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 * 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_create_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 %#llx to %#llx\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 LASSERT(d->opd_pre->osp_pre_delorphan_sent != 0);
577 spin_lock(&d->opd_pre_lock);
578 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
579 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
580 grow = d->opd_pre_create_count;
581 spin_unlock(&d->opd_pre_lock);
583 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
586 *fid = d->opd_pre_last_created_fid;
587 rc = osp_precreate_fids(env, d, fid, &grow);
589 /* Current seq has been used up*/
590 if (!osp_is_fid_client(d)) {
591 osp_pre_update_status(d, -ENOSPC);
594 wake_up(&d->opd_pre_waitq);
598 if (!osp_is_fid_client(d)) {
599 /* Non-FID client will always send seq 0 because of
601 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
605 fid_to_ostid(fid, &body->oa.o_oi);
606 body->oa.o_valid = OBD_MD_FLGROUP;
608 ptlrpc_request_set_replen(req);
610 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
613 rc = ptlrpc_queue_wait(req);
615 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
619 LASSERT(req->rq_transno == 0);
621 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
623 GOTO(out_req, rc = -EPROTO);
625 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
628 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
629 CERROR("%s: precreate fid "DFID" < local used fid "DFID
630 ": rc = %d\n", d->opd_obd->obd_name,
631 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
632 GOTO(out_req, rc = -ESTALE);
635 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
637 spin_lock(&d->opd_pre_lock);
639 /* the OST has not managed to create all the
640 * objects we asked for */
641 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
642 d->opd_pre_create_slow = 1;
644 /* the OST is able to keep up with the work,
645 * we could consider increasing create_count
646 * next time if needed */
647 d->opd_pre_create_slow = 0;
650 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
651 fid_to_ostid(fid, &body->oa.o_oi);
653 d->opd_pre_last_created_fid = *fid;
654 spin_unlock(&d->opd_pre_lock);
656 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
657 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
658 PFID(&d->opd_pre_last_created_fid));
660 /* now we can wakeup all users awaiting for objects */
661 osp_pre_update_status(d, rc);
662 wake_up(&d->opd_pre_user_waitq);
664 ptlrpc_req_finished(req);
669 * Get last precreated object from target (OST)
671 * Sends synchronous RPC to the target (OST) to learn the last precreated
672 * object. This later is used to remove all unused objects (cleanup orphan
673 * procedure). Also, the next object after one we got will be used as a
674 * starting point for the new precreates.
676 * \param[in] env LU environment provided by the caller
677 * \param[in] d OSP device
679 * \retval 0 on success
680 * \retval negative negated errno on error
682 static int osp_get_lastfid_from_ost(const struct lu_env *env,
683 struct osp_device *d)
685 struct ptlrpc_request *req = NULL;
686 struct obd_import *imp;
687 struct lu_fid *last_fid;
692 imp = d->opd_obd->u.cli.cl_import;
695 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
699 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
700 sizeof(KEY_LAST_FID));
702 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
704 ptlrpc_request_free(req);
708 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
709 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
711 req->rq_no_delay = req->rq_no_resend = 1;
712 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
713 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
715 ptlrpc_request_set_replen(req);
717 rc = ptlrpc_queue_wait(req);
719 /* bad-bad OST.. let sysadm sort this out */
720 if (rc == -ENOTSUPP) {
721 CERROR("%s: server does not support FID: rc = %d\n",
722 d->opd_obd->obd_name, -ENOTSUPP);
724 ptlrpc_set_import_active(imp, 0);
728 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
729 if (last_fid == NULL) {
730 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
731 GOTO(out, rc = -EPROTO);
734 if (!fid_is_sane(last_fid)) {
735 CERROR("%s: Got insane last_fid "DFID"\n",
736 d->opd_obd->obd_name, PFID(last_fid));
737 GOTO(out, rc = -EPROTO);
740 /* Only update the last used fid, if the OST has objects for
741 * this sequence, i.e. fid_oid > 0 */
742 if (fid_oid(last_fid) > 0)
743 d->opd_last_used_fid = *last_fid;
745 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
749 ptlrpc_req_finished(req);
754 * Cleanup orphans on OST
756 * This function is called in a contex of a dedicated thread handling
757 * all the precreation suff. The function waits till local recovery
758 * is complete, then identify all the unreferenced objects (orphans)
759 * using the highest ID referenced by a local and the highest object
760 * precreated by the target. The found range is a subject to removal
761 * using specially flagged RPC. During this process OSP is marked
762 * unavailable for new objects.
764 * \param[in] env LU environment provided by the caller
765 * \param[in] d OSP device
767 * \retval 0 on success
768 * \retval negative negated errno on error
770 static int osp_precreate_cleanup_orphans(struct lu_env *env,
771 struct osp_device *d)
773 struct osp_thread_info *osi = osp_env_info(env);
774 struct lu_fid *last_fid = &osi->osi_fid;
775 struct ptlrpc_request *req = NULL;
776 struct obd_import *imp;
777 struct ost_body *body;
778 struct l_wait_info lwi = { 0 };
779 int update_status = 0;
787 * wait for local recovery to finish, so we can cleanup orphans.
788 * orphans are all objects since "last used" (assigned).
789 * consider reserved objects as created otherwise we can get into
790 * a livelock when one blocked thread holding a reservation can
791 * block recovery. see LU-8367 for the details. in some cases this
792 * can result in gaps (i.e. leaked objects), but we've got LFSCK...
794 * do not allow new reservations because they may end up getting
795 * orphans being cleaned up below. so we block new reservations.
797 spin_lock(&d->opd_pre_lock);
798 d->opd_pre_recovering = 1;
799 spin_unlock(&d->opd_pre_lock);
801 * The locking above makes sure the opd_pre_reserved check below will
802 * catch all osp_precreate_reserve() calls who find
803 * "!opd_pre_recovering".
805 l_wait_event(d->opd_pre_waitq, d->opd_recovery_completed ||
806 !osp_precreate_running(d) || d->opd_got_disconnected,
808 if (!osp_precreate_running(d) || d->opd_got_disconnected)
809 GOTO(out, rc = -EAGAIN);
811 *last_fid = d->opd_last_used_fid;
812 /* The OSP should already get the valid seq now */
813 LASSERT(!fid_is_zero(last_fid));
814 if (fid_oid(&d->opd_last_used_fid) < 2) {
815 /* lastfid looks strange... ask OST */
816 rc = osp_get_lastfid_from_ost(env, d);
821 imp = d->opd_obd->u.cli.cl_import;
824 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
826 GOTO(out, rc = -ENOMEM);
828 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
830 ptlrpc_request_free(req);
835 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
837 GOTO(out, rc = -EPROTO);
839 body->oa.o_flags = 0;
840 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
842 /* unless this is the very first DELORPHAN (when we really
843 * can destroy some orphans), just tell OST to recreate
844 * missing objects in our precreate pool */
845 spin_lock(&d->opd_pre_lock);
846 if (d->opd_pre->osp_pre_delorphan_sent) {
847 fid = d->opd_pre_last_created_fid;
849 fid = d->opd_last_used_fid;
850 body->oa.o_flags = OBD_FL_DELORPHAN;
852 spin_unlock(&d->opd_pre_lock);
853 fid_to_ostid(&fid, &body->oa.o_oi);
855 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
856 d->opd_obd->obd_name, PFID(&fid));
858 ptlrpc_request_set_replen(req);
860 /* Don't resend the delorphan req */
861 req->rq_no_resend = req->rq_no_delay = 1;
863 rc = ptlrpc_queue_wait(req);
869 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
871 GOTO(out, rc = -EPROTO);
874 * OST provides us with id new pool starts from in body->oa.o_id
876 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
878 spin_lock(&d->opd_pre_lock);
879 diff = osp_fid_diff(&fid, last_fid);
881 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
882 d->opd_pre_last_created_fid = *last_fid;
884 d->opd_pre_create_count = OST_MIN_PRECREATE;
885 d->opd_pre_last_created_fid = *last_fid;
888 * This empties the pre-creation pool and effectively blocks any new
891 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
892 LUSTRE_DATA_SEQ_MAX_WIDTH);
893 if (d->opd_pre->osp_pre_delorphan_sent == 0)
894 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
895 d->opd_pre_create_slow = 0;
896 spin_unlock(&d->opd_pre_lock);
897 d->opd_pre->osp_pre_delorphan_sent = 1;
899 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
900 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
901 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
904 ptlrpc_req_finished(req);
907 * If rc is zero, the pre-creation window should have been emptied.
908 * Since waking up the herd would be useless without pre-created
909 * objects, we defer the signal to osp_precreate_send() in that case.
913 CERROR("%s: cannot cleanup orphans: rc = %d\n",
914 d->opd_obd->obd_name, rc);
915 /* we can't proceed from here, OST seem to
916 * be in a bad shape, better to wait for
917 * a new instance of the server and repeat
918 * from the beginning. notify possible waiters
919 * this OSP isn't quite functional yet */
920 osp_pre_update_status(d, rc);
922 wake_up(&d->opd_pre_user_waitq);
925 spin_lock(&d->opd_pre_lock);
926 d->opd_pre_recovering = 0;
927 spin_unlock(&d->opd_pre_lock);
934 * Update precreate status using statfs data
936 * The function decides whether this OSP should be used for new objects.
937 * IOW, whether this OST is used up or has some free space. Cached statfs
938 * data is used to make this decision. If the latest result of statfs
939 * request (rc argument) is not success, then just mark OSP unavailable
942 * Add a bit of hysteresis so this flag isn't continually flapping,
943 * and ensure that new files don't get extremely fragmented due to
944 * only a small amount of available space in the filesystem.
945 * We want to set the ENOSPC when there is less than reserved size
946 * free and clear it when there is at least 2*reserved size free space.
947 * the function updates current precreation status used: functional or not
949 * \param[in] d OSP device
950 * \param[in] rc new precreate status for device \a d
952 * \retval 0 on success
953 * \retval negative negated errno on error
955 void osp_pre_update_status(struct osp_device *d, int rc)
957 struct obd_statfs *msfs = &d->opd_statfs;
958 int old = d->opd_pre_status;
961 d->opd_pre_status = rc;
965 if (likely(msfs->os_type)) {
966 if (unlikely(d->opd_reserved_mb_high == 0 &&
967 d->opd_reserved_mb_low == 0)) {
968 /* Use ~0.1% by default to disable object allocation,
969 * and ~0.2% to enable, size in MB, set both watermark
971 spin_lock(&d->opd_pre_lock);
972 if (d->opd_reserved_mb_high == 0 &&
973 d->opd_reserved_mb_low == 0) {
974 d->opd_reserved_mb_low =
975 ((msfs->os_bsize >> 10) *
976 msfs->os_blocks) >> 20;
977 if (d->opd_reserved_mb_low == 0)
978 d->opd_reserved_mb_low = 1;
979 d->opd_reserved_mb_high =
980 (d->opd_reserved_mb_low << 1) + 1;
982 spin_unlock(&d->opd_pre_lock);
985 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
986 if (msfs->os_ffree < 32)
987 msfs->os_state |= OS_STATE_ENOINO;
988 else if (msfs->os_ffree > 64)
989 msfs->os_state &= ~OS_STATE_ENOINO;
991 if (available < d->opd_reserved_mb_low)
992 msfs->os_state |= OS_STATE_ENOSPC;
993 else if (available > d->opd_reserved_mb_high)
994 msfs->os_state &= ~OS_STATE_ENOSPC;
995 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
996 d->opd_pre_status = -ENOSPC;
998 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
999 "free, %llu avail, %llu MB avail, %u "
1000 "hwm -> %d: rc = %d\n",
1001 d->opd_obd->obd_name, msfs->os_blocks,
1002 msfs->os_bfree, msfs->os_bavail,
1003 available, d->opd_reserved_mb_high,
1004 d->opd_pre_status, rc);
1006 "non-committed changes: %u, in progress: %u\n",
1007 atomic_read(&d->opd_syn_changes),
1008 atomic_read(&d->opd_syn_rpc_in_progress));
1009 } else if (unlikely(old == -ENOSPC)) {
1010 d->opd_pre_status = 0;
1011 spin_lock(&d->opd_pre_lock);
1012 d->opd_pre_create_slow = 0;
1013 d->opd_pre_create_count = OST_MIN_PRECREATE;
1014 spin_unlock(&d->opd_pre_lock);
1015 wake_up(&d->opd_pre_waitq);
1017 CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1018 " free, %llu avail, %lluMB avail, %u lwm"
1019 " -> %d: rc = %d\n", d->opd_obd->obd_name,
1020 msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1021 available, d->opd_reserved_mb_low,
1022 d->opd_pre_status, rc);
1026 wake_up(&d->opd_pre_user_waitq);
1030 * Initialize FID for precreation
1032 * For a just created new target, a new sequence should be taken.
1033 * The function checks there is no IDIF in use (if the target was
1034 * added with the older version of Lustre), then requests a new
1035 * sequence from FLDB using the regular protocol. Then this new
1036 * sequence is stored on a persisten storage synchronously to prevent
1037 * possible object leakage (for the detail see the description for
1038 * osp_precreate_rollover_new_seq()).
1040 * \param[in] osp OSP device
1042 * \retval 0 on success
1043 * \retval negative negated errno on error
1045 int osp_init_pre_fid(struct osp_device *osp)
1048 struct osp_thread_info *osi;
1049 struct lu_client_seq *cli_seq;
1050 struct lu_fid *last_fid;
1054 LASSERT(osp->opd_pre != NULL);
1056 /* Let's check if the current last_seq/fid is valid,
1057 * otherwise request new sequence from the controller */
1058 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1059 /* Non-MDT0 can only use normal sequence for
1061 if (fid_is_norm(&osp->opd_last_used_fid))
1064 /* Initially MDT0 will start with IDIF, after
1065 * that it will request new sequence from the
1067 if (fid_is_idif(&osp->opd_last_used_fid) ||
1068 fid_is_norm(&osp->opd_last_used_fid))
1072 if (!fid_is_zero(&osp->opd_last_used_fid))
1073 CWARN("%s: invalid last used fid "DFID
1074 ", try to get new sequence.\n",
1075 osp->opd_obd->obd_name,
1076 PFID(&osp->opd_last_used_fid));
1078 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1080 CERROR("%s: init env error: rc = %d\n",
1081 osp->opd_obd->obd_name, rc);
1085 osi = osp_env_info(&env);
1086 last_fid = &osi->osi_fid;
1088 /* For a freshed fs, it will allocate a new sequence first */
1089 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1090 cli_seq = osp->opd_obd->u.cli.cl_seq;
1091 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1093 CERROR("%s: alloc fid error: rc = %d\n",
1094 osp->opd_obd->obd_name, rc);
1098 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1100 last_fid->f_oid = 1;
1101 last_fid->f_ver = 0;
1103 spin_lock(&osp->opd_pre_lock);
1104 osp->opd_last_used_fid = *last_fid;
1105 osp->opd_pre_used_fid = *last_fid;
1106 osp->opd_pre_last_created_fid = *last_fid;
1107 spin_unlock(&osp->opd_pre_lock);
1108 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1110 CERROR("%s: write fid error: rc = %d\n",
1111 osp->opd_obd->obd_name, rc);
1120 * The core of precreate functionality
1122 * The function implements the main precreation loop. Basically it
1123 * involves connecting to the target, precerate FID initialization,
1124 * identifying and removing orphans, then serving precreation. As
1125 * part of the latter, the thread is responsible for statfs data
1126 * updates. The precreation is mostly driven by another threads
1127 * asking for new OST objects - those askers wake the thread when
1128 * the number of precreated objects reach low watermark.
1129 * After a disconnect, the sequence above repeats. This is keep going
1130 * until the thread is requested to stop.
1132 * \param[in] _arg private data the thread (OSP device to handle)
1134 * \retval 0 on success
1135 * \retval negative negated errno on error
1137 static int osp_precreate_thread(void *_arg)
1139 struct osp_device *d = _arg;
1140 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1141 struct l_wait_info lwi = { 0 };
1142 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1143 back_to_sleep, NULL);
1149 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1151 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1156 spin_lock(&d->opd_pre_lock);
1157 thread->t_flags = SVC_RUNNING;
1158 spin_unlock(&d->opd_pre_lock);
1159 wake_up(&thread->t_ctl_waitq);
1161 while (osp_precreate_running(d)) {
1163 * need to be connected to OST
1165 while (osp_precreate_running(d)) {
1166 if (d->opd_pre_recovering &&
1167 d->opd_imp_connected &&
1168 !d->opd_got_disconnected)
1170 l_wait_event(d->opd_pre_waitq,
1171 !osp_precreate_running(d) ||
1172 d->opd_new_connection,
1175 if (!d->opd_new_connection)
1178 d->opd_new_connection = 0;
1179 d->opd_got_disconnected = 0;
1183 if (!osp_precreate_running(d))
1186 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1187 /* Sigh, fid client is not ready yet */
1188 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1191 /* Init fid for osp_precreate if necessary */
1192 rc = osp_init_pre_fid(d);
1194 class_export_put(d->opd_exp);
1195 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1196 CERROR("%s: init pre fid error: rc = %d\n",
1197 d->opd_obd->obd_name, rc);
1201 if (osp_statfs_update(d)) {
1202 l_wait_event(d->opd_pre_waitq,
1203 !osp_precreate_running(d), &lwi2);
1208 * Clean up orphans or recreate missing objects.
1210 rc = osp_precreate_cleanup_orphans(&env, d);
1212 schedule_timeout_interruptible(cfs_time_seconds(1));
1216 * connected, can handle precreates now
1218 while (osp_precreate_running(d)) {
1219 l_wait_event(d->opd_pre_waitq,
1220 !osp_precreate_running(d) ||
1221 osp_precreate_near_empty(&env, d) ||
1222 osp_statfs_need_update(d) ||
1223 d->opd_got_disconnected, &lwi);
1225 if (!osp_precreate_running(d))
1228 /* something happened to the connection
1229 * have to start from the beginning */
1230 if (d->opd_got_disconnected)
1233 if (osp_statfs_need_update(d))
1234 if (osp_statfs_update(d))
1237 /* To avoid handling different seq in precreate/orphan
1238 * cleanup, it will hold precreate until current seq is
1240 if (unlikely(osp_precreate_end_seq(&env, d) &&
1241 !osp_create_end_seq(&env, d)))
1244 if (unlikely(osp_precreate_end_seq(&env, d) &&
1245 osp_create_end_seq(&env, d))) {
1246 LCONSOLE_INFO("%s:%#llx is used up."
1247 " Update to new seq\n",
1248 d->opd_obd->obd_name,
1249 fid_seq(&d->opd_pre_last_created_fid));
1250 rc = osp_precreate_rollover_new_seq(&env, d);
1255 if (osp_precreate_near_empty(&env, d)) {
1256 rc = osp_precreate_send(&env, d);
1257 /* osp_precreate_send() sets opd_pre_status
1258 * in case of error, that prevent the using of
1260 if (rc < 0 && rc != -ENOSPC &&
1261 rc != -ETIMEDOUT && rc != -ENOTCONN)
1262 CERROR("%s: cannot precreate objects:"
1264 d->opd_obd->obd_name, rc);
1269 thread->t_flags = SVC_STOPPED;
1271 wake_up(&thread->t_ctl_waitq);
1277 * Check when to stop to wait for precreate objects.
1279 * The caller wanting a new OST object can't wait undefinitely. The
1280 * function checks for few conditions including available new OST
1281 * objects, disconnected OST, lack of space with no pending destroys,
1282 * etc. IOW, it checks whether the current OSP state is good to keep
1283 * waiting or it's better to give up.
1285 * \param[in] env LU environment provided by the caller
1286 * \param[in] d OSP device
1288 * \retval 0 - keep waiting, 1 - no luck
1290 static int osp_precreate_ready_condition(const struct lu_env *env,
1291 struct osp_device *d)
1293 if (d->opd_pre_recovering)
1296 /* ready if got enough precreated objects */
1297 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1298 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1301 /* ready if OST reported no space and no destroys in progress */
1302 if (atomic_read(&d->opd_syn_changes) +
1303 atomic_read(&d->opd_syn_rpc_in_progress) == 0 &&
1304 d->opd_pre_status == -ENOSPC)
1307 /* Bail out I/O fails to OST */
1308 if (d->opd_pre_status != 0 &&
1309 d->opd_pre_status != -EAGAIN &&
1310 d->opd_pre_status != -ENODEV &&
1311 d->opd_pre_status != -ENOTCONN &&
1312 d->opd_pre_status != -ENOSPC) {
1314 if (d->opd_pre_status != -EIO)
1315 CERROR("%s: precreate failed opd_pre_status %d\n",
1316 d->opd_obd->obd_name, d->opd_pre_status);
1323 static int osp_precreate_timeout_condition(void *data)
1325 struct osp_device *d = data;
1327 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1328 "reserved=%llu, syn_changes=%u, "
1329 "syn_rpc_in_progress=%d, status=%d\n",
1330 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1331 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1332 atomic_read(&d->opd_syn_changes),
1333 atomic_read(&d->opd_syn_rpc_in_progress),
1340 * Reserve object in precreate pool
1342 * When the caller wants to create a new object on this target (target
1343 * represented by the given OSP), it should declare this intention using
1344 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1345 * to reserve an object in the existing precreated pool or wait up to
1346 * obd_timeout for the available object to appear in the pool (a dedicated
1347 * thread will be doing real precreation in background). The object can be
1348 * consumed later with osp_precreate_get_fid() or be released with call to
1349 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1350 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1351 * If the space on the target is short and there is a pending object destroy,
1352 * then the function forces local commit to speedup space release (see
1353 * osp_sync.c for the details).
1355 * \param[in] env LU environment provided by the caller
1356 * \param[in] d OSP device
1358 * \retval 0 on success
1359 * \retval -ENOSPC when no space on OST
1360 * \retval -EAGAIN try later, slow precreation in progress
1361 * \retval -EIO when no access to OST
1363 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1365 struct l_wait_info lwi;
1366 cfs_time_t expire = cfs_time_shift(obd_timeout);
1371 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1372 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1373 PFID(&d->opd_pre_used_fid));
1375 /* opd_pre_max_create_count 0 to not use specified OST. */
1376 if (d->opd_pre_max_create_count == 0)
1379 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_OSP_PRECREATE_WAIT)) {
1380 if (d->opd_index == cfs_fail_val)
1381 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_OSP_PRECREATE_WAIT,
1387 * - preallocation is done
1388 * - no free space expected soon
1389 * - can't connect to OST for too long (obd_timeout)
1390 * - OST can allocate fid sequence.
1392 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1393 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1396 * increase number of precreations
1398 precreated = osp_objs_precreated(env, d);
1399 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1400 d->opd_pre_create_slow == 0 &&
1401 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1402 spin_lock(&d->opd_pre_lock);
1403 d->opd_pre_create_slow = 1;
1404 d->opd_pre_create_count *= 2;
1405 spin_unlock(&d->opd_pre_lock);
1408 spin_lock(&d->opd_pre_lock);
1409 precreated = osp_objs_precreated(env, d);
1410 if (precreated > d->opd_pre_reserved &&
1411 !d->opd_pre_recovering) {
1412 d->opd_pre_reserved++;
1413 spin_unlock(&d->opd_pre_lock);
1416 /* XXX: don't wake up if precreation is in progress */
1417 if (osp_precreate_near_empty_nolock(env, d) &&
1418 !osp_precreate_end_seq_nolock(env, d))
1419 wake_up(&d->opd_pre_waitq);
1423 spin_unlock(&d->opd_pre_lock);
1426 * all precreated objects have been used and no-space
1427 * status leave us no chance to succeed very soon
1428 * but if there is destroy in progress, then we should
1429 * wait till that is done - some space might be released
1431 if (unlikely(rc == -ENOSPC)) {
1432 if (atomic_read(&d->opd_syn_changes)) {
1433 /* force local commit to release space */
1434 dt_commit_async(env, d->opd_storage);
1436 if (atomic_read(&d->opd_syn_rpc_in_progress)) {
1437 /* just wait till destroys are done */
1438 /* see l_wait_even() few lines below */
1440 if (atomic_read(&d->opd_syn_changes) +
1441 atomic_read(&d->opd_syn_rpc_in_progress) == 0) {
1442 /* no hope for free space */
1447 /* XXX: don't wake up if precreation is in progress */
1448 wake_up(&d->opd_pre_waitq);
1450 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1451 osp_precreate_timeout_condition, d);
1452 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1457 l_wait_event(d->opd_pre_user_waitq,
1458 osp_precreate_ready_condition(env, d), &lwi);
1465 * Get a FID from precreation pool
1467 * The function is a companion for osp_precreate_reserve() - it assigns
1468 * a specific FID from the precreate. The function should be called only
1469 * if the call to osp_precreate_reserve() was successful. The function
1470 * updates a local storage to remember the highest object ID referenced
1471 * by the node in the given sequence.
1473 * A very importan details: this is supposed to be called once the
1474 * transaction is started, so on-disk update will be atomic with the
1475 * data (like LOVEA) refering this object. Then the object won't be leaked:
1476 * either it's referenced by the committed transaction or it's a subject
1477 * to the orphan cleanup procedure.
1479 * \param[in] env LU environment provided by the caller
1480 * \param[in] d OSP device
1481 * \param[out] fid generated FID
1483 * \retval 0 on success
1484 * \retval negative negated errno on error
1486 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1489 /* grab next id from the pool */
1490 spin_lock(&d->opd_pre_lock);
1492 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1493 &d->opd_pre_last_created_fid) < 0,
1494 "next fid "DFID" last created fid "DFID"\n",
1495 PFID(&d->opd_pre_used_fid),
1496 PFID(&d->opd_pre_last_created_fid));
1498 d->opd_pre_used_fid.f_oid++;
1499 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1500 d->opd_pre_reserved--;
1502 * last_used_id must be changed along with getting new id otherwise
1503 * we might miscalculate gap causing object loss or leak
1505 osp_update_last_fid(d, fid);
1506 spin_unlock(&d->opd_pre_lock);
1509 * probably main thread suspended orphan cleanup till
1510 * all reservations are released, see comment in
1511 * osp_precreate_thread() just before orphan cleanup
1513 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1514 wake_up(&d->opd_pre_waitq);
1520 * Set size regular attribute on an object
1522 * When a striping is created late, it's possible that size is already
1523 * initialized on the file. Then the new striping should inherit size
1524 * from the file. The function sets size on the object using the regular
1525 * protocol (OST_PUNCH).
1526 * XXX: should be re-implemented using OUT ?
1528 * \param[in] env LU environment provided by the caller
1529 * \param[in] dt object
1530 * \param[in] size size to set.
1532 * \retval 0 on success
1533 * \retval negative negated errno on error
1535 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1538 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1539 struct ptlrpc_request *req = NULL;
1540 struct obd_import *imp;
1541 struct ost_body *body;
1542 struct obdo *oa = NULL;
1547 imp = d->opd_obd->u.cli.cl_import;
1550 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1554 /* XXX: capa support? */
1555 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1556 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1558 ptlrpc_request_free(req);
1563 * XXX: decide how do we do here with resend
1564 * if we don't resend, then client may see wrong file size
1565 * if we do resend, then MDS thread can get stuck for quite long
1566 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1567 * (see LU-7975 and sanity/test_27F use cases)
1568 * but let's decide not to resend/delay this truncate request to OST
1569 * and allow Client to decide to resend, in a less agressive way from
1570 * after_reply(), by returning -EINPROGRESS instead of
1571 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1572 * end of this routine
1574 req->rq_no_resend = req->rq_no_delay = 1;
1576 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1577 ptlrpc_at_set_req_timeout(req);
1581 GOTO(out, rc = -ENOMEM);
1583 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1586 oa->o_blocks = OBD_OBJECT_EOF;
1587 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1588 OBD_MD_FLID | OBD_MD_FLGROUP;
1590 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1592 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1594 /* XXX: capa support? */
1595 /* osc_pack_capa(req, body, capa); */
1597 ptlrpc_request_set_replen(req);
1599 rc = ptlrpc_queue_wait(req);
1601 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1602 * since we have decided not to resend/delay, but this could
1603 * lead to wrong size to be seen at Client side and even process
1604 * trying to open to exit/fail if not itself handling -EAGAIN.
1605 * So it should be better to return -EINPROGRESS instead and
1606 * leave the decision to resend at Client side in after_reply()
1608 if (rc == -EWOULDBLOCK) {
1610 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1611 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1614 CERROR("can't punch object: %d\n", rc);
1618 ptlrpc_req_finished(req);
1625 * Initialize precreation functionality of OSP
1627 * Prepares all the internal structures and starts the precreate thread
1629 * \param[in] d OSP device
1631 * \retval 0 on success
1632 * \retval negative negated errno on error
1634 int osp_init_precreate(struct osp_device *d)
1636 struct l_wait_info lwi = { 0 };
1637 struct task_struct *task;
1641 OBD_ALLOC_PTR(d->opd_pre);
1642 if (d->opd_pre == NULL)
1645 /* initially precreation isn't ready */
1646 d->opd_pre_status = -EAGAIN;
1647 fid_zero(&d->opd_pre_used_fid);
1648 d->opd_pre_used_fid.f_oid = 1;
1649 fid_zero(&d->opd_pre_last_created_fid);
1650 d->opd_pre_last_created_fid.f_oid = 1;
1651 d->opd_pre_reserved = 0;
1652 d->opd_got_disconnected = 1;
1653 d->opd_pre_create_slow = 0;
1654 d->opd_pre_create_count = OST_MIN_PRECREATE;
1655 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1656 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1657 d->opd_reserved_mb_high = 0;
1658 d->opd_reserved_mb_low = 0;
1660 spin_lock_init(&d->opd_pre_lock);
1661 init_waitqueue_head(&d->opd_pre_waitq);
1662 init_waitqueue_head(&d->opd_pre_user_waitq);
1663 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1666 * Initialize statfs-related things
1668 d->opd_statfs_maxage = 5; /* default update interval */
1669 d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1670 CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1671 (unsigned long long)cfs_time_current(),
1672 (unsigned long long)d->opd_statfs_fresh_till);
1673 setup_timer(&d->opd_statfs_timer, osp_statfs_timer_cb,
1677 * start thread handling precreation and statfs updates
1679 task = kthread_run(osp_precreate_thread, d,
1680 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1682 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1683 RETURN(PTR_ERR(task));
1686 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1687 osp_precreate_running(d) || osp_precreate_stopped(d),
1694 * Finish precreate functionality of OSP
1697 * Asks all the activity (the thread, update timer) to stop, then
1698 * wait till that is done.
1700 * \param[in] d OSP device
1702 void osp_precreate_fini(struct osp_device *d)
1704 struct ptlrpc_thread *thread;
1708 del_timer(&d->opd_statfs_timer);
1710 if (d->opd_pre == NULL)
1713 thread = &d->opd_pre_thread;
1715 thread->t_flags = SVC_STOPPING;
1716 wake_up(&d->opd_pre_waitq);
1718 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1720 OBD_FREE_PTR(d->opd_pre);