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, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/osp/osp_precreate.c
33 * Lustre OST Proxy Device
35 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
36 * Author: Mikhail Pershin <mike.pershin@intel.com>
37 * Author: Di Wang <di.wang@intel.com>
40 #define DEBUG_SUBSYSTEM S_MDS
42 #include <linux/kthread.h>
44 #include <lustre_obdo.h>
46 #include "osp_internal.h"
49 * there are two specific states to take care about:
51 * = import is disconnected =
53 * = import is inactive =
54 * in this case osp_declare_create() returns an error
59 * Check whether statfs data is expired
61 * OSP device caches statfs data for the target, the function checks
62 * whether the data is expired or not.
64 * \param[in] d OSP device
66 * \retval 0 - not expired, 1 - expired
68 static inline int osp_statfs_need_update(struct osp_device *d)
70 return !ktime_before(ktime_get(), d->opd_statfs_fresh_till);
74 * OSP tries to maintain pool of available objects so that calls to create
75 * objects don't block most of time
77 * each time OSP gets connected to OST, we should start from precreation cleanup
79 static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data)
81 struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
84 /* invalidate statfs data so osp_precreate_thread() can refresh */
85 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
87 wake_up(&d->opd_pre_waitq);
90 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs);
93 * The function updates current precreation status if broken, and
94 * updates that cached statfs state if functional, then wakes up waiters.
95 * We don't clear opd_pre_status directly here, but rather leave this
96 * to osp_pre_update_msfs() to do if everything is OK so that we don't
97 * have a race to clear opd_pre_status and then set it to -ENOSPC again.
99 * \param[in] d OSP device
100 * \param[in] msfs statfs data
101 * \param[in] rc new precreate status for device \a d
103 static void osp_pre_update_status_msfs(struct osp_device *d,
104 struct obd_statfs *msfs, int rc)
106 CDEBUG(D_INFO, "%s: Updating status = %d\n", d->opd_obd->obd_name, rc);
108 d->opd_pre_status = rc;
110 osp_pre_update_msfs(d, msfs);
112 wake_up_all(&d->opd_pre_user_waitq);
115 /* Pass in the old statfs data in case the limits have changed */
116 void osp_pre_update_status(struct osp_device *d, int rc)
118 osp_pre_update_status_msfs(d, &d->opd_statfs, rc);
123 * RPC interpret callback for OST_STATFS RPC
125 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
126 * replied by the target. It's used to maintain statfs cache for the target.
127 * The function fills data from the reply if successful and schedules another
130 * \param[in] env LU environment provided by the caller
131 * \param[in] req RPC replied
132 * \param[in] aa callback data
133 * \param[in] rc RPC result
135 * \retval 0 on success
136 * \retval negative negated errno on error
138 static int osp_statfs_interpret(const struct lu_env *env,
139 struct ptlrpc_request *req, void *args, int rc)
141 union ptlrpc_async_args *aa = args;
142 struct obd_import *imp = req->rq_import;
143 struct obd_statfs *msfs;
144 struct obd_statfs *sfs;
145 struct osp_device *d;
150 aa = ptlrpc_req_async_args(aa, req);
151 d = aa->pointer_arg[0];
157 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
159 GOTO(out, rc = -EPROTO);
162 osp_pre_update_status_msfs(d, msfs, 0);
164 osp_pre_update_msfs(d, msfs);
166 /* schedule next update */
167 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
168 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
169 mod_timer(&d->opd_statfs_timer,
170 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
171 d->opd_statfs_update_in_progress = 0;
173 sfs = &d->opd_statfs;
175 "%s (%p): %llu blocks, %llu free, %llu avail, %u bsize, %u reserved mb low, %u reserved mb high, %u reserved ino low, %u reserved ino high, %llu files, %llu free files %#x\n",
176 d->opd_obd->obd_name, d, sfs->os_blocks, sfs->os_bfree,
177 sfs->os_bavail, sfs->os_bsize, d->opd_reserved_mb_low,
178 d->opd_reserved_mb_high, d->opd_reserved_ino_low,
179 d->opd_reserved_ino_high, sfs->os_files, sfs->os_ffree,
184 /* couldn't update statfs, try again with a small delay */
185 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
186 d->opd_statfs_update_in_progress = 0;
187 if (d->opd_pre && d->opd_pre_task)
188 wake_up(&d->opd_pre_waitq);
190 if (req->rq_import_generation == imp->imp_generation)
191 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
192 d->opd_obd->obd_name, rc);
197 * Send OST_STATFS RPC
199 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
200 * Also disables scheduled updates as times OSP may need to refresh
201 * statfs data before expiration. The function doesn't block, instead
202 * an interpretation callback osp_statfs_interpret() is used.
204 * \param[in] env LU environment provided by the caller
205 * \param[in] d OSP device
207 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
209 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
210 struct ptlrpc_request *req;
211 struct obd_import *imp;
212 union ptlrpc_async_args *aa;
217 CDEBUG(D_CACHE, "going to update statfs\n");
219 imp = d->opd_obd->u.cli.cl_import;
222 req = ptlrpc_request_alloc(imp,
223 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
227 rc = ptlrpc_request_pack(req,
228 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
229 d->opd_pre ? OST_STATFS : MDS_STATFS);
231 ptlrpc_request_free(req);
234 ptlrpc_request_set_replen(req);
236 req->rq_request_portal = OST_CREATE_PORTAL;
237 ptlrpc_at_set_req_timeout(req);
239 req->rq_interpret_reply = osp_statfs_interpret;
240 aa = ptlrpc_req_async_args(aa, req);
241 aa->pointer_arg[0] = d;
244 * no updates till reply
246 timer_delete(&d->opd_statfs_timer);
247 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
248 d->opd_statfs_update_in_progress = 1;
250 ptlrpcd_add_req(req);
252 /* we still want to sync changes if no new changes are coming */
253 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
256 if (atomic_read(&d->opd_sync_changes)) {
259 th = dt_trans_create(env, d->opd_storage);
261 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
264 rc = dt_trans_start_local(env, d->opd_storage, th);
266 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
267 d->opd_obd->obd_name,
268 atomic_read(&d->opd_sync_changes));
269 osp_sync_add_commit_cb_1s(env, d, th);
271 dt_trans_stop(env, d->opd_storage, th);
279 * Schedule an immediate update for statfs data
281 * If cached statfs data claim no free space, but OSP has got a request to
282 * destroy an object (so release some space probably), then we may need to
283 * refresh cached statfs data sooner than planned. The function checks there
284 * is no statfs update going and schedules immediate update if so.
285 * XXX: there might be a case where removed object(s) do not add free space (empty
286 * object). If the number of such deletions is high, then we can start to update
287 * statfs too often causing a RPC storm. some throttling is needed...
289 * \param[in] d OSP device where statfs data needs to be refreshed
291 void osp_statfs_need_now(struct osp_device *d)
293 if (!d->opd_statfs_update_in_progress) {
295 * if current status is -ENOSPC (lack of free space on OST)
296 * then we should poll OST immediately once object destroy
299 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
300 timer_delete(&d->opd_statfs_timer);
301 wake_up(&d->opd_pre_waitq);
306 * Check pool of precreated objects is nearly empty
308 * We should not wait till the pool of the precreated objects is exhausted,
309 * because then there will be a long period of OSP being unavailable for the
310 * new creations due to lenghty precreate RPC. Instead we ask for another
311 * precreation ahead and hopefully have it ready before the current pool is
312 * empty. Notice this function relies on an external locking.
314 * \param[in] d OSP device
316 * \retval 0 - current pool is good enough, 1 - time to precreate
318 static inline int osp_precreate_near_empty_nolock(struct osp_device *d)
320 int window = osp_objs_precreated(d);
322 /* don't consider new precreation till OST is healty and
324 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2 ||
325 d->opd_force_creation) && (d->opd_pre_status == 0));
329 * Check pool of precreated objects
331 * This is protected version of osp_precreate_near_empty_nolock(), check that
334 * \param[in] d OSP device
336 * \retval 0 - current pool is good enough, 1 - time to precreate
338 static inline int osp_precreate_near_empty(struct osp_device *d)
342 if (d->opd_pre == NULL)
345 /* XXX: do we really need locking here? */
346 spin_lock(&d->opd_pre_lock);
347 rc = osp_precreate_near_empty_nolock(d);
348 spin_unlock(&d->opd_pre_lock);
353 * Write FID into into last_oid/last_seq file
355 * The function stores the sequence and the in-sequence id into two dedicated
356 * files. The sync argument can be used to request synchronous commit, so the
357 * function won't return until the updates are committed.
359 * \param[in] env LU environment provided by the caller
360 * \param[in] osp OSP device
361 * \param[in] fid fid where sequence/id is taken
362 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
364 * \retval 0 on success
365 * \retval negative negated errno on error
367 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
368 struct lu_fid *fid, int sync)
370 struct osp_thread_info *oti = osp_env_info(env);
371 struct lu_buf *lb_oid = &oti->osi_lb;
372 struct lu_buf *lb_oseq = &oti->osi_lb2;
380 if (osp->opd_storage->dd_rdonly)
383 /* Note: through f_oid is only 32 bits, it will also write 64 bits
384 * for oid to keep compatibility with the previous version. */
386 osp_objid_buf_prep(lb_oid, &oid_off,
387 &oid, osp->opd_index);
389 osp_objseq_buf_prep(lb_oseq, &oseq_off,
390 &fid->f_seq, osp->opd_index);
392 th = dt_trans_create(env, osp->opd_storage);
397 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
398 lb_oid, oid_off, th);
402 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
403 lb_oseq, oseq_off, th);
407 rc = dt_trans_start_local(env, osp->opd_storage, th);
411 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
414 CERROR("%s: can not write to last seq file: rc = %d\n",
415 osp->opd_obd->obd_name, rc);
418 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
421 CERROR("%s: can not write to last seq file: rc = %d\n",
422 osp->opd_obd->obd_name, rc);
426 dt_trans_stop(env, osp->opd_storage, th);
430 static void osp_update_fldb_cache(const struct lu_env *env,
431 struct osp_device *osp,
434 struct lu_seq_range range = { 0 };
435 struct lu_server_fld *server_fld;
436 struct lu_site *site;
438 site = osp->opd_storage->dd_lu_dev.ld_site;
439 server_fld = lu_site2seq(site)->ss_server_fld;
443 fld_range_set_type(&range, LU_SEQ_RANGE_ANY);
444 fld_server_lookup(env, server_fld, fid_seq(fid), &range);
448 * Switch to another sequence
450 * When a current sequence has no available IDs left, OSP has to switch to
451 * another new sequence. OSP requests it using the regular FLDB protocol
452 * and stores synchronously before that is used in precreate. This is needed
453 * to basically have the sequences referenced (not orphaned), otherwise it's
454 * possible that OST has some objects precreated and the clients have data
455 * written to it, but after MDT failover nobody refers those objects and OSP
456 * has no idea that the sequence need cleanup to be done.
457 * While this is very expensive operation, it's supposed to happen infrequently
458 * because sequence has LUSTRE_DATA_SEQ_MAX_WIDTH=32M objects by default.
460 * \param[in] env LU environment provided by the caller
461 * \param[in] osp OSP device
463 * \retval 0 on success
464 * \retval negative negated errno on error
466 static int osp_precreate_rollover_new_seq(struct lu_env *env,
467 struct osp_device *osp)
469 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
470 struct lu_fid *last_fid = &osp->opd_last_used_fid;
474 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
476 CERROR("%s: alloc fid error: rc = %d\n",
477 osp->opd_obd->obd_name, rc);
483 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
484 "fid "DFID", last_fid "DFID"\n", PFID(fid),
487 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
489 CERROR("%s: Can not update oid/seq file: rc = %d\n",
490 osp->opd_obd->obd_name, rc);
494 LCONSOLE(D_INFO, "%s: update sequence from %#llx to %#llx\n",
495 osp->opd_obd->obd_name, fid_seq(last_fid),
497 /* Update last_xxx to the new seq */
498 spin_lock(&osp->opd_pre_lock);
499 osp->opd_last_used_fid = *fid;
500 osp_fid_to_obdid(fid, &osp->opd_last_id);
501 osp->opd_gap_start_fid = *fid;
502 osp->opd_pre_used_fid = *fid;
503 osp->opd_pre_last_created_fid = *fid;
504 spin_unlock(&osp->opd_pre_lock);
507 osp_update_fldb_cache(env, osp, fid);
513 * Find IDs available in current sequence
515 * The function calculates the highest possible ID and the number of IDs
516 * available in the current sequence OSP is using. The number is limited
517 * artifically by the caller (grow param) and the number of IDs available
518 * in the sequence by nature. The function doesn't require an external
521 * \param[in] env LU environment provided by the caller
522 * \param[in] osp OSP device
523 * \param[in] fid FID the caller wants to start with
524 * \param[in] grow how many the caller wants
525 * \param[out] fid the highest calculated FID
526 * \param[out] grow the number of available IDs calculated
528 * \retval 0 on success, 1 - the sequence is empty
530 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
531 struct lu_fid *fid, int *grow)
533 struct osp_thread_info *osi = osp_env_info(env);
534 __u64 seq_width = osp->opd_pre_seq_width;
538 if (fid_is_idif(fid)) {
539 struct lu_fid *last_fid;
540 struct ost_id *oi = &osi->osi_oi;
543 spin_lock(&osp->opd_pre_lock);
544 last_fid = &osp->opd_pre_last_created_fid;
545 fid_to_ostid(last_fid, oi);
546 end = min(ostid_id(oi) + *grow, min(IDIF_MAX_OID, seq_width));
547 *grow = end - ostid_id(oi);
548 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
549 spin_unlock(&osp->opd_pre_lock);
551 if (*grow == 0 || rc)
554 ostid_to_fid(fid, oi, osp->opd_index);
558 spin_lock(&osp->opd_pre_lock);
559 *fid = osp->opd_pre_last_created_fid;
561 end = min((end + *grow), min(OBIF_MAX_OID, seq_width));
562 *grow = end - fid->f_oid;
563 fid->f_oid += end - fid->f_oid;
564 spin_unlock(&osp->opd_pre_lock);
566 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
567 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
569 return *grow > 0 ? 0 : 1;
573 * Prepare and send precreate RPC
575 * The function finds how many objects should be precreated. Then allocates,
576 * prepares and schedules precreate RPC synchronously. Upon reply the function
577 * wakes up the threads waiting for the new objects on this target. If the
578 * target wasn't able to create all the objects requested, then the next
579 * precreate will be asking for fewer objects (i.e. slow precreate down).
581 * \param[in] env LU environment provided by the caller
582 * \param[in] d OSP device
584 * \retval 0 on success
585 * \retval negative negated errno on error
587 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
589 struct osp_thread_info *oti = osp_env_info(env);
590 struct ptlrpc_request *req;
591 struct obd_import *imp;
592 struct ost_body *body;
594 struct lu_fid *fid = &oti->osi_fid;
597 /* don't precreate new objects till OST healthy and has free space */
598 if (unlikely(d->opd_pre_status)) {
599 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
600 d->opd_obd->obd_name, d->opd_pre_status);
605 * if not connection/initialization is compeleted, ignore
607 imp = d->opd_obd->u.cli.cl_import;
610 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
613 req->rq_request_portal = OST_CREATE_PORTAL;
615 /* We should not resend create request - anyway we will have delorphan
616 * and kill these objects.
617 * Only needed for MDS+OSS rolling upgrade interop with 2.16+older.
619 if (unlikely(!imp_connect_replay_create(imp)))
620 req->rq_no_delay = req->rq_no_resend = 1;
622 /* Delorphan happens only with a first MDT-OST connect. resend/replay
623 * handles objects creation on reconnects, no need to do delorhpan
627 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
629 ptlrpc_request_free(req);
633 spin_lock(&d->opd_pre_lock);
634 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
635 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
636 grow = d->opd_pre_create_count;
637 spin_unlock(&d->opd_pre_lock);
639 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
642 *fid = d->opd_pre_last_created_fid;
643 rc = osp_precreate_fids(env, d, fid, &grow);
645 /* Current seq has been used up*/
646 GOTO(out_req, rc = -ENOSPC);
648 if (!osp_is_fid_client(d)) {
649 /* Non-FID client will always send seq 0 because of
651 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
655 fid_to_ostid(fid, &body->oa.o_oi);
656 body->oa.o_valid = OBD_MD_FLGROUP;
658 ptlrpc_request_set_replen(req);
660 if (CFS_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
663 rc = ptlrpc_queue_wait(req);
665 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
668 /* have osp_precreate_reserve() to wait for repeat */
673 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
675 GOTO(out_req, rc = -EPROTO);
677 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
680 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
681 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
682 ": rc = %d\n", d->opd_obd->obd_name,
683 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
684 GOTO(out_req, rc = -ESTALE);
687 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
689 spin_lock(&d->opd_pre_lock);
691 /* the OST has not managed to create all the
692 * objects we asked for */
693 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
694 d->opd_pre_create_slow = 1;
696 /* the OST is able to keep up with the work,
697 * we could consider increasing create_count
698 * next time if needed */
699 d->opd_pre_create_slow = 0;
702 if ((body->oa.o_valid & OBD_MD_FLSIZE) && body->oa.o_size)
703 d->opd_pre_seq_width = body->oa.o_size;
705 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
706 fid_to_ostid(fid, &body->oa.o_oi);
708 d->opd_pre_last_created_fid = *fid;
709 d->opd_force_creation = false;
710 spin_unlock(&d->opd_pre_lock);
712 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
713 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
714 PFID(&d->opd_pre_last_created_fid));
716 /* now we can wakeup all users awaiting for objects */
717 osp_pre_update_status(d, rc);
719 ptlrpc_req_finished(req);
722 osp_update_fldb_cache(env, d, fid);
728 * Get last precreated object from target (OST)
730 * Sends synchronous RPC to the target (OST) to learn the last precreated
731 * object. This later is used to remove all unused objects (cleanup orphan
732 * procedure). Also, the next object after one we got will be used as a
733 * starting point for the new precreates.
735 * \param[in] env LU environment provided by the caller
736 * \param[in] d OSP device
737 * \param[in] update update or not update last used fid
739 * \retval 0 on success
740 * \retval negative negated errno on error
742 static int osp_get_lastfid_from_ost(const struct lu_env *env,
743 struct osp_device *d, bool update)
745 struct ptlrpc_request *req = NULL;
746 struct obd_import *imp;
747 struct lu_fid *last_fid;
752 imp = d->opd_obd->u.cli.cl_import;
755 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
759 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
760 sizeof(KEY_LAST_FID));
762 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
764 ptlrpc_request_free(req);
768 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
769 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
771 req->rq_no_delay = req->rq_no_resend = 1;
772 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
773 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
775 ptlrpc_request_set_replen(req);
777 rc = ptlrpc_queue_wait(req);
779 /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld),
780 * let sysadm sort this * out.
783 ptlrpc_set_import_active(imp, 0);
787 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
788 if (last_fid == NULL) {
789 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
790 GOTO(out, rc = -EPROTO);
793 if (!fid_is_sane(last_fid)) {
794 CERROR("%s: Got insane last_fid "DFID"\n",
795 d->opd_obd->obd_name, PFID(last_fid));
796 GOTO(out, rc = -EPROTO);
799 /* Only update the last used fid, if the OST has objects for
800 * this sequence, i.e. fid_oid > 0 */
801 if (fid_oid(last_fid) > 0 && update)
802 d->opd_last_used_fid = *last_fid;
804 if (fid_seq(last_fid) == fid_seq(&d->opd_last_used_fid)) {
805 if (fid_oid(last_fid) == 0 ||
806 (fid_seq_is_norm(fid_seq(last_fid)) &&
807 fid_oid(last_fid) == LUSTRE_FID_INIT_OID)) {
808 /* reformatted OST, it requires creation request
809 * to recreate objects
811 spin_lock(&d->opd_pre_lock);
812 d->opd_force_creation = true;
813 d->opd_pre_create_count = OST_MIN_PRECREATE;
814 spin_unlock(&d->opd_pre_lock);
817 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
821 ptlrpc_req_finished(req);
826 * Cleanup orphans on OST
828 * This function is called in a contex of a dedicated thread handling
829 * all the precreation suff. The function waits till local recovery
830 * is complete, then identify all the unreferenced objects (orphans)
831 * using the highest ID referenced by a local and the highest object
832 * precreated by the target. The found range is a subject to removal
833 * using specially flagged RPC. During this process OSP is marked
834 * unavailable for new objects.
836 * \param[in] env LU environment provided by the caller
837 * \param[in] d OSP device
839 * \retval 0 on success
840 * \retval negative negated errno on error
842 static int osp_precreate_cleanup_orphans(struct lu_env *env,
843 struct osp_device *d)
845 struct osp_thread_info *osi = osp_env_info(env);
846 struct lu_fid *last_fid = &osi->osi_fid;
847 struct ptlrpc_request *req = NULL;
848 struct obd_import *imp = d->opd_obd->u.cli.cl_import;
849 struct ost_body *body;
850 int update_status = 0;
857 * Do cleanup orphans only with a first connection, after that
858 * all precreate requests uses resend/replay flags to support OST
859 * failover/reconnect.
861 if (d->opd_cleanup_orphans_done && imp_connect_replay_create(imp)) {
862 rc = osp_get_lastfid_from_ost(env, d, false);
866 * wait for local recovery to finish, so we can cleanup orphans
867 * orphans are all objects since "last used" (assigned), but
868 * there might be objects reserved and in some cases they won't
869 * be used. we can't cleanup them till we're sure they won't be
870 * used. also can't we allow new reservations because they may
871 * end up getting orphans being cleaned up below. so we block
872 * new reservations and wait till all reserved objects either
875 spin_lock(&d->opd_pre_lock);
876 d->opd_pre_recovering = 1;
877 spin_unlock(&d->opd_pre_lock);
879 * The locking above makes sure the opd_pre_reserved check below will
880 * catch all osp_precreate_reserve() calls who find
881 * "!opd_pre_recovering".
883 wait_event_idle(d->opd_pre_waitq,
884 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
885 !d->opd_pre_task || d->opd_got_disconnected);
886 if (!d->opd_pre_task || d->opd_got_disconnected)
887 GOTO(out, rc = -EAGAIN);
889 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
890 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
892 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val);
894 *last_fid = d->opd_last_used_fid;
895 /* The OSP should already get the valid seq now */
896 LASSERT(!fid_is_zero(last_fid));
897 if (fid_oid(&d->opd_last_used_fid) < 2) {
898 /* lastfid looks strange... ask OST */
899 rc = osp_get_lastfid_from_ost(env, d, true);
904 imp = d->opd_obd->u.cli.cl_import;
907 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
909 GOTO(out, rc = -ENOMEM);
911 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
913 ptlrpc_request_free(req);
918 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
920 GOTO(out, rc = -EPROTO);
922 body->oa.o_flags = OBD_FL_DELORPHAN;
923 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
925 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
927 ptlrpc_request_set_replen(req);
929 /* Don't resend the delorphan req */
930 req->rq_no_resend = req->rq_no_delay = 1;
932 rc = ptlrpc_queue_wait(req);
938 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
940 GOTO(out, rc = -EPROTO);
943 * OST provides us with id new pool starts from in body->oa.o_id
945 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
947 spin_lock(&d->opd_pre_lock);
948 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
950 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
951 d->opd_pre_last_created_fid = d->opd_last_used_fid;
953 d->opd_pre_create_count = OST_MIN_PRECREATE;
954 d->opd_pre_last_created_fid = *last_fid;
957 * This empties the pre-creation pool and effectively blocks any new
960 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
961 d->opd_pre_create_slow = 0;
962 if ((body->oa.o_valid & OBD_MD_FLSIZE) && body->oa.o_size)
963 d->opd_pre_seq_width = body->oa.o_size;
964 spin_unlock(&d->opd_pre_lock);
966 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
967 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
968 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
971 ptlrpc_req_finished(req);
975 * If rc is zero, the pre-creation window should have been emptied.
976 * Since waking up the herd would be useless without pre-created
977 * objects, we defer the signal to osp_precreate_send() in that case.
981 CERROR("%s: cannot cleanup orphans: rc = %d\n",
982 d->opd_obd->obd_name, rc);
983 /* we can't proceed from here, OST seem to
984 * be in a bad shape, better to wait for
985 * a new instance of the server and repeat
986 * from the beginning. notify possible waiters
987 * this OSP isn't quite functional yet */
988 osp_pre_update_status(d, rc);
990 wake_up_all(&d->opd_pre_user_waitq);
993 spin_lock(&d->opd_pre_lock);
994 d->opd_pre_recovering = 0;
995 spin_unlock(&d->opd_pre_lock);
996 d->opd_cleanup_orphans_done = true;
1003 * Update precreate status using statfs data
1005 * The function decides whether this OSP should be used for new objects.
1006 * IOW, whether this OST is used up or has some free space. Cached statfs
1007 * data is used to make this decision. If the latest result of statfs
1008 * request (rc argument) is not success, then just mark OSP unavailable
1011 * The new statfs data is passed in \a msfs and needs to be stored into
1012 * opd_statfs, but only after the various flags in os_state are set, so
1013 * that the new statfs data is not visible without appropriate flags set.
1014 * As such, there is no need to clear the flags here, since this is called
1015 * with new statfs data, and they should not be cleared if sent from OST.
1017 * Add a bit of hysteresis so this flag isn't continually flapping, and
1018 * ensure that new files don't get extremely fragmented due to only a
1019 * small amount of available space in the filesystem. We want to set
1020 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1021 * reserved size free, and still copy them from the old state when there
1022 * is less than 2*reserved size free space or inodes.
1024 * \param[in] d OSP device
1025 * \param[in] msfs statfs data
1027 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1029 u32 old_state = d->opd_statfs.os_state;
1032 /* statfs structure not initialized yet */
1033 if (unlikely(!msfs->os_type))
1036 /* if the low and high watermarks have not been initialized yet */
1037 if (unlikely(d->opd_reserved_mb_high == 0 &&
1038 d->opd_reserved_mb_low == 0)) {
1039 /* Use ~0.1% by default to disable object allocation,
1040 * and ~0.2% to enable, size in MB, set both watermark
1042 spin_lock(&d->opd_pre_lock);
1043 if (d->opd_reserved_mb_high == 0 &&
1044 d->opd_reserved_mb_low == 0) {
1045 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1046 msfs->os_blocks) >> 20;
1047 if (d->opd_reserved_mb_low < 1)
1048 d->opd_reserved_mb_low = 1;
1049 d->opd_reserved_mb_high =
1050 (d->opd_reserved_mb_low << 1) + 1;
1052 spin_unlock(&d->opd_pre_lock);
1055 if (unlikely(d->opd_reserved_ino_high == 0 &&
1056 d->opd_reserved_ino_low == 0)) {
1057 /* Use ~0.0001% by default to disallow distributed transactions,
1058 * and ~0.0002% to allow, set both watermark
1060 spin_lock(&d->opd_pre_lock);
1061 if (d->opd_reserved_ino_high == 0 &&
1062 d->opd_reserved_ino_low == 0) {
1063 d->opd_reserved_ino_low = msfs->os_ffree >> 20;
1064 if (d->opd_reserved_ino_low < 32)
1065 d->opd_reserved_ino_low = 32;
1066 d->opd_reserved_ino_high =
1067 (d->opd_reserved_ino_low << 1) + 1;
1069 spin_unlock(&d->opd_pre_lock);
1072 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1073 if (msfs->os_ffree < d->opd_reserved_ino_low)
1074 msfs->os_state |= OS_STATFS_ENOINO;
1075 else if (msfs->os_ffree <= d->opd_reserved_ino_high)
1076 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1077 /* else don't clear flags in new msfs->os_state sent from OST */
1079 if (available_mb < d->opd_reserved_mb_low)
1080 msfs->os_state |= OS_STATFS_ENOSPC;
1081 else if (available_mb <= d->opd_reserved_mb_high)
1082 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1083 /* else don't clear flags in new msfs->os_state sent from OST */
1086 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1087 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1088 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1089 msfs->os_files, msfs->os_ffree, msfs->os_state,
1090 d->opd_pre ? d->opd_pre_status : 0);
1095 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1096 d->opd_pre_status = -ENOSPC;
1097 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1098 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1099 d->opd_obd->obd_name, msfs->os_state,
1101 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1102 atomic_read(&d->opd_sync_changes),
1103 atomic_read(&d->opd_sync_rpcs_in_progress));
1104 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1105 d->opd_pre_status = 0;
1106 spin_lock(&d->opd_pre_lock);
1107 d->opd_pre_create_slow = 0;
1108 d->opd_pre_create_count = OST_MIN_PRECREATE;
1109 spin_unlock(&d->opd_pre_lock);
1110 wake_up(&d->opd_pre_waitq);
1113 "%s: available: state=%x: rc = %d\n",
1114 d->opd_obd->obd_name, msfs->os_state,
1117 /* we only get here if rc == 0 in the caller */
1118 d->opd_pre_status = 0;
1121 /* Object precreation skipped on OST if manually disabled */
1122 if (d->opd_pre_max_create_count == 0)
1123 msfs->os_state |= OS_STATFS_NOCREATE;
1124 /* else don't clear flags in new msfs->os_state sent from OST */
1127 /* copy only new statfs state to make it visible to MDS threads */
1128 if (&d->opd_statfs != msfs)
1129 d->opd_statfs = *msfs;
1133 * Initialize FID for precreation
1135 * For a just created new target, a new sequence should be taken.
1136 * The function checks there is no IDIF in use (if the target was
1137 * added with the older version of Lustre), then requests a new
1138 * sequence from FLDB using the regular protocol. Then this new
1139 * sequence is stored on a persisten storage synchronously to prevent
1140 * possible object leakage (for the detail see the description for
1141 * osp_precreate_rollover_new_seq()).
1143 * \param[in] osp OSP device
1145 * \retval 0 on success
1146 * \retval negative negated errno on error
1148 static int osp_init_pre_fid(struct lu_env *env, struct osp_device *osp)
1150 struct osp_thread_info *osi;
1151 struct lu_client_seq *cli_seq;
1152 struct lu_fid *last_fid;
1156 LASSERT(osp->opd_pre != NULL);
1158 /* Let's check if the current last_seq/fid is valid,
1159 * otherwise request new sequence from the controller */
1160 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1161 /* Non-MDT0 can only use normal sequence for
1163 if (fid_is_norm(&osp->opd_last_used_fid))
1166 /* Initially MDT0 will start with IDIF, after
1167 * that it will request new sequence from the
1169 if (fid_is_idif(&osp->opd_last_used_fid) ||
1170 fid_is_norm(&osp->opd_last_used_fid))
1174 if (!fid_is_zero(&osp->opd_last_used_fid))
1175 CWARN("%s: invalid last used fid "DFID
1176 ", try to get new sequence.\n",
1177 osp->opd_obd->obd_name,
1178 PFID(&osp->opd_last_used_fid));
1180 osi = osp_env_info(env);
1181 last_fid = &osi->osi_fid;
1183 /* For a freshed fs, it will allocate a new sequence first */
1184 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1185 cli_seq = osp->opd_obd->u.cli.cl_seq;
1186 rc = seq_client_get_seq(env, cli_seq, &last_fid->f_seq);
1188 CERROR("%s: alloc fid error: rc = %d\n",
1189 osp->opd_obd->obd_name, rc);
1193 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1195 last_fid->f_oid = 1;
1196 last_fid->f_ver = 0;
1198 spin_lock(&osp->opd_pre_lock);
1199 osp->opd_last_used_fid = *last_fid;
1200 osp->opd_pre_used_fid = *last_fid;
1201 osp->opd_pre_last_created_fid = *last_fid;
1202 spin_unlock(&osp->opd_pre_lock);
1203 rc = osp_write_last_oid_seq_files(env, osp, last_fid, 1);
1205 CERROR("%s: write fid error: rc = %d\n",
1206 osp->opd_obd->obd_name, rc);
1214 struct osp_device *opta_dev;
1215 struct lu_env opta_env;
1216 struct completion *opta_started;
1219 * The core of precreate functionality
1221 * The function implements the main precreation loop. Basically it
1222 * involves connecting to the target, precerate FID initialization,
1223 * identifying and removing orphans, then serving precreation. As
1224 * part of the latter, the thread is responsible for statfs data
1225 * updates. The precreation is mostly driven by another threads
1226 * asking for new OST objects - those askers wake the thread when
1227 * the number of precreated objects reach low watermark.
1228 * After a disconnect, the sequence above repeats. This is keep going
1229 * until the thread is requested to stop.
1231 * \param[in] _arg private data the thread (OSP device to handle)
1233 * \retval 0 on success
1234 * \retval negative negated errno on error
1236 static int osp_precreate_thread(void *_args)
1238 struct opt_args *args = _args;
1239 struct osp_device *d = args->opta_dev;
1240 struct lu_env *env = &args->opta_env;
1245 complete(args->opta_started);
1247 /* wait for connection from the layers above */
1248 wait_event_idle(d->opd_pre_waitq,
1249 kthread_should_stop() ||
1250 d->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1252 while (!kthread_should_stop()) {
1254 * need to be connected to OST
1256 while (!kthread_should_stop()) {
1257 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1258 d->opd_imp_connected &&
1259 !d->opd_got_disconnected)
1261 wait_event_idle(d->opd_pre_waitq,
1262 kthread_should_stop() ||
1263 d->opd_new_connection);
1265 if (!d->opd_new_connection)
1268 CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
1270 d->opd_new_connection = 0;
1271 d->opd_got_disconnected = 0;
1275 if (kthread_should_stop())
1279 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1280 LASSERT(d->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1282 /* Init fid for osp_precreate if necessary */
1283 rc = osp_init_pre_fid(env, d);
1285 class_export_put(d->opd_exp);
1286 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1287 CERROR("%s: init pre fid error: rc = %d\n",
1288 d->opd_obd->obd_name, rc);
1293 if (osp_statfs_update(env, d)) {
1294 if (wait_event_idle_timeout(d->opd_pre_waitq,
1295 kthread_should_stop(),
1296 cfs_time_seconds(5)) == 0)
1297 l_wait_event_abortable(
1299 kthread_should_stop());
1305 * Clean up orphans or recreate missing objects.
1307 rc = osp_precreate_cleanup_orphans(env, d);
1309 schedule_timeout_interruptible(cfs_time_seconds(1));
1315 * connected, can handle precreates now
1317 while (!kthread_should_stop()) {
1318 wait_event_idle(d->opd_pre_waitq,
1319 kthread_should_stop() ||
1320 (osp_precreate_near_empty(d) &&
1321 !(osp_precreate_end_seq(d) &&
1322 osp_objs_precreated(d) != 0)) ||
1323 osp_statfs_need_update(d) ||
1324 d->opd_got_disconnected);
1326 if (kthread_should_stop())
1329 /* something happened to the connection
1330 * have to start from the beginning */
1331 if (d->opd_got_disconnected)
1334 if (osp_statfs_need_update(d))
1335 if (osp_statfs_update(env, d))
1338 if (d->opd_pre == NULL)
1341 /* To avoid handling different seq in precreate/orphan
1342 * cleanup, it will hold precreate until current seq is
1344 if (unlikely(osp_precreate_end_seq(d))) {
1345 if (osp_objs_precreated(d) == 0) {
1346 rc = osp_precreate_rollover_new_seq(env, d);
1354 if (osp_precreate_near_empty(d)) {
1355 rc = osp_precreate_send(env, d);
1356 /* osp_precreate_send() sets opd_pre_status
1357 * in case of error, that prevent the using of
1359 if (rc < 0 && rc != -ENOSPC &&
1360 rc != -ETIMEDOUT && rc != -ENOTCONN)
1361 CERROR("%s: cannot precreate objects:"
1363 d->opd_obd->obd_name, rc);
1375 * Check when to stop to wait for precreate objects.
1377 * The caller wanting a new OST object can't wait undefinitely. The
1378 * function checks for few conditions including available new OST
1379 * objects, disconnected OST, lack of space with no pending destroys,
1380 * etc. IOW, it checks whether the current OSP state is good to keep
1381 * waiting or it's better to give up.
1383 * \param[in] env LU environment provided by the caller
1384 * \param[in] d OSP device
1386 * \retval 0 - keep waiting, 1 - no luck
1388 static int osp_precreate_ready_condition(const struct lu_env *env,
1389 struct osp_device *d)
1391 /* Bail out I/O fails to OST */
1392 if (d->opd_pre_status != 0 &&
1393 d->opd_pre_status != -EAGAIN &&
1394 d->opd_pre_status != -ENODEV &&
1395 d->opd_pre_status != -ENOTCONN &&
1396 d->opd_pre_status != -ENOSPC) {
1398 if (d->opd_pre_status != -EIO)
1399 CERROR("%s: precreate failed opd_pre_status %d\n",
1400 d->opd_obd->obd_name, d->opd_pre_status);
1404 if (d->opd_pre_recovering || d->opd_force_creation)
1407 /* ready if got enough precreated objects */
1408 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1409 if (d->opd_pre_reserved + 1 < osp_objs_precreated(d))
1412 /* ready if OST reported no space and no destroys in progress */
1413 if (atomic_read(&d->opd_sync_changes) +
1414 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1415 d->opd_pre_status == -ENOSPC)
1422 * Reserve object in precreate pool
1424 * When the caller wants to create a new object on this target (target
1425 * represented by the given OSP), it should declare this intention using
1426 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1427 * to reserve an object in the existing precreated pool or wait up to
1428 * obd_timeout for the available object to appear in the pool (a dedicated
1429 * thread will be doing real precreation in background). The object can be
1430 * consumed later with osp_precreate_get_fid() or be released with call to
1431 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1432 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1433 * If the space on the target is short and there is a pending object destroy,
1434 * then the function forces local commit to speedup space release (see
1435 * osp_sync.c for the details).
1437 * \param[in] env LU environment provided by the caller
1438 * \param[in] d OSP device
1440 * \retval 0 on success
1441 * \retval -ENOSPC when no space on OST
1442 * \retval -EAGAIN try later, slow precreation in progress
1443 * \retval -EIO when no access to OST
1445 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
1448 time64_t expire = ktime_get_seconds() + obd_timeout;
1449 int precreated, rc, synced = 0;
1453 LASSERTF(osp_objs_precreated(d) >= 0, "Last created FID "DFID
1454 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1455 PFID(&d->opd_pre_used_fid));
1457 /* opd_pre_max_create_count 0 to not use specified OST. */
1458 if (d->opd_pre_max_create_count == 0)
1463 * - preallocation is done
1464 * - no free space expected soon
1465 * - can't connect to OST for too long (obd_timeout)
1466 * - OST can allocate fid sequence.
1468 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1469 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1472 * increase number of precreations
1474 precreated = osp_objs_precreated(d);
1475 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1476 d->opd_pre_create_slow == 0 &&
1477 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1478 spin_lock(&d->opd_pre_lock);
1479 d->opd_pre_create_slow = 1;
1480 d->opd_pre_create_count *= 2;
1481 spin_unlock(&d->opd_pre_lock);
1484 spin_lock(&d->opd_pre_lock);
1485 precreated = osp_objs_precreated(d);
1486 if (!d->opd_pre_recovering && !d->opd_force_creation) {
1487 if (precreated > d->opd_pre_reserved) {
1488 d->opd_pre_reserved++;
1489 spin_unlock(&d->opd_pre_lock);
1493 * XXX: don't wake up if precreation
1496 if (osp_precreate_near_empty_nolock(d) &&
1497 !osp_precreate_end_seq_nolock(d))
1498 wake_up(&d->opd_pre_waitq);
1501 } else if (unlikely(precreated &&
1502 osp_precreate_end_seq_nolock(d))) {
1504 * precreate pool is reaching the end of the
1505 * current seq, and doesn't have enough objects
1508 spin_unlock(&d->opd_pre_lock);
1512 spin_unlock(&d->opd_pre_lock);
1515 * all precreated objects have been used and no-space
1516 * status leave us no chance to succeed very soon
1517 * but if there is destroy in progress, then we should
1518 * wait till that is done - some space might be released
1520 if (unlikely(rc == -ENOSPC)) {
1521 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1522 /* force local commit to release space */
1523 dt_commit_async(env, d->opd_storage);
1524 osp_sync_check_for_work(d);
1527 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1528 /* just wait till destroys are done
1529 * see wait_event_idle_timeout() below
1532 if (atomic_read(&d->opd_sync_changes) +
1533 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1534 /* no hope for free space */
1539 /* XXX: don't wake up if precreation is in progress */
1540 wake_up(&d->opd_pre_waitq);
1542 if (ktime_get_seconds() >= expire) {
1548 LASSERT(d->opd_pre);
1553 CDEBUG(D_INFO, "%s: Sleeping on objects\n",
1554 d->opd_obd->obd_name);
1555 if (wait_event_idle_timeout(
1556 d->opd_pre_user_waitq,
1557 osp_precreate_ready_condition(env, d),
1558 cfs_time_seconds(obd_timeout)) == 0) {
1560 "%s: slow creates, last="DFID", next="DFID", "
1561 "reserved=%llu, sync_changes=%u, "
1562 "sync_rpcs_in_progress=%d, status=%d\n",
1563 d->opd_obd->obd_name,
1564 PFID(&d->opd_pre_last_created_fid),
1565 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1566 atomic_read(&d->opd_sync_changes),
1567 atomic_read(&d->opd_sync_rpcs_in_progress),
1570 CDEBUG(D_INFO, "%s: Waked up, status=%d\n",
1571 d->opd_obd->obd_name, d->opd_pre_status);
1579 * Get a FID from precreation pool
1581 * The function is a companion for osp_precreate_reserve() - it assigns
1582 * a specific FID from the precreate. The function should be called only
1583 * if the call to osp_precreate_reserve() was successful. The function
1584 * updates a local storage to remember the highest object ID referenced
1585 * by the node in the given sequence.
1587 * A very importan details: this is supposed to be called once the
1588 * transaction is started, so on-disk update will be atomic with the
1589 * data (like LOVEA) refering this object. Then the object won't be leaked:
1590 * either it's referenced by the committed transaction or it's a subject
1591 * to the orphan cleanup procedure.
1593 * \param[in] env LU environment provided by the caller
1594 * \param[in] d OSP device
1595 * \param[out] fid generated FID
1597 * \retval 0 on success
1598 * \retval negative negated errno on error
1600 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1603 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1605 /* grab next id from the pool */
1606 spin_lock(&d->opd_pre_lock);
1608 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1609 &d->opd_pre_last_created_fid) < 0,
1610 "next fid "DFID" > last created fid "DFID"\n",
1611 PFID(&d->opd_pre_used_fid),
1612 PFID(&d->opd_pre_last_created_fid));
1614 /* Non-IDIF FIDs shouldn't get here with OID == OBIF_MAX_OID. For IDIF,
1615 * f_oid wraps and "f_seq" (holding high 16 bits of ID) needs increment
1617 if (fid_is_idif(pre_used_fid) &&
1618 unlikely(fid_oid(pre_used_fid) == OBIF_MAX_OID)) {
1620 __u32 idx = fid_idif_ost_idx(pre_used_fid);
1622 fid_to_ostid(pre_used_fid, &oi);
1624 ostid_to_fid(pre_used_fid, &oi, idx);
1626 pre_used_fid->f_oid++;
1629 memcpy(fid, pre_used_fid, sizeof(*fid));
1630 d->opd_pre_reserved--;
1632 * last_used_id must be changed along with getting new id otherwise
1633 * we might miscalculate gap causing object loss or leak
1635 osp_update_last_fid(d, fid);
1636 spin_unlock(&d->opd_pre_lock);
1639 * probably main thread suspended orphan cleanup till
1640 * all reservations are released, see comment in
1641 * osp_precreate_thread() just before orphan cleanup
1643 if (unlikely(d->opd_pre_reserved == 0 &&
1644 (d->opd_pre_recovering || d->opd_pre_status)))
1645 wake_up(&d->opd_pre_waitq);
1651 * Set size regular attribute on an object
1653 * When a striping is created late, it's possible that size is already
1654 * initialized on the file. Then the new striping should inherit size
1655 * from the file. The function sets size on the object using the regular
1656 * protocol (OST_PUNCH).
1657 * XXX: should be re-implemented using OUT ?
1659 * \param[in] env LU environment provided by the caller
1660 * \param[in] dt object
1661 * \param[in] size size to set.
1663 * \retval 0 on success
1664 * \retval negative negated errno on error
1666 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1669 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1670 struct ptlrpc_request *req = NULL;
1671 struct obd_import *imp;
1672 struct ost_body *body;
1673 struct obdo *oa = NULL;
1678 imp = d->opd_obd->u.cli.cl_import;
1681 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1685 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1687 ptlrpc_request_free(req);
1692 * XXX: decide how do we do here with resend
1693 * if we don't resend, then client may see wrong file size
1694 * if we do resend, then MDS thread can get stuck for quite long
1695 * and if we don't resend, then client will also get -EAGAIN !!
1696 * (see LU-7975 and sanity/test_27F use cases)
1697 * but let's decide not to resend/delay this truncate request to OST
1698 * and allow Client to decide to resend, in a less agressive way from
1699 * after_reply(), by returning -EINPROGRESS instead of
1700 * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
1701 * end of this routine
1703 req->rq_no_resend = req->rq_no_delay = 1;
1705 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1706 ptlrpc_at_set_req_timeout(req);
1710 GOTO(out, rc = -ENOMEM);
1712 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1715 oa->o_blocks = OBD_OBJECT_EOF;
1716 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1717 OBD_MD_FLID | OBD_MD_FLGROUP;
1719 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1721 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1723 /* XXX: capa support? */
1724 /* osc_pack_capa(req, body, capa); */
1726 ptlrpc_request_set_replen(req);
1728 rc = ptlrpc_queue_wait(req);
1730 /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
1731 * since we have decided not to resend/delay, but this could
1732 * lead to wrong size to be seen at Client side and even process
1733 * trying to open to exit/fail if not itself handling -EAGAIN.
1734 * So it should be better to return -EINPROGRESS instead and
1735 * leave the decision to resend at Client side in after_reply()
1737 if (rc == -EAGAIN) {
1739 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1740 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1743 CERROR("can't punch object: %d\n", rc);
1747 ptlrpc_req_finished(req);
1754 * Initialize precreation functionality of OSP
1756 * Prepares all the internal structures and starts the precreate thread
1758 * \param[in] d OSP device
1760 * \retval 0 on success
1761 * \retval negative negated errno on error
1763 int osp_init_precreate(struct osp_device *d)
1767 OBD_ALLOC_PTR(d->opd_pre);
1768 if (d->opd_pre == NULL)
1771 /* initially precreation isn't ready */
1772 init_waitqueue_head(&d->opd_pre_user_waitq);
1773 d->opd_pre_status = -EAGAIN;
1774 fid_zero(&d->opd_pre_used_fid);
1775 d->opd_pre_used_fid.f_oid = 1;
1776 fid_zero(&d->opd_pre_last_created_fid);
1777 d->opd_pre_last_created_fid.f_oid = 1;
1779 d->opd_pre_reserved = 0;
1780 d->opd_pre_seq_width = LUSTRE_DATA_SEQ_MAX_WIDTH;
1781 d->opd_got_disconnected = 1;
1782 d->opd_pre_create_slow = 0;
1783 d->opd_pre_create_count = OST_MIN_PRECREATE;
1784 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1785 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1786 d->opd_reserved_mb_high = 0;
1787 d->opd_reserved_mb_low = 0;
1788 d->opd_cleanup_orphans_done = false;
1789 d->opd_force_creation = false;
1795 * Finish precreate functionality of OSP
1798 * Asks all the activity (the thread, update timer) to stop, then
1799 * wait till that is done.
1801 * \param[in] d OSP device
1803 void osp_precreate_fini(struct osp_device *d)
1807 if (d->opd_pre == NULL)
1810 OBD_FREE_PTR(d->opd_pre);
1816 int osp_init_statfs(struct osp_device *d)
1818 struct task_struct *task;
1819 struct opt_args *args;
1820 DECLARE_COMPLETION_ONSTACK(started);
1825 spin_lock_init(&d->opd_pre_lock);
1826 init_waitqueue_head(&d->opd_pre_waitq);
1829 * Initialize statfs-related things
1831 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1832 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1833 1000 * NSEC_PER_SEC);
1834 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1836 ktime_to_ns(d->opd_statfs_fresh_till));
1837 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1838 (unsigned long)d, 0);
1840 if (d->opd_storage->dd_rdonly)
1843 OBD_ALLOC_PTR(args);
1847 args->opta_started = &started;
1848 rc = lu_env_init(&args->opta_env,
1849 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1851 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1858 * start thread handling precreation and statfs updates
1860 task = kthread_create(osp_precreate_thread, args,
1861 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1863 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1864 lu_env_fini(&args->opta_env);
1866 RETURN(PTR_ERR(task));
1868 d->opd_pre_task = task;
1869 wake_up_process(task);
1870 wait_for_completion(&started);
1875 void osp_statfs_fini(struct osp_device *d)
1877 struct task_struct *task = d->opd_pre_task;
1880 timer_delete(&d->opd_statfs_timer);
1882 d->opd_pre_task = NULL;