linux/Documentation/vm/page-types.c
<<
>>
Prefs
   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
 103static 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
 148static int              opt_raw;        /* for kernel developers */
 149static int              opt_list;       /* list pages (in ranges) */
 150static int              opt_no_summary; /* don't show summary */
 151static pid_t            opt_pid;        /* process to walk */
 152
 153#define MAX_ADDR_RANGES 1024
 154static int              nr_addr_ranges;
 155static unsigned long    opt_offset[MAX_ADDR_RANGES];
 156static unsigned long    opt_size[MAX_ADDR_RANGES];
 157
 158#define MAX_VMAS        10240
 159static int              nr_vmas;
 160static unsigned long    pg_start[MAX_VMAS];
 161static unsigned long    pg_end[MAX_VMAS];
 162
 163#define MAX_BIT_FILTERS 64
 164static int              nr_bit_filters;
 165static uint64_t         opt_mask[MAX_BIT_FILTERS];
 166static uint64_t         opt_bits[MAX_BIT_FILTERS];
 167
 168static int              page_size;
 169
 170static int              pagemap_fd;
 171static int              kpageflags_fd;
 172
 173static int              opt_hwpoison;
 174static int              opt_unpoison;
 175
 176static char             *hwpoison_debug_fs = "/debug/hwpoison";
 177static int              hwpoison_inject_fd;
 178static 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
 185static unsigned long    total_pages;
 186static unsigned long    nr_pages[HASH_SIZE];
 187static 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
 206static unsigned long pages2mb(unsigned long pages)
 207{
 208        return (pages * page_size) >> 20;
 209}
 210
 211static 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
 221static 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
 237static 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
 263static 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
 270static 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
 277static 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
 294static 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("unkown flag bit %d\n", i);
 305                        continue;
 306                }
 307                buf[j++] = present ? page_flag_names[i][0] : '_';
 308        }
 309
 310        return buf;
 311}
 312
 313static 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
 337static 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
 364static 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
 372static 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
 398static 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
 415static 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
 434static 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
 446static 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
 460static 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
 475static 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
 489static 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
 507static 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
 534static 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 */
 557static 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        unsigned 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)
 581static 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
 606static 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
 628static 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
 638static 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
 661static 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
 670static 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"            -a|--addr    addr-spec    Walk a range of pages\n"
 678"            -b|--bits    bits-spec    Walk pages with specified bits\n"
 679"            -p|--pid     pid          Walk process address space\n"
 680#if 0 /* planned features */
 681"            -f|--file    filename     Walk file address space\n"
 682#endif
 683"            -l|--list                 Show page details in ranges\n"
 684"            -L|--list-each            Show page details one by one\n"
 685"            -N|--no-summary           Don't show summay info\n"
 686"            -X|--hwpoison             hwpoison pages\n"
 687"            -x|--unpoison             unpoison pages\n"
 688"            -h|--help                 Show this usage message\n"
 689"addr-spec:\n"
 690"            N                         one page at offset N (unit: pages)\n"
 691"            N+M                       pages range from N to N+M-1\n"
 692"            N,M                       pages range from N to M-1\n"
 693"            N,                        pages range from N to end\n"
 694"            ,M                        pages range from 0 to M-1\n"
 695"bits-spec:\n"
 696"            bit1,bit2                 (flags & (bit1|bit2)) != 0\n"
 697"            bit1,bit2=bit1            (flags & (bit1|bit2)) == bit1\n"
 698"            bit1,~bit2                (flags & (bit1|bit2)) == bit1\n"
 699"            =bit1,bit2                flags == (bit1|bit2)\n"
 700"bit-names:\n"
 701        );
 702
 703        for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 704                if (!page_flag_names[i])
 705                        continue;
 706                printf("%16s%s", page_flag_names[i] + 2,
 707                                 page_flag_type(1ULL << i));
 708                if (++j > 3) {
 709                        j = 0;
 710                        putchar('\n');
 711                }
 712        }
 713        printf("\n                                   "
 714                "(r) raw mode bits  (o) overloaded bits\n");
 715}
 716
 717static unsigned long long parse_number(const char *str)
 718{
 719        unsigned long long n;
 720
 721        n = strtoll(str, NULL, 0);
 722
 723        if (n == 0 && str[0] != '0')
 724                fatal("invalid name or number: %s\n", str);
 725
 726        return n;
 727}
 728
 729static void parse_pid(const char *str)
 730{
 731        FILE *file;
 732        char buf[5000];
 733
 734        opt_pid = parse_number(str);
 735
 736        sprintf(buf, "/proc/%d/pagemap", opt_pid);
 737        pagemap_fd = checked_open(buf, O_RDONLY);
 738
 739        sprintf(buf, "/proc/%d/maps", opt_pid);
 740        file = fopen(buf, "r");
 741        if (!file) {
 742                perror(buf);
 743                exit(EXIT_FAILURE);
 744        }
 745
 746        while (fgets(buf, sizeof(buf), file) != NULL) {
 747                unsigned long vm_start;
 748                unsigned long vm_end;
 749                unsigned long long pgoff;
 750                int major, minor;
 751                char r, w, x, s;
 752                unsigned long ino;
 753                int n;
 754
 755                n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 756                           &vm_start,
 757                           &vm_end,
 758                           &r, &w, &x, &s,
 759                           &pgoff,
 760                           &major, &minor,
 761                           &ino);
 762                if (n < 10) {
 763                        fprintf(stderr, "unexpected line: %s\n", buf);
 764                        continue;
 765                }
 766                pg_start[nr_vmas] = vm_start / page_size;
 767                pg_end[nr_vmas] = vm_end / page_size;
 768                if (++nr_vmas >= MAX_VMAS) {
 769                        fprintf(stderr, "too many VMAs\n");
 770                        break;
 771                }
 772        }
 773        fclose(file);
 774}
 775
 776static void parse_file(const char *name)
 777{
 778}
 779
 780static void parse_addr_range(const char *optarg)
 781{
 782        unsigned long offset;
 783        unsigned long size;
 784        char *p;
 785
 786        p = strchr(optarg, ',');
 787        if (!p)
 788                p = strchr(optarg, '+');
 789
 790        if (p == optarg) {
 791                offset = 0;
 792                size   = parse_number(p + 1);
 793        } else if (p) {
 794                offset = parse_number(optarg);
 795                if (p[1] == '\0')
 796                        size = ULONG_MAX;
 797                else {
 798                        size = parse_number(p + 1);
 799                        if (*p == ',') {
 800                                if (size < offset)
 801                                        fatal("invalid range: %lu,%lu\n",
 802                                                        offset, size);
 803                                size -= offset;
 804                        }
 805                }
 806        } else {
 807                offset = parse_number(optarg);
 808                size   = 1;
 809        }
 810
 811        add_addr_range(offset, size);
 812}
 813
 814static void add_bits_filter(uint64_t mask, uint64_t bits)
 815{
 816        if (nr_bit_filters >= MAX_BIT_FILTERS)
 817                fatal("too much bit filters\n");
 818
 819        opt_mask[nr_bit_filters] = mask;
 820        opt_bits[nr_bit_filters] = bits;
 821        nr_bit_filters++;
 822}
 823
 824static uint64_t parse_flag_name(const char *str, int len)
 825{
 826        int i;
 827
 828        if (!*str || !len)
 829                return 0;
 830
 831        if (len <= 8 && !strncmp(str, "compound", len))
 832                return BITS_COMPOUND;
 833
 834        for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 835                if (!page_flag_names[i])
 836                        continue;
 837                if (!strncmp(str, page_flag_names[i] + 2, len))
 838                        return 1ULL << i;
 839        }
 840
 841        return parse_number(str);
 842}
 843
 844static uint64_t parse_flag_names(const char *str, int all)
 845{
 846        const char *p    = str;
 847        uint64_t   flags = 0;
 848
 849        while (1) {
 850                if (*p == ',' || *p == '=' || *p == '\0') {
 851                        if ((*str != '~') || (*str == '~' && all && *++str))
 852                                flags |= parse_flag_name(str, p - str);
 853                        if (*p != ',')
 854                                break;
 855                        str = p + 1;
 856                }
 857                p++;
 858        }
 859
 860        return flags;
 861}
 862
 863static void parse_bits_mask(const char *optarg)
 864{
 865        uint64_t mask;
 866        uint64_t bits;
 867        const char *p;
 868
 869        p = strchr(optarg, '=');
 870        if (p == optarg) {
 871                mask = KPF_ALL_BITS;
 872                bits = parse_flag_names(p + 1, 0);
 873        } else if (p) {
 874                mask = parse_flag_names(optarg, 0);
 875                bits = parse_flag_names(p + 1, 0);
 876        } else if (strchr(optarg, '~')) {
 877                mask = parse_flag_names(optarg, 1);
 878                bits = parse_flag_names(optarg, 0);
 879        } else {
 880                mask = parse_flag_names(optarg, 0);
 881                bits = KPF_ALL_BITS;
 882        }
 883
 884        add_bits_filter(mask, bits);
 885}
 886
 887
 888static struct option opts[] = {
 889        { "raw"       , 0, NULL, 'r' },
 890        { "pid"       , 1, NULL, 'p' },
 891        { "file"      , 1, NULL, 'f' },
 892        { "addr"      , 1, NULL, 'a' },
 893        { "bits"      , 1, NULL, 'b' },
 894        { "list"      , 0, NULL, 'l' },
 895        { "list-each" , 0, NULL, 'L' },
 896        { "no-summary", 0, NULL, 'N' },
 897        { "hwpoison"  , 0, NULL, 'X' },
 898        { "unpoison"  , 0, NULL, 'x' },
 899        { "help"      , 0, NULL, 'h' },
 900        { NULL        , 0, NULL, 0 }
 901};
 902
 903int main(int argc, char *argv[])
 904{
 905        int c;
 906
 907        page_size = getpagesize();
 908
 909        while ((c = getopt_long(argc, argv,
 910                                "rp:f:a:b:lLNXxh", opts, NULL)) != -1) {
 911                switch (c) {
 912                case 'r':
 913                        opt_raw = 1;
 914                        break;
 915                case 'p':
 916                        parse_pid(optarg);
 917                        break;
 918                case 'f':
 919                        parse_file(optarg);
 920                        break;
 921                case 'a':
 922                        parse_addr_range(optarg);
 923                        break;
 924                case 'b':
 925                        parse_bits_mask(optarg);
 926                        break;
 927                case 'l':
 928                        opt_list = 1;
 929                        break;
 930                case 'L':
 931                        opt_list = 2;
 932                        break;
 933                case 'N':
 934                        opt_no_summary = 1;
 935                        break;
 936                case 'X':
 937                        opt_hwpoison = 1;
 938                        prepare_hwpoison_fd();
 939                        break;
 940                case 'x':
 941                        opt_unpoison = 1;
 942                        prepare_hwpoison_fd();
 943                        break;
 944                case 'h':
 945                        usage();
 946                        exit(0);
 947                default:
 948                        usage();
 949                        exit(1);
 950                }
 951        }
 952
 953        if (opt_list && opt_pid)
 954                printf("voffset\t");
 955        if (opt_list == 1)
 956                printf("offset\tlen\tflags\n");
 957        if (opt_list == 2)
 958                printf("offset\tflags\n");
 959
 960        walk_addr_ranges();
 961
 962        if (opt_list == 1)
 963                show_page_range(0, 0, 0);  /* drain the buffer */
 964
 965        if (opt_no_summary)
 966                return 0;
 967
 968        if (opt_list)
 969                printf("\n\n");
 970
 971        show_summary();
 972
 973        return 0;
 974}
 975