lustre/utils/llsom_sync.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; if not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2017, DDN Storage Corporation.
  24  */
  25 /*
  26  * lustre/utils/llsom_sync.c
  27  *
  28  * Tool for sync the LSOM xattr.
  29  *
  30  * Author: Qian Yingjin <qian@ddn.com>
  31  */
  32
  33 #include <stdlib.h>
  34 #include <errno.h>
  35 #include <getopt.h>
  36 #include <unistd.h>
  37 #include <fcntl.h>
  38 #include <poll.h>
  39 #include <assert.h>
  40 #include <sys/stat.h>
  41 #include <sys/types.h>
  42 #include <time.h>
  43 #include <linux/unistd.h>
  44 #include <linux/kernel.h>
  45 #include <sys/sysinfo.h>
  46 #include <linux/lustre/lustre_user.h>
  47 #include <lustre/lustreapi.h>
  48 #include <linux/lustre/lustre_idl.h>
  49 #include <linux/lustre/lustre_fid.h>
  50 #include <libcfs/util/hash.h>
  51 #include <libcfs/util/list.h>
  52 #include <libcfs/util/parser.h>
  53
  54 #define container_of(ptr, type, member) ({                      \
  55         const typeof(((type *) 0)->member) * __mptr = (ptr);     \
  56         (type *) ((char *) __mptr - offsetof(type, member)); })
  57
  58 #define CHLG_POLL_INTV  60
  59 #define REC_MIN_AGE     600
  60 #define DEF_CACHE_SIZE  (256 * 1048576) /* 256MB */
  61
  62 struct options {
  63         const char      *o_chlg_user;
  64         const char      *o_mdtname;
  65         const char      *o_mntpt;
  66         bool             o_daemonize;
  67         bool             o_data_sync;
  68         int              o_verbose;
  69         int              o_intv;
  70         int              o_min_age;
  71         unsigned long    o_cached_fid_hiwm; /* high watermark */
  72         unsigned long    o_batch_sync_cnt;
  73 };
  74
  75 struct options opt;
  76
  77 struct fid_rec {
  78         struct hlist_node       fr_node;
  79         struct list_head        fr_link;
  80         lustre_fid              fr_fid;
  81         __u64                   fr_time;
  82         __u64                   fr_index;
  83 };
  84
  85 static const int fid_hash_shift = 6;
  86
  87 #define FID_HASH_ENTRIES        (1 << fid_hash_shift)
  88 #define FID_ON_HASH(f)          (!hlist_unhashed(&(f)->fr_node))
  89
  90 #if __BITS_PER_LONG == 32
  91 #define FID_HASH_FN(f)  (hash_long(fid_flatten32(f), fid_hash_shift))
  92 #elif __BITS_PER_LONG == 64
  93 #define FID_HASH_FN(f)  (hash_long(fid_flatten(f), fid_hash_shift))
  94 #else
  95 #error Wordsize not 32 or 64
  96 #endif
  97
  98 struct lsom_head {
  99         struct hlist_head       *lh_hash;
 100         struct list_head         lh_list; /* ordered list by record index */
 101         unsigned long            lh_cached_count;
 102 } head;
 103
 104 static void usage(char *prog)
 105 {
 106         printf("\nUsage: %s [options] -u <userid> -m <mdtdev> <mntpt>\n"
 107                "options:\n"
 108                "\t-d, --daemonize\n"
 109                "\t-i, --interval, poll interval in second\n"
 110                "\t-a, --min-age, min age before a record is processed.\n"
 111                "\t-c, --max-cache, percentage of the memroy used for cache.\n"
 112                "\t-s, --sync, data sync when update LSOM xattr\n"
 113                "\t-v, --verbose, produce more verbose ouput\n",
 114                prog);
 115         exit(0);
 116 }
 117
 118 static inline __u64 fid_flatten(const struct lu_fid *fid)
 119 {
 120         __u64 ino;
 121         __u64 seq;
 122
 123         if (fid_is_igif(fid)) {
 124                 ino = lu_igif_ino(fid);
 125                 return ino;
 126         }
 127
 128         seq = fid_seq(fid);
 129
 130         ino = (seq << 24) + ((seq >> 24) & 0xffffff0000ULL) + fid_oid(fid);
 131
 132         return ino ?: fid_oid(fid);
 133 }
 134
 135 /**
 136  * map fid to 32 bit value for ino on 32bit systems.
 137  */
 138 static inline __u32 fid_flatten32(const struct lu_fid *fid)
 139 {
 140         __u32 ino;
 141         __u64 seq;
 142
 143         if (fid_is_igif(fid)) {
 144                 ino = lu_igif_ino(fid);
 145                 return ino;
 146         }
 147
 148         seq = fid_seq(fid) - FID_SEQ_START;
 149
 150         /* Map the high bits of the OID into higher bits of the inode number so
 151          * that inodes generated at about the same time have a reduced chance
 152          * of collisions. This will give a period of 2^12 = 1024 unique clients
 153          * (from SEQ) and up to min(LUSTRE_SEQ_MAX_WIDTH, 2^20) = 128k objects
 154          * (from OID), or up to 128M inodes without collisions for new files.
 155          */
 156         ino = ((seq & 0x000fffffULL) << 12) + ((seq >> 8) & 0xfffff000) +
 157               (seq >> (64 - (40-8)) & 0xffffff00) +
 158               (fid_oid(fid) & 0xff000fff) + ((fid_oid(fid) & 0x00fff000) << 8);
 159
 160         return ino ?: fid_oid(fid);
 161 }
 162
 163 static inline bool fid_eq(const lustre_fid *f1, const lustre_fid *f2)
 164 {
 165         return f1->f_seq == f2->f_seq && f1->f_oid == f2->f_oid &&
 166                f1->f_ver == f2->f_ver;
 167 }
 168
 169 static void fid_hash_del(struct fid_rec *f)
 170 {
 171         if (FID_ON_HASH(f))
 172                 hlist_del_init(&f->fr_node);
 173 }
 174
 175 static void fid_hash_add(struct fid_rec *f)
 176 {
 177         assert(!FID_ON_HASH(f));
 178         hlist_add_head(&f->fr_node, &head.lh_hash[FID_HASH_FN(&f->fr_fid)]);
 179 }
 180
 181 static struct fid_rec *fid_hash_find(const lustre_fid *fid)
 182 {
 183         struct hlist_head *hash_list;
 184         struct hlist_node *entry, *next;
 185         struct fid_rec *f;
 186
 187         hash_list = &head.lh_hash[FID_HASH_FN(fid)];
 188         hlist_for_each_entry_safe(f, entry, next, hash_list, fr_node) {
 189                 assert(FID_ON_HASH(f));
 190                 if (fid_eq(fid, &f->fr_fid))
 191                         return f;
 192         }
 193
 194         return NULL;
 195 }
 196
 197 static int lsom_setup(void)
 198 {
 199         int i;
 200
 201         /* set llapi message level */
 202         llapi_msg_set_level(opt.o_verbose);
 203
 204         memset(&head, 0, sizeof(head));
 205         head.lh_hash = malloc(sizeof(struct hlist_head) * FID_HASH_ENTRIES);
 206         if (head.lh_hash == NULL) {
 207                 llapi_err_noerrno(LLAPI_MSG_ERROR,
 208                                  "failed to alloc memory for hash (%zu).",
 209                                  sizeof(struct hlist_head) * FID_HASH_ENTRIES);
 210                 return -ENOMEM;
 211         }
 212
 213         for (i = 0; i < FID_HASH_ENTRIES; i++)
 214                 INIT_HLIST_HEAD(&head.lh_hash[i]);
 215
 216         INIT_LIST_HEAD(&head.lh_list);
 217         return 0;
 218 }
 219
 220 static void lsom_cleanup(void)
 221 {
 222         free(head.lh_hash);
 223 }
 224
 225 static int lsom_update_one(struct fid_rec *f)
 226 {
 227         struct stat st;
 228         int fd;
 229         int rc = 0;
 230
 231         fd = llapi_open_by_fid(opt.o_mntpt, &f->fr_fid,
 232                                O_RDONLY | O_NOATIME);
 233         if (fd < 0) {
 234                 rc = -errno;
 235
 236                 /* The file may be deleted, clean the corresponding
 237                  * changelog record and ignore this error.
 238                  */
 239                 if (rc == -ENOENT)
 240                         goto clean_up;
 241
 242                 llapi_error(LLAPI_MSG_ERROR, rc,
 243                             "llapi_open_by_fid for " DFID " failed",
 244                             PFID(&f->fr_fid));
 245                 return rc;
 246         }
 247
 248         if (opt.o_data_sync) {
 249                 __u64 dv;
 250
 251                 /* Flush dirty pages from clients */
 252                 rc = llapi_get_data_version(fd, &dv, LL_DV_RD_FLUSH);
 253                 if (rc < 0)
 254                         llapi_error(LLAPI_MSG_ERROR, errno,
 255                                     "failed to sync data for " DFID,
 256                                     PFID(&f->fr_fid));
 257                 /* ignore this error, continue to sync lsom data */
 258         }
 259
 260         rc = fstat(fd, &st);
 261         if (rc < 0) {
 262                 llapi_error(LLAPI_MSG_ERROR, rc, "failed to stat FID: " DFID,
 263                             PFID(&f->fr_fid));
 264                 return rc;
 265         }
 266
 267         /* After call fstat(), it already gets OST attrs to the client,
 268          * when close the file, MDS will update the LSOM data itself
 269          * according the size and blocks information from the client.
 270          */
 271         close(fd);
 272
 273         llapi_printf(LLAPI_MSG_DEBUG,
 274                      "record %llu:%llu, updated LSOM for fid " DFID
 275                      " size:%lu blocks:%lu\n", f->fr_time, f->fr_index,
 276                      PFID(&f->fr_fid), st.st_size, st.st_blocks);
 277
 278 clean_up:
 279         rc = llapi_changelog_clear(opt.o_mdtname,
 280                                    opt.o_chlg_user, f->fr_index);
 281         if (rc)
 282                 llapi_error(LLAPI_MSG_ERROR, rc,
 283                             "failed to clear changelog record: %s:%llu",
 284                             opt.o_chlg_user, f->fr_index);
 285         return rc;
 286 }
 287
 288 static int lsom_start_update(int count)
 289 {
 290         int rc = 0;
 291         int i = 0;
 292
 293         llapi_printf(LLAPI_MSG_INFO, "Start to sync %d records.\n", count);
 294
 295         while (i < count) {
 296                 struct fid_rec *f;
 297
 298                 f = list_entry(head.lh_list.next, struct fid_rec, fr_link);
 299                 rc = lsom_update_one(f);
 300                 if (rc == 0) {
 301                         list_del_init(&f->fr_link);
 302                         fid_hash_del(f);
 303                         free(f);
 304                         head.lh_cached_count--;
 305                         i++;
 306                 } else {
 307                         goto out;
 308                 }
 309         }
 310
 311 out:
 312         return rc;
 313 }
 314
 315 static int lsom_check_sync(void)
 316 {
 317         int rc = 0;
 318         int count;
 319
 320 repeated:
 321         count = 0;
 322         if (list_empty(&head.lh_list))
 323                 return 0;
 324
 325         if (head.lh_cached_count > opt.o_cached_fid_hiwm)
 326                 count = opt.o_batch_sync_cnt;
 327         else {
 328                 struct fid_rec *f;
 329                 time_t now;
 330
 331                 /* When the first record in the list was not being
 332                  * processed for a long time (more than o_min_age),
 333                  * pop the record, start to handle it immediately.
 334                  */
 335                 now = time(NULL);
 336                 f = list_entry(head.lh_list.next, struct fid_rec, fr_link);
 337                 if (now > ((f->fr_time >> 30) + opt.o_min_age))
 338                         count = 1;
 339         }
 340
 341         if (count > 0)
 342                 rc = lsom_start_update(count);
 343
 344         if (rc == 0 && count == 1)
 345                 goto repeated;
 346
 347         return rc;
 348 }
 349
 350 static void lsom_sort_record_list(struct fid_rec *f)
 351 {
 352         struct list_head *pos;
 353         bool need_move = false;
 354
 355         for (pos = f->fr_link.next; pos != &head.lh_list; pos = pos->next) {
 356                 struct fid_rec *rec = list_entry(pos, struct fid_rec, fr_link);
 357
 358                 if (f->fr_index > rec->fr_index) {
 359                         need_move = true;
 360                         continue;
 361                 } else {
 362                         break;
 363                 }
 364         }
 365
 366         if (need_move)
 367                 list_move_tail(&f->fr_link, pos);
 368 }
 369
 370 static int process_record(struct changelog_rec *rec)
 371 {
 372         __u64 index = rec->cr_index;
 373         int rc = 0;
 374
 375         if (rec->cr_type == CL_CLOSE || rec->cr_type == CL_TRUNC ||
 376             rec->cr_type == CL_SETATTR) {
 377                 struct fid_rec *f;
 378
 379                 f = fid_hash_find(&rec->cr_tfid);
 380                 if (f == NULL) {
 381                         f = malloc(sizeof(struct fid_rec));
 382                         if (f == NULL) {
 383                                 rc = -ENOMEM;
 384                                 llapi_error(LLAPI_MSG_ERROR, rc,
 385                                             "failed to alloc memory for fid_rec");
 386                                 return rc;
 387                         }
 388
 389                         f->fr_fid = rec->cr_tfid;
 390                         f->fr_index = index;
 391                         f->fr_time = rec->cr_time;
 392                         INIT_HLIST_NODE(&f->fr_node);
 393                         fid_hash_add(f);
 394                         /*
 395                          * The newly changelog record index is processed in the
 396                          * ascending order, so it is safe to put the record at
 397                          * the tail of the ordered list.
 398                          */
 399                         list_add_tail(&f->fr_link, &head.lh_list);
 400                         head.lh_cached_count++;
 401                 } else {
 402                         f->fr_index = index;
 403                         lsom_sort_record_list(f);
 404                 }
 405         }
 406
 407         llapi_printf(LLAPI_MSG_DEBUG, "Processed changelog record index:%llu "
 408                      "type:%s(0x%x) FID:"DFID"\n", index,
 409                      changelog_type2str(__le32_to_cpu(rec->cr_type)),
 410                      __le32_to_cpu(rec->cr_type), PFID(&rec->cr_tfid));
 411
 412         return rc;
 413 }
 414
 415 static unsigned long get_fid_cache_size(int pct)
 416 {
 417         struct sysinfo sinfo;
 418         unsigned long cache_size;
 419         int rc;
 420
 421         rc = sysinfo(&sinfo);
 422         if (rc) {
 423                 llapi_error(LLAPI_MSG_ERROR, rc, "failed to get sysinfo");
 424                 /* ignore this error, just pick some reasonable static
 425                  * limit for the cache size (e.g. 256MB, default value).
 426                  */
 427                 cache_size = DEF_CACHE_SIZE;
 428         } else {
 429                 /* maximum cached fid size is tunned according to total
 430                  * memory size, e.g. 5% of the memroy.
 431                  */
 432                 cache_size = sinfo.totalram * pct / 100;
 433         }
 434
 435         return cache_size;
 436 }
 437
 438 int main(int argc, char **argv)
 439 {
 440         int                      c;
 441         int                      rc;
 442         void                    *chglog_hdlr;
 443         struct changelog_rec    *rec;
 444         bool                     stop = 0;
 445         int                      ret = 0;
 446         unsigned long            cache_size = DEF_CACHE_SIZE;
 447         char                     fsname[MAX_OBD_NAME + 1];
 448         static struct option options[] = {
 449                 { "mdt", required_argument, NULL, 'm' },
 450                 { "user", required_argument, 0, 'u'},
 451                 { "daemonize", no_argument, NULL, 'd'},
 452                 { "interval", required_argument, NULL, 'i'},
 453                 { "min-age", required_argument, NULL, 'a'},
 454                 { "max-cache", required_argument, NULL, 'c'},
 455                 { "verbose", no_argument, NULL, 'v'},
 456                 { "sync", no_argument, NULL, 's'},
 457                 { "help", no_argument, NULL, 'h' },
 458                 { NULL }
 459         };
 460
 461         memset(&opt, 0, sizeof(opt));
 462         opt.o_data_sync = false;
 463         opt.o_verbose = LLAPI_MSG_INFO;
 464         opt.o_intv = CHLG_POLL_INTV;
 465         opt.o_min_age = REC_MIN_AGE;
 466
 467         while ((c = getopt_long(argc, argv, "u:hm:dsi:a:c:v", options, NULL))
 468                != EOF) {
 469                 switch (c) {
 470                 default:
 471                         rc = -EINVAL;
 472                         llapi_error(LLAPI_MSG_ERROR, rc,
 473                                     "%s: unknown option '-%c'\n",
 474                                     argv[0], optopt);
 475                         return rc;
 476                 case 'u':
 477                         opt.o_chlg_user = optarg;
 478                         break;
 479                 case 'h':
 480                         usage(argv[0]);
 481                         break;
 482                 case 'm':
 483                         opt.o_mdtname = optarg;
 484                         break;
 485                 case 'd':
 486                         opt.o_daemonize = true;
 487                         break;
 488                 case 'i':
 489                         opt.o_intv = atoi(optarg);
 490                         if (opt.o_intv < 0) {
 491                                 rc = -EINVAL;
 492                                 llapi_error(LLAPI_MSG_ERROR, rc,
 493                                             "bad value for -i %s", optarg);
 494                                 return rc;
 495                         }
 496                         break;
 497                 case 'a':
 498                         opt.o_min_age = atoi(optarg);
 499                         if (opt.o_min_age < 0) {
 500                                 rc = -EINVAL;
 501                                 llapi_error(LLAPI_MSG_ERROR, rc,
 502                                             "bad value for -a %s", optarg);
 503                                 return rc;
 504                         }
 505                         break;
 506                 case 'c':
 507                         rc = Parser_size(&cache_size, optarg);
 508                         if (rc < 0) {
 509                                 rc = -EINVAL;
 510                                 llapi_error(LLAPI_MSG_ERROR, rc,
 511                                             "bad valud for -c '%s'", optarg);
 512                                 return rc;
 513                         }
 514
 515                         /* For value < 100, it is taken as the percentage of
 516                          * total memory instead.
 517                          */
 518                         if (cache_size < 100)
 519                                 cache_size = get_fid_cache_size(cache_size);
 520                         llapi_printf(LLAPI_MSG_INFO, "Cache size: %lu\n",
 521                                      cache_size);
 522                         break;
 523                 case 'v':
 524                         opt.o_verbose++;
 525                         break;
 526                 case 's':
 527                         opt.o_data_sync = true;
 528                         break;
 529                 }
 530         }
 531
 532         if (argc != optind + 1) {
 533                 llapi_err_noerrno(LLAPI_MSG_ERROR,
 534                                   "%s: no mount point specified\n", argv[0]);
 535                 usage(argv[0]);
 536         }
 537
 538         opt.o_mntpt = argv[optind];
 539         rc = llapi_search_fsname(opt.o_mntpt, fsname);
 540         if (rc < 0) {
 541                 llapi_error(LLAPI_MSG_ERROR, rc,
 542                             "cannot find a Lustre file system mounted at '%s'",
 543                             opt.o_mntpt);
 544                 return rc;
 545         }
 546
 547         if (!opt.o_mdtname)
 548                 usage(argv[0]);
 549
 550         if (!opt.o_chlg_user)
 551                 usage(argv[0]);
 552
 553         if (opt.o_daemonize) {
 554                 rc = daemon(1, 1);
 555                 if (rc < 0) {
 556                         rc = -errno;
 557                         llapi_error(LLAPI_MSG_ERROR, rc, "cannot daemonize");
 558                         return rc;
 559                 }
 560
 561                 setbuf(stdout, NULL);
 562         }
 563
 564         opt.o_cached_fid_hiwm = cache_size / sizeof(struct fid_rec);
 565         opt.o_batch_sync_cnt = opt.o_cached_fid_hiwm / 2;
 566
 567         rc = lsom_setup();
 568         if (rc < 0)
 569                 return rc;
 570
 571         while (!stop) {
 572                 bool eof = false;
 573
 574                 llapi_printf(LLAPI_MSG_DEBUG, "Start receiving records\n");
 575                 rc = llapi_changelog_start(&chglog_hdlr,
 576                                            CHANGELOG_FLAG_BLOCK |
 577                                            CHANGELOG_FLAG_JOBID |
 578                                            CHANGELOG_FLAG_EXTRA_FLAGS,
 579                                            opt.o_mdtname, 0);
 580                 if (rc) {
 581                         llapi_error(LLAPI_MSG_ERROR, rc,
 582                                     "unable to open changelog of MDT [%s]\n",
 583                                     opt.o_mdtname);
 584                         return rc;
 585                 }
 586
 587                 while (!eof && !stop) {
 588                         rc = llapi_changelog_recv(chglog_hdlr, &rec);
 589                         switch (rc) {
 590                         case 0:
 591                                 rc = process_record(rec);
 592                                 if (rc) {
 593                                         llapi_error(LLAPI_MSG_ERROR, rc,
 594                                                     "failed to process record");
 595                                         ret = rc;
 596                                 }
 597
 598                                 llapi_changelog_free(&rec);
 599
 600                                 rc = lsom_check_sync();
 601                                 if (rc) {
 602                                         stop = true;
 603                                         ret = rc;
 604                                 }
 605
 606                                 break;
 607                         case 1: /* EOF */
 608                                 llapi_printf(LLAPI_MSG_DEBUG,
 609                                              "finished reading [%s]\n",
 610                                              opt.o_mdtname);
 611                                 eof = true;
 612                                 break;
 613                         case -EINVAL: /* FS unmounted */
 614                         case -EPROTO:  /* error in KUC channel */
 615                         default:
 616                                 stop = true;
 617                                 llapi_error(LLAPI_MSG_ERROR, rc,
 618                                             "failed to get changelog record");
 619                                 ret = rc;
 620                                 break;
 621                         }
 622                 }
 623
 624                 /* reach EOF of changelog */
 625                 rc = llapi_changelog_fini(&chglog_hdlr);
 626                 if (rc) {
 627                         llapi_error(LLAPI_MSG_ERROR, rc,
 628                                     "unable to close changelog of MDT [%s]",
 629                                     opt.o_mdtname);
 630                         ret = rc;
 631                         return rc;
 632                 }
 633
 634                 if (opt.o_daemonize) {
 635                         sleep(opt.o_intv);
 636
 637                         rc = lsom_check_sync();
 638                         if (rc) {
 639                                 stop = true;
 640                                 ret = rc;
 641                         }
 642                 } else {
 643                         lsom_start_update(head.lh_cached_count);
 644                         stop = true;
 645                 }
 646         }
 647
 648         lsom_cleanup();
 649         return ret;
 650 }