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;
576 LASSERT(env != NULL);
578 fo = ofd_object_find(env, ofd, fid);
583 ofd_info(env)->fti_obj = fo;
585 if (oa->o_valid & OBD_MD_FLATIME)
586 ofd_handle_atime(env, ofd, fo, oa->o_atime);
588 ofd_read_lock(env, fo);
589 if (!ofd_object_exists(fo))
590 GOTO(unlock, rc = -ENOENT);
592 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
593 rc = ofd_verify_ff(env, fo, oa);
598 if (ptlrpc_connection_is_local(exp->exp_connection))
599 dbt |= DT_BUFS_TYPE_LOCAL;
601 for (*nr_local = 0, i = 0, j = 0; i < niocount; i++) {
603 if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
604 rnb[i].rnb_len = 100 * 1024 * 1024;
606 rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
607 lnb + j, maxlnb, dbt);
608 if (unlikely(rc < 0))
610 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
611 /* correct index for local buffers to continue with */
615 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
616 tot_bytes += rnb[i].rnb_len;
619 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
620 rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
624 ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes);
628 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
630 ofd_read_unlock(env, fo);
631 ofd_object_put(env, fo);
636 * Prepare buffers for write request processing.
638 * This function converts remote buffers from client to local buffers
639 * and prepares the latter. If there is recovery in progress and required
640 * object is missing then it can be re-created before write.
642 * \param[in] env execution environment
643 * \param[in] exp OBD export of client
644 * \param[in] ofd OFD device
645 * \param[in] fid FID of object
646 * \param[in] la object attributes
647 * \param[in] oa OBDO structure from client
648 * \param[in] objcount always 1
649 * \param[in] obj object data
650 * \param[in] rnb remote buffers
651 * \param[in] nr_local number of local buffers
652 * \param[in] lnb local buffers
653 * \param[in] jobid job ID name
655 * \retval 0 on successful prepare
656 * \retval negative value on error
658 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
659 struct ofd_device *ofd, const struct lu_fid *fid,
660 struct lu_attr *la, struct obdo *oa,
661 int objcount, struct obd_ioobj *obj,
662 struct niobuf_remote *rnb, int *nr_local,
663 struct niobuf_local *lnb, char *jobid)
665 struct ofd_object *fo;
666 int i, j, k, rc = 0, tot_bytes = 0;
667 enum dt_bufs_type dbt = DT_BUFS_TYPE_WRITE;
668 int maxlnb = *nr_local;
671 LASSERT(env != NULL);
672 LASSERT(objcount == 1);
674 if (unlikely(exp->exp_obd->obd_recovering)) {
675 u64 seq = fid_seq(fid);
676 u64 oid = fid_oid(fid);
677 struct ofd_seq *oseq;
679 oseq = ofd_seq_load(env, ofd, seq);
681 CERROR("%s: Can't find FID Sequence %#llx: rc = %d\n",
682 ofd_name(ofd), seq, (int)PTR_ERR(oseq));
683 GOTO(out, rc = -EINVAL);
686 if (oid > ofd_seq_last_oid(oseq)) {
690 mutex_lock(&oseq->os_create_lock);
691 diff = oid - ofd_seq_last_oid(oseq);
693 /* Do sync create if the seq is about to used up */
694 if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) {
695 if (unlikely(oid >= IDIF_MAX_OID - 1))
697 } else if (fid_seq_is_norm(seq)) {
699 LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
702 CERROR("%s : invalid o_seq "DOSTID"\n",
703 ofd_name(ofd), POSTID(&oa->o_oi));
704 mutex_unlock(&oseq->os_create_lock);
705 ofd_seq_put(env, oseq);
706 GOTO(out, rc = -EINVAL);
710 u64 next_id = ofd_seq_last_oid(oseq) + 1;
711 int count = ofd_precreate_batch(ofd, diff);
713 rc = ofd_precreate_objects(env, ofd, next_id,
716 mutex_unlock(&oseq->os_create_lock);
717 ofd_seq_put(env, oseq);
724 mutex_unlock(&oseq->os_create_lock);
727 ofd_seq_put(env, oseq);
730 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
731 * space back if possible, we have to do this outside of the lock as
732 * grant preparation may need to sync whole fs thus wait for all the
733 * transactions to complete. */
734 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
736 fo = ofd_object_find(env, ofd, fid);
738 GOTO(out, rc = PTR_ERR(fo));
741 ofd_info(env)->fti_obj = fo;
743 ofd_read_lock(env, fo);
744 if (!ofd_object_exists(fo)) {
745 CERROR("%s: BRW to missing obj "DOSTID"\n",
746 exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
747 ofd_read_unlock(env, fo);
748 ofd_object_put(env, fo);
749 GOTO(out, rc = -ENOENT);
752 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
753 rc = ofd_verify_ff(env, fo, oa);
755 ofd_read_unlock(env, fo);
756 ofd_object_put(env, fo);
761 /* need to verify layout version */
762 if (oa->o_valid & OBD_MD_LAYOUT_VERSION) {
763 rc = ofd_verify_layout_version(env, fo, oa);
765 ofd_read_unlock(env, fo);
766 ofd_object_put(env, fo);
770 oa->o_valid &= ~OBD_MD_LAYOUT_VERSION;
773 if (ptlrpc_connection_is_local(exp->exp_connection))
774 dbt |= DT_BUFS_TYPE_LOCAL;
776 /* parse remote buffers to local buffers and prepare the latter */
777 for (*nr_local = 0, i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
778 if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
779 rnb[i].rnb_len += PAGE_SIZE;
780 rc = dt_bufs_get(env, ofd_object_child(fo),
781 rnb + i, lnb + j, maxlnb, dbt);
782 if (unlikely(rc < 0))
784 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
785 /* correct index for local buffers to continue with */
786 for (k = 0; k < rc; k++) {
787 lnb[j+k].lnb_flags = rnb[i].rnb_flags;
788 lnb[j+k].lnb_flags &= ~OBD_BRW_LOCALS;
789 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
790 lnb[j+k].lnb_rc = -ENOSPC;
795 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
796 tot_bytes += rnb[i].rnb_len;
798 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
800 rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
801 if (unlikely(rc != 0))
804 ofd_read_unlock(env, fo);
805 ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes);
808 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
809 ofd_read_unlock(env, fo);
810 ofd_object_put(env, fo);
811 /* tgt_grant_prepare_write() was called, so we must commit */
812 tgt_grant_commit(exp, oa->o_grant_used, rc);
814 /* let's still process incoming grant information packed in the oa,
815 * but without enforcing grant since we won't proceed with the write.
816 * Just like a read request actually. */
817 tgt_grant_prepare_read(env, exp, oa);
822 * Prepare bulk IO requests for processing.
824 * This function does initial checks of IO and calls corresponding
825 * functions for read/write processing.
827 * \param[in] env execution environment
828 * \param[in] cmd IO type (read/write)
829 * \param[in] exp OBD export of client
830 * \param[in] oa OBDO structure from request
831 * \param[in] objcount always 1
832 * \param[in] obj object data
833 * \param[in] rnb remote buffers
834 * \param[in] nr_local number of local buffers
835 * \param[in] lnb local buffers
837 * \retval 0 on successful prepare
838 * \retval negative value on error
840 int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
841 struct obdo *oa, int objcount, struct obd_ioobj *obj,
842 struct niobuf_remote *rnb, int *nr_local,
843 struct niobuf_local *lnb)
845 struct tgt_session_info *tsi = tgt_ses_info(env);
846 struct ofd_device *ofd = ofd_exp(exp);
847 struct ofd_thread_info *info;
849 const struct lu_fid *fid = &oa->o_oi.oi_fid;
852 if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
853 CERROR("%s: bulk has too many pages %d, which exceeds the maximum pages per RPC of %d\n",
854 exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
858 if (tgt_ses_req(tsi) == NULL) { /* echo client case */
859 info = ofd_info_init(env, exp);
862 info = tsi2ofd_info(tsi);
863 jobid = tsi->tsi_jobid;
868 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) {
869 struct ofd_seq *oseq;
871 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
873 CERROR("%s: Can not find seq for "DOSTID
874 ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
879 if (oseq->os_destroys_in_progress == 0) {
880 /* don't fail lookups for orphan recovery, it causes
881 * later LBUGs when objects still exist during
883 ofd_seq_put(env, oseq);
886 ofd_seq_put(env, oseq);
889 LASSERT(objcount == 1);
890 LASSERT(obj->ioo_bufcnt > 0);
892 if (cmd == OBD_BRW_WRITE) {
893 la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
894 rc = ofd_preprw_write(env, exp, ofd, fid, &info->fti_attr, oa,
895 objcount, obj, rnb, nr_local, lnb, jobid);
896 } else if (cmd == OBD_BRW_READ) {
897 tgt_grant_prepare_read(env, exp, oa);
898 rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa,
899 obj->ioo_bufcnt, rnb, nr_local, lnb,
902 CERROR("%s: wrong cmd %d received!\n",
903 exp->exp_obd->obd_name, cmd);
910 * Drop reference on local buffers for read bulk IO.
912 * This will free all local buffers use by this read request.
914 * \param[in] env execution environment
915 * \param[in] ofd OFD device
916 * \param[in] fid object FID
917 * \param[in] objcount always 1
918 * \param[in] niocount number of local buffers
919 * \param[in] lnb local buffers
921 * \retval 0 on successful execution
922 * \retval negative value on error
925 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
926 const struct lu_fid *fid, int objcount, int niocount,
927 struct niobuf_local *lnb)
929 struct ofd_object *fo;
933 LASSERT(niocount > 0);
935 fo = ofd_info(env)->fti_obj;
937 LASSERT(ofd_object_exists(fo));
938 dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
940 ofd_read_unlock(env, fo);
941 ofd_object_put(env, fo);
947 * Set attributes of object during write bulk IO processing.
949 * Change object attributes and write parent FID into extended
950 * attributes when needed.
952 * \param[in] env execution environment
953 * \param[in] ofd OFD device
954 * \param[in] ofd_obj OFD object
955 * \param[in] la object attributes
958 * \retval 0 on successful attributes update
959 * \retval negative value on error
962 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
963 struct ofd_object *ofd_obj, struct lu_attr *la,
966 struct ofd_thread_info *info = ofd_info(env);
967 struct filter_fid *ff = &info->fti_mds_fid;
968 __u64 valid = la->la_valid;
970 struct dt_object *dt_obj;
978 dt_obj = ofd_object_child(ofd_obj);
979 LASSERT(dt_obj != NULL);
981 la->la_valid &= LA_UID | LA_GID | LA_PROJID;
983 rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
987 if (!la->la_valid && !(oa->o_valid &
988 (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
989 /* no attributes to set */
992 th = ofd_trans_create(env, ofd);
994 GOTO(out, rc = PTR_ERR(th));
997 rc = dt_declare_attr_set(env, dt_obj, la, th);
1002 if (oa->o_valid & (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT |
1003 OBD_MD_LAYOUT_VERSION)) {
1004 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
1005 XATTR_NAME_FID, 0, th);
1009 /* We don't need a transno for this operation which will be re-executed
1010 * anyway when the OST_WRITE (with a transno assigned) is replayed */
1011 rc = dt_trans_start_local(env, ofd->ofd_osd , th);
1015 ofd_read_lock(env, ofd_obj);
1017 rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
1019 GOTO(out_unlock, rc);
1021 if (!la->la_valid && !(oa->o_valid &
1022 (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
1023 /* no attributes to set */
1024 GOTO(out_unlock, rc = 0);
1028 /* set uid/gid/projid */
1030 rc = dt_attr_set(env, dt_obj, la, th);
1032 GOTO(out_unlock, rc);
1035 fl = ofd_object_ff_update(env, ofd_obj, oa, ff);
1037 GOTO(out_unlock, rc = fl);
1039 /* set filter fid EA.
1040 * FIXME: it holds read lock of ofd object to modify the XATTR_NAME_FID
1041 * while the write lock should be held. However, it should work because
1042 * write RPCs only modify ff_{parent,layout} and those information will
1043 * be the same from all the write RPCs. The reason that fl is not used
1044 * in dt_xattr_set() is to allow this race. */
1045 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
1046 GOTO(out_unlock, rc);
1047 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
1048 ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
1049 else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
1050 le32_add_cpu(&ff->ff_parent.f_oid, -1);
1052 info->fti_buf.lb_buf = ff;
1053 info->fti_buf.lb_len = sizeof(*ff);
1054 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, 0, th);
1056 filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff));
1058 GOTO(out_unlock, rc);
1061 ofd_read_unlock(env, ofd_obj);
1063 dt_trans_stop(env, ofd->ofd_osd, th);
1065 la->la_valid = valid;
1069 struct ofd_soft_sync_callback {
1070 struct dt_txn_commit_cb ossc_cb;
1071 struct obd_export *ossc_exp;
1075 * Callback function for "soft sync" update.
1077 * Reset fed_soft_sync_count upon committing the "soft_sync" update.
1078 * See ofd_soft_sync_cb_add() below for more details on soft sync.
1080 * \param[in] env execution environment
1081 * \param[in] th transaction handle
1082 * \param[in] cb callback data
1083 * \param[in] err error code
1085 static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th,
1086 struct dt_txn_commit_cb *cb, int err)
1088 struct ofd_soft_sync_callback *ossc;
1090 ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb);
1092 CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp);
1093 atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0);
1095 class_export_cb_put(ossc->ossc_exp);
1100 * Add callback for "soft sync" processing.
1102 * The "soft sync" mechanism does asynchronous commit when OBD_BRW_SOFT_SYNC
1103 * flag is set in client buffers. The intention is for this operation to
1104 * commit pages belonging to a client which has "too many" outstanding
1105 * unstable pages in its cache. See LU-2139 for details.
1107 * This function adds callback to be called when commit is done.
1109 * \param[in] th transaction handle
1110 * \param[in] exp OBD export of client
1112 * \retval 0 on successful callback adding
1113 * \retval negative value on error
1115 static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
1117 struct ofd_soft_sync_callback *ossc;
1118 struct dt_txn_commit_cb *dcb;
1121 OBD_ALLOC_PTR(ossc);
1125 ossc->ossc_exp = class_export_cb_get(exp);
1127 dcb = &ossc->ossc_cb;
1128 dcb->dcb_func = ofd_cb_soft_sync;
1129 INIT_LIST_HEAD(&dcb->dcb_linkage);
1130 strlcpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name));
1132 rc = dt_trans_cb_add(th, dcb);
1134 class_export_cb_put(exp);
1142 * Commit bulk IO buffers to the storage.
1144 * This function finalizes write IO processing by writing data to the disk.
1145 * That write can be synchronous or asynchronous depending on buffers flags.
1147 * \param[in] env execution environment
1148 * \param[in] exp OBD export of client
1149 * \param[in] ofd OFD device
1150 * \param[in] fid FID of object
1151 * \param[in] la object attributes
1152 * \param[in] ff parent FID of object
1153 * \param[in] objcount always 1
1154 * \param[in] niocount number of local buffers
1155 * \param[in] lnb local buffers
1156 * \param[in] granted grant space consumed for the bulk I/O
1157 * \param[in] old_rc result of processing at this point
1159 * \retval 0 on successful commit
1160 * \retval negative value on error
1163 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
1164 struct ofd_device *ofd, const struct lu_fid *fid,
1165 struct lu_attr *la, struct obdo *oa, int objcount,
1166 int niocount, struct niobuf_local *lnb,
1167 unsigned long granted, int old_rc)
1169 struct ofd_thread_info *info = ofd_info(env);
1170 struct filter_export_data *fed = &exp->exp_filter_data;
1171 struct ofd_object *fo;
1172 struct dt_object *o;
1178 bool soft_sync = false;
1179 bool cb_registered = false;
1180 bool fake_write = false;
1184 LASSERT(objcount == 1);
1186 fo = ofd_info(env)->fti_obj;
1187 LASSERT(fo != NULL);
1189 o = ofd_object_child(fo);
1193 GOTO(out, rc = old_rc);
1194 if (!ofd_object_exists(fo))
1195 GOTO(out, rc = -ENOENT);
1198 * The first write to each object must set some attributes. It is
1199 * important to set the uid/gid before calling
1200 * dt_declare_write_commit() since quota enforcement is now handled in
1203 rc = ofd_write_attr_set(env, ofd, fo, la, oa);
1207 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
1209 /* do fake write, to simulate the write case for performance testing */
1210 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW)) {
1211 struct niobuf_local *last = &lnb[niocount - 1];
1212 __u64 file_size = last->lnb_file_offset + last->lnb_len;
1213 __u64 valid = la->la_valid;
1215 la->la_valid = LA_SIZE;
1217 rc = dt_attr_get(env, o, la);
1218 if (rc < 0 && rc != -ENOENT)
1221 if (file_size < la->la_size)
1222 file_size = la->la_size;
1224 /* dirty inode by setting file size */
1225 la->la_valid = valid | LA_SIZE;
1226 la->la_size = file_size;
1232 th = ofd_trans_create(env, ofd);
1234 GOTO(out, rc = PTR_ERR(th));
1236 th->th_sync |= ofd->ofd_sync_journal;
1237 if (th->th_sync == 0) {
1238 for (i = 0; i < niocount; i++) {
1239 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
1243 if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC)
1248 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
1249 GOTO(out_stop, rc = -EINPROGRESS);
1251 if (likely(!fake_write)) {
1252 rc = dt_declare_write_commit(env, o, lnb, niocount, th);
1257 /* don't update atime on disk if it is older */
1258 if (la->la_valid & LA_ATIME && la->la_atime <= fo->ofo_atime_ondisk)
1259 la->la_valid &= ~LA_ATIME;
1262 /* update [mac]time if needed */
1263 rc = dt_declare_attr_set(env, o, la, th);
1268 rc = ofd_trans_start(env, ofd, fo, th);
1272 ofd_read_lock(env, fo);
1273 if (!ofd_object_exists(fo))
1274 GOTO(out_unlock, rc = -ENOENT);
1276 if (likely(!fake_write)) {
1277 rc = dt_write_commit(env, o, lnb, niocount, th);
1279 GOTO(out_unlock, rc);
1282 /* Don't update timestamps if this write is older than a
1283 * setattr which modifies the timestamps. b=10150 */
1284 if (la->la_valid && tgt_fmd_check(exp, fid, info->fti_xid)) {
1285 rc = dt_attr_set(env, o, la, th);
1287 GOTO(out_unlock, rc);
1288 if (la->la_valid & LA_ATIME)
1289 fo->ofo_atime_ondisk = la->la_atime;
1292 /* get attr to return */
1293 rc = dt_attr_get(env, o, la);
1296 ofd_read_unlock(env, fo);
1298 /* Force commit to make the just-deleted blocks
1299 * reusable. LU-456 */
1303 /* do this before trans stop in case commit has finished */
1304 if (!th->th_sync && soft_sync && !cb_registered) {
1305 ofd_soft_sync_cb_add(th, exp);
1306 cb_registered = true;
1309 if (rc == 0 && granted > 0) {
1310 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
1314 rc2 = ofd_trans_stop(env, ofd, th, rc);
1317 if (rc == -ENOSPC && retries++ < 3) {
1318 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
1324 /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
1325 atomic_set(&fed->fed_soft_sync_count, 0);
1326 else if (atomic_inc_return(&fed->fed_soft_sync_count) ==
1327 ofd->ofd_soft_sync_limit)
1328 dt_commit_async(env, ofd->ofd_osd);
1331 dt_bufs_put(env, o, lnb, niocount);
1332 ofd_object_put(env, fo);
1334 tgt_grant_commit(exp, granted, old_rc);
1339 * Commit bulk IO to the storage.
1341 * This is companion function to the ofd_preprw(). It finishes bulk IO
1342 * request processing by committing buffers to the storage (WRITE) and/or
1343 * freeing those buffers (read/write). See ofd_commitrw_read() and
1344 * ofd_commitrw_write() for details about each type of IO.
1346 * \param[in] env execution environment
1347 * \param[in] cmd IO type (READ/WRITE)
1348 * \param[in] exp OBD export of client
1349 * \param[in] oa OBDO structure from client
1350 * \param[in] objcount always 1
1351 * \param[in] obj object data
1352 * \param[in] rnb remote buffers
1353 * \param[in] npages number of local buffers
1354 * \param[in] lnb local buffers
1355 * \param[in] old_rc result of processing at this point
1357 * \retval 0 on successful commit
1358 * \retval negative value on error
1360 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
1361 struct obdo *oa, int objcount, struct obd_ioobj *obj,
1362 struct niobuf_remote *rnb, int npages,
1363 struct niobuf_local *lnb, int old_rc)
1365 struct ofd_thread_info *info = ofd_info(env);
1366 struct ofd_device *ofd = ofd_exp(exp);
1367 const struct lu_fid *fid = &oa->o_oi.oi_fid;
1368 struct ldlm_namespace *ns = ofd->ofd_namespace;
1369 struct ldlm_resource *rs = NULL;
1373 LASSERT(npages > 0);
1375 if (cmd == OBD_BRW_WRITE) {
1376 struct lu_nodemap *nodemap;
1378 valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID |
1379 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
1380 la_from_obdo(&info->fti_attr, oa, valid);
1382 rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
1383 oa, objcount, npages, lnb,
1384 oa->o_grant_used, old_rc);
1386 obdo_from_la(oa, &info->fti_attr,
1387 OFD_VALID_FLAGS | LA_GID | LA_UID |
1390 obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID |
1393 /* don't report overquota flag if we failed before reaching
1395 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
1396 /* return the overquota flags to client */
1397 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
1398 if (oa->o_valid & OBD_MD_FLFLAGS)
1399 oa->o_flags |= OBD_FL_NO_USRQUOTA;
1401 oa->o_flags = OBD_FL_NO_USRQUOTA;
1404 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
1405 if (oa->o_valid & OBD_MD_FLFLAGS)
1406 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
1408 oa->o_flags = OBD_FL_NO_GRPQUOTA;
1410 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
1411 if (oa->o_valid & OBD_MD_FLFLAGS)
1412 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
1414 oa->o_flags = OBD_FL_NO_PRJQUOTA;
1417 oa->o_valid |= OBD_MD_FLFLAGS;
1418 oa->o_valid |= OBD_MD_FLALLQUOTA;
1422 * Update LVB after writing finish for server lock, see
1423 * comments in ldlm_lock_decref_internal(), If this is a
1424 * local lock on a server namespace and this was the last
1425 * reference, lock will be destroyed directly thus there
1426 * is no chance for ldlm_request_cancel() to update lvb.
1428 if (rc == 0 && (rnb[0].rnb_flags & OBD_BRW_SRVLOCK)) {
1429 ost_fid_build_resid(fid, &info->fti_resid);
1430 rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
1433 ldlm_res_lvbo_update(rs, NULL, 1);
1434 ldlm_resource_putref(rs);
1438 /* Convert back to client IDs. LU-9671.
1439 * nodemap_get_from_exp() may fail due to nodemap deactivated,
1440 * server ID will be returned back to client in that case. */
1441 nodemap = nodemap_get_from_exp(exp);
1442 if (nodemap != NULL && !IS_ERR(nodemap)) {
1443 oa->o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
1444 NODEMAP_FS_TO_CLIENT,
1446 oa->o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
1447 NODEMAP_FS_TO_CLIENT,
1449 nodemap_putref(nodemap);
1451 } else if (cmd == OBD_BRW_READ) {
1452 rc = ofd_commitrw_read(env, ofd, fid, objcount,