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 ofd_read_lock(env, fo);
584 if (!ofd_object_exists(fo))
585 GOTO(unlock, rc = -ENOENT);
587 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
588 rc = ofd_verify_ff(env, fo, oa);
593 if (ptlrpc_connection_is_local(exp->exp_connection))
594 dbt |= DT_BUFS_TYPE_LOCAL;
599 for (*nr_local = 0, i = 0, j = 0; i < niocount; i++) {
600 begin = min_t(__u64, begin, rnb[i].rnb_offset);
601 end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
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);
623 ofd_read_unlock(env, fo);
627 .f_seq = oa->o_parent_seq,
628 .f_oid = oa->o_parent_oid,
629 .f_ver = oa->o_stripe_idx,
639 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
641 ofd_read_unlock(env, fo);
642 ofd_object_put(env, fo);
647 * Prepare buffers for write request processing.
649 * This function converts remote buffers from client to local buffers
650 * and prepares the latter. If there is recovery in progress and required
651 * object is missing then it can be re-created before write.
653 * \param[in] env execution environment
654 * \param[in] exp OBD export of client
655 * \param[in] ofd OFD device
656 * \param[in] fid FID of object
657 * \param[in] la object attributes
658 * \param[in] oa OBDO structure from client
659 * \param[in] objcount always 1
660 * \param[in] obj object data
661 * \param[in] rnb remote buffers
662 * \param[in] nr_local number of local buffers
663 * \param[in] lnb local buffers
664 * \param[in] jobid job ID name
666 * \retval 0 on successful prepare
667 * \retval negative value on error
669 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
670 struct ofd_device *ofd, const struct lu_fid *fid,
671 struct lu_attr *la, struct obdo *oa,
672 int objcount, struct obd_ioobj *obj,
673 struct niobuf_remote *rnb, int *nr_local,
674 struct niobuf_local *lnb)
676 struct ofd_object *fo;
677 int i, j, k, rc = 0, tot_bytes = 0;
678 enum dt_bufs_type dbt = DT_BUFS_TYPE_WRITE;
679 int maxlnb = *nr_local;
683 LASSERT(env != NULL);
684 LASSERT(objcount == 1);
686 if (unlikely(exp->exp_obd->obd_recovering)) {
687 u64 seq = ostid_seq(&oa->o_oi);
688 u64 oid = ostid_id(&oa->o_oi);
689 struct ofd_seq *oseq;
691 oseq = ofd_seq_load(env, ofd, seq);
693 CERROR("%s: Can't find FID Sequence %#llx: rc = %d\n",
694 ofd_name(ofd), seq, (int)PTR_ERR(oseq));
695 GOTO(out, rc = -EINVAL);
698 if (oid > ofd_seq_last_oid(oseq)) {
702 mutex_lock(&oseq->os_create_lock);
703 diff = oid - ofd_seq_last_oid(oseq);
705 /* Do sync create if the seq is about to used up */
706 if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) {
707 if (unlikely(oid >= IDIF_MAX_OID - 1))
709 } else if (fid_seq_is_norm(seq)) {
711 LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
714 CERROR("%s : invalid o_seq "DOSTID"\n",
715 ofd_name(ofd), POSTID(&oa->o_oi));
716 mutex_unlock(&oseq->os_create_lock);
717 ofd_seq_put(env, oseq);
718 GOTO(out, rc = -EINVAL);
722 u64 next_id = ofd_seq_last_oid(oseq) + 1;
723 int count = ofd_precreate_batch(ofd, diff);
725 rc = ofd_precreate_objects(env, ofd, next_id,
728 mutex_unlock(&oseq->os_create_lock);
729 ofd_seq_put(env, oseq);
736 mutex_unlock(&oseq->os_create_lock);
739 ofd_seq_put(env, oseq);
742 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
743 * space back if possible, we have to do this outside of the lock as
744 * grant preparation may need to sync whole fs thus wait for all the
745 * transactions to complete. */
746 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
748 fo = ofd_object_find(env, ofd, fid);
750 GOTO(out, rc = PTR_ERR(fo));
753 ofd_info(env)->fti_obj = fo;
755 if (!ofd_object_exists(fo)) {
756 CERROR("%s: BRW to missing obj "DOSTID"\n",
757 exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
758 ofd_object_put(env, fo);
759 GOTO(out, rc = -ENOENT);
762 if (ptlrpc_connection_is_local(exp->exp_connection))
763 dbt |= DT_BUFS_TYPE_LOCAL;
768 /* parse remote buffers to local buffers and prepare the latter */
769 for (*nr_local = 0, i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
770 begin = min_t(__u64, begin, rnb[i].rnb_offset);
771 end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
773 if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
774 rnb[i].rnb_len += PAGE_SIZE;
775 rc = dt_bufs_get(env, ofd_object_child(fo),
776 rnb + i, lnb + j, maxlnb, dbt);
777 if (unlikely(rc < 0))
778 GOTO(err_nolock, rc);
779 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
780 /* correct index for local buffers to continue with */
781 for (k = 0; k < rc; k++) {
782 lnb[j+k].lnb_flags = rnb[i].rnb_flags;
783 lnb[j+k].lnb_flags &= ~OBD_BRW_LOCALS;
784 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
785 lnb[j+k].lnb_rc = -ENOSPC;
790 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
791 tot_bytes += rnb[i].rnb_len;
793 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
795 ofd_read_lock(env, fo);
796 if (!ofd_object_exists(fo)) {
797 CERROR("%s: BRW to missing obj "DOSTID": rc = -ENOENT\n",
798 exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
799 GOTO(err, rc = -ENOENT);
802 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
803 rc = ofd_verify_ff(env, fo, oa);
808 /* need to verify layout version */
809 if (oa->o_valid & OBD_MD_LAYOUT_VERSION) {
810 rc = ofd_verify_layout_version(env, fo, oa);
815 rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
816 if (unlikely(rc != 0))
819 ofd_read_unlock(env, fo);
823 .f_seq = oa->o_parent_seq,
824 .f_oid = oa->o_parent_oid,
825 .f_ver = oa->o_stripe_idx,
835 ofd_read_unlock(env, fo);
837 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
838 ofd_object_put(env, fo);
839 /* tgt_grant_prepare_write() was called, so we must commit */
840 tgt_grant_commit(exp, oa->o_grant_used, rc);
842 /* let's still process incoming grant information packed in the oa,
843 * but without enforcing grant since we won't proceed with the write.
844 * Just like a read request actually. */
845 tgt_grant_prepare_read(env, exp, oa);
850 * Prepare bulk IO requests for processing.
852 * This function does initial checks of IO and calls corresponding
853 * functions for read/write processing.
855 * \param[in] env execution environment
856 * \param[in] cmd IO type (read/write)
857 * \param[in] exp OBD export of client
858 * \param[in] oa OBDO structure from request
859 * \param[in] objcount always 1
860 * \param[in] obj object data
861 * \param[in] rnb remote buffers
862 * \param[in] nr_local number of local buffers
863 * \param[in] lnb local buffers
865 * \retval 0 on successful prepare
866 * \retval negative value on error
868 int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
869 struct obdo *oa, int objcount, struct obd_ioobj *obj,
870 struct niobuf_remote *rnb, int *nr_local,
871 struct niobuf_local *lnb)
873 struct tgt_session_info *tsi = tgt_ses_info(env);
874 struct ofd_device *ofd = ofd_exp(exp);
875 struct ofd_thread_info *info;
876 const struct lu_fid *fid = &oa->o_oi.oi_fid;
879 if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
880 CERROR("%s: bulk has too many pages %d, which exceeds the maximum pages per RPC of %d\n",
881 exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
885 if (tgt_ses_req(tsi) == NULL) { /* echo client case */
886 info = ofd_info_init(env, exp);
888 info = tsi2ofd_info(tsi);
893 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) {
894 struct ofd_seq *oseq;
896 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
898 CERROR("%s: Can not find seq for "DOSTID
899 ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
904 if (oseq->os_destroys_in_progress == 0) {
905 /* don't fail lookups for orphan recovery, it causes
906 * later LBUGs when objects still exist during
908 ofd_seq_put(env, oseq);
911 ofd_seq_put(env, oseq);
914 LASSERT(objcount == 1);
915 LASSERT(obj->ioo_bufcnt > 0);
917 if (cmd == OBD_BRW_WRITE) {
918 la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
919 rc = ofd_preprw_write(env, exp, ofd, fid, &info->fti_attr, oa,
920 objcount, obj, rnb, nr_local, lnb);
921 } else if (cmd == OBD_BRW_READ) {
922 tgt_grant_prepare_read(env, exp, oa);
923 rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa,
924 obj->ioo_bufcnt, rnb, nr_local, lnb);
926 CERROR("%s: wrong cmd %d received!\n",
927 exp->exp_obd->obd_name, cmd);
934 * Drop reference on local buffers for read bulk IO.
936 * This will free all local buffers use by this read request.
938 * \param[in] env execution environment
939 * \param[in] ofd OFD device
940 * \param[in] fid object FID
941 * \param[in] objcount always 1
942 * \param[in] niocount number of local buffers
943 * \param[in] lnb local buffers
945 * \retval 0 on successful execution
946 * \retval negative value on error
949 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
950 const struct lu_fid *fid, int objcount, int niocount,
951 struct niobuf_local *lnb)
953 struct ofd_object *fo;
957 LASSERT(niocount > 0);
959 fo = ofd_info(env)->fti_obj;
961 LASSERT(ofd_object_exists(fo));
962 dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
964 ofd_object_put(env, fo);
970 * Set attributes of object during write bulk IO processing.
972 * Change object attributes and write parent FID into extended
973 * attributes when needed.
975 * \param[in] env execution environment
976 * \param[in] ofd OFD device
977 * \param[in] ofd_obj OFD object
978 * \param[in] la object attributes
981 * \retval 0 on successful attributes update
982 * \retval negative value on error
985 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
986 struct ofd_object *ofd_obj, struct lu_attr *la,
989 struct ofd_thread_info *info = ofd_info(env);
990 struct filter_fid *ff = &info->fti_mds_fid;
991 __u64 valid = la->la_valid;
993 struct dt_object *dt_obj;
1001 dt_obj = ofd_object_child(ofd_obj);
1002 LASSERT(dt_obj != NULL);
1004 la->la_valid &= LA_UID | LA_GID | LA_PROJID;
1006 rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
1010 if (!la->la_valid && !(oa->o_valid &
1011 (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
1012 /* no attributes to set */
1015 th = ofd_trans_create(env, ofd);
1017 GOTO(out, rc = PTR_ERR(th));
1020 rc = dt_declare_attr_set(env, dt_obj, la, th);
1025 if (oa->o_valid & (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT |
1026 OBD_MD_LAYOUT_VERSION)) {
1027 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
1028 XATTR_NAME_FID, 0, th);
1032 /* We don't need a transno for this operation which will be re-executed
1033 * anyway when the OST_WRITE (with a transno assigned) is replayed */
1034 rc = dt_trans_start_local(env, ofd->ofd_osd , th);
1038 ofd_read_lock(env, ofd_obj);
1040 rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
1042 GOTO(out_unlock, rc);
1044 if (!la->la_valid && !(oa->o_valid &
1045 (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
1046 /* no attributes to set */
1047 GOTO(out_unlock, rc = 0);
1051 /* set uid/gid/projid */
1053 rc = dt_attr_set(env, dt_obj, la, th);
1055 GOTO(out_unlock, rc);
1058 fl = ofd_object_ff_update(env, ofd_obj, oa, ff);
1060 GOTO(out_unlock, rc = fl);
1062 /* set filter fid EA.
1063 * FIXME: it holds read lock of ofd object to modify the XATTR_NAME_FID
1064 * while the write lock should be held. However, it should work because
1065 * write RPCs only modify ff_{parent,layout} and those information will
1066 * be the same from all the write RPCs. The reason that fl is not used
1067 * in dt_xattr_set() is to allow this race. */
1068 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
1069 GOTO(out_unlock, rc);
1070 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
1071 ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
1072 else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
1073 le32_add_cpu(&ff->ff_parent.f_oid, -1);
1075 info->fti_buf.lb_buf = ff;
1076 info->fti_buf.lb_len = sizeof(*ff);
1077 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, 0, th);
1079 filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff));
1081 GOTO(out_unlock, rc);
1084 ofd_read_unlock(env, ofd_obj);
1086 dt_trans_stop(env, ofd->ofd_osd, th);
1088 la->la_valid = valid;
1092 struct ofd_soft_sync_callback {
1093 struct dt_txn_commit_cb ossc_cb;
1094 struct obd_export *ossc_exp;
1098 * Callback function for "soft sync" update.
1100 * Reset fed_soft_sync_count upon committing the "soft_sync" update.
1101 * See ofd_soft_sync_cb_add() below for more details on soft sync.
1103 * \param[in] env execution environment
1104 * \param[in] th transaction handle
1105 * \param[in] cb callback data
1106 * \param[in] err error code
1108 static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th,
1109 struct dt_txn_commit_cb *cb, int err)
1111 struct ofd_soft_sync_callback *ossc;
1113 ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb);
1115 CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp);
1116 atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0);
1118 class_export_cb_put(ossc->ossc_exp);
1123 * Add callback for "soft sync" processing.
1125 * The "soft sync" mechanism does asynchronous commit when OBD_BRW_SOFT_SYNC
1126 * flag is set in client buffers. The intention is for this operation to
1127 * commit pages belonging to a client which has "too many" outstanding
1128 * unstable pages in its cache. See LU-2139 for details.
1130 * This function adds callback to be called when commit is done.
1132 * \param[in] th transaction handle
1133 * \param[in] exp OBD export of client
1135 * \retval 0 on successful callback adding
1136 * \retval negative value on error
1138 static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
1140 struct ofd_soft_sync_callback *ossc;
1141 struct dt_txn_commit_cb *dcb;
1144 OBD_ALLOC_PTR(ossc);
1148 ossc->ossc_exp = class_export_cb_get(exp);
1150 dcb = &ossc->ossc_cb;
1151 dcb->dcb_func = ofd_cb_soft_sync;
1152 INIT_LIST_HEAD(&dcb->dcb_linkage);
1153 strlcpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name));
1155 rc = dt_trans_cb_add(th, dcb);
1157 class_export_cb_put(exp);
1165 * Commit bulk IO buffers to the storage.
1167 * This function finalizes write IO processing by writing data to the disk.
1168 * That write can be synchronous or asynchronous depending on buffers flags.
1170 * \param[in] env execution environment
1171 * \param[in] exp OBD export of client
1172 * \param[in] ofd OFD device
1173 * \param[in] fid FID of object
1174 * \param[in] la object attributes
1175 * \param[in] ff parent FID of object
1176 * \param[in] objcount always 1
1177 * \param[in] niocount number of local buffers
1178 * \param[in] lnb local buffers
1179 * \param[in] granted grant space consumed for the bulk I/O
1180 * \param[in] old_rc result of processing at this point
1182 * \retval 0 on successful commit
1183 * \retval negative value on error
1186 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
1187 struct ofd_device *ofd, const struct lu_fid *fid,
1188 struct lu_attr *la, struct obdo *oa, int objcount,
1189 int niocount, struct niobuf_local *lnb,
1190 unsigned long granted, int old_rc)
1192 struct ofd_thread_info *info = ofd_info(env);
1193 struct filter_export_data *fed = &exp->exp_filter_data;
1194 struct ofd_object *fo;
1195 struct dt_object *o;
1201 bool soft_sync = false;
1202 bool cb_registered = false;
1203 bool fake_write = false;
1207 LASSERT(objcount == 1);
1209 fo = ofd_info(env)->fti_obj;
1210 LASSERT(fo != NULL);
1212 o = ofd_object_child(fo);
1216 GOTO(out, rc = old_rc);
1217 if (!ofd_object_exists(fo))
1218 GOTO(out, rc = -ENOENT);
1221 * The first write to each object must set some attributes. It is
1222 * important to set the uid/gid before calling
1223 * dt_declare_write_commit() since quota enforcement is now handled in
1226 rc = ofd_write_attr_set(env, ofd, fo, la, oa);
1230 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
1232 /* do fake write, to simulate the write case for performance testing */
1233 if (OBD_FAIL_CHECK_QUIET(OBD_FAIL_OST_FAKE_RW)) {
1234 struct niobuf_local *last = &lnb[niocount - 1];
1235 __u64 file_size = last->lnb_file_offset + last->lnb_len;
1236 __u64 valid = la->la_valid;
1238 la->la_valid = LA_SIZE;
1240 rc = dt_attr_get(env, o, la);
1241 if (rc < 0 && rc != -ENOENT)
1244 if (file_size < la->la_size)
1245 file_size = la->la_size;
1247 /* dirty inode by setting file size */
1248 la->la_valid = valid | LA_SIZE;
1249 la->la_size = file_size;
1255 th = ofd_trans_create(env, ofd);
1257 GOTO(out, rc = PTR_ERR(th));
1259 th->th_sync |= ofd->ofd_sync_journal;
1260 if (th->th_sync == 0) {
1261 for (i = 0; i < niocount; i++) {
1262 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
1266 if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC)
1271 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
1272 GOTO(out_stop, rc = -EINPROGRESS);
1274 if (likely(!fake_write)) {
1275 rc = dt_declare_write_commit(env, o, lnb, niocount, th);
1280 /* don't update atime on disk if it is older */
1281 if (la->la_valid & LA_ATIME && la->la_atime <= fo->ofo_atime_ondisk)
1282 la->la_valid &= ~LA_ATIME;
1285 /* update [mac]time if needed */
1286 rc = dt_declare_attr_set(env, o, la, th);
1291 rc = ofd_trans_start(env, ofd, fo, th);
1295 ofd_read_lock(env, fo);
1296 if (!ofd_object_exists(fo))
1297 GOTO(out_unlock, rc = -ENOENT);
1299 /* Don't update timestamps if this write is older than a
1300 * setattr which modifies the timestamps. b=10150 */
1301 if (la->la_valid && tgt_fmd_check(exp, fid, info->fti_xid)) {
1302 rc = dt_attr_set(env, o, la, th);
1304 GOTO(out_unlock, rc);
1305 if (la->la_valid & LA_ATIME)
1306 fo->ofo_atime_ondisk = la->la_atime;
1309 if (likely(!fake_write)) {
1310 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_OST_WR_ATTR_DELAY,
1311 OBD_FAIL_ONCE, cfs_fail_val);
1312 rc = dt_write_commit(env, o, lnb, niocount, th, oa->o_size);
1314 restart = th->th_restart_tran;
1315 GOTO(out_unlock, rc);
1319 /* get attr to return */
1320 rc = dt_attr_get(env, o, la);
1323 ofd_read_unlock(env, fo);
1325 /* Force commit to make the just-deleted blocks
1326 * reusable. LU-456 */
1330 /* do this before trans stop in case commit has finished */
1331 if (!th->th_sync && soft_sync && !cb_registered) {
1332 ofd_soft_sync_cb_add(th, exp);
1333 cb_registered = true;
1336 if (rc == 0 && granted > 0) {
1337 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
1341 rc2 = ofd_trans_stop(env, ofd, th, restart ? 0 : rc);
1344 if (rc == -ENOSPC && retries++ < 3) {
1345 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
1353 if (retries % 10000 == 0)
1354 CERROR("%s: restart IO write too many times: %d\n",
1355 ofd_name(ofd), retries);
1356 CDEBUG(D_INODE, "retry transaction, retries:%d\n",
1361 /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
1362 atomic_set(&fed->fed_soft_sync_count, 0);
1363 else if (atomic_inc_return(&fed->fed_soft_sync_count) ==
1364 ofd->ofd_soft_sync_limit)
1365 dt_commit_async(env, ofd->ofd_osd);
1368 dt_bufs_put(env, o, lnb, niocount);
1369 ofd_object_put(env, fo);
1371 tgt_grant_commit(exp, granted, old_rc);
1376 * Commit bulk IO to the storage.
1378 * This is companion function to the ofd_preprw(). It finishes bulk IO
1379 * request processing by committing buffers to the storage (WRITE) and/or
1380 * freeing those buffers (read/write). See ofd_commitrw_read() and
1381 * ofd_commitrw_write() for details about each type of IO.
1383 * \param[in] env execution environment
1384 * \param[in] cmd IO type (READ/WRITE)
1385 * \param[in] exp OBD export of client
1386 * \param[in] oa OBDO structure from client
1387 * \param[in] objcount always 1
1388 * \param[in] obj object data
1389 * \param[in] rnb remote buffers
1390 * \param[in] npages number of local buffers
1391 * \param[in] lnb local buffers
1392 * \param[in] old_rc result of processing at this point
1394 * \retval 0 on successful commit
1395 * \retval negative value on error
1397 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
1398 struct obdo *oa, int objcount, struct obd_ioobj *obj,
1399 struct niobuf_remote *rnb, int npages,
1400 struct niobuf_local *lnb, int old_rc, int nob, ktime_t kstart)
1402 struct tgt_session_info *tsi = tgt_ses_info(env);
1403 struct ofd_thread_info *info = ofd_info(env);
1404 struct ofd_device *ofd = ofd_exp(exp);
1405 const struct lu_fid *fid = &oa->o_oi.oi_fid;
1406 struct ldlm_namespace *ns = ofd->ofd_namespace;
1407 struct ldlm_resource *rs = NULL;
1411 int root_squash = 0;
1413 LASSERT(npages > 0);
1415 if (tgt_ses_req(tsi) == NULL) { /* echo client case */
1418 jobid = tsi->tsi_jobid;
1421 if (cmd == OBD_BRW_WRITE) {
1422 struct lu_nodemap *nodemap;
1423 __u32 mapped_uid, mapped_gid, mapped_projid;
1425 /* doing this before the commit operation places the counter
1426 * update almost immediately after reply to the client, which
1427 * gives reasonable time stats and lets us use the actual
1428 * bytes of i/o (rather than requested)
1430 ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE_BYTES, jobid, nob);
1431 ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid,
1432 ktime_us_delta(ktime_get(), kstart));
1434 mapped_uid = oa->o_uid;
1435 mapped_gid = oa->o_gid;
1436 mapped_projid = oa->o_projid;
1437 nodemap = nodemap_get_from_exp(exp);
1438 if (!IS_ERR(nodemap)) {
1439 mapped_uid = nodemap_map_id(nodemap, NODEMAP_UID,
1440 NODEMAP_FS_TO_CLIENT,
1442 mapped_gid = nodemap_map_id(nodemap, NODEMAP_GID,
1443 NODEMAP_FS_TO_CLIENT,
1445 mapped_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
1446 NODEMAP_FS_TO_CLIENT,
1448 } else if (old_rc == 0) {
1449 old_rc = PTR_ERR(nodemap);
1452 if (!IS_ERR_OR_NULL(nodemap)) {
1453 /* do not bypass quota enforcement if squashed uid */
1454 if (unlikely(mapped_uid == nodemap->nm_squash_uid)) {
1457 for (idx = 0; idx < npages; idx++)
1458 lnb[idx].lnb_flags &=
1459 ~OBD_BRW_SYS_RESOURCE;
1462 nodemap_putref(nodemap);
1465 valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID |
1466 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
1467 la_from_obdo(&info->fti_attr, oa, valid);
1469 rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
1470 oa, objcount, npages, lnb,
1471 oa->o_grant_used, old_rc);
1473 obdo_from_la(oa, &info->fti_attr,
1474 OFD_VALID_FLAGS | LA_GID | LA_UID |
1477 obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID |
1480 /* don't report overquota flag if we failed before reaching
1482 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
1483 /* return the overquota flags to client */
1484 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
1485 if (oa->o_valid & OBD_MD_FLFLAGS)
1486 oa->o_flags |= OBD_FL_NO_USRQUOTA;
1488 oa->o_flags = OBD_FL_NO_USRQUOTA;
1491 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
1492 if (oa->o_valid & OBD_MD_FLFLAGS)
1493 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
1495 oa->o_flags = OBD_FL_NO_GRPQUOTA;
1497 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
1498 if (oa->o_valid & OBD_MD_FLFLAGS)
1499 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
1501 oa->o_flags = OBD_FL_NO_PRJQUOTA;
1505 oa->o_flags |= OBD_FL_ROOT_SQUASH;
1507 oa->o_valid |= OBD_MD_FLFLAGS;
1508 oa->o_valid |= OBD_MD_FLALLQUOTA;
1512 * Update LVB after writing finish for server lock, see
1513 * comments in ldlm_lock_decref_internal(), If this is a
1514 * local lock on a server namespace and this was the last
1515 * reference, lock will be destroyed directly thus there
1516 * is no chance for ldlm_request_cancel() to update lvb.
1518 if (rc == 0 && (rnb[0].rnb_flags & OBD_BRW_SRVLOCK)) {
1519 ost_fid_build_resid(fid, &info->fti_resid);
1520 rs = ldlm_resource_get(ns, &info->fti_resid,
1523 ldlm_res_lvbo_update(rs, NULL, 1);
1524 ldlm_resource_putref(rs);
1528 /* Convert back to client IDs. LU-9671.
1529 * nodemap_get_from_exp() may fail due to nodemap deactivated,
1530 * server ID will be returned back to client in that case. */
1531 oa->o_uid = mapped_uid;
1532 oa->o_gid = mapped_gid;
1533 oa->o_projid = mapped_projid;
1534 } else if (cmd == OBD_BRW_READ) {
1535 /* see comment on LPROC_OFD_STATS_WRITE_BYTES usage above */
1536 ofd_counter_incr(exp, LPROC_OFD_STATS_READ_BYTES, jobid, nob);
1537 ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid,
1538 ktime_us_delta(ktime_get(), kstart));
1540 rc = ofd_commitrw_read(env, ofd, fid, objcount,