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, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/ofd/ofd_obd.c
33 * This file contains OBD API methods for OBD Filter Device (OFD) which are
34 * used for export handling, configuration purposes and recovery.
35 * Several methods are used by ECHO client only since it still uses OBD API.
36 * Such methods have _echo_ prefix in name.
38 * Author: Andreas Dilger <andreas.dilger@intel.com>
39 * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
40 * Author: Mikhail Pershin <mike.pershin@intel.com>
43 #define DEBUG_SUBSYSTEM S_FILTER
45 #include "ofd_internal.h"
46 #include <obd_cksum.h>
47 #include <uapi/linux/lustre/lustre_ioctl.h>
48 #include <lustre_quota.h>
49 #include <lustre_lfsck.h>
50 #include <lustre_nodemap.h>
51 #include <obd_support.h>
54 * Initialize OFD per-export statistics.
56 * This function sets up procfs entries for various OFD export counters. These
57 * counters are for per-client statistics tracked on the server.
59 * \param[in] ofd OFD device
60 * \param[in] exp OBD export
61 * \param[in] client_nid NID of client
63 * \retval 0 if successful
64 * \retval negative value on error
66 static int ofd_export_stats_init(struct ofd_device *ofd,
67 struct obd_export *exp,
68 struct lnet_nid *client_nid)
70 struct obd_device *obd = ofd_obd(ofd);
71 struct nid_stat *stats;
76 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
77 /* Self-export gets no proc entry */
80 rc = lprocfs_exp_setup(exp, client_nid);
82 /* Mask error for already created /proc entries */
83 RETURN(rc == -EALREADY ? 0 : rc);
85 stats = exp->exp_nid_stats;
86 stats->nid_stats = lprocfs_stats_alloc(LPROC_OFD_STATS_LAST,
87 LPROCFS_STATS_FLAG_NOPERCPU);
88 if (!stats->nid_stats)
91 ofd_stats_counter_init(stats->nid_stats, 0, LPROCFS_CNTR_HISTOGRAM);
93 rc = lprocfs_stats_register(stats->nid_proc, "stats", stats->nid_stats);
95 lprocfs_stats_free(&stats->nid_stats);
99 rc = lprocfs_nid_ldlm_stats_init(stats);
108 * Match client and OST server connection feature flags.
110 * Compute the compatibility flags for a connection request based on
111 * features mutually supported by client and server.
113 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
114 * must not be updated here, otherwise a partially initialized value may
115 * be exposed. After the connection request is successfully processed,
116 * the top-level tgt_connect() request handler atomically updates the export
117 * connect flags from the obd_connect_data::ocd_connect_flags field of the
118 * reply. \see tgt_connect().
120 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
121 * full struct obd_connect_data. So care must be taken when accessing fields
122 * that are not present in struct obd_connect_data_v1. See LU-16.
124 * \param[in] env execution environment
125 * \param[in] exp the obd_export associated with this
127 * \param[in] data stores data for this connect request
128 * \param[in] new_connection is this connection new or not
130 * \retval 0 if success
131 * \retval -EPROTO client and server feature requirements are
133 * \retval -EBADF OST index in connect request doesn't match
136 static int ofd_parse_connect_data(const struct lu_env *env,
137 struct obd_export *exp,
138 struct obd_connect_data *data,
141 struct ofd_device *ofd = ofd_exp(exp);
142 struct filter_export_data *fed = &exp->exp_filter_data;
148 "%s: cli %s/%p ocd_connect_flags: %#llx ocd_version: %x ocd_grant: %d ocd_index: %u ocd_group %u\n",
149 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
150 data->ocd_connect_flags, data->ocd_version,
151 data->ocd_grant, data->ocd_index, data->ocd_group);
153 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
154 CWARN("!!! This export (nid %s) used object group %d earlier; now it's trying to use group %d! This could be a bug in the MDS. Please report to https://jira.whamcloud.com/\n",
155 obd_export_nid2str(exp), fed->fed_group,
159 fed->fed_group = data->ocd_group;
161 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
163 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
164 data->ocd_connect_flags2 &= OST_CONNECT_SUPPORTED2;
166 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
167 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
168 CDEBUG(D_HA, "%s: Received MDS connection for group %u\n",
169 exp->exp_obd->obd_name, data->ocd_group);
170 else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
173 /* Determine optimal brw size before calculating grant */
174 if (CFS_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
175 data->ocd_brw_size = 65536;
176 } else if (OCD_HAS_FLAG(data, BRW_SIZE)) {
177 if (data->ocd_brw_size > ofd->ofd_brw_size)
178 data->ocd_brw_size = ofd->ofd_brw_size;
179 if (data->ocd_brw_size == 0) {
180 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx ocd_version: %x ocd_grant: %d ocd_index: %u ocd_brw_size is unexpectedly zero, network data corruption? Refusing connection of this client\n",
181 exp->exp_obd->obd_name,
182 exp->exp_client_uuid.uuid,
183 exp, data->ocd_connect_flags, data->ocd_version,
184 data->ocd_grant, data->ocd_index);
189 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
190 struct dt_device_param *ddp = &ofd->ofd_lut.lut_dt_conf;
192 /* client is reporting its page size, for future use */
193 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
194 data->ocd_grant_blkbits = ofd->ofd_lut.lut_tgd.tgd_blockbits;
196 * ddp_inodespace may not be power-of-two value, eg. for ldiskfs
197 * it's LDISKFS_DIR_REC_LEN(20) = 28.
199 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
200 /* ocd_grant_tax_kb is in 1K byte blocks */
201 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
202 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
206 * Save connect_data we have so far because tgt_grant_connect()
207 * uses it to calculate grant, and we want to save the client
208 * version before it is overwritten by LUSTRE_VERSION_CODE.
210 exp->exp_connect_data = *data;
211 if (OCD_HAS_FLAG(data, GRANT))
212 tgt_grant_connect(env, exp, data, new_connection);
214 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
215 struct lr_server_data *lsd = &ofd->ofd_lut.lut_lsd;
216 int index = lsd->lsd_osd_index;
218 if (index != data->ocd_index) {
219 LCONSOLE_ERROR_MSG(0x136,
220 "Connection from %s to index %u doesn't match actual OST index %u in last_rcvd file, bad configuration?\n",
221 obd_export_nid2str(exp), index,
225 if (!(lsd->lsd_feature_compat & OBD_COMPAT_OST)) {
226 /* this will only happen on the first connect */
227 lsd->lsd_feature_compat |= OBD_COMPAT_OST;
229 * sync is not needed here as tgt_client_new will
230 * set exp_need_sync flag
232 tgt_server_data_update(env, &ofd->ofd_lut, 0);
236 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
237 __u32 cksum_types = data->ocd_cksum_types;
239 tgt_mask_cksum_types(&ofd->ofd_lut, &data->ocd_cksum_types);
241 if (unlikely(data->ocd_cksum_types == 0)) {
242 CERROR("%s: Connect with checksum support but no ocd_cksum_types is set\n",
243 exp->exp_obd->obd_name);
248 "%s: cli %s supports cksum type %x, return %x\n",
249 exp->exp_obd->obd_name, obd_export_nid2str(exp),
250 cksum_types, data->ocd_cksum_types);
253 * This client does not support OBD_CONNECT_CKSUM.
254 * Report failure to negotiate checksum at connect
257 "%s: cli %s does not support OBD_CONNECT_CKSUM\n",
258 exp->exp_obd->obd_name, obd_export_nid2str(exp));
261 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
262 data->ocd_maxbytes = ofd->ofd_lut.lut_dt_conf.ddp_maxbytes;
264 data->ocd_version = LUSTRE_VERSION_CODE;
266 if (OCD_HAS_FLAG(data, PINGLESS)) {
267 if (ptlrpc_pinger_suppress_pings()) {
268 spin_lock(&exp->exp_obd->obd_dev_lock);
269 list_del_init(&exp->exp_obd_chain_timed);
270 spin_unlock(&exp->exp_obd->obd_dev_lock);
272 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
276 if (!ofd->ofd_lut.lut_dt_conf.ddp_has_lseek_data_hole)
277 data->ocd_connect_flags2 &= ~OBD_CONNECT2_LSEEK;
283 * Re-initialize export upon client reconnection.
285 * This function parses connection data from reconnect and resets
288 * \param[in] env execution environment
289 * \param[in] exp OBD export
290 * \param[in] obd OFD device
291 * \param[in] cluuid NID of client
292 * \param[in] data connection data from request
293 * \param[in] localdata client NID
295 * \retval 0 if successful
296 * \retval negative value on error
298 static int ofd_obd_reconnect(const struct lu_env *env, struct obd_export *exp,
299 struct obd_device *obd, struct obd_uuid *cluuid,
300 struct obd_connect_data *data,
303 struct ofd_device *ofd;
304 struct lnet_nid *client_nid = localdata;
309 if (!exp || !obd || !cluuid)
312 rc = nodemap_add_member(client_nid, exp);
313 if (rc != 0 && rc != -EEXIST)
316 ofd = ofd_dev(obd->obd_lu_dev);
318 rc = ofd_parse_connect_data(env, exp, data, false);
320 ofd_export_stats_init(ofd, exp, client_nid);
322 nodemap_del_member(exp);
328 * Initialize new client connection.
330 * This function handles new connection to the OFD. The new export is
331 * created (in context of class_connect()) and persistent client data is
332 * initialized on storage.
334 * \param[in] env execution environment
335 * \param[out] _exp stores pointer to new export
336 * \param[in] obd OFD device
337 * \param[in] cluuid client UUID
338 * \param[in] data connection data from request
339 * \param[in] localdata client NID
341 * \retval 0 if successful
342 * \retval negative value on error
344 static int ofd_obd_connect(const struct lu_env *env, struct obd_export **_exp,
345 struct obd_device *obd, struct obd_uuid *cluuid,
346 struct obd_connect_data *data, void *localdata)
348 struct obd_export *exp;
349 struct ofd_device *ofd;
350 struct lustre_handle conn = { 0 };
351 struct lnet_nid *client_nid = localdata;
356 if (!_exp || !obd || !cluuid)
359 ofd = ofd_dev(obd->obd_lu_dev);
361 rc = class_connect(&conn, obd, cluuid);
365 exp = class_conn2export(&conn);
366 LASSERT(exp != NULL);
369 rc = nodemap_add_member(client_nid, exp);
370 if (rc != 0 && rc != -EEXIST)
374 "%s: cannot find nodemap for client %s: nid is null\n",
375 obd->obd_name, cluuid->uuid);
378 rc = ofd_parse_connect_data(env, exp, data, true);
382 if (obd->obd_replayable) {
383 struct tg_export_data *ted = &exp->exp_target_data;
385 memcpy(ted->ted_lcd->lcd_uuid, cluuid,
386 sizeof(ted->ted_lcd->lcd_uuid));
387 rc = tgt_client_new(env, exp);
390 ofd_export_stats_init(ofd, exp, client_nid);
393 CDEBUG(D_HA, "%s: get connection from MDS %d\n", obd->obd_name,
394 data ? data->ocd_group : -1);
398 class_disconnect(exp);
399 nodemap_del_member(exp);
408 * Disconnect a connected client.
410 * This function terminates the client connection. The client export is
411 * disconnected (cleaned up) and client data on persistent storage is removed.
413 * \param[in] exp OBD export
415 * \retval 0 if successful
416 * \retval negative value on error
418 int ofd_obd_disconnect(struct obd_export *exp)
420 struct ofd_device *ofd = ofd_exp(exp);
427 class_export_get(exp);
429 rc = server_disconnect_export(exp);
431 tgt_grant_discard(exp);
433 if (!(exp->exp_flags & OBD_OPT_FORCE))
434 tgt_grant_sanity_check(ofd_obd(ofd), __func__);
436 /* Do not erase record for recoverable client. */
437 if (exp->exp_obd->obd_replayable &&
438 (!exp->exp_obd->obd_fail || exp->exp_failed)) {
439 rc = lu_env_init(&env, LCT_DT_THREAD);
443 tgt_client_del(&env, exp);
447 nodemap_del_member(exp);
448 class_export_put(exp);
453 * Implementation of obd_ops::o_init_export.
455 * This function is called from class_new_export() and initializes
456 * the OFD-specific data for new export.
458 * \param[in] exp OBD export
460 * \retval 0 if successful
461 * \retval negative value on error
463 static int ofd_init_export(struct obd_export *exp)
467 atomic_set(&exp->exp_filter_data.fed_soft_sync_count, 0);
468 spin_lock(&exp->exp_lock);
469 exp->exp_connecting = 1;
470 spin_unlock(&exp->exp_lock);
472 /* self-export doesn't need client data and ldlm initialization */
473 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
474 &exp->exp_client_uuid)))
477 rc = tgt_client_alloc(exp);
479 ldlm_init_export(exp);
481 CERROR("%s: Can't initialize export: rc %d\n",
482 exp->exp_obd->obd_name, rc);
487 * Implementation of obd_ops::o_destroy_export.
489 * This function is called from class_export_destroy() to cleanup
490 * the OFD-specific data for export being destroyed.
492 * \param[in] exp OBD export
494 * \retval 0 if successful
495 * \retval negative value on error
497 static int ofd_destroy_export(struct obd_export *exp)
499 struct ofd_device *ofd = ofd_exp(exp);
501 if (exp->exp_target_data.ted_pending)
502 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
503 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
504 exp, exp->exp_target_data.ted_pending);
506 target_destroy_export(exp);
508 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
509 &exp->exp_client_uuid)))
512 ldlm_destroy_export(exp);
513 tgt_client_free(exp);
516 * discard grants once we're sure no more
517 * interaction with the client is possible
519 tgt_grant_discard(exp);
521 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
522 ofd->ofd_lut.lut_tgd.tgd_tot_granted_clients--;
524 if (!(exp->exp_flags & OBD_OPT_FORCE))
525 tgt_grant_sanity_check(exp->exp_obd, __func__);
531 * Notify all devices in server stack about recovery completion.
533 * This function calls ldo_recovery_complete() for all lower devices in the
534 * server stack so they will be prepared for normal operations.
536 * \param[in] env execution environment
537 * \param[in] ofd OFD device
539 * \retval 0 if successful
540 * \retval negative value on error
542 int ofd_postrecov(const struct lu_env *env, struct ofd_device *ofd)
544 struct lu_device *ldev = &ofd->ofd_dt_dev.dd_lu_dev;
547 CDEBUG(D_HA, "%s: recovery is over\n", ofd_name(ofd));
549 if (!ofd->ofd_skip_lfsck && !ofd->ofd_osd->dd_rdonly) {
550 struct lfsck_start_param lsp;
552 lsp.lsp_start = NULL;
553 lsp.lsp_index_valid = 0;
554 rc = lfsck_start(env, ofd->ofd_osd, &lsp);
555 if (rc != 0 && rc != -EALREADY)
556 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
560 return ldev->ld_ops->ldo_recovery_complete(env, ldev);
564 * Implementation of obd_ops::o_postrecov.
566 * This function is called from target_finish_recovery() upon recovery
569 * \param[in] obd OBD device of OFD
571 * \retval 0 if successful
572 * \retval negative value on error
574 static int ofd_obd_postrecov(struct obd_device *obd)
577 struct lu_device *ldev = obd->obd_lu_dev;
582 rc = lu_env_init(&env, LCT_DT_THREAD);
585 ofd_info_init(&env, obd->obd_self_export);
587 rc = ofd_postrecov(&env, ofd_dev(ldev));
594 * Implementation of obd_ops::o_set_info_async.
596 * This function is not called from request handler, it is only used by
597 * class_notify_sptlrpc_conf() locally by direct obd_set_info_async() call.
598 * \see ofd_set_info_hdl() for request handler function.
600 * \param[in] env execution environment
601 * \param[in] exp OBD export of OFD device
602 * \param[in] keylen length of \a key
603 * \param[in] key key name
604 * \param[in] vallen length of \a val
605 * \param[in] val the \a key value
606 * \param[in] set not used in OFD
608 * \retval 0 if successful
609 * \retval negative value on error
611 static int ofd_set_info_async(const struct lu_env *env, struct obd_export *exp,
612 __u32 keylen, void *key, __u32 vallen, void *val,
613 struct ptlrpc_request_set *set)
620 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
624 if (KEY_IS(KEY_SPTLRPC_CONF)) {
625 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
627 CERROR("%s: Unsupported key %s\n",
628 exp->exp_obd->obd_name, (char *)key);
635 * Implementation of obd_ops::o_get_info.
637 * This function is not called from request handler, it is only used by
638 * direct call from nrs_orr_range_fill_physical() in ptlrpc, see LU-3239.
640 * \see ofd_get_info_hdl() for request handler function.
642 * \param[in] env execution environment
643 * \param[in] exp OBD export of OFD device
644 * \param[in] keylen length of \a key
645 * \param[in] key key name
646 * \param[out] vallen length of key value
647 * \param[out] val the key value to return
649 * \retval 0 if successful
650 * \retval negative value on error
652 static int ofd_get_info(const struct lu_env *env, struct obd_export *exp,
653 __u32 keylen, void *key, __u32 *vallen, void *val)
655 struct ofd_thread_info *info;
656 struct ofd_device *ofd;
657 struct ll_fiemap_info_key *fm_key = key;
658 struct fiemap *fiemap = val;
664 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
670 if (KEY_IS(KEY_FIEMAP)) {
671 info = ofd_info_init(env, exp);
673 rc = ostid_to_fid(&info->fti_fid, &fm_key->lfik_oa.o_oi,
674 ofd->ofd_lut.lut_lsd.lsd_osd_index);
678 rc = ofd_fiemap_get(env, ofd, &info->fti_fid, fiemap);
680 CERROR("%s: not supported key %s\n",
681 ofd_name(ofd), (char *)key);
689 * Implementation of obd_ops::o_statfs.
691 * This function returns information about a storage file system.
692 * It is called from several places by using the OBD API as well as
693 * by direct call, e.g. from request handler.
695 * \see ofd_statfs_hdl() for request handler function.
697 * Report also the state of the OST to the caller in osfs->os_state
698 * (OS_STATFS_READONLY, OS_STATFS_DEGRADED).
700 * \param[in] env execution environment
701 * \param[in] exp OBD export of OFD device
702 * \param[out] osfs statistic data to return
703 * \param[in] max_age maximum age for cached data
704 * \param[in] flags not used in OFD
706 * \retval 0 if successful
707 * \retval negative value on error
709 int ofd_statfs(const struct lu_env *env, struct obd_export *exp,
710 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
712 struct obd_device *obd = class_exp2obd(exp);
713 struct ofd_device *ofd = ofd_exp(exp);
714 struct tg_grants_data *tgd = &ofd->ofd_lut.lut_tgd;
715 int current_blockbits;
720 rc = tgt_statfs_internal(env, &ofd->ofd_lut, osfs, max_age, NULL);
724 /* tgd_blockbit is recordsize bits set during mkfs.
725 * This once set does not change. However, 'zfs set'
726 * can be used to change the OST blocksize. Instead
727 * of using cached value of 'tgd_blockbit' always
728 * calculate the blocksize bits which may have
731 current_blockbits = fls64(osfs->os_bsize) - 1;
734 * at least try to account for cached pages. its still racy and
735 * might be under-reporting if clients haven't announced their
736 * caches with brw recently
738 CDEBUG(D_SUPER | D_CACHE,
739 "blocks cached %llu granted %llu pending %llu free %llu avail %llu\n",
740 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
741 tgd->tgd_tot_pending,
742 osfs->os_bfree << current_blockbits,
743 osfs->os_bavail << current_blockbits);
745 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
746 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
747 osfs->os_bsize - 1) >> current_blockbits));
750 * The QoS code on the MDS does not care about space reserved for
751 * precreate, so take it out.
753 if (exp_connect_flags(exp) & OBD_CONNECT_MDS) {
754 struct tg_export_data *ted;
756 ted = &obd->obd_self_export->exp_target_data;
757 osfs->os_granted = min_t(u64, osfs->os_bavail,
758 ted->ted_grant >> current_blockbits);
759 osfs->os_bavail -= osfs->os_granted;
762 tgt_grant_sanity_check(obd, __func__);
763 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
764 "%llu objects: %llu free; state %x\n",
765 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
766 osfs->os_files, osfs->os_ffree, osfs->os_state);
768 if (CFS_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOINO,
769 ofd->ofd_lut.lut_lsd.lsd_osd_index)) {
770 /* Reduce free inode count to zero, but keep "used" intact */
771 osfs->os_files -= osfs->os_ffree;
772 osfs->os_ffree -= osfs->os_ffree;
775 /* OS_STATFS_READONLY can be set by OSD already, only add flags */
776 if (ofd->ofd_raid_degraded)
777 osfs->os_state |= OS_STATFS_DEGRADED;
779 if (ofd->ofd_lut.lut_no_create)
780 osfs->os_state |= OS_STATFS_NOCREATE;
782 if (obd->obd_self_export != exp && !exp_grant_param_supp(exp) &&
783 current_blockbits > COMPAT_BSIZE_SHIFT) {
785 * clients which don't support OBD_CONNECT_GRANT_PARAM
786 * should not see a block size > page size, otherwise
787 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
788 * block size which is the biggest block size known to work
789 * with all client's page size.
791 osfs->os_blocks <<= current_blockbits - COMPAT_BSIZE_SHIFT;
792 osfs->os_bfree <<= current_blockbits - COMPAT_BSIZE_SHIFT;
793 osfs->os_bavail <<= current_blockbits - COMPAT_BSIZE_SHIFT;
794 osfs->os_granted <<= current_blockbits - COMPAT_BSIZE_SHIFT;
795 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
798 if (CFS_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOSPC,
799 ofd->ofd_lut.lut_lsd.lsd_osd_index)) {
800 /* Reduce free blocks count near zero, but keep "used" intact */
801 osfs->os_bavail -= osfs->os_bavail - 2;
802 osfs->os_blocks -= osfs->os_bfree - 2;
803 osfs->os_bfree -= osfs->os_bfree - 2;
812 * Implementation of obd_ops::o_setattr.
814 * This function is only used by ECHO client when it is run on top of OFD,
815 * \see ofd_setattr_hdl() for request handler function.
817 * \param[in] env execution environment
818 * \param[in] exp OBD export of OFD device
819 * \param[in] oa setattr parameters
821 * \retval 0 if successful
822 * \retval negative value on error
824 static int ofd_echo_setattr(const struct lu_env *env, struct obd_export *exp,
827 struct ofd_thread_info *info;
828 struct ofd_device *ofd = ofd_exp(exp);
829 struct ldlm_namespace *ns = ofd->ofd_namespace;
830 struct ldlm_resource *res;
831 struct ofd_object *fo;
832 struct lu_fid *fid = &oa->o_oi.oi_fid;
833 ktime_t kstart = ktime_get();
838 info = ofd_info_init(env, exp);
840 ost_fid_build_resid(fid, &info->fti_resid);
843 * This would be very bad - accidentally truncating a file when
844 * changing the time or similar - bug 12203.
846 if (oa->o_valid & OBD_MD_FLSIZE) {
847 static char mdsinum[48];
849 if (oa->o_valid & OBD_MD_FLFID)
850 snprintf(mdsinum, sizeof(mdsinum) - 1,
851 "of parent "DFID, oa->o_parent_seq,
852 oa->o_parent_oid, 0);
856 CERROR("%s: setattr from %s trying to truncate object "DFID
857 " %s\n", ofd_name(ofd), obd_export_nid2str(exp),
859 GOTO(out, rc = -EPERM);
862 fo = ofd_object_find_exists(env, ofd, fid);
864 CERROR("%s: can't find object "DFID"\n",
865 ofd_name(ofd), PFID(fid));
866 GOTO(out, rc = PTR_ERR(fo));
869 la_from_obdo(&info->fti_attr, oa, oa->o_valid);
870 info->fti_attr.la_valid &= ~LA_TYPE;
872 /* setting objects attributes (including owner/group) */
873 rc = ofd_attr_set(env, fo, &info->fti_attr, oa);
875 GOTO(out_unlock, rc);
877 ofd_counter_incr(exp, LPROC_OFD_STATS_SETATTR, NULL,
878 ktime_us_delta(ktime_get(), kstart));
881 ofd_object_put(env, fo);
885 * we do not call this before to avoid lu_object_find() in
886 * ->lvbo_update() holding another reference on the object.
887 * otherwise concurrent destroy can make the object unavailable
888 * for 2nd lu_object_find() waiting for the first reference
891 res = ldlm_resource_get(ns, &info->fti_resid, LDLM_EXTENT, 0);
893 ldlm_res_lvbo_update(res, NULL, 0);
894 ldlm_resource_putref(res);
902 * Destroy OFD object by its FID.
904 * Supplemental function to destroy object by FID, it is used by request
905 * handler and by ofd_echo_destroy() below to find object by FID, lock it
906 * and call ofd_destroy() finally.
908 * \param[in] env execution environment
909 * \param[in] ofd OFD device
910 * \param[in] fid FID of object
911 * \param[in] orphan set if object being destroyed is an orphan
913 * \retval 0 if successful
914 * \retval negative value on error
916 int ofd_destroy_by_fid(const struct lu_env *env, struct ofd_device *ofd,
917 const struct lu_fid *fid, int orphan)
919 struct ofd_thread_info *info = ofd_info(env);
920 struct lustre_handle lockh;
921 union ldlm_policy_data policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
922 struct ofd_object *fo;
923 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
928 fo = ofd_object_find_exists(env, ofd, fid);
933 * Tell the clients that the object is gone now and that they should
934 * throw away any cached pages.
936 ost_fid_build_resid(fid, &info->fti_resid);
937 rc = ldlm_cli_enqueue_local(env, ofd->ofd_namespace, &info->fti_resid,
938 LDLM_EXTENT, &policy, LCK_PW, &flags,
939 ldlm_blocking_ast, ldlm_completion_ast,
940 NULL, NULL, 0, LVB_T_NONE, NULL, &lockh);
942 /* We only care about the side-effects, just drop the lock. */
944 ldlm_lock_decref(&lockh, LCK_PW);
948 rc = ofd_destroy(env, fo, orphan);
951 ofd_object_put(env, fo);
956 * Implementation of obd_ops::o_destroy.
958 * This function is only used by ECHO client when it is run on top of OFD,
959 * \see ofd_destroy_hdl() for request handler function.
961 * \param[in] env execution environment
962 * \param[in] exp OBD export of OFD device
963 * \param[in] oa obdo structure with FID
965 * Note: this is OBD API method which is common API for server OBDs and
966 * client OBDs. Thus some parameters used in client OBDs may not be used
967 * on server OBDs and vice versa.
969 * \retval 0 if successful
970 * \retval negative value on error
972 static int ofd_echo_destroy(const struct lu_env *env, struct obd_export *exp,
975 struct ofd_device *ofd = ofd_exp(exp);
976 struct lu_fid *fid = &oa->o_oi.oi_fid;
981 ofd_info_init(env, exp);
983 rc = ofd_validate_seq(exp, ostid_seq(&oa->o_oi));
987 CDEBUG(D_HA, "%s: Destroy object "DFID"\n", ofd_name(ofd), PFID(fid));
989 rc = ofd_destroy_by_fid(env, ofd, fid, 0);
991 CDEBUG(D_INODE, "%s: destroying non-existent object "DFID"\n",
992 ofd_name(ofd), PFID(fid));
994 } else if (rc != 0) {
995 CERROR("%s: error destroying object "DFID": %d\n",
996 ofd_name(ofd), PFID(fid), rc);
1005 * Implementation of obd_ops::o_create.
1007 * This function is only used by ECHO client when it is run on top of OFD
1008 * and just creates an object.
1009 * \see ofd_create_hdl() for request handler function.
1011 * \param[in] env execution environment
1012 * \param[in] exp OBD export of OFD device
1013 * \param[in] oa obdo structure with FID sequence to use
1015 * Note: this is OBD API method which is common API for server OBDs and
1016 * client OBDs. Thus some parameters used in client OBDs may not be used
1017 * on server OBDs and vice versa.
1019 * \retval 0 if successful
1020 * \retval negative value on error
1022 static int ofd_echo_create(const struct lu_env *env, struct obd_export *exp,
1025 struct ofd_device *ofd = ofd_exp(exp);
1026 u64 seq = ostid_seq(&oa->o_oi);
1027 struct ofd_seq *oseq;
1036 if (ofd->ofd_lut.lut_no_create)
1039 ofd_info_init(env, exp);
1041 LASSERT(seq == FID_SEQ_ECHO);
1042 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
1044 CDEBUG(D_INFO, "ofd_create("DOSTID")\n", POSTID(&oa->o_oi));
1046 down_read(&ofd->ofd_lastid_rwsem);
1048 * Currently, for safe, we do not distinguish which LAST_ID is broken,
1049 * we may do that in the future.
1050 * Return -ENOSPC until the LAST_ID rebuilt.
1052 if (unlikely(ofd->ofd_lastid_rebuilding))
1053 GOTO(out_sem, rc = -ENOSPC);
1055 rc = ofd_validate_seq(exp, seq);
1059 oseq = ofd_seq_load(env, ofd, seq);
1061 CERROR("%s: Can't find FID Sequence %#llx: rc = %ld\n",
1062 ofd_name(ofd), seq, PTR_ERR(oseq));
1063 GOTO(out_sem, rc = -EINVAL);
1066 mutex_lock(&oseq->os_create_lock);
1067 granted = tgt_grant_create(env, ofd_obd(ofd)->obd_self_export, &diff);
1072 "%s: failed to acquire grant space for precreate (%lld): rc = %d\n",
1073 ofd_name(ofd), diff, rc);
1078 next_id = ofd_seq_last_oid(oseq) + 1;
1079 count = ofd_precreate_batch(ofd, (int)diff);
1081 rc = ofd_precreate_objects(env, ofd, next_id, oseq, count, 0, false);
1083 CERROR("%s: unable to precreate: rc = %d\n",
1086 rc = ostid_set_id(&oa->o_oi, ofd_seq_last_oid(oseq));
1088 CERROR("%s: Bad %llu to set " DOSTID " : rc %d\n",
1090 (unsigned long long)ofd_seq_last_oid(oseq),
1091 POSTID(&oa->o_oi), rc);
1093 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
1096 tgt_grant_commit(ofd_obd(ofd)->obd_self_export, granted, rc);
1098 mutex_unlock(&oseq->os_create_lock);
1099 ofd_seq_put(env, oseq);
1102 up_read(&ofd->ofd_lastid_rwsem);
1107 * Implementation of obd_ops::o_getattr.
1109 * This function is only used by ECHO client when it is run on top of OFD
1110 * and returns attributes of object.
1111 * \see ofd_getattr_hdl() for request handler function.
1113 * \param[in] env execution environment
1114 * \param[in] exp OBD export of OFD device
1115 * \param[in,out] oa contains FID of object to get attributes from and
1116 * is used to return attributes back
1118 * \retval 0 if successful
1119 * \retval negative value on error
1121 static int ofd_echo_getattr(const struct lu_env *env, struct obd_export *exp,
1124 struct ofd_device *ofd = ofd_exp(exp);
1125 struct ofd_thread_info *info;
1126 struct lu_fid *fid = &oa->o_oi.oi_fid;
1127 struct ofd_object *fo;
1132 info = ofd_info_init(env, exp);
1134 fo = ofd_object_find_exists(env, ofd, fid);
1136 GOTO(out, rc = PTR_ERR(fo));
1138 LASSERT(fo != NULL);
1139 rc = ofd_attr_get(env, fo, &info->fti_attr);
1140 oa->o_valid = OBD_MD_FLID;
1144 obdo_from_la(oa, &info->fti_attr,
1145 OFD_VALID_FLAGS | LA_UID | LA_GID | LA_PROJID);
1147 /* Store object version in reply */
1148 curr_version = dt_version_get(env, ofd_object_child(fo));
1149 if ((__s64)curr_version != -EOPNOTSUPP) {
1150 oa->o_valid |= OBD_MD_FLDATAVERSION;
1151 oa->o_data_version = curr_version;
1155 ofd_object_put(env, fo);
1161 * Get object version for OBD_IOC_GET_OBJ_VERSION ioctl.
1163 * This is supplemental function for ofd_iocontrol() to return object
1164 * version for lctl tool.
1166 * \param[in] env execution environment
1167 * \param[in] ofd OFD device
1168 * \param[out] karg ioctl data
1170 * \retval 0 if successful
1171 * \retval negative value on error
1173 static int ofd_ioc_get_obj_version(const struct lu_env *env,
1174 struct ofd_device *ofd, void *karg)
1176 struct obd_ioctl_data *data = karg;
1178 struct ofd_object *fo;
1179 dt_obj_version_t version;
1183 if (!data || !data->ioc_inlbuf2 || data->ioc_inllen2 != sizeof(version))
1184 GOTO(out, rc = -EINVAL);
1186 if (data->ioc_inlbuf1 && data->ioc_inllen1 == sizeof(fid)) {
1187 fid = *(struct lu_fid *)data->ioc_inlbuf1;
1188 } else if (data->ioc_inlbuf3 &&
1189 data->ioc_inllen3 == sizeof(__u64) &&
1190 data->ioc_inlbuf4 &&
1191 data->ioc_inllen4 == sizeof(__u64)) {
1192 struct ost_id ostid = { };
1194 ostid_set_seq(&ostid, *(__u64 *)data->ioc_inlbuf4);
1195 rc = ostid_set_id(&ostid, *(__u64 *)data->ioc_inlbuf3);
1198 rc = ostid_to_fid(&fid, &ostid,
1199 ofd->ofd_lut.lut_lsd.lsd_osd_index);
1203 GOTO(out, rc = -EINVAL);
1206 if (!fid_is_sane(&fid))
1207 GOTO(out, rc = -EINVAL);
1209 fo = ofd_object_find(env, ofd, &fid);
1211 GOTO(out, rc = PTR_ERR(fo));
1213 if (!ofd_object_exists(fo))
1214 GOTO(out_fo, rc = -ENOENT);
1216 if (lu_object_remote(&fo->ofo_obj.do_lu))
1217 GOTO(out_fo, rc = -EREMOTE);
1219 version = dt_version_get(env, ofd_object_child(fo));
1221 GOTO(out_fo, rc = -EIO);
1223 *(dt_obj_version_t *)data->ioc_inlbuf2 = version;
1227 ofd_object_put(env, fo);
1233 * Implementation of obd_ops::o_iocontrol.
1235 * This is OFD ioctl handling function which is primary interface for
1236 * Lustre tools like lfs, lctl and lfsck.
1238 * \param[in] cmd ioctl command
1239 * \param[in] exp OBD export of OFD
1240 * \param[in] len not used
1241 * \param[in,out] karg buffer with data
1242 * \param[in] uarg not used
1244 * \retval 0 if successful
1245 * \retval negative value on error
1247 static int ofd_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1248 void *karg, void __user *uarg)
1251 struct ofd_device *ofd = ofd_exp(exp);
1252 struct obd_device *obd = ofd_obd(ofd);
1253 struct obd_ioctl_data *data;
1257 CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
1258 obd->obd_name, cmd, len, karg, uarg);
1259 rc = lu_env_init(&env, LCT_DT_THREAD);
1263 /* handle commands that don't use @karg first */
1266 case OBD_IOC_ABORT_RECOVERY:
1267 LCONSOLE_WARN("%s: Aborting recovery\n", obd->obd_name);
1268 obd->obd_abort_recovery = 1;
1269 target_stop_recovery_thread(obd);
1272 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
1273 rc = dt_sync(&env, ofd->ofd_osd);
1275 case OBD_IOC_SET_READONLY:
1276 rc = dt_sync(&env, ofd->ofd_osd);
1278 rc = dt_ro(&env, ofd->ofd_osd);
1282 if (unlikely(karg == NULL)) {
1283 OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL", rc = -EINVAL);
1289 case OBD_IOC_START_LFSCK: {
1290 struct lfsck_start_param lsp;
1292 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
1293 lsp.lsp_index_valid = 0;
1294 rc = lfsck_start(&env, ofd->ofd_osd, &lsp);
1297 case OBD_IOC_STOP_LFSCK: {
1298 struct lfsck_stop stop;
1300 stop.ls_status = LS_STOPPED;
1301 /* Old lfsck utils may pass NULL @stop. */
1302 if (!data->ioc_inlbuf1)
1306 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
1308 rc = lfsck_stop(&env, ofd->ofd_osd, &stop);
1311 case OBD_IOC_GET_OBJ_VERSION:
1312 rc = ofd_ioc_get_obj_version(&env, ofd, karg);
1315 rc = OBD_IOC_ERROR(obd->obd_name, cmd, "unrecognized", -ENOTTY);
1324 * Implementation of obd_ops::o_precleanup.
1326 * This function stops device activity before shutting it down. It is called
1327 * from a cleanup function upon forceful device cleanup. For OFD there are no
1328 * special actions, it just invokes target_recovery_cleanup().
1330 * \param[in] obd OBD device of OFD
1334 static int ofd_precleanup(struct obd_device *obd)
1337 target_cleanup_recovery(obd);
1342 * Implementation of obd_ops::o_health_check.
1344 * This function checks the OFD device health - ability to respond on
1345 * incoming requests. There are two health_check methods:
1346 * - get statfs from the OSD. It checks just responsiveness of
1348 * - do write attempt on bottom device to check it is fully operational and
1349 * is not stuck. This is expensive method and must be manually enabled
1350 * by setting the enable_health_write sysfs parameter.
1352 * \param[in] nul not used
1353 * \param[in] obd OBD device of OFD
1355 * \retval 0 if successful
1356 * \retval negative value in case of error
1358 static int ofd_health_check(const struct lu_env *nul, struct obd_device *obd)
1360 struct ofd_device *ofd = ofd_dev(obd->obd_lu_dev);
1361 struct ofd_thread_info *info;
1367 /* obd_proc_read_health pass NULL env, we need real one */
1368 rc = lu_env_init(&env, LCT_DT_THREAD);
1372 info = ofd_info_init(&env, NULL);
1373 rc = dt_statfs(&env, ofd->ofd_osd, &info->fti_u.osfs);
1377 if (info->fti_u.osfs.os_state & OS_STATFS_READONLY)
1378 GOTO(out, rc = -EROFS);
1380 if (!obd_enable_health_write)
1383 OBD_ALLOC(info->fti_buf.lb_buf, PAGE_SIZE);
1384 if (!info->fti_buf.lb_buf)
1385 GOTO(out, rc = -ENOMEM);
1387 info->fti_buf.lb_len = PAGE_SIZE;
1390 th = dt_trans_create(&env, ofd->ofd_osd);
1392 GOTO(out, rc = PTR_ERR(th));
1394 rc = dt_declare_record_write(&env, ofd->ofd_health_check_file,
1395 &info->fti_buf, info->fti_off, th);
1399 th->th_sync = 1; /* sync IO is needed */
1400 rc = dt_trans_start_local(&env, ofd->ofd_osd, th);
1405 rc = dt_record_write(&env, ofd->ofd_health_check_file,
1406 &info->fti_buf, &info->fti_off,
1410 rc1 = dt_trans_stop(&env, ofd->ofd_osd, th);
1413 OBD_FREE(info->fti_buf.lb_buf, PAGE_SIZE);
1415 CDEBUG(D_INFO, "write 1 page synchronously for checking io rc %d\n",
1423 const struct obd_ops ofd_obd_ops = {
1424 .o_owner = THIS_MODULE,
1425 .o_connect = ofd_obd_connect,
1426 .o_reconnect = ofd_obd_reconnect,
1427 .o_disconnect = ofd_obd_disconnect,
1428 .o_create = ofd_echo_create,
1429 .o_statfs = ofd_statfs,
1430 .o_setattr = ofd_echo_setattr,
1431 .o_preprw = ofd_preprw,
1432 .o_commitrw = ofd_commitrw,
1433 .o_destroy = ofd_echo_destroy,
1434 .o_init_export = ofd_init_export,
1435 .o_destroy_export = ofd_destroy_export,
1436 .o_postrecov = ofd_obd_postrecov,
1437 .o_getattr = ofd_echo_getattr,
1438 .o_iocontrol = ofd_iocontrol,
1439 .o_precleanup = ofd_precleanup,
1440 .o_health_check = ofd_health_check,
1441 .o_set_info_async = ofd_set_info_async,
1442 .o_get_info = ofd_get_info,