tools/vm/page-types.c

   1 /*
   2  * page-types: Tool for querying page flags
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License as published by the Free
   6  * Software Foundation; version 2.
   7  *
   8  * This program is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should find a copy of v2 of the GNU General Public License somewhere on
  14  * your Linux system; if not, write to the Free Software Foundation, Inc., 59
  15  * Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Copyright (C) 2009 Intel corporation
  18  *
  19  * Authors: Wu Fengguang <[email protected]>
  20  */
  21
  22 #define _FILE_OFFSET_BITS 64
  23 #define _GNU_SOURCE
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <unistd.h>
  27 #include <stdint.h>
  28 #include <stdarg.h>
  29 #include <string.h>
  30 #include <getopt.h>
  31 #include <limits.h>
  32 #include <assert.h>
  33 #include <ftw.h>
  34 #include <time.h>
  35 #include <setjmp.h>
  36 #include <signal.h>
  37 #include <sys/types.h>
  38 #include <sys/errno.h>
  39 #include <sys/fcntl.h>
  40 #include <sys/mount.h>
  41 #include <sys/statfs.h>
  42 #include <sys/mman.h>
  43 #include "../../include/uapi/linux/magic.h"
  44 #include "../../include/uapi/linux/kernel-page-flags.h"
  45 #include <api/fs/fs.h>
  46
  47 #ifndef MAX_PATH
  48 # define MAX_PATH 256
  49 #endif
  50
  51 #ifndef STR
  52 # define _STR(x) #x
  53 # define STR(x) _STR(x)
  54 #endif
  55
  56 /*
  57  * pagemap kernel ABI bits
  58  */
  59
  60 #define PM_ENTRY_BYTES          8
  61 #define PM_PFRAME_BITS          55
  62 #define PM_PFRAME_MASK          ((1LL << PM_PFRAME_BITS) - 1)
  63 #define PM_PFRAME(x)            ((x) & PM_PFRAME_MASK)
  64 #define MAX_SWAPFILES_SHIFT     5
  65 #define PM_SWAP_OFFSET(x)       (((x) & PM_PFRAME_MASK) >> MAX_SWAPFILES_SHIFT)
  66 #define PM_SOFT_DIRTY           (1ULL << 55)
  67 #define PM_MMAP_EXCLUSIVE       (1ULL << 56)
  68 #define PM_FILE                 (1ULL << 61)
  69 #define PM_SWAP                 (1ULL << 62)
  70 #define PM_PRESENT              (1ULL << 63)
  71
  72 /*
  73  * kernel page flags
  74  */
  75
  76 #define KPF_BYTES               8
  77 #define PROC_KPAGEFLAGS         "/proc/kpageflags"
  78 #define PROC_KPAGECOUNT         "/proc/kpagecount"
  79 #define PROC_KPAGECGROUP        "/proc/kpagecgroup"
  80
  81 #define SYS_KERNEL_MM_PAGE_IDLE "/sys/kernel/mm/page_idle/bitmap"
  82
  83 /* [32-] kernel hacking assistances */
  84 #define KPF_RESERVED            32
  85 #define KPF_MLOCKED             33
  86 #define KPF_MAPPEDTODISK        34
  87 #define KPF_PRIVATE             35
  88 #define KPF_PRIVATE_2           36
  89 #define KPF_OWNER_PRIVATE       37
  90 #define KPF_ARCH                38
  91 #define KPF_UNCACHED            39
  92 #define KPF_SOFTDIRTY           40
  93
  94 /* [48-] take some arbitrary free slots for expanding overloaded flags
  95  * not part of kernel API
  96  */
  97 #define KPF_READAHEAD           48
  98 #define KPF_SLOB_FREE           49
  99 #define KPF_SLUB_FROZEN         50
 100 #define KPF_SLUB_DEBUG          51
 101 #define KPF_FILE                61
 102 #define KPF_SWAP                62
 103 #define KPF_MMAP_EXCLUSIVE      63
 104
 105 #define KPF_ALL_BITS            ((uint64_t)~0ULL)
 106 #define KPF_HACKERS_BITS        (0xffffULL << 32)
 107 #define KPF_OVERLOADED_BITS     (0xffffULL << 48)
 108 #define BIT(name)               (1ULL << KPF_##name)
 109 #define BITS_COMPOUND           (BIT(COMPOUND_HEAD) | BIT(COMPOUND_TAIL))
 110
 111 static const char * const page_flag_names[] = {
 112         [KPF_LOCKED]            = "L:locked",
 113         [KPF_ERROR]             = "E:error",
 114         [KPF_REFERENCED]        = "R:referenced",
 115         [KPF_UPTODATE]          = "U:uptodate",
 116         [KPF_DIRTY]             = "D:dirty",
 117         [KPF_LRU]               = "l:lru",
 118         [KPF_ACTIVE]            = "A:active",
 119         [KPF_SLAB]              = "S:slab",
 120         [KPF_WRITEBACK]         = "W:writeback",
 121         [KPF_RECLAIM]           = "I:reclaim",
 122         [KPF_BUDDY]             = "B:buddy",
 123
 124         [KPF_MMAP]              = "M:mmap",
 125         [KPF_ANON]              = "a:anonymous",
 126         [KPF_SWAPCACHE]         = "s:swapcache",
 127         [KPF_SWAPBACKED]        = "b:swapbacked",
 128         [KPF_COMPOUND_HEAD]     = "H:compound_head",
 129         [KPF_COMPOUND_TAIL]     = "T:compound_tail",
 130         [KPF_HUGE]              = "G:huge",
 131         [KPF_UNEVICTABLE]       = "u:unevictable",
 132         [KPF_HWPOISON]          = "X:hwpoison",
 133         [KPF_NOPAGE]            = "n:nopage",
 134         [KPF_KSM]               = "x:ksm",
 135         [KPF_THP]               = "t:thp",
 136         [KPF_BALLOON]           = "o:balloon",
 137         [KPF_PGTABLE]           = "g:pgtable",
 138         [KPF_ZERO_PAGE]         = "z:zero_page",
 139         [KPF_IDLE]              = "i:idle_page",
 140
 141         [KPF_RESERVED]          = "r:reserved",
 142         [KPF_MLOCKED]           = "m:mlocked",
 143         [KPF_MAPPEDTODISK]      = "d:mappedtodisk",
 144         [KPF_PRIVATE]           = "P:private",
 145         [KPF_PRIVATE_2]         = "p:private_2",
 146         [KPF_OWNER_PRIVATE]     = "O:owner_private",
 147         [KPF_ARCH]              = "h:arch",
 148         [KPF_UNCACHED]          = "c:uncached",
 149         [KPF_SOFTDIRTY]         = "f:softdirty",
 150
 151         [KPF_READAHEAD]         = "I:readahead",
 152         [KPF_SLOB_FREE]         = "P:slob_free",
 153         [KPF_SLUB_FROZEN]       = "A:slub_frozen",
 154         [KPF_SLUB_DEBUG]        = "E:slub_debug",
 155
 156         [KPF_FILE]              = "F:file",
 157         [KPF_SWAP]              = "w:swap",
 158         [KPF_MMAP_EXCLUSIVE]    = "1:mmap_exclusive",
 159 };
 160
 161
 162 /*
 163  * data structures
 164  */
 165
 166 static int              opt_raw;        /* for kernel developers */
 167 static int              opt_list;       /* list pages (in ranges) */
 168 static int              opt_mark_idle;  /* set accessed bit */
 169 static int              opt_no_summary; /* don't show summary */
 170 static pid_t            opt_pid;        /* process to walk */
 171 const char              *opt_file;      /* file or directory path */
 172 static uint64_t         opt_cgroup;     /* cgroup inode */
 173 static int              opt_list_cgroup;/* list page cgroup */
 174 static int              opt_list_mapcnt;/* list page map count */
 175 static const char       *opt_kpageflags;/* kpageflags file to parse */
 176
 177 #define MAX_ADDR_RANGES 1024
 178 static int              nr_addr_ranges;
 179 static unsigned long    opt_offset[MAX_ADDR_RANGES];
 180 static unsigned long    opt_size[MAX_ADDR_RANGES];
 181
 182 #define MAX_VMAS        10240
 183 static int              nr_vmas;
 184 static unsigned long    pg_start[MAX_VMAS];
 185 static unsigned long    pg_end[MAX_VMAS];
 186
 187 #define MAX_BIT_FILTERS 64
 188 static int              nr_bit_filters;
 189 static uint64_t         opt_mask[MAX_BIT_FILTERS];
 190 static uint64_t         opt_bits[MAX_BIT_FILTERS];
 191
 192 static int              page_size;
 193
 194 static int              pagemap_fd;
 195 static int              kpageflags_fd;
 196 static int              kpagecount_fd = -1;
 197 static int              kpagecgroup_fd = -1;
 198 static int              page_idle_fd = -1;
 199
 200 static int              opt_hwpoison;
 201 static int              opt_unpoison;
 202
 203 static const char       *hwpoison_debug_fs;
 204 static int              hwpoison_inject_fd;
 205 static int              hwpoison_forget_fd;
 206
 207 #define HASH_SHIFT      13
 208 #define HASH_SIZE       (1 << HASH_SHIFT)
 209 #define HASH_MASK       (HASH_SIZE - 1)
 210 #define HASH_KEY(flags) (flags & HASH_MASK)
 211
 212 static unsigned long    total_pages;
 213 static unsigned long    nr_pages[HASH_SIZE];
 214 static uint64_t         page_flags[HASH_SIZE];
 215
 216
 217 /*
 218  * helper functions
 219  */
 220
 221 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 222
 223 #define min_t(type, x, y) ({                    \
 224         type __min1 = (x);                      \
 225         type __min2 = (y);                      \
 226         __min1 < __min2 ? __min1 : __min2; })
 227
 228 #define max_t(type, x, y) ({                    \
 229         type __max1 = (x);                      \
 230         type __max2 = (y);                      \
 231         __max1 > __max2 ? __max1 : __max2; })
 232
 233 static unsigned long pages2mb(unsigned long pages)
 234 {
 235         return (pages * page_size) >> 20;
 236 }
 237
 238 static void fatal(const char *x, ...)
 239 {
 240         va_list ap;
 241
 242         va_start(ap, x);
 243         vfprintf(stderr, x, ap);
 244         va_end(ap);
 245         exit(EXIT_FAILURE);
 246 }
 247
 248 static int checked_open(const char *pathname, int flags)
 249 {
 250         int fd = open(pathname, flags);
 251
 252         if (fd < 0) {
 253                 perror(pathname);
 254                 exit(EXIT_FAILURE);
 255         }
 256
 257         return fd;
 258 }
 259
 260 /*
 261  * pagemap/kpageflags routines
 262  */
 263
 264 static unsigned long do_u64_read(int fd, const char *name,
 265                                  uint64_t *buf,
 266                                  unsigned long index,
 267                                  unsigned long count)
 268 {
 269         long bytes;
 270
 271         if (index > ULONG_MAX / 8)
 272                 fatal("index overflow: %lu\n", index);
 273
 274         bytes = pread(fd, buf, count * 8, (off_t)index * 8);
 275         if (bytes < 0) {
 276                 perror(name);
 277                 exit(EXIT_FAILURE);
 278         }
 279         if (bytes % 8)
 280                 fatal("partial read: %lu bytes\n", bytes);
 281
 282         return bytes / 8;
 283 }
 284
 285 static unsigned long kpageflags_read(uint64_t *buf,
 286                                      unsigned long index,
 287                                      unsigned long pages)
 288 {
 289         return do_u64_read(kpageflags_fd, opt_kpageflags, buf, index, pages);
 290 }
 291
 292 static unsigned long kpagecgroup_read(uint64_t *buf,
 293                                       unsigned long index,
 294                                       unsigned long pages)
 295 {
 296         if (kpagecgroup_fd < 0)
 297                 return pages;
 298
 299         return do_u64_read(kpagecgroup_fd, opt_kpageflags, buf, index, pages);
 300 }
 301
 302 static unsigned long kpagecount_read(uint64_t *buf,
 303                                      unsigned long index,
 304                                      unsigned long pages)
 305 {
 306         return kpagecount_fd < 0 ? pages :
 307                 do_u64_read(kpagecount_fd, PROC_KPAGECOUNT,
 308                             buf, index, pages);
 309 }
 310
 311 static unsigned long pagemap_read(uint64_t *buf,
 312                                   unsigned long index,
 313                                   unsigned long pages)
 314 {
 315         return do_u64_read(pagemap_fd, "/proc/pid/pagemap", buf, index, pages);
 316 }
 317
 318 static unsigned long pagemap_pfn(uint64_t val)
 319 {
 320         unsigned long pfn;
 321
 322         if (val & PM_PRESENT)
 323                 pfn = PM_PFRAME(val);
 324         else
 325                 pfn = 0;
 326
 327         return pfn;
 328 }
 329
 330 static unsigned long pagemap_swap_offset(uint64_t val)
 331 {
 332         return val & PM_SWAP ? PM_SWAP_OFFSET(val) : 0;
 333 }
 334
 335 /*
 336  * page flag names
 337  */
 338
 339 static char *page_flag_name(uint64_t flags)
 340 {
 341         static char buf[65];
 342         int present;
 343         size_t i, j;
 344
 345         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 346                 present = (flags >> i) & 1;
 347                 if (!page_flag_names[i]) {
 348                         if (present)
 349                                 fatal("unknown flag bit %d\n", i);
 350                         continue;
 351                 }
 352                 buf[j++] = present ? page_flag_names[i][0] : '_';
 353         }
 354
 355         return buf;
 356 }
 357
 358 static char *page_flag_longname(uint64_t flags)
 359 {
 360         static char buf[1024];
 361         size_t i, n;
 362
 363         for (i = 0, n = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 364                 if (!page_flag_names[i])
 365                         continue;
 366                 if ((flags >> i) & 1)
 367                         n += snprintf(buf + n, sizeof(buf) - n, "%s,",
 368                                         page_flag_names[i] + 2);
 369         }
 370         if (n)
 371                 n--;
 372         buf[n] = '\0';
 373
 374         return buf;
 375 }
 376
 377
 378 /*
 379  * page list and summary
 380  */
 381
 382 static void show_page_range(unsigned long voffset, unsigned long offset,
 383                             unsigned long size, uint64_t flags,
 384                             uint64_t cgroup, uint64_t mapcnt)
 385 {
 386         static uint64_t      flags0;
 387         static uint64_t      cgroup0;
 388         static uint64_t      mapcnt0;
 389         static unsigned long voff;
 390         static unsigned long index;
 391         static unsigned long count;
 392
 393         if (flags == flags0 && cgroup == cgroup0 && mapcnt == mapcnt0 &&
 394             offset == index + count && size && voffset == voff + count) {
 395                 count += size;
 396                 return;
 397         }
 398
 399         if (count) {
 400                 if (opt_pid)
 401                         printf("%lx\t", voff);
 402                 if (opt_file)
 403                         printf("%lu\t", voff);
 404                 if (opt_list_cgroup)
 405                         printf("@%llu\t", (unsigned long long)cgroup0);
 406                 if (opt_list_mapcnt)
 407                         printf("%lu\t", mapcnt0);
 408                 printf("%lx\t%lx\t%s\n",
 409                                 index, count, page_flag_name(flags0));
 410         }
 411
 412         flags0 = flags;
 413         cgroup0 = cgroup;
 414         mapcnt0 = mapcnt;
 415         index  = offset;
 416         voff   = voffset;
 417         count  = size;
 418 }
 419
 420 static void flush_page_range(void)
 421 {
 422         show_page_range(0, 0, 0, 0, 0, 0);
 423 }
 424
 425 static void show_page(unsigned long voffset, unsigned long offset,
 426                       uint64_t flags, uint64_t cgroup, uint64_t mapcnt)
 427 {
 428         if (opt_pid)
 429                 printf("%lx\t", voffset);
 430         if (opt_file)
 431                 printf("%lu\t", voffset);
 432         if (opt_list_cgroup)
 433                 printf("@%llu\t", (unsigned long long)cgroup);
 434         if (opt_list_mapcnt)
 435                 printf("%lu\t", mapcnt);
 436
 437         printf("%lx\t%s\n", offset, page_flag_name(flags));
 438 }
 439
 440 static void show_summary(void)
 441 {
 442         size_t i;
 443
 444         printf("             flags\tpage-count       MB"
 445                 "  symbolic-flags\t\t\tlong-symbolic-flags\n");
 446
 447         for (i = 0; i < ARRAY_SIZE(nr_pages); i++) {
 448                 if (nr_pages[i])
 449                         printf("0x%016llx\t%10lu %8lu  %s\t%s\n",
 450                                 (unsigned long long)page_flags[i],
 451                                 nr_pages[i],
 452                                 pages2mb(nr_pages[i]),
 453                                 page_flag_name(page_flags[i]),
 454                                 page_flag_longname(page_flags[i]));
 455         }
 456
 457         printf("             total\t%10lu %8lu\n",
 458                         total_pages, pages2mb(total_pages));
 459 }
 460
 461
 462 /*
 463  * page flag filters
 464  */
 465
 466 static int bit_mask_ok(uint64_t flags)
 467 {
 468         int i;
 469
 470         for (i = 0; i < nr_bit_filters; i++) {
 471                 if (opt_bits[i] == KPF_ALL_BITS) {
 472                         if ((flags & opt_mask[i]) == 0)
 473                                 return 0;
 474                 } else {
 475                         if ((flags & opt_mask[i]) != opt_bits[i])
 476                                 return 0;
 477                 }
 478         }
 479
 480         return 1;
 481 }
 482
 483 static uint64_t expand_overloaded_flags(uint64_t flags, uint64_t pme)
 484 {
 485         /* SLOB/SLUB overload several page flags */
 486         if (flags & BIT(SLAB)) {
 487                 if (flags & BIT(PRIVATE))
 488                         flags ^= BIT(PRIVATE) | BIT(SLOB_FREE);
 489                 if (flags & BIT(ACTIVE))
 490                         flags ^= BIT(ACTIVE) | BIT(SLUB_FROZEN);
 491                 if (flags & BIT(ERROR))
 492                         flags ^= BIT(ERROR) | BIT(SLUB_DEBUG);
 493         }
 494
 495         /* PG_reclaim is overloaded as PG_readahead in the read path */
 496         if ((flags & (BIT(RECLAIM) | BIT(WRITEBACK))) == BIT(RECLAIM))
 497                 flags ^= BIT(RECLAIM) | BIT(READAHEAD);
 498
 499         if (pme & PM_SOFT_DIRTY)
 500                 flags |= BIT(SOFTDIRTY);
 501         if (pme & PM_FILE)
 502                 flags |= BIT(FILE);
 503         if (pme & PM_SWAP)
 504                 flags |= BIT(SWAP);
 505         if (pme & PM_MMAP_EXCLUSIVE)
 506                 flags |= BIT(MMAP_EXCLUSIVE);
 507
 508         return flags;
 509 }
 510
 511 static uint64_t well_known_flags(uint64_t flags)
 512 {
 513         /* hide flags intended only for kernel hacker */
 514         flags &= ~KPF_HACKERS_BITS;
 515
 516         /* hide non-hugeTLB compound pages */
 517         if ((flags & BITS_COMPOUND) && !(flags & BIT(HUGE)))
 518                 flags &= ~BITS_COMPOUND;
 519
 520         return flags;
 521 }
 522
 523 static uint64_t kpageflags_flags(uint64_t flags, uint64_t pme)
 524 {
 525         if (opt_raw)
 526                 flags = expand_overloaded_flags(flags, pme);
 527         else
 528                 flags = well_known_flags(flags);
 529
 530         return flags;
 531 }
 532
 533 /*
 534  * page actions
 535  */
 536
 537 static void prepare_hwpoison_fd(void)
 538 {
 539         char buf[MAX_PATH + 1];
 540
 541         hwpoison_debug_fs = debugfs__mount();
 542         if (!hwpoison_debug_fs) {
 543                 perror("mount debugfs");
 544                 exit(EXIT_FAILURE);
 545         }
 546
 547         if (opt_hwpoison && !hwpoison_inject_fd) {
 548                 snprintf(buf, MAX_PATH, "%s/hwpoison/corrupt-pfn",
 549                         hwpoison_debug_fs);
 550                 hwpoison_inject_fd = checked_open(buf, O_WRONLY);
 551         }
 552
 553         if (opt_unpoison && !hwpoison_forget_fd) {
 554                 snprintf(buf, MAX_PATH, "%s/hwpoison/unpoison-pfn",
 555                         hwpoison_debug_fs);
 556                 hwpoison_forget_fd = checked_open(buf, O_WRONLY);
 557         }
 558 }
 559
 560 static int hwpoison_page(unsigned long offset)
 561 {
 562         char buf[100];
 563         int len;
 564
 565         len = sprintf(buf, "0x%lx\n", offset);
 566         len = write(hwpoison_inject_fd, buf, len);
 567         if (len < 0) {
 568                 perror("hwpoison inject");
 569                 return len;
 570         }
 571         return 0;
 572 }
 573
 574 static int unpoison_page(unsigned long offset)
 575 {
 576         char buf[100];
 577         int len;
 578
 579         len = sprintf(buf, "0x%lx\n", offset);
 580         len = write(hwpoison_forget_fd, buf, len);
 581         if (len < 0) {
 582                 perror("hwpoison forget");
 583                 return len;
 584         }
 585         return 0;
 586 }
 587
 588 static int mark_page_idle(unsigned long offset)
 589 {
 590         static unsigned long off;
 591         static uint64_t buf;
 592         int len;
 593
 594         if ((offset / 64 == off / 64) || buf == 0) {
 595                 buf |= 1UL << (offset % 64);
 596                 off = offset;
 597                 return 0;
 598         }
 599
 600         len = pwrite(page_idle_fd, &buf, 8, 8 * (off / 64));
 601         if (len < 0) {
 602                 perror("mark page idle");
 603                 return len;
 604         }
 605
 606         buf = 1UL << (offset % 64);
 607         off = offset;
 608
 609         return 0;
 610 }
 611
 612 /*
 613  * page frame walker
 614  */
 615
 616 static size_t hash_slot(uint64_t flags)
 617 {
 618         size_t k = HASH_KEY(flags);
 619         size_t i;
 620
 621         /* Explicitly reserve slot 0 for flags 0: the following logic
 622          * cannot distinguish an unoccupied slot from slot (flags==0).
 623          */
 624         if (flags == 0)
 625                 return 0;
 626
 627         /* search through the remaining (HASH_SIZE-1) slots */
 628         for (i = 1; i < ARRAY_SIZE(page_flags); i++, k++) {
 629                 if (!k || k >= ARRAY_SIZE(page_flags))
 630                         k = 1;
 631                 if (page_flags[k] == 0) {
 632                         page_flags[k] = flags;
 633                         return k;
 634                 }
 635                 if (page_flags[k] == flags)
 636                         return k;
 637         }
 638
 639         fatal("hash table full: bump up HASH_SHIFT?\n");
 640         exit(EXIT_FAILURE);
 641 }
 642
 643 static void add_page(unsigned long voffset, unsigned long offset,
 644                      uint64_t flags, uint64_t cgroup, uint64_t mapcnt,
 645                      uint64_t pme)
 646 {
 647         flags = kpageflags_flags(flags, pme);
 648
 649         if (!bit_mask_ok(flags))
 650                 return;
 651
 652         if (opt_cgroup && cgroup != (uint64_t)opt_cgroup)
 653                 return;
 654
 655         if (opt_hwpoison)
 656                 hwpoison_page(offset);
 657         if (opt_unpoison)
 658                 unpoison_page(offset);
 659
 660         if (opt_mark_idle)
 661                 mark_page_idle(offset);
 662
 663         if (opt_list == 1)
 664                 show_page_range(voffset, offset, 1, flags, cgroup, mapcnt);
 665         else if (opt_list == 2)
 666                 show_page(voffset, offset, flags, cgroup, mapcnt);
 667
 668         nr_pages[hash_slot(flags)]++;
 669         total_pages++;
 670 }
 671
 672 #define KPAGEFLAGS_BATCH        (64 << 10)      /* 64k pages */
 673 static void walk_pfn(unsigned long voffset,
 674                      unsigned long index,
 675                      unsigned long count,
 676                      uint64_t pme)
 677 {
 678         uint64_t buf[KPAGEFLAGS_BATCH];
 679         uint64_t cgi[KPAGEFLAGS_BATCH];
 680         uint64_t cnt[KPAGEFLAGS_BATCH];
 681         unsigned long batch;
 682         unsigned long pages;
 683         unsigned long i;
 684
 685         /*
 686          * kpagecgroup_read() reads only if kpagecgroup were opened, but
 687          * /proc/kpagecgroup might even not exist, so it's better to fill
 688          * them with zeros here.
 689          */
 690         if (count == 1)
 691                 cgi[0] = 0;
 692         else
 693                 memset(cgi, 0, sizeof cgi);
 694
 695         while (count) {
 696                 batch = min_t(unsigned long, count, KPAGEFLAGS_BATCH);
 697                 pages = kpageflags_read(buf, index, batch);
 698                 if (pages == 0)
 699                         break;
 700
 701                 if (kpagecgroup_read(cgi, index, pages) != pages)
 702                         fatal("kpagecgroup returned fewer pages than expected");
 703
 704                 if (kpagecount_read(cnt, index, pages) != pages)
 705                         fatal("kpagecount returned fewer pages than expected");
 706
 707                 for (i = 0; i < pages; i++)
 708                         add_page(voffset + i, index + i,
 709                                  buf[i], cgi[i], cnt[i], pme);
 710
 711                 index += pages;
 712                 count -= pages;
 713         }
 714 }
 715
 716 static void walk_swap(unsigned long voffset, uint64_t pme)
 717 {
 718         uint64_t flags = kpageflags_flags(0, pme);
 719
 720         if (!bit_mask_ok(flags))
 721                 return;
 722
 723         if (opt_cgroup)
 724                 return;
 725
 726         if (opt_list == 1)
 727                 show_page_range(voffset, pagemap_swap_offset(pme),
 728                                 1, flags, 0, 0);
 729         else if (opt_list == 2)
 730                 show_page(voffset, pagemap_swap_offset(pme), flags, 0, 0);
 731
 732         nr_pages[hash_slot(flags)]++;
 733         total_pages++;
 734 }
 735
 736 #define PAGEMAP_BATCH   (64 << 10)
 737 static void walk_vma(unsigned long index, unsigned long count)
 738 {
 739         uint64_t buf[PAGEMAP_BATCH];
 740         unsigned long batch;
 741         unsigned long pages;
 742         unsigned long pfn;
 743         unsigned long i;
 744
 745         while (count) {
 746                 batch = min_t(unsigned long, count, PAGEMAP_BATCH);
 747                 pages = pagemap_read(buf, index, batch);
 748                 if (pages == 0)
 749                         break;
 750
 751                 for (i = 0; i < pages; i++) {
 752                         pfn = pagemap_pfn(buf[i]);
 753                         if (pfn)
 754                                 walk_pfn(index + i, pfn, 1, buf[i]);
 755                         if (buf[i] & PM_SWAP)
 756                                 walk_swap(index + i, buf[i]);
 757                 }
 758
 759                 index += pages;
 760                 count -= pages;
 761         }
 762 }
 763
 764 static void walk_task(unsigned long index, unsigned long count)
 765 {
 766         const unsigned long end = index + count;
 767         unsigned long start;
 768         int i = 0;
 769
 770         while (index < end) {
 771
 772                 while (pg_end[i] <= index)
 773                         if (++i >= nr_vmas)
 774                                 return;
 775                 if (pg_start[i] >= end)
 776                         return;
 777
 778                 start = max_t(unsigned long, pg_start[i], index);
 779                 index = min_t(unsigned long, pg_end[i], end);
 780
 781                 assert(start < index);
 782                 walk_vma(start, index - start);
 783         }
 784 }
 785
 786 static void add_addr_range(unsigned long offset, unsigned long size)
 787 {
 788         if (nr_addr_ranges >= MAX_ADDR_RANGES)
 789                 fatal("too many addr ranges\n");
 790
 791         opt_offset[nr_addr_ranges] = offset;
 792         opt_size[nr_addr_ranges] = min_t(unsigned long, size, ULONG_MAX-offset);
 793         nr_addr_ranges++;
 794 }
 795
 796 static void walk_addr_ranges(void)
 797 {
 798         int i;
 799
 800         kpageflags_fd = checked_open(opt_kpageflags, O_RDONLY);
 801
 802         if (!nr_addr_ranges)
 803                 add_addr_range(0, ULONG_MAX);
 804
 805         for (i = 0; i < nr_addr_ranges; i++)
 806                 if (!opt_pid)
 807                         walk_pfn(opt_offset[i], opt_offset[i], opt_size[i], 0);
 808                 else
 809                         walk_task(opt_offset[i], opt_size[i]);
 810
 811         if (opt_mark_idle)
 812                 mark_page_idle(0);
 813
 814         close(kpageflags_fd);
 815 }
 816
 817
 818 /*
 819  * user interface
 820  */
 821
 822 static const char *page_flag_type(uint64_t flag)
 823 {
 824         if (flag & KPF_HACKERS_BITS)
 825                 return "(r)";
 826         if (flag & KPF_OVERLOADED_BITS)
 827                 return "(o)";
 828         return "   ";
 829 }
 830
 831 static void usage(void)
 832 {
 833         size_t i, j;
 834
 835         printf(
 836 "page-types [options]\n"
 837 "            -r|--raw                   Raw mode, for kernel developers\n"
 838 "            -d|--describe flags        Describe flags\n"
 839 "            -a|--addr    addr-spec     Walk a range of pages\n"
 840 "            -b|--bits    bits-spec     Walk pages with specified bits\n"
 841 "            -c|--cgroup  path|@inode   Walk pages within memory cgroup\n"
 842 "            -p|--pid     pid           Walk process address space\n"
 843 "            -f|--file    filename      Walk file address space\n"
 844 "            -i|--mark-idle             Mark pages idle\n"
 845 "            -l|--list                  Show page details in ranges\n"
 846 "            -L|--list-each             Show page details one by one\n"
 847 "            -C|--list-cgroup           Show cgroup inode for pages\n"
 848 "            -M|--list-mapcnt           Show page map count\n"
 849 "            -N|--no-summary            Don't show summary info\n"
 850 "            -X|--hwpoison              hwpoison pages\n"
 851 "            -x|--unpoison              unpoison pages\n"
 852 "            -F|--kpageflags filename   kpageflags file to parse\n"
 853 "            -h|--help                  Show this usage message\n"
 854 "flags:\n"
 855 "            0x10                       bitfield format, e.g.\n"
 856 "            anon                       bit-name, e.g.\n"
 857 "            0x10,anon                  comma-separated list, e.g.\n"
 858 "addr-spec:\n"
 859 "            N                          one page at offset N (unit: pages)\n"
 860 "            N+M                        pages range from N to N+M-1\n"
 861 "            N,M                        pages range from N to M-1\n"
 862 "            N,                         pages range from N to end\n"
 863 "            ,M                         pages range from 0 to M-1\n"
 864 "bits-spec:\n"
 865 "            bit1,bit2                  (flags & (bit1|bit2)) != 0\n"
 866 "            bit1,bit2=bit1             (flags & (bit1|bit2)) == bit1\n"
 867 "            bit1,~bit2                 (flags & (bit1|bit2)) == bit1\n"
 868 "            =bit1,bit2                 flags == (bit1|bit2)\n"
 869 "bit-names:\n"
 870         );
 871
 872         for (i = 0, j = 0; i < ARRAY_SIZE(page_flag_names); i++) {
 873                 if (!page_flag_names[i])
 874                         continue;
 875                 printf("%16s%s", page_flag_names[i] + 2,
 876                                  page_flag_type(1ULL << i));
 877                 if (++j > 3) {
 878                         j = 0;
 879                         putchar('\n');
 880                 }
 881         }
 882         printf("\n                                   "
 883                 "(r) raw mode bits  (o) overloaded bits\n");
 884 }
 885
 886 static unsigned long long parse_number(const char *str)
 887 {
 888         unsigned long long n;
 889
 890         n = strtoll(str, NULL, 0);
 891
 892         if (n == 0 && str[0] != '0')
 893                 fatal("invalid name or number: %s\n", str);
 894
 895         return n;
 896 }
 897
 898 static void parse_pid(const char *str)
 899 {
 900         FILE *file;
 901         char buf[5000];
 902
 903         opt_pid = parse_number(str);
 904
 905         sprintf(buf, "/proc/%d/pagemap", opt_pid);
 906         pagemap_fd = checked_open(buf, O_RDONLY);
 907
 908         sprintf(buf, "/proc/%d/maps", opt_pid);
 909         file = fopen(buf, "r");
 910         if (!file) {
 911                 perror(buf);
 912                 exit(EXIT_FAILURE);
 913         }
 914
 915         while (fgets(buf, sizeof(buf), file) != NULL) {
 916                 unsigned long vm_start;
 917                 unsigned long vm_end;
 918                 unsigned long long pgoff;
 919                 int major, minor;
 920                 char r, w, x, s;
 921                 unsigned long ino;
 922                 int n;
 923
 924                 n = sscanf(buf, "%lx-%lx %c%c%c%c %llx %x:%x %lu",
 925                            &vm_start,
 926                            &vm_end,
 927                            &r, &w, &x, &s,
 928                            &pgoff,
 929                            &major, &minor,
 930                            &ino);
 931                 if (n < 10) {
 932                         fprintf(stderr, "unexpected line: %s\n", buf);
 933                         continue;
 934                 }
 935                 pg_start[nr_vmas] = vm_start / page_size;
 936                 pg_end[nr_vmas] = vm_end / page_size;
 937                 if (++nr_vmas >= MAX_VMAS) {
 938                         fprintf(stderr, "too many VMAs\n");
 939                         break;
 940                 }
 941         }
 942         fclose(file);
 943 }
 944
 945 static void show_file(const char *name, const struct stat *st)
 946 {
 947         unsigned long long size = st->st_size;
 948         char atime[64], mtime[64];
 949         long now = time(NULL);
 950
 951         printf("%s\tInode: %u\tSize: %llu (%llu pages)\n",
 952                         name, (unsigned)st->st_ino,
 953                         size, (size + page_size - 1) / page_size);
 954
 955         strftime(atime, sizeof(atime), "%c", localtime(&st->st_atime));
 956         strftime(mtime, sizeof(mtime), "%c", localtime(&st->st_mtime));
 957
 958         printf("Modify: %s (%ld seconds ago)\nAccess: %s (%ld seconds ago)\n",
 959                         mtime, now - st->st_mtime,
 960                         atime, now - st->st_atime);
 961 }
 962
 963 static sigjmp_buf sigbus_jmp;
 964
 965 static void * volatile sigbus_addr;
 966
 967 static void sigbus_handler(int sig, siginfo_t *info, void *ucontex)
 968 {
 969         (void)sig;
 970         (void)ucontex;
 971         sigbus_addr = info ? info->si_addr : NULL;
 972         siglongjmp(sigbus_jmp, 1);
 973 }
 974
 975 static struct sigaction sigbus_action = {
 976         .sa_sigaction = sigbus_handler,
 977         .sa_flags = SA_SIGINFO,
 978 };
 979
 980 static void walk_file(const char *name, const struct stat *st)
 981 {
 982         uint8_t vec[PAGEMAP_BATCH];
 983         uint64_t buf[PAGEMAP_BATCH], flags;
 984         uint64_t cgroup = 0;
 985         uint64_t mapcnt = 0;
 986         unsigned long nr_pages, pfn, i;
 987         off_t off, end = st->st_size;
 988         int fd;
 989         ssize_t len;
 990         void *ptr;
 991         int first = 1;
 992
 993         fd = checked_open(name, O_RDONLY|O_NOATIME|O_NOFOLLOW);
 994
 995         for (off = 0; off < end; off += len) {
 996                 nr_pages = (end - off + page_size - 1) / page_size;
 997                 if (nr_pages > PAGEMAP_BATCH)
 998                         nr_pages = PAGEMAP_BATCH;
 999                 len = nr_pages * page_size;
1000
1001                 ptr = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, off);
1002                 if (ptr == MAP_FAILED)
1003                         fatal("mmap failed: %s", name);
1004
1005                 /* determine cached pages */
1006                 if (mincore(ptr, len, vec))
1007                         fatal("mincore failed: %s", name);
1008
1009                 /* turn off readahead */
1010                 if (madvise(ptr, len, MADV_RANDOM))
1011                         fatal("madvice failed: %s", name);
1012
1013                 if (sigsetjmp(sigbus_jmp, 1)) {
1014                         end = off + sigbus_addr ? sigbus_addr - ptr : 0;
1015                         fprintf(stderr, "got sigbus at offset %lld: %s\n",
1016                                         (long long)end, name);
1017                         goto got_sigbus;
1018                 }
1019
1020                 /* populate ptes */
1021                 for (i = 0; i < nr_pages ; i++) {
1022                         if (vec[i] & 1)
1023                                 (void)*(volatile int *)(ptr + i * page_size);
1024                 }
1025 got_sigbus:
1026
1027                 /* turn off harvesting reference bits */
1028                 if (madvise(ptr, len, MADV_SEQUENTIAL))
1029                         fatal("madvice failed: %s", name);
1030
1031                 if (pagemap_read(buf, (unsigned long)ptr / page_size,
1032                                         nr_pages) != nr_pages)
1033                         fatal("cannot read pagemap");
1034
1035                 munmap(ptr, len);
1036
1037                 for (i = 0; i < nr_pages; i++) {
1038                         pfn = pagemap_pfn(buf[i]);
1039                         if (!pfn)
1040                                 continue;
1041                         if (!kpageflags_read(&flags, pfn, 1))
1042                                 continue;
1043                         if (!kpagecgroup_read(&cgroup, pfn, 1))
1044                                 fatal("kpagecgroup_read failed");
1045                         if (!kpagecount_read(&mapcnt, pfn, 1))
1046                                 fatal("kpagecount_read failed");
1047                         if (first && opt_list) {
1048                                 first = 0;
1049                                 flush_page_range();
1050                                 show_file(name, st);
1051                         }
1052                         add_page(off / page_size + i, pfn,
1053                                  flags, cgroup, mapcnt, buf[i]);
1054                 }
1055         }
1056
1057         close(fd);
1058 }
1059
1060 int walk_tree(const char *name, const struct stat *st, int type, struct FTW *f)
1061 {
1062         (void)f;
1063         switch (type) {
1064         case FTW_F:
1065                 if (S_ISREG(st->st_mode))
1066                         walk_file(name, st);
1067                 break;
1068         case FTW_DNR:
1069                 fprintf(stderr, "cannot read dir: %s\n", name);
1070                 break;
1071         }
1072         return 0;
1073 }
1074
1075 static void walk_page_cache(void)
1076 {
1077         struct stat st;
1078
1079         kpageflags_fd = checked_open(opt_kpageflags, O_RDONLY);
1080         pagemap_fd = checked_open("/proc/self/pagemap", O_RDONLY);
1081         sigaction(SIGBUS, &sigbus_action, NULL);
1082
1083         if (stat(opt_file, &st))
1084                 fatal("stat failed: %s\n", opt_file);
1085
1086         if (S_ISREG(st.st_mode)) {
1087                 walk_file(opt_file, &st);
1088         } else if (S_ISDIR(st.st_mode)) {
1089                 /* do not follow symlinks and mountpoints */
1090                 if (nftw(opt_file, walk_tree, 64, FTW_MOUNT | FTW_PHYS) < 0)
1091                         fatal("nftw failed: %s\n", opt_file);
1092         } else
1093                 fatal("unhandled file type: %s\n", opt_file);
1094
1095         close(kpageflags_fd);
1096         close(pagemap_fd);
1097         signal(SIGBUS, SIG_DFL);
1098 }
1099
1100 static void parse_file(const char *name)
1101 {
1102         opt_file = name;
1103 }
1104
1105 static void parse_cgroup(const char *path)
1106 {
1107         if (path[0] == '@') {
1108                 opt_cgroup = parse_number(path + 1);
1109                 return;
1110         }
1111
1112         struct stat st;
1113
1114         if (stat(path, &st))
1115                 fatal("stat failed: %s: %m\n", path);
1116
1117         if (!S_ISDIR(st.st_mode))
1118                 fatal("cgroup supposed to be a directory: %s\n", path);
1119
1120         opt_cgroup = st.st_ino;
1121 }
1122
1123 static void parse_addr_range(const char *optarg)
1124 {
1125         unsigned long offset;
1126         unsigned long size;
1127         char *p;
1128
1129         p = strchr(optarg, ',');
1130         if (!p)
1131                 p = strchr(optarg, '+');
1132
1133         if (p == optarg) {
1134                 offset = 0;
1135                 size   = parse_number(p + 1);
1136         } else if (p) {
1137                 offset = parse_number(optarg);
1138                 if (p[1] == '\0')
1139                         size = ULONG_MAX;
1140                 else {
1141                         size = parse_number(p + 1);
1142                         if (*p == ',') {
1143                                 if (size < offset)
1144                                         fatal("invalid range: %lu,%lu\n",
1145                                                         offset, size);
1146                                 size -= offset;
1147                         }
1148                 }
1149         } else {
1150                 offset = parse_number(optarg);
1151                 size   = 1;
1152         }
1153
1154         add_addr_range(offset, size);
1155 }
1156
1157 static void add_bits_filter(uint64_t mask, uint64_t bits)
1158 {
1159         if (nr_bit_filters >= MAX_BIT_FILTERS)
1160                 fatal("too much bit filters\n");
1161
1162         opt_mask[nr_bit_filters] = mask;
1163         opt_bits[nr_bit_filters] = bits;
1164         nr_bit_filters++;
1165 }
1166
1167 static uint64_t parse_flag_name(const char *str, int len)
1168 {
1169         size_t i;
1170
1171         if (!*str || !len)
1172                 return 0;
1173
1174         if (len <= 8 && !strncmp(str, "compound", len))
1175                 return BITS_COMPOUND;
1176
1177         for (i = 0; i < ARRAY_SIZE(page_flag_names); i++) {
1178                 if (!page_flag_names[i])
1179                         continue;
1180                 if (!strncmp(str, page_flag_names[i] + 2, len))
1181                         return 1ULL << i;
1182         }
1183
1184         return parse_number(str);
1185 }
1186
1187 static uint64_t parse_flag_names(const char *str, int all)
1188 {
1189         const char *p    = str;
1190         uint64_t   flags = 0;
1191
1192         while (1) {
1193                 if (*p == ',' || *p == '=' || *p == '\0') {
1194                         if ((*str != '~') || (*str == '~' && all && *++str))
1195                                 flags |= parse_flag_name(str, p - str);
1196                         if (*p != ',')
1197                                 break;
1198                         str = p + 1;
1199                 }
1200                 p++;
1201         }
1202
1203         return flags;
1204 }
1205
1206 static void parse_bits_mask(const char *optarg)
1207 {
1208         uint64_t mask;
1209         uint64_t bits;
1210         const char *p;
1211
1212         p = strchr(optarg, '=');
1213         if (p == optarg) {
1214                 mask = KPF_ALL_BITS;
1215                 bits = parse_flag_names(p + 1, 0);
1216         } else if (p) {
1217                 mask = parse_flag_names(optarg, 0);
1218                 bits = parse_flag_names(p + 1, 0);
1219         } else if (strchr(optarg, '~')) {
1220                 mask = parse_flag_names(optarg, 1);
1221                 bits = parse_flag_names(optarg, 0);
1222         } else {
1223                 mask = parse_flag_names(optarg, 0);
1224                 bits = KPF_ALL_BITS;
1225         }
1226
1227         add_bits_filter(mask, bits);
1228 }
1229
1230 static void parse_kpageflags(const char *name)
1231 {
1232         opt_kpageflags = name;
1233 }
1234
1235 static void describe_flags(const char *optarg)
1236 {
1237         uint64_t flags = parse_flag_names(optarg, 0);
1238
1239         printf("0x%016llx\t%s\t%s\n",
1240                 (unsigned long long)flags,
1241                 page_flag_name(flags),
1242                 page_flag_longname(flags));
1243 }
1244
1245 static const struct option opts[] = {
1246         { "raw"       , 0, NULL, 'r' },
1247         { "pid"       , 1, NULL, 'p' },
1248         { "file"      , 1, NULL, 'f' },
1249         { "addr"      , 1, NULL, 'a' },
1250         { "bits"      , 1, NULL, 'b' },
1251         { "cgroup"    , 1, NULL, 'c' },
1252         { "describe"  , 1, NULL, 'd' },
1253         { "mark-idle" , 0, NULL, 'i' },
1254         { "list"      , 0, NULL, 'l' },
1255         { "list-each" , 0, NULL, 'L' },
1256         { "list-cgroup", 0, NULL, 'C' },
1257         { "list-mapcnt", 0, NULL, 'M' },
1258         { "no-summary", 0, NULL, 'N' },
1259         { "hwpoison"  , 0, NULL, 'X' },
1260         { "unpoison"  , 0, NULL, 'x' },
1261         { "kpageflags", 0, NULL, 'F' },
1262         { "help"      , 0, NULL, 'h' },
1263         { NULL        , 0, NULL, 0 }
1264 };
1265
1266 int main(int argc, char *argv[])
1267 {
1268         int c;
1269
1270         page_size = getpagesize();
1271
1272         while ((c = getopt_long(argc, argv,
1273                                 "rp:f:a:b:d:c:CilLMNXxF:h",
1274                                 opts, NULL)) != -1) {
1275                 switch (c) {
1276                 case 'r':
1277                         opt_raw = 1;
1278                         break;
1279                 case 'p':
1280                         parse_pid(optarg);
1281                         break;
1282                 case 'f':
1283                         parse_file(optarg);
1284                         break;
1285                 case 'a':
1286                         parse_addr_range(optarg);
1287                         break;
1288                 case 'b':
1289                         parse_bits_mask(optarg);
1290                         break;
1291                 case 'c':
1292                         parse_cgroup(optarg);
1293                         break;
1294                 case 'C':
1295                         opt_list_cgroup = 1;
1296                         break;
1297                 case 'd':
1298                         describe_flags(optarg);
1299                         exit(0);
1300                 case 'i':
1301                         opt_mark_idle = 1;
1302                         break;
1303                 case 'l':
1304                         opt_list = 1;
1305                         break;
1306                 case 'L':
1307                         opt_list = 2;
1308                         break;
1309                 case 'M':
1310                         opt_list_mapcnt = 1;
1311                         break;
1312                 case 'N':
1313                         opt_no_summary = 1;
1314                         break;
1315                 case 'X':
1316                         opt_hwpoison = 1;
1317                         prepare_hwpoison_fd();
1318                         break;
1319                 case 'x':
1320                         opt_unpoison = 1;
1321                         prepare_hwpoison_fd();
1322                         break;
1323                 case 'F':
1324                         parse_kpageflags(optarg);
1325                         break;
1326                 case 'h':
1327                         usage();
1328                         exit(0);
1329                 default:
1330                         usage();
1331                         exit(1);
1332                 }
1333         }
1334
1335         if (!opt_kpageflags)
1336                 opt_kpageflags = PROC_KPAGEFLAGS;
1337
1338         if (opt_cgroup || opt_list_cgroup)
1339                 kpagecgroup_fd = checked_open(PROC_KPAGECGROUP, O_RDONLY);
1340
1341         if (opt_list && opt_list_mapcnt)
1342                 kpagecount_fd = checked_open(PROC_KPAGECOUNT, O_RDONLY);
1343
1344         if (opt_mark_idle && opt_file)
1345                 page_idle_fd = checked_open(SYS_KERNEL_MM_PAGE_IDLE, O_RDWR);
1346
1347         if (opt_list && opt_pid)
1348                 printf("voffset\t");
1349         if (opt_list && opt_file)
1350                 printf("foffset\t");
1351         if (opt_list && opt_list_cgroup)
1352                 printf("cgroup\t");
1353         if (opt_list && opt_list_mapcnt)
1354                 printf("map-cnt\t");
1355
1356         if (opt_list == 1)
1357                 printf("offset\tlen\tflags\n");
1358         if (opt_list == 2)
1359                 printf("offset\tflags\n");
1360
1361         if (opt_file)
1362                 walk_page_cache();
1363         else
1364                 walk_addr_ranges();
1365
1366         if (opt_list == 1)
1367                 flush_page_range();
1368
1369         if (opt_no_summary)
1370                 return 0;
1371
1372         if (opt_list)
1373                 printf("\n\n");
1374
1375         show_summary();
1376
1377         if (opt_list_mapcnt)
1378                 close(kpagecount_fd);
1379
1380         if (page_idle_fd >= 0)
1381                 close(page_idle_fd);
1382
1383         return 0;
1384 }