4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2017, DDN Storage Corporation.
26 * lustre/utils/llsom_sync.c
28 * Tool for sync the LSOM xattr.
30 * Author: Qian Yingjin <qian@ddn.com>
41 #include <sys/types.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>
55 #define PATH_MAX (4096)
58 #define container_of(ptr, type, member) ({ \
59 const typeof(((type *) 0)->member) * __mptr = (ptr); \
60 (type *) ((char *) __mptr - offsetof(type, member)); })
62 #define CHLG_POLL_INTV 60
63 #define REC_MIN_AGE 600
64 #define DEF_CACHE_SIZE (256 * 1048576) /* 256MB */
67 const char *o_chlg_user;
68 const char *o_mdtname;
75 unsigned long o_cached_fid_hiwm; /* high watermark */
76 unsigned long o_batch_sync_cnt;
82 struct hlist_node fr_node;
83 struct list_head fr_link;
89 static const int fid_hash_shift = 6;
91 #define FID_HASH_ENTRIES (1 << fid_hash_shift)
92 #define FID_ON_HASH(f) (!hlist_unhashed(&(f)->fr_node))
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))
99 #error Wordsize not 32 or 64
103 struct hlist_head *lh_hash;
104 struct list_head lh_list; /* ordered list by record index */
105 unsigned long lh_cached_count;
108 static void usage(char *prog)
110 printf("\nUsage: %s [options] -u <userid> -m <mdtdev> <mntpt>\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",
122 static inline __u64 fid_flatten(const struct lu_fid *fid)
127 if (fid_is_igif(fid)) {
128 ino = lu_igif_ino(fid);
134 ino = (seq << 24) + ((seq >> 24) & 0xffffff0000ULL) + fid_oid(fid);
136 return ino ?: fid_oid(fid);
140 * map fid to 32 bit value for ino on 32bit systems.
142 static inline __u32 fid_flatten32(const struct lu_fid *fid)
147 if (fid_is_igif(fid)) {
148 ino = lu_igif_ino(fid);
152 seq = fid_seq(fid) - FID_SEQ_START;
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.
160 ino = ((seq & 0x000fffffULL) << 12) + ((seq >> 8) & 0xfffff000) +
161 (seq >> (64 - (40-8)) & 0xffffff00) +
162 (fid_oid(fid) & 0xff000fff) + ((fid_oid(fid) & 0x00fff000) << 8);
164 return ino ?: fid_oid(fid);
167 static inline bool fid_eq(const lustre_fid *f1, const lustre_fid *f2)
169 return f1->f_seq == f2->f_seq && f1->f_oid == f2->f_oid &&
170 f1->f_ver == f2->f_ver;
173 static void fid_hash_del(struct fid_rec *f)
176 hlist_del_init(&f->fr_node);
179 static void fid_hash_add(struct fid_rec *f)
181 assert(!FID_ON_HASH(f));
182 hlist_add_head(&f->fr_node, &head.lh_hash[FID_HASH_FN(&f->fr_fid)]);
185 static struct fid_rec *fid_hash_find(const lustre_fid *fid)
187 struct hlist_head *hash_list;
188 struct hlist_node *entry, *next;
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))
201 static int lsom_setup(void)
205 /* set llapi message level */
206 llapi_msg_set_level(opt.o_verbose);
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);
217 for (i = 0; i < FID_HASH_ENTRIES; i++)
218 INIT_HLIST_HEAD(&head.lh_hash[i]);
220 INIT_LIST_HEAD(&head.lh_list);
224 static void lsom_cleanup(void)
229 static int lsom_update_one(struct fid_rec *f)
235 fd = llapi_open_by_fid(opt.o_mntpt, &f->fr_fid,
236 O_RDONLY | O_NOATIME);
240 /* The file may be deleted, clean the corresponding
241 * changelog record and ignore this error.
246 llapi_error(LLAPI_MSG_ERROR, rc,
247 "llapi_open_by_fid for " DFID " failed",
252 if (opt.o_data_sync) {
255 /* Flush dirty pages from clients */
256 rc = llapi_get_data_version(fd, &dv, LL_DV_RD_FLUSH);
258 llapi_error(LLAPI_MSG_ERROR, errno,
259 "failed to sync data for " DFID,
261 /* ignore this error, continue to sync lsom data */
266 llapi_error(LLAPI_MSG_ERROR, rc, "failed to stat FID: " DFID,
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.
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);
283 rc = llapi_changelog_clear(opt.o_mdtname,
284 opt.o_chlg_user, f->fr_index);
286 llapi_error(LLAPI_MSG_ERROR, rc,
287 "failed to clear changelog record: %s:%llu",
288 opt.o_chlg_user, f->fr_index);
292 static int lsom_start_update(int count)
297 llapi_printf(LLAPI_MSG_INFO, "Start to sync %d records.\n", count);
302 f = list_entry(head.lh_list.next, struct fid_rec, fr_link);
303 rc = lsom_update_one(f);
305 list_del_init(&f->fr_link);
308 head.lh_cached_count--;
319 static int lsom_check_sync(void)
326 if (list_empty(&head.lh_list))
329 if (head.lh_cached_count > opt.o_cached_fid_hiwm)
330 count = opt.o_batch_sync_cnt;
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.
340 f = list_entry(head.lh_list.next, struct fid_rec, fr_link);
341 if (now > ((f->fr_time >> 30) + opt.o_min_age))
346 rc = lsom_start_update(count);
348 if (rc == 0 && count == 1)
354 static void lsom_sort_record_list(struct fid_rec *f)
356 struct list_head *pos;
357 bool need_move = false;
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);
362 if (f->fr_index > rec->fr_index) {
371 list_move_tail(&f->fr_link, pos);
374 static int process_record(struct changelog_rec *rec)
376 __u64 index = rec->cr_index;
379 if (rec->cr_type == CL_CLOSE || rec->cr_type == CL_TRUNC ||
380 rec->cr_type == CL_SETATTR) {
383 f = fid_hash_find(&rec->cr_tfid);
385 f = malloc(sizeof(struct fid_rec));
388 llapi_error(LLAPI_MSG_ERROR, rc,
389 "failed to alloc memory for fid_rec");
393 f->fr_fid = rec->cr_tfid;
395 f->fr_time = rec->cr_time;
396 INIT_HLIST_NODE(&f->fr_node);
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.
403 list_add_tail(&f->fr_link, &head.lh_list);
404 head.lh_cached_count++;
407 lsom_sort_record_list(f);
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));
419 static unsigned long get_fid_cache_size(int pct)
421 struct sysinfo sinfo;
422 unsigned long cache_size;
425 rc = sysinfo(&sinfo);
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).
431 cache_size = DEF_CACHE_SIZE;
433 /* maximum cached fid size is tunned according to total
434 * memory size, e.g. 5% of the memroy.
436 cache_size = sinfo.totalram * pct / 100;
442 int main(int argc, char **argv)
447 struct changelog_rec *rec;
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' },
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;
471 while ((c = getopt_long(argc, argv, "u:hm:dsi:a:c:v", options, NULL))
476 llapi_error(LLAPI_MSG_ERROR, rc,
477 "%s: unknown option '-%c'\n",
481 opt.o_chlg_user = optarg;
487 opt.o_mdtname = optarg;
490 opt.o_daemonize = true;
493 opt.o_intv = atoi(optarg);
494 if (opt.o_intv < 0) {
496 llapi_error(LLAPI_MSG_ERROR, rc,
497 "bad value for -i %s", optarg);
502 opt.o_min_age = atoi(optarg);
503 if (opt.o_min_age < 0) {
505 llapi_error(LLAPI_MSG_ERROR, rc,
506 "bad value for -a %s", optarg);
511 rc = Parser_size(&cache_size, optarg);
514 llapi_error(LLAPI_MSG_ERROR, rc,
515 "bad valud for -c '%s'", optarg);
519 /* For value < 100, it is taken as the percentage of
520 * total memory instead.
522 if (cache_size < 100)
523 cache_size = get_fid_cache_size(cache_size);
524 llapi_printf(LLAPI_MSG_INFO, "Cache size: %lu\n",
531 opt.o_data_sync = true;
536 if (argc != optind + 1) {
537 llapi_err_noerrno(LLAPI_MSG_ERROR,
538 "%s: no mount point specified\n", argv[0]);
542 opt.o_mntpt = argv[optind];
543 rc = llapi_search_fsname(opt.o_mntpt, fsname);
545 llapi_error(LLAPI_MSG_ERROR, rc,
546 "cannot find a Lustre file system mounted at '%s'",
554 if (!opt.o_chlg_user)
557 if (opt.o_daemonize) {
561 llapi_error(LLAPI_MSG_ERROR, rc, "cannot daemonize");
565 setbuf(stdout, NULL);
568 opt.o_cached_fid_hiwm = cache_size / sizeof(struct fid_rec);
569 opt.o_batch_sync_cnt = opt.o_cached_fid_hiwm / 2;
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,
585 llapi_error(LLAPI_MSG_ERROR, rc,
586 "unable to open changelog of MDT [%s]\n",
591 while (!eof && !stop) {
592 rc = llapi_changelog_recv(chglog_hdlr, &rec);
595 rc = process_record(rec);
597 llapi_error(LLAPI_MSG_ERROR, rc,
598 "failed to process record");
602 llapi_changelog_free(&rec);
604 rc = lsom_check_sync();
612 llapi_printf(LLAPI_MSG_DEBUG,
613 "finished reading [%s]\n",
617 case -EINVAL: /* FS unmounted */
618 case -EPROTO: /* error in KUC channel */
621 llapi_error(LLAPI_MSG_ERROR, rc,
622 "failed to get changelog record");
628 /* reach EOF of changelog */
629 rc = llapi_changelog_fini(&chglog_hdlr);
631 llapi_error(LLAPI_MSG_ERROR, rc,
632 "unable to close changelog of MDT [%s]",
638 if (opt.o_daemonize) {
641 rc = lsom_check_sync();
647 lsom_start_update(head.lh_cached_count);