Documentation/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
  36
  37 /*
  38  * pagemap kernel ABI bits
  39  */
  40
  41 #define PM_ENTRY_BYTES      sizeof(uint64_t)
  42 #define PM_STATUS_BITS      3
  43 #define PM_STATUS_OFFSET    (64 - PM_STATUS_BITS)
  44 #define PM_STATUS_MASK      (((1LL << PM_STATUS_BITS) - 1) << PM_STATUS_OFFSET)
  45 #define PM_STATUS(nr)       (((nr) << PM_STATUS_OFFSET) & PM_STATUS_MASK)
  46 #define PM_PSHIFT_BITS      6
  47 #define PM_PSHIFT_OFFSET    (PM_STATUS_OFFSET - PM_PSHIFT_BITS)
  48 #define PM_PSHIFT_MASK      (((1LL << PM_PSHIFT_BITS) - 1) << PM_PSHIFT_OFFSET)
  49 #define PM_PSHIFT(x)        (((u64) (x) << PM_PSHIFT_OFFSET) & PM_PSHIFT_MASK)
  50 #define PM_PFRAME_MASK      ((1LL << PM_PSHIFT_OFFSET) - 1)
  51 #define PM_PFRAME(x)        ((x) & PM_PFRAME_MASK)
  52
  53 #define PM_PRESENT          PM_STATUS(4LL)
  54 #define PM_SWAP             PM_STATUS(2LL)
  55
  56
  57 /*
  58  * kernel page flags
  59  */
  60
  61 #define KPF_BYTES               8
  62 #define PROC_KPAGEFLAGS         "/proc/kpageflags"
  63
  64 /* copied from kpageflags_read() */
  65 #define KPF_LOCKED              0
  66 #define KPF_ERROR               1
  67 #define KPF_REFERENCED          2
  68 #define KPF_UPTODATE            3
  69 #define KPF_DIRTY               4
  70 #define KPF_LRU                 5
  71 #define KPF_ACTIVE              6
  72 #define KPF_SLAB                7
  73 #define KPF_WRITEBACK           8
  74 #define KPF_RECLAIM             9
  75 #define KPF_BUDDY               10
  76
  77 /* [11-20] new additions in 2.6.31 */
  78 #define KPF_MMAP                11
  79 #define KPF_ANON                12
  80 #define KPF_SWAPCACHE           13
  81 #define KPF_SWAPBACKED          14
  82 #define KPF_COMPOUND_HEAD       15
  83 #define KPF_COMPOUND_TAIL       16
  84 #define KPF_HUGE                17
  85 #define KPF_UNEVICTABLE         18
  86 #define KPF_HWPOISON            19
  87 #define KPF_NOPAGE              20
  88 #define KPF_KSM                 21
  89
  90 /* [32-] kernel hacking assistances */
  91 #define KPF_RESERVED            32
  92 #define KPF_MLOCKED             33
  93 #define KPF_MAPPEDTODISK        34
  94 #define KPF_PRIVATE             35
  95 #define KPF_PRIVATE_2           36
  96 #define KPF_OWNER_PRIVATE       37
  97 #define KPF_ARCH                38
  98 #define KPF_UNCACHED            39
  99
 100 /* [48-] take some arbitrary free slots for expanding overloaded flags
 101  * not part of kernel API
 102  */
 103 #define KPF_READAHEAD           48
 104 #define KPF_SLOB_FREE           49
 105 #define KPF_SLUB_FROZEN         50
 106 #define KPF_SLUB_DEBUG          51
 107
 108 #define KPF_ALL_BITS            ((uint64_t)~0ULL)
 109 #define KPF_HACKERS_BITS        (0xffffULL << 32)
 110 #define KPF_OVERLOADED_BITS     (0xffffULL << 48)
 111 #define BIT(name)               (1ULL << KPF_##name)
 112 #define BITS_COMPOUND           (BIT(COMPOUND_HEAD) | BIT(COMPOUND_TAIL))
 113
 114 static const char *page_flag_names[] = {
 115         [KPF_LOCKED]            = "L:locked",
 116         [KPF_ERROR]             = "E:error",
 117         [KPF_REFERENCED]        = "R:referenced",
 118         [KPF_UPTODATE]          = "U:uptodate",
 119         [KPF_DIRTY]             = "D:dirty",
 120         [KPF_LRU]               = "l:lru",
 121         [KPF_ACTIVE]            = "A:active",
 122         [KPF_SLAB]              = "S:slab",
 123         [KPF_WRITEBACK]         = "W:writeback",
 124         [KPF_RECLAIM]           = "I:reclaim",
 125         [KPF_BUDDY]             = "B:buddy",
 126
 127         [KPF_MMAP]              = "M:mmap",
 128         [KPF_ANON]              = "a:anonymous",
 129         [KPF_SWAPCACHE]         = "s:swapcache",
 130         [KPF_SWAPBACKED]        = "b:swapbacked",
 131         [KPF_COMPOUND_HEAD]     = "H:compound_head",
 132         [KPF_COMPOUND_TAIL]     = "T:compound_tail",
 133         [KPF_HUGE]              = "G:huge",
 134         [KPF_UNEVICTABLE]       = "u:unevictable",
 135         [KPF_HWPOISON]          = "X:hwpoison",
 136         [KPF_NOPAGE]            = "n:nopage",
 137         [KPF_KSM]               = "x:ksm",
 138
 139         [KPF_RESERVED]          = "r:reserved",
 140         [KPF_MLOCKED]           = "m:mlocked",
 141         [KPF_MAPPEDTODISK]      = "d:mappedtodisk",
 142         [KPF_PRIVATE]           = "P:private",
 143         [KPF_PRIVATE_2]         = "p:private_2",
 144         [KPF_OWNER_PRIVATE]     = "O:owner_private",
 145         [KPF_ARCH]              = "h:arch",
 146         [KPF_UNCACHED]          = "c:uncached",
 147
 148         [KPF_READAHEAD]         = "I:readahead",
 149         [KPF_SLOB_FREE]         = "P:slob_free",
 150         [KPF_SLUB_FROZEN]       = "A:slub_frozen",
 151         [KPF_SLUB_DEBUG]        = "E:slub_debug",
 152 };
 153
 154
 155 /*
 156  * data structures
 157  */
 158
 159 static int              opt_raw;        /* for kernel developers */
 160 static int              opt_list;       /* list pages (in ranges) */
 161 static int              opt_no_summary; /* don't show summary */
 162 static pid_t            opt_pid;        /* process to walk */
 163
 164 #define MAX_ADDR_RANGES 1024
 165 static int              nr_addr_ranges;
 166 static unsigned long    opt_offset[MAX_ADDR_RANGES];
 167 static unsigned long    opt_size[MAX_ADDR_RANGES];
 168
 169 #define MAX_VMAS        10240
 170 static int              nr_vmas;
 171 static unsigned long    pg_start[MAX_VMAS];
 172 static unsigned long    pg_end[MAX_VMAS];
 173
 174 #define MAX_BIT_FILTERS 64
 175 static int              nr_bit_filters;
 176 static uint64_t         opt_mask[MAX_BIT_FILTERS];
 177 static uint64_t         opt_bits[MAX_BIT_FILTERS];
 178
 179 static int              page_size;
 180
 181 static int              pagemap_fd;
 182 static int              kpageflags_fd;
 183
 184 static int              opt_hwpoison;
 185 static int              opt_unpoison;
 186
 187 static const char       hwpoison_debug_fs[] = "/debug/hwpoison";
 188 static int              hwpoison_inject_fd;
 189 static int              hwpoison_forget_fd;
 190
 191 #define HASH_SHIFT      13
 192 #define HASH_SIZE       (1 << HASH_SHIFT)
 193 #define HASH_MASK       (HASH_SIZE - 1)
 194 #define HASH_KEY(flags) (flags & HASH_MASK)
 195
 196 static unsigned long    total_pages;
 197 static unsigned long    nr_pages[HASH_SIZE];
 198 static uint64_t         page_flags[HASH_SIZE];
 199
 200
 201 /*
 202  * helper functions
 203  */
 204
 205 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 206
 207 #define min_t(type, x, y) ({                    \
 208         type __min1 = (x);                      \
 209         type __min2 = (y);                      \
 210         __min1 < __min2 ? __min1 : __min2; })
 211
 212 #define max_t(type, x, y) ({                    \
 213         type __max1 = (x);                      \
 214         type __max2 = (y);                      \
 215         __max1 > __max2 ? __max1 : __max2; })
 216
 217 static unsigned long pages2mb(unsigned long pages)
 218 {
 219         return (pages * page_size) >> 20;
 220 }
 221
 222 static void fatal(const char *x, ...)
 223 {
 224         va_list ap;
 225
 226         va_start(ap, x);
 227         vfprintf(stderr, x, ap);
 228         va_end(ap);
 229         exit(EXIT_FAILURE);
 230 }
 231
 232 static int checked_open(const char *pathname, int flags)
 233 {
 234         int fd = open(pathname, flags);
 235
 236         if (fd < 0) {
 237                 perror(pathname);
 238                 exit(EXIT_FAILURE);
 239         }
 240
 241         return fd;
 242 }
 243
 244 /*
 245  * pagemap/kpageflags routines
 246  */
 247
 248 static unsigned long do_u64_read(int fd, char *name,
 249                                  uint64_t *buf,
 250                                  unsigned long index,
 251                                  unsigned long count)
 252 {
 253         long bytes;
 254
 255         if (index > ULONG_MAX / 8)
 256                 fatal("index overflow: %lu\n", index);
 257
 258         if (lseek(fd, index * 8, SEEK_SET) < 0) {
 259                 perror(name);
 260                 exit(EXIT_FAILURE);
 261         }
 262
 263         bytes = read(fd, buf, count * 8);
 264         if (bytes < 0) {
 265                 perror(name);
 266                 exit(EXIT_FAILURE);
 267         }
 268         if (bytes % 8)
 269                 fatal("partial read: %lu bytes\n", bytes);
 270
 271         return bytes / 8;
 272 }
 273
 274 static unsigned long kpageflags_read(uint64_t *buf,
 275                                      unsigned long index,
 276                                      unsigned long pages)
 277 {
 278         return do_u64_read(kpageflags_fd, PROC_KPAGEFLAGS, buf, index, pages);
 279 }
 280
 281 static unsigned long pagemap_read(uint64_t *buf,
 282                                   unsigned long index,
 283                                   unsigned long pages)
 284 {
 285         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 286 }
 287
 288 static unsigned long pagemap_pfn(uint64_t val)
 289 {
 290         unsigned long pfn;
 291
 292         if (val & PM_PRESENT)
 293                 pfn = PM_PFRAME(val);
 294         else
 295                 pfn = 0;
 296
 297         return pfn;
 298 }
 299
 300
 301 /*
 302  * page flag names
 303  */
 304
 305 static char *page_flag_name(uint64_t flags)
 306 {
 307         static char buf[65];
 308         int present;
 309         int i, j;
 310
 311         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 312                 present = (flags >> i) & 1;
 313                 if (!page_flag_names[i]) {
 314                         if (present)
 315                                 fatal("unknown flag bit %d\n", i);
 316                         continue;
 317                 }
 318                 buf[j++] = present ? page_flag_names[i][0] : '_';
 319         }
 320
 321         return buf;
 322 }
 323
 324 static char *page_flag_longname(uint64_t flags)
 325 {
 326         static char buf[1024];
 327         int i, n;
 328
 329         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 330                 if (!page_flag_names[i])
 331                         continue;
 332                 if ((flags >> i) & 1)
 333                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 334                                         page_flag_names[i] + 2);
 335         }
 336         if (n)
 337                 n--;
 338         buf[n] = '\0';
 339
 340         return buf;
 341 }
 342
 343
 344 /*
 345  * page list and summary
 346  */
 347
 348 static void show_page_range(unsigned long voffset,
 349                             unsigned long offset, uint64_t flags)
 350 {
 351         static uint64_t      flags0;
 352         static unsigned long voff;
 353         static unsigned long index;
 354         static unsigned long count;
 355
 356         if (flags == flags0 && offset == index + count &&
 357             (!opt_pid || voffset == voff + count)) {
 358                 count++;
 359                 return;
 360         }
 361
 362         if (count) {
 363                 if (opt_pid)
 364                         printf("%lx\t", voff);
 365                 printf("%lx\t%lx\t%s\n",
 366                                 index, count, page_flag_name(flags0));
 367         }
 368
 369         flags0 = flags;
 370         index  = offset;
 371         voff   = voffset;
 372         count  = 1;
 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         printf("%lx\t%s\n", offset, page_flag_name(flags));
 381 }
 382
 383 static void show_summary(void)
 384 {
 385         int i;
 386
 387         printf("             flags\tpage-count       MB"
 388                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 389
 390         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 391                 if (nr_pages[i])
 392                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 393                                 (unsigned long long)page_flags[i],
 394                                 nr_pages[i],
 395                                 pages2mb(nr_pages[i]),
 396                                 page_flag_name(page_flags[i]),
 397                                 page_flag_longname(page_flags[i]));
 398         }
 399
 400         printf("             total\t%10lu %8lu\n",
 401                         total_pages, pages2mb(total_pages));
 402 }
 403
 404
 405 /*
 406  * page flag filters
 407  */
 408
 409 static int bit_mask_ok(uint64_t flags)
 410 {
 411         int i;
 412
 413         for (i = 0; i < nr_bit_filters; i++) {
 414                 if (opt_bits[i] == KPF_ALL_BITS) {
 415                         if ((flags & opt_mask[i]) == 0)
 416                                 return 0;
 417                 } else {
 418                         if ((flags & opt_mask[i]) != opt_bits[i])
 419                                 return 0;
 420                 }
 421         }
 422
 423         return 1;
 424 }
 425
 426 static uint64_t expand_overloaded_flags(uint64_t flags)
 427 {
 428         /* SLOB/SLUB overload several page flags */
 429         if (flags & BIT(SLAB)) {
 430                 if (flags & BIT(PRIVATE))
 431                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 432                 if (flags & BIT(ACTIVE))
 433                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 434                 if (flags & BIT(ERROR))
 435                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 436         }
 437
 438         /* PG_reclaim is overloaded as PG_readahead in the read path */
 439         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 440                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 441
 442         return flags;
 443 }
 444
 445 static uint64_t well_known_flags(uint64_t flags)
 446 {
 447         /* hide flags intended only for kernel hacker */
 448         flags &= ~KPF_HACKERS_BITS;
 449
 450         /* hide non-hugeTLB compound pages */
 451         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 452                 flags &= ~BITS_COMPOUND;
 453
 454         return flags;
 455 }
 456
 457 static uint64_t kpageflags_flags(uint64_t flags)
 458 {
 459         flags = expand_overloaded_flags(flags);
 460
 461         if (!opt_raw)
 462                 flags = well_known_flags(flags);
 463
 464         return flags;
 465 }
 466
 467 /*
 468  * page actions
 469  */
 470
 471 static void prepare_hwpoison_fd(void)
 472 {
 473         char buf[100];
 474
 475         if (opt_hwpoison && !hwpoison_inject_fd) {
 476                 sprintf(buf, "%s/corrupt-pfn", hwpoison_debug_fs);
 477                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 478         }
 479
 480         if (opt_unpoison && !hwpoison_forget_fd) {
 481                 sprintf(buf, "%s/renew-pfn", hwpoison_debug_fs);
 482                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 483         }
 484 }
 485
 486 static int hwpoison_page(unsigned long offset)
 487 {
 488         char buf[100];
 489         int len;
 490
 491         len = sprintf(buf, "0x%lx\n", offset);
 492         len = write(hwpoison_inject_fd, buf, len);
 493         if (len < 0) {
 494                 perror("hwpoison inject");
 495                 return len;
 496         }
 497         return 0;
 498 }
 499
 500 static int unpoison_page(unsigned long offset)
 501 {
 502         char buf[100];
 503         int len;
 504
 505         len = sprintf(buf, "0x%lx\n", offset);
 506         len = write(hwpoison_forget_fd, buf, len);
 507         if (len < 0) {
 508                 perror("hwpoison forget");
 509                 return len;
 510         }
 511         return 0;
 512 }
 513
 514 /*
 515  * page frame walker
 516  */
 517
 518 static int hash_slot(uint64_t flags)
 519 {
 520         int k = HASH_KEY(flags);
 521         int i;
 522
 523         /* Explicitly reserve slot 0 for flags 0: the following logic
 524          * cannot distinguish an unoccupied slot from slot (flags==0).
 525          */
 526         if (flags == 0)
 527                 return 0;
 528
 529         /* search through the remaining (HASH_SIZE-1) slots */
 530         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 531                 if (!k || k >= ARRAY_SIZE(page_flags))
 532                         k = 1;
 533                 if (page_flags[k] == 0) {
 534                         page_flags[k] = flags;
 535                         return k;
 536                 }
 537                 if (page_flags[k] == flags)
 538                         return k;
 539         }
 540
 541         fatal("hash table full: bump up HASH_SHIFT?\n");
 542         exit(EXIT_FAILURE);
 543 }
 544
 545 static void add_page(unsigned long voffset,
 546                      unsigned long offset, uint64_t flags)
 547 {
 548         flags = kpageflags_flags(flags);
 549
 550         if (!bit_mask_ok(flags))
 551                 return;
 552
 553         if (opt_hwpoison)
 554                 hwpoison_page(offset);
 555         if (opt_unpoison)
 556                 unpoison_page(offset);
 557
 558         if (opt_list == 1)
 559                 show_page_range(voffset, offset, flags);
 560         else if (opt_list == 2)
 561                 show_page(voffset, offset, flags);
 562
 563         nr_pages[hash_slot(flags)]++;
 564         total_pages++;
 565 }
 566
 567 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 568 static void walk_pfn(unsigned long voffset,
 569                      unsigned long index,
 570                      unsigned long count)
 571 {
 572         uint64_t buf[KPAGEFLAGS_BATCH];
 573         unsigned long batch;
 574         long pages;
 575         unsigned long i;
 576
 577         while (count) {
 578                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 579                 pages = kpageflags_read(buf, index, batch);
 580                 if (pages == 0)
 581                         break;
 582
 583                 for (i = 0; i < pages; i++)
 584                         add_page(voffset + i, index + i, buf[i]);
 585
 586                 index += pages;
 587                 count -= pages;
 588         }
 589 }
 590
 591 #define PAGEMAP_BATCH   (64 << 10)
 592 static void walk_vma(unsigned long index, unsigned long count)
 593 {
 594         uint64_t buf[PAGEMAP_BATCH];
 595         unsigned long batch;
 596         unsigned long pages;
 597         unsigned long pfn;
 598         unsigned long i;
 599
 600         while (count) {
 601                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 602                 pages = pagemap_read(buf, index, batch);
 603                 if (pages == 0)
 604                         break;
 605
 606                 for (i = 0; i < pages; i++) {
 607                         pfn = pagemap_pfn(buf[i]);
 608                         if (pfn)
 609                                 walk_pfn(index + i, pfn, 1);
 610                 }
 611
 612                 index += pages;
 613                 count -= pages;
 614         }
 615 }
 616
 617 static void walk_task(unsigned long index, unsigned long count)
 618 {
 619         const unsigned long end = index + count;
 620         unsigned long start;
 621         int i = 0;
 622
 623         while (index < end) {
 624
 625                 while (pg_end[i] <= index)
 626                         if (++i >= nr_vmas)
 627                                 return;
 628                 if (pg_start[i] >= end)
 629                         return;
 630
 631                 start = max_t(unsigned long, pg_start[i], index);
 632                 index = min_t(unsigned long, pg_end[i], end);
 633
 634                 assert(start < index);
 635                 walk_vma(start, index - start);
 636         }
 637 }
 638
 639 static void add_addr_range(unsigned long offset, unsigned long size)
 640 {
 641         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 642                 fatal("too many addr ranges\n");
 643
 644         opt_offset[nr_addr_ranges] = offset;
 645         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 646         nr_addr_ranges++;
 647 }
 648
 649 static void walk_addr_ranges(void)
 650 {
 651         int i;
 652
 653         kpageflags_fd = checked_open(PROC_KPAGEFLAGS, O_RDONLY);
 654
 655         if (!nr_addr_ranges)
 656                 add_addr_range(0, ULONG_MAX);
 657
 658         for (i = 0; i < nr_addr_ranges; i++)
 659                 if (!opt_pid)
 660                         walk_pfn(0, opt_offset[i], opt_size[i]);
 661                 else
 662                         walk_task(opt_offset[i], opt_size[i]);
 663
 664         close(kpageflags_fd);
 665 }
 666
 667
 668 /*
 669  * user interface
 670  */
 671
 672 static const char *page_flag_type(uint64_t flag)
 673 {
 674         if (flag & KPF_HACKERS_BITS)
 675                 return "(r)";
 676         if (flag & KPF_OVERLOADED_BITS)
 677                 return "(o)";
 678         return "   ";
 679 }
 680
 681 static void usage(void)
 682 {
 683         int i, j;
 684
 685         printf(
 686 "page-types [options]\n"
 687 "            -r|--raw                   Raw mode, for kernel developers\n"
 688 "            -d|--describe flags        Describe flags\n"
 689 "            -a|--addr    addr-spec     Walk a range of pages\n"
 690 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 691 "            -p|--pid     pid           Walk process address space\n"
 692 #if 0 /* planned features */
 693 "            -f|--file    filename      Walk file address space\n"
 694 #endif
 695 "            -l|--list                  Show page details in ranges\n"
 696 "            -L|--list-each             Show page details one by one\n"
 697 "            -N|--no-summary            Don't show summary info\n"
 698 "            -X|--hwpoison              hwpoison pages\n"
 699 "            -x|--unpoison              unpoison pages\n"
 700 "            -h|--help                  Show this usage message\n"
 701 "flags:\n"
 702 "            0x10                       bitfield format, e.g.\n"
 703 "            anon                       bit-name, e.g.\n"
 704 "            0x10,anon                  comma-separated list, e.g.\n"
 705 "addr-spec:\n"
 706 "            N                          one page at offset N (unit: pages)\n"
 707 "            N+M                        pages range from N to N+M-1\n"
 708 "            N,M                        pages range from N to M-1\n"
 709 "            N,                         pages range from N to end\n"
 710 "            ,M                         pages range from 0 to M-1\n"
 711 "bits-spec:\n"
 712 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 713 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 714 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 715 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 716 "bit-names:\n"
 717         );
 718
 719         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 720                 if (!page_flag_names[i])
 721                         continue;
 722                 printf("%16s%s", page_flag_names[i] + 2,
 723                                  page_flag_type(1ULL << i));
 724                 if (++j > 3) {
 725                         j = 0;
 726                         putchar('\n');
 727                 }
 728         }
 729         printf("\n                                   "
 730                 "(r) raw mode bits  (o) overloaded bits\n");
 731 }
 732
 733 static unsigned long long parse_number(const char *str)
 734 {
 735         unsigned long long n;
 736
 737         n = strtoll(str, NULL, 0);
 738
 739         if (n == 0 && str[0] != '0')
 740                 fatal("invalid name or number: %s\n", str);
 741
 742         return n;
 743 }
 744
 745 static void parse_pid(const char *str)
 746 {
 747         FILE *file;
 748         char buf[5000];
 749
 750         opt_pid = parse_number(str);
 751
 752         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 753         pagemap_fd = checked_open(buf, O_RDONLY);
 754
 755         sprintf(buf, "/proc/%d/maps", opt_pid);
 756         file = fopen(buf, "r");
 757         if (!file) {
 758                 perror(buf);
 759                 exit(EXIT_FAILURE);
 760         }
 761
 762         while (fgets(buf, sizeof(buf), file) != NULL) {
 763                 unsigned long vm_start;
 764                 unsigned long vm_end;
 765                 unsigned long long pgoff;
 766                 int major, minor;
 767                 char r, w, x, s;
 768                 unsigned long ino;
 769                 int n;
 770
 771                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 772                            &vm_start,
 773                            &vm_end,
 774                            &r, &w, &x, &s,
 775                            &pgoff,
 776                            &major, &minor,
 777                            &ino);
 778                 if (n < 10) {
 779                         fprintf(stderr, "unexpected line: %s\n", buf);
 780                         continue;
 781                 }
 782                 pg_start[nr_vmas] = vm_start / page_size;
 783                 pg_end[nr_vmas] = vm_end / page_size;
 784                 if (++nr_vmas >= MAX_VMAS) {
 785                         fprintf(stderr, "too many VMAs\n");
 786                         break;
 787                 }
 788         }
 789         fclose(file);
 790 }
 791
 792 static void parse_file(const char *name)
 793 {
 794 }
 795
 796 static void parse_addr_range(const char *optarg)
 797 {
 798         unsigned long offset;
 799         unsigned long size;
 800         char *p;
 801
 802         p = strchr(optarg, ',');
 803         if (!p)
 804                 p = strchr(optarg, '+');
 805
 806         if (p == optarg) {
 807                 offset = 0;
 808                 size   = parse_number(p + 1);
 809         } else if (p) {
 810                 offset = parse_number(optarg);
 811                 if (p[1] == '\0')
 812                         size = ULONG_MAX;
 813                 else {
 814                         size = parse_number(p + 1);
 815                         if (*p == ',') {
 816                                 if (size < offset)
 817                                         fatal("invalid range: %lu,%lu\n",
 818                                                         offset, size);
 819                                 size -= offset;
 820                         }
 821                 }
 822         } else {
 823                 offset = parse_number(optarg);
 824                 size   = 1;
 825         }
 826
 827         add_addr_range(offset, size);
 828 }
 829
 830 static void add_bits_filter(uint64_t mask, uint64_t bits)
 831 {
 832         if (nr_bit_filters >= MAX_BIT_FILTERS)
 833                 fatal("too much bit filters\n");
 834
 835         opt_mask[nr_bit_filters] = mask;
 836         opt_bits[nr_bit_filters] = bits;
 837         nr_bit_filters++;
 838 }
 839
 840 static uint64_t parse_flag_name(const char *str, int len)
 841 {
 842         int i;
 843
 844         if (!*str || !len)
 845                 return 0;
 846
 847         if (len <= 8 && !strncmp(str, "compound", len))
 848                 return BITS_COMPOUND;
 849
 850         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 851                 if (!page_flag_names[i])
 852                         continue;
 853                 if (!strncmp(str, page_flag_names[i] + 2, len))
 854                         return 1ULL << i;
 855         }
 856
 857         return parse_number(str);
 858 }
 859
 860 static uint64_t parse_flag_names(const char *str, int all)
 861 {
 862         const char *p    = str;
 863         uint64_t   flags = 0;
 864
 865         while (1) {
 866                 if (*p == ',' || *p == '=' || *p == '\0') {
 867                         if ((*str != '~') || (*str == '~' && all && *++str))
 868                                 flags |= parse_flag_name(str, p - str);
 869                         if (*p != ',')
 870                                 break;
 871                         str = p + 1;
 872                 }
 873                 p++;
 874         }
 875
 876         return flags;
 877 }
 878
 879 static void parse_bits_mask(const char *optarg)
 880 {
 881         uint64_t mask;
 882         uint64_t bits;
 883         const char *p;
 884
 885         p = strchr(optarg, '=');
 886         if (p == optarg) {
 887                 mask = KPF_ALL_BITS;
 888                 bits = parse_flag_names(p + 1, 0);
 889         } else if (p) {
 890                 mask = parse_flag_names(optarg, 0);
 891                 bits = parse_flag_names(p + 1, 0);
 892         } else if (strchr(optarg, '~')) {
 893                 mask = parse_flag_names(optarg, 1);
 894                 bits = parse_flag_names(optarg, 0);
 895         } else {
 896                 mask = parse_flag_names(optarg, 0);
 897                 bits = KPF_ALL_BITS;
 898         }
 899
 900         add_bits_filter(mask, bits);
 901 }
 902
 903 static void describe_flags(const char *optarg)
 904 {
 905         uint64_t flags = parse_flag_names(optarg, 0);
 906
 907         printf("0x%016llx\t%s\t%s\n",
 908                 (unsigned long long)flags,
 909                 page_flag_name(flags),
 910                 page_flag_longname(flags));
 911 }
 912
 913 static const struct option opts[] = {
 914         { "raw"       , 0, NULL, 'r' },
 915         { "pid"       , 1, NULL, 'p' },
 916         { "file"      , 1, NULL, 'f' },
 917         { "addr"      , 1, NULL, 'a' },
 918         { "bits"      , 1, NULL, 'b' },
 919         { "describe"  , 1, NULL, 'd' },
 920         { "list"      , 0, NULL, 'l' },
 921         { "list-each" , 0, NULL, 'L' },
 922         { "no-summary", 0, NULL, 'N' },
 923         { "hwpoison"  , 0, NULL, 'X' },
 924         { "unpoison"  , 0, NULL, 'x' },
 925         { "help"      , 0, NULL, 'h' },
 926         { NULL        , 0, NULL, 0 }
 927 };
 928
 929 int main(int argc, char *argv[])
 930 {
 931         int c;
 932
 933         page_size = getpagesize();
 934
 935         while ((c = getopt_long(argc, argv,
 936                                 "rp:f:a:b:d:lLNXxh", opts, NULL)) != -1) {
 937                 switch (c) {
 938                 case 'r':
 939                         opt_raw = 1;
 940                         break;
 941                 case 'p':
 942                         parse_pid(optarg);
 943                         break;
 944                 case 'f':
 945                         parse_file(optarg);
 946                         break;
 947                 case 'a':
 948                         parse_addr_range(optarg);
 949                         break;
 950                 case 'b':
 951                         parse_bits_mask(optarg);
 952                         break;
 953                 case 'd':
 954                         describe_flags(optarg);
 955                         exit(0);
 956                 case 'l':
 957                         opt_list = 1;
 958                         break;
 959                 case 'L':
 960                         opt_list = 2;
 961                         break;
 962                 case 'N':
 963                         opt_no_summary = 1;
 964                         break;
 965                 case 'X':
 966                         opt_hwpoison = 1;
 967                         prepare_hwpoison_fd();
 968                         break;
 969                 case 'x':
 970                         opt_unpoison = 1;
 971                         prepare_hwpoison_fd();
 972                         break;
 973                 case 'h':
 974                         usage();
 975                         exit(0);
 976                 default:
 977                         usage();
 978                         exit(1);
 979                 }
 980         }
 981
 982         if (opt_list && opt_pid)
 983                 printf("voffset\t");
 984         if (opt_list == 1)
 985                 printf("offset\tlen\tflags\n");
 986         if (opt_list == 2)
 987                 printf("offset\tflags\n");
 988
 989         walk_addr_ranges();
 990
 991         if (opt_list == 1)
 992                 show_page_range(0, 0, 0);  /* drain the buffer */
 993
 994         if (opt_no_summary)
 995                 return 0;
 996
 997         if (opt_list)
 998                 printf("\n\n");
 999
1000         show_summary();
1001
1002         return 0;
1003 }