SAFE public projects git trees. - safe/jmp/linux-2.6/blob - Documentation/vm/page-types.c

   1 /*
   2  * page-types: Tool for querying page flags
   3  *
   4  * Copyright (C) 2009 Intel corporation
   5  *
   6  * Authors: Wu Fengguang <fengguang.wu@intel.com>
   7  *
   8  * Released under the General Public License (GPL).
   9  */
  10
  11 #define _LARGEFILE64_SOURCE
  12 #include <stdio.h>
  13 #include <stdlib.h>
  14 #include <unistd.h>
  15 #include <stdint.h>
  16 #include <stdarg.h>
  17 #include <string.h>
  18 #include <getopt.h>
  19 #include <limits.h>
  20 #include <assert.h>
  21 #include <sys/types.h>
  22 #include <sys/errno.h>
  23 #include <sys/fcntl.h>
  24
  25
  26 /*
  27  * pagemap kernel ABI bits
  28  */
  29
  30 #define PM_ENTRY_BYTES      sizeof(uint64_t)
  31 #define PM_STATUS_BITS      3
  32 #define PM_STATUS_OFFSET    (64 - PM_STATUS_BITS)
  33 #define PM_STATUS_MASK      (((1LL << PM_STATUS_BITS) - 1) << PM_STATUS_OFFSET)
  34 #define PM_STATUS(nr)       (((nr) << PM_STATUS_OFFSET) & PM_STATUS_MASK)
  35 #define PM_PSHIFT_BITS      6
  36 #define PM_PSHIFT_OFFSET    (PM_STATUS_OFFSET - PM_PSHIFT_BITS)
  37 #define PM_PSHIFT_MASK      (((1LL << PM_PSHIFT_BITS) - 1) << PM_PSHIFT_OFFSET)
  38 #define PM_PSHIFT(x)        (((u64) (x) << PM_PSHIFT_OFFSET) & PM_PSHIFT_MASK)
  39 #define PM_PFRAME_MASK      ((1LL << PM_PSHIFT_OFFSET) - 1)
  40 #define PM_PFRAME(x)        ((x) & PM_PFRAME_MASK)
  41
  42 #define PM_PRESENT          PM_STATUS(4LL)
  43 #define PM_SWAP             PM_STATUS(2LL)
  44
  45
  46 /*
  47  * kernel page flags
  48  */
  49
  50 #define KPF_BYTES               8
  51 #define PROC_KPAGEFLAGS         "/proc/kpageflags"
  52
  53 /* copied from kpageflags_read() */
  54 #define KPF_LOCKED              0
  55 #define KPF_ERROR               1
  56 #define KPF_REFERENCED          2
  57 #define KPF_UPTODATE            3
  58 #define KPF_DIRTY               4
  59 #define KPF_LRU                 5
  60 #define KPF_ACTIVE              6
  61 #define KPF_SLAB                7
  62 #define KPF_WRITEBACK           8
  63 #define KPF_RECLAIM             9
  64 #define KPF_BUDDY               10
  65
  66 /* [11-20] new additions in 2.6.31 */
  67 #define KPF_MMAP                11
  68 #define KPF_ANON                12
  69 #define KPF_SWAPCACHE           13
  70 #define KPF_SWAPBACKED          14
  71 #define KPF_COMPOUND_HEAD       15
  72 #define KPF_COMPOUND_TAIL       16
  73 #define KPF_HUGE                17
  74 #define KPF_UNEVICTABLE         18
  75 #define KPF_HWPOISON            19
  76 #define KPF_NOPAGE              20
  77 #define KPF_KSM                 21
  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
  89 /* [48-] take some arbitrary free slots for expanding overloaded flags
  90  * not part of kernel API
  91  */
  92 #define KPF_READAHEAD           48
  93 #define KPF_SLOB_FREE           49
  94 #define KPF_SLUB_FROZEN         50
  95 #define KPF_SLUB_DEBUG          51
  96
  97 #define KPF_ALL_BITS            ((uint64_t)~0ULL)
  98 #define KPF_HACKERS_BITS        (0xffffULL << 32)
  99 #define KPF_OVERLOADED_BITS     (0xffffULL << 48)
 100 #define BIT(name)               (1ULL << KPF_##name)
 101 #define BITS_COMPOUND           (BIT(COMPOUND_HEAD) | BIT(COMPOUND_TAIL))
 102
 103 static const char *page_flag_names[] = {
 104         [KPF_LOCKED]            = "L:locked",
 105         [KPF_ERROR]             = "E:error",
 106         [KPF_REFERENCED]        = "R:referenced",
 107         [KPF_UPTODATE]          = "U:uptodate",
 108         [KPF_DIRTY]             = "D:dirty",
 109         [KPF_LRU]               = "l:lru",
 110         [KPF_ACTIVE]            = "A:active",
 111         [KPF_SLAB]              = "S:slab",
 112         [KPF_WRITEBACK]         = "W:writeback",
 113         [KPF_RECLAIM]           = "I:reclaim",
 114         [KPF_BUDDY]             = "B:buddy",
 115
 116         [KPF_MMAP]              = "M:mmap",
 117         [KPF_ANON]              = "a:anonymous",
 118         [KPF_SWAPCACHE]         = "s:swapcache",
 119         [KPF_SWAPBACKED]        = "b:swapbacked",
 120         [KPF_COMPOUND_HEAD]     = "H:compound_head",
 121         [KPF_COMPOUND_TAIL]     = "T:compound_tail",
 122         [KPF_HUGE]              = "G:huge",
 123         [KPF_UNEVICTABLE]       = "u:unevictable",
 124         [KPF_HWPOISON]          = "X:hwpoison",
 125         [KPF_NOPAGE]            = "n:nopage",
 126         [KPF_KSM]               = "x:ksm",
 127
 128         [KPF_RESERVED]          = "r:reserved",
 129         [KPF_MLOCKED]           = "m:mlocked",
 130         [KPF_MAPPEDTODISK]      = "d:mappedtodisk",
 131         [KPF_PRIVATE]           = "P:private",
 132         [KPF_PRIVATE_2]         = "p:private_2",
 133         [KPF_OWNER_PRIVATE]     = "O:owner_private",
 134         [KPF_ARCH]              = "h:arch",
 135         [KPF_UNCACHED]          = "c:uncached",
 136
 137         [KPF_READAHEAD]         = "I:readahead",
 138         [KPF_SLOB_FREE]         = "P:slob_free",
 139         [KPF_SLUB_FROZEN]       = "A:slub_frozen",
 140         [KPF_SLUB_DEBUG]        = "E:slub_debug",
 141 };
 142
 143
 144 /*
 145  * data structures
 146  */
 147
 148 static int              opt_raw;        /* for kernel developers */
 149 static int              opt_list;       /* list pages (in ranges) */
 150 static int              opt_no_summary; /* don't show summary */
 151 static pid_t            opt_pid;        /* process to walk */
 152
 153 #define MAX_ADDR_RANGES 1024
 154 static int              nr_addr_ranges;
 155 static unsigned long    opt_offset[MAX_ADDR_RANGES];
 156 static unsigned long    opt_size[MAX_ADDR_RANGES];
 157
 158 #define MAX_VMAS        10240
 159 static int              nr_vmas;
 160 static unsigned long    pg_start[MAX_VMAS];
 161 static unsigned long    pg_end[MAX_VMAS];
 162
 163 #define MAX_BIT_FILTERS 64
 164 static int              nr_bit_filters;
 165 static uint64_t         opt_mask[MAX_BIT_FILTERS];
 166 static uint64_t         opt_bits[MAX_BIT_FILTERS];
 167
 168 static int              page_size;
 169
 170 static int              pagemap_fd;
 171 static int              kpageflags_fd;
 172
 173 static int              opt_hwpoison;
 174 static int              opt_unpoison;
 175
 176 static const char       hwpoison_debug_fs[] = "/debug/hwpoison";
 177 static int              hwpoison_inject_fd;
 178 static int              hwpoison_forget_fd;
 179
 180 #define HASH_SHIFT      13
 181 #define HASH_SIZE       (1 << HASH_SHIFT)
 182 #define HASH_MASK       (HASH_SIZE - 1)
 183 #define HASH_KEY(flags) (flags & HASH_MASK)
 184
 185 static unsigned long    total_pages;
 186 static unsigned long    nr_pages[HASH_SIZE];
 187 static uint64_t         page_flags[HASH_SIZE];
 188
 189
 190 /*
 191  * helper functions
 192  */
 193
 194 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 195
 196 #define min_t(type, x, y) ({                    \
 197         type __min1 = (x);                      \
 198         type __min2 = (y);                      \
 199         __min1 < __min2 ? __min1 : __min2; })
 200
 201 #define max_t(type, x, y) ({                    \
 202         type __max1 = (x);                      \
 203         type __max2 = (y);                      \
 204         __max1 > __max2 ? __max1 : __max2; })
 205
 206 static unsigned long pages2mb(unsigned long pages)
 207 {
 208         return (pages * page_size) >> 20;
 209 }
 210
 211 static void fatal(const char *x, ...)
 212 {
 213         va_list ap;
 214
 215         va_start(ap, x);
 216         vfprintf(stderr, x, ap);
 217         va_end(ap);
 218         exit(EXIT_FAILURE);
 219 }
 220
 221 static int checked_open(const char *pathname, int flags)
 222 {
 223         int fd = open(pathname, flags);
 224
 225         if (fd < 0) {
 226                 perror(pathname);
 227                 exit(EXIT_FAILURE);
 228         }
 229
 230         return fd;
 231 }
 232
 233 /*
 234  * pagemap/kpageflags routines
 235  */
 236
 237 static unsigned long do_u64_read(int fd, char *name,
 238                                  uint64_t *buf,
 239                                  unsigned long index,
 240                                  unsigned long count)
 241 {
 242         long bytes;
 243
 244         if (index > ULONG_MAX / 8)
 245                 fatal("index overflow: %lu\n", index);
 246
 247         if (lseek(fd, index * 8, SEEK_SET) < 0) {
 248                 perror(name);
 249                 exit(EXIT_FAILURE);
 250         }
 251
 252         bytes = read(fd, buf, count * 8);
 253         if (bytes < 0) {
 254                 perror(name);
 255                 exit(EXIT_FAILURE);
 256         }
 257         if (bytes % 8)
 258                 fatal("partial read: %lu bytes\n", bytes);
 259
 260         return bytes / 8;
 261 }
 262
 263 static unsigned long kpageflags_read(uint64_t *buf,
 264                                      unsigned long index,
 265                                      unsigned long pages)
 266 {
 267         return do_u64_read(kpageflags_fd, PROC_KPAGEFLAGS, buf, index, pages);
 268 }
 269
 270 static unsigned long pagemap_read(uint64_t *buf,
 271                                   unsigned long index,
 272                                   unsigned long pages)
 273 {
 274         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 275 }
 276
 277 static unsigned long pagemap_pfn(uint64_t val)
 278 {
 279         unsigned long pfn;
 280
 281         if (val & PM_PRESENT)
 282                 pfn = PM_PFRAME(val);
 283         else
 284                 pfn = 0;
 285
 286         return pfn;
 287 }
 288
 289
 290 /*
 291  * page flag names
 292  */
 293
 294 static char *page_flag_name(uint64_t flags)
 295 {
 296         static char buf[65];
 297         int present;
 298         int i, j;
 299
 300         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 301                 present = (flags >> i) & 1;
 302                 if (!page_flag_names[i]) {
 303                         if (present)
 304                                 fatal("unknown flag bit %d\n", i);
 305                         continue;
 306                 }
 307                 buf[j++] = present ? page_flag_names[i][0] : '_';
 308         }
 309
 310         return buf;
 311 }
 312
 313 static char *page_flag_longname(uint64_t flags)
 314 {
 315         static char buf[1024];
 316         int i, n;
 317
 318         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 319                 if (!page_flag_names[i])
 320                         continue;
 321                 if ((flags >> i) & 1)
 322                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 323                                         page_flag_names[i] + 2);
 324         }
 325         if (n)
 326                 n--;
 327         buf[n] = '\0';
 328
 329         return buf;
 330 }
 331
 332
 333 /*
 334  * page list and summary
 335  */
 336
 337 static void show_page_range(unsigned long voffset,
 338                             unsigned long offset, uint64_t flags)
 339 {
 340         static uint64_t      flags0;
 341         static unsigned long voff;
 342         static unsigned long index;
 343         static unsigned long count;
 344
 345         if (flags == flags0 && offset == index + count &&
 346             (!opt_pid || voffset == voff + count)) {
 347                 count++;
 348                 return;
 349         }
 350
 351         if (count) {
 352                 if (opt_pid)
 353                         printf("%lx\t", voff);
 354                 printf("%lx\t%lx\t%s\n",
 355                                 index, count, page_flag_name(flags0));
 356         }
 357
 358         flags0 = flags;
 359         index  = offset;
 360         voff   = voffset;
 361         count  = 1;
 362 }
 363
 364 static void show_page(unsigned long voffset,
 365                       unsigned long offset, uint64_t flags)
 366 {
 367         if (opt_pid)
 368                 printf("%lx\t", voffset);
 369         printf("%lx\t%s\n", offset, page_flag_name(flags));
 370 }
 371
 372 static void show_summary(void)
 373 {
 374         int i;
 375
 376         printf("             flags\tpage-count       MB"
 377                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 378
 379         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 380                 if (nr_pages[i])
 381                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 382                                 (unsigned long long)page_flags[i],
 383                                 nr_pages[i],
 384                                 pages2mb(nr_pages[i]),
 385                                 page_flag_name(page_flags[i]),
 386                                 page_flag_longname(page_flags[i]));
 387         }
 388
 389         printf("             total\t%10lu %8lu\n",
 390                         total_pages, pages2mb(total_pages));
 391 }
 392
 393
 394 /*
 395  * page flag filters
 396  */
 397
 398 static int bit_mask_ok(uint64_t flags)
 399 {
 400         int i;
 401
 402         for (i = 0; i < nr_bit_filters; i++) {
 403                 if (opt_bits[i] == KPF_ALL_BITS) {
 404                         if ((flags & opt_mask[i]) == 0)
 405                                 return 0;
 406                 } else {
 407                         if ((flags & opt_mask[i]) != opt_bits[i])
 408                                 return 0;
 409                 }
 410         }
 411
 412         return 1;
 413 }
 414
 415 static uint64_t expand_overloaded_flags(uint64_t flags)
 416 {
 417         /* SLOB/SLUB overload several page flags */
 418         if (flags & BIT(SLAB)) {
 419                 if (flags & BIT(PRIVATE))
 420                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 421                 if (flags & BIT(ACTIVE))
 422                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 423                 if (flags & BIT(ERROR))
 424                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 425         }
 426
 427         /* PG_reclaim is overloaded as PG_readahead in the read path */
 428         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 429                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 430
 431         return flags;
 432 }
 433
 434 static uint64_t well_known_flags(uint64_t flags)
 435 {
 436         /* hide flags intended only for kernel hacker */
 437         flags &= ~KPF_HACKERS_BITS;
 438
 439         /* hide non-hugeTLB compound pages */
 440         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 441                 flags &= ~BITS_COMPOUND;
 442
 443         return flags;
 444 }
 445
 446 static uint64_t kpageflags_flags(uint64_t flags)
 447 {
 448         flags = expand_overloaded_flags(flags);
 449
 450         if (!opt_raw)
 451                 flags = well_known_flags(flags);
 452
 453         return flags;
 454 }
 455
 456 /*
 457  * page actions
 458  */
 459
 460 static void prepare_hwpoison_fd(void)
 461 {
 462         char buf[100];
 463
 464         if (opt_hwpoison && !hwpoison_inject_fd) {
 465                 sprintf(buf, "%s/corrupt-pfn", hwpoison_debug_fs);
 466                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 467         }
 468
 469         if (opt_unpoison && !hwpoison_forget_fd) {
 470                 sprintf(buf, "%s/renew-pfn", hwpoison_debug_fs);
 471                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 472         }
 473 }
 474
 475 static int hwpoison_page(unsigned long offset)
 476 {
 477         char buf[100];
 478         int len;
 479
 480         len = sprintf(buf, "0x%lx\n", offset);
 481         len = write(hwpoison_inject_fd, buf, len);
 482         if (len < 0) {
 483                 perror("hwpoison inject");
 484                 return len;
 485         }
 486         return 0;
 487 }
 488
 489 static int unpoison_page(unsigned long offset)
 490 {
 491         char buf[100];
 492         int len;
 493
 494         len = sprintf(buf, "0x%lx\n", offset);
 495         len = write(hwpoison_forget_fd, buf, len);
 496         if (len < 0) {
 497                 perror("hwpoison forget");
 498                 return len;
 499         }
 500         return 0;
 501 }
 502
 503 /*
 504  * page frame walker
 505  */
 506
 507 static int hash_slot(uint64_t flags)
 508 {
 509         int k = HASH_KEY(flags);
 510         int i;
 511
 512         /* Explicitly reserve slot 0 for flags 0: the following logic
 513          * cannot distinguish an unoccupied slot from slot (flags==0).
 514          */
 515         if (flags == 0)
 516                 return 0;
 517
 518         /* search through the remaining (HASH_SIZE-1) slots */
 519         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 520                 if (!k || k >= ARRAY_SIZE(page_flags))
 521                         k = 1;
 522                 if (page_flags[k] == 0) {
 523                         page_flags[k] = flags;
 524                         return k;
 525                 }
 526                 if (page_flags[k] == flags)
 527                         return k;
 528         }
 529
 530         fatal("hash table full: bump up HASH_SHIFT?\n");
 531         exit(EXIT_FAILURE);
 532 }
 533
 534 static void add_page(unsigned long voffset,
 535                      unsigned long offset, uint64_t flags)
 536 {
 537         flags = kpageflags_flags(flags);
 538
 539         if (!bit_mask_ok(flags))
 540                 return;
 541
 542         if (opt_hwpoison)
 543                 hwpoison_page(offset);
 544         if (opt_unpoison)
 545                 unpoison_page(offset);
 546
 547         if (opt_list == 1)
 548                 show_page_range(voffset, offset, flags);
 549         else if (opt_list == 2)
 550                 show_page(voffset, offset, flags);
 551
 552         nr_pages[hash_slot(flags)]++;
 553         total_pages++;
 554 }
 555
 556 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 557 static void walk_pfn(unsigned long voffset,
 558                      unsigned long index,
 559                      unsigned long count)
 560 {
 561         uint64_t buf[KPAGEFLAGS_BATCH];
 562         unsigned long batch;
 563         long pages;
 564         unsigned long i;
 565
 566         while (count) {
 567                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 568                 pages = kpageflags_read(buf, index, batch);
 569                 if (pages == 0)
 570                         break;
 571
 572                 for (i = 0; i < pages; i++)
 573                         add_page(voffset + i, index + i, buf[i]);
 574
 575                 index += pages;
 576                 count -= pages;
 577         }
 578 }
 579
 580 #define PAGEMAP_BATCH   (64 << 10)
 581 static void walk_vma(unsigned long index, unsigned long count)
 582 {
 583         uint64_t buf[PAGEMAP_BATCH];
 584         unsigned long batch;
 585         unsigned long pages;
 586         unsigned long pfn;
 587         unsigned long i;
 588
 589         while (count) {
 590                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 591                 pages = pagemap_read(buf, index, batch);
 592                 if (pages == 0)
 593                         break;
 594
 595                 for (i = 0; i < pages; i++) {
 596                         pfn = pagemap_pfn(buf[i]);
 597                         if (pfn)
 598                                 walk_pfn(index + i, pfn, 1);
 599                 }
 600
 601                 index += pages;
 602                 count -= pages;
 603         }
 604 }
 605
 606 static void walk_task(unsigned long index, unsigned long count)
 607 {
 608         const unsigned long end = index + count;
 609         unsigned long start;
 610         int i = 0;
 611
 612         while (index < end) {
 613
 614                 while (pg_end[i] <= index)
 615                         if (++i >= nr_vmas)
 616                                 return;
 617                 if (pg_start[i] >= end)
 618                         return;
 619
 620                 start = max_t(unsigned long, pg_start[i], index);
 621                 index = min_t(unsigned long, pg_end[i], end);
 622
 623                 assert(start < index);
 624                 walk_vma(start, index - start);
 625         }
 626 }
 627
 628 static void add_addr_range(unsigned long offset, unsigned long size)
 629 {
 630         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 631                 fatal("too many addr ranges\n");
 632
 633         opt_offset[nr_addr_ranges] = offset;
 634         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 635         nr_addr_ranges++;
 636 }
 637
 638 static void walk_addr_ranges(void)
 639 {
 640         int i;
 641
 642         kpageflags_fd = checked_open(PROC_KPAGEFLAGS, O_RDONLY);
 643
 644         if (!nr_addr_ranges)
 645                 add_addr_range(0, ULONG_MAX);
 646
 647         for (i = 0; i < nr_addr_ranges; i++)
 648                 if (!opt_pid)
 649                         walk_pfn(0, opt_offset[i], opt_size[i]);
 650                 else
 651                         walk_task(opt_offset[i], opt_size[i]);
 652
 653         close(kpageflags_fd);
 654 }
 655
 656
 657 /*
 658  * user interface
 659  */
 660
 661 static const char *page_flag_type(uint64_t flag)
 662 {
 663         if (flag & KPF_HACKERS_BITS)
 664                 return "(r)";
 665         if (flag & KPF_OVERLOADED_BITS)
 666                 return "(o)";
 667         return "   ";
 668 }
 669
 670 static void usage(void)
 671 {
 672         int i, j;
 673
 674         printf(
 675 "page-types [options]\n"
 676 "            -r|--raw                   Raw mode, for kernel developers\n"
 677 "            -d|--describe flags        Describe flags\n"
 678 "            -a|--addr    addr-spec     Walk a range of pages\n"
 679 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 680 "            -p|--pid     pid           Walk process address space\n"
 681 #if 0 /* planned features */
 682 "            -f|--file    filename      Walk file address space\n"
 683 #endif
 684 "            -l|--list                  Show page details in ranges\n"
 685 "            -L|--list-each             Show page details one by one\n"
 686 "            -N|--no-summary            Don't show summay info\n"
 687 "            -X|--hwpoison              hwpoison pages\n"
 688 "            -x|--unpoison              unpoison pages\n"
 689 "            -h|--help                  Show this usage message\n"
 690 "flags:\n"
 691 "            0x10                       bitfield format, e.g.\n"
 692 "            anon                       bit-name, e.g.\n"
 693 "            0x10,anon                  comma-separated list, e.g.\n"
 694 "addr-spec:\n"
 695 "            N                          one page at offset N (unit: pages)\n"
 696 "            N+M                        pages range from N to N+M-1\n"
 697 "            N,M                        pages range from N to M-1\n"
 698 "            N,                         pages range from N to end\n"
 699 "            ,M                         pages range from 0 to M-1\n"
 700 "bits-spec:\n"
 701 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 702 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 703 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 704 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 705 "bit-names:\n"
 706         );
 707
 708         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 709                 if (!page_flag_names[i])
 710                         continue;
 711                 printf("%16s%s", page_flag_names[i] + 2,
 712                                  page_flag_type(1ULL << i));
 713                 if (++j > 3) {
 714                         j = 0;
 715                         putchar('\n');
 716                 }
 717         }
 718         printf("\n                                   "
 719                 "(r) raw mode bits  (o) overloaded bits\n");
 720 }
 721
 722 static unsigned long long parse_number(const char *str)
 723 {
 724         unsigned long long n;
 725
 726         n = strtoll(str, NULL, 0);
 727
 728         if (n == 0 && str[0] != '0')
 729                 fatal("invalid name or number: %s\n", str);
 730
 731         return n;
 732 }
 733
 734 static void parse_pid(const char *str)
 735 {
 736         FILE *file;
 737         char buf[5000];
 738
 739         opt_pid = parse_number(str);
 740
 741         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 742         pagemap_fd = checked_open(buf, O_RDONLY);
 743
 744         sprintf(buf, "/proc/%d/maps", opt_pid);
 745         file = fopen(buf, "r");
 746         if (!file) {
 747                 perror(buf);
 748                 exit(EXIT_FAILURE);
 749         }
 750
 751         while (fgets(buf, sizeof(buf), file) != NULL) {
 752                 unsigned long vm_start;
 753                 unsigned long vm_end;
 754                 unsigned long long pgoff;
 755                 int major, minor;
 756                 char r, w, x, s;
 757                 unsigned long ino;
 758                 int n;
 759
 760                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 761                            &vm_start,
 762                            &vm_end,
 763                            &r, &w, &x, &s,
 764                            &pgoff,
 765                            &major, &minor,
 766                            &ino);
 767                 if (n < 10) {
 768                         fprintf(stderr, "unexpected line: %s\n", buf);
 769                         continue;
 770                 }
 771                 pg_start[nr_vmas] = vm_start / page_size;
 772                 pg_end[nr_vmas] = vm_end / page_size;
 773                 if (++nr_vmas >= MAX_VMAS) {
 774                         fprintf(stderr, "too many VMAs\n");
 775                         break;
 776                 }
 777         }
 778         fclose(file);
 779 }
 780
 781 static void parse_file(const char *name)
 782 {
 783 }
 784
 785 static void parse_addr_range(const char *optarg)
 786 {
 787         unsigned long offset;
 788         unsigned long size;
 789         char *p;
 790
 791         p = strchr(optarg, ',');
 792         if (!p)
 793                 p = strchr(optarg, '+');
 794
 795         if (p == optarg) {
 796                 offset = 0;
 797                 size   = parse_number(p + 1);
 798         } else if (p) {
 799                 offset = parse_number(optarg);
 800                 if (p[1] == '\0')
 801                         size = ULONG_MAX;
 802                 else {
 803                         size = parse_number(p + 1);
 804                         if (*p == ',') {
 805                                 if (size < offset)
 806                                         fatal("invalid range: %lu,%lu\n",
 807                                                         offset, size);
 808                                 size -= offset;
 809                         }
 810                 }
 811         } else {
 812                 offset = parse_number(optarg);
 813                 size   = 1;
 814         }
 815
 816         add_addr_range(offset, size);
 817 }
 818
 819 static void add_bits_filter(uint64_t mask, uint64_t bits)
 820 {
 821         if (nr_bit_filters >= MAX_BIT_FILTERS)
 822                 fatal("too much bit filters\n");
 823
 824         opt_mask[nr_bit_filters] = mask;
 825         opt_bits[nr_bit_filters] = bits;
 826         nr_bit_filters++;
 827 }
 828
 829 static uint64_t parse_flag_name(const char *str, int len)
 830 {
 831         int i;
 832
 833         if (!*str || !len)
 834                 return 0;
 835
 836         if (len <= 8 && !strncmp(str, "compound", len))
 837                 return BITS_COMPOUND;
 838
 839         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 840                 if (!page_flag_names[i])
 841                         continue;
 842                 if (!strncmp(str, page_flag_names[i] + 2, len))
 843                         return 1ULL << i;
 844         }
 845
 846         return parse_number(str);
 847 }
 848
 849 static uint64_t parse_flag_names(const char *str, int all)
 850 {
 851         const char *p    = str;
 852         uint64_t   flags = 0;
 853
 854         while (1) {
 855                 if (*p == ',' || *p == '=' || *p == '\0') {
 856                         if ((*str != '~') || (*str == '~' && all && *++str))
 857                                 flags |= parse_flag_name(str, p - str);
 858                         if (*p != ',')
 859                                 break;
 860                         str = p + 1;
 861                 }
 862                 p++;
 863         }
 864
 865         return flags;
 866 }
 867
 868 static void parse_bits_mask(const char *optarg)
 869 {
 870         uint64_t mask;
 871         uint64_t bits;
 872         const char *p;
 873
 874         p = strchr(optarg, '=');
 875         if (p == optarg) {
 876                 mask = KPF_ALL_BITS;
 877                 bits = parse_flag_names(p + 1, 0);
 878         } else if (p) {
 879                 mask = parse_flag_names(optarg, 0);
 880                 bits = parse_flag_names(p + 1, 0);
 881         } else if (strchr(optarg, '~')) {
 882                 mask = parse_flag_names(optarg, 1);
 883                 bits = parse_flag_names(optarg, 0);
 884         } else {
 885                 mask = parse_flag_names(optarg, 0);
 886                 bits = KPF_ALL_BITS;
 887         }
 888
 889         add_bits_filter(mask, bits);
 890 }
 891
 892 static void describe_flags(const char *optarg)
 893 {
 894         uint64_t flags = parse_flag_names(optarg, 0);
 895
 896         printf("0x%016llx\t%s\t%s\n",
 897                 (unsigned long long)flags,
 898                 page_flag_name(flags),
 899                 page_flag_longname(flags));
 900 }
 901
 902 static const struct option opts[] = {
 903         { "raw"       , 0, NULL, 'r' },
 904         { "pid"       , 1, NULL, 'p' },
 905         { "file"      , 1, NULL, 'f' },
 906         { "addr"      , 1, NULL, 'a' },
 907         { "bits"      , 1, NULL, 'b' },
 908         { "describe"  , 1, NULL, 'd' },
 909         { "list"      , 0, NULL, 'l' },
 910         { "list-each" , 0, NULL, 'L' },
 911         { "no-summary", 0, NULL, 'N' },
 912         { "hwpoison"  , 0, NULL, 'X' },
 913         { "unpoison"  , 0, NULL, 'x' },
 914         { "help"      , 0, NULL, 'h' },
 915         { NULL        , 0, NULL, 0 }
 916 };
 917
 918 int main(int argc, char *argv[])
 919 {
 920         int c;
 921
 922         page_size = getpagesize();
 923
 924         while ((c = getopt_long(argc, argv,
 925                                 "rp:f:a:b:d:lLNXxh", opts, NULL)) != -1) {
 926                 switch (c) {
 927                 case 'r':
 928                         opt_raw = 1;
 929                         break;
 930                 case 'p':
 931                         parse_pid(optarg);
 932                         break;
 933                 case 'f':
 934                         parse_file(optarg);
 935                         break;
 936                 case 'a':
 937                         parse_addr_range(optarg);
 938                         break;
 939                 case 'b':
 940                         parse_bits_mask(optarg);
 941                         break;
 942                 case 'd':
 943                         describe_flags(optarg);
 944                         exit(0);
 945                 case 'l':
 946                         opt_list = 1;
 947                         break;
 948                 case 'L':
 949                         opt_list = 2;
 950                         break;
 951                 case 'N':
 952                         opt_no_summary = 1;
 953                         break;
 954                 case 'X':
 955                         opt_hwpoison = 1;
 956                         prepare_hwpoison_fd();
 957                         break;
 958                 case 'x':
 959                         opt_unpoison = 1;
 960                         prepare_hwpoison_fd();
 961                         break;
 962                 case 'h':
 963                         usage();
 964                         exit(0);
 965                 default:
 966                         usage();
 967                         exit(1);
 968                 }
 969         }
 970
 971         if (opt_list && opt_pid)
 972                 printf("voffset\t");
 973         if (opt_list == 1)
 974                 printf("offset\tlen\tflags\n");
 975         if (opt_list == 2)
 976                 printf("offset\tflags\n");
 977
 978         walk_addr_ranges();
 979
 980         if (opt_list == 1)
 981                 show_page_range(0, 0, 0);  /* drain the buffer */
 982
 983         if (opt_no_summary)
 984                 return 0;
 985
 986         if (opt_list)
 987                 printf("\n\n");
 988
 989         show_summary();
 990
 991         return 0;
 992 }