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,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License version 2 for more details. A copy is
14 * included in the COPYING file that accompanied this code.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 * Copyright (c) 2017, Intel Corporation.
25 * lustre/mgs/mgs_barrier.c
27 * Author: Fan, Yong <fan.yong@intel.com>
30 #define DEBUG_SUBSYSTEM S_MGS
31 #define D_MGS D_CONFIG
33 #include <uapi/linux/lustre/lustre_ioctl.h>
34 #include <lustre_swab.h>
35 #include <uapi/linux/lustre/lustre_barrier_user.h>
37 #include "mgs_internal.h"
40 * Handle the barrier lock glimpse reply.
42 * The barrier lock glimpse reply contains the target MDT's index and
43 * the barrier operation status on such MDT. With such infomation. If
44 * the MDT given barrier status is the expected one, then set related
45 * 'fsdb''s barrier bitmap; otherwise record the failure or status.
47 * \param[in] env pointer to the thread context
48 * \param[in] req pointer to the glimpse callback RPC request
49 * \param[in] data pointer the async glimpse callback data
50 * \param[in] rc the glimpse callback RPC return value
52 * \retval 0 for success
53 * \retval negative error number on failure
55 static int mgs_barrier_gl_interpret_reply(const struct lu_env *env,
56 struct ptlrpc_request *req,
59 struct ldlm_cb_async_args *ca = data;
60 struct fs_db *fsdb = ca->ca_set_arg->gl_interpret_data;
61 struct barrier_lvb *lvb;
66 /* The lock is useless, cancel it. */
67 ldlm_lock_cancel(ca->ca_lock);
74 lvb = req_capsule_server_swab_get(&req->rq_pill, &RMF_DLM_LVB,
75 lustre_swab_barrier_lvb);
77 GOTO(out, rc = -EPROTO);
79 if (lvb->lvb_status == fsdb->fsdb_barrier_expected) {
80 if (unlikely(lvb->lvb_index > INDEX_MAP_SIZE))
83 set_bit(lvb->lvb_index, fsdb->fsdb_barrier_map);
84 } else if (likely(!test_bit(lvb->lvb_index, fsdb->fsdb_barrier_map))) {
85 fsdb->fsdb_barrier_result = lvb->lvb_status;
92 fsdb->fsdb_barrier_result = rc;
98 * Send glimpse callback to the barrier locks holders.
100 * The glimpse callback takes the current barrier status. The barrier locks
101 * holders (on the MDTs) will take related barrier actions according to the
102 * given barrier status, then return their local barrier status.
104 * \param[in] env pointer to the thread context
105 * \param[in] mgs pointer to the MGS device
106 * \param[in] fsdb pointer the barrier 'fsdb'
107 * \param[in] timeout indicate when the barrier will be expired
108 * \param[in] expected the expected barrier status on remote servers (MDTs)
110 * \retval positive number for unexpected barrier status
111 * \retval 0 for success
112 * \retval negative error number on failure
114 static int mgs_barrier_glimpse_lock(const struct lu_env *env,
115 struct mgs_device *mgs,
117 __u32 timeout, __u32 expected)
119 union ldlm_gl_desc *desc = &mgs_env_info(env)->mgi_gl_desc;
120 struct ldlm_res_id res_id;
121 struct ldlm_resource *res;
122 struct ldlm_glimpse_work *work;
123 struct ldlm_glimpse_work *tmp;
125 struct list_head *pos;
130 LASSERT(fsdb->fsdb_mdt_count > 0);
132 rc = mgc_logname2resid(fsdb->fsdb_name, &res_id, CONFIG_T_BARRIER);
136 res = ldlm_resource_get(mgs->mgs_obd->obd_namespace, NULL, &res_id,
139 RETURN(PTR_ERR(res));
141 fsdb->fsdb_barrier_result = 0;
142 fsdb->fsdb_barrier_expected = expected;
143 desc->barrier_desc.lgbd_status = fsdb->fsdb_barrier_status;
144 desc->barrier_desc.lgbd_timeout = timeout;
147 list_for_each_entry(work, &gl_list, gl_list) {
151 LDLM_LOCK_RELEASE(work->gl_lock);
152 work->gl_lock = NULL;
155 /* It is not big issue to alloc more work item than needed. */
156 for (i = 0; i < fsdb->fsdb_mdt_count; i++) {
159 GOTO(out, rc = -ENOMEM);
161 list_add_tail(&work->gl_list, &gl_list);
164 work = list_entry(gl_list.next, struct ldlm_glimpse_work, gl_list);
167 list_for_each(pos, &res->lr_granted) {
168 struct ldlm_lock *lock = list_entry(pos, struct ldlm_lock,
171 work->gl_lock = LDLM_LOCK_GET(lock);
173 work->gl_desc = desc;
174 work->gl_interpret_reply = mgs_barrier_gl_interpret_reply;
175 work->gl_interpret_data = fsdb;
177 if (unlikely(work->gl_list.next == &gl_list)) {
178 if (likely(pos->next == &res->lr_granted))
182 /* The granted locks are more than the MDTs count. */
186 work = list_entry(work->gl_list.next, struct ldlm_glimpse_work,
191 /* The MDTs count may be more than the granted locks. */
192 list_for_each_entry_safe_reverse(work, tmp, &gl_list, gl_list) {
196 list_del(&work->gl_list);
200 if (!list_empty(&gl_list))
201 rc = ldlm_glimpse_locks(res, &gl_list);
208 list_for_each_entry_safe(work, tmp, &gl_list, gl_list) {
209 list_del(&work->gl_list);
211 LDLM_LOCK_RELEASE(work->gl_lock);
215 ldlm_resource_putref(res);
217 rc = fsdb->fsdb_barrier_result;
222 static void mgs_barrier_bitmap_setup(struct mgs_device *mgs,
223 struct fs_db *b_fsdb,
226 struct fs_db *c_fsdb;
228 c_fsdb = mgs_find_fsdb(mgs, name);
229 if (likely(c_fsdb)) {
230 memcpy(b_fsdb->fsdb_mdt_index_map,
231 c_fsdb->fsdb_mdt_index_map, INDEX_MAP_SIZE);
232 b_fsdb->fsdb_mdt_count = c_fsdb->fsdb_mdt_count;
233 mgs_put_fsdb(mgs, c_fsdb);
237 static bool mgs_barrier_done(struct fs_db *fsdb)
241 for (i = 0; i < INDEX_MAP_SIZE * 8; i++) {
242 if (test_bit(i, fsdb->fsdb_mdt_index_map) &&
243 !test_bit(i, fsdb->fsdb_barrier_map))
250 bool mgs_barrier_expired(struct fs_db *fsdb, time64_t timeout)
252 time64_t expired = fsdb->fsdb_barrier_latest_create_time + timeout;
254 return expired > ktime_get_real_seconds();
258 * Create the barrier for the given instance.
260 * We use two-phases barrier to guarantee that after the barrier setup:
261 * 1) All the server side pending async modification RPCs have been flushed.
262 * 2) Any subsequent modification will be blocked.
263 * 3) All async transactions on the MDTs have been committed.
265 * For phase1, we do the following:
267 * Firstly, it sets barrier flag on the instance that will block subsequent
268 * modifications from clients. (Note: server sponsored modification will be
269 * allowed for flush pending modifications)
271 * Secondly, it will flush all pending modification via dt_sync(), such as
272 * async OST-object destroy, async OST-object owner changes, and so on.
274 * If there are some on-handling clients sponsored modifications during the
275 * barrier creating, then related modifications may cause pending requests
276 * after the first dt_sync(), so call dt_sync() again after all on-handling
277 * modifications done.
279 * With the phase1 barrier set, all pending cross-servers modification RPCs
280 * have been flushed to remote servers, and any new modification will be
281 * blocked. But it does not guarantees that all the updates have been
282 * committed to storage on remote servers. So when all the instances have
283 * done phase1 barrier successfully, the MGS will notify all instances to
284 * do the phase2 barrier as following:
286 * Every barrier instance will call dt_sync() to make all async transactions
287 * to be committed locally.
289 * \param[in] env pointer to the thread context
290 * \param[in] mgs pointer to the MGS device
291 * \param[in] bc pointer the barrier control structure
293 * \retval 0 for success
294 * \retval negative error number on failure
296 static int mgs_barrier_freeze(const struct lu_env *env,
297 struct mgs_device *mgs,
298 struct barrier_ctl *bc)
300 char *name = mgs_env_info(env)->mgi_fsname;
308 snprintf(name, sizeof(mgs_env_info(env)->mgi_fsname) - 1, "%s-%s",
309 bc->bc_name, BARRIER_FILENAME);
311 down_write(&mgs->mgs_barrier_rwsem);
312 mutex_lock(&mgs->mgs_mutex);
314 rc = mgs_find_or_make_fsdb_nolock(env, mgs, name, &fsdb);
316 mutex_unlock(&mgs->mgs_mutex);
317 up_write(&mgs->mgs_barrier_rwsem);
321 if (unlikely(fsdb->fsdb_mdt_count == 0)) {
322 mgs_barrier_bitmap_setup(mgs, fsdb, bc->bc_name);
324 /* fsdb was just created, ensure that fsdb_barrier_disabled is
326 if (fsdb->fsdb_mdt_count > 0) {
327 struct obd_export *exp;
328 struct obd_device *mgs_obd = mgs->mgs_obd;
330 spin_lock(&mgs_obd->obd_dev_lock);
331 list_for_each_entry(exp, &mgs_obd->obd_exports,
333 __u64 flags = exp_connect_flags(exp);
334 if (!!(flags & OBD_CONNECT_MDS_MDS) &&
335 !(flags & OBD_CONNECT_BARRIER)) {
336 fsdb->fsdb_barrier_disabled = 1;
340 spin_unlock(&mgs_obd->obd_dev_lock);
344 mutex_lock(&fsdb->fsdb_mutex);
345 mutex_unlock(&mgs->mgs_mutex);
347 switch (fsdb->fsdb_barrier_status) {
357 if (mgs_barrier_expired(fsdb, fsdb->fsdb_barrier_timeout)) {
366 if (fsdb->fsdb_barrier_disabled) {
368 } else if (unlikely(fsdb->fsdb_mdt_count == 0)) {
371 fsdb->fsdb_barrier_latest_create_time =
372 ktime_get_real_seconds();
373 fsdb->fsdb_barrier_status = BS_FREEZING_P1;
374 if (bc->bc_timeout != 0)
375 fsdb->fsdb_barrier_timeout = bc->bc_timeout;
377 fsdb->fsdb_barrier_timeout =
378 BARRIER_TIMEOUT_DEFAULT;
379 memset(fsdb->fsdb_barrier_map, 0, INDEX_MAP_SIZE);
383 LCONSOLE_WARN("%s: found unexpected barrier status %u\n",
384 bc->bc_name, fsdb->fsdb_barrier_status);
392 left = fsdb->fsdb_barrier_timeout;
395 mutex_unlock(&fsdb->fsdb_mutex);
396 up_write(&mgs->mgs_barrier_rwsem);
398 CFS_FAIL_TIMEOUT(OBD_FAIL_BARRIER_DELAY, cfs_fail_val);
400 rc = mgs_barrier_glimpse_lock(env, mgs, fsdb, left,
401 phase1 ? BS_FREEZING_P1 : BS_FROZEN);
402 down_write(&mgs->mgs_barrier_rwsem);
403 mutex_lock(&fsdb->fsdb_mutex);
406 left = fsdb->fsdb_barrier_latest_create_time +
407 fsdb->fsdb_barrier_timeout - ktime_get_real_seconds();
409 fsdb->fsdb_barrier_status = BS_EXPIRED;
411 GOTO(out, rc = -ETIME);
414 LASSERTF(fsdb->fsdb_barrier_status ==
415 (phase1 ? BS_FREEZING_P1 : BS_FREEZING_P2),
416 "unexpected barrier status %u\n",
417 fsdb->fsdb_barrier_status);
419 if (rc == -ETIMEDOUT) {
420 fsdb->fsdb_barrier_status = BS_EXPIRED;
423 fsdb->fsdb_barrier_status = rc;
426 fsdb->fsdb_barrier_status = BS_FAILED;
427 } else if (mgs_barrier_done(fsdb)) {
429 fsdb->fsdb_barrier_status = BS_FREEZING_P2;
430 memset(fsdb->fsdb_barrier_map, 0,
436 fsdb->fsdb_barrier_status = BS_FROZEN;
439 fsdb->fsdb_barrier_status = BS_FAILED;
446 mutex_unlock(&fsdb->fsdb_mutex);
447 up_write(&mgs->mgs_barrier_rwsem);
449 memset(fsdb->fsdb_barrier_map, 0, INDEX_MAP_SIZE);
450 mgs_barrier_glimpse_lock(env, mgs, fsdb, 0, BS_THAWED);
453 mgs_put_fsdb(mgs, fsdb);
458 static int mgs_barrier_thaw(const struct lu_env *env,
459 struct mgs_device *mgs,
460 struct barrier_ctl *bc)
462 char *name = mgs_env_info(env)->mgi_fsname;
467 snprintf(name, sizeof(mgs_env_info(env)->mgi_fsname) - 1, "%s-%s",
468 bc->bc_name, BARRIER_FILENAME);
470 down_write(&mgs->mgs_barrier_rwsem);
471 mutex_lock(&mgs->mgs_mutex);
473 rc = mgs_find_or_make_fsdb_nolock(env, mgs, name, &fsdb);
475 mutex_unlock(&mgs->mgs_mutex);
476 up_write(&mgs->mgs_barrier_rwsem);
480 if (unlikely(fsdb->fsdb_mdt_count == 0)) {
481 mgs_barrier_bitmap_setup(mgs, fsdb, bc->bc_name);
483 /* fsdb was just created, ensure that fsdb_barrier_disabled is
485 if (fsdb->fsdb_mdt_count > 0) {
486 struct obd_export *exp;
487 struct obd_device *mgs_obd = mgs->mgs_obd;
489 spin_lock(&mgs_obd->obd_dev_lock);
490 list_for_each_entry(exp, &mgs_obd->obd_exports,
492 __u64 flags = exp_connect_flags(exp);
493 if (!!(flags & OBD_CONNECT_MDS_MDS) &&
494 !(flags & OBD_CONNECT_BARRIER)) {
495 fsdb->fsdb_barrier_disabled = 1;
499 spin_unlock(&mgs_obd->obd_dev_lock);
503 mutex_lock(&fsdb->fsdb_mutex);
504 mutex_unlock(&mgs->mgs_mutex);
506 switch (fsdb->fsdb_barrier_status) {
520 case BS_EXPIRED: /* The barrier on some MDT(s) may be expired,
521 * but may be not on others. Destory anyway. */
523 if (unlikely(fsdb->fsdb_mdt_count == 0)) {
526 fsdb->fsdb_barrier_status = BS_THAWING;
527 memset(fsdb->fsdb_barrier_map, 0, INDEX_MAP_SIZE);
531 LCONSOLE_WARN("%s: found unexpected barrier status %u\n",
532 bc->bc_name, fsdb->fsdb_barrier_status);
540 mutex_unlock(&fsdb->fsdb_mutex);
541 up_write(&mgs->mgs_barrier_rwsem);
543 CFS_FAIL_TIMEOUT(OBD_FAIL_BARRIER_DELAY, cfs_fail_val);
545 rc = mgs_barrier_glimpse_lock(env, mgs, fsdb, 0, BS_THAWED);
546 down_write(&mgs->mgs_barrier_rwsem);
547 mutex_lock(&fsdb->fsdb_mutex);
549 LASSERTF(fsdb->fsdb_barrier_status == BS_THAWING,
550 "unexpected barrier status %u\n",
551 fsdb->fsdb_barrier_status);
554 fsdb->fsdb_barrier_status = rc;
557 fsdb->fsdb_barrier_status = BS_FAILED;
558 } else if (mgs_barrier_done(fsdb)) {
559 fsdb->fsdb_barrier_status = BS_THAWED;
561 fsdb->fsdb_barrier_status = BS_FAILED;
568 mutex_unlock(&fsdb->fsdb_mutex);
569 up_write(&mgs->mgs_barrier_rwsem);
570 mgs_put_fsdb(mgs, fsdb);
575 static int mgs_barrier_stat(const struct lu_env *env,
576 struct mgs_device *mgs,
577 struct barrier_ctl *bc)
579 char *name = mgs_env_info(env)->mgi_fsname;
583 snprintf(name, sizeof(mgs_env_info(env)->mgi_fsname) - 1, "%s-%s",
584 bc->bc_name, BARRIER_FILENAME);
586 mutex_lock(&mgs->mgs_mutex);
588 fsdb = mgs_find_fsdb(mgs, name);
590 mutex_lock(&fsdb->fsdb_mutex);
591 mutex_unlock(&mgs->mgs_mutex);
593 bc->bc_status = fsdb->fsdb_barrier_status;
594 if (bc->bc_status == BS_FREEZING_P1 ||
595 bc->bc_status == BS_FREEZING_P2 ||
596 bc->bc_status == BS_FROZEN) {
597 if (mgs_barrier_expired(fsdb, fsdb->fsdb_barrier_timeout))
599 fsdb->fsdb_barrier_latest_create_time +
600 fsdb->fsdb_barrier_timeout -
601 ktime_get_real_seconds();
603 bc->bc_status = fsdb->fsdb_barrier_status =
607 mutex_unlock(&fsdb->fsdb_mutex);
608 mgs_put_fsdb(mgs, fsdb);
610 mutex_unlock(&mgs->mgs_mutex);
612 bc->bc_status = BS_INIT;
618 static int mgs_barrier_rescan(const struct lu_env *env,
619 struct mgs_device *mgs,
620 struct barrier_ctl *bc)
622 char *name = mgs_env_info(env)->mgi_fsname;
623 struct fs_db *b_fsdb;
624 struct fs_db *c_fsdb;
628 down_write(&mgs->mgs_barrier_rwsem);
629 mutex_lock(&mgs->mgs_mutex);
631 c_fsdb = mgs_find_fsdb(mgs, bc->bc_name);
632 if (!c_fsdb || unlikely(c_fsdb->fsdb_mdt_count == 0)) {
633 mutex_unlock(&mgs->mgs_mutex);
634 up_write(&mgs->mgs_barrier_rwsem);
639 snprintf(name, sizeof(mgs_env_info(env)->mgi_fsname) - 1, "%s-%s",
640 bc->bc_name, BARRIER_FILENAME);
641 rc = mgs_find_or_make_fsdb_nolock(env, mgs, name, &b_fsdb);
643 mutex_unlock(&mgs->mgs_mutex);
644 up_write(&mgs->mgs_barrier_rwsem);
645 mgs_put_fsdb(mgs, c_fsdb);
649 if (unlikely(b_fsdb->fsdb_mdt_count == 0 &&
650 c_fsdb->fsdb_mdt_count > 0)) {
651 /* fsdb was just created, ensure that fsdb_barrier_disabled is
653 struct obd_export *exp;
654 struct obd_device *mgs_obd = mgs->mgs_obd;
656 spin_lock(&mgs_obd->obd_dev_lock);
657 list_for_each_entry(exp, &mgs_obd->obd_exports,
659 __u64 flags = exp_connect_flags(exp);
660 if (!!(flags & OBD_CONNECT_MDS_MDS) &&
661 !(flags & OBD_CONNECT_BARRIER)) {
662 b_fsdb->fsdb_barrier_disabled = 1;
666 spin_unlock(&mgs_obd->obd_dev_lock);
669 mutex_lock(&b_fsdb->fsdb_mutex);
670 mutex_lock(&c_fsdb->fsdb_mutex);
671 mutex_unlock(&mgs->mgs_mutex);
673 switch (b_fsdb->fsdb_barrier_status) {
683 if (mgs_barrier_expired(b_fsdb, b_fsdb->fsdb_barrier_timeout)) {
692 b_fsdb->fsdb_barrier_latest_create_time = ktime_get_real_seconds();
693 b_fsdb->fsdb_barrier_status = BS_RESCAN;
694 memcpy(b_fsdb->fsdb_mdt_index_map, c_fsdb->fsdb_mdt_index_map,
696 memset(b_fsdb->fsdb_barrier_map, 0, INDEX_MAP_SIZE);
697 b_fsdb->fsdb_mdt_count = c_fsdb->fsdb_mdt_count;
700 LCONSOLE_WARN("%s: found unexpected barrier status %u\n",
701 bc->bc_name, b_fsdb->fsdb_barrier_status);
706 mutex_unlock(&c_fsdb->fsdb_mutex);
707 mgs_put_fsdb(mgs, c_fsdb);
713 mutex_unlock(&b_fsdb->fsdb_mutex);
714 up_write(&mgs->mgs_barrier_rwsem);
715 rc = mgs_barrier_glimpse_lock(env, mgs, b_fsdb, 0, BS_INIT);
716 down_write(&mgs->mgs_barrier_rwsem);
717 mutex_lock(&b_fsdb->fsdb_mutex);
719 LASSERTF(b_fsdb->fsdb_barrier_status == BS_RESCAN,
720 "unexpected barrier status %u\n",
721 b_fsdb->fsdb_barrier_status);
724 b_fsdb->fsdb_barrier_status = rc;
726 } else if (rc == -ETIMEDOUT &&
727 mgs_barrier_expired(b_fsdb, bc->bc_timeout)) {
728 memset(b_fsdb->fsdb_barrier_map, 0, INDEX_MAP_SIZE);
731 } else if (rc < 0 && rc != -ETIMEDOUT && rc != -ENODEV) {
732 b_fsdb->fsdb_barrier_status = BS_FAILED;
736 b_fsdb->fsdb_mdt_count = 0;
740 for (i = 0; i < INDEX_MAP_SIZE * 8; i++) {
741 if (test_bit(i, b_fsdb->fsdb_barrier_map)) {
742 b_fsdb->fsdb_mdt_count++;
743 } else if (test_bit(i, b_fsdb->fsdb_mdt_index_map)) {
744 b_fsdb->fsdb_mdt_count++;
749 bc->bc_total = b_fsdb->fsdb_mdt_count;
750 memcpy(b_fsdb->fsdb_mdt_index_map,
751 b_fsdb->fsdb_barrier_map, INDEX_MAP_SIZE);
752 b_fsdb->fsdb_barrier_status = BS_INIT;
758 mutex_unlock(&b_fsdb->fsdb_mutex);
759 up_write(&mgs->mgs_barrier_rwsem);
760 mgs_put_fsdb(mgs, b_fsdb);
765 int mgs_iocontrol_barrier(const struct lu_env *env,
766 struct mgs_device *mgs,
767 struct obd_ioctl_data *data)
769 struct barrier_ctl *bc = (struct barrier_ctl *)(data->ioc_inlbuf1);
773 if (unlikely(bc->bc_version != BARRIER_VERSION_V1))
776 if (unlikely(bc->bc_name[0] == '\0' ||
777 strnlen(bc->bc_name, sizeof(bc->bc_name)) > 8))
780 /* NOT allow barrier operations during recovery. */
781 if (unlikely(mgs->mgs_obd->obd_recovering))
784 switch (bc->bc_cmd) {
786 rc = mgs_barrier_freeze(env, mgs, bc);
789 rc = mgs_barrier_thaw(env, mgs, bc);
792 rc = mgs_barrier_stat(env, mgs, bc);
795 rc = mgs_barrier_rescan(env, mgs, bc);