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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osp/osp_sync.c
34 * Lustre OST Proxy Device
36 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
37 * Author: Mikhail Pershin <mike.pershin@intel.com>
38 * Author: Di Wang <di.wang@intel.com>
41 #define DEBUG_SUBSYSTEM S_MDS
43 #include <linux/kthread.h>
45 #include <lustre_obdo.h>
47 #include "osp_internal.h"
50 * there are two specific states to take care about:
52 * = import is disconnected =
54 * = import is inactive =
55 * in this case osp_declare_object_create() returns an error
61 * Check whether statfs data is expired
63 * OSP device caches statfs data for the target, the function checks
64 * whether the data is expired or not.
66 * \param[in] d OSP device
68 * \retval 0 - not expired, 1 - expired
70 static inline int osp_statfs_need_update(struct osp_device *d)
72 return !cfs_time_before(cfs_time_current(),
73 d->opd_statfs_fresh_till);
77 * OSP tries to maintain pool of available objects so that calls to create
78 * objects don't block most of time
80 * each time OSP gets connected to OST, we should start from precreation cleanup
82 static inline bool osp_precreate_running(struct osp_device *d)
84 return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
87 static inline bool osp_precreate_stopped(struct osp_device *d)
89 return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
92 static void osp_statfs_timer_cb(unsigned long _d)
94 struct osp_device *d = (struct osp_device *) _d;
97 if (d->opd_pre != NULL && osp_precreate_running(d))
98 wake_up(&d->opd_pre_waitq);
102 * RPC interpret callback for OST_STATFS RPC
104 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
105 * replied by the target. It's used to maintain statfs cache for the target.
106 * The function fills data from the reply if successful and schedules another
109 * \param[in] env LU environment provided by the caller
110 * \param[in] req RPC replied
111 * \param[in] aa callback data
112 * \param[in] rc RPC result
114 * \retval 0 on success
115 * \retval negative negated errno on error
117 static int osp_statfs_interpret(const struct lu_env *env,
118 struct ptlrpc_request *req,
119 union ptlrpc_async_args *aa, int rc)
121 struct obd_import *imp = req->rq_import;
122 struct obd_statfs *msfs;
123 struct osp_device *d;
127 aa = ptlrpc_req_async_args(req);
128 d = aa->pointer_arg[0];
134 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
136 GOTO(out, rc = -EPROTO);
138 d->opd_statfs = *msfs;
140 osp_pre_update_status(d, rc);
142 /* schedule next update */
143 d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
144 mod_timer(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
145 d->opd_statfs_update_in_progress = 0;
147 CDEBUG(D_CACHE, "updated statfs %p\n", d);
151 /* couldn't update statfs, try again as soon as possible */
152 if (d->opd_pre != NULL && osp_precreate_running(d))
153 wake_up(&d->opd_pre_waitq);
155 if (req->rq_import_generation == imp->imp_generation)
156 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
157 d->opd_obd->obd_name, rc);
162 * Send OST_STATFS RPC
164 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
165 * Also disables scheduled updates as times OSP may need to refresh
166 * statfs data before expiration. The function doesn't block, instead
167 * an interpretation callback osp_statfs_interpret() is used.
169 * \param[in] d OSP device
171 static int osp_statfs_update(struct osp_device *d)
173 struct ptlrpc_request *req;
174 struct obd_import *imp;
175 union ptlrpc_async_args *aa;
180 CDEBUG(D_CACHE, "going to update statfs\n");
182 imp = d->opd_obd->u.cli.cl_import;
185 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
189 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
191 ptlrpc_request_free(req);
194 ptlrpc_request_set_replen(req);
195 req->rq_request_portal = OST_CREATE_PORTAL;
196 ptlrpc_at_set_req_timeout(req);
198 req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
199 aa = ptlrpc_req_async_args(req);
200 aa->pointer_arg[0] = d;
203 * no updates till reply
205 del_timer(&d->opd_statfs_timer);
206 d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
207 d->opd_statfs_update_in_progress = 1;
209 ptlrpcd_add_req(req);
215 * Schedule an immediate update for statfs data
217 * If cached statfs data claim no free space, but OSP has got a request to
218 * destroy an object (so release some space probably), then we may need to
219 * refresh cached statfs data sooner than planned. The function checks there
220 * is no statfs update going and schedules immediate update if so.
221 * XXX: there might be a case where removed object(s) do not add free space (empty
222 * object). If the number of such deletions is high, then we can start to update
223 * statfs too often causing a RPC storm. some throttling is needed...
225 * \param[in] d OSP device where statfs data needs to be refreshed
227 void osp_statfs_need_now(struct osp_device *d)
229 if (!d->opd_statfs_update_in_progress) {
231 * if current status is -ENOSPC (lack of free space on OST)
232 * then we should poll OST immediately once object destroy
235 d->opd_statfs_fresh_till = cfs_time_shift(-1);
236 del_timer(&d->opd_statfs_timer);
237 wake_up(&d->opd_pre_waitq);
242 * Check pool of precreated objects is nearly empty
244 * We should not wait till the pool of the precreated objects is exhausted,
245 * because then there will be a long period of OSP being unavailable for the
246 * new creations due to lenghty precreate RPC. Instead we ask for another
247 * precreation ahead and hopefully have it ready before the current pool is
248 * empty. Notice this function relies on an external locking.
250 * \param[in] env LU environment provided by the caller
251 * \param[in] d OSP device
253 * \retval 0 - current pool is good enough, 1 - time to precreate
255 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
256 struct osp_device *d)
258 int window = osp_objs_precreated(env, d);
260 /* don't consider new precreation till OST is healty and
262 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
263 (d->opd_pre_status == 0));
267 * Check pool of precreated objects
269 * This is protected version of osp_precreate_near_empty_nolock(), check that
272 * \param[in] env LU environment provided by the caller
273 * \param[in] d OSP device
275 * \retval 0 - current pool is good enough, 1 - time to precreate
277 static inline int osp_precreate_near_empty(const struct lu_env *env,
278 struct osp_device *d)
282 /* XXX: do we really need locking here? */
283 spin_lock(&d->opd_pre_lock);
284 rc = osp_precreate_near_empty_nolock(env, d);
285 spin_unlock(&d->opd_pre_lock);
290 * Check given sequence is empty
292 * Returns a binary result whether the given sequence has some IDs left
293 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
294 * version of that function.
296 * \param[in] env LU environment provided by the caller
297 * \param[in] osp OSP device
299 * \retval 0 - current sequence has no IDs, 1 - otherwise
301 static inline int osp_create_end_seq(const struct lu_env *env,
302 struct osp_device *osp)
304 struct lu_fid *fid = &osp->opd_pre_used_fid;
307 spin_lock(&osp->opd_pre_lock);
308 rc = osp_fid_end_seq(env, fid);
309 spin_unlock(&osp->opd_pre_lock);
314 * Write FID into into last_oid/last_seq file
316 * The function stores the sequence and the in-sequence id into two dedicated
317 * files. The sync argument can be used to request synchronous commit, so the
318 * function won't return until the updates are committed.
320 * \param[in] env LU environment provided by the caller
321 * \param[in] osp OSP device
322 * \param[in] fid fid where sequence/id is taken
323 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
325 * \retval 0 on success
326 * \retval negative negated errno on error
328 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
329 struct lu_fid *fid, int sync)
331 struct osp_thread_info *oti = osp_env_info(env);
332 struct lu_buf *lb_oid = &oti->osi_lb;
333 struct lu_buf *lb_oseq = &oti->osi_lb2;
340 if (osp->opd_storage->dd_rdonly)
343 /* Note: through f_oid is only 32 bits, it will also write 64 bits
344 * for oid to keep compatibility with the previous version. */
345 lb_oid->lb_buf = &fid->f_oid;
346 lb_oid->lb_len = sizeof(u64);
347 oid_off = sizeof(u64) * osp->opd_index;
349 lb_oseq->lb_buf = &fid->f_seq;
350 lb_oseq->lb_len = sizeof(u64);
351 oseq_off = sizeof(u64) * osp->opd_index;
353 th = dt_trans_create(env, osp->opd_storage);
358 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
359 lb_oid, oid_off, th);
363 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
364 lb_oseq, oseq_off, th);
368 rc = dt_trans_start_local(env, osp->opd_storage, th);
372 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
375 CERROR("%s: can not write to last seq file: rc = %d\n",
376 osp->opd_obd->obd_name, rc);
379 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
382 CERROR("%s: can not write to last seq file: rc = %d\n",
383 osp->opd_obd->obd_name, rc);
387 dt_trans_stop(env, osp->opd_storage, th);
392 * Switch to another sequence
394 * When a current sequence has no available IDs left, OSP has to switch to
395 * another new sequence. OSP requests it using the regular FLDB protocol
396 * and stores synchronously before that is used in precreated. This is needed
397 * to basically have the sequences referenced (not orphaned), otherwise it's
398 * possible that OST has some objects precreated and the clients have data
399 * written to it, but after MDT failover nobody refers those objects and OSP
400 * has no idea that the sequence need cleanup to be done.
401 * While this is very expensive operation, it's supposed to happen very very
402 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
404 * \param[in] env LU environment provided by the caller
405 * \param[in] osp OSP device
407 * \retval 0 on success
408 * \retval negative negated errno on error
410 static int osp_precreate_rollover_new_seq(struct lu_env *env,
411 struct osp_device *osp)
413 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
414 struct lu_fid *last_fid = &osp->opd_last_used_fid;
418 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
420 CERROR("%s: alloc fid error: rc = %d\n",
421 osp->opd_obd->obd_name, rc);
427 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
428 "fid "DFID", last_fid "DFID"\n", PFID(fid),
431 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
433 CERROR("%s: Can not update oid/seq file: rc = %d\n",
434 osp->opd_obd->obd_name, rc);
438 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
439 osp->opd_obd->obd_name, fid_seq(last_fid),
441 /* Update last_xxx to the new seq */
442 spin_lock(&osp->opd_pre_lock);
443 osp->opd_last_used_fid = *fid;
444 osp->opd_gap_start_fid = *fid;
445 osp->opd_pre_used_fid = *fid;
446 osp->opd_pre_last_created_fid = *fid;
447 spin_unlock(&osp->opd_pre_lock);
453 * Find IDs available in current sequence
455 * The function calculates the highest possible ID and the number of IDs
456 * available in the current sequence OSP is using. The number is limited
457 * artifically by the caller (grow param) and the number of IDs available
458 * in the sequence by nature. The function doesn't require an external
461 * \param[in] env LU environment provided by the caller
462 * \param[in] osp OSP device
463 * \param[in] fid FID the caller wants to start with
464 * \param[in] grow how many the caller wants
465 * \param[out] fid the highest calculated FID
466 * \param[out] grow the number of available IDs calculated
468 * \retval 0 on success, 1 - the sequence is empty
470 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
471 struct lu_fid *fid, int *grow)
473 struct osp_thread_info *osi = osp_env_info(env);
477 if (fid_is_idif(fid)) {
478 struct lu_fid *last_fid;
479 struct ost_id *oi = &osi->osi_oi;
482 spin_lock(&osp->opd_pre_lock);
483 last_fid = &osp->opd_pre_last_created_fid;
484 fid_to_ostid(last_fid, oi);
485 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
486 *grow = end - ostid_id(oi);
487 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
488 spin_unlock(&osp->opd_pre_lock);
490 if (*grow == 0 || rc)
493 ostid_to_fid(fid, oi, osp->opd_index);
497 spin_lock(&osp->opd_pre_lock);
498 *fid = osp->opd_pre_last_created_fid;
500 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
501 *grow = end - fid->f_oid;
502 fid->f_oid += end - fid->f_oid;
503 spin_unlock(&osp->opd_pre_lock);
505 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
506 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
508 return *grow > 0 ? 0 : 1;
512 * Prepare and send precreate RPC
514 * The function finds how many objects should be precreated. Then allocates,
515 * prepares and schedules precreate RPC synchronously. Upon reply the function
516 * wake ups the threads waiting for the new objects on this target. If the
517 * target wasn't able to create all the objects requested, then the next
518 * precreate will be asking less objects (i.e. slow precreate down).
520 * \param[in] env LU environment provided by the caller
521 * \param[in] d OSP device
523 * \retval 0 on success
524 * \retval negative negated errno on error
526 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
528 struct osp_thread_info *oti = osp_env_info(env);
529 struct ptlrpc_request *req;
530 struct obd_import *imp;
531 struct ost_body *body;
533 struct lu_fid *fid = &oti->osi_fid;
536 /* don't precreate new objects till OST healthy and has free space */
537 if (unlikely(d->opd_pre_status)) {
538 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
539 d->opd_obd->obd_name, d->opd_pre_status);
544 * if not connection/initialization is compeleted, ignore
546 imp = d->opd_obd->u.cli.cl_import;
549 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
552 req->rq_request_portal = OST_CREATE_PORTAL;
553 /* we should not resend create request - anyway we will have delorphan
554 * and kill these objects */
555 req->rq_no_delay = req->rq_no_resend = 1;
557 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
559 ptlrpc_request_free(req);
563 spin_lock(&d->opd_pre_lock);
564 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
565 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
566 grow = d->opd_pre_create_count;
567 spin_unlock(&d->opd_pre_lock);
569 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
572 *fid = d->opd_pre_last_created_fid;
573 rc = osp_precreate_fids(env, d, fid, &grow);
575 /* Current seq has been used up*/
576 if (!osp_is_fid_client(d)) {
577 osp_pre_update_status(d, -ENOSPC);
580 wake_up(&d->opd_pre_waitq);
584 if (!osp_is_fid_client(d)) {
585 /* Non-FID client will always send seq 0 because of
587 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
591 fid_to_ostid(fid, &body->oa.o_oi);
592 body->oa.o_valid = OBD_MD_FLGROUP;
594 ptlrpc_request_set_replen(req);
596 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
599 rc = ptlrpc_queue_wait(req);
601 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
605 LASSERT(req->rq_transno == 0);
607 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
609 GOTO(out_req, rc = -EPROTO);
611 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
614 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
615 CERROR("%s: precreate fid "DFID" < local used fid "DFID
616 ": rc = %d\n", d->opd_obd->obd_name,
617 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
618 GOTO(out_req, rc = -ESTALE);
621 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
623 spin_lock(&d->opd_pre_lock);
625 /* the OST has not managed to create all the
626 * objects we asked for */
627 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
628 d->opd_pre_create_slow = 1;
630 /* the OST is able to keep up with the work,
631 * we could consider increasing create_count
632 * next time if needed */
633 d->opd_pre_create_slow = 0;
636 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
637 fid_to_ostid(fid, &body->oa.o_oi);
639 d->opd_pre_last_created_fid = *fid;
640 spin_unlock(&d->opd_pre_lock);
642 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
643 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
644 PFID(&d->opd_pre_last_created_fid));
646 /* now we can wakeup all users awaiting for objects */
647 osp_pre_update_status(d, rc);
648 wake_up(&d->opd_pre_user_waitq);
650 ptlrpc_req_finished(req);
655 * Get last precreated object from target (OST)
657 * Sends synchronous RPC to the target (OST) to learn the last precreated
658 * object. This later is used to remove all unused objects (cleanup orphan
659 * procedure). Also, the next object after one we got will be used as a
660 * starting point for the new precreates.
662 * \param[in] env LU environment provided by the caller
663 * \param[in] d OSP device
665 * \retval 0 on success
666 * \retval negative negated errno on error
668 static int osp_get_lastfid_from_ost(const struct lu_env *env,
669 struct osp_device *d)
671 struct ptlrpc_request *req = NULL;
672 struct obd_import *imp;
673 struct lu_fid *last_fid;
678 imp = d->opd_obd->u.cli.cl_import;
681 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
685 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
686 sizeof(KEY_LAST_FID));
688 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
690 ptlrpc_request_free(req);
694 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
695 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
697 req->rq_no_delay = req->rq_no_resend = 1;
698 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
699 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
701 ptlrpc_request_set_replen(req);
703 rc = ptlrpc_queue_wait(req);
705 /* bad-bad OST.. let sysadm sort this out */
706 if (rc == -ENOTSUPP) {
707 CERROR("%s: server does not support FID: rc = %d\n",
708 d->opd_obd->obd_name, -ENOTSUPP);
710 ptlrpc_set_import_active(imp, 0);
714 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
715 if (last_fid == NULL) {
716 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
717 GOTO(out, rc = -EPROTO);
720 if (!fid_is_sane(last_fid)) {
721 CERROR("%s: Got insane last_fid "DFID"\n",
722 d->opd_obd->obd_name, PFID(last_fid));
723 GOTO(out, rc = -EPROTO);
726 /* Only update the last used fid, if the OST has objects for
727 * this sequence, i.e. fid_oid > 0 */
728 if (fid_oid(last_fid) > 0)
729 d->opd_last_used_fid = *last_fid;
731 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
735 ptlrpc_req_finished(req);
740 * Cleanup orphans on OST
742 * This function is called in a contex of a dedicated thread handling
743 * all the precreation suff. The function waits till local recovery
744 * is complete, then identify all the unreferenced objects (orphans)
745 * using the highest ID referenced by a local and the highest object
746 * precreated by the target. The found range is a subject to removal
747 * using specially flagged RPC. During this process OSP is marked
748 * unavailable for new objects.
750 * \param[in] env LU environment provided by the caller
751 * \param[in] d OSP device
753 * \retval 0 on success
754 * \retval negative negated errno on error
756 static int osp_precreate_cleanup_orphans(struct lu_env *env,
757 struct osp_device *d)
759 struct osp_thread_info *osi = osp_env_info(env);
760 struct lu_fid *last_fid = &osi->osi_fid;
762 struct ptlrpc_request *req = NULL;
763 struct obd_import *imp;
764 struct ost_body *body;
765 struct l_wait_info lwi = { 0 };
766 int update_status = 0;
773 * wait for local recovery to finish, so we can cleanup orphans
774 * orphans are all objects since "last used" (assigned). we do not
775 * block waiting for all reservations as this can lead to a deadlock
776 * see LU-8972 for the details.
778 spin_lock(&d->opd_pre_lock);
779 d->opd_pre_recovering = 1;
780 spin_unlock(&d->opd_pre_lock);
782 l_wait_event(d->opd_pre_waitq, d->opd_recovery_completed ||
783 !osp_precreate_running(d) || d->opd_got_disconnected,
785 if (!osp_precreate_running(d) || d->opd_got_disconnected)
786 GOTO(out, rc = -EAGAIN);
788 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
789 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
791 *last_fid = d->opd_last_used_fid;
792 /* The OSP should already get the valid seq now */
793 LASSERT(!fid_is_zero(last_fid));
794 if (fid_oid(&d->opd_last_used_fid) < 2) {
795 /* lastfid looks strange... ask OST */
796 LCONSOLE_WARN("%s: refresh last id\n", d->opd_obd->obd_name);
797 rc = osp_get_lastfid_from_ost(env, d);
802 imp = d->opd_obd->u.cli.cl_import;
805 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
807 GOTO(out, rc = -ENOMEM);
809 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
811 ptlrpc_request_free(req);
816 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
818 GOTO(out, rc = -EPROTO);
820 body->oa.o_flags = OBD_FL_DELORPHAN;
821 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
823 /* cleanup objects upto used+reserved as we do not
824 * want to block the orphan cleanup procedure */
825 spin_lock(&d->opd_pre_lock);
826 if (fid_seq(&d->opd_pre_used_fid) != 0) {
827 tmp = d->opd_pre_used_fid;
828 tmp.f_oid += d->opd_pre_reserved;
829 /* shrink current precreate window to let reserved
830 * already objects be created and block new
832 d->opd_pre_last_created_fid = tmp;
834 tmp = d->opd_last_used_fid;
836 fid_to_ostid(&tmp, &body->oa.o_oi);
837 spin_unlock(&d->opd_pre_lock);
839 ptlrpc_request_set_replen(req);
841 /* Don't resend the delorphan req */
842 req->rq_no_resend = req->rq_no_delay = 1;
844 rc = ptlrpc_queue_wait(req);
850 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
852 GOTO(out, rc = -EPROTO);
854 /* OST provides us with id new pool starts from in body->oa.o_id */
855 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
857 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
858 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
859 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
862 ptlrpc_req_finished(req);
865 * If rc is zero, the pre-creation window should have been emptied.
866 * Since waking up the herd would be useless without pre-created
867 * objects, we defer the signal to osp_precreate_send() in that case.
871 CERROR("%s: cannot cleanup orphans: rc = %d\n",
872 d->opd_obd->obd_name, rc);
873 /* we can't proceed from here, OST seem to
874 * be in a bad shape, better to wait for
875 * a new instance of the server and repeat
876 * from the beginning. notify possible waiters
877 * this OSP isn't quite functional yet */
878 osp_pre_update_status(d, rc);
880 wake_up(&d->opd_pre_user_waitq);
885 spin_lock(&d->opd_pre_lock);
886 d->opd_pre_recovering = 0;
887 spin_unlock(&d->opd_pre_lock);
889 /* now we wait until all reserved objects are consumed or released,
890 * so that the window doesn't change. otherwise we can get objects
892 l_wait_event(d->opd_pre_waitq, d->opd_pre_reserved == 0 ||
893 !osp_precreate_running(d) || d->opd_got_disconnected, &lwi);
894 if (!osp_precreate_running(d))
897 spin_lock(&d->opd_pre_lock);
898 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
900 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
901 d->opd_pre_last_created_fid = d->opd_last_used_fid;
903 d->opd_pre_create_count = OST_MIN_PRECREATE;
904 d->opd_pre_last_created_fid = *last_fid;
907 * This empties the pre-creation pool and effectively blocks any new
910 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
911 LUSTRE_DATA_SEQ_MAX_WIDTH);
912 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
913 d->opd_pre_create_slow = 0;
914 spin_unlock(&d->opd_pre_lock);
921 * Update precreate status using statfs data
923 * The function decides whether this OSP should be used for new objects.
924 * IOW, whether this OST is used up or has some free space. Cached statfs
925 * data is used to make this decision. If the latest result of statfs
926 * request (rc argument) is not success, then just mark OSP unavailable
929 * Add a bit of hysteresis so this flag isn't continually flapping,
930 * and ensure that new files don't get extremely fragmented due to
931 * only a small amount of available space in the filesystem.
932 * We want to set the ENOSPC when there is less than reserved size
933 * free and clear it when there is at least 2*reserved size free space.
934 * the function updates current precreation status used: functional or not
936 * \param[in] d OSP device
937 * \param[in] rc new precreate status for device \a d
939 * \retval 0 on success
940 * \retval negative negated errno on error
942 void osp_pre_update_status(struct osp_device *d, int rc)
944 struct obd_statfs *msfs = &d->opd_statfs;
945 int old = d->opd_pre_status;
948 d->opd_pre_status = rc;
952 if (likely(msfs->os_type)) {
953 if (unlikely(d->opd_reserved_mb_high == 0 &&
954 d->opd_reserved_mb_low == 0)) {
955 /* Use ~0.1% by default to disable object allocation,
956 * and ~0.2% to enable, size in MB, set both watermark
958 spin_lock(&d->opd_pre_lock);
959 if (d->opd_reserved_mb_high == 0 &&
960 d->opd_reserved_mb_low == 0) {
961 d->opd_reserved_mb_low =
962 ((msfs->os_bsize >> 10) *
963 msfs->os_blocks) >> 20;
964 if (d->opd_reserved_mb_low == 0)
965 d->opd_reserved_mb_low = 1;
966 d->opd_reserved_mb_high =
967 (d->opd_reserved_mb_low << 1) + 1;
969 spin_unlock(&d->opd_pre_lock);
972 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
973 if (msfs->os_ffree < 32)
974 msfs->os_state |= OS_STATE_ENOINO;
975 else if (msfs->os_ffree > 64)
976 msfs->os_state &= ~OS_STATE_ENOINO;
978 if (available < d->opd_reserved_mb_low)
979 msfs->os_state |= OS_STATE_ENOSPC;
980 else if (available > d->opd_reserved_mb_high)
981 msfs->os_state &= ~OS_STATE_ENOSPC;
982 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
983 d->opd_pre_status = -ENOSPC;
985 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
986 "free, %llu avail, %llu MB avail, %u "
987 "hwm -> %d: rc = %d\n",
988 d->opd_obd->obd_name, msfs->os_blocks,
989 msfs->os_bfree, msfs->os_bavail,
990 available, d->opd_reserved_mb_high,
991 d->opd_pre_status, rc);
993 "non-committed changes: %u, in progress: %u\n",
994 atomic_read(&d->opd_syn_changes),
995 atomic_read(&d->opd_syn_rpc_in_progress));
996 } else if (unlikely(old == -ENOSPC)) {
997 d->opd_pre_status = 0;
998 spin_lock(&d->opd_pre_lock);
999 d->opd_pre_create_slow = 0;
1000 d->opd_pre_create_count = OST_MIN_PRECREATE;
1001 spin_unlock(&d->opd_pre_lock);
1002 wake_up(&d->opd_pre_waitq);
1004 CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1005 " free, %llu avail, %lluMB avail, %u lwm"
1006 " -> %d: rc = %d\n", d->opd_obd->obd_name,
1007 msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1008 available, d->opd_reserved_mb_low,
1009 d->opd_pre_status, rc);
1013 wake_up(&d->opd_pre_user_waitq);
1017 * Initialize FID for precreation
1019 * For a just created new target, a new sequence should be taken.
1020 * The function checks there is no IDIF in use (if the target was
1021 * added with the older version of Lustre), then requests a new
1022 * sequence from FLDB using the regular protocol. Then this new
1023 * sequence is stored on a persisten storage synchronously to prevent
1024 * possible object leakage (for the detail see the description for
1025 * osp_precreate_rollover_new_seq()).
1027 * \param[in] osp OSP device
1029 * \retval 0 on success
1030 * \retval negative negated errno on error
1032 int osp_init_pre_fid(struct osp_device *osp)
1035 struct osp_thread_info *osi;
1036 struct lu_client_seq *cli_seq;
1037 struct lu_fid *last_fid;
1041 LASSERT(osp->opd_pre != NULL);
1043 /* Let's check if the current last_seq/fid is valid,
1044 * otherwise request new sequence from the controller */
1045 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1046 /* Non-MDT0 can only use normal sequence for
1048 if (fid_is_norm(&osp->opd_last_used_fid))
1051 /* Initially MDT0 will start with IDIF, after
1052 * that it will request new sequence from the
1054 if (fid_is_idif(&osp->opd_last_used_fid) ||
1055 fid_is_norm(&osp->opd_last_used_fid))
1059 if (!fid_is_zero(&osp->opd_last_used_fid))
1060 CWARN("%s: invalid last used fid "DFID
1061 ", try to get new sequence.\n",
1062 osp->opd_obd->obd_name,
1063 PFID(&osp->opd_last_used_fid));
1065 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1067 CERROR("%s: init env error: rc = %d\n",
1068 osp->opd_obd->obd_name, rc);
1072 osi = osp_env_info(&env);
1073 last_fid = &osi->osi_fid;
1075 /* For a freshed fs, it will allocate a new sequence first */
1076 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1077 cli_seq = osp->opd_obd->u.cli.cl_seq;
1078 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1080 CERROR("%s: alloc fid error: rc = %d\n",
1081 osp->opd_obd->obd_name, rc);
1085 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1087 last_fid->f_oid = 1;
1088 last_fid->f_ver = 0;
1090 spin_lock(&osp->opd_pre_lock);
1091 osp->opd_last_used_fid = *last_fid;
1092 osp->opd_pre_used_fid = *last_fid;
1093 osp->opd_pre_last_created_fid = *last_fid;
1094 spin_unlock(&osp->opd_pre_lock);
1095 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1097 CERROR("%s: write fid error: rc = %d\n",
1098 osp->opd_obd->obd_name, rc);
1107 * The core of precreate functionality
1109 * The function implements the main precreation loop. Basically it
1110 * involves connecting to the target, precerate FID initialization,
1111 * identifying and removing orphans, then serving precreation. As
1112 * part of the latter, the thread is responsible for statfs data
1113 * updates. The precreation is mostly driven by another threads
1114 * asking for new OST objects - those askers wake the thread when
1115 * the number of precreated objects reach low watermark.
1116 * After a disconnect, the sequence above repeats. This is keep going
1117 * until the thread is requested to stop.
1119 * \param[in] _arg private data the thread (OSP device to handle)
1121 * \retval 0 on success
1122 * \retval negative negated errno on error
1124 static int osp_precreate_thread(void *_arg)
1126 struct osp_device *d = _arg;
1127 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1128 struct l_wait_info lwi = { 0 };
1129 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1130 back_to_sleep, NULL);
1136 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1138 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1141 spin_lock(&d->opd_pre_lock);
1142 thread->t_flags = SVC_STOPPED;
1143 spin_unlock(&d->opd_pre_lock);
1144 wake_up(&thread->t_ctl_waitq);
1149 spin_lock(&d->opd_pre_lock);
1150 thread->t_flags = SVC_RUNNING;
1151 spin_unlock(&d->opd_pre_lock);
1152 wake_up(&thread->t_ctl_waitq);
1154 while (osp_precreate_running(d)) {
1156 * need to be connected to OST
1158 while (osp_precreate_running(d)) {
1159 if (d->opd_pre_recovering &&
1160 d->opd_imp_connected &&
1161 !d->opd_got_disconnected)
1163 l_wait_event(d->opd_pre_waitq,
1164 !osp_precreate_running(d) ||
1165 d->opd_new_connection,
1168 if (!d->opd_new_connection)
1171 d->opd_new_connection = 0;
1172 d->opd_got_disconnected = 0;
1176 if (!osp_precreate_running(d))
1179 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1180 /* Sigh, fid client is not ready yet */
1181 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1184 /* Init fid for osp_precreate if necessary */
1185 rc = osp_init_pre_fid(d);
1187 class_export_put(d->opd_exp);
1188 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1189 CERROR("%s: init pre fid error: rc = %d\n",
1190 d->opd_obd->obd_name, rc);
1194 if (osp_statfs_update(d)) {
1195 l_wait_event(d->opd_pre_waitq,
1196 !osp_precreate_running(d), &lwi2);
1201 * Clean up orphans or recreate missing objects.
1203 rc = osp_precreate_cleanup_orphans(&env, d);
1205 schedule_timeout_interruptible(cfs_time_seconds(1));
1209 * connected, can handle precreates now
1211 while (osp_precreate_running(d)) {
1212 l_wait_event(d->opd_pre_waitq,
1213 !osp_precreate_running(d) ||
1214 osp_precreate_near_empty(&env, d) ||
1215 osp_statfs_need_update(d) ||
1216 d->opd_got_disconnected, &lwi);
1218 if (!osp_precreate_running(d))
1221 /* something happened to the connection
1222 * have to start from the beginning */
1223 if (d->opd_got_disconnected)
1226 if (osp_statfs_need_update(d))
1227 if (osp_statfs_update(d))
1230 /* To avoid handling different seq in precreate/orphan
1231 * cleanup, it will hold precreate until current seq is
1233 if (unlikely(osp_precreate_end_seq(&env, d) &&
1234 !osp_create_end_seq(&env, d)))
1237 if (unlikely(osp_precreate_end_seq(&env, d) &&
1238 osp_create_end_seq(&env, d))) {
1239 LCONSOLE_INFO("%s:%#llx is used up."
1240 " Update to new seq\n",
1241 d->opd_obd->obd_name,
1242 fid_seq(&d->opd_pre_last_created_fid));
1243 rc = osp_precreate_rollover_new_seq(&env, d);
1248 if (osp_precreate_near_empty(&env, d)) {
1249 rc = osp_precreate_send(&env, d);
1250 /* osp_precreate_send() sets opd_pre_status
1251 * in case of error, that prevent the using of
1253 if (rc < 0 && rc != -ENOSPC &&
1254 rc != -ETIMEDOUT && rc != -ENOTCONN)
1255 CERROR("%s: cannot precreate objects:"
1257 d->opd_obd->obd_name, rc);
1262 thread->t_flags = SVC_STOPPED;
1264 wake_up(&thread->t_ctl_waitq);
1270 * Check when to stop to wait for precreate objects.
1272 * The caller wanting a new OST object can't wait undefinitely. The
1273 * function checks for few conditions including available new OST
1274 * objects, disconnected OST, lack of space with no pending destroys,
1275 * etc. IOW, it checks whether the current OSP state is good to keep
1276 * waiting or it's better to give up.
1278 * \param[in] env LU environment provided by the caller
1279 * \param[in] d OSP device
1281 * \retval 0 - keep waiting, 1 - no luck
1283 static int osp_precreate_ready_condition(const struct lu_env *env,
1284 struct osp_device *d)
1286 if (d->opd_pre_recovering)
1289 /* ready if got enough precreated objects */
1290 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1291 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1294 /* ready if OST reported no space and no destroys in progress */
1295 if (atomic_read(&d->opd_syn_changes) +
1296 atomic_read(&d->opd_syn_rpc_in_progress) == 0 &&
1297 d->opd_pre_status == -ENOSPC)
1300 /* Bail out I/O fails to OST */
1301 if (d->opd_pre_status != 0 &&
1302 d->opd_pre_status != -EAGAIN &&
1303 d->opd_pre_status != -ENODEV &&
1304 d->opd_pre_status != -ENOTCONN &&
1305 d->opd_pre_status != -ENOSPC) {
1307 if (d->opd_pre_status != -EIO)
1308 CERROR("%s: precreate failed opd_pre_status %d\n",
1309 d->opd_obd->obd_name, d->opd_pre_status);
1316 static int osp_precreate_timeout_condition(void *data)
1318 struct osp_device *d = data;
1320 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1321 "reserved=%llu, syn_changes=%u, "
1322 "syn_rpc_in_progress=%d, status=%d\n",
1323 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1324 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1325 atomic_read(&d->opd_syn_changes),
1326 atomic_read(&d->opd_syn_rpc_in_progress),
1333 * Reserve object in precreate pool
1335 * When the caller wants to create a new object on this target (target
1336 * represented by the given OSP), it should declare this intention using
1337 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1338 * to reserve an object in the existing precreated pool or wait up to
1339 * obd_timeout for the available object to appear in the pool (a dedicated
1340 * thread will be doing real precreation in background). The object can be
1341 * consumed later with osp_precreate_get_fid() or be released with call to
1342 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1343 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1344 * If the space on the target is short and there is a pending object destroy,
1345 * then the function forces local commit to speedup space release (see
1346 * osp_sync.c for the details).
1348 * \param[in] env LU environment provided by the caller
1349 * \param[in] d OSP device
1351 * \retval 0 on success
1352 * \retval -ENOSPC when no space on OST
1353 * \retval -EAGAIN try later, slow precreation in progress
1354 * \retval -EIO when no access to OST
1356 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1358 struct l_wait_info lwi;
1359 cfs_time_t expire = cfs_time_shift(obd_timeout);
1364 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1365 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1366 PFID(&d->opd_pre_used_fid));
1368 /* opd_pre_max_create_count 0 to not use specified OST. */
1369 if (d->opd_pre_max_create_count == 0)
1372 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_OSP_PRECREATE_WAIT)) {
1373 if (d->opd_index == cfs_fail_val)
1374 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_OSP_PRECREATE_WAIT,
1380 * - preallocation is done
1381 * - no free space expected soon
1382 * - can't connect to OST for too long (obd_timeout)
1383 * - OST can allocate fid sequence.
1385 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1386 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1389 * increase number of precreations
1391 precreated = osp_objs_precreated(env, d);
1392 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1393 d->opd_pre_create_slow == 0 &&
1394 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1395 spin_lock(&d->opd_pre_lock);
1396 d->opd_pre_create_slow = 1;
1397 d->opd_pre_create_count *= 2;
1398 spin_unlock(&d->opd_pre_lock);
1401 spin_lock(&d->opd_pre_lock);
1402 precreated = osp_objs_precreated(env, d);
1403 if (precreated > d->opd_pre_reserved &&
1404 !d->opd_pre_recovering) {
1405 d->opd_pre_reserved++;
1406 spin_unlock(&d->opd_pre_lock);
1409 /* XXX: don't wake up if precreation is in progress */
1410 if (osp_precreate_near_empty_nolock(env, d) &&
1411 !osp_precreate_end_seq_nolock(env, d))
1412 wake_up(&d->opd_pre_waitq);
1416 spin_unlock(&d->opd_pre_lock);
1419 * all precreated objects have been used and no-space
1420 * status leave us no chance to succeed very soon
1421 * but if there is destroy in progress, then we should
1422 * wait till that is done - some space might be released
1424 if (unlikely(rc == -ENOSPC)) {
1425 if (atomic_read(&d->opd_syn_changes)) {
1426 /* force local commit to release space */
1427 dt_commit_async(env, d->opd_storage);
1429 if (atomic_read(&d->opd_syn_rpc_in_progress)) {
1430 /* just wait till destroys are done */
1431 /* see l_wait_even() few lines below */
1433 if (atomic_read(&d->opd_syn_changes) +
1434 atomic_read(&d->opd_syn_rpc_in_progress) == 0) {
1435 /* no hope for free space */
1440 /* XXX: don't wake up if precreation is in progress */
1441 wake_up(&d->opd_pre_waitq);
1443 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1444 osp_precreate_timeout_condition, d);
1445 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1450 l_wait_event(d->opd_pre_user_waitq,
1451 osp_precreate_ready_condition(env, d), &lwi);
1458 * Get a FID from precreation pool
1460 * The function is a companion for osp_precreate_reserve() - it assigns
1461 * a specific FID from the precreate. The function should be called only
1462 * if the call to osp_precreate_reserve() was successful. The function
1463 * updates a local storage to remember the highest object ID referenced
1464 * by the node in the given sequence.
1466 * A very importan details: this is supposed to be called once the
1467 * transaction is started, so on-disk update will be atomic with the
1468 * data (like LOVEA) refering this object. Then the object won't be leaked:
1469 * either it's referenced by the committed transaction or it's a subject
1470 * to the orphan cleanup procedure.
1472 * \param[in] env LU environment provided by the caller
1473 * \param[in] d OSP device
1474 * \param[out] fid generated FID
1476 * \retval 0 on success
1477 * \retval negative negated errno on error
1479 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1482 /* grab next id from the pool */
1483 spin_lock(&d->opd_pre_lock);
1485 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1486 &d->opd_pre_last_created_fid) < 0,
1487 "next fid "DFID" last created fid "DFID"\n",
1488 PFID(&d->opd_pre_used_fid),
1489 PFID(&d->opd_pre_last_created_fid));
1491 d->opd_pre_used_fid.f_oid++;
1492 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1493 d->opd_pre_reserved--;
1495 * last_used_id must be changed along with getting new id otherwise
1496 * we might miscalculate gap causing object loss or leak
1498 osp_update_last_fid(d, fid);
1499 spin_unlock(&d->opd_pre_lock);
1502 * probably main thread suspended orphan cleanup till
1503 * all reservations are released, see comment in
1504 * osp_precreate_thread() just before orphan cleanup
1506 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1507 wake_up(&d->opd_pre_waitq);
1513 * Set size regular attribute on an object
1515 * When a striping is created late, it's possible that size is already
1516 * initialized on the file. Then the new striping should inherit size
1517 * from the file. The function sets size on the object using the regular
1518 * protocol (OST_PUNCH).
1519 * XXX: should be re-implemented using OUT ?
1521 * \param[in] env LU environment provided by the caller
1522 * \param[in] dt object
1523 * \param[in] size size to set.
1525 * \retval 0 on success
1526 * \retval negative negated errno on error
1528 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1531 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1532 struct ptlrpc_request *req = NULL;
1533 struct obd_import *imp;
1534 struct ost_body *body;
1535 struct obdo *oa = NULL;
1540 imp = d->opd_obd->u.cli.cl_import;
1543 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1547 /* XXX: capa support? */
1548 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1549 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1551 ptlrpc_request_free(req);
1556 * XXX: decide how do we do here with resend
1557 * if we don't resend, then client may see wrong file size
1558 * if we do resend, then MDS thread can get stuck for quite long
1559 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1560 * (see LU-7975 and sanity/test_27F use cases)
1561 * but let's decide not to resend/delay this truncate request to OST
1562 * and allow Client to decide to resend, in a less agressive way from
1563 * after_reply(), by returning -EINPROGRESS instead of
1564 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1565 * end of this routine
1567 req->rq_no_resend = req->rq_no_delay = 1;
1569 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1570 ptlrpc_at_set_req_timeout(req);
1574 GOTO(out, rc = -ENOMEM);
1576 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1579 oa->o_blocks = OBD_OBJECT_EOF;
1580 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1581 OBD_MD_FLID | OBD_MD_FLGROUP;
1583 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1585 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1587 /* XXX: capa support? */
1588 /* osc_pack_capa(req, body, capa); */
1590 ptlrpc_request_set_replen(req);
1592 rc = ptlrpc_queue_wait(req);
1594 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1595 * since we have decided not to resend/delay, but this could
1596 * lead to wrong size to be seen at Client side and even process
1597 * trying to open to exit/fail if not itself handling -EAGAIN.
1598 * So it should be better to return -EINPROGRESS instead and
1599 * leave the decision to resend at Client side in after_reply()
1601 if (rc == -EWOULDBLOCK) {
1603 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1604 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1607 CERROR("can't punch object: %d\n", rc);
1611 ptlrpc_req_finished(req);
1618 * Initialize precreation functionality of OSP
1620 * Prepares all the internal structures and starts the precreate thread
1622 * \param[in] d OSP device
1624 * \retval 0 on success
1625 * \retval negative negated errno on error
1627 int osp_init_precreate(struct osp_device *d)
1629 struct l_wait_info lwi = { 0 };
1630 struct task_struct *task;
1634 OBD_ALLOC_PTR(d->opd_pre);
1635 if (d->opd_pre == NULL)
1638 /* initially precreation isn't ready */
1639 d->opd_pre_status = -EAGAIN;
1640 fid_zero(&d->opd_pre_used_fid);
1641 d->opd_pre_used_fid.f_oid = 1;
1642 fid_zero(&d->opd_pre_last_created_fid);
1643 d->opd_pre_last_created_fid.f_oid = 1;
1644 d->opd_pre_reserved = 0;
1645 d->opd_got_disconnected = 1;
1646 d->opd_pre_create_slow = 0;
1647 d->opd_pre_create_count = OST_MIN_PRECREATE;
1648 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1649 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1650 d->opd_reserved_mb_high = 0;
1651 d->opd_reserved_mb_low = 0;
1653 spin_lock_init(&d->opd_pre_lock);
1654 init_waitqueue_head(&d->opd_pre_waitq);
1655 init_waitqueue_head(&d->opd_pre_user_waitq);
1656 thread_set_flags(&d->opd_pre_thread, SVC_INIT);
1657 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1660 * Initialize statfs-related things
1662 d->opd_statfs_maxage = 5; /* default update interval */
1663 d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1664 CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1665 (unsigned long long)cfs_time_current(),
1666 (unsigned long long)d->opd_statfs_fresh_till);
1667 setup_timer(&d->opd_statfs_timer, osp_statfs_timer_cb,
1670 if (d->opd_storage->dd_rdonly)
1674 * start thread handling precreation and statfs updates
1676 task = kthread_run(osp_precreate_thread, d,
1677 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1679 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1680 RETURN(PTR_ERR(task));
1683 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1684 osp_precreate_running(d) || osp_precreate_stopped(d),
1691 * Finish precreate functionality of OSP
1694 * Asks all the activity (the thread, update timer) to stop, then
1695 * wait till that is done.
1697 * \param[in] d OSP device
1699 void osp_precreate_fini(struct osp_device *d)
1701 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1704 del_timer(&d->opd_statfs_timer);
1706 if (d->opd_pre == NULL)
1709 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1710 thread->t_flags = SVC_STOPPING;
1711 wake_up(&d->opd_pre_waitq);
1712 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1715 OBD_FREE_PTR(d->opd_pre);