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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osp/osp_sync.c
38 * Lustre OST Proxy Device
40 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
41 * Author: Mikhail Pershin <mike.pershin@intel.com>
42 * Author: Di Wang <di.wang@intel.com>
45 #define DEBUG_SUBSYSTEM S_MDS
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 cfs_timer_arm(&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 cfs_timer_disarm(&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, PDL_POLICY_ROUND, -1);
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 then 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 cfs_timer_disarm(&d->opd_statfs_timer);
237 wake_up(&d->opd_pre_waitq);
242 * Return number of precreated objects
244 * A simple helper to calculate the number of precreated objects on the device.
246 * \param[in] env LU environment provided by the caller
247 * \param[in] osp OSP device
249 * \retval the number of the precreated objects
251 static inline int osp_objs_precreated(const struct lu_env *env,
252 struct osp_device *osp)
254 return osp_fid_diff(&osp->opd_pre_last_created_fid,
255 &osp->opd_pre_used_fid);
259 * Check pool of precreated objects is nearly empty
261 * We should not wait till the pool of the precreated objects is exhausted,
262 * because then there will be a long period of OSP being unavailable for the
263 * new creations due to lenghty precreate RPC. Instead we ask for another
264 * precreation ahead and hopefully have it ready before the current pool is
265 * empty. Notice this function relies on an external locking.
267 * \param[in] env LU environment provided by the caller
268 * \param[in] d OSP device
270 * \retval 0 - current pool is good enough, 1 - time to precreate
272 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
273 struct osp_device *d)
275 int window = osp_objs_precreated(env, d);
277 /* don't consider new precreation till OST is healty and
279 return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) &&
280 (d->opd_pre_status == 0));
284 * Check pool of precreated objects
286 * This is protected version of osp_precreate_near_empty_nolock(), check that
289 * \param[in] env LU environment provided by the caller
290 * \param[in] d OSP device
292 * \retval 0 - current pool is good enough, 1 - time to precreate
294 static inline int osp_precreate_near_empty(const struct lu_env *env,
295 struct osp_device *d)
299 /* XXX: do we really need locking here? */
300 spin_lock(&d->opd_pre_lock);
301 rc = osp_precreate_near_empty_nolock(env, d);
302 spin_unlock(&d->opd_pre_lock);
307 * Check given sequence is empty
309 * Returns a binary result whether the given sequence has some IDs left
310 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
311 * version of that function.
313 * \param[in] env LU environment provided by the caller
314 * \param[in] osp OSP device
316 * \retval 0 - current sequence has no IDs, 1 - otherwise
318 static inline int osp_create_end_seq(const struct lu_env *env,
319 struct osp_device *osp)
321 struct lu_fid *fid = &osp->opd_pre_used_fid;
324 spin_lock(&osp->opd_pre_lock);
325 rc = osp_fid_end_seq(env, fid);
326 spin_unlock(&osp->opd_pre_lock);
331 * Write FID into into last_oid/last_seq file
333 * The function stores the sequence and the in-sequence id into two dedicated
334 * files. The sync argument can be used to request synchronous commit, so the
335 * function won't return until the updates are committed.
337 * \param[in] env LU environment provided by the caller
338 * \param[in] osp OSP device
339 * \param[in] fid fid where sequence/id is taken
340 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
342 * \retval 0 on success
343 * \retval negative negated errno on error
345 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
346 struct lu_fid *fid, int sync)
348 struct osp_thread_info *oti = osp_env_info(env);
349 struct lu_buf *lb_oid = &oti->osi_lb;
350 struct lu_buf *lb_oseq = &oti->osi_lb2;
357 /* Note: through f_oid is only 32 bits, it will also write 64 bits
358 * for oid to keep compatibility with the previous version. */
359 lb_oid->lb_buf = &fid->f_oid;
360 lb_oid->lb_len = sizeof(u64);
361 oid_off = sizeof(u64) * osp->opd_index;
363 lb_oseq->lb_buf = &fid->f_seq;
364 lb_oseq->lb_len = sizeof(u64);
365 oseq_off = sizeof(u64) * osp->opd_index;
367 th = dt_trans_create(env, osp->opd_storage);
372 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
373 lb_oid, oid_off, th);
377 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
378 lb_oseq, oseq_off, th);
382 rc = dt_trans_start_local(env, osp->opd_storage, th);
386 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
389 CERROR("%s: can not write to last seq file: rc = %d\n",
390 osp->opd_obd->obd_name, rc);
393 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
396 CERROR("%s: can not write to last seq file: rc = %d\n",
397 osp->opd_obd->obd_name, rc);
401 dt_trans_stop(env, osp->opd_storage, th);
406 * Switch to another sequence
408 * When a current sequence has no available IDs left, OSP has to switch to
409 * another new sequence. OSP requests it using the regular FLDB protocol
410 * and stores synchronously before that is used in precreated. This is needed
411 * to basically have the sequences referenced (not orphaned), otherwise it's
412 * possible that OST has some objects precreated and the clients have data
413 * written to it, but after MDT failover nobody refers those objects and OSP
414 * has no idea that the sequence need cleanup to be done.
415 * While this is very expensive operation, it's supposed to happen very very
416 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
418 * \param[in] env LU environment provided by the caller
419 * \param[in] osp OSP device
421 * \retval 0 on success
422 * \retval negative negated errno on error
424 static int osp_precreate_rollover_new_seq(struct lu_env *env,
425 struct osp_device *osp)
427 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
428 struct lu_fid *last_fid = &osp->opd_last_used_fid;
432 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
434 CERROR("%s: alloc fid error: rc = %d\n",
435 osp->opd_obd->obd_name, rc);
441 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
442 "fid "DFID", last_fid "DFID"\n", PFID(fid),
445 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
447 CERROR("%s: Can not update oid/seq file: rc = %d\n",
448 osp->opd_obd->obd_name, rc);
452 LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n",
453 osp->opd_obd->obd_name, fid_seq(last_fid),
455 /* Update last_xxx to the new seq */
456 spin_lock(&osp->opd_pre_lock);
457 osp->opd_last_used_fid = *fid;
458 osp->opd_gap_start_fid = *fid;
459 osp->opd_pre_used_fid = *fid;
460 osp->opd_pre_last_created_fid = *fid;
461 spin_unlock(&osp->opd_pre_lock);
467 * Find IDs available in current sequence
469 * The function calculates the highest possible ID and the number of IDs
470 * available in the current sequence OSP is using. The number is limited
471 * artifically by the caller (grow param) and the number of IDs available
472 * in the sequence by nature. The function doesn't require an external
475 * \param[in] env LU environment provided by the caller
476 * \param[in] osp OSP device
477 * \param[in] fid FID the caller wants to start with
478 * \param[in] grow how many the caller wants
479 * \param[out] fid the highest calculated FID
480 * \param[out] grow the number of available IDs calculated
482 * \retval 0 on success, 1 - the sequence is empty
484 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
485 struct lu_fid *fid, int *grow)
487 struct osp_thread_info *osi = osp_env_info(env);
491 if (fid_is_idif(fid)) {
492 struct lu_fid *last_fid;
493 struct ost_id *oi = &osi->osi_oi;
495 spin_lock(&osp->opd_pre_lock);
496 last_fid = &osp->opd_pre_last_created_fid;
497 fid_to_ostid(last_fid, oi);
498 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
499 *grow = end - ostid_id(oi);
500 ostid_set_id(oi, ostid_id(oi) + *grow);
501 spin_unlock(&osp->opd_pre_lock);
506 ostid_to_fid(fid, oi, osp->opd_index);
510 spin_lock(&osp->opd_pre_lock);
511 *fid = osp->opd_pre_last_created_fid;
513 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
514 *grow = end - fid->f_oid;
515 fid->f_oid += end - fid->f_oid;
516 spin_unlock(&osp->opd_pre_lock);
518 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
519 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
521 return *grow > 0 ? 0 : 1;
525 * Prepare and send precreate RPC
527 * The function finds how many objects should be precreated. Then allocates,
528 * prepares and schedules precreate RPC synchronously. Upon reply the function
529 * wake ups the threads waiting for the new objects on this target. If the
530 * target wasn't able to create all the objects requested, then the next
531 * precreate will be asking less objects (i.e. slow precreate down).
533 * \param[in] env LU environment provided by the caller
534 * \param[in] d OSP device
536 * \retval 0 on success
537 * \retval negative negated errno on error
539 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
541 struct osp_thread_info *oti = osp_env_info(env);
542 struct ptlrpc_request *req;
543 struct obd_import *imp;
544 struct ost_body *body;
546 struct lu_fid *fid = &oti->osi_fid;
549 /* don't precreate new objects till OST healthy and has free space */
550 if (unlikely(d->opd_pre_status)) {
551 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
552 d->opd_obd->obd_name, d->opd_pre_status);
557 * if not connection/initialization is compeleted, ignore
559 imp = d->opd_obd->u.cli.cl_import;
562 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
565 req->rq_request_portal = OST_CREATE_PORTAL;
566 /* we should not resend create request - anyway we will have delorphan
567 * and kill these objects */
568 req->rq_no_delay = req->rq_no_resend = 1;
570 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
572 ptlrpc_request_free(req);
576 spin_lock(&d->opd_pre_lock);
577 if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
578 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
579 grow = d->opd_pre_grow_count;
580 spin_unlock(&d->opd_pre_lock);
582 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
585 *fid = d->opd_pre_last_created_fid;
586 rc = osp_precreate_fids(env, d, fid, &grow);
588 /* Current seq has been used up*/
589 if (!osp_is_fid_client(d)) {
590 osp_pre_update_status(d, -ENOSPC);
593 wake_up(&d->opd_pre_waitq);
597 if (!osp_is_fid_client(d)) {
598 /* Non-FID client will always send seq 0 because of
600 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
604 fid_to_ostid(fid, &body->oa.o_oi);
605 body->oa.o_valid = OBD_MD_FLGROUP;
607 ptlrpc_request_set_replen(req);
609 rc = ptlrpc_queue_wait(req);
611 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
615 LASSERT(req->rq_transno == 0);
617 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
619 GOTO(out_req, rc = -EPROTO);
621 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
622 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
623 CERROR("%s: precreate fid "DFID" < local used fid "DFID
624 ": rc = %d\n", d->opd_obd->obd_name,
625 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
626 GOTO(out_req, rc = -ESTALE);
629 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
631 spin_lock(&d->opd_pre_lock);
633 /* the OST has not managed to create all the
634 * objects we asked for */
635 d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
636 d->opd_pre_grow_slow = 1;
638 /* the OST is able to keep up with the work,
639 * we could consider increasing grow_count
640 * next time if needed */
641 d->opd_pre_grow_slow = 0;
644 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
645 fid_to_ostid(fid, &body->oa.o_oi);
647 d->opd_pre_last_created_fid = *fid;
648 spin_unlock(&d->opd_pre_lock);
650 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
651 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
652 PFID(&d->opd_pre_last_created_fid));
654 /* now we can wakeup all users awaiting for objects */
655 osp_pre_update_status(d, rc);
656 wake_up(&d->opd_pre_user_waitq);
658 ptlrpc_req_finished(req);
663 * Get last precreated object from target (OST)
665 * Sends synchronous RPC to the target (OST) to learn the last precreated
666 * object. This later is used to remove all unused objects (cleanup orphan
667 * procedure). Also, the next object after one we got will be used as a
668 * starting point for the new precreates.
670 * \param[in] env LU environment provided by the caller
671 * \param[in] d OSP device
673 * \retval 0 on success
674 * \retval negative negated errno on error
676 static int osp_get_lastfid_from_ost(const struct lu_env *env,
677 struct osp_device *d)
679 struct ptlrpc_request *req = NULL;
680 struct obd_import *imp;
681 struct lu_fid *last_fid;
686 imp = d->opd_obd->u.cli.cl_import;
689 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
693 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
694 sizeof(KEY_LAST_FID));
696 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
698 ptlrpc_request_free(req);
702 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
703 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
705 req->rq_no_delay = req->rq_no_resend = 1;
706 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
707 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
709 ptlrpc_request_set_replen(req);
711 rc = ptlrpc_queue_wait(req);
713 /* bad-bad OST.. let sysadm sort this out */
714 if (rc == -ENOTSUPP) {
715 CERROR("%s: server does not support FID: rc = %d\n",
716 d->opd_obd->obd_name, -ENOTSUPP);
718 ptlrpc_set_import_active(imp, 0);
722 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
723 if (last_fid == NULL) {
724 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
725 GOTO(out, rc = -EPROTO);
728 if (!fid_is_sane(last_fid)) {
729 CERROR("%s: Got insane last_fid "DFID"\n",
730 d->opd_obd->obd_name, PFID(last_fid));
731 GOTO(out, rc = -EPROTO);
734 /* Only update the last used fid, if the OST has objects for
735 * this sequence, i.e. fid_oid > 0 */
736 if (fid_oid(last_fid) > 0)
737 d->opd_last_used_fid = *last_fid;
739 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
743 ptlrpc_req_finished(req);
748 * Cleanup orphans on OST
750 * This function is called in a contex of a dedicated thread handling
751 * all the precreation suff. The function waits till local recovery
752 * is complete, then identify all the unreferenced objects (orphans)
753 * using the highest ID referenced by a local and the highest object
754 * precreated by the target. The found range is a subject to removal
755 * using specially flagged RPC. During this process OSP is marked
756 * unavailable for new objects.
758 * \param[in] env LU environment provided by the caller
759 * \param[in] d OSP device
761 * \retval 0 on success
762 * \retval negative negated errno on error
764 static int osp_precreate_cleanup_orphans(struct lu_env *env,
765 struct osp_device *d)
767 struct osp_thread_info *osi = osp_env_info(env);
768 struct lu_fid *last_fid = &osi->osi_fid;
769 struct ptlrpc_request *req = NULL;
770 struct obd_import *imp;
771 struct ost_body *body;
772 struct l_wait_info lwi = { 0 };
773 int update_status = 0;
780 * wait for local recovery to finish, so we can cleanup orphans
781 * orphans are all objects since "last used" (assigned), but
782 * there might be objects reserved and in some cases they won't
783 * be used. we can't cleanup them till we're sure they won't be
784 * used. also can't we allow new reservations because they may
785 * end up getting orphans being cleaned up below. so we block
786 * new reservations and wait till all reserved objects either
789 spin_lock(&d->opd_pre_lock);
790 d->opd_pre_recovering = 1;
791 spin_unlock(&d->opd_pre_lock);
793 * The locking above makes sure the opd_pre_reserved check below will
794 * catch all osp_precreate_reserve() calls who find
795 * "!opd_pre_recovering".
797 l_wait_event(d->opd_pre_waitq,
798 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
799 !osp_precreate_running(d) || d->opd_got_disconnected,
801 if (!osp_precreate_running(d) || d->opd_got_disconnected)
802 GOTO(out, rc = -EAGAIN);
804 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
805 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
807 *last_fid = d->opd_last_used_fid;
808 /* The OSP should already get the valid seq now */
809 LASSERT(!fid_is_zero(last_fid));
810 if (fid_oid(&d->opd_last_used_fid) < 2) {
811 /* lastfid looks strange... ask OST */
812 rc = osp_get_lastfid_from_ost(env, d);
817 imp = d->opd_obd->u.cli.cl_import;
820 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
822 GOTO(out, rc = -ENOMEM);
824 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
826 ptlrpc_request_free(req);
831 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
833 GOTO(out, rc = -EPROTO);
835 body->oa.o_flags = OBD_FL_DELORPHAN;
836 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
838 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
840 ptlrpc_request_set_replen(req);
842 /* Don't resend the delorphan req */
843 req->rq_no_resend = req->rq_no_delay = 1;
845 rc = ptlrpc_queue_wait(req);
851 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
853 GOTO(out, rc = -EPROTO);
856 * OST provides us with id new pool starts from in body->oa.o_id
858 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
860 spin_lock(&d->opd_pre_lock);
861 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
863 d->opd_pre_grow_count = OST_MIN_PRECREATE + diff;
864 d->opd_pre_last_created_fid = d->opd_last_used_fid;
866 d->opd_pre_grow_count = OST_MIN_PRECREATE;
867 d->opd_pre_last_created_fid = *last_fid;
870 * This empties the pre-creation pool and effectively blocks any new
873 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
874 LUSTRE_DATA_SEQ_MAX_WIDTH);
875 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
876 d->opd_pre_grow_slow = 0;
877 spin_unlock(&d->opd_pre_lock);
879 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
880 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
881 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
884 ptlrpc_req_finished(req);
886 spin_lock(&d->opd_pre_lock);
887 d->opd_pre_recovering = 0;
888 spin_unlock(&d->opd_pre_lock);
891 * If rc is zero, the pre-creation window should have been emptied.
892 * Since waking up the herd would be useless without pre-created
893 * objects, we defer the signal to osp_precreate_send() in that case.
897 CERROR("%s: cannot cleanup orphans: rc = %d\n",
898 d->opd_obd->obd_name, rc);
899 /* we can't proceed from here, OST seem to
900 * be in a bad shape, better to wait for
901 * a new instance of the server and repeat
902 * from the beginning. notify possible waiters
903 * this OSP isn't quite functional yet */
904 osp_pre_update_status(d, rc);
906 wake_up(&d->opd_pre_user_waitq);
914 * Update precreate status using statfs data
916 * The function decides whether this OSP should be used for new objects.
917 * IOW, whether this OST is used up or has some free space. Cached statfs
918 * data is used to make this decision. If the latest result of statfs
919 * request (rc argument) is not success, then just mark OSP unavailable
922 * Add a bit of hysteresis so this flag isn't continually flapping,
923 * and ensure that new files don't get extremely fragmented due to
924 * only a small amount of available space in the filesystem.
925 * We want to set the NOSPC flag when there is less than ~0.1% free
926 * and clear it when there is at least ~0.2% free space, so:
927 * avail < ~0.1% max max = avail + used
928 * 1025 * avail < avail + used used = blocks - free
929 * 1024 * avail < used
930 * 1024 * avail < blocks - free
931 * avail < ((blocks - free) >> 10)
933 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
934 * lose that amount of space so in those cases we report no space left
935 * if their is less than 1 GB left.
936 * the function updates current precreation status used: functional or not
938 * \param[in] d OSP device
939 * \param[in] rc new precreate status for device \a d
941 * \retval 0 on success
942 * \retval negative negated errno on error
944 void osp_pre_update_status(struct osp_device *d, int rc)
946 struct obd_statfs *msfs = &d->opd_statfs;
947 int old = d->opd_pre_status;
950 d->opd_pre_status = rc;
954 if (likely(msfs->os_type)) {
955 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
957 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
958 d->opd_pre_status = -ENOSPC;
960 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
961 LPU64" free, "LPU64" used, "LPU64" "
962 "avail -> %d: rc = %d\n",
963 d->opd_obd->obd_name, msfs->os_blocks,
964 msfs->os_bfree, used, msfs->os_bavail,
965 d->opd_pre_status, rc);
967 "non-commited changes: %lu, in progress: %u\n",
968 d->opd_syn_changes, d->opd_syn_rpc_in_progress);
969 } else if (old == -ENOSPC) {
970 d->opd_pre_status = 0;
971 spin_lock(&d->opd_pre_lock);
972 d->opd_pre_grow_slow = 0;
973 d->opd_pre_grow_count = OST_MIN_PRECREATE;
974 spin_unlock(&d->opd_pre_lock);
975 wake_up(&d->opd_pre_waitq);
976 CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
977 " free, "LPU64" used, "LPU64" avail -> %d: "
978 "rc = %d\n", d->opd_obd->obd_name,
979 msfs->os_blocks, msfs->os_bfree, used,
980 msfs->os_bavail, d->opd_pre_status, rc);
985 wake_up(&d->opd_pre_user_waitq);
989 * Initialize FID for precreation
991 * For a just created new target, a new sequence should be taken.
992 * The function checks there is no IDIF in use (if the target was
993 * added with the older version of Lustre), then requests a new
994 * sequence from FLDB using the regular protocol. Then this new
995 * sequence is stored on a persisten storage synchronously to prevent
996 * possible object leakage (for the detail see the description for
997 * osp_precreate_rollover_new_seq()).
999 * \param[in] osp OSP device
1001 * \retval 0 on success
1002 * \retval negative negated errno on error
1004 int osp_init_pre_fid(struct osp_device *osp)
1007 struct osp_thread_info *osi;
1008 struct lu_client_seq *cli_seq;
1009 struct lu_fid *last_fid;
1013 LASSERT(osp->opd_pre != NULL);
1015 /* Return if last_used fid has been initialized */
1016 if (!fid_is_zero(&osp->opd_last_used_fid))
1019 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1021 CERROR("%s: init env error: rc = %d\n",
1022 osp->opd_obd->obd_name, rc);
1026 osi = osp_env_info(&env);
1027 last_fid = &osi->osi_fid;
1029 /* For a freshed fs, it will allocate a new sequence first */
1030 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1031 cli_seq = osp->opd_obd->u.cli.cl_seq;
1032 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1034 CERROR("%s: alloc fid error: rc = %d\n",
1035 osp->opd_obd->obd_name, rc);
1039 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1041 last_fid->f_oid = 1;
1042 last_fid->f_ver = 0;
1044 spin_lock(&osp->opd_pre_lock);
1045 osp->opd_last_used_fid = *last_fid;
1046 osp->opd_pre_used_fid = *last_fid;
1047 osp->opd_pre_last_created_fid = *last_fid;
1048 spin_unlock(&osp->opd_pre_lock);
1049 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1051 CERROR("%s: write fid error: rc = %d\n",
1052 osp->opd_obd->obd_name, rc);
1061 * The core of precreate functionality
1063 * The function implements the main precreation loop. Basically it
1064 * involves connecting to the target, precerate FID initialization,
1065 * identifying and removing orphans, then serving precreation. As
1066 * part of the latter, the thread is responsible for statfs data
1067 * updates. The precreation is mostly driven by another threads
1068 * asking for new OST objects - those askers wake the thread when
1069 * the number of precreated objects reach low watermark.
1070 * After a disconnect, the sequence above repeats. This is keep going
1071 * until the thread is requested to stop.
1073 * \param[in] _arg private data the thread (OSP device to handle)
1075 * \retval 0 on success
1076 * \retval negative negated errno on error
1078 static int osp_precreate_thread(void *_arg)
1080 struct osp_device *d = _arg;
1081 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1082 struct l_wait_info lwi = { 0 };
1088 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1090 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1095 spin_lock(&d->opd_pre_lock);
1096 thread->t_flags = SVC_RUNNING;
1097 spin_unlock(&d->opd_pre_lock);
1098 wake_up(&thread->t_ctl_waitq);
1100 while (osp_precreate_running(d)) {
1102 * need to be connected to OST
1104 while (osp_precreate_running(d)) {
1105 l_wait_event(d->opd_pre_waitq,
1106 !osp_precreate_running(d) ||
1107 d->opd_new_connection,
1110 if (!d->opd_new_connection)
1113 d->opd_new_connection = 0;
1114 d->opd_got_disconnected = 0;
1118 if (!osp_precreate_running(d))
1121 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1122 /* Sigh, fid client is not ready yet */
1123 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1126 /* Init fid for osp_precreate if necessary */
1127 rc = osp_init_pre_fid(d);
1129 class_export_put(d->opd_exp);
1130 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1131 CERROR("%s: init pre fid error: rc = %d\n",
1132 d->opd_obd->obd_name, rc);
1136 osp_statfs_update(d);
1139 * Clean up orphans or recreate missing objects.
1141 rc = osp_precreate_cleanup_orphans(&env, d);
1145 * connected, can handle precreates now
1147 while (osp_precreate_running(d)) {
1148 l_wait_event(d->opd_pre_waitq,
1149 !osp_precreate_running(d) ||
1150 osp_precreate_near_empty(&env, d) ||
1151 osp_statfs_need_update(d) ||
1152 d->opd_got_disconnected, &lwi);
1154 if (!osp_precreate_running(d))
1157 /* something happened to the connection
1158 * have to start from the beginning */
1159 if (d->opd_got_disconnected)
1162 if (osp_statfs_need_update(d))
1163 osp_statfs_update(d);
1165 /* To avoid handling different seq in precreate/orphan
1166 * cleanup, it will hold precreate until current seq is
1168 if (unlikely(osp_precreate_end_seq(&env, d) &&
1169 !osp_create_end_seq(&env, d)))
1172 if (unlikely(osp_precreate_end_seq(&env, d) &&
1173 osp_create_end_seq(&env, d))) {
1174 LCONSOLE_INFO("%s:"LPX64" is used up."
1175 " Update to new seq\n",
1176 d->opd_obd->obd_name,
1177 fid_seq(&d->opd_pre_last_created_fid));
1178 rc = osp_precreate_rollover_new_seq(&env, d);
1183 if (osp_precreate_near_empty(&env, d)) {
1184 rc = osp_precreate_send(&env, d);
1185 /* osp_precreate_send() sets opd_pre_status
1186 * in case of error, that prevent the using of
1188 if (rc < 0 && rc != -ENOSPC &&
1189 rc != -ETIMEDOUT && rc != -ENOTCONN)
1190 CERROR("%s: cannot precreate objects:"
1192 d->opd_obd->obd_name, rc);
1197 thread->t_flags = SVC_STOPPED;
1199 wake_up(&thread->t_ctl_waitq);
1205 * Check when to stop to wait for precreate objects.
1207 * The caller wanting a new OST object can't wait undefinitely. The
1208 * function checks for few conditions including available new OST
1209 * objects, disconnected OST, lack of space with no pending destroys,
1210 * etc. IOW, it checks whether the current OSP state is good to keep
1211 * waiting or it's better to give up.
1213 * \param[in] env LU environment provided by the caller
1214 * \param[in] d OSP device
1216 * \retval 0 - keep waiting, 1 - no luck
1218 static int osp_precreate_ready_condition(const struct lu_env *env,
1219 struct osp_device *d)
1221 if (d->opd_pre_recovering)
1224 /* ready if got enough precreated objects */
1225 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1226 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1229 /* ready if OST reported no space and no destroys in progress */
1230 if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
1231 d->opd_pre_status == -ENOSPC)
1234 /* Bail out I/O fails to OST */
1235 if (d->opd_pre_status != 0 &&
1236 d->opd_pre_status != -EAGAIN &&
1237 d->opd_pre_status != -ENODEV &&
1238 d->opd_pre_status != -ENOSPC) {
1240 if (d->opd_pre_status != -EIO)
1241 CERROR("%s: precreate failed opd_pre_status %d\n",
1242 d->opd_obd->obd_name, d->opd_pre_status);
1249 static int osp_precreate_timeout_condition(void *data)
1251 struct osp_device *d = data;
1253 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1254 "reserved="LPU64", syn_changes=%lu, "
1255 "syn_rpc_in_progress=%d, status=%d\n",
1256 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1257 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1258 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1265 * Reserve object in precreate pool
1267 * When the caller wants to create a new object on this target (target
1268 * represented by the given OSP), it should declare this intention using
1269 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1270 * to reserve an object in the existing precreated pool or wait up to
1271 * obd_timeout for the available object to appear in the pool (a dedicated
1272 * thread will be doing real precreation in background). The object can be
1273 * consumed later with osp_precreate_get_fid() or be released with call to
1274 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1275 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1276 * If the space on the target is short and there is a pending object destroy,
1277 * then the function forces local commit to speedup space release (see
1278 * osp_sync.c for the details).
1280 * \param[in] env LU environment provided by the caller
1281 * \param[in] d OSP device
1283 * \retval 0 on success
1284 * \retval -ENOSPC when no space on OST
1285 * \retval -EAGAIN try later, slow precreation in progress
1286 * \retval -EIO when no access to OST
1288 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1290 struct l_wait_info lwi;
1291 cfs_time_t expire = cfs_time_shift(obd_timeout);
1296 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1297 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1298 PFID(&d->opd_pre_used_fid));
1302 * - preallocation is done
1303 * - no free space expected soon
1304 * - can't connect to OST for too long (obd_timeout)
1305 * - OST can allocate fid sequence.
1307 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1308 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1311 * increase number of precreations
1313 precreated = osp_objs_precreated(env, d);
1314 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
1315 d->opd_pre_grow_slow == 0 &&
1316 precreated <= (d->opd_pre_grow_count / 4 + 1)) {
1317 spin_lock(&d->opd_pre_lock);
1318 d->opd_pre_grow_slow = 1;
1319 d->opd_pre_grow_count *= 2;
1320 spin_unlock(&d->opd_pre_lock);
1323 spin_lock(&d->opd_pre_lock);
1324 precreated = osp_objs_precreated(env, d);
1325 if (precreated > d->opd_pre_reserved &&
1326 !d->opd_pre_recovering) {
1327 d->opd_pre_reserved++;
1328 spin_unlock(&d->opd_pre_lock);
1331 /* XXX: don't wake up if precreation is in progress */
1332 if (osp_precreate_near_empty_nolock(env, d) &&
1333 !osp_precreate_end_seq_nolock(env, d))
1334 wake_up(&d->opd_pre_waitq);
1338 spin_unlock(&d->opd_pre_lock);
1341 * all precreated objects have been used and no-space
1342 * status leave us no chance to succeed very soon
1343 * but if there is destroy in progress, then we should
1344 * wait till that is done - some space might be released
1346 if (unlikely(rc == -ENOSPC)) {
1347 if (d->opd_syn_changes) {
1348 /* force local commit to release space */
1349 dt_commit_async(env, d->opd_storage);
1351 if (d->opd_syn_rpc_in_progress) {
1352 /* just wait till destroys are done */
1353 /* see l_wait_even() few lines below */
1355 if (d->opd_syn_changes +
1356 d->opd_syn_rpc_in_progress == 0) {
1357 /* no hope for free space */
1362 /* XXX: don't wake up if precreation is in progress */
1363 wake_up(&d->opd_pre_waitq);
1365 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1366 osp_precreate_timeout_condition, d);
1367 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1372 l_wait_event(d->opd_pre_user_waitq,
1373 osp_precreate_ready_condition(env, d), &lwi);
1380 * Get a FID from precreation pool
1382 * The function is a companion for osp_precreate_reserve() - it assigns
1383 * a specific FID from the precreate. The function should be called only
1384 * if the call to osp_precreate_reserve() was successful. The function
1385 * updates a local storage to remember the highest object ID referenced
1386 * by the node in the given sequence.
1388 * A very importan details: this is supposed to be called once the
1389 * transaction is started, so on-disk update will be atomic with the
1390 * data (like LOVEA) refering this object. Then the object won't be leaked:
1391 * either it's referenced by the committed transaction or it's a subject
1392 * to the orphan cleanup procedure.
1394 * \param[in] env LU environment provided by the caller
1395 * \param[in] d OSP device
1396 * \param[out] fid generated FID
1398 * \retval 0 on success
1399 * \retval negative negated errno on error
1401 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1404 /* grab next id from the pool */
1405 spin_lock(&d->opd_pre_lock);
1407 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1408 &d->opd_pre_last_created_fid) < 0,
1409 "next fid "DFID" last created fid "DFID"\n",
1410 PFID(&d->opd_pre_used_fid),
1411 PFID(&d->opd_pre_last_created_fid));
1413 d->opd_pre_used_fid.f_oid++;
1414 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1415 d->opd_pre_reserved--;
1417 * last_used_id must be changed along with getting new id otherwise
1418 * we might miscalculate gap causing object loss or leak
1420 osp_update_last_fid(d, fid);
1421 spin_unlock(&d->opd_pre_lock);
1424 * probably main thread suspended orphan cleanup till
1425 * all reservations are released, see comment in
1426 * osp_precreate_thread() just before orphan cleanup
1428 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1429 wake_up(&d->opd_pre_waitq);
1435 * Set size regular attribute on an object
1437 * When a striping is created late, it's possible that size is already
1438 * initialized on the file. Then the new striping should inherit size
1439 * from the file. The function sets size on the object using the regular
1440 * protocol (OST_PUNCH).
1441 * XXX: should be re-implemented using OUT ?
1443 * \param[in] env LU environment provided by the caller
1444 * \param[in] dt object
1445 * \param[in] size size to set.
1447 * \retval 0 on success
1448 * \retval negative negated errno on error
1450 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1453 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1454 struct ptlrpc_request *req = NULL;
1455 struct obd_import *imp;
1456 struct ost_body *body;
1457 struct obdo *oa = NULL;
1462 imp = d->opd_obd->u.cli.cl_import;
1465 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1469 /* XXX: capa support? */
1470 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1471 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1473 ptlrpc_request_free(req);
1478 * XXX: decide how do we do here with resend
1479 * if we don't resend, then client may see wrong file size
1480 * if we do resend, then MDS thread can get stuck for quite long
1482 req->rq_no_resend = req->rq_no_delay = 1;
1484 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1485 ptlrpc_at_set_req_timeout(req);
1489 GOTO(out, rc = -ENOMEM);
1491 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1494 oa->o_blocks = OBD_OBJECT_EOF;
1495 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1496 OBD_MD_FLID | OBD_MD_FLGROUP;
1498 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1500 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1502 /* XXX: capa support? */
1503 /* osc_pack_capa(req, body, capa); */
1505 ptlrpc_request_set_replen(req);
1507 rc = ptlrpc_queue_wait(req);
1509 CERROR("can't punch object: %d\n", rc);
1511 ptlrpc_req_finished(req);
1518 * Initialize precreation functionality of OSP
1520 * Prepares all the internal structures and starts the precreate thread
1522 * \param[in] d OSP device
1524 * \retval 0 on success
1525 * \retval negative negated errno on error
1527 int osp_init_precreate(struct osp_device *d)
1529 struct l_wait_info lwi = { 0 };
1530 struct task_struct *task;
1534 OBD_ALLOC_PTR(d->opd_pre);
1535 if (d->opd_pre == NULL)
1538 /* initially precreation isn't ready */
1539 d->opd_pre_status = -EAGAIN;
1540 fid_zero(&d->opd_pre_used_fid);
1541 d->opd_pre_used_fid.f_oid = 1;
1542 fid_zero(&d->opd_pre_last_created_fid);
1543 d->opd_pre_last_created_fid.f_oid = 1;
1544 d->opd_pre_reserved = 0;
1545 d->opd_got_disconnected = 1;
1546 d->opd_pre_grow_slow = 0;
1547 d->opd_pre_grow_count = OST_MIN_PRECREATE;
1548 d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
1549 d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
1551 spin_lock_init(&d->opd_pre_lock);
1552 init_waitqueue_head(&d->opd_pre_waitq);
1553 init_waitqueue_head(&d->opd_pre_user_waitq);
1554 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1557 * Initialize statfs-related things
1559 d->opd_statfs_maxage = 5; /* default update interval */
1560 d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1561 CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1562 (unsigned long long)cfs_time_current(),
1563 (unsigned long long)d->opd_statfs_fresh_till);
1564 cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
1567 * start thread handling precreation and statfs updates
1569 task = kthread_run(osp_precreate_thread, d,
1570 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1572 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1573 RETURN(PTR_ERR(task));
1576 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1577 osp_precreate_running(d) || osp_precreate_stopped(d),
1584 * Finish precreate functionality of OSP
1587 * Asks all the activity (the thread, update timer) to stop, then
1588 * wait till that is done.
1590 * \param[in] d OSP device
1592 void osp_precreate_fini(struct osp_device *d)
1594 struct ptlrpc_thread *thread;
1598 cfs_timer_disarm(&d->opd_statfs_timer);
1600 if (d->opd_pre == NULL)
1603 thread = &d->opd_pre_thread;
1605 thread->t_flags = SVC_STOPPING;
1606 wake_up(&d->opd_pre_waitq);
1608 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1610 OBD_FREE_PTR(d->opd_pre);