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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file provides functions to handle IO requests from clients and
35 * also LFSCK routines to check parent file identifier (PFID) consistency.
37 * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
38 * Author: Fan Yong <fan.yong@intel.com>
41 #define DEBUG_SUBSYSTEM S_FILTER
43 #include <linux/kthread.h>
44 #include "ofd_internal.h"
45 #include <lustre_nodemap.h>
47 struct ofd_inconsistency_item {
48 struct list_head oii_list;
49 struct ofd_object *oii_obj;
50 struct filter_fid oii_ff;
54 * Verify single object for parent FID consistency.
56 * Part of LFSCK processing which checks single object PFID stored in extended
57 * attribute (XATTR) against real FID of MDT parent object received by LFSCK.
58 * This verifies that the OST object is being referenced by only a single MDT
61 * \param[in] env execution environment
62 * \param[in] ofd OFD device
63 * \param[in] oii object-related local data
64 * \param[in] lrl LFSCK request data
66 static void ofd_inconsistency_verify_one(const struct lu_env *env,
67 struct ofd_device *ofd,
68 struct ofd_inconsistency_item *oii,
69 struct lfsck_req_local *lrl)
71 struct ofd_object *fo = oii->oii_obj;
72 struct filter_fid *client_ff = &oii->oii_ff;
73 struct filter_fid *local_ff = &fo->ofo_ff;
76 LASSERT(fo->ofo_pfid_checking);
77 LASSERT(!fo->ofo_pfid_verified);
79 lrl->lrl_fid = fo->ofo_header.loh_fid; /* OST-object itself FID. */
80 lrl->lrl_ff_client = *client_ff; /* client given PFID. */
81 lrl->lrl_ff_local = *local_ff; /* OST local stored PFID. */
83 rc = lfsck_in_notify_local(env, ofd->ofd_osd, lrl, NULL);
84 ofd_write_lock(env, fo);
85 switch (lrl->lrl_status) {
90 CDEBUG(D_LFSCK, "%s: fail to verify OST local stored "
91 "PFID xattr for "DFID", the client given PFID "
92 DFID", OST local stored PFID "DFID": rc = %d\n",
93 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
94 PFID(&client_ff->ff_parent),
95 PFID(&local_ff->ff_parent), rc);
97 fo->ofo_pfid_verified = 1;
99 case LPVS_INCONSISTENT:
102 ofd->ofd_inconsistency_self_detected++;
104 CDEBUG(D_LFSCK, "%s: fail to verify the client given "
105 "PFID for "DFID", the client given PFID "DFID
106 ", local stored PFID "DFID": rc = %d\n",
107 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
108 PFID(&client_ff->ff_parent),
109 PFID(&local_ff->ff_parent), rc);
111 CDEBUG(D_LFSCK, "%s: both the client given PFID and "
112 "the OST local stored PFID are stale for the "
113 "OST-object "DFID", client given PFID is "DFID
114 ", local stored PFID is "DFID"\n",
115 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
116 PFID(&client_ff->ff_parent),
117 PFID(&local_ff->ff_parent));
119 case LPVS_INCONSISTENT_TOFIX:
120 ofd->ofd_inconsistency_self_detected++;
122 ofd->ofd_inconsistency_self_repaired++;
123 CDEBUG(D_LFSCK, "%s: fixed the staled OST PFID xattr "
124 "for "DFID", with the client given PFID "DFID
125 ", the old stored PFID "DFID"\n",
126 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
127 PFID(&client_ff->ff_parent),
128 PFID(&local_ff->ff_parent));
130 CDEBUG(D_LFSCK, "%s: fail to fix the OST PFID xattr "
131 "for "DFID", client given PFID "DFID", local "
132 "stored PFID "DFID": rc = %d\n",
133 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
134 PFID(&client_ff->ff_parent),
135 PFID(&local_ff->ff_parent), rc);
137 local_ff->ff_parent = client_ff->ff_parent;
138 fo->ofo_pfid_verified = 1;
143 fo->ofo_pfid_checking = 0;
144 ofd_write_unlock(env, fo);
146 ofd_object_put(env, fo);
151 * Verification thread to check parent FID consistency.
153 * Kernel thread to check consistency of parent FID for any
154 * new item added for checking by ofd_add_inconsistency_item().
156 * \param[in] args OFD device
158 * \retval 0 on successful thread termination
159 * \retval negative value if thread can't start
161 static int ofd_inconsistency_verification_main(void *args)
164 struct ofd_device *ofd = args;
165 struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread;
166 struct ofd_inconsistency_item *oii;
167 struct lfsck_req_local *lrl = NULL;
171 rc = lu_env_init(&env, LCT_DT_THREAD);
172 spin_lock(&ofd->ofd_inconsistency_lock);
173 thread_set_flags(thread, rc ? SVC_STOPPED : SVC_RUNNING);
174 wake_up_all(&thread->t_ctl_waitq);
175 spin_unlock(&ofd->ofd_inconsistency_lock);
181 GOTO(out_unlocked, rc = -ENOMEM);
183 lrl->lrl_event = LEL_PAIRS_VERIFY_LOCAL;
184 lrl->lrl_active = LFSCK_TYPE_LAYOUT;
186 spin_lock(&ofd->ofd_inconsistency_lock);
188 if (unlikely(!thread_is_running(thread)))
191 while (!list_empty(&ofd->ofd_inconsistency_list)) {
192 oii = list_entry(ofd->ofd_inconsistency_list.next,
193 struct ofd_inconsistency_item,
195 list_del_init(&oii->oii_list);
196 spin_unlock(&ofd->ofd_inconsistency_lock);
197 ofd_inconsistency_verify_one(&env, ofd, oii, lrl);
198 spin_lock(&ofd->ofd_inconsistency_lock);
201 spin_unlock(&ofd->ofd_inconsistency_lock);
202 wait_event_idle(thread->t_ctl_waitq,
203 !list_empty(&ofd->ofd_inconsistency_list) ||
204 !thread_is_running(thread));
205 spin_lock(&ofd->ofd_inconsistency_lock);
208 while (!list_empty(&ofd->ofd_inconsistency_list)) {
209 struct ofd_object *fo;
211 oii = list_entry(ofd->ofd_inconsistency_list.next,
212 struct ofd_inconsistency_item,
214 list_del_init(&oii->oii_list);
216 spin_unlock(&ofd->ofd_inconsistency_lock);
218 ofd_write_lock(&env, fo);
219 fo->ofo_pfid_checking = 0;
220 ofd_write_unlock(&env, fo);
222 ofd_object_put(&env, fo);
224 spin_lock(&ofd->ofd_inconsistency_lock);
232 spin_lock(&ofd->ofd_inconsistency_lock);
234 thread_set_flags(thread, SVC_STOPPED);
235 wake_up_all(&thread->t_ctl_waitq);
236 spin_unlock(&ofd->ofd_inconsistency_lock);
243 * Start parent FID verification thread.
245 * See ofd_inconsistency_verification_main().
247 * \param[in] ofd OFD device
249 * \retval 0 on successful start of thread
250 * \retval negative value on error
252 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd)
254 struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread;
255 struct task_struct *task;
258 spin_lock(&ofd->ofd_inconsistency_lock);
259 if (unlikely(thread_is_running(thread))) {
260 spin_unlock(&ofd->ofd_inconsistency_lock);
265 thread_set_flags(thread, 0);
266 spin_unlock(&ofd->ofd_inconsistency_lock);
267 task = kthread_run(ofd_inconsistency_verification_main, ofd,
268 "inconsistency_verification");
271 CERROR("%s: cannot start self_repair thread: rc = %d\n",
275 wait_event_idle(thread->t_ctl_waitq,
276 thread_is_running(thread) ||
277 thread_is_stopped(thread));
284 * Stop parent FID verification thread.
286 * \param[in] ofd OFD device
288 * \retval 0 on successful start of thread
289 * \retval -EALREADY if thread is already stopped
291 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd)
293 struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread;
295 spin_lock(&ofd->ofd_inconsistency_lock);
296 if (thread_is_init(thread) || thread_is_stopped(thread)) {
297 spin_unlock(&ofd->ofd_inconsistency_lock);
302 thread_set_flags(thread, SVC_STOPPING);
303 spin_unlock(&ofd->ofd_inconsistency_lock);
304 wake_up_all(&thread->t_ctl_waitq);
305 wait_event_idle(thread->t_ctl_waitq,
306 thread_is_stopped(thread));
312 * Add new item for parent FID verification.
314 * Prepare new verification item and pass it to the dedicated
315 * verification thread for further processing.
317 * \param[in] env execution environment
318 * \param[in] fo OFD object
319 * \param[in] oa OBDO structure with PFID
321 static void ofd_add_inconsistency_item(const struct lu_env *env,
322 struct ofd_object *fo, struct obdo *oa)
324 struct ofd_device *ofd = ofd_obj2dev(fo);
325 struct ofd_inconsistency_item *oii;
326 struct filter_fid *ff;
333 INIT_LIST_HEAD(&oii->oii_list);
334 lu_object_get(&fo->ofo_obj.do_lu);
337 ff->ff_parent.f_seq = oa->o_parent_seq;
338 ff->ff_parent.f_oid = oa->o_parent_oid;
339 ff->ff_parent.f_stripe_idx = oa->o_stripe_idx;
340 ff->ff_layout = oa->o_layout;
342 spin_lock(&ofd->ofd_inconsistency_lock);
343 if (fo->ofo_pfid_checking || fo->ofo_pfid_verified) {
344 spin_unlock(&ofd->ofd_inconsistency_lock);
350 fo->ofo_pfid_checking = 1;
351 if (list_empty(&ofd->ofd_inconsistency_list))
353 list_add_tail(&oii->oii_list, &ofd->ofd_inconsistency_list);
354 spin_unlock(&ofd->ofd_inconsistency_lock);
356 wake_up_all(&ofd->ofd_inconsistency_thread.t_ctl_waitq);
358 /* XXX: When the found inconsistency exceeds some threshold,
359 * we can trigger the LFSCK to scan part of the system
360 * or the whole system, which depends on how to define
361 * the threshold, a simple way maybe like that: define
362 * the absolute value of how many inconsisteny allowed
363 * to be repaired via self detect/repair mechanism, if
364 * exceeded, then trigger the LFSCK to scan the layout
365 * inconsistency within the whole system. */
369 * Verify parent FID of an object.
371 * Check the parent FID is sane and start extended
372 * verification procedure otherwise.
374 * \param[in] env execution environment
375 * \param[in] fo OFD object
376 * \param[in] oa OBDO structure with PFID
378 * \retval 0 on successful verification
379 * \retval -EINPROGRESS if PFID is being repaired
380 * \retval -EPERM if PFID was verified but still insane
382 int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
385 struct lu_fid *pfid = &fo->ofo_ff.ff_parent;
389 if (fid_is_sane(pfid)) {
390 if (likely(oa->o_parent_seq == pfid->f_seq &&
391 oa->o_parent_oid == pfid->f_oid &&
392 oa->o_stripe_idx == pfid->f_stripe_idx))
395 if (fo->ofo_pfid_verified)
399 /* The OST-object may be inconsistent, and we need further verification.
400 * To avoid block the RPC service thread, return -EINPROGRESS to client
401 * and make it retry later. */
402 if (fo->ofo_pfid_checking)
403 RETURN(-EINPROGRESS);
405 rc = ofd_object_ff_load(env, fo);
412 if (likely(oa->o_parent_seq == pfid->f_seq &&
413 oa->o_parent_oid == pfid->f_oid &&
414 oa->o_stripe_idx == pfid->f_stripe_idx))
417 /* Push it to the dedicated thread for further verification. */
418 ofd_add_inconsistency_item(env, fo, oa);
420 RETURN(-EINPROGRESS);
424 * FLR: verify the layout version of object.
426 * \param[in] env execution environment
427 * \param[in] fo OFD object
428 * \param[in] oa OBDO structure with layout version
430 * \retval 0 on successful verification
431 * \retval -EINPROGRESS layout version is in transfer
432 * \retval -ESTALE the layout version on client is stale
434 int ofd_verify_layout_version(const struct lu_env *env,
435 struct ofd_object *fo, const struct obdo *oa)
437 __u32 layout_version;
441 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_SKIP_LV_CHECK)))
444 rc = ofd_object_ff_load(env, fo);
451 layout_version = fo->ofo_ff.ff_layout_version;
452 if (oa->o_layout_version >= layout_version &&
453 oa->o_layout_version <= layout_version + fo->ofo_ff.ff_range)
456 /* normal traffic, decide if to return ESTALE or EINPROGRESS */
457 layout_version &= ~LU_LAYOUT_RESYNC;
459 /* this update is not legitimate */
460 if ((oa->o_layout_version & ~LU_LAYOUT_RESYNC) <= layout_version)
461 GOTO(out, rc = -ESTALE);
463 /* layout version may not be transmitted yet */
464 if ((oa->o_layout_version & ~LU_LAYOUT_RESYNC) > layout_version)
465 GOTO(out, rc = -EINPROGRESS);
470 CDEBUG(D_INODE, DFID " verify layout version: %u vs. %u/%u, rc: %d\n",
471 PFID(lu_object_fid(&fo->ofo_obj.do_lu)),
472 oa->o_layout_version, fo->ofo_ff.ff_layout_version,
473 fo->ofo_ff.ff_range, rc);
479 * Lazy ATIME update to refresh atime every ofd_atime_diff
480 * seconds so that external scanning tool can see it actual
481 * within that period and be able to identify accessed files
483 static void ofd_handle_atime(const struct lu_env *env, struct ofd_device *ofd,
484 struct ofd_object *fo, time64_t atime)
491 if (ofd->ofd_atime_diff == 0)
494 la = &ofd_info(env)->fti_attr2;
495 o = ofd_object_child(fo);
497 if (unlikely(fo->ofo_atime_ondisk == 0)) {
498 rc = dt_attr_get(env, o, la);
501 LASSERT(la->la_valid & LA_ATIME);
502 if (la->la_atime == 0)
503 la->la_atime = la->la_mtime;
504 fo->ofo_atime_ondisk = la->la_atime;
506 if (atime - fo->ofo_atime_ondisk < ofd->ofd_atime_diff)
509 /* atime hasn't been updated too long, update it */
510 fo->ofo_atime_ondisk = atime;
512 th = ofd_trans_create(env, ofd);
514 CERROR("%s: cannot create transaction: rc = %d\n",
515 ofd_name(ofd), (int)PTR_ERR(th));
519 la->la_valid = LA_ATIME;
520 rc = dt_declare_attr_set(env, o, la, th);
524 rc = dt_trans_start_local(env, ofd->ofd_osd , th);
526 CERROR("%s: cannot start transaction: rc = %d\n",
531 ofd_read_lock(env, fo);
532 if (ofd_object_exists(fo)) {
533 la->la_atime = fo->ofo_atime_ondisk;
534 rc = dt_attr_set(env, o, la, th);
537 ofd_read_unlock(env, fo);
540 ofd_trans_stop(env, ofd, th, rc);
544 * Prepare buffers for read request processing.
546 * This function converts remote buffers from client to local buffers
547 * and prepares the latter.
549 * \param[in] env execution environment
550 * \param[in] exp OBD export of client
551 * \param[in] ofd OFD device
552 * \param[in] fid FID of object
553 * \param[in] la object attributes
554 * \param[in] oa OBDO structure from client
555 * \param[in] niocount number of remote buffers
556 * \param[in] rnb remote buffers
557 * \param[in] nr_local number of local buffers
558 * \param[in] lnb local buffers
559 * \param[in] jobid job ID name
561 * \retval 0 on successful prepare
562 * \retval negative value on error
564 static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
565 struct ofd_device *ofd, const struct lu_fid *fid,
566 struct lu_attr *la, struct obdo *oa, int niocount,
567 struct niobuf_remote *rnb, int *nr_local,
568 struct niobuf_local *lnb, char *jobid)
570 struct ofd_object *fo;
571 int i, j, rc, tot_bytes = 0;
572 enum dt_bufs_type dbt = DT_BUFS_TYPE_READ;
573 int maxlnb = *nr_local;
577 LASSERT(env != NULL);
579 fo = ofd_object_find(env, ofd, fid);
584 ofd_info(env)->fti_obj = fo;
586 if (oa->o_valid & OBD_MD_FLATIME)
587 ofd_handle_atime(env, ofd, fo, oa->o_atime);
589 ofd_read_lock(env, fo);
590 if (!ofd_object_exists(fo))
591 GOTO(unlock, rc = -ENOENT);
593 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
594 rc = ofd_verify_ff(env, fo, oa);
599 if (ptlrpc_connection_is_local(exp->exp_connection))
600 dbt |= DT_BUFS_TYPE_LOCAL;
605 for (*nr_local = 0, i = 0, j = 0; i < niocount; i++) {
606 begin = min_t(__u64, begin, rnb[i].rnb_offset);
607 end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
609 if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
610 rnb[i].rnb_len = 100 * 1024 * 1024;
612 rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
613 lnb + j, maxlnb, dbt);
614 if (unlikely(rc < 0))
616 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
617 /* correct index for local buffers to continue with */
621 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
622 tot_bytes += rnb[i].rnb_len;
625 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
626 rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
632 .f_seq = oa->o_parent_seq,
633 .f_oid = oa->o_parent_oid,
634 .f_ver = oa->o_stripe_idx,
641 ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes);
645 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
647 ofd_read_unlock(env, fo);
648 ofd_object_put(env, fo);
653 * Prepare buffers for write request processing.
655 * This function converts remote buffers from client to local buffers
656 * and prepares the latter. If there is recovery in progress and required
657 * object is missing then it can be re-created before write.
659 * \param[in] env execution environment
660 * \param[in] exp OBD export of client
661 * \param[in] ofd OFD device
662 * \param[in] fid FID of object
663 * \param[in] la object attributes
664 * \param[in] oa OBDO structure from client
665 * \param[in] objcount always 1
666 * \param[in] obj object data
667 * \param[in] rnb remote buffers
668 * \param[in] nr_local number of local buffers
669 * \param[in] lnb local buffers
670 * \param[in] jobid job ID name
672 * \retval 0 on successful prepare
673 * \retval negative value on error
675 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
676 struct ofd_device *ofd, const struct lu_fid *fid,
677 struct lu_attr *la, struct obdo *oa,
678 int objcount, struct obd_ioobj *obj,
679 struct niobuf_remote *rnb, int *nr_local,
680 struct niobuf_local *lnb, char *jobid)
682 struct ofd_object *fo;
683 int i, j, k, rc = 0, tot_bytes = 0;
684 enum dt_bufs_type dbt = DT_BUFS_TYPE_WRITE;
685 int maxlnb = *nr_local;
689 LASSERT(env != NULL);
690 LASSERT(objcount == 1);
692 if (unlikely(exp->exp_obd->obd_recovering)) {
693 u64 seq = fid_seq(fid);
694 u64 oid = fid_oid(fid);
695 struct ofd_seq *oseq;
697 oseq = ofd_seq_load(env, ofd, seq);
699 CERROR("%s: Can't find FID Sequence %#llx: rc = %d\n",
700 ofd_name(ofd), seq, (int)PTR_ERR(oseq));
701 GOTO(out, rc = -EINVAL);
704 if (oid > ofd_seq_last_oid(oseq)) {
708 mutex_lock(&oseq->os_create_lock);
709 diff = oid - ofd_seq_last_oid(oseq);
711 /* Do sync create if the seq is about to used up */
712 if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) {
713 if (unlikely(oid >= IDIF_MAX_OID - 1))
715 } else if (fid_seq_is_norm(seq)) {
717 LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
720 CERROR("%s : invalid o_seq "DOSTID"\n",
721 ofd_name(ofd), POSTID(&oa->o_oi));
722 mutex_unlock(&oseq->os_create_lock);
723 ofd_seq_put(env, oseq);
724 GOTO(out, rc = -EINVAL);
728 u64 next_id = ofd_seq_last_oid(oseq) + 1;
729 int count = ofd_precreate_batch(ofd, diff);
731 rc = ofd_precreate_objects(env, ofd, next_id,
734 mutex_unlock(&oseq->os_create_lock);
735 ofd_seq_put(env, oseq);
742 mutex_unlock(&oseq->os_create_lock);
745 ofd_seq_put(env, oseq);
748 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
749 * space back if possible, we have to do this outside of the lock as
750 * grant preparation may need to sync whole fs thus wait for all the
751 * transactions to complete. */
752 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
754 fo = ofd_object_find(env, ofd, fid);
756 GOTO(out, rc = PTR_ERR(fo));
759 ofd_info(env)->fti_obj = fo;
761 ofd_read_lock(env, fo);
762 if (!ofd_object_exists(fo)) {
763 CERROR("%s: BRW to missing obj "DOSTID"\n",
764 exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
765 ofd_read_unlock(env, fo);
766 ofd_object_put(env, fo);
767 GOTO(out, rc = -ENOENT);
770 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
771 rc = ofd_verify_ff(env, fo, oa);
773 ofd_read_unlock(env, fo);
774 ofd_object_put(env, fo);
779 /* need to verify layout version */
780 if (oa->o_valid & OBD_MD_LAYOUT_VERSION) {
781 rc = ofd_verify_layout_version(env, fo, oa);
783 ofd_read_unlock(env, fo);
784 ofd_object_put(env, fo);
788 oa->o_valid &= ~OBD_MD_LAYOUT_VERSION;
791 if (ptlrpc_connection_is_local(exp->exp_connection))
792 dbt |= DT_BUFS_TYPE_LOCAL;
797 /* parse remote buffers to local buffers and prepare the latter */
798 for (*nr_local = 0, i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
799 begin = min_t(__u64, begin, rnb[i].rnb_offset);
800 end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
802 if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
803 rnb[i].rnb_len += PAGE_SIZE;
804 rc = dt_bufs_get(env, ofd_object_child(fo),
805 rnb + i, lnb + j, maxlnb, dbt);
806 if (unlikely(rc < 0))
808 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
809 /* correct index for local buffers to continue with */
810 for (k = 0; k < rc; k++) {
811 lnb[j+k].lnb_flags = rnb[i].rnb_flags;
812 lnb[j+k].lnb_flags &= ~OBD_BRW_LOCALS;
813 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
814 lnb[j+k].lnb_rc = -ENOSPC;
819 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
820 tot_bytes += rnb[i].rnb_len;
822 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
824 rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
825 if (unlikely(rc != 0))
828 ofd_read_unlock(env, fo);
832 .f_seq = oa->o_parent_seq,
833 .f_oid = oa->o_parent_oid,
834 .f_ver = oa->o_stripe_idx,
841 ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes);
844 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
845 ofd_read_unlock(env, fo);
846 ofd_object_put(env, fo);
847 /* tgt_grant_prepare_write() was called, so we must commit */
848 tgt_grant_commit(exp, oa->o_grant_used, rc);
850 /* let's still process incoming grant information packed in the oa,
851 * but without enforcing grant since we won't proceed with the write.
852 * Just like a read request actually. */
853 tgt_grant_prepare_read(env, exp, oa);
858 * Prepare bulk IO requests for processing.
860 * This function does initial checks of IO and calls corresponding
861 * functions for read/write processing.
863 * \param[in] env execution environment
864 * \param[in] cmd IO type (read/write)
865 * \param[in] exp OBD export of client
866 * \param[in] oa OBDO structure from request
867 * \param[in] objcount always 1
868 * \param[in] obj object data
869 * \param[in] rnb remote buffers
870 * \param[in] nr_local number of local buffers
871 * \param[in] lnb local buffers
873 * \retval 0 on successful prepare
874 * \retval negative value on error
876 int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
877 struct obdo *oa, int objcount, struct obd_ioobj *obj,
878 struct niobuf_remote *rnb, int *nr_local,
879 struct niobuf_local *lnb)
881 struct tgt_session_info *tsi = tgt_ses_info(env);
882 struct ofd_device *ofd = ofd_exp(exp);
883 struct ofd_thread_info *info;
885 const struct lu_fid *fid = &oa->o_oi.oi_fid;
888 if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
889 CERROR("%s: bulk has too many pages %d, which exceeds the maximum pages per RPC of %d\n",
890 exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
894 if (tgt_ses_req(tsi) == NULL) { /* echo client case */
895 info = ofd_info_init(env, exp);
898 info = tsi2ofd_info(tsi);
899 jobid = tsi->tsi_jobid;
904 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) {
905 struct ofd_seq *oseq;
907 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
909 CERROR("%s: Can not find seq for "DOSTID
910 ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
915 if (oseq->os_destroys_in_progress == 0) {
916 /* don't fail lookups for orphan recovery, it causes
917 * later LBUGs when objects still exist during
919 ofd_seq_put(env, oseq);
922 ofd_seq_put(env, oseq);
925 LASSERT(objcount == 1);
926 LASSERT(obj->ioo_bufcnt > 0);
928 if (cmd == OBD_BRW_WRITE) {
929 la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
930 rc = ofd_preprw_write(env, exp, ofd, fid, &info->fti_attr, oa,
931 objcount, obj, rnb, nr_local, lnb, jobid);
932 } else if (cmd == OBD_BRW_READ) {
933 tgt_grant_prepare_read(env, exp, oa);
934 rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa,
935 obj->ioo_bufcnt, rnb, nr_local, lnb,
938 CERROR("%s: wrong cmd %d received!\n",
939 exp->exp_obd->obd_name, cmd);
946 * Drop reference on local buffers for read bulk IO.
948 * This will free all local buffers use by this read request.
950 * \param[in] env execution environment
951 * \param[in] ofd OFD device
952 * \param[in] fid object FID
953 * \param[in] objcount always 1
954 * \param[in] niocount number of local buffers
955 * \param[in] lnb local buffers
957 * \retval 0 on successful execution
958 * \retval negative value on error
961 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
962 const struct lu_fid *fid, int objcount, int niocount,
963 struct niobuf_local *lnb)
965 struct ofd_object *fo;
969 LASSERT(niocount > 0);
971 fo = ofd_info(env)->fti_obj;
973 LASSERT(ofd_object_exists(fo));
974 dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
976 ofd_read_unlock(env, fo);
977 ofd_object_put(env, fo);
983 * Set attributes of object during write bulk IO processing.
985 * Change object attributes and write parent FID into extended
986 * attributes when needed.
988 * \param[in] env execution environment
989 * \param[in] ofd OFD device
990 * \param[in] ofd_obj OFD object
991 * \param[in] la object attributes
994 * \retval 0 on successful attributes update
995 * \retval negative value on error
998 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
999 struct ofd_object *ofd_obj, struct lu_attr *la,
1002 struct ofd_thread_info *info = ofd_info(env);
1003 struct filter_fid *ff = &info->fti_mds_fid;
1004 __u64 valid = la->la_valid;
1006 struct dt_object *dt_obj;
1014 dt_obj = ofd_object_child(ofd_obj);
1015 LASSERT(dt_obj != NULL);
1017 la->la_valid &= LA_UID | LA_GID | LA_PROJID;
1019 rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
1023 if (!la->la_valid && !(oa->o_valid &
1024 (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
1025 /* no attributes to set */
1028 th = ofd_trans_create(env, ofd);
1030 GOTO(out, rc = PTR_ERR(th));
1033 rc = dt_declare_attr_set(env, dt_obj, la, th);
1038 if (oa->o_valid & (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT |
1039 OBD_MD_LAYOUT_VERSION)) {
1040 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
1041 XATTR_NAME_FID, 0, th);
1045 /* We don't need a transno for this operation which will be re-executed
1046 * anyway when the OST_WRITE (with a transno assigned) is replayed */
1047 rc = dt_trans_start_local(env, ofd->ofd_osd , th);
1051 ofd_read_lock(env, ofd_obj);
1053 rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
1055 GOTO(out_unlock, rc);
1057 if (!la->la_valid && !(oa->o_valid &
1058 (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
1059 /* no attributes to set */
1060 GOTO(out_unlock, rc = 0);
1064 /* set uid/gid/projid */
1066 rc = dt_attr_set(env, dt_obj, la, th);
1068 GOTO(out_unlock, rc);
1071 fl = ofd_object_ff_update(env, ofd_obj, oa, ff);
1073 GOTO(out_unlock, rc = fl);
1075 /* set filter fid EA.
1076 * FIXME: it holds read lock of ofd object to modify the XATTR_NAME_FID
1077 * while the write lock should be held. However, it should work because
1078 * write RPCs only modify ff_{parent,layout} and those information will
1079 * be the same from all the write RPCs. The reason that fl is not used
1080 * in dt_xattr_set() is to allow this race. */
1081 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
1082 GOTO(out_unlock, rc);
1083 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
1084 ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
1085 else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
1086 le32_add_cpu(&ff->ff_parent.f_oid, -1);
1088 info->fti_buf.lb_buf = ff;
1089 info->fti_buf.lb_len = sizeof(*ff);
1090 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, 0, th);
1092 filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff));
1094 GOTO(out_unlock, rc);
1097 ofd_read_unlock(env, ofd_obj);
1099 dt_trans_stop(env, ofd->ofd_osd, th);
1101 la->la_valid = valid;
1105 struct ofd_soft_sync_callback {
1106 struct dt_txn_commit_cb ossc_cb;
1107 struct obd_export *ossc_exp;
1111 * Callback function for "soft sync" update.
1113 * Reset fed_soft_sync_count upon committing the "soft_sync" update.
1114 * See ofd_soft_sync_cb_add() below for more details on soft sync.
1116 * \param[in] env execution environment
1117 * \param[in] th transaction handle
1118 * \param[in] cb callback data
1119 * \param[in] err error code
1121 static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th,
1122 struct dt_txn_commit_cb *cb, int err)
1124 struct ofd_soft_sync_callback *ossc;
1126 ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb);
1128 CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp);
1129 atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0);
1131 class_export_cb_put(ossc->ossc_exp);
1136 * Add callback for "soft sync" processing.
1138 * The "soft sync" mechanism does asynchronous commit when OBD_BRW_SOFT_SYNC
1139 * flag is set in client buffers. The intention is for this operation to
1140 * commit pages belonging to a client which has "too many" outstanding
1141 * unstable pages in its cache. See LU-2139 for details.
1143 * This function adds callback to be called when commit is done.
1145 * \param[in] th transaction handle
1146 * \param[in] exp OBD export of client
1148 * \retval 0 on successful callback adding
1149 * \retval negative value on error
1151 static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
1153 struct ofd_soft_sync_callback *ossc;
1154 struct dt_txn_commit_cb *dcb;
1157 OBD_ALLOC_PTR(ossc);
1161 ossc->ossc_exp = class_export_cb_get(exp);
1163 dcb = &ossc->ossc_cb;
1164 dcb->dcb_func = ofd_cb_soft_sync;
1165 INIT_LIST_HEAD(&dcb->dcb_linkage);
1166 strlcpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name));
1168 rc = dt_trans_cb_add(th, dcb);
1170 class_export_cb_put(exp);
1178 * Commit bulk IO buffers to the storage.
1180 * This function finalizes write IO processing by writing data to the disk.
1181 * That write can be synchronous or asynchronous depending on buffers flags.
1183 * \param[in] env execution environment
1184 * \param[in] exp OBD export of client
1185 * \param[in] ofd OFD device
1186 * \param[in] fid FID of object
1187 * \param[in] la object attributes
1188 * \param[in] ff parent FID of object
1189 * \param[in] objcount always 1
1190 * \param[in] niocount number of local buffers
1191 * \param[in] lnb local buffers
1192 * \param[in] granted grant space consumed for the bulk I/O
1193 * \param[in] old_rc result of processing at this point
1195 * \retval 0 on successful commit
1196 * \retval negative value on error
1199 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
1200 struct ofd_device *ofd, const struct lu_fid *fid,
1201 struct lu_attr *la, struct obdo *oa, int objcount,
1202 int niocount, struct niobuf_local *lnb,
1203 unsigned long granted, int old_rc)
1205 struct ofd_thread_info *info = ofd_info(env);
1206 struct filter_export_data *fed = &exp->exp_filter_data;
1207 struct ofd_object *fo;
1208 struct dt_object *o;
1214 bool soft_sync = false;
1215 bool cb_registered = false;
1216 bool fake_write = false;
1220 LASSERT(objcount == 1);
1222 fo = ofd_info(env)->fti_obj;
1223 LASSERT(fo != NULL);
1225 o = ofd_object_child(fo);
1229 GOTO(out, rc = old_rc);
1230 if (!ofd_object_exists(fo))
1231 GOTO(out, rc = -ENOENT);
1234 * The first write to each object must set some attributes. It is
1235 * important to set the uid/gid before calling
1236 * dt_declare_write_commit() since quota enforcement is now handled in
1239 rc = ofd_write_attr_set(env, ofd, fo, la, oa);
1243 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
1245 /* do fake write, to simulate the write case for performance testing */
1246 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW)) {
1247 struct niobuf_local *last = &lnb[niocount - 1];
1248 __u64 file_size = last->lnb_file_offset + last->lnb_len;
1249 __u64 valid = la->la_valid;
1251 la->la_valid = LA_SIZE;
1253 rc = dt_attr_get(env, o, la);
1254 if (rc < 0 && rc != -ENOENT)
1257 if (file_size < la->la_size)
1258 file_size = la->la_size;
1260 /* dirty inode by setting file size */
1261 la->la_valid = valid | LA_SIZE;
1262 la->la_size = file_size;
1268 th = ofd_trans_create(env, ofd);
1270 GOTO(out, rc = PTR_ERR(th));
1272 th->th_sync |= ofd->ofd_sync_journal;
1273 if (th->th_sync == 0) {
1274 for (i = 0; i < niocount; i++) {
1275 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
1279 if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC)
1284 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
1285 GOTO(out_stop, rc = -EINPROGRESS);
1287 if (likely(!fake_write)) {
1288 rc = dt_declare_write_commit(env, o, lnb, niocount, th);
1293 /* don't update atime on disk if it is older */
1294 if (la->la_valid & LA_ATIME && la->la_atime <= fo->ofo_atime_ondisk)
1295 la->la_valid &= ~LA_ATIME;
1298 /* update [mac]time if needed */
1299 rc = dt_declare_attr_set(env, o, la, th);
1304 rc = ofd_trans_start(env, ofd, fo, th);
1308 ofd_read_lock(env, fo);
1309 if (!ofd_object_exists(fo))
1310 GOTO(out_unlock, rc = -ENOENT);
1312 if (likely(!fake_write)) {
1313 rc = dt_write_commit(env, o, lnb, niocount, th);
1315 GOTO(out_unlock, rc);
1318 /* Don't update timestamps if this write is older than a
1319 * setattr which modifies the timestamps. b=10150 */
1320 if (la->la_valid && tgt_fmd_check(exp, fid, info->fti_xid)) {
1321 rc = dt_attr_set(env, o, la, th);
1323 GOTO(out_unlock, rc);
1324 if (la->la_valid & LA_ATIME)
1325 fo->ofo_atime_ondisk = la->la_atime;
1328 /* get attr to return */
1329 rc = dt_attr_get(env, o, la);
1332 ofd_read_unlock(env, fo);
1334 /* Force commit to make the just-deleted blocks
1335 * reusable. LU-456 */
1339 /* do this before trans stop in case commit has finished */
1340 if (!th->th_sync && soft_sync && !cb_registered) {
1341 ofd_soft_sync_cb_add(th, exp);
1342 cb_registered = true;
1345 if (rc == 0 && granted > 0) {
1346 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
1350 rc2 = ofd_trans_stop(env, ofd, th, rc);
1353 if (rc == -ENOSPC && retries++ < 3) {
1354 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
1360 /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
1361 atomic_set(&fed->fed_soft_sync_count, 0);
1362 else if (atomic_inc_return(&fed->fed_soft_sync_count) ==
1363 ofd->ofd_soft_sync_limit)
1364 dt_commit_async(env, ofd->ofd_osd);
1367 dt_bufs_put(env, o, lnb, niocount);
1368 ofd_object_put(env, fo);
1370 tgt_grant_commit(exp, granted, old_rc);
1375 * Commit bulk IO to the storage.
1377 * This is companion function to the ofd_preprw(). It finishes bulk IO
1378 * request processing by committing buffers to the storage (WRITE) and/or
1379 * freeing those buffers (read/write). See ofd_commitrw_read() and
1380 * ofd_commitrw_write() for details about each type of IO.
1382 * \param[in] env execution environment
1383 * \param[in] cmd IO type (READ/WRITE)
1384 * \param[in] exp OBD export of client
1385 * \param[in] oa OBDO structure from client
1386 * \param[in] objcount always 1
1387 * \param[in] obj object data
1388 * \param[in] rnb remote buffers
1389 * \param[in] npages number of local buffers
1390 * \param[in] lnb local buffers
1391 * \param[in] old_rc result of processing at this point
1393 * \retval 0 on successful commit
1394 * \retval negative value on error
1396 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
1397 struct obdo *oa, int objcount, struct obd_ioobj *obj,
1398 struct niobuf_remote *rnb, int npages,
1399 struct niobuf_local *lnb, int old_rc)
1401 struct ofd_thread_info *info = ofd_info(env);
1402 struct ofd_device *ofd = ofd_exp(exp);
1403 const struct lu_fid *fid = &oa->o_oi.oi_fid;
1404 struct ldlm_namespace *ns = ofd->ofd_namespace;
1405 struct ldlm_resource *rs = NULL;
1409 LASSERT(npages > 0);
1411 if (cmd == OBD_BRW_WRITE) {
1412 struct lu_nodemap *nodemap;
1414 valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID |
1415 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
1416 la_from_obdo(&info->fti_attr, oa, valid);
1418 rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
1419 oa, objcount, npages, lnb,
1420 oa->o_grant_used, old_rc);
1422 obdo_from_la(oa, &info->fti_attr,
1423 OFD_VALID_FLAGS | LA_GID | LA_UID |
1426 obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID |
1429 /* don't report overquota flag if we failed before reaching
1431 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
1432 /* return the overquota flags to client */
1433 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
1434 if (oa->o_valid & OBD_MD_FLFLAGS)
1435 oa->o_flags |= OBD_FL_NO_USRQUOTA;
1437 oa->o_flags = OBD_FL_NO_USRQUOTA;
1440 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
1441 if (oa->o_valid & OBD_MD_FLFLAGS)
1442 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
1444 oa->o_flags = OBD_FL_NO_GRPQUOTA;
1446 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
1447 if (oa->o_valid & OBD_MD_FLFLAGS)
1448 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
1450 oa->o_flags = OBD_FL_NO_PRJQUOTA;
1453 oa->o_valid |= OBD_MD_FLFLAGS;
1454 oa->o_valid |= OBD_MD_FLALLQUOTA;
1458 * Update LVB after writing finish for server lock, see
1459 * comments in ldlm_lock_decref_internal(), If this is a
1460 * local lock on a server namespace and this was the last
1461 * reference, lock will be destroyed directly thus there
1462 * is no chance for ldlm_request_cancel() to update lvb.
1464 if (rc == 0 && (rnb[0].rnb_flags & OBD_BRW_SRVLOCK)) {
1465 ost_fid_build_resid(fid, &info->fti_resid);
1466 rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
1469 ldlm_res_lvbo_update(rs, NULL, 1);
1470 ldlm_resource_putref(rs);
1474 /* Convert back to client IDs. LU-9671.
1475 * nodemap_get_from_exp() may fail due to nodemap deactivated,
1476 * server ID will be returned back to client in that case. */
1477 nodemap = nodemap_get_from_exp(exp);
1478 if (nodemap != NULL && !IS_ERR(nodemap)) {
1479 oa->o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
1480 NODEMAP_FS_TO_CLIENT,
1482 oa->o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
1483 NODEMAP_FS_TO_CLIENT,
1485 nodemap_putref(nodemap);
1487 } else if (cmd == OBD_BRW_READ) {
1488 rc = ofd_commitrw_read(env, ofd, fid, objcount,