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/
33 * This file provides functions to handle IO requests from clients and
34 * also LFSCK routines to check parent file identifier (PFID) consistency.
36 * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
37 * Author: Fan Yong <fan.yong@intel.com>
40 #define DEBUG_SUBSYSTEM S_FILTER
42 #include <linux/kthread.h>
43 #include "ofd_internal.h"
44 #include <lustre_nodemap.h>
46 struct ofd_inconsistency_item {
47 struct list_head oii_list;
48 struct ofd_object *oii_obj;
49 struct filter_fid oii_ff;
53 * Verify single object for parent FID consistency.
55 * Part of LFSCK processing which checks single object PFID stored in extended
56 * attribute (XATTR) against real FID of MDT parent object received by LFSCK.
57 * This verifies that the OST object is being referenced by only a single MDT
60 * \param[in] env execution environment
61 * \param[in] ofd OFD device
62 * \param[in] oii object-related local data
63 * \param[in] lrl LFSCK request data
65 static void ofd_inconsistency_verify_one(const struct lu_env *env,
66 struct ofd_device *ofd,
67 struct ofd_inconsistency_item *oii,
68 struct lfsck_req_local *lrl)
70 struct ofd_object *fo = oii->oii_obj;
71 struct filter_fid *client_ff = &oii->oii_ff;
72 struct filter_fid *local_ff = &fo->ofo_ff;
75 LASSERT(fo->ofo_pfid_checking);
76 LASSERT(!fo->ofo_pfid_verified);
78 lrl->lrl_fid = fo->ofo_header.loh_fid; /* OST-object itself FID. */
79 lrl->lrl_ff_client = *client_ff; /* client given PFID. */
80 lrl->lrl_ff_local = *local_ff; /* OST local stored PFID. */
82 rc = lfsck_in_notify_local(env, ofd->ofd_osd, lrl, NULL);
83 ofd_write_lock(env, fo);
84 switch (lrl->lrl_status) {
89 CDEBUG(D_LFSCK, "%s: fail to verify OST local stored "
90 "PFID xattr for "DFID", the client given PFID "
91 DFID", OST local stored PFID "DFID": rc = %d\n",
92 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
93 PFID(&client_ff->ff_parent),
94 PFID(&local_ff->ff_parent), rc);
96 fo->ofo_pfid_verified = 1;
98 case LPVS_INCONSISTENT:
101 ofd->ofd_inconsistency_self_detected++;
103 CDEBUG(D_LFSCK, "%s: fail to verify the client given "
104 "PFID for "DFID", the client given PFID "DFID
105 ", local stored PFID "DFID": rc = %d\n",
106 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
107 PFID(&client_ff->ff_parent),
108 PFID(&local_ff->ff_parent), rc);
110 CDEBUG(D_LFSCK, "%s: both the client given PFID and "
111 "the OST local stored PFID are stale for the "
112 "OST-object "DFID", client given PFID is "DFID
113 ", local stored PFID is "DFID"\n",
114 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
115 PFID(&client_ff->ff_parent),
116 PFID(&local_ff->ff_parent));
118 case LPVS_INCONSISTENT_TOFIX:
119 ofd->ofd_inconsistency_self_detected++;
121 ofd->ofd_inconsistency_self_repaired++;
122 CDEBUG(D_LFSCK, "%s: fixed the staled OST PFID xattr "
123 "for "DFID", with the client given PFID "DFID
124 ", the old stored PFID "DFID"\n",
125 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
126 PFID(&client_ff->ff_parent),
127 PFID(&local_ff->ff_parent));
129 CDEBUG(D_LFSCK, "%s: fail to fix the OST PFID xattr "
130 "for "DFID", client given PFID "DFID", local "
131 "stored PFID "DFID": rc = %d\n",
132 ofd_name(ofd), PFID(&fo->ofo_header.loh_fid),
133 PFID(&client_ff->ff_parent),
134 PFID(&local_ff->ff_parent), rc);
136 local_ff->ff_parent = client_ff->ff_parent;
137 fo->ofo_pfid_verified = 1;
142 fo->ofo_pfid_checking = 0;
143 ofd_write_unlock(env, fo);
145 ofd_object_put(env, fo);
150 struct ofd_device *od_ofd;
151 struct lu_env od_env;
152 struct lfsck_req_local od_lrl;
153 struct completion *od_started;
157 #define TASK_IDLE TASK_INTERRUPTIBLE
161 * Verification thread to check parent FID consistency.
163 * Kernel thread to check consistency of parent FID for any
164 * new item added for checking by ofd_add_inconsistency_item().
166 * \param[in] args OFD device
168 * \retval 0 on successful thread termination
169 * \retval negative value if thread can't start
171 static int ofd_inconsistency_verification_main(void *_args)
173 struct oivm_args *args = _args;
174 struct lu_env *env = &args->od_env;
175 struct ofd_device *ofd = args->od_ofd;
176 struct ofd_inconsistency_item *oii;
177 struct lfsck_req_local *lrl = &args->od_lrl;
180 lrl->lrl_event = LEL_PAIRS_VERIFY_LOCAL;
181 lrl->lrl_active = LFSCK_TYPE_LAYOUT;
182 complete(args->od_started);
184 spin_lock(&ofd->ofd_inconsistency_lock);
185 while (({set_current_state(TASK_IDLE);
186 !kthread_should_stop(); })) {
188 while (!list_empty(&ofd->ofd_inconsistency_list)) {
189 __set_current_state(TASK_RUNNING);
190 oii = list_entry(ofd->ofd_inconsistency_list.next,
191 struct ofd_inconsistency_item,
193 list_del_init(&oii->oii_list);
194 spin_unlock(&ofd->ofd_inconsistency_lock);
195 ofd_inconsistency_verify_one(env, ofd, oii, lrl);
196 spin_lock(&ofd->ofd_inconsistency_lock);
199 spin_unlock(&ofd->ofd_inconsistency_lock);
201 spin_lock(&ofd->ofd_inconsistency_lock);
203 __set_current_state(TASK_RUNNING);
205 while (!list_empty(&ofd->ofd_inconsistency_list)) {
206 struct ofd_object *fo;
208 oii = list_entry(ofd->ofd_inconsistency_list.next,
209 struct ofd_inconsistency_item,
211 list_del_init(&oii->oii_list);
213 spin_unlock(&ofd->ofd_inconsistency_lock);
215 ofd_write_lock(env, fo);
216 fo->ofo_pfid_checking = 0;
217 ofd_write_unlock(env, fo);
219 ofd_object_put(env, fo);
221 spin_lock(&ofd->ofd_inconsistency_lock);
224 spin_unlock(&ofd->ofd_inconsistency_lock);
226 lu_env_fini(&args->od_env);
232 * Start parent FID verification thread.
234 * See ofd_inconsistency_verification_main().
236 * \param[in] ofd OFD device
238 * \retval 0 on successful start of thread
239 * \retval negative value on error
241 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd)
243 struct task_struct *task;
244 struct oivm_args *args;
245 DECLARE_COMPLETION_ONSTACK(started);
248 if (ofd->ofd_inconsistency_task)
254 rc = lu_env_init(&args->od_env, LCT_DT_THREAD);
261 args->od_started = &started;
262 task = kthread_create(ofd_inconsistency_verification_main, args,
263 "inconsistency_verification");
266 CERROR("%s: cannot start self_repair thread: rc = %d\n",
270 spin_lock(&ofd->ofd_inconsistency_lock);
271 if (ofd->ofd_inconsistency_task)
274 ofd->ofd_inconsistency_task = task;
275 spin_unlock(&ofd->ofd_inconsistency_lock);
280 wake_up_process(task);
281 wait_for_completion(&started);
285 lu_env_fini(&args->od_env);
293 * Stop parent FID verification thread.
295 * \param[in] ofd OFD device
297 * \retval 0 on successful start of thread
298 * \retval -EALREADY if thread is already stopped
300 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd)
302 struct task_struct *task;
304 spin_lock(&ofd->ofd_inconsistency_lock);
305 task = ofd->ofd_inconsistency_task;
306 ofd->ofd_inconsistency_task = NULL;
307 spin_unlock(&ofd->ofd_inconsistency_lock);
317 * Add new item for parent FID verification.
319 * Prepare new verification item and pass it to the dedicated
320 * verification thread for further processing.
322 * \param[in] env execution environment
323 * \param[in] fo OFD object
324 * \param[in] oa OBDO structure with PFID
326 static void ofd_add_inconsistency_item(const struct lu_env *env,
327 struct ofd_object *fo, struct obdo *oa)
329 struct ofd_device *ofd = ofd_obj2dev(fo);
330 struct ofd_inconsistency_item *oii;
331 struct filter_fid *ff;
338 INIT_LIST_HEAD(&oii->oii_list);
339 lu_object_get(&fo->ofo_obj.do_lu);
342 ff->ff_parent.f_seq = oa->o_parent_seq;
343 ff->ff_parent.f_oid = oa->o_parent_oid;
344 ff->ff_parent.f_stripe_idx = oa->o_stripe_idx;
345 ff->ff_layout = oa->o_layout;
347 spin_lock(&ofd->ofd_inconsistency_lock);
348 if (fo->ofo_pfid_checking || fo->ofo_pfid_verified) {
349 spin_unlock(&ofd->ofd_inconsistency_lock);
355 fo->ofo_pfid_checking = 1;
356 if (list_empty(&ofd->ofd_inconsistency_list))
358 list_add_tail(&oii->oii_list, &ofd->ofd_inconsistency_list);
359 if (wakeup && ofd->ofd_inconsistency_task)
360 wake_up_process(ofd->ofd_inconsistency_task);
361 spin_unlock(&ofd->ofd_inconsistency_lock);
363 /* XXX: When the found inconsistency exceeds some threshold,
364 * we can trigger the LFSCK to scan part of the system
365 * or the whole system, which depends on how to define
366 * the threshold, a simple way maybe like that: define
367 * the absolute value of how many inconsisteny allowed
368 * to be repaired via self detect/repair mechanism, if
369 * exceeded, then trigger the LFSCK to scan the layout
370 * inconsistency within the whole system. */
374 * Verify parent FID of an object.
376 * Check the parent FID is sane and start extended
377 * verification procedure otherwise.
379 * \param[in] env execution environment
380 * \param[in] fo OFD object
381 * \param[in] oa OBDO structure with PFID
383 * \retval 0 on successful verification
384 * \retval -EINPROGRESS if PFID is being repaired
385 * \retval -EPERM if PFID was verified but still insane
387 int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
390 struct lu_fid *pfid = &fo->ofo_ff.ff_parent;
394 if (fid_is_sane(pfid)) {
395 if (likely(oa->o_parent_seq == pfid->f_seq &&
396 oa->o_parent_oid == pfid->f_oid &&
397 oa->o_stripe_idx == pfid->f_stripe_idx))
400 if (fo->ofo_pfid_verified)
404 /* The OST-object may be inconsistent, and we need further verification.
405 * To avoid block the RPC service thread, return -EINPROGRESS to client
406 * and make it retry later. */
407 if (fo->ofo_pfid_checking)
408 RETURN(-EINPROGRESS);
410 rc = ofd_object_ff_load(env, fo);
417 if (likely(oa->o_parent_seq == pfid->f_seq &&
418 oa->o_parent_oid == pfid->f_oid &&
419 oa->o_stripe_idx == pfid->f_stripe_idx))
422 /* Push it to the dedicated thread for further verification. */
423 ofd_add_inconsistency_item(env, fo, oa);
425 RETURN(-EINPROGRESS);
429 * FLR: verify the layout version of object.
431 * \param[in] env execution environment
432 * \param[in] fo OFD object
433 * \param[in] oa OBDO structure with layout version
435 * \retval 0 on successful verification
436 * \retval -EINPROGRESS layout version is in transfer
437 * \retval -ESTALE the layout version on client is stale
439 int ofd_verify_layout_version(const struct lu_env *env,
440 struct ofd_object *fo, const struct obdo *oa)
445 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_SKIP_LV_CHECK)))
448 rc = ofd_object_ff_load(env, fo);
456 * this update is not legitimate, whose layout version is older than
459 if (ofd_layout_version_less(oa->o_layout_version,
460 fo->ofo_ff.ff_layout_version))
461 GOTO(out, rc = -ESTALE);
464 CDEBUG(D_INODE, DFID " verify layout version: %u vs. %u/%u: rc = %d\n",
465 PFID(lu_object_fid(&fo->ofo_obj.do_lu)),
466 oa->o_layout_version, fo->ofo_ff.ff_layout_version,
467 fo->ofo_ff.ff_range, rc);
473 * Lazy ATIME update to refresh atime every ofd_atime_diff
474 * seconds so that external scanning tool can see it actual
475 * within that period and be able to identify accessed files
477 static void ofd_handle_atime(const struct lu_env *env, struct ofd_device *ofd,
478 struct ofd_object *fo, time64_t atime)
485 if (ofd->ofd_atime_diff == 0)
488 la = &ofd_info(env)->fti_attr2;
489 o = ofd_object_child(fo);
491 if (unlikely(fo->ofo_atime_ondisk == 0)) {
492 rc = dt_attr_get(env, o, la);
495 LASSERT(la->la_valid & LA_ATIME);
496 if (la->la_atime == 0)
497 la->la_atime = la->la_mtime;
498 fo->ofo_atime_ondisk = la->la_atime;
500 if (atime - fo->ofo_atime_ondisk < ofd->ofd_atime_diff)
503 /* atime hasn't been updated too long, update it */
504 fo->ofo_atime_ondisk = atime;
506 th = ofd_trans_create(env, ofd);
508 CERROR("%s: cannot create transaction: rc = %d\n",
509 ofd_name(ofd), (int)PTR_ERR(th));
513 la->la_valid = LA_ATIME;
514 rc = dt_declare_attr_set(env, o, la, th);
518 rc = dt_trans_start_local(env, ofd->ofd_osd , th);
520 CERROR("%s: cannot start transaction: rc = %d\n",
525 ofd_read_lock(env, fo);
526 if (ofd_object_exists(fo)) {
527 la->la_atime = fo->ofo_atime_ondisk;
528 rc = dt_attr_set(env, o, la, th);
531 ofd_read_unlock(env, fo);
534 ofd_trans_stop(env, ofd, th, rc);
538 * Prepare buffers for read request processing.
540 * This function converts remote buffers from client to local buffers
541 * and prepares the latter.
543 * \param[in] env execution environment
544 * \param[in] exp OBD export of client
545 * \param[in] ofd OFD device
546 * \param[in] fid FID of object
547 * \param[in] la object attributes
548 * \param[in] oa OBDO structure from client
549 * \param[in] niocount number of remote buffers
550 * \param[in] rnb remote buffers
551 * \param[in] nr_local number of local buffers
552 * \param[in] lnb local buffers
553 * \param[in] jobid job ID name
555 * \retval 0 on successful prepare
556 * \retval negative value on error
558 static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
559 struct ofd_device *ofd, const struct lu_fid *fid,
560 struct lu_attr *la, struct obdo *oa, int niocount,
561 struct niobuf_remote *rnb, int *nr_local,
562 struct niobuf_local *lnb)
564 struct ofd_object *fo;
565 int i, j, rc, tot_bytes = 0;
566 enum dt_bufs_type dbt = DT_BUFS_TYPE_READ;
567 int maxlnb = *nr_local;
571 LASSERT(env != NULL);
573 fo = ofd_object_find(env, ofd, fid);
578 ofd_info(env)->fti_obj = fo;
580 if (oa->o_valid & OBD_MD_FLATIME)
581 ofd_handle_atime(env, ofd, fo, oa->o_atime);
583 if (!ofd_object_exists(fo))
584 GOTO(obj_put, rc = -ENOENT);
586 if (ptlrpc_connection_is_local(exp->exp_connection))
587 dbt |= DT_BUFS_TYPE_LOCAL;
592 for (*nr_local = 0, i = 0, j = 0; i < niocount; i++) {
593 begin = min_t(__u64, begin, rnb[i].rnb_offset);
594 end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
596 if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
597 rnb[i].rnb_len = 100 * 1024 * 1024;
599 rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
600 lnb + j, maxlnb, dbt);
601 if (unlikely(rc < 0))
603 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
604 /* correct index for local buffers to continue with */
608 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
609 tot_bytes += rnb[i].rnb_len;
612 ofd_read_lock(env, fo);
613 if (!ofd_object_exists(fo))
614 GOTO(unlock, rc = -ENOENT);
616 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
617 rc = ofd_verify_ff(env, fo, oa);
622 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
623 rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
626 ofd_read_unlock(env, fo);
630 .f_seq = oa->o_parent_seq,
631 .f_oid = oa->o_parent_oid,
632 .f_ver = oa->o_stripe_idx,
642 ofd_read_unlock(env, fo);
644 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
646 ofd_object_put(env, fo);
651 * Prepare buffers for write request processing.
653 * This function converts remote buffers from client to local buffers
654 * and prepares the latter. If there is recovery in progress and required
655 * object is missing then it can be re-created before write.
657 * \param[in] env execution environment
658 * \param[in] exp OBD export of client
659 * \param[in] ofd OFD device
660 * \param[in] fid FID of object
661 * \param[in] la object attributes
662 * \param[in] oa OBDO structure from client
663 * \param[in] objcount always 1
664 * \param[in] obj object data
665 * \param[in] rnb remote buffers
666 * \param[in] nr_local number of local buffers
667 * \param[in] lnb local buffers
668 * \param[in] jobid job ID name
670 * \retval 0 on successful prepare
671 * \retval negative value on error
673 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
674 struct ofd_device *ofd, const struct lu_fid *fid,
675 struct lu_attr *la, struct obdo *oa,
676 int objcount, struct obd_ioobj *obj,
677 struct niobuf_remote *rnb, int *nr_local,
678 struct niobuf_local *lnb)
680 struct ofd_object *fo;
681 int i, j, k, rc = 0, tot_bytes = 0;
682 enum dt_bufs_type dbt = DT_BUFS_TYPE_WRITE;
683 int maxlnb = *nr_local;
687 LASSERT(env != NULL);
688 LASSERT(objcount == 1);
690 if (unlikely(exp->exp_obd->obd_recovering)) {
691 u64 seq = ostid_seq(&oa->o_oi);
692 u64 oid = ostid_id(&oa->o_oi);
693 struct ofd_seq *oseq;
695 oseq = ofd_seq_load(env, ofd, seq);
697 CERROR("%s: Can't find FID Sequence %#llx: rc = %d\n",
698 ofd_name(ofd), seq, (int)PTR_ERR(oseq));
699 GOTO(out, rc = -EINVAL);
702 if (oid > ofd_seq_last_oid(oseq)) {
706 mutex_lock(&oseq->os_create_lock);
707 diff = oid - ofd_seq_last_oid(oseq);
709 /* Do sync create if the seq is about to used up */
710 if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) {
711 if (unlikely(oid >= IDIF_MAX_OID - 1))
713 } else if (fid_seq_is_norm(seq)) {
715 LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
718 CERROR("%s : invalid o_seq "DOSTID"\n",
719 ofd_name(ofd), POSTID(&oa->o_oi));
720 mutex_unlock(&oseq->os_create_lock);
721 ofd_seq_put(env, oseq);
722 GOTO(out, rc = -EINVAL);
726 u64 next_id = ofd_seq_last_oid(oseq) + 1;
727 int count = ofd_precreate_batch(ofd, diff);
729 rc = ofd_precreate_objects(env, ofd, next_id,
732 mutex_unlock(&oseq->os_create_lock);
733 ofd_seq_put(env, oseq);
740 mutex_unlock(&oseq->os_create_lock);
743 ofd_seq_put(env, oseq);
746 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
747 * space back if possible, we have to do this outside of the lock as
748 * grant preparation may need to sync whole fs thus wait for all the
749 * transactions to complete. */
750 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
752 fo = ofd_object_find(env, ofd, fid);
754 GOTO(out, rc = PTR_ERR(fo));
757 ofd_info(env)->fti_obj = fo;
759 if (!ofd_object_exists(fo)) {
760 CERROR("%s: BRW to missing obj "DOSTID"\n",
761 exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
762 ofd_object_put(env, fo);
763 GOTO(out, rc = -ENOENT);
766 if (ptlrpc_connection_is_local(exp->exp_connection))
767 dbt |= DT_BUFS_TYPE_LOCAL;
772 /* parse remote buffers to local buffers and prepare the latter */
773 for (*nr_local = 0, i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
774 begin = min_t(__u64, begin, rnb[i].rnb_offset);
775 end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
777 if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
778 rnb[i].rnb_len += PAGE_SIZE;
779 rc = dt_bufs_get(env, ofd_object_child(fo),
780 rnb + i, lnb + j, maxlnb, dbt);
781 if (unlikely(rc < 0))
782 GOTO(err_nolock, rc);
783 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
784 /* correct index for local buffers to continue with */
785 for (k = 0; k < rc; k++) {
786 lnb[j+k].lnb_flags = rnb[i].rnb_flags;
787 lnb[j+k].lnb_flags &= ~OBD_BRW_LOCALS;
788 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
789 lnb[j+k].lnb_rc = -ENOSPC;
794 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
795 tot_bytes += rnb[i].rnb_len;
797 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
799 ofd_read_lock(env, fo);
800 if (!ofd_object_exists(fo)) {
801 CERROR("%s: BRW to missing obj "DOSTID": rc = -ENOENT\n",
802 exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
803 GOTO(err, rc = -ENOENT);
806 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
807 rc = ofd_verify_ff(env, fo, oa);
812 /* need to verify layout version */
813 if (oa->o_valid & OBD_MD_LAYOUT_VERSION) {
814 rc = ofd_verify_layout_version(env, fo, oa);
819 rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
820 if (unlikely(rc != 0))
823 ofd_read_unlock(env, fo);
827 .f_seq = oa->o_parent_seq,
828 .f_oid = oa->o_parent_oid,
829 .f_ver = oa->o_stripe_idx,
839 ofd_read_unlock(env, fo);
841 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
842 ofd_object_put(env, fo);
843 /* tgt_grant_prepare_write() was called, so we must commit */
844 tgt_grant_commit(exp, oa->o_grant_used, rc);
846 /* let's still process incoming grant information packed in the oa,
847 * but without enforcing grant since we won't proceed with the write.
848 * Just like a read request actually. */
849 tgt_grant_prepare_read(env, exp, oa);
854 * Prepare bulk IO requests for processing.
856 * This function does initial checks of IO and calls corresponding
857 * functions for read/write processing.
859 * \param[in] env execution environment
860 * \param[in] cmd IO type (read/write)
861 * \param[in] exp OBD export of client
862 * \param[in] oa OBDO structure from request
863 * \param[in] objcount always 1
864 * \param[in] obj object data
865 * \param[in] rnb remote buffers
866 * \param[in] nr_local number of local buffers
867 * \param[in] lnb local buffers
869 * \retval 0 on successful prepare
870 * \retval negative value on error
872 int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
873 struct obdo *oa, int objcount, struct obd_ioobj *obj,
874 struct niobuf_remote *rnb, int *nr_local,
875 struct niobuf_local *lnb)
877 struct tgt_session_info *tsi = tgt_ses_info(env);
878 struct ofd_device *ofd = ofd_exp(exp);
879 struct ofd_thread_info *info;
880 const struct lu_fid *fid = &oa->o_oi.oi_fid;
883 if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
884 CERROR("%s: bulk has too many pages %d, which exceeds the maximum pages per RPC of %d\n",
885 exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
889 if (tgt_ses_req(tsi) == NULL) { /* echo client case */
890 info = ofd_info_init(env, exp);
892 info = tsi2ofd_info(tsi);
897 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) {
898 struct ofd_seq *oseq;
900 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
902 CERROR("%s: Can not find seq for "DOSTID
903 ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
908 if (oseq->os_destroys_in_progress == 0) {
909 /* don't fail lookups for orphan recovery, it causes
910 * later LBUGs when objects still exist during
912 ofd_seq_put(env, oseq);
915 ofd_seq_put(env, oseq);
918 LASSERT(objcount == 1);
919 LASSERT(obj->ioo_bufcnt > 0);
921 if (cmd == OBD_BRW_WRITE) {
922 la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
923 rc = ofd_preprw_write(env, exp, ofd, fid, &info->fti_attr, oa,
924 objcount, obj, rnb, nr_local, lnb);
925 } else if (cmd == OBD_BRW_READ) {
926 tgt_grant_prepare_read(env, exp, oa);
927 rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa,
928 obj->ioo_bufcnt, rnb, nr_local, lnb);
930 CERROR("%s: wrong cmd %d received!\n",
931 exp->exp_obd->obd_name, cmd);
938 * Drop reference on local buffers for read bulk IO.
940 * This will free all local buffers use by this read request.
942 * \param[in] env execution environment
943 * \param[in] ofd OFD device
944 * \param[in] fid object FID
945 * \param[in] objcount always 1
946 * \param[in] niocount number of local buffers
947 * \param[in] lnb local buffers
949 * \retval 0 on successful execution
950 * \retval negative value on error
953 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
954 const struct lu_fid *fid, int objcount, int niocount,
955 struct niobuf_local *lnb)
957 struct ofd_object *fo;
961 LASSERT(niocount > 0);
963 fo = ofd_info(env)->fti_obj;
965 LASSERT(ofd_object_exists(fo));
966 dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
968 ofd_object_put(env, fo);
974 * Set attributes of object during write bulk IO processing.
976 * Change object attributes and write parent FID into extended
977 * attributes when needed.
979 * \param[in] env execution environment
980 * \param[in] ofd OFD device
981 * \param[in] ofd_obj OFD object
982 * \param[in] la object attributes
985 * \retval 0 on successful attributes update
986 * \retval negative value on error
989 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
990 struct ofd_object *ofd_obj, struct lu_attr *la,
993 struct ofd_thread_info *info = ofd_info(env);
994 struct filter_fid *ff = &info->fti_mds_fid;
995 __u64 valid = la->la_valid;
997 struct dt_object *dt_obj;
1005 dt_obj = ofd_object_child(ofd_obj);
1006 LASSERT(dt_obj != NULL);
1008 la->la_valid &= LA_UID | LA_GID | LA_PROJID;
1010 rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
1014 if (!la->la_valid && !(oa->o_valid &
1015 (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
1016 /* no attributes to set */
1019 th = ofd_trans_create(env, ofd);
1021 GOTO(out, rc = PTR_ERR(th));
1024 rc = dt_declare_attr_set(env, dt_obj, la, th);
1029 if (oa->o_valid & (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT |
1030 OBD_MD_LAYOUT_VERSION)) {
1031 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
1032 XATTR_NAME_FID, 0, th);
1036 /* We don't need a transno for this operation which will be re-executed
1037 * anyway when the OST_WRITE (with a transno assigned) is replayed */
1038 rc = dt_trans_start_local(env, ofd->ofd_osd , th);
1042 ofd_read_lock(env, ofd_obj);
1044 rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
1046 GOTO(out_unlock, rc);
1048 if (!la->la_valid && !(oa->o_valid &
1049 (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
1050 /* no attributes to set */
1051 GOTO(out_unlock, rc = 0);
1055 /* set uid/gid/projid */
1057 rc = dt_attr_set(env, dt_obj, la, th);
1059 GOTO(out_unlock, rc);
1062 fl = ofd_object_ff_update(env, ofd_obj, oa, ff);
1064 GOTO(out_unlock, rc = fl);
1066 /* set filter fid EA.
1067 * FIXME: it holds read lock of ofd object to modify the XATTR_NAME_FID
1068 * while the write lock should be held. However, it should work because
1069 * write RPCs only modify ff_{parent,layout} and those information will
1070 * be the same from all the write RPCs. The reason that fl is not used
1071 * in dt_xattr_set() is to allow this race. */
1072 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
1073 GOTO(out_unlock, rc);
1074 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
1075 ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
1076 else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
1077 le32_add_cpu(&ff->ff_parent.f_oid, -1);
1079 info->fti_buf.lb_buf = ff;
1080 info->fti_buf.lb_len = sizeof(*ff);
1081 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, 0, th);
1083 filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff));
1085 GOTO(out_unlock, rc);
1088 ofd_read_unlock(env, ofd_obj);
1090 dt_trans_stop(env, ofd->ofd_osd, th);
1092 la->la_valid = valid;
1096 struct ofd_soft_sync_callback {
1097 struct dt_txn_commit_cb ossc_cb;
1098 struct obd_export *ossc_exp;
1102 * Callback function for "soft sync" update.
1104 * Reset fed_soft_sync_count upon committing the "soft_sync" update.
1105 * See ofd_soft_sync_cb_add() below for more details on soft sync.
1107 * \param[in] env execution environment
1108 * \param[in] th transaction handle
1109 * \param[in] cb callback data
1110 * \param[in] err error code
1112 static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th,
1113 struct dt_txn_commit_cb *cb, int err)
1115 struct ofd_soft_sync_callback *ossc;
1117 ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb);
1119 CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp);
1120 atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0);
1122 class_export_cb_put(ossc->ossc_exp);
1127 * Add callback for "soft sync" processing.
1129 * The "soft sync" mechanism does asynchronous commit when OBD_BRW_SOFT_SYNC
1130 * flag is set in client buffers. The intention is for this operation to
1131 * commit pages belonging to a client which has "too many" outstanding
1132 * unstable pages in its cache. See LU-2139 for details.
1134 * This function adds callback to be called when commit is done.
1136 * \param[in] th transaction handle
1137 * \param[in] exp OBD export of client
1139 * \retval 0 on successful callback adding
1140 * \retval negative value on error
1142 static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
1144 struct ofd_soft_sync_callback *ossc;
1145 struct dt_txn_commit_cb *dcb;
1148 OBD_ALLOC_PTR(ossc);
1152 ossc->ossc_exp = class_export_cb_get(exp);
1154 dcb = &ossc->ossc_cb;
1155 dcb->dcb_func = ofd_cb_soft_sync;
1156 INIT_LIST_HEAD(&dcb->dcb_linkage);
1157 strlcpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name));
1159 rc = dt_trans_cb_add(th, dcb);
1161 class_export_cb_put(exp);
1169 * Commit bulk IO buffers to the storage.
1171 * This function finalizes write IO processing by writing data to the disk.
1172 * That write can be synchronous or asynchronous depending on buffers flags.
1174 * \param[in] env execution environment
1175 * \param[in] exp OBD export of client
1176 * \param[in] ofd OFD device
1177 * \param[in] fid FID of object
1178 * \param[in] la object attributes
1179 * \param[in] ff parent FID of object
1180 * \param[in] objcount always 1
1181 * \param[in] niocount number of local buffers
1182 * \param[in] lnb local buffers
1183 * \param[in] granted grant space consumed for the bulk I/O
1184 * \param[in] old_rc result of processing at this point
1186 * \retval 0 on successful commit
1187 * \retval negative value on error
1190 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
1191 struct ofd_device *ofd, const struct lu_fid *fid,
1192 struct lu_attr *la, struct obdo *oa, int objcount,
1193 int niocount, struct niobuf_local *lnb,
1194 unsigned long granted, int old_rc)
1196 struct ofd_thread_info *info = ofd_info(env);
1197 struct filter_export_data *fed = &exp->exp_filter_data;
1198 struct ofd_object *fo;
1199 struct dt_object *o;
1205 bool soft_sync = false;
1206 bool cb_registered = false;
1207 bool fake_write = false;
1211 LASSERT(objcount == 1);
1213 fo = ofd_info(env)->fti_obj;
1214 LASSERT(fo != NULL);
1216 o = ofd_object_child(fo);
1220 GOTO(out, rc = old_rc);
1221 if (!ofd_object_exists(fo))
1222 GOTO(out, rc = -ENOENT);
1225 * The first write to each object must set some attributes. It is
1226 * important to set the uid/gid before calling
1227 * dt_declare_write_commit() since quota enforcement is now handled in
1230 rc = ofd_write_attr_set(env, ofd, fo, la, oa);
1234 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
1236 /* do fake write, to simulate the write case for performance testing */
1237 if (OBD_FAIL_CHECK_QUIET(OBD_FAIL_OST_FAKE_RW)) {
1238 struct niobuf_local *last = &lnb[niocount - 1];
1239 __u64 file_size = last->lnb_file_offset + last->lnb_len;
1240 __u64 valid = la->la_valid;
1242 la->la_valid = LA_SIZE;
1244 rc = dt_attr_get(env, o, la);
1245 if (rc < 0 && rc != -ENOENT)
1248 if (file_size < la->la_size)
1249 file_size = la->la_size;
1251 /* dirty inode by setting file size */
1252 la->la_valid = valid | LA_SIZE;
1253 la->la_size = file_size;
1259 th = ofd_trans_create(env, ofd);
1261 GOTO(out, rc = PTR_ERR(th));
1263 th->th_sync |= ofd->ofd_sync_journal;
1264 if (th->th_sync == 0) {
1265 for (i = 0; i < niocount; i++) {
1266 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
1270 if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC)
1275 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
1276 GOTO(out_stop, rc = -EINPROGRESS);
1278 if (likely(!fake_write)) {
1279 rc = dt_declare_write_commit(env, o, lnb, niocount, th);
1284 /* don't update atime on disk if it is older */
1285 if (la->la_valid & LA_ATIME && la->la_atime <= fo->ofo_atime_ondisk)
1286 la->la_valid &= ~LA_ATIME;
1289 /* update [mac]time if needed */
1290 rc = dt_declare_attr_set(env, o, la, th);
1295 rc = ofd_trans_start(env, ofd, fo, th);
1299 ofd_read_lock(env, fo);
1300 if (!ofd_object_exists(fo))
1301 GOTO(out_unlock, rc = -ENOENT);
1303 /* Don't update timestamps if this write is older than a
1304 * setattr which modifies the timestamps. b=10150 */
1305 if (la->la_valid && tgt_fmd_check(exp, fid, info->fti_xid)) {
1306 rc = dt_attr_set(env, o, la, th);
1308 GOTO(out_unlock, rc);
1309 if (la->la_valid & LA_ATIME)
1310 fo->ofo_atime_ondisk = la->la_atime;
1313 if (likely(!fake_write)) {
1314 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_OST_WR_ATTR_DELAY,
1315 OBD_FAIL_ONCE, cfs_fail_val);
1316 rc = dt_write_commit(env, o, lnb, niocount, th, oa->o_size);
1318 restart = th->th_restart_tran;
1319 GOTO(out_unlock, rc);
1323 /* get attr to return */
1324 rc = dt_attr_get(env, o, la);
1327 ofd_read_unlock(env, fo);
1329 /* Force commit to make the just-deleted blocks
1330 * reusable. LU-456 */
1334 /* do this before trans stop in case commit has finished */
1335 if (!th->th_sync && soft_sync && !cb_registered) {
1336 ofd_soft_sync_cb_add(th, exp);
1337 cb_registered = true;
1340 if (rc == 0 && granted > 0) {
1341 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
1345 rc2 = ofd_trans_stop(env, ofd, th, restart ? 0 : rc);
1348 if (rc == -ENOSPC && retries++ < 3) {
1349 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
1357 if (retries % 10000 == 0)
1358 CERROR("%s: restart IO write too many times: %d\n",
1359 ofd_name(ofd), retries);
1360 CDEBUG(D_INODE, "retry transaction, retries:%d\n",
1365 /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
1366 atomic_set(&fed->fed_soft_sync_count, 0);
1367 else if (atomic_inc_return(&fed->fed_soft_sync_count) ==
1368 ofd->ofd_soft_sync_limit)
1369 dt_commit_async(env, ofd->ofd_osd);
1372 dt_bufs_put(env, o, lnb, niocount);
1373 ofd_object_put(env, fo);
1375 tgt_grant_commit(exp, granted, old_rc);
1380 * Commit bulk IO to the storage.
1382 * This is companion function to the ofd_preprw(). It finishes bulk IO
1383 * request processing by committing buffers to the storage (WRITE) and/or
1384 * freeing those buffers (read/write). See ofd_commitrw_read() and
1385 * ofd_commitrw_write() for details about each type of IO.
1387 * \param[in] env execution environment
1388 * \param[in] cmd IO type (READ/WRITE)
1389 * \param[in] exp OBD export of client
1390 * \param[in] oa OBDO structure from client
1391 * \param[in] objcount always 1
1392 * \param[in] obj object data
1393 * \param[in] rnb remote buffers
1394 * \param[in] npages number of local buffers
1395 * \param[in] lnb local buffers
1396 * \param[in] old_rc result of processing at this point
1398 * \retval 0 on successful commit
1399 * \retval negative value on error
1401 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
1402 struct obdo *oa, int objcount, struct obd_ioobj *obj,
1403 struct niobuf_remote *rnb, int npages,
1404 struct niobuf_local *lnb, int old_rc, int nob, ktime_t kstart)
1406 struct tgt_session_info *tsi = tgt_ses_info(env);
1407 struct ofd_thread_info *info = ofd_info(env);
1408 struct ofd_device *ofd = ofd_exp(exp);
1409 const struct lu_fid *fid = &oa->o_oi.oi_fid;
1410 struct ldlm_namespace *ns = ofd->ofd_namespace;
1411 struct ldlm_resource *rs = NULL;
1415 int root_squash = 0;
1417 LASSERT(npages > 0);
1419 if (tgt_ses_req(tsi) == NULL) { /* echo client case */
1422 jobid = tsi->tsi_jobid;
1425 if (cmd == OBD_BRW_WRITE) {
1426 struct lu_nodemap *nodemap;
1427 __u32 mapped_uid, mapped_gid, mapped_projid;
1429 /* doing this before the commit operation places the counter
1430 * update almost immediately after reply to the client, which
1431 * gives reasonable time stats and lets us use the actual
1432 * bytes of i/o (rather than requested)
1434 ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE_BYTES, jobid, nob);
1435 ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid,
1436 ktime_us_delta(ktime_get(), kstart));
1438 mapped_uid = oa->o_uid;
1439 mapped_gid = oa->o_gid;
1440 mapped_projid = oa->o_projid;
1441 nodemap = nodemap_get_from_exp(exp);
1442 if (!IS_ERR(nodemap)) {
1443 mapped_uid = nodemap_map_id(nodemap, NODEMAP_UID,
1444 NODEMAP_FS_TO_CLIENT,
1446 mapped_gid = nodemap_map_id(nodemap, NODEMAP_GID,
1447 NODEMAP_FS_TO_CLIENT,
1449 mapped_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
1450 NODEMAP_FS_TO_CLIENT,
1452 } else if (old_rc == 0) {
1453 old_rc = PTR_ERR(nodemap);
1456 if (!IS_ERR_OR_NULL(nodemap)) {
1457 /* do not bypass quota enforcement if squashed uid */
1458 if (unlikely(mapped_uid == nodemap->nm_squash_uid)) {
1461 for (idx = 0; idx < npages; idx++)
1462 lnb[idx].lnb_flags &=
1463 ~OBD_BRW_SYS_RESOURCE;
1466 nodemap_putref(nodemap);
1469 valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID |
1470 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
1471 la_from_obdo(&info->fti_attr, oa, valid);
1473 rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
1474 oa, objcount, npages, lnb,
1475 oa->o_grant_used, old_rc);
1477 obdo_from_la(oa, &info->fti_attr,
1478 OFD_VALID_FLAGS | LA_GID | LA_UID |
1481 obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID |
1484 /* don't report overquota flag if we failed before reaching
1486 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
1487 /* return the overquota flags to client */
1488 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
1489 if (oa->o_valid & OBD_MD_FLFLAGS)
1490 oa->o_flags |= OBD_FL_NO_USRQUOTA;
1492 oa->o_flags = OBD_FL_NO_USRQUOTA;
1495 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
1496 if (oa->o_valid & OBD_MD_FLFLAGS)
1497 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
1499 oa->o_flags = OBD_FL_NO_GRPQUOTA;
1501 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
1502 if (oa->o_valid & OBD_MD_FLFLAGS)
1503 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
1505 oa->o_flags = OBD_FL_NO_PRJQUOTA;
1509 oa->o_flags |= OBD_FL_ROOT_SQUASH;
1511 oa->o_valid |= OBD_MD_FLFLAGS;
1512 oa->o_valid |= OBD_MD_FLALLQUOTA;
1516 * Update LVB after writing finish for server lock, see
1517 * comments in ldlm_lock_decref_internal(), If this is a
1518 * local lock on a server namespace and this was the last
1519 * reference, lock will be destroyed directly thus there
1520 * is no chance for ldlm_request_cancel() to update lvb.
1522 if (rc == 0 && (rnb[0].rnb_flags & OBD_BRW_SRVLOCK)) {
1523 ost_fid_build_resid(fid, &info->fti_resid);
1524 rs = ldlm_resource_get(ns, &info->fti_resid,
1527 ldlm_res_lvbo_update(rs, NULL, 1);
1528 ldlm_resource_putref(rs);
1532 /* Convert back to client IDs. LU-9671.
1533 * nodemap_get_from_exp() may fail due to nodemap deactivated,
1534 * server ID will be returned back to client in that case. */
1535 oa->o_uid = mapped_uid;
1536 oa->o_gid = mapped_gid;
1537 oa->o_projid = mapped_projid;
1538 } else if (cmd == OBD_BRW_READ) {
1539 /* see comment on LPROC_OFD_STATS_WRITE_BYTES usage above */
1540 ofd_counter_incr(exp, LPROC_OFD_STATS_READ_BYTES, jobid, nob);
1541 ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid,
1542 ktime_us_delta(ktime_get(), kstart));
1544 rc = ofd_commitrw_read(env, ofd, fid, objcount,