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) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ofd/ofd_obd.c
34 * This file contains OBD API methods for OBD Filter Device (OFD) which are
35 * used for export handling, configuration purposes and recovery.
36 * Several methods are used by ECHO client only since it still uses OBD API.
37 * Such methods have _echo_ prefix in name.
39 * Author: Andreas Dilger <andreas.dilger@intel.com>
40 * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
41 * Author: Mikhail Pershin <mike.pershin@intel.com>
44 #define DEBUG_SUBSYSTEM S_FILTER
46 #include <lustre/lustre_idl.h>
47 #include "ofd_internal.h"
48 #include <obd_cksum.h>
49 #include <lustre_ioctl.h>
50 #include <lustre_quota.h>
51 #include <lustre_lfsck.h>
52 #include <lustre_nodemap.h>
55 * Initialize OFD per-export statistics.
57 * This function sets up procfs entries for various OFD export counters. These
58 * counters are for per-client statistics tracked on the server.
60 * \param[in] ofd OFD device
61 * \param[in] exp OBD export
62 * \param[in] client_nid NID of client
64 * \retval 0 if successful
65 * \retval negative value on error
67 static int ofd_export_stats_init(struct ofd_device *ofd,
68 struct obd_export *exp,
69 lnet_nid_t *client_nid)
71 struct obd_device *obd = ofd_obd(ofd);
72 struct nid_stat *stats;
76 LASSERT(obd->obd_uses_nid_stats);
78 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
79 /* Self-export gets no proc entry */
82 rc = lprocfs_exp_setup(exp, client_nid);
84 /* Mask error for already created /proc entries */
85 RETURN(rc == -EALREADY ? 0 : rc);
87 stats = exp->exp_nid_stats;
88 stats->nid_stats = lprocfs_alloc_stats(NUM_OBD_STATS +
90 LPROCFS_STATS_FLAG_NOPERCPU);
91 if (stats->nid_stats == NULL)
94 lprocfs_init_ops_stats(LPROC_OFD_STATS_LAST, stats->nid_stats);
96 ofd_stats_counter_init(stats->nid_stats);
98 rc = lprocfs_register_stats(stats->nid_proc, "stats", stats->nid_stats);
100 lprocfs_free_stats(&stats->nid_stats);
104 rc = lprocfs_nid_ldlm_stats_init(stats);
113 * Match client and OST server connection feature flags.
115 * Compute the compatibility flags for a connection request based on
116 * features mutually supported by client and server.
118 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
119 * must not be updated here, otherwise a partially initialized value may
120 * be exposed. After the connection request is successfully processed,
121 * the top-level tgt_connect() request handler atomically updates the export
122 * connect flags from the obd_connect_data::ocd_connect_flags field of the
123 * reply. \see tgt_connect().
125 * \param[in] env execution environment
126 * \param[in] exp the obd_export associated with this
128 * \param[in] data stores data for this connect request
129 * \param[in] new_connection is this connection new or not
131 * \retval 0 if success
132 * \retval -EPROTO client and server feature requirements are
134 * \retval -EBADF OST index in connect request doesn't match
137 static int ofd_parse_connect_data(const struct lu_env *env,
138 struct obd_export *exp,
139 struct obd_connect_data *data,
142 struct ofd_device *ofd = ofd_exp(exp);
143 struct filter_export_data *fed = &exp->exp_filter_data;
148 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
149 " ocd_version: %x ocd_grant: %d ocd_index: %u"
151 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
152 data->ocd_connect_flags, data->ocd_version,
153 data->ocd_grant, data->ocd_index, data->ocd_group);
155 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
156 CWARN("!!! This export (nid %s) used object group %d "
157 "earlier; now it's trying to use group %d! This could "
158 "be a bug in the MDS. Please report to "
159 "https://jira.hpdd.intel.com/\n",
160 obd_export_nid2str(exp), fed->fed_group,
164 fed->fed_group = data->ocd_group;
166 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
167 data->ocd_version = LUSTRE_VERSION_CODE;
169 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
170 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
171 CDEBUG(D_HA, "%s: Received MDS connection for group %u\n",
172 exp->exp_obd->obd_name, data->ocd_group);
173 else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
176 if (ofd_grant_param_supp(exp)) {
177 exp->exp_filter_data.fed_pagesize = data->ocd_blocksize;
178 /* ocd_{blocksize,inodespace} are log2 values */
179 data->ocd_blocksize = ofd->ofd_blockbits;
180 data->ocd_inodespace = ofd->ofd_dt_conf.ddp_inodespace;
181 /* ocd_grant_extent is in 1K blocks */
182 data->ocd_grant_extent = ofd->ofd_dt_conf.ddp_grant_frag >> 10;
185 if (data->ocd_connect_flags & OBD_CONNECT_GRANT)
186 data->ocd_grant = ofd_grant_connect(env, exp, data->ocd_grant,
189 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
190 struct lr_server_data *lsd = &ofd->ofd_lut.lut_lsd;
191 int index = lsd->lsd_osd_index;
193 if (index != data->ocd_index) {
194 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
195 " %u doesn't match actual OST index"
196 " %u in last_rcvd file, bad "
198 obd_export_nid2str(exp), index,
202 if (!(lsd->lsd_feature_compat & OBD_COMPAT_OST)) {
203 /* this will only happen on the first connect */
204 lsd->lsd_feature_compat |= OBD_COMPAT_OST;
205 /* sync is not needed here as lut_client_add will
206 * set exp_need_sync flag */
207 tgt_server_data_update(env, &ofd->ofd_lut, 0);
210 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
211 data->ocd_brw_size = 65536;
212 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
213 data->ocd_brw_size = min(data->ocd_brw_size,
214 (__u32)DT_MAX_BRW_SIZE);
215 if (data->ocd_brw_size == 0) {
216 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
217 " ocd_version: %x ocd_grant: %d ocd_index: %u "
218 "ocd_brw_size is unexpectedly zero, "
219 "network data corruption?"
220 "Refusing connection of this client\n",
221 exp->exp_obd->obd_name,
222 exp->exp_client_uuid.uuid,
223 exp, data->ocd_connect_flags, data->ocd_version,
224 data->ocd_grant, data->ocd_index);
229 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
230 __u32 cksum_types = data->ocd_cksum_types;
232 /* The client set in ocd_cksum_types the checksum types it
233 * supports. We have to mask off the algorithms that we don't
235 data->ocd_cksum_types &= cksum_types_supported_server();
237 if (unlikely(data->ocd_cksum_types == 0)) {
238 CERROR("%s: Connect with checksum support but no "
239 "ocd_cksum_types is set\n",
240 exp->exp_obd->obd_name);
244 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
245 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
246 cksum_types, data->ocd_cksum_types);
248 /* This client does not support OBD_CONNECT_CKSUM
249 * fall back to CRC32 */
250 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
251 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
252 exp->exp_obd->obd_name, obd_export_nid2str(exp));
255 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
256 data->ocd_maxbytes = ofd->ofd_dt_conf.ddp_maxbytes;
258 if (OCD_HAS_FLAG(data, PINGLESS)) {
259 if (ptlrpc_pinger_suppress_pings()) {
260 spin_lock(&exp->exp_obd->obd_dev_lock);
261 list_del_init(&exp->exp_obd_chain_timed);
262 spin_unlock(&exp->exp_obd->obd_dev_lock);
264 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
272 * Re-initialize export upon client reconnection.
274 * This function parses connection data from reconnect and resets
277 * \param[in] env execution environment
278 * \param[in] exp OBD export
279 * \param[in] obd OFD device
280 * \param[in] cluuid NID of client
281 * \param[in] data connection data from request
282 * \param[in] localdata client NID
284 * \retval 0 if successful
285 * \retval negative value on error
287 static int ofd_obd_reconnect(const struct lu_env *env, struct obd_export *exp,
288 struct obd_device *obd, struct obd_uuid *cluuid,
289 struct obd_connect_data *data,
292 struct ofd_device *ofd;
297 if (exp == NULL || obd == NULL || cluuid == NULL)
300 rc = nodemap_add_member(*(lnet_nid_t *)client_nid, exp);
301 if (rc != 0 && rc != -EEXIST)
304 ofd = ofd_dev(obd->obd_lu_dev);
306 rc = ofd_parse_connect_data(env, exp, data, false);
308 ofd_export_stats_init(ofd, exp, client_nid);
310 nodemap_del_member(exp);
316 * Initialize new client connection.
318 * This function handles new connection to the OFD. The new export is
319 * created (in context of class_connect()) and persistent client data is
320 * initialized on storage.
322 * \param[in] env execution environment
323 * \param[out] _exp stores pointer to new export
324 * \param[in] obd OFD device
325 * \param[in] cluuid client UUID
326 * \param[in] data connection data from request
327 * \param[in] localdata client NID
329 * \retval 0 if successful
330 * \retval negative value on error
332 static int ofd_obd_connect(const struct lu_env *env, struct obd_export **_exp,
333 struct obd_device *obd, struct obd_uuid *cluuid,
334 struct obd_connect_data *data, void *localdata)
336 struct obd_export *exp;
337 struct ofd_device *ofd;
338 struct lustre_handle conn = { 0 };
342 if (_exp == NULL || obd == NULL || cluuid == NULL)
345 ofd = ofd_dev(obd->obd_lu_dev);
347 rc = class_connect(&conn, obd, cluuid);
351 exp = class_conn2export(&conn);
352 LASSERT(exp != NULL);
354 if (localdata != NULL) {
355 rc = nodemap_add_member(*(lnet_nid_t *)localdata, exp);
356 if (rc != 0 && rc != -EEXIST)
359 CDEBUG(D_HA, "%s: cannot find nodemap for client %s: "
360 "nid is null\n", obd->obd_name, cluuid->uuid);
363 rc = ofd_parse_connect_data(env, exp, data, true);
367 if (obd->obd_replayable) {
368 struct tg_export_data *ted = &exp->exp_target_data;
370 memcpy(ted->ted_lcd->lcd_uuid, cluuid,
371 sizeof(ted->ted_lcd->lcd_uuid));
372 rc = tgt_client_new(env, exp);
375 ofd_export_stats_init(ofd, exp, localdata);
378 CDEBUG(D_HA, "%s: get connection from MDS %d\n", obd->obd_name,
379 data ? data->ocd_group : -1);
383 class_disconnect(exp);
384 nodemap_del_member(exp);
393 * Disconnect a connected client.
395 * This function terminates the client connection. The client export is
396 * disconnected (cleaned up) and client data on persistent storage is removed.
398 * \param[in] exp OBD export
400 * \retval 0 if successful
401 * \retval negative value on error
403 int ofd_obd_disconnect(struct obd_export *exp)
405 struct ofd_device *ofd = ofd_exp(exp);
412 class_export_get(exp);
414 if (!(exp->exp_flags & OBD_OPT_FORCE))
415 ofd_grant_sanity_check(ofd_obd(ofd), __FUNCTION__);
417 rc = server_disconnect_export(exp);
419 ofd_grant_discard(exp);
421 /* Do not erase record for recoverable client. */
422 if (exp->exp_obd->obd_replayable &&
423 (!exp->exp_obd->obd_fail || exp->exp_failed)) {
424 rc = lu_env_init(&env, LCT_DT_THREAD);
428 tgt_client_del(&env, exp);
432 nodemap_del_member(exp);
433 class_export_put(exp);
438 * Implementation of obd_ops::o_init_export.
440 * This function is called from class_new_export() and initializes
441 * the OFD-specific data for new export.
443 * \param[in] exp OBD export
445 * \retval 0 if successful
446 * \retval negative value on error
448 static int ofd_init_export(struct obd_export *exp)
452 spin_lock_init(&exp->exp_filter_data.fed_lock);
453 INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
454 atomic_set(&exp->exp_filter_data.fed_soft_sync_count, 0);
455 spin_lock(&exp->exp_lock);
456 exp->exp_connecting = 1;
457 spin_unlock(&exp->exp_lock);
459 /* self-export doesn't need client data and ldlm initialization */
460 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
461 &exp->exp_client_uuid)))
464 rc = tgt_client_alloc(exp);
466 ldlm_init_export(exp);
468 CERROR("%s: Can't initialize export: rc %d\n",
469 exp->exp_obd->obd_name, rc);
474 * Implementation of obd_ops::o_destroy_export.
476 * This function is called from class_export_destroy() to cleanup
477 * the OFD-specific data for export being destroyed.
479 * \param[in] exp OBD export
481 * \retval 0 if successful
482 * \retval negative value on error
484 static int ofd_destroy_export(struct obd_export *exp)
486 struct ofd_device *ofd = ofd_exp(exp);
488 if (exp->exp_filter_data.fed_pending)
489 CERROR("%s: cli %s/%p has %lu pending on destroyed export"
490 "\n", exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
491 exp, exp->exp_filter_data.fed_pending);
493 target_destroy_export(exp);
495 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
496 &exp->exp_client_uuid)))
499 ldlm_destroy_export(exp);
500 tgt_client_free(exp);
502 ofd_fmd_cleanup(exp);
505 * discard grants once we're sure no more
506 * interaction with the client is possible
508 ofd_grant_discard(exp);
509 ofd_fmd_cleanup(exp);
511 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT_SHRINK) {
512 if (ofd->ofd_tot_granted_clients > 0)
513 ofd->ofd_tot_granted_clients --;
516 if (!(exp->exp_flags & OBD_OPT_FORCE))
517 ofd_grant_sanity_check(exp->exp_obd, __FUNCTION__);
519 LASSERT(list_empty(&exp->exp_filter_data.fed_mod_list));
524 * Notify all devices in server stack about recovery completion.
526 * This function calls ldo_recovery_complete() for all lower devices in the
527 * server stack so they will be prepared for normal operations.
529 * \param[in] env execution environment
530 * \param[in] ofd OFD device
532 * \retval 0 if successful
533 * \retval negative value on error
535 int ofd_postrecov(const struct lu_env *env, struct ofd_device *ofd)
537 struct lu_device *ldev = &ofd->ofd_dt_dev.dd_lu_dev;
540 CDEBUG(D_HA, "%s: recovery is over\n", ofd_name(ofd));
542 if (!ofd->ofd_skip_lfsck) {
543 struct lfsck_start_param lsp;
545 lsp.lsp_start = NULL;
546 lsp.lsp_index_valid = 0;
547 rc = lfsck_start(env, ofd->ofd_osd, &lsp);
548 if (rc != 0 && rc != -EALREADY)
549 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
553 return ldev->ld_ops->ldo_recovery_complete(env, ldev);
557 * Implementation of obd_ops::o_postrecov.
559 * This function is called from target_finish_recovery() upon recovery
562 * \param[in] obd OBD device of OFD
564 * \retval 0 if successful
565 * \retval negative value on error
567 static int ofd_obd_postrecov(struct obd_device *obd)
570 struct lu_device *ldev = obd->obd_lu_dev;
575 rc = lu_env_init(&env, LCT_DT_THREAD);
578 ofd_info_init(&env, obd->obd_self_export);
580 rc = ofd_postrecov(&env, ofd_dev(ldev));
587 * Implementation of obd_ops::o_set_info_async.
589 * This function is not called from request handler, it is only used by
590 * class_notify_sptlrpc_conf() locally by direct obd_set_info_async() call.
591 * \see ofd_set_info_hdl() for request handler function.
593 * \param[in] env execution environment
594 * \param[in] exp OBD export of OFD device
595 * \param[in] keylen length of \a key
596 * \param[in] key key name
597 * \param[in] vallen length of \a val
598 * \param[in] val the \a key value
599 * \param[in] set not used in OFD
601 * \retval 0 if successful
602 * \retval negative value on error
604 static int ofd_set_info_async(const struct lu_env *env, struct obd_export *exp,
605 __u32 keylen, void *key, __u32 vallen, void *val,
606 struct ptlrpc_request_set *set)
612 if (exp->exp_obd == NULL) {
613 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
617 if (KEY_IS(KEY_SPTLRPC_CONF)) {
618 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
620 CERROR("%s: Unsupported key %s\n",
621 exp->exp_obd->obd_name, (char*)key);
628 * Implementation of obd_ops::o_get_info.
630 * This function is not called from request handler, it is only used by
631 * direct call from nrs_orr_range_fill_physical() in ptlrpc, see LU-3239.
633 * \see ofd_get_info_hdl() for request handler function.
635 * \param[in] env execution environment
636 * \param[in] exp OBD export of OFD device
637 * \param[in] keylen length of \a key
638 * \param[in] key key name
639 * \param[out] vallen length of key value
640 * \param[out] val the key value to return
642 * \retval 0 if successful
643 * \retval negative value on error
645 static int ofd_get_info(const struct lu_env *env, struct obd_export *exp,
646 __u32 keylen, void *key, __u32 *vallen, void *val)
648 struct ofd_thread_info *info;
649 struct ofd_device *ofd;
650 struct ll_fiemap_info_key *fm_key = key;
651 struct fiemap *fiemap = val;
656 if (exp->exp_obd == NULL) {
657 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
663 if (KEY_IS(KEY_FIEMAP)) {
664 info = ofd_info_init(env, exp);
666 rc = ostid_to_fid(&info->fti_fid, &fm_key->lfik_oa.o_oi,
667 ofd->ofd_lut.lut_lsd.lsd_osd_index);
671 rc = ofd_fiemap_get(env, ofd, &info->fti_fid, fiemap);
673 CERROR("%s: not supported key %s\n", ofd_name(ofd), (char*)key);
681 * Get file system statistics of OST server.
683 * Helper function for ofd_statfs(), also used by grant code.
684 * Implements caching for statistics to avoid calling OSD device each time.
686 * \param[in] env execution environment
687 * \param[in] ofd OFD device
688 * \param[out] osfs statistic data to return
689 * \param[in] max_age maximum age for cached data
690 * \param[in] from_cache show that data was get from cache or not
692 * \retval 0 if successful
693 * \retval negative value on error
695 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
696 struct obd_statfs *osfs, __u64 max_age, int *from_cache)
701 spin_lock(&ofd->ofd_osfs_lock);
702 if (cfs_time_before_64(ofd->ofd_osfs_age, max_age) || max_age == 0) {
705 /* statfs data are too old, get up-to-date one.
706 * we must be cautious here since multiple threads might be
707 * willing to update statfs data concurrently and we must
708 * grant that cached statfs data are always consistent */
710 if (ofd->ofd_statfs_inflight == 0)
711 /* clear inflight counter if no users, although it would
712 * take a while to overflow this 64-bit counter ... */
713 ofd->ofd_osfs_inflight = 0;
714 /* notify ofd_grant_commit() that we want to track writes
715 * completed as of now */
716 ofd->ofd_statfs_inflight++;
717 /* record value of inflight counter before running statfs to
718 * compute the diff once statfs is completed */
719 unstable = ofd->ofd_osfs_inflight;
720 spin_unlock(&ofd->ofd_osfs_lock);
722 /* statfs can sleep ... hopefully not for too long since we can
723 * call it fairly often as space fills up */
724 rc = dt_statfs(env, ofd->ofd_osd, osfs);
728 spin_lock(&ofd->ofd_grant_lock);
729 spin_lock(&ofd->ofd_osfs_lock);
730 /* calculate how much space was written while we released the
732 unstable = ofd->ofd_osfs_inflight - unstable;
733 ofd->ofd_osfs_unstable = 0;
735 /* some writes committed while we were running statfs
736 * w/o the ofd_osfs_lock. Those ones got added to
737 * the cached statfs data that we are about to crunch.
738 * Take them into account in the new statfs data */
739 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
740 unstable >> ofd->ofd_blockbits);
741 /* However, we don't really know if those writes got
742 * accounted in the statfs call, so tell
743 * ofd_grant_space_left() there is some uncertainty
744 * on the accounting of those writes.
745 * The purpose is to prevent spurious error messages in
746 * ofd_grant_space_left() since those writes might be
747 * accounted twice. */
748 ofd->ofd_osfs_unstable += unstable;
750 /* similarly, there is some uncertainty on write requests
751 * between prepare & commit */
752 ofd->ofd_osfs_unstable += ofd->ofd_tot_pending;
753 spin_unlock(&ofd->ofd_grant_lock);
755 /* finally udpate cached statfs data */
756 ofd->ofd_osfs = *osfs;
757 ofd->ofd_osfs_age = cfs_time_current_64();
759 ofd->ofd_statfs_inflight--; /* stop tracking */
760 if (ofd->ofd_statfs_inflight == 0)
761 ofd->ofd_osfs_inflight = 0;
762 spin_unlock(&ofd->ofd_osfs_lock);
767 /* use cached statfs data */
768 *osfs = ofd->ofd_osfs;
769 spin_unlock(&ofd->ofd_osfs_lock);
780 * Implementation of obd_ops::o_statfs.
782 * This function returns information about a storage file system.
783 * It is called from several places by using the OBD API as well as
784 * by direct call, e.g. from request handler.
786 * \see ofd_statfs_hdl() for request handler function.
788 * Report also the state of the OST to the caller in osfs->os_state
789 * (OS_STATE_READONLY, OS_STATE_DEGRADED).
791 * \param[in] env execution environment
792 * \param[in] exp OBD export of OFD device
793 * \param[out] osfs statistic data to return
794 * \param[in] max_age maximum age for cached data
795 * \param[in] flags not used in OFD
797 * \retval 0 if successful
798 * \retval negative value on error
800 int ofd_statfs(const struct lu_env *env, struct obd_export *exp,
801 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
803 struct obd_device *obd = class_exp2obd(exp);
804 struct ofd_device *ofd = ofd_exp(exp);
809 rc = ofd_statfs_internal(env, ofd, osfs, max_age, NULL);
813 /* at least try to account for cached pages. its still racy and
814 * might be under-reporting if clients haven't announced their
815 * caches with brw recently */
817 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
818 " pending "LPU64" free "LPU64" avail "LPU64"\n",
819 ofd->ofd_tot_dirty, ofd->ofd_tot_granted, ofd->ofd_tot_pending,
820 osfs->os_bfree << ofd->ofd_blockbits,
821 osfs->os_bavail << ofd->ofd_blockbits);
823 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
824 ((ofd->ofd_tot_dirty + ofd->ofd_tot_pending +
825 osfs->os_bsize - 1) >> ofd->ofd_blockbits));
827 /* The QoS code on the MDS does not care about space reserved for
828 * precreate, so take it out. */
829 if (exp_connect_flags(exp) & OBD_CONNECT_MDS) {
830 struct filter_export_data *fed;
832 fed = &obd->obd_self_export->exp_filter_data;
833 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
834 fed->fed_grant >> ofd->ofd_blockbits);
837 ofd_grant_sanity_check(obd, __FUNCTION__);
838 CDEBUG(D_CACHE, LPU64" blocks: "LPU64" free, "LPU64" avail; "
839 LPU64" objects: "LPU64" free; state %x\n",
840 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
841 osfs->os_files, osfs->os_ffree, osfs->os_state);
843 if (OBD_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOINO,
844 ofd->ofd_lut.lut_lsd.lsd_osd_index))
847 /* OS_STATE_READONLY can be set by OSD already */
848 if (ofd->ofd_raid_degraded)
849 osfs->os_state |= OS_STATE_DEGRADED;
851 if (obd->obd_self_export != exp && ofd_grant_compat(exp, ofd)) {
852 /* clients which don't support OBD_CONNECT_GRANT_PARAM
853 * should not see a block size > page size, otherwise
854 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
855 * block size which is the biggest block size known to work
856 * with all client's page size. */
857 osfs->os_blocks <<= ofd->ofd_blockbits - COMPAT_BSIZE_SHIFT;
858 osfs->os_bfree <<= ofd->ofd_blockbits - COMPAT_BSIZE_SHIFT;
859 osfs->os_bavail <<= ofd->ofd_blockbits - COMPAT_BSIZE_SHIFT;
860 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
863 if (OBD_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOSPC,
864 ofd->ofd_lut.lut_lsd.lsd_osd_index))
865 osfs->os_bfree = osfs->os_bavail = 2;
873 * Implementation of obd_ops::o_setattr.
875 * This function is only used by ECHO client when it is run on top of OFD,
876 * \see ofd_setattr_hdl() for request handler function.
878 * \param[in] env execution environment
879 * \param[in] exp OBD export of OFD device
880 * \param[in] oa setattr parameters
882 * \retval 0 if successful
883 * \retval negative value on error
885 static int ofd_echo_setattr(const struct lu_env *env, struct obd_export *exp,
888 struct ofd_thread_info *info;
889 struct ofd_device *ofd = ofd_exp(exp);
890 struct ldlm_namespace *ns = ofd->ofd_namespace;
891 struct ldlm_resource *res;
892 struct ofd_object *fo;
893 struct lu_fid *fid = &oa->o_oi.oi_fid;
894 struct filter_fid *ff = NULL;
899 info = ofd_info_init(env, exp);
901 ost_fid_build_resid(fid, &info->fti_resid);
903 /* This would be very bad - accidentally truncating a file when
904 * changing the time or similar - bug 12203. */
905 if (oa->o_valid & OBD_MD_FLSIZE) {
906 static char mdsinum[48];
908 if (oa->o_valid & OBD_MD_FLFID)
909 snprintf(mdsinum, sizeof(mdsinum) - 1,
910 "of parent "DFID, oa->o_parent_seq,
911 oa->o_parent_oid, 0);
915 CERROR("%s: setattr from %s trying to truncate object "DFID
916 " %s\n", ofd_name(ofd), obd_export_nid2str(exp),
918 GOTO(out, rc = -EPERM);
921 fo = ofd_object_find_exists(env, ofd, fid);
923 CERROR("%s: can't find object "DFID"\n",
924 ofd_name(ofd), PFID(fid));
925 GOTO(out, rc = PTR_ERR(fo));
928 la_from_obdo(&info->fti_attr, oa, oa->o_valid);
929 info->fti_attr.la_valid &= ~LA_TYPE;
931 if (oa->o_valid & OBD_MD_FLFID) {
932 ff = &info->fti_mds_fid;
933 ofd_prepare_fidea(ff, oa);
936 /* setting objects attributes (including owner/group) */
937 rc = ofd_attr_set(env, fo, &info->fti_attr, ff);
939 GOTO(out_unlock, rc);
941 ofd_counter_incr(exp, LPROC_OFD_STATS_SETATTR, NULL, 1);
944 ofd_object_put(env, fo);
947 /* we do not call this before to avoid lu_object_find() in
948 * ->lvbo_update() holding another reference on the object.
949 * otherwise concurrent destroy can make the object unavailable
950 * for 2nd lu_object_find() waiting for the first reference
951 * to go... deadlock! */
952 res = ldlm_resource_get(ns, NULL, &info->fti_resid, LDLM_EXTENT, 0);
954 ldlm_res_lvbo_update(res, NULL, 0);
955 ldlm_resource_putref(res);
963 * Destroy OFD object by its FID.
965 * Supplemental function to destroy object by FID, it is used by request
966 * handler and by ofd_echo_destroy() below to find object by FID, lock it
967 * and call ofd_object_destroy() finally.
969 * \param[in] env execution environment
970 * \param[in] ofd OFD device
971 * \param[in] fid FID of object
972 * \param[in] orphan set if object being destroyed is an orphan
974 * \retval 0 if successful
975 * \retval negative value on error
977 int ofd_destroy_by_fid(const struct lu_env *env, struct ofd_device *ofd,
978 const struct lu_fid *fid, int orphan)
980 struct ofd_thread_info *info = ofd_info(env);
981 struct lustre_handle lockh;
982 union ldlm_policy_data policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
983 struct ofd_object *fo;
984 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
989 fo = ofd_object_find_exists(env, ofd, fid);
993 /* Tell the clients that the object is gone now and that they should
994 * throw away any cached pages. */
995 ost_fid_build_resid(fid, &info->fti_resid);
996 rc = ldlm_cli_enqueue_local(ofd->ofd_namespace, &info->fti_resid,
997 LDLM_EXTENT, &policy, LCK_PW, &flags,
998 ldlm_blocking_ast, ldlm_completion_ast,
999 NULL, NULL, 0, LVB_T_NONE, NULL, &lockh);
1001 /* We only care about the side-effects, just drop the lock. */
1003 ldlm_lock_decref(&lockh, LCK_PW);
1005 LASSERT(fo != NULL);
1007 rc = ofd_object_destroy(env, fo, orphan);
1010 ofd_object_put(env, fo);
1015 * Implementation of obd_ops::o_destroy.
1017 * This function is only used by ECHO client when it is run on top of OFD,
1018 * \see ofd_destroy_hdl() for request handler function.
1020 * \param[in] env execution environment
1021 * \param[in] exp OBD export of OFD device
1022 * \param[in] oa obdo structure with FID
1024 * Note: this is OBD API method which is common API for server OBDs and
1025 * client OBDs. Thus some parameters used in client OBDs may not be used
1026 * on server OBDs and vice versa.
1028 * \retval 0 if successful
1029 * \retval negative value on error
1031 static int ofd_echo_destroy(const struct lu_env *env, struct obd_export *exp,
1034 struct ofd_device *ofd = ofd_exp(exp);
1035 struct lu_fid *fid = &oa->o_oi.oi_fid;
1040 ofd_info_init(env, exp);
1042 CDEBUG(D_HA, "%s: Destroy object "DFID"\n", ofd_name(ofd), PFID(fid));
1044 rc = ofd_destroy_by_fid(env, ofd, fid, 0);
1045 if (rc == -ENOENT) {
1046 CDEBUG(D_INODE, "%s: destroying non-existent object "DFID"\n",
1047 ofd_name(ofd), PFID(fid));
1049 } else if (rc != 0) {
1050 CERROR("%s: error destroying object "DFID": %d\n",
1051 ofd_name(ofd), PFID(fid), rc);
1060 * Implementation of obd_ops::o_create.
1062 * This function is only used by ECHO client when it is run on top of OFD
1063 * and just creates an object.
1064 * \see ofd_create_hdl() for request handler function.
1066 * \param[in] env execution environment
1067 * \param[in] exp OBD export of OFD device
1068 * \param[in] oa obdo structure with FID sequence to use
1070 * Note: this is OBD API method which is common API for server OBDs and
1071 * client OBDs. Thus some parameters used in client OBDs may not be used
1072 * on server OBDs and vice versa.
1074 * \retval 0 if successful
1075 * \retval negative value on error
1077 static int ofd_echo_create(const struct lu_env *env, struct obd_export *exp,
1080 struct ofd_device *ofd = ofd_exp(exp);
1081 struct ofd_thread_info *info;
1082 u64 seq = ostid_seq(&oa->o_oi);
1083 struct ofd_seq *oseq;
1084 int rc = 0, diff = 1;
1091 info = ofd_info_init(env, exp);
1093 LASSERT(seq == FID_SEQ_ECHO);
1094 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
1096 CDEBUG(D_INFO, "ofd_create("DOSTID")\n", POSTID(&oa->o_oi));
1098 down_read(&ofd->ofd_lastid_rwsem);
1099 /* Currently, for safe, we do not distinguish which LAST_ID is broken,
1100 * we may do that in the future.
1101 * Return -ENOSPC until the LAST_ID rebuilt. */
1102 if (unlikely(ofd->ofd_lastid_rebuilding))
1103 GOTO(out_sem, rc = -ENOSPC);
1105 oseq = ofd_seq_load(env, ofd, seq);
1107 CERROR("%s: Can't find FID Sequence "LPX64": rc = %ld\n",
1108 ofd_name(ofd), seq, PTR_ERR(oseq));
1109 GOTO(out_sem, rc = -EINVAL);
1112 mutex_lock(&oseq->os_create_lock);
1113 granted = ofd_grant_create(env, ofd_obd(ofd)->obd_self_export, &diff);
1117 CDEBUG(D_HA, "%s: failed to acquire grant space for "
1118 "precreate (%d): rc = %d\n", ofd_name(ofd), diff, rc);
1123 next_id = ofd_seq_last_oid(oseq) + 1;
1124 count = ofd_precreate_batch(ofd, diff);
1126 rc = ofd_precreate_objects(env, ofd, next_id, oseq, count, 0);
1128 CERROR("%s: unable to precreate: rc = %d\n",
1131 ostid_set_id(&oa->o_oi, ofd_seq_last_oid(oseq));
1132 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
1136 ofd_grant_commit(ofd_obd(ofd)->obd_self_export, granted, rc);
1138 mutex_unlock(&oseq->os_create_lock);
1139 ofd_seq_put(env, oseq);
1142 up_read(&ofd->ofd_lastid_rwsem);
1147 * Implementation of obd_ops::o_getattr.
1149 * This function is only used by ECHO client when it is run on top of OFD
1150 * and returns attributes of object.
1151 * \see ofd_getattr_hdl() for request handler function.
1153 * \param[in] env execution environment
1154 * \param[in] exp OBD export of OFD device
1155 * \param[in,out] oa contains FID of object to get attributes from and
1156 * is used to return attributes back
1158 * \retval 0 if successful
1159 * \retval negative value on error
1161 static int ofd_echo_getattr(const struct lu_env *env, struct obd_export *exp,
1164 struct ofd_device *ofd = ofd_exp(exp);
1165 struct ofd_thread_info *info;
1166 struct lu_fid *fid = &oa->o_oi.oi_fid;
1167 struct ofd_object *fo;
1172 info = ofd_info_init(env, exp);
1174 fo = ofd_object_find_exists(env, ofd, fid);
1176 GOTO(out, rc = PTR_ERR(fo));
1178 LASSERT(fo != NULL);
1179 rc = ofd_attr_get(env, fo, &info->fti_attr);
1180 oa->o_valid = OBD_MD_FLID;
1184 obdo_from_la(oa, &info->fti_attr,
1185 OFD_VALID_FLAGS | LA_UID | LA_GID);
1187 /* Store object version in reply */
1188 curr_version = dt_version_get(env, ofd_object_child(fo));
1189 if ((__s64)curr_version != -EOPNOTSUPP) {
1190 oa->o_valid |= OBD_MD_FLDATAVERSION;
1191 oa->o_data_version = curr_version;
1195 ofd_object_put(env, fo);
1201 * Get object version for OBD_IOC_GET_OBJ_VERSION ioctl.
1203 * This is supplemental function for ofd_iocontrol() to return object
1204 * version for lctl tool.
1206 * \param[in] env execution environment
1207 * \param[in] ofd OFD device
1208 * \param[out] karg ioctl data
1210 * \retval 0 if successful
1211 * \retval negative value on error
1213 static int ofd_ioc_get_obj_version(const struct lu_env *env,
1214 struct ofd_device *ofd, void *karg)
1216 struct obd_ioctl_data *data = karg;
1218 struct ofd_object *fo;
1219 dt_obj_version_t version;
1224 if (data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
1225 GOTO(out, rc = -EINVAL);
1227 if (data->ioc_inlbuf1 != NULL && data->ioc_inllen1 == sizeof(fid)) {
1228 fid = *(struct lu_fid *)data->ioc_inlbuf1;
1229 } else if (data->ioc_inlbuf3 != NULL &&
1230 data->ioc_inllen3 == sizeof(__u64) &&
1231 data->ioc_inlbuf4 != NULL &&
1232 data->ioc_inllen4 == sizeof(__u64)) {
1233 struct ost_id ostid;
1235 ostid_set_seq(&ostid, *(__u64 *)data->ioc_inlbuf4);
1236 ostid_set_id(&ostid, *(__u64 *)data->ioc_inlbuf3);
1237 rc = ostid_to_fid(&fid, &ostid,
1238 ofd->ofd_lut.lut_lsd.lsd_osd_index);
1242 GOTO(out, rc = -EINVAL);
1245 if (!fid_is_sane(&fid))
1246 GOTO(out, rc = -EINVAL);
1248 fo = ofd_object_find(env, ofd, &fid);
1250 GOTO(out, rc = PTR_ERR(fo));
1252 if (!ofd_object_exists(fo))
1253 GOTO(out_fo, rc = -ENOENT);
1255 if (lu_object_remote(&fo->ofo_obj.do_lu))
1256 GOTO(out_fo, rc = -EREMOTE);
1258 version = dt_version_get(env, ofd_object_child(fo));
1260 GOTO(out_fo, rc = -EIO);
1262 *(dt_obj_version_t *)data->ioc_inlbuf2 = version;
1266 ofd_object_put(env, fo);
1272 * Implementation of obd_ops::o_iocontrol.
1274 * This is OFD ioctl handling function which is primary interface for
1275 * Lustre tools like lfs, lctl and lfsck.
1277 * \param[in] cmd ioctl command
1278 * \param[in] exp OBD export of OFD
1279 * \param[in] len not used
1280 * \param[in,out] karg buffer with data
1281 * \param[in] uarg not used
1283 * \retval 0 if successful
1284 * \retval negative value on error
1286 static int ofd_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1287 void *karg, void *uarg)
1290 struct ofd_device *ofd = ofd_exp(exp);
1291 struct obd_device *obd = ofd_obd(ofd);
1296 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
1297 rc = lu_env_init(&env, LCT_DT_THREAD);
1302 case OBD_IOC_ABORT_RECOVERY:
1303 CERROR("%s: aborting recovery\n", obd->obd_name);
1304 obd->obd_force_abort_recovery = 1;
1305 target_stop_recovery_thread(obd);
1308 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
1309 rc = dt_sync(&env, ofd->ofd_osd);
1311 case OBD_IOC_SET_READONLY:
1312 rc = dt_sync(&env, ofd->ofd_osd);
1314 rc = dt_ro(&env, ofd->ofd_osd);
1316 case OBD_IOC_START_LFSCK: {
1317 struct obd_ioctl_data *data = karg;
1318 struct lfsck_start_param lsp;
1320 if (unlikely(data == NULL)) {
1325 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
1326 lsp.lsp_index_valid = 0;
1327 rc = lfsck_start(&env, ofd->ofd_osd, &lsp);
1330 case OBD_IOC_STOP_LFSCK: {
1331 struct obd_ioctl_data *data = karg;
1332 struct lfsck_stop stop;
1334 stop.ls_status = LS_STOPPED;
1335 /* Old lfsck utils may pass NULL @stop. */
1336 if (data->ioc_inlbuf1 == NULL)
1340 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
1342 rc = lfsck_stop(&env, ofd->ofd_osd, &stop);
1345 case OBD_IOC_GET_OBJ_VERSION:
1346 rc = ofd_ioc_get_obj_version(&env, ofd, karg);
1349 CERROR("%s: not supported cmd = %#x\n", obd->obd_name, cmd);
1358 * Implementation of obd_ops::o_precleanup.
1360 * This function stops device activity before shutting it down. It is called
1361 * from a cleanup function upon forceful device cleanup. For OFD there are no
1362 * special actions, it just invokes target_recovery_cleanup().
1364 * \param[in] obd OBD device of OFD
1368 static int ofd_precleanup(struct obd_device *obd)
1371 target_cleanup_recovery(obd);
1376 * Implementation of obd_ops::o_ping.
1378 * This is OFD-specific part of OBD_PING request handling.
1379 * It controls Filter Modification Data (FMD) expiration each time PING is
1382 * \see ofd_fmd_expire() and ofd_fmd.c for details
1384 * \param[in] env execution environment
1385 * \param[in] exp OBD export of client
1389 static int ofd_ping(const struct lu_env *env, struct obd_export *exp)
1391 ofd_fmd_expire(exp);
1396 * Implementation of obd_ops::o_health_check.
1398 * This function checks the OFD device health - ability to respond on
1399 * incoming requests. There are two health_check methods:
1400 * - get statfs from the OSD. It checks just responsiveness of
1402 * - do write attempt on bottom device to check it is fully operational and
1403 * is not stuck. This is expensive method and requires special configuration
1404 * option --enable-health-write while building Lustre, it is turned off
1407 * \param[in] nul not used
1408 * \param[in] obd OBD device of OFD
1410 * \retval 0 if successful
1411 * \retval negative value in case of error
1413 static int ofd_health_check(const struct lu_env *nul, struct obd_device *obd)
1415 struct ofd_device *ofd = ofd_dev(obd->obd_lu_dev);
1416 struct ofd_thread_info *info;
1418 #ifdef USE_HEALTH_CHECK_WRITE
1423 /* obd_proc_read_health pass NULL env, we need real one */
1424 rc = lu_env_init(&env, LCT_DT_THREAD);
1428 info = ofd_info_init(&env, NULL);
1429 rc = dt_statfs(&env, ofd->ofd_osd, &info->fti_u.osfs);
1433 if (info->fti_u.osfs.os_state == OS_STATE_READONLY)
1434 GOTO(out, rc = -EROFS);
1436 #ifdef USE_HEALTH_CHECK_WRITE
1437 OBD_ALLOC(info->fti_buf.lb_buf, PAGE_CACHE_SIZE);
1438 if (info->fti_buf.lb_buf == NULL)
1439 GOTO(out, rc = -ENOMEM);
1441 info->fti_buf.lb_len = PAGE_CACHE_SIZE;
1444 th = dt_trans_create(&env, ofd->ofd_osd);
1446 GOTO(out, rc = PTR_ERR(th));
1448 rc = dt_declare_record_write(&env, ofd->ofd_health_check_file,
1449 &info->fti_buf, info->fti_off, th);
1451 th->th_sync = 1; /* sync IO is needed */
1452 rc = dt_trans_start_local(&env, ofd->ofd_osd, th);
1454 rc = dt_record_write(&env, ofd->ofd_health_check_file,
1455 &info->fti_buf, &info->fti_off,
1458 dt_trans_stop(&env, ofd->ofd_osd, th);
1460 OBD_FREE(info->fti_buf.lb_buf, PAGE_CACHE_SIZE);
1462 CDEBUG(D_INFO, "write 1 page synchronously for checking io rc %d\n",rc);
1469 struct obd_ops ofd_obd_ops = {
1470 .o_owner = THIS_MODULE,
1471 .o_connect = ofd_obd_connect,
1472 .o_reconnect = ofd_obd_reconnect,
1473 .o_disconnect = ofd_obd_disconnect,
1474 .o_create = ofd_echo_create,
1475 .o_statfs = ofd_statfs,
1476 .o_setattr = ofd_echo_setattr,
1477 .o_preprw = ofd_preprw,
1478 .o_commitrw = ofd_commitrw,
1479 .o_destroy = ofd_echo_destroy,
1480 .o_init_export = ofd_init_export,
1481 .o_destroy_export = ofd_destroy_export,
1482 .o_postrecov = ofd_obd_postrecov,
1483 .o_getattr = ofd_echo_getattr,
1484 .o_iocontrol = ofd_iocontrol,
1485 .o_precleanup = ofd_precleanup,
1487 .o_health_check = ofd_health_check,
1488 .o_set_info_async = ofd_set_info_async,
1489 .o_get_info = ofd_get_info,