int              pagemap_fd;
 176 static int              kpageflags_fd;
 177
 178 static int              opt_hwpoison;
 179 static int              opt_unpoison;
 180
 181 static char             hwpoison_debug_fs[MAX_PATH+1];
 182 static int              hwpoison_inject_fd;
 183 static int              hwpoison_forget_fd;
 184
 185 #define HASH_SHIFT      13
 186 #define HASH_SIZE       (1 << HASH_SHIFT)
 187 #define HASH_MASK       (HASH_SIZE - 1)
 188 #define HASH_KEY(flags) (flags & HASH_MASK)
 189
 190 static unsigned long    total_pages;
 191 static unsigned long    nr_pages[HASH_SIZE];
 192 static uint64_t         page_flags[HASH_SIZE];
 193
 194
 195 /*
 196  * helper functions
 197  */
 198
 199 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 200
 201 #define min_t(type, x, y) ({                    \
 202         type __min1 = (x);                      \
 203         type __min2 = (y);                      \
 204         __min1 < __min2 ? __min1 : __min2; })
 205
 206 #define max_t(type, x, y) ({                    \
 207         type __max1 = (x);                      \
 208         type __max2 = (y);                      \
 209         __max1 > __max2 ? __max1 : __max2; })
 210
 211 static unsigned long pages2mb(unsigned long pages)
 212 {
 213         return (pages * page_size) >> 20;
 214 }
 215
 216 static void fatal(const char *x, ...)
 217 {
 218         va_list ap;
 219
 220         va_start(ap, x);
 221         vfprintf(stderr, x, ap);
 222         va_end(ap);
 223         exit(EXIT_FAILURE);
 224 }
 225
 226 static int checked_open(const char *pathname, int flags)
 227 {
 228         int fd = open(pathname, flags);
 229
 230         if (fd < 0) {
 231                 perror(pathname);
 232                 exit(EXIT_FAILURE);
 233         }
 234
 235         return fd;
 236 }
 237
 238 /*
 239  * pagemap/kpageflags routines
 240  */
 241
 242 static unsigned long do_u64_read(int fd, char *name,
 243                                  uint64_t *buf,
 244                                  unsigned long index,
 245                                  unsigned long count)
 246 {
 247         long bytes;
 248
 249         if (index > ULONG_MAX / 8)
 250                 fatal("index overflow: %lu\n", index);
 251
 252         if (lseek(fd, index * 8, SEEK_SET) < 0) {
 253                 perror(name);
 254                 exit(EXIT_FAILURE);
 255         }
 256
 257         bytes = read(fd, buf, count * 8);
 258         if (bytes < 0) {
 259                 perror(name);
 260                 exit(EXIT_FAILURE);
 261         }
 262         if (bytes % 8)
 263                 fatal("partial read: %lu bytes\n", bytes);
 264
 265         return bytes / 8;
 266 }
 267
 268 static unsigned long kpageflags_read(uint64_t *buf,
 269                                      unsigned long index,
 270                                      unsigned long pages)
 271 {
 272         return do_u64_read(kpageflags_fd, PROC_KPAGEFLAGS, buf, index, pages);
 273 }
 274
 275 static unsigned long pagemap_read(uint64_t *buf,
 276                                   unsigned long index,
 277                                   unsigned long pages)
 278 {
 279         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 280 }
 281
 282 static unsigned long pagemap_pfn(uint64_t val)
 283 {
 284         unsigned long pfn;
 285
 286         if (val & PM_PRESENT)
 287                 pfn = PM_PFRAME(val);
 288         else
 289                 pfn = 0;
 290
 291         return pfn;
 292 }
 293
 294
 295 /*
 296  * page flag names
 297  */
 298
 299 static char *page_flag_name(uint64_t flags)
 300 {
 301         static char buf[65];
 302         int present;
 303         size_t i, j;
 304
 305         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 306                 present = (flags >> i) & 1;
 307                 if (!page_flag_names[i]) {
 308                         if (present)
 309                                 fatal("unknown flag bit %d\n", i);
 310                         continue;
 311                 }
 312                 buf[j++] = present ? page_flag_names[i][0] : '_';
 313         }
 314
 315         return buf;
 316 }
 317
 318 static char *page_flag_longname(uint64_t flags)
 319 {
 320         static char buf[1024];
 321         size_t i, n;
 322
 323         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 324                 if (!page_flag_names[i])
 325                         continue;
 326                 if ((flags >> i) & 1)
 327                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 328                                         page_flag_names[i] + 2);
 329         }
 330         if (n)
 331                 n--;
 332         buf[n] = '\0';
 333
 334         return buf;
 335 }
 336
 337
 338 /*
 339  * page list and summary
 340  */
 341
 342 static void show_page_range(unsigned long voffset,
 343                             unsigned long offset, uint64_t flags)
 344 {
 345         static uint64_t      flags0;
 346         static unsigned long voff;
 347         static unsigned long index;
 348         static unsigned long count;
 349
 350         if (flags == flags0 && offset == index + count &&
 351             (!opt_pid || voffset == voff + count)) {
 352                 count++;
 353                 return;
 354         }
 355
 356         if (count) {
 357                 if (opt_pid)
 358                         printf("%lx\t", voff);
 359                 printf("%lx\t%lx\t%s\n",
 360                                 index, count, page_flag_name(flags0));
 361         }
 362
 363         flags0 = flags;
 364         index  = offset;
 365         voff   = voffset;
 366         count  = 1;
 367 }
 368
 369 static void show_page(unsigned long voffset,
 370                       unsigned long offset, uint64_t flags)
 371 {
 372         if (opt_pid)
 373                 printf("%lx\t", voffset);
 374         printf("%lx\t%s\n", offset, page_flag_name(flags));
 375 }
 376
 377 static void show_summary(void)
 378 {
 379         size_t i;
 380
 381         printf("             flags\tpage-count       MB"
 382                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 383
 384         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 385                 if (nr_pages[i])
 386                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 387                                 (unsigned long long)page_flags[i],
 388                                 nr_pages[i],
 389                                 pages2mb(nr_pages[i]),
 390                                 page_flag_name(page_flags[i]),
 391                                 page_flag_longname(page_flags[i]));
 392         }
 393
 394         printf("             total\t%10lu %8lu\n",
 395                         total_pages, pages2mb(total_pages));
 396 }
 397
 398
 399 /*
 400  * page flag filters
 401  */
 402
 403 static int bit_mask_ok(uint64_t flags)
 404 {
 405         int i;
 406
 407         for (i = 0; i < nr_bit_filters; i++) {
 408                 if (opt_bits[i] == KPF_ALL_BITS) {
 409                         if ((flags & opt_mask[i]) == 0)
 410                                 return 0;
 411                 } else {
 412                         if ((flags & opt_mask[i]) != opt_bits[i])
 413                                 return 0;
 414                 }
 415         }
 416
 417         return 1;
 418 }
 419
 420 static uint64_t expand_overloaded_flags(uint64_t flags)
 421 {
 422         /* SLOB/SLUB overload several page flags */
 423         if (flags & BIT(SLAB)) {
 424                 if (flags & BIT(PRIVATE))
 425                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 426                 if (flags & BIT(ACTIVE))
 427                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 428                 if (flags & BIT(ERROR))
 429                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 430         }
 431
 432         /* PG_reclaim is overloaded as PG_readahead in the read path */
 433         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 434                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 435
 436         return flags;
 437 }
 438
 439 static uint64_t well_known_flags(uint64_t flags)
 440 {
 441         /* hide flags intended only for kernel hacker */
 442         flags &= ~KPF_HACKERS_BITS;
 443
 444         /* hide non-hugeTLB compound pages */
 445         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 446                 flags &= ~BITS_COMPOUND;
 447
 448         return flags;
 449 }
 450
 451 static uint64_t kpageflags_flags(uint64_t flags)
 452 {
 453         flags = expand_overloaded_flags(flags);
 454
 455         if (!opt_raw)
 456                 flags = well_known_flags(flags);
 457
 458         return flags;
 459 }
 460
 461 /* verify that a mountpoint is actually a debugfs instance */
 462 static int debugfs_valid_mountpoint(const char *debugfs)
 463 {
 464         struct statfs st_fs;
 465
 466         if (statfs(debugfs, &st_fs) < 0)
 467                 return -ENOENT;
 468         else if (st_fs.f_type != (long) DEBUGFS_MAGIC)
 469                 return -ENOENT;
 470
 471         return 0;
 472 }
 473
 474 /* find the path to the mounted debugfs */
 475 static const char *debugfs_find_mountpoint(void)
 476 {
 477         const char *const *ptr;
 478         char type[100];
 479         FILE *fp;
 480
 481         ptr = debugfs_known_mountpoints;
 482         while (*ptr) {
 483                 if (debugfs_valid_mountpoint(*ptr) == 0) {
 484                         strcpy(hwpoison_debug_fs, *ptr);
 485                         return hwpoison_debug_fs;
 486                 }
 487                 ptr++;
 488         }
 489
 490         /* give up and parse /proc/mounts */
 491         fp = fopen("/proc/mounts", "r");
 492         if (fp == NULL)
 493                 perror("Can't open /proc/mounts for read");
 494
 495         while (fscanf(fp, "%*s %"
 496                       STR(MAX_PATH)
 497                       "s %99s %*s %*d %*d\n",
 498                       hwpoison_debug_fs, type) == 2) {
 499                 if (strcmp(type, "debugfs") == 0)
 500                         break;
 501         }
 502         fclose(fp);
 503
 504         if (strcmp(type, "debugfs") != 0)
 505                 return NULL;
 506
 507         return hwpoison_debug_fs;
 508 }
 509
 510 /* mount the debugfs somewhere if it's not mounted */
 511
 512 static void debugfs_mount(void)
 513 {
 514         const char *const *ptr;
 515
 516         /* see if it's already mounted */
 517         if (debugfs_find_mountpoint())
 518                 return;
 519
 520         ptr = debugfs_known_mountpoints;
 521         while (*ptr) {
 522                 if (mount(NULL, *ptr, "debugfs", 0, NULL) == 0) {
 523                         /* save the mountpoint */
 524                         strcpy(hwpoison_debug_fs, *ptr);
 525                         break;
 526                 }
 527                 ptr++;
 528         }
 529
 530         if (*ptr == NULL) {
 531                 perror("mount debugfs");
 532                 exit(EXIT_FAILURE);
 533         }
 534 }
 535
 536 /*
 537  * page actions
 538  */
 539
 540 static void prepare_hwpoison_fd(void)
 541 {
 542         char buf[MAX_PATH + 1];
 543
 544         debugfs_mount();
 545
 546         if (opt_hwpoison && !hwpoison_inject_fd) {
 547                 snprintf(buf, MAX_PATH, "%s/hwpoison/corrupt-pfn",
 548                         hwpoison_debug_fs);
 549                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 550         }
 551
 552         if (opt_unpoison && !hwpoison_forget_fd) {
 553                 snprintf(buf, MAX_PATH, "%s/hwpoison/unpoison-pfn",
 554                         hwpoison_debug_fs);
 555                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 556         }
 557 }
 558
 559 static int hwpoison_page(unsigned long offset)
 560 {
 561         char buf[100];
 562         int len;
 563
 564         len = sprintf(buf, "0x%lx\n", offset);
 565         len = write(hwpoison_inject_fd, buf, len);
 566         if (len < 0) {
 567                 perror("hwpoison inject");
 568                 return len;
 569         }
 570         return 0;
 571 }
 572
 573 static int unpoison_page(unsigned long offset)
 574 {
 575         char buf[100];
 576         int len;
 577
 578         len = sprintf(buf, "0x%lx\n", offset);
 579         len = write(hwpoison_forget_fd, buf, len);
 580         if (len < 0) {
 581                 perror("hwpoison forget");
 582                 return len;
 583         }
 584         return 0;
 585 }
 586
 587 /*
 588  * page frame walker
 589  */
 590
 591 static size_t hash_slot(uint64_t flags)
 592 {
 593         size_t k = HASH_KEY(flags);
 594         size_t i;
 595
 596         /* Explicitly reserve slot 0 for flags 0: the following logic
 597          * cannot distinguish an unoccupied slot from slot (flags==0).
 598          */
 599         if (flags == 0)
 600                 return 0;
 601
 602         /* search through the remaining (HASH_SIZE-1) slots */
 603         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 604                 if (!k || k >= ARRAY_SIZE(page_flags))
 605                         k = 1;
 606                 if (page_flags[k] == 0) {
 607                         page_flags[k] = flags;
 608                         return k;
 609                 }
 610                 if (page_flags[k] == flags)
 611                         return k;
 612         }
 613
 614         fatal("hash table full: bump up HASH_SHIFT?\n");
 615         exit(EXIT_FAILURE);
 616 }
 617
 618 static void add_page(unsigned long voffset,
 619                      unsigned long offset, uint64_t flags)
 620 {
 621         flags = kpageflags_flags(flags);
 622
 623         if (!bit_mask_ok(flags))
 624                 return;
 625
 626         if (opt_hwpoison)
 627                 hwpoison_page(offset);
 628         if (opt_unpoison)
 629                 unpoison_page(offset);
 630
 631         if (opt_list == 1)
 632                 show_page_range(voffset, offset, flags);
 633         else if (opt_list == 2)
 634                 show_page(voffset, offset, flags);
 635
 636         nr_pages[hash_slot(flags)]++;
 637         total_pages++;
 638 }
 639
 640 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 641 static void walk_pfn(unsigned long voffset,
 642                      unsigned long index,
 643                      unsigned long count)
 644 {
 645         uint64_t buf[KPAGEFLAGS_BATCH];
 646         unsigned long batch;
 647         unsigned long pages;
 648         unsigned long i;
 649
 650         while (count) {
 651                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 652                 pages = kpageflags_read(buf, index, batch);
 653                 if (pages == 0)
 654                         break;
 655
 656                 for (i = 0; i < pages; i++)
 657                         add_page(voffset + i, index + i, buf[i]);
 658
 659                 index += pages;
 660                 count -= pages;
 661         }
 662 }
 663
 664 #define PAGEMAP_BATCH   (64 << 10)
 665 static void walk_vma(unsigned long index, unsigned long count)
 666 {
 667         uint64_t buf[PAGEMAP_BATCH];
 668         unsigned long batch;
 669         unsigned long pages;
 670         unsigned long pfn;
 671         unsigned long i;
 672
 673         while (count) {
 674                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 675                 pages = pagemap_read(buf, index, batch);
 676                 if (pages == 0)
 677                         break;
 678
 679                 for (i = 0; i < pages; i++) {
 680                         pfn = pagemap_pfn(buf[i]);
 681                         if (pfn)
 682                                 walk_pfn(index + i, pfn, 1);
 683                 }
 684
 685                 index += pages;
 686                 count -= pages;
 687         }
 688 }
 689
 690 static void walk_task(unsigned long index, unsigned long count)
 691 {
 692         const unsigned long end = index + count;
 693         unsigned long start;
 694         int i = 0;
 695
 696         while (index < end) {
 697
 698                 while (pg_end[i] <= index)
 699                         if (++i >= nr_vmas)
 700                                 return;
 701                 if (pg_start[i] >= end)
 702                         return;
 703
 704                 start = max_t(unsigned long, pg_start[i], index);
 705                 index = min_t(unsigned long, pg_end[i], end);
 706
 707                 assert(start < index);
 708                 walk_vma(start, index - start);
 709         }
 710 }
 711
 712 static void add_addr_range(unsigned long offset, unsigned long size)
 713 {
 714         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 715                 fatal("too many addr ranges\n");
 716
 717         opt_offset[nr_addr_ranges] = offset;
 718         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 719         nr_addr_ranges++;
 720 }
 721
 722 static void walk_addr_ranges(void)
 723 {
 724         int i;
 725
 726         kpageflags_fd = checked_open(PROC_KPAGEFLAGS, O_RDONLY);
 727
 728         if (!nr_addr_ranges)
 729                 add_addr_range(0, ULONG_MAX);
 730
 731         for (i = 0; i < nr_addr_ranges; i++)
 732                 if (!opt_pid)
 733                         walk_pfn(0, opt_offset[i], opt_size[i]);
 734                 else
 735                         walk_task(opt_offset[i], opt_size[i]);
 736
 737         close(kpageflags_fd);
 738 }
 739
 740
 741 /*
 742  * user interface
 743  */
 744
 745 static const char *page_flag_type(uint64_t flag)
 746 {
 747         if (flag & KPF_HACKERS_BITS)
 748                 return "(r)";
 749         if (flag & KPF_OVERLOADED_BITS)
 750                 return "(o)";
 751         return "   ";
 752 }
 753
 754 static void usage(void)
 755 {
 756         size_t i, j;
 757
 758         printf(
 759 "page-types [options]\n"
 760 "            -r|--raw                   Raw mode, for kernel developers\n"
 761 "            -d|--describe flags        Describe flags\n"
 762 "            -a|--addr    addr-spec     Walk a range of pages\n"
 763 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 764 "            -p|--pid     pid           Walk process address space\n"
 765 #if 0 /* planned features */
 766 "            -f|--file    filename      Walk file address space\n"
 767 #endif
 768 "            -l|--list                  Show page details in ranges\n"
 769 "            -L|--list-each             Show page details one by one\n"
 770 "            -N|--no-summary            Don't show summary info\n"
 771 "            -X|--hwpoison              hwpoison pages\n"
 772 "            -x|--unpoison              unpoison pages\n"
 773 "            -h|--help                  Show this usage message\n"
 774 "flags:\n"
 775 "            0x10                       bitfield format, e.g.\n"
 776 "            anon                       bit-name, e.g.\n"
 777 "            0x10,anon                  comma-separated list, e.g.\n"
 778 "addr-spec:\n"
 779 "            N                          one page at offset N (unit: pages)\n"
 780 "            N+M                        pages range from N to N+M-1\n"
 781 "            N,M                        pages range from N to M-1\n"
 782 "            N,                         pages range from N to end\n"
 783 "            ,M                         pages range from 0 to M-1\n"
 784 "bits-spec:\n"
 785 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 786 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 787 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 788 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 789 "bit-names:\n"
 790         );
 791
 792         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 793                 if (!page_flag_names[i])
 794                         continue;
 795                 printf("%16s%s", page_flag_names[i] + 2,
 796                                  page_flag_type(1ULL << i));
 797                 if (++j > 3) {
 798                         j = 0;
 799                         putchar('\n');
 800                 }
 801         }
 802         printf("\n                                   "
 803                 "(r) raw mode bits  (o) overloaded bits\n");
 804 }
 805
 806 static unsigned long long parse_number(const char *str)
 807 {
 808         unsigned long long n;
 809
 810         n = strtoll(str, NULL, 0);
 811
 812         if (n == 0 && str[0] != '0')
 813                 fatal("invalid name or number: %s\n", str);
 814
 815         return n;
 816 }
 817
 818 static void parse_pid(const char *str)
 819 {
 820         FILE *file;
 821         char buf[5000];
 822
 823         opt_pid = parse_number(str);
 824
 825         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 826         pagemap_fd = checked_open(buf, O_RDONLY);
 827
 828         sprintf(buf, "/proc/%d/maps", opt_pid);
 829         file = fopen(buf, "r");
 830         if (!file) {
 831                 perror(buf);
 832                 exit(EXIT_FAILURE);
 833         }
 834
 835         while (fgets(buf, sizeof(buf), file) != NULL) {
 836                 unsigned long vm_start;
 837                 unsigned long vm_end;
 838                 unsigned long long pgoff;
 839                 int major, minor;
 840                 char r, w, x, s;
 841                 unsigned long ino;
 842                 int n;
 843
 844                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 845                            &vm_start,
 846                            &vm_end,
 847                            &r, &w, &x, &s,
 848                            &pgoff,
 849                            &major, &minor,
 850                            &ino);
 851                 if (n < 10) {
 852                         fprintf(stderr, "unexpected line: %s\n", buf);
 853                         continue;
 854                 }
 855                 pg_start[nr_vmas] = vm_start / page_size;
 856                 pg_end[nr_vmas] = vm_end / page_size;
 857                 if (++nr_vmas >= MAX_VMAS) {
 858                         fprintf(stderr, "too many VMAs\n");
 859                         break;
 860                 }
 861         }
 862         fclose(file);
 863 }
 864
 865 static void parse_file(const char *name)
 866 {
 867 }
 868
 869 static void parse_addr_range(const char *optarg)
 870 {
 871         unsigned long offset;
 872         unsigned long size;
 873         char *p;
 874
 875         p = strchr(optarg, ',');
 876         if (!p)
 877                 p = strchr(optarg, '+');
 878
 879         if (p == optarg) {
 880                 offset = 0;
 881                 size   = parse_number(p + 1);
 882         } else if (p) {
 883                 offset = parse_number(optarg);
 884                 if (p[1] == '\0')
 885                         size = ULONG_MAX;
 886                 else {
 887                         size = parse_number(p + 1);
 888                         if (*p == ',') {
 889                                 if (size < offset)
 890                                         fatal("invalid range: %lu,%lu\n",
 891                                                         offset, size);
 892                                 size -= offset;
 893                         }
 894                 }
 895         } else {
 896                 offset = parse_number(optarg);
 897                 size   = 1;
 898         }
 899
 900         add_addr_range(offset, size);
 901 }
 902
 903 static void add_bits_filter(uint64_t mask, uint64_t bits)
 904 {
 905         if (nr_bit_filters >= MAX_BIT_FILTERS)
 906                 fatal("too much bit filters\n");
 907
 908         opt_mask[nr_bit_filters] = mask;
 909         opt_bits[nr_bit_filters] = bits;
 910         nr_bit_filters++;
 911 }
 912
 913 static uint64_t parse_flag_name(const char *str, int len)
 914 {
 915         size_t i;
 916
 917         if (!*str || !len)
 918                 return 0;
 919
 920         if (len <= 8 && !strncmp(str, "compound", len))
 921                 return BITS_COMPOUND;
 922
 923         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 924                 if (!page_flag_names[i])
 925                         continue;
 926                 if (!strncmp(str, page_flag_names[i] + 2, len))
 927                         return 1ULL << i;
 928         }
 929
 930         return parse_number(str);
 931 }
 932
 933 static uint64_t parse_flag_names(const char *str, int all)
 934 {
 935         const char *p    = str;
 936         uint64_t   flags = 0;
 937
 938         while (1) {
 939                 if (*p == ',' || *p == '=' || *p == '\0') {
 940                         if ((*str != '~') || (*str == '~' && all && *++str))
 941                                 flags |= parse_flag_name(str, p - str);
 942                         if (*p != ',')
 943                                 break;
 944                         str = p + 1;
 945                 }
 946                 p++;
 947         }
 948
 949         return flags;
 950 }
 951
 952 static void parse_bits_mask(const char *optarg)
 953 {
 954         uint64_t mask;
 955         uint64_t bits;
 956         const char *p;
 957
 958         p = strchr(optarg, '=');
 959         if (p == optarg) {
 960                 mask = KPF_ALL_BITS;
 961                 bits = parse_flag_names(p + 1, 0);
 962         } else if (p) {
 963                 mask = parse_flag_names(optarg, 0);
 964                 bits = parse_flag_names(p + 1, 0);
 965         } else if (strchr(optarg, '~')) {
 966                 mask = parse_flag_names(optarg, 1);
 967                 bits = parse_flag_names(optarg, 0);
 968         } else {
 969                 mask = parse_flag_names(optarg, 0);
 970                 bits = KPF_ALL_BITS;
 971         }
 972
 973         add_bits_filter(mask, bits);
 974 }
 975
 976 static void describe_flags(const char *optarg)
 977 {
 978         uint64_t flags = parse_flag_names(optarg, 0);
 979
 980         printf("0x%016llx\t%s\t%s\n",
 981                 (unsigned long long)flags,
 982                 page_flag_name(flags),
 983                 page_flag_longname(flags));
 984 }
 985
 986 static const struct option opts[] = {
 987         { "raw"       , 0, NULL, 'r' },
 988         { "pid"       , 1, NULL, 'p' },
 989         { "file"      , 1, NULL, 'f' },
 990         { "addr"      , 1, NULL, 'a' },
 991         { "bits"      , 1, NULL, 'b' },
 992         { "describe"  , 1, NULL, 'd' },
 993         { "list"      , 0, NULL, 'l' },
 994         { "list-each" , 0, NULL, 'L' },
 995         { "no-summary", 0, NULL, 'N' },
 996         { "hwpoison"  , 0, NULL, 'X' },
 997         { "unpoison"  , 0, NULL, 'x' },
 998         { "help"      , 0, NULL, 'h' },
 999         { NULL        , 0, NULL, 0 }
1000 };
1001
1002 int main(int argc, char *argv[])
1003 {
1004         int c;
1005
1006         page_size = getpagesize();
1007
1008         while ((c = getopt_long(argc, argv,
1009                                 "rp:f:a:b:d:lLNXxh", opts, NULL)) != -1) {
1010                 switch (c) {
1011                 case 'r':
1012                         opt_raw = 1;
1013                         break;
1014                 case 'p':
1015                         parse_pid(optarg);
1016                         break;
1017                 case 'f':
1018                         parse_file(optarg);
1019                         break;
1020                 case 'a':
1021                         parse_addr_range(optarg);
1022                         break;
1023                 case 'b':
1024                         parse_bits_mask(optarg);
1025                         break;
1026                 case 'd':
1027                         describe_flags(optarg);
1028                         exit(0);
1029                 case 'l':
1030                         opt_list = 1;
1031                         break;
1032                 case 'L':
1033                         opt_list = 2;
1034                         break;
1035                 case 'N':
1036                         opt_no_summary = 1;
1037                         break;
1038                 case 'X':
1039                         opt_hwpoison = 1;
1040                         prepare_hwpoison_fd();
1041                         break;
1042                 case 'x':
1043                         opt_unpoison = 1;
1044                         prepare_hwpoison_fd();
1045                         break;
1046                 case 'h':
1047                         usage();
1048                         exit(0);
1049                 default:
1050                         usage();
1051                         exit(1);
1052                 }
1053         }
1054
1055         if (opt_list && opt_pid)
1056                 printf("voffset\t");
1057         if (opt_list == 1)
1058                 printf("offset\tlen\tflags\n");
1059         if (opt_list == 2)
1060                 printf("offset\tflags\n");
1061
1062         walk_addr_ranges();
1063
1064         if (opt_list == 1)
1065                 show_page_range(0, 0, 0);  /* drain the buffer */
1066
1067         if (opt_no_summary)
1068                 return 0;
1069
1070         if (opt_list)
1071                 printf("\n\n");
1072
1073         show_summary();
1074
1075         return 0;
1076 }