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 #ifndef PATH_MAX
  55 #define PATH_MAX (4096)
  56 #endif
  57
  58 #define container_of(ptr, type, member) ({                      \
  59         const typeof(((type *) 0)->member) * __mptr = (ptr);     \
  60         (type *) ((char *) __mptr - offsetof(type, member)); })
  61
  62 #define CHLG_POLL_INTV  60
  63 #define REC_MIN_AGE     600
  64 #define DEF_CACHE_SIZE  (256 * 1048576) /* 256MB */
  65
  66 struct options {
  67         const char      *o_chlg_user;
  68         const char      *o_mdtname;
  69         const char      *o_mntpt;
  70         bool             o_daemonize;
  71         bool             o_data_sync;
  72         int              o_verbose;
  73         int              o_intv;
  74         int              o_min_age;
  75         unsigned long    o_cached_fid_hiwm; /* high watermark */
  76         unsigned long    o_batch_sync_cnt;
  77 };
  78
  79 struct options opt;
  80
  81 struct fid_rec {
  82         struct hlist_node       fr_node;
  83         struct list_head        fr_link;
  84         lustre_fid              fr_fid;
  85         __u64                   fr_time;
  86         __u64                   fr_index;
  87 };
  88
  89 static const int fid_hash_shift = 6;
  90
  91 #define FID_HASH_ENTRIES        (1 << fid_hash_shift)
  92 #define FID_ON_HASH(f)          (!hlist_unhashed(&(f)->fr_node))
  93
  94 #if __BITS_PER_LONG == 32
  95 #define FID_HASH_FN(f)  (hash_long(fid_flatten32(f), fid_hash_shift))
  96 #elif __BITS_PER_LONG == 64
  97 #define FID_HASH_FN(f)  (hash_long(fid_flatten(f), fid_hash_shift))
  98 #else
  99 #error Wordsize not 32 or 64
 100 #endif
 101
 102 struct lsom_head {
 103         struct hlist_head       *lh_hash;
 104         struct list_head         lh_list; /* ordered list by record index */
 105         unsigned long            lh_cached_count;
 106 } head;
 107
 108 static void usage(char *prog)
 109 {
 110         printf("\nUsage: %s [options] -u <userid> -m <mdtdev> <mntpt>\n"
 111                "options:\n"
 112                "\t-d, --daemonize\n"
 113                "\t-i, --interval, poll interval in second\n"
 114                "\t-a, --min-age, min age before a record is processed.\n"
 115                "\t-c, --max-cache, percentage of the memroy used for cache.\n"
 116                "\t-s, --sync, data sync when update LSOM xattr\n"
 117                "\t-v, --verbose, produce more verbose ouput\n",
 118                prog);
 119         exit(0);
 120 }
 121
 122 static inline __u64 fid_flatten(const struct lu_fid *fid)
 123 {
 124         __u64 ino;
 125         __u64 seq;
 126
 127         if (fid_is_igif(fid)) {
 128                 ino = lu_igif_ino(fid);
 129                 return ino;
 130         }
 131
 132         seq = fid_seq(fid);
 133
 134         ino = (seq << 24) + ((seq >> 24) & 0xffffff0000ULL) + fid_oid(fid);
 135
 136         return ino ?: fid_oid(fid);
 137 }
 138
 139 /**
 140  * map fid to 32 bit value for ino on 32bit systems.
 141  */
 142 static inline __u32 fid_flatten32(const struct lu_fid *fid)
 143 {
 144         __u32 ino;
 145         __u64 seq;
 146
 147         if (fid_is_igif(fid)) {
 148                 ino = lu_igif_ino(fid);
 149                 return ino;
 150         }
 151
 152         seq = fid_seq(fid) - FID_SEQ_START;
 153
 154         /* Map the high bits of the OID into higher bits of the inode number so
 155          * that inodes generated at about the same time have a reduced chance
 156          * of collisions. This will give a period of 2^12 = 1024 unique clients
 157          * (from SEQ) and up to min(LUSTRE_SEQ_MAX_WIDTH, 2^20) = 128k objects
 158          * (from OID), or up to 128M inodes without collisions for new files.
 159          */
 160         ino = ((seq & 0x000fffffULL) << 12) + ((seq >> 8) & 0xfffff000) +
 161               (seq >> (64 - (40-8)) & 0xffffff00) +
 162               (fid_oid(fid) & 0xff000fff) + ((fid_oid(fid) & 0x00fff000) << 8);
 163
 164         return ino ?: fid_oid(fid);
 165 }
 166
 167 static inline bool fid_eq(const lustre_fid *f1, const lustre_fid *f2)
 168 {
 169         return f1->f_seq == f2->f_seq && f1->f_oid == f2->f_oid &&
 170                f1->f_ver == f2->f_ver;
 171 }
 172
 173 static void fid_hash_del(struct fid_rec *f)
 174 {
 175         if (FID_ON_HASH(f))
 176                 hlist_del_init(&f->fr_node);
 177 }
 178
 179 static void fid_hash_add(struct fid_rec *f)
 180 {
 181         assert(!FID_ON_HASH(f));
 182         hlist_add_head(&f->fr_node, &head.lh_hash[FID_HASH_FN(&f->fr_fid)]);
 183 }
 184
 185 static struct fid_rec *fid_hash_find(const lustre_fid *fid)
 186 {
 187         struct hlist_head *hash_list;
 188         struct hlist_node *entry, *next;
 189         struct fid_rec *f;
 190
 191         hash_list = &head.lh_hash[FID_HASH_FN(fid)];
 192         hlist_for_each_entry_safe(f, entry, next, hash_list, fr_node) {
 193                 assert(FID_ON_HASH(f));
 194                 if (fid_eq(fid, &f->fr_fid))
 195                         return f;
 196         }
 197
 198         return NULL;
 199 }
 200
 201 static int lsom_setup(void)
 202 {
 203         int i;
 204
 205         /* set llapi message level */
 206         llapi_msg_set_level(opt.o_verbose);
 207
 208         memset(&head, 0, sizeof(head));
 209         head.lh_hash = malloc(sizeof(struct hlist_head) * FID_HASH_ENTRIES);
 210         if (head.lh_hash == NULL) {
 211                 llapi_err_noerrno(LLAPI_MSG_ERROR,
 212                                  "failed to alloc memory for hash (%zu).",
 213                                  sizeof(struct hlist_head) * FID_HASH_ENTRIES);
 214                 return -ENOMEM;
 215         }
 216
 217         for (i = 0; i < FID_HASH_ENTRIES; i++)
 218                 INIT_HLIST_HEAD(&head.lh_hash[i]);
 219
 220         INIT_LIST_HEAD(&head.lh_list);
 221         return 0;
 222 }
 223
 224 static void lsom_cleanup(void)
 225 {
 226         free(head.lh_hash);
 227 }
 228
 229 static int lsom_update_one(struct fid_rec *f)
 230 {
 231         struct stat st;
 232         int fd;
 233         int rc = 0;
 234
 235         fd = llapi_open_by_fid(opt.o_mntpt, &f->fr_fid,
 236                                O_RDONLY | O_NOATIME);
 237         if (fd < 0) {
 238                 rc = -errno;
 239
 240                 /* The file may be deleted, clean the corresponding
 241                  * changelog record and ignore this error.
 242                  */
 243                 if (rc == -ENOENT)
 244                         goto clean_up;
 245
 246                 llapi_error(LLAPI_MSG_ERROR, rc,
 247                             "llapi_open_by_fid for " DFID " failed",
 248                             PFID(&f->fr_fid));
 249                 return rc;
 250         }
 251
 252         if (opt.o_data_sync) {
 253                 __u64 dv;
 254
 255                 /* Flush dirty pages from clients */
 256                 rc = llapi_get_data_version(fd, &dv, LL_DV_RD_FLUSH);
 257                 if (rc < 0)
 258                         llapi_error(LLAPI_MSG_ERROR, errno,
 259                                     "failed to sync data for " DFID,
 260                                     PFID(&f->fr_fid));
 261                 /* ignore this error, continue to sync lsom data */
 262         }
 263
 264         rc = fstat(fd, &st);
 265         if (rc < 0) {
 266                 llapi_error(LLAPI_MSG_ERROR, rc, "failed to stat FID: " DFID,
 267                             PFID(&f->fr_fid));
 268                 return rc;
 269         }
 270
 271         /* After call fstat(), it already gets OST attrs to the client,
 272          * when close the file, MDS will update the LSOM data itself
 273          * according the size and blocks information from the client.
 274          */
 275         close(fd);
 276
 277         llapi_printf(LLAPI_MSG_DEBUG,
 278                      "record %llu:%llu, updated LSOM for fid " DFID
 279                      " size:%lu blocks:%lu\n", f->fr_time, f->fr_index,
 280                      PFID(&f->fr_fid), st.st_size, st.st_blocks);
 281
 282 clean_up:
 283         rc = llapi_changelog_clear(opt.o_mdtname,
 284                                    opt.o_chlg_user, f->fr_index);
 285         if (rc)
 286                 llapi_error(LLAPI_MSG_ERROR, rc,
 287                             "failed to clear changelog record: %s:%llu",
 288                             opt.o_chlg_user, f->fr_index);
 289         return rc;
 290 }
 291
 292 static int lsom_start_update(int count)
 293 {
 294         int rc = 0;
 295         int i = 0;
 296
 297         llapi_printf(LLAPI_MSG_INFO, "Start to sync %d records.\n", count);
 298
 299         while (i < count) {
 300                 struct fid_rec *f;
 301
 302                 f = list_entry(head.lh_list.next, struct fid_rec, fr_link);
 303                 rc = lsom_update_one(f);
 304                 if (rc == 0) {
 305                         list_del_init(&f->fr_link);
 306                         fid_hash_del(f);
 307                         free(f);
 308                         head.lh_cached_count--;
 309                         i++;
 310                 } else {
 311                         goto out;
 312                 }
 313         }
 314
 315 out:
 316         return rc;
 317 }
 318
 319 static int lsom_check_sync(void)
 320 {
 321         int rc = 0;
 322         int count;
 323
 324 repeated:
 325         count = 0;
 326         if (list_empty(&head.lh_list))
 327                 return 0;
 328
 329         if (head.lh_cached_count > opt.o_cached_fid_hiwm)
 330                 count = opt.o_batch_sync_cnt;
 331         else {
 332                 struct fid_rec *f;
 333                 time_t now;
 334
 335                 /* When the first record in the list was not being
 336                  * processed for a long time (more than o_min_age),
 337                  * pop the record, start to handle it immediately.
 338                  */
 339                 now = time(NULL);
 340                 f = list_entry(head.lh_list.next, struct fid_rec, fr_link);
 341                 if (now > ((f->fr_time >> 30) + opt.o_min_age))
 342                         count = 1;
 343         }
 344
 345         if (count > 0)
 346                 rc = lsom_start_update(count);
 347
 348         if (rc == 0 && count == 1)
 349                 goto repeated;
 350
 351         return rc;
 352 }
 353
 354 static void lsom_sort_record_list(struct fid_rec *f)
 355 {
 356         struct list_head *pos;
 357         bool need_move = false;
 358
 359         for (pos = f->fr_link.next; pos != &head.lh_list; pos = pos->next) {
 360                 struct fid_rec *rec = list_entry(pos, struct fid_rec, fr_link);
 361
 362                 if (f->fr_index > rec->fr_index) {
 363                         need_move = true;
 364                         continue;
 365                 } else {
 366                         break;
 367                 }
 368         }
 369
 370         if (need_move)
 371                 list_move_tail(&f->fr_link, pos);
 372 }
 373
 374 static int process_record(struct changelog_rec *rec)
 375 {
 376         __u64 index = rec->cr_index;
 377         int rc = 0;
 378
 379         if (rec->cr_type == CL_CLOSE || rec->cr_type == CL_TRUNC ||
 380             rec->cr_type == CL_SETATTR) {
 381                 struct fid_rec *f;
 382
 383                 f = fid_hash_find(&rec->cr_tfid);
 384                 if (f == NULL) {
 385                         f = malloc(sizeof(struct fid_rec));
 386                         if (f == NULL) {
 387                                 rc = -ENOMEM;
 388                                 llapi_error(LLAPI_MSG_ERROR, rc,
 389                                             "failed to alloc memory for fid_rec");
 390                                 return rc;
 391                         }
 392
 393                         f->fr_fid = rec->cr_tfid;
 394                         f->fr_index = index;
 395                         f->fr_time = rec->cr_time;
 396                         INIT_HLIST_NODE(&f->fr_node);
 397                         fid_hash_add(f);
 398                         /*
 399                          * The newly changelog record index is processed in the
 400                          * ascending order, so it is safe to put the record at
 401                          * the tail of the ordered list.
 402                          */
 403                         list_add_tail(&f->fr_link, &head.lh_list);
 404                         head.lh_cached_count++;
 405                 } else {
 406                         f->fr_index = index;
 407                         lsom_sort_record_list(f);
 408                 }
 409         }
 410
 411         llapi_printf(LLAPI_MSG_DEBUG, "Processed changelog record index:%llu "
 412                      "type:%s(0x%x) FID:"DFID"\n", index,
 413                      changelog_type2str(__le32_to_cpu(rec->cr_type)),
 414                      __le32_to_cpu(rec->cr_type), PFID(&rec->cr_tfid));
 415
 416         return rc;
 417 }
 418
 419 static unsigned long get_fid_cache_size(int pct)
 420 {
 421         struct sysinfo sinfo;
 422         unsigned long cache_size;
 423         int rc;
 424
 425         rc = sysinfo(&sinfo);
 426         if (rc) {
 427                 llapi_error(LLAPI_MSG_ERROR, rc, "failed to get sysinfo");
 428                 /* ignore this error, just pick some reasonable static
 429                  * limit for the cache size (e.g. 256MB, default value).
 430                  */
 431                 cache_size = DEF_CACHE_SIZE;
 432         } else {
 433                 /* maximum cached fid size is tunned according to total
 434                  * memory size, e.g. 5% of the memroy.
 435                  */
 436                 cache_size = sinfo.totalram * pct / 100;
 437         }
 438
 439         return cache_size;
 440 }
 441
 442 int main(int argc, char **argv)
 443 {
 444         int                      c;
 445         int                      rc;
 446         void                    *chglog_hdlr;
 447         struct changelog_rec    *rec;
 448         bool                     stop = 0;
 449         int                      ret = 0;
 450         unsigned long            cache_size = DEF_CACHE_SIZE;
 451         char                     fsname[MAX_OBD_NAME + 1];
 452         static struct option options[] = {
 453                 { "mdt", required_argument, NULL, 'm' },
 454                 { "user", required_argument, 0, 'u'},
 455                 { "daemonize", no_argument, NULL, 'd'},
 456                 { "interval", required_argument, NULL, 'i'},
 457                 { "min-age", required_argument, NULL, 'a'},
 458                 { "max-cache", required_argument, NULL, 'c'},
 459                 { "verbose", no_argument, NULL, 'v'},
 460                 { "sync", no_argument, NULL, 's'},
 461                 { "help", no_argument, NULL, 'h' },
 462                 { NULL }
 463         };
 464
 465         memset(&opt, 0, sizeof(opt));
 466         opt.o_data_sync = false;
 467         opt.o_verbose = LLAPI_MSG_INFO;
 468         opt.o_intv = CHLG_POLL_INTV;
 469         opt.o_min_age = REC_MIN_AGE;
 470
 471         while ((c = getopt_long(argc, argv, "u:hm:dsi:a:c:v", options, NULL))
 472                != EOF) {
 473                 switch (c) {
 474                 default:
 475                         rc = -EINVAL;
 476                         llapi_error(LLAPI_MSG_ERROR, rc,
 477                                     "%s: unknown option '-%c'\n",
 478                                     argv[0], optopt);
 479                         return rc;
 480                 case 'u':
 481                         opt.o_chlg_user = optarg;
 482                         break;
 483                 case 'h':
 484                         usage(argv[0]);
 485                         break;
 486                 case 'm':
 487                         opt.o_mdtname = optarg;
 488                         break;
 489                 case 'd':
 490                         opt.o_daemonize = true;
 491                         break;
 492                 case 'i':
 493                         opt.o_intv = atoi(optarg);
 494                         if (opt.o_intv < 0) {
 495                                 rc = -EINVAL;
 496                                 llapi_error(LLAPI_MSG_ERROR, rc,
 497                                             "bad value for -i %s", optarg);
 498                                 return rc;
 499                         }
 500                         break;
 501                 case 'a':
 502                         opt.o_min_age = atoi(optarg);
 503                         if (opt.o_min_age < 0) {
 504                                 rc = -EINVAL;
 505                                 llapi_error(LLAPI_MSG_ERROR, rc,
 506                                             "bad value for -a %s", optarg);
 507                                 return rc;
 508                         }
 509                         break;
 510                 case 'c':
 511                         rc = Parser_size(&cache_size, optarg);
 512                         if (rc < 0) {
 513                                 rc = -EINVAL;
 514                                 llapi_error(LLAPI_MSG_ERROR, rc,
 515                                             "bad valud for -c '%s'", optarg);
 516                                 return rc;
 517                         }
 518
 519                         /* For value < 100, it is taken as the percentage of
 520                          * total memory instead.
 521                          */
 522                         if (cache_size < 100)
 523                                 cache_size = get_fid_cache_size(cache_size);
 524                         llapi_printf(LLAPI_MSG_INFO, "Cache size: %lu\n",
 525                                      cache_size);
 526                         break;
 527                 case 'v':
 528                         opt.o_verbose++;
 529                         break;
 530                 case 's':
 531                         opt.o_data_sync = true;
 532                         break;
 533                 }
 534         }
 535
 536         if (argc != optind + 1) {
 537                 llapi_err_noerrno(LLAPI_MSG_ERROR,
 538                                   "%s: no mount point specified\n", argv[0]);
 539                 usage(argv[0]);
 540         }
 541
 542         opt.o_mntpt = argv[optind];
 543         rc = llapi_search_fsname(opt.o_mntpt, fsname);
 544         if (rc < 0) {
 545                 llapi_error(LLAPI_MSG_ERROR, rc,
 546                             "cannot find a Lustre file system mounted at '%s'",
 547                             opt.o_mntpt);
 548                 return rc;
 549         }
 550
 551         if (!opt.o_mdtname)
 552                 usage(argv[0]);
 553
 554         if (!opt.o_chlg_user)
 555                 usage(argv[0]);
 556
 557         if (opt.o_daemonize) {
 558                 rc = daemon(1, 1);
 559                 if (rc < 0) {
 560                         rc = -errno;
 561                         llapi_error(LLAPI_MSG_ERROR, rc, "cannot daemonize");
 562                         return rc;
 563                 }
 564
 565                 setbuf(stdout, NULL);
 566         }
 567
 568         opt.o_cached_fid_hiwm = cache_size / sizeof(struct fid_rec);
 569         opt.o_batch_sync_cnt = opt.o_cached_fid_hiwm / 2;
 570
 571         rc = lsom_setup();
 572         if (rc < 0)
 573                 return rc;
 574
 575         while (!stop) {
 576                 bool eof = false;
 577
 578                 llapi_printf(LLAPI_MSG_DEBUG, "Start receiving records\n");
 579                 rc = llapi_changelog_start(&chglog_hdlr,
 580                                            CHANGELOG_FLAG_BLOCK |
 581                                            CHANGELOG_FLAG_JOBID |
 582                                            CHANGELOG_FLAG_EXTRA_FLAGS,
 583                                            opt.o_mdtname, 0);
 584                 if (rc) {
 585                         llapi_error(LLAPI_MSG_ERROR, rc,
 586                                     "unable to open changelog of MDT [%s]\n",
 587                                     opt.o_mdtname);
 588                         return rc;
 589                 }
 590
 591                 while (!eof && !stop) {
 592                         rc = llapi_changelog_recv(chglog_hdlr, &rec);
 593                         switch (rc) {
 594                         case 0:
 595                                 rc = process_record(rec);
 596                                 if (rc) {
 597                                         llapi_error(LLAPI_MSG_ERROR, rc,
 598                                                     "failed to process record");
 599                                         ret = rc;
 600                                 }
 601
 602                                 llapi_changelog_free(&rec);
 603
 604                                 rc = lsom_check_sync();
 605                                 if (rc) {
 606                                         stop = true;
 607                                         ret = rc;
 608                                 }
 609
 610                                 break;
 611                         case 1: /* EOF */
 612                                 llapi_printf(LLAPI_MSG_DEBUG,
 613                                              "finished reading [%s]\n",
 614                                              opt.o_mdtname);
 615                                 eof = true;
 616                                 break;
 617                         case -EINVAL: /* FS unmounted */
 618                         case -EPROTO:  /* error in KUC channel */
 619                         default:
 620                                 stop = true;
 621                                 llapi_error(LLAPI_MSG_ERROR, rc,
 622                                             "failed to get changelog record");
 623                                 ret = rc;
 624                                 break;
 625                         }
 626                 }
 627
 628                 /* reach EOF of changelog */
 629                 rc = llapi_changelog_fini(&chglog_hdlr);
 630                 if (rc) {
 631                         llapi_error(LLAPI_MSG_ERROR, rc,
 632                                     "unable to close changelog of MDT [%s]",
 633                                     opt.o_mdtname);
 634                         ret = rc;
 635                         return rc;
 636                 }
 637
 638                 if (opt.o_daemonize) {
 639                         sleep(opt.o_intv);
 640
 641                         rc = lsom_check_sync();
 642                         if (rc) {
 643                                 stop = true;
 644                                 ret = rc;
 645                         }
 646                 } else {
 647                         lsom_start_update(head.lh_cached_count);
 648                         stop = true;
 649                 }
 650         }
 651
 652         lsom_cleanup();
 653         return ret;
 654 }