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 d->opd_statfs = *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;
174 CDEBUG(D_CACHE, "%s (%p): %llu blocks, %llu free, %llu avail, "
175 "%u bsize, %u reserved mb low, %u reserved mb high,"
176 "%llu files, %llu free files\n", d->opd_obd->obd_name, d,
177 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail, sfs->os_bsize,
178 d->opd_reserved_mb_low, d->opd_reserved_mb_high,
179 sfs->os_files, sfs->os_ffree);
183 /* couldn't update statfs, try again with a small delay */
184 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
185 d->opd_statfs_update_in_progress = 0;
186 if (d->opd_pre && d->opd_pre_task)
187 wake_up(&d->opd_pre_waitq);
189 if (req->rq_import_generation == imp->imp_generation)
190 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
191 d->opd_obd->obd_name, rc);
196 * Send OST_STATFS RPC
198 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
199 * Also disables scheduled updates as times OSP may need to refresh
200 * statfs data before expiration. The function doesn't block, instead
201 * an interpretation callback osp_statfs_interpret() is used.
203 * \param[in] d OSP device
205 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
207 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
208 struct ptlrpc_request *req;
209 struct obd_import *imp;
210 union ptlrpc_async_args *aa;
215 CDEBUG(D_CACHE, "going to update statfs\n");
217 imp = d->opd_obd->u.cli.cl_import;
220 req = ptlrpc_request_alloc(imp,
221 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
225 rc = ptlrpc_request_pack(req,
226 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
227 d->opd_pre ? OST_STATFS : MDS_STATFS);
229 ptlrpc_request_free(req);
232 ptlrpc_request_set_replen(req);
234 req->rq_request_portal = OST_CREATE_PORTAL;
235 ptlrpc_at_set_req_timeout(req);
237 req->rq_interpret_reply = osp_statfs_interpret;
238 aa = ptlrpc_req_async_args(aa, req);
239 aa->pointer_arg[0] = d;
242 * no updates till reply
244 del_timer(&d->opd_statfs_timer);
245 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
246 d->opd_statfs_update_in_progress = 1;
248 ptlrpcd_add_req(req);
250 /* we still want to sync changes if no new changes are coming */
251 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
254 if (atomic_read(&d->opd_sync_changes)) {
257 th = dt_trans_create(env, d->opd_storage);
259 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
262 rc = dt_trans_start_local(env, d->opd_storage, th);
264 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
265 d->opd_obd->obd_name,
266 atomic_read(&d->opd_sync_changes));
267 osp_sync_add_commit_cb_1s(env, d, th);
269 dt_trans_stop(env, d->opd_storage, th);
277 * Schedule an immediate update for statfs data
279 * If cached statfs data claim no free space, but OSP has got a request to
280 * destroy an object (so release some space probably), then we may need to
281 * refresh cached statfs data sooner than planned. The function checks there
282 * is no statfs update going and schedules immediate update if so.
283 * XXX: there might be a case where removed object(s) do not add free space (empty
284 * object). If the number of such deletions is high, then we can start to update
285 * statfs too often causing a RPC storm. some throttling is needed...
287 * \param[in] d OSP device where statfs data needs to be refreshed
289 void osp_statfs_need_now(struct osp_device *d)
291 if (!d->opd_statfs_update_in_progress) {
293 * if current status is -ENOSPC (lack of free space on OST)
294 * then we should poll OST immediately once object destroy
297 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
298 del_timer(&d->opd_statfs_timer);
299 wake_up(&d->opd_pre_waitq);
304 * Return number of precreated objects
306 * A simple helper to calculate the number of precreated objects on the device.
308 * \param[in] env LU environment provided by the caller
309 * \param[in] osp OSP device
311 * \retval the number of the precreated objects
313 static inline int osp_objs_precreated(const struct lu_env *env,
314 struct osp_device *osp)
316 return osp_fid_diff(&osp->opd_pre_last_created_fid,
317 &osp->opd_pre_used_fid);
321 * Check pool of precreated objects is nearly empty
323 * We should not wait till the pool of the precreated objects is exhausted,
324 * because then there will be a long period of OSP being unavailable for the
325 * new creations due to lenghty precreate RPC. Instead we ask for another
326 * precreation ahead and hopefully have it ready before the current pool is
327 * empty. Notice this function relies on an external locking.
329 * \param[in] env LU environment provided by the caller
330 * \param[in] d OSP device
332 * \retval 0 - current pool is good enough, 1 - time to precreate
334 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
335 struct osp_device *d)
337 int window = osp_objs_precreated(env, d);
339 /* don't consider new precreation till OST is healty and
341 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
342 (d->opd_pre_status == 0));
346 * Check pool of precreated objects
348 * This is protected version of osp_precreate_near_empty_nolock(), check that
351 * \param[in] env LU environment provided by the caller
352 * \param[in] d OSP device
354 * \retval 0 - current pool is good enough, 1 - time to precreate
356 static inline int osp_precreate_near_empty(const struct lu_env *env,
357 struct osp_device *d)
361 if (d->opd_pre == NULL)
364 /* XXX: do we really need locking here? */
365 spin_lock(&d->opd_pre_lock);
366 rc = osp_precreate_near_empty_nolock(env, d);
367 spin_unlock(&d->opd_pre_lock);
372 * Check given sequence is empty
374 * Returns a binary result whether the given sequence has some IDs left
375 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
376 * version of that function.
378 * \param[in] env LU environment provided by the caller
379 * \param[in] osp OSP device
381 * \retval 0 - current sequence has no IDs, 1 - otherwise
383 static inline int osp_create_end_seq(const struct lu_env *env,
384 struct osp_device *osp)
386 struct lu_fid *fid = &osp->opd_pre_used_fid;
389 spin_lock(&osp->opd_pre_lock);
390 rc = osp_fid_end_seq(env, fid);
391 spin_unlock(&osp->opd_pre_lock);
396 * Write FID into into last_oid/last_seq file
398 * The function stores the sequence and the in-sequence id into two dedicated
399 * files. The sync argument can be used to request synchronous commit, so the
400 * function won't return until the updates are committed.
402 * \param[in] env LU environment provided by the caller
403 * \param[in] osp OSP device
404 * \param[in] fid fid where sequence/id is taken
405 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
407 * \retval 0 on success
408 * \retval negative negated errno on error
410 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
411 struct lu_fid *fid, int sync)
413 struct osp_thread_info *oti = osp_env_info(env);
414 struct lu_buf *lb_oid = &oti->osi_lb;
415 struct lu_buf *lb_oseq = &oti->osi_lb2;
423 if (osp->opd_storage->dd_rdonly)
426 /* Note: through f_oid is only 32 bits, it will also write 64 bits
427 * for oid to keep compatibility with the previous version. */
429 osp_objid_buf_prep(lb_oid, &oid_off,
430 &oid, osp->opd_index);
432 osp_objseq_buf_prep(lb_oseq, &oseq_off,
433 &fid->f_seq, osp->opd_index);
435 th = dt_trans_create(env, osp->opd_storage);
440 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
441 lb_oid, oid_off, th);
445 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
446 lb_oseq, oseq_off, th);
450 rc = dt_trans_start_local(env, osp->opd_storage, th);
454 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
457 CERROR("%s: can not write to last seq file: rc = %d\n",
458 osp->opd_obd->obd_name, rc);
461 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
464 CERROR("%s: can not write to last seq file: rc = %d\n",
465 osp->opd_obd->obd_name, rc);
469 dt_trans_stop(env, osp->opd_storage, th);
474 * Switch to another sequence
476 * When a current sequence has no available IDs left, OSP has to switch to
477 * another new sequence. OSP requests it using the regular FLDB protocol
478 * and stores synchronously before that is used in precreated. This is needed
479 * to basically have the sequences referenced (not orphaned), otherwise it's
480 * possible that OST has some objects precreated and the clients have data
481 * written to it, but after MDT failover nobody refers those objects and OSP
482 * has no idea that the sequence need cleanup to be done.
483 * While this is very expensive operation, it's supposed to happen very very
484 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
486 * \param[in] env LU environment provided by the caller
487 * \param[in] osp OSP device
489 * \retval 0 on success
490 * \retval negative negated errno on error
492 static int osp_precreate_rollover_new_seq(struct lu_env *env,
493 struct osp_device *osp)
495 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
496 struct lu_fid *last_fid = &osp->opd_last_used_fid;
500 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
502 CERROR("%s: alloc fid error: rc = %d\n",
503 osp->opd_obd->obd_name, rc);
509 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
510 "fid "DFID", last_fid "DFID"\n", PFID(fid),
513 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
515 CERROR("%s: Can not update oid/seq file: rc = %d\n",
516 osp->opd_obd->obd_name, rc);
520 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
521 osp->opd_obd->obd_name, fid_seq(last_fid),
523 /* Update last_xxx to the new seq */
524 spin_lock(&osp->opd_pre_lock);
525 osp->opd_last_used_fid = *fid;
526 osp_fid_to_obdid(fid, &osp->opd_last_id);
527 osp->opd_gap_start_fid = *fid;
528 osp->opd_pre_used_fid = *fid;
529 osp->opd_pre_last_created_fid = *fid;
530 spin_unlock(&osp->opd_pre_lock);
536 * Find IDs available in current sequence
538 * The function calculates the highest possible ID and the number of IDs
539 * available in the current sequence OSP is using. The number is limited
540 * artifically by the caller (grow param) and the number of IDs available
541 * in the sequence by nature. The function doesn't require an external
544 * \param[in] env LU environment provided by the caller
545 * \param[in] osp OSP device
546 * \param[in] fid FID the caller wants to start with
547 * \param[in] grow how many the caller wants
548 * \param[out] fid the highest calculated FID
549 * \param[out] grow the number of available IDs calculated
551 * \retval 0 on success, 1 - the sequence is empty
553 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
554 struct lu_fid *fid, int *grow)
556 struct osp_thread_info *osi = osp_env_info(env);
560 if (fid_is_idif(fid)) {
561 struct lu_fid *last_fid;
562 struct ost_id *oi = &osi->osi_oi;
565 spin_lock(&osp->opd_pre_lock);
566 last_fid = &osp->opd_pre_last_created_fid;
567 fid_to_ostid(last_fid, oi);
568 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
569 *grow = end - ostid_id(oi);
570 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
571 spin_unlock(&osp->opd_pre_lock);
573 if (*grow == 0 || rc)
576 ostid_to_fid(fid, oi, osp->opd_index);
580 spin_lock(&osp->opd_pre_lock);
581 *fid = osp->opd_pre_last_created_fid;
583 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
584 *grow = end - fid->f_oid;
585 fid->f_oid += end - fid->f_oid;
586 spin_unlock(&osp->opd_pre_lock);
588 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
589 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
591 return *grow > 0 ? 0 : 1;
595 * Prepare and send precreate RPC
597 * The function finds how many objects should be precreated. Then allocates,
598 * prepares and schedules precreate RPC synchronously. Upon reply the function
599 * wakes up the threads waiting for the new objects on this target. If the
600 * target wasn't able to create all the objects requested, then the next
601 * precreate will be asking for fewer objects (i.e. slow precreate down).
603 * \param[in] env LU environment provided by the caller
604 * \param[in] d OSP device
606 * \retval 0 on success
607 * \retval negative negated errno on error
609 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
611 struct osp_thread_info *oti = osp_env_info(env);
612 struct ptlrpc_request *req;
613 struct obd_import *imp;
614 struct ost_body *body;
616 struct lu_fid *fid = &oti->osi_fid;
619 /* don't precreate new objects till OST healthy and has free space */
620 if (unlikely(d->opd_pre_status)) {
621 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
622 d->opd_obd->obd_name, d->opd_pre_status);
627 * if not connection/initialization is compeleted, ignore
629 imp = d->opd_obd->u.cli.cl_import;
632 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
635 req->rq_request_portal = OST_CREATE_PORTAL;
636 /* we should not resend create request - anyway we will have delorphan
637 * and kill these objects */
638 req->rq_no_delay = req->rq_no_resend = 1;
640 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
642 ptlrpc_request_free(req);
646 spin_lock(&d->opd_pre_lock);
647 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
648 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
649 grow = d->opd_pre_create_count;
650 spin_unlock(&d->opd_pre_lock);
652 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
655 *fid = d->opd_pre_last_created_fid;
656 rc = osp_precreate_fids(env, d, fid, &grow);
658 /* Current seq has been used up*/
659 GOTO(out_req, rc = -ENOSPC);
661 if (!osp_is_fid_client(d)) {
662 /* Non-FID client will always send seq 0 because of
664 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
668 fid_to_ostid(fid, &body->oa.o_oi);
669 body->oa.o_valid = OBD_MD_FLGROUP;
671 ptlrpc_request_set_replen(req);
673 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
676 rc = ptlrpc_queue_wait(req);
678 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
681 /* have osp_precreate_reserve() to wait for repeat */
685 LASSERT(req->rq_transno == 0);
687 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
689 GOTO(out_req, rc = -EPROTO);
691 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
694 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
695 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
696 ": rc = %d\n", d->opd_obd->obd_name,
697 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
698 GOTO(out_req, rc = -ESTALE);
701 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
703 spin_lock(&d->opd_pre_lock);
705 /* the OST has not managed to create all the
706 * objects we asked for */
707 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
708 d->opd_pre_create_slow = 1;
710 /* the OST is able to keep up with the work,
711 * we could consider increasing create_count
712 * next time if needed */
713 d->opd_pre_create_slow = 0;
716 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
717 fid_to_ostid(fid, &body->oa.o_oi);
719 d->opd_pre_last_created_fid = *fid;
720 spin_unlock(&d->opd_pre_lock);
722 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
723 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
724 PFID(&d->opd_pre_last_created_fid));
726 /* now we can wakeup all users awaiting for objects */
727 osp_pre_update_status(d, rc);
729 /* pause to let osp_precreate_reserve to go first */
730 CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_PRECREATE_PAUSE, 2);
732 ptlrpc_req_finished(req);
737 * Get last precreated object from target (OST)
739 * Sends synchronous RPC to the target (OST) to learn the last precreated
740 * object. This later is used to remove all unused objects (cleanup orphan
741 * procedure). Also, the next object after one we got will be used as a
742 * starting point for the new precreates.
744 * \param[in] env LU environment provided by the caller
745 * \param[in] d OSP device
747 * \retval 0 on success
748 * \retval negative negated errno on error
750 static int osp_get_lastfid_from_ost(const struct lu_env *env,
751 struct osp_device *d)
753 struct ptlrpc_request *req = NULL;
754 struct obd_import *imp;
755 struct lu_fid *last_fid;
760 imp = d->opd_obd->u.cli.cl_import;
763 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
767 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
768 sizeof(KEY_LAST_FID));
770 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
772 ptlrpc_request_free(req);
776 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
777 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
779 req->rq_no_delay = req->rq_no_resend = 1;
780 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
781 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
783 ptlrpc_request_set_replen(req);
785 rc = ptlrpc_queue_wait(req);
787 /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld),
788 * let sysadm sort this * out.
791 ptlrpc_set_import_active(imp, 0);
795 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
796 if (last_fid == NULL) {
797 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
798 GOTO(out, rc = -EPROTO);
801 if (!fid_is_sane(last_fid)) {
802 CERROR("%s: Got insane last_fid "DFID"\n",
803 d->opd_obd->obd_name, PFID(last_fid));
804 GOTO(out, rc = -EPROTO);
807 /* Only update the last used fid, if the OST has objects for
808 * this sequence, i.e. fid_oid > 0 */
809 if (fid_oid(last_fid) > 0)
810 d->opd_last_used_fid = *last_fid;
812 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
816 ptlrpc_req_finished(req);
821 * Cleanup orphans on OST
823 * This function is called in a contex of a dedicated thread handling
824 * all the precreation suff. The function waits till local recovery
825 * is complete, then identify all the unreferenced objects (orphans)
826 * using the highest ID referenced by a local and the highest object
827 * precreated by the target. The found range is a subject to removal
828 * using specially flagged RPC. During this process OSP is marked
829 * unavailable for new objects.
831 * \param[in] env LU environment provided by the caller
832 * \param[in] d OSP device
834 * \retval 0 on success
835 * \retval negative negated errno on error
837 static int osp_precreate_cleanup_orphans(struct lu_env *env,
838 struct osp_device *d)
840 struct osp_thread_info *osi = osp_env_info(env);
841 struct lu_fid *last_fid = &osi->osi_fid;
842 struct ptlrpc_request *req = NULL;
843 struct obd_import *imp;
844 struct ost_body *body;
845 int update_status = 0;
852 * wait for local recovery to finish, so we can cleanup orphans
853 * orphans are all objects since "last used" (assigned), but
854 * there might be objects reserved and in some cases they won't
855 * be used. we can't cleanup them till we're sure they won't be
856 * used. also can't we allow new reservations because they may
857 * end up getting orphans being cleaned up below. so we block
858 * new reservations and wait till all reserved objects either
861 spin_lock(&d->opd_pre_lock);
862 d->opd_pre_recovering = 1;
863 spin_unlock(&d->opd_pre_lock);
865 * The locking above makes sure the opd_pre_reserved check below will
866 * catch all osp_precreate_reserve() calls who find
867 * "!opd_pre_recovering".
869 wait_event_idle(d->opd_pre_waitq,
870 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
871 !d->opd_pre_task || d->opd_got_disconnected);
872 if (!d->opd_pre_task || d->opd_got_disconnected)
873 GOTO(out, rc = -EAGAIN);
875 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
876 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
878 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val);
880 *last_fid = d->opd_last_used_fid;
881 /* The OSP should already get the valid seq now */
882 LASSERT(!fid_is_zero(last_fid));
883 if (fid_oid(&d->opd_last_used_fid) < 2 ||
884 OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
885 /* lastfid looks strange... ask OST */
886 rc = osp_get_lastfid_from_ost(env, d);
891 imp = d->opd_obd->u.cli.cl_import;
894 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
896 GOTO(out, rc = -ENOMEM);
898 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
900 ptlrpc_request_free(req);
905 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
907 GOTO(out, rc = -EPROTO);
909 body->oa.o_flags = OBD_FL_DELORPHAN;
910 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
912 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
914 ptlrpc_request_set_replen(req);
916 /* Don't resend the delorphan req */
917 req->rq_no_resend = req->rq_no_delay = 1;
919 rc = ptlrpc_queue_wait(req);
925 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
927 GOTO(out, rc = -EPROTO);
930 * OST provides us with id new pool starts from in body->oa.o_id
932 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
934 spin_lock(&d->opd_pre_lock);
935 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
937 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
938 d->opd_pre_last_created_fid = d->opd_last_used_fid;
940 d->opd_pre_create_count = OST_MIN_PRECREATE;
941 d->opd_pre_last_created_fid = *last_fid;
944 * This empties the pre-creation pool and effectively blocks any new
947 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
948 LUSTRE_DATA_SEQ_MAX_WIDTH);
949 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
950 d->opd_pre_create_slow = 0;
951 spin_unlock(&d->opd_pre_lock);
953 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
954 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
955 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
958 ptlrpc_req_finished(req);
962 * If rc is zero, the pre-creation window should have been emptied.
963 * Since waking up the herd would be useless without pre-created
964 * objects, we defer the signal to osp_precreate_send() in that case.
968 CERROR("%s: cannot cleanup orphans: rc = %d\n",
969 d->opd_obd->obd_name, rc);
970 /* we can't proceed from here, OST seem to
971 * be in a bad shape, better to wait for
972 * a new instance of the server and repeat
973 * from the beginning. notify possible waiters
974 * this OSP isn't quite functional yet */
975 osp_pre_update_status(d, rc);
977 wake_up_all(&d->opd_pre_user_waitq);
980 spin_lock(&d->opd_pre_lock);
981 d->opd_pre_recovering = 0;
982 spin_unlock(&d->opd_pre_lock);
989 * Update precreate status using statfs data
991 * The function decides whether this OSP should be used for new objects.
992 * IOW, whether this OST is used up or has some free space. Cached statfs
993 * data is used to make this decision. If the latest result of statfs
994 * request (rc argument) is not success, then just mark OSP unavailable
997 * The new statfs data is passed in \a msfs and needs to be stored into
998 * opd_statfs, but only after the various flags in os_state are set, so
999 * that the new statfs data is not visible without appropriate flags set.
1000 * As such, there is no need to clear the flags here, since this is called
1001 * with new statfs data, and they should not be cleared if sent from OST.
1003 * Add a bit of hysteresis so this flag isn't continually flapping, and
1004 * ensure that new files don't get extremely fragmented due to only a
1005 * small amount of available space in the filesystem. We want to set
1006 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1007 * reserved size free, and still copy them from the old state when there
1008 * is less than 2*reserved size free space or inodes.
1010 * \param[in] d OSP device
1011 * \param[in] msfs statfs data
1013 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1015 u32 old_state = d->opd_statfs.os_state;
1016 u32 reserved_ino_low = 32; /* could be tunable in the future */
1017 u32 reserved_ino_high = reserved_ino_low * 2;
1020 /* statfs structure not initialized yet */
1021 if (unlikely(!msfs->os_type))
1024 /* if the low and high watermarks have not been initialized yet */
1025 if (unlikely(d->opd_reserved_mb_high == 0 &&
1026 d->opd_reserved_mb_low == 0)) {
1027 /* Use ~0.1% by default to disable object allocation,
1028 * and ~0.2% to enable, size in MB, set both watermark
1030 spin_lock(&d->opd_pre_lock);
1031 if (d->opd_reserved_mb_high == 0 &&
1032 d->opd_reserved_mb_low == 0) {
1033 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1034 msfs->os_blocks) >> 20;
1035 if (d->opd_reserved_mb_low == 0)
1036 d->opd_reserved_mb_low = 1;
1037 d->opd_reserved_mb_high =
1038 (d->opd_reserved_mb_low << 1) + 1;
1040 spin_unlock(&d->opd_pre_lock);
1043 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1044 if (msfs->os_ffree < reserved_ino_low)
1045 msfs->os_state |= OS_STATFS_ENOINO;
1046 else if (msfs->os_ffree <= reserved_ino_high)
1047 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1048 /* else don't clear flags in new msfs->os_state sent from OST */
1051 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1052 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1053 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1054 msfs->os_files, msfs->os_ffree, msfs->os_state,
1056 if (available_mb < d->opd_reserved_mb_low)
1057 msfs->os_state |= OS_STATFS_ENOSPC;
1058 else if (available_mb <= d->opd_reserved_mb_high)
1059 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1060 /* else don't clear flags in new msfs->os_state sent from OST */
1062 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1063 d->opd_pre_status = -ENOSPC;
1064 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1065 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1066 d->opd_obd->obd_name, msfs->os_state,
1068 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1069 atomic_read(&d->opd_sync_changes),
1070 atomic_read(&d->opd_sync_rpcs_in_progress));
1071 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1072 d->opd_pre_status = 0;
1073 spin_lock(&d->opd_pre_lock);
1074 d->opd_pre_create_slow = 0;
1075 d->opd_pre_create_count = OST_MIN_PRECREATE;
1076 spin_unlock(&d->opd_pre_lock);
1077 wake_up(&d->opd_pre_waitq);
1080 "%s: available: state=%x: rc = %d\n",
1081 d->opd_obd->obd_name, msfs->os_state,
1084 /* we only get here if rc == 0 in the caller */
1085 d->opd_pre_status = 0;
1088 /* Object precreation skipped on OST if manually disabled */
1089 if (d->opd_pre_max_create_count == 0)
1090 msfs->os_state |= OS_STATFS_NOPRECREATE;
1091 /* else don't clear flags in new msfs->os_state sent from OST */
1093 /* copy only new statfs state to make it visible to MDS threads */
1094 if (&d->opd_statfs != msfs)
1095 d->opd_statfs = *msfs;
1099 * Initialize FID for precreation
1101 * For a just created new target, a new sequence should be taken.
1102 * The function checks there is no IDIF in use (if the target was
1103 * added with the older version of Lustre), then requests a new
1104 * sequence from FLDB using the regular protocol. Then this new
1105 * sequence is stored on a persisten storage synchronously to prevent
1106 * possible object leakage (for the detail see the description for
1107 * osp_precreate_rollover_new_seq()).
1109 * \param[in] osp OSP device
1111 * \retval 0 on success
1112 * \retval negative negated errno on error
1114 int osp_init_pre_fid(struct osp_device *osp)
1117 struct osp_thread_info *osi;
1118 struct lu_client_seq *cli_seq;
1119 struct lu_fid *last_fid;
1123 LASSERT(osp->opd_pre != NULL);
1125 /* Let's check if the current last_seq/fid is valid,
1126 * otherwise request new sequence from the controller */
1127 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1128 /* Non-MDT0 can only use normal sequence for
1130 if (fid_is_norm(&osp->opd_last_used_fid))
1133 /* Initially MDT0 will start with IDIF, after
1134 * that it will request new sequence from the
1136 if (fid_is_idif(&osp->opd_last_used_fid) ||
1137 fid_is_norm(&osp->opd_last_used_fid))
1141 if (!fid_is_zero(&osp->opd_last_used_fid))
1142 CWARN("%s: invalid last used fid "DFID
1143 ", try to get new sequence.\n",
1144 osp->opd_obd->obd_name,
1145 PFID(&osp->opd_last_used_fid));
1147 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1149 CERROR("%s: init env error: rc = %d\n",
1150 osp->opd_obd->obd_name, rc);
1154 osi = osp_env_info(&env);
1155 last_fid = &osi->osi_fid;
1157 /* For a freshed fs, it will allocate a new sequence first */
1158 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1159 cli_seq = osp->opd_obd->u.cli.cl_seq;
1160 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1162 CERROR("%s: alloc fid error: rc = %d\n",
1163 osp->opd_obd->obd_name, rc);
1167 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1169 last_fid->f_oid = 1;
1170 last_fid->f_ver = 0;
1172 spin_lock(&osp->opd_pre_lock);
1173 osp->opd_last_used_fid = *last_fid;
1174 osp->opd_pre_used_fid = *last_fid;
1175 osp->opd_pre_last_created_fid = *last_fid;
1176 spin_unlock(&osp->opd_pre_lock);
1177 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1179 CERROR("%s: write fid error: rc = %d\n",
1180 osp->opd_obd->obd_name, rc);
1189 struct osp_device *opta_dev;
1190 struct lu_env opta_env;
1191 struct completion *opta_started;
1194 * The core of precreate functionality
1196 * The function implements the main precreation loop. Basically it
1197 * involves connecting to the target, precerate FID initialization,
1198 * identifying and removing orphans, then serving precreation. As
1199 * part of the latter, the thread is responsible for statfs data
1200 * updates. The precreation is mostly driven by another threads
1201 * asking for new OST objects - those askers wake the thread when
1202 * the number of precreated objects reach low watermark.
1203 * After a disconnect, the sequence above repeats. This is keep going
1204 * until the thread is requested to stop.
1206 * \param[in] _arg private data the thread (OSP device to handle)
1208 * \retval 0 on success
1209 * \retval negative negated errno on error
1211 static int osp_precreate_thread(void *_args)
1213 struct opt_args *args = _args;
1214 struct osp_device *d = args->opta_dev;
1215 struct lu_env *env = &args->opta_env;
1220 complete(args->opta_started);
1221 while (!kthread_should_stop()) {
1223 * need to be connected to OST
1225 while (!kthread_should_stop()) {
1226 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1227 d->opd_imp_connected &&
1228 !d->opd_got_disconnected)
1230 wait_event_idle(d->opd_pre_waitq,
1231 kthread_should_stop() ||
1232 d->opd_new_connection);
1234 if (!d->opd_new_connection)
1237 OBD_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
1239 d->opd_new_connection = 0;
1240 d->opd_got_disconnected = 0;
1244 if (kthread_should_stop())
1248 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1249 /* Sigh, fid client is not ready yet */
1250 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1253 /* Init fid for osp_precreate if necessary */
1254 rc = osp_init_pre_fid(d);
1256 class_export_put(d->opd_exp);
1257 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1258 CERROR("%s: init pre fid error: rc = %d\n",
1259 d->opd_obd->obd_name, rc);
1264 if (osp_statfs_update(env, d)) {
1265 if (wait_event_idle_timeout(d->opd_pre_waitq,
1266 kthread_should_stop(),
1267 cfs_time_seconds(5)) == 0)
1268 l_wait_event_abortable(
1270 kthread_should_stop());
1276 * Clean up orphans or recreate missing objects.
1278 rc = osp_precreate_cleanup_orphans(env, d);
1280 schedule_timeout_interruptible(cfs_time_seconds(1));
1286 * connected, can handle precreates now
1288 while (!kthread_should_stop()) {
1289 wait_event_idle(d->opd_pre_waitq,
1290 kthread_should_stop() ||
1291 osp_precreate_near_empty(env, d) ||
1292 osp_statfs_need_update(d) ||
1293 d->opd_got_disconnected);
1295 if (kthread_should_stop())
1298 /* something happened to the connection
1299 * have to start from the beginning */
1300 if (d->opd_got_disconnected)
1303 if (osp_statfs_need_update(d))
1304 if (osp_statfs_update(env, d))
1307 if (d->opd_pre == NULL)
1310 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
1311 d->opd_pre_recovering = 1;
1315 /* To avoid handling different seq in precreate/orphan
1316 * cleanup, it will hold precreate until current seq is
1318 if (unlikely(osp_precreate_end_seq(env, d) &&
1319 !osp_create_end_seq(env, d)))
1322 if (unlikely(osp_precreate_end_seq(env, d) &&
1323 osp_create_end_seq(env, d))) {
1324 LCONSOLE_INFO("%s:%#llx is used up."
1325 " Update to new seq\n",
1326 d->opd_obd->obd_name,
1327 fid_seq(&d->opd_pre_last_created_fid));
1328 rc = osp_precreate_rollover_new_seq(env, d);
1333 if (osp_precreate_near_empty(env, d)) {
1334 rc = osp_precreate_send(env, d);
1335 /* osp_precreate_send() sets opd_pre_status
1336 * in case of error, that prevent the using of
1338 if (rc < 0 && rc != -ENOSPC &&
1339 rc != -ETIMEDOUT && rc != -ENOTCONN)
1340 CERROR("%s: cannot precreate objects:"
1342 d->opd_obd->obd_name, rc);
1354 * Check when to stop to wait for precreate objects.
1356 * The caller wanting a new OST object can't wait undefinitely. The
1357 * function checks for few conditions including available new OST
1358 * objects, disconnected OST, lack of space with no pending destroys,
1359 * etc. IOW, it checks whether the current OSP state is good to keep
1360 * waiting or it's better to give up.
1362 * \param[in] env LU environment provided by the caller
1363 * \param[in] d OSP device
1365 * \retval 0 - keep waiting, 1 - no luck
1367 static int osp_precreate_ready_condition(const struct lu_env *env,
1368 struct osp_device *d)
1370 /* Bail out I/O fails to OST */
1371 if (d->opd_pre_status != 0 &&
1372 d->opd_pre_status != -EAGAIN &&
1373 d->opd_pre_status != -ENODEV &&
1374 d->opd_pre_status != -ENOTCONN &&
1375 d->opd_pre_status != -ENOSPC) {
1377 if (d->opd_pre_status != -EIO)
1378 CERROR("%s: precreate failed opd_pre_status %d\n",
1379 d->opd_obd->obd_name, d->opd_pre_status);
1383 if (d->opd_pre_recovering)
1386 /* ready if got enough precreated objects */
1387 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1388 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1391 /* ready if OST reported no space and no destroys in progress */
1392 if (atomic_read(&d->opd_sync_changes) +
1393 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1394 d->opd_pre_status == -ENOSPC)
1401 * Reserve object in precreate pool
1403 * When the caller wants to create a new object on this target (target
1404 * represented by the given OSP), it should declare this intention using
1405 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1406 * to reserve an object in the existing precreated pool or wait up to
1407 * obd_timeout for the available object to appear in the pool (a dedicated
1408 * thread will be doing real precreation in background). The object can be
1409 * consumed later with osp_precreate_get_fid() or be released with call to
1410 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1411 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1412 * If the space on the target is short and there is a pending object destroy,
1413 * then the function forces local commit to speedup space release (see
1414 * osp_sync.c for the details).
1416 * \param[in] env LU environment provided by the caller
1417 * \param[in] d OSP device
1419 * \retval 0 on success
1420 * \retval -ENOSPC when no space on OST
1421 * \retval -EAGAIN try later, slow precreation in progress
1422 * \retval -EIO when no access to OST
1424 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
1427 time64_t expire = ktime_get_seconds() + obd_timeout;
1428 int precreated, rc, synced = 0;
1432 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1433 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1434 PFID(&d->opd_pre_used_fid));
1436 /* opd_pre_max_create_count 0 to not use specified OST. */
1437 if (d->opd_pre_max_create_count == 0)
1442 * - preallocation is done
1443 * - no free space expected soon
1444 * - can't connect to OST for too long (obd_timeout)
1445 * - OST can allocate fid sequence.
1447 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1448 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1451 * increase number of precreations
1453 precreated = osp_objs_precreated(env, d);
1454 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1455 d->opd_pre_create_slow == 0 &&
1456 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1457 spin_lock(&d->opd_pre_lock);
1458 d->opd_pre_create_slow = 1;
1459 d->opd_pre_create_count *= 2;
1460 spin_unlock(&d->opd_pre_lock);
1463 spin_lock(&d->opd_pre_lock);
1464 precreated = osp_objs_precreated(env, d);
1465 if (precreated > d->opd_pre_reserved &&
1466 !d->opd_pre_recovering) {
1467 d->opd_pre_reserved++;
1468 spin_unlock(&d->opd_pre_lock);
1471 /* XXX: don't wake up if precreation is in progress */
1472 if (osp_precreate_near_empty_nolock(env, d) &&
1473 !osp_precreate_end_seq_nolock(env, d))
1474 wake_up(&d->opd_pre_waitq);
1478 spin_unlock(&d->opd_pre_lock);
1481 * all precreated objects have been used and no-space
1482 * status leave us no chance to succeed very soon
1483 * but if there is destroy in progress, then we should
1484 * wait till that is done - some space might be released
1486 if (unlikely(rc == -ENOSPC)) {
1487 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1488 /* force local commit to release space */
1489 dt_commit_async(env, d->opd_storage);
1490 osp_sync_check_for_work(d);
1493 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1494 /* just wait till destroys are done
1495 * see wait_event_idle_timeout() below
1498 if (atomic_read(&d->opd_sync_changes) +
1499 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1500 /* no hope for free space */
1505 /* XXX: don't wake up if precreation is in progress */
1506 wake_up(&d->opd_pre_waitq);
1508 if (ktime_get_seconds() >= expire) {
1514 LASSERT(d->opd_pre);
1519 CDEBUG(D_INFO, "%s: Sleeping on objects\n",
1520 d->opd_obd->obd_name);
1521 if (wait_event_idle_timeout(
1522 d->opd_pre_user_waitq,
1523 osp_precreate_ready_condition(env, d),
1524 cfs_time_seconds(obd_timeout)) == 0) {
1526 "%s: slow creates, last="DFID", next="DFID", "
1527 "reserved=%llu, sync_changes=%u, "
1528 "sync_rpcs_in_progress=%d, status=%d\n",
1529 d->opd_obd->obd_name,
1530 PFID(&d->opd_pre_last_created_fid),
1531 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1532 atomic_read(&d->opd_sync_changes),
1533 atomic_read(&d->opd_sync_rpcs_in_progress),
1536 CDEBUG(D_INFO, "%s: Waked up, status=%d\n",
1537 d->opd_obd->obd_name, d->opd_pre_status);
1545 * Get a FID from precreation pool
1547 * The function is a companion for osp_precreate_reserve() - it assigns
1548 * a specific FID from the precreate. The function should be called only
1549 * if the call to osp_precreate_reserve() was successful. The function
1550 * updates a local storage to remember the highest object ID referenced
1551 * by the node in the given sequence.
1553 * A very importan details: this is supposed to be called once the
1554 * transaction is started, so on-disk update will be atomic with the
1555 * data (like LOVEA) refering this object. Then the object won't be leaked:
1556 * either it's referenced by the committed transaction or it's a subject
1557 * to the orphan cleanup procedure.
1559 * \param[in] env LU environment provided by the caller
1560 * \param[in] d OSP device
1561 * \param[out] fid generated FID
1563 * \retval 0 on success
1564 * \retval negative negated errno on error
1566 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1569 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1570 /* grab next id from the pool */
1571 spin_lock(&d->opd_pre_lock);
1573 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1574 &d->opd_pre_last_created_fid) < 0,
1575 "next fid "DFID" last created fid "DFID"\n",
1576 PFID(&d->opd_pre_used_fid),
1577 PFID(&d->opd_pre_last_created_fid));
1580 * When sequence is used up, new one should be allocated in
1581 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1584 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1585 "next fid "DFID" last created fid "DFID"\n",
1586 PFID(&d->opd_pre_used_fid),
1587 PFID(&d->opd_pre_last_created_fid));
1588 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1589 if (fid_is_idif(pre_used_fid) &&
1590 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1591 pre_used_fid->f_seq++;
1593 d->opd_pre_used_fid.f_oid++;
1594 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1595 d->opd_pre_reserved--;
1597 * last_used_id must be changed along with getting new id otherwise
1598 * we might miscalculate gap causing object loss or leak
1600 osp_update_last_fid(d, fid);
1601 spin_unlock(&d->opd_pre_lock);
1604 * probably main thread suspended orphan cleanup till
1605 * all reservations are released, see comment in
1606 * osp_precreate_thread() just before orphan cleanup
1608 if (unlikely(d->opd_pre_reserved == 0 &&
1609 (d->opd_pre_recovering || d->opd_pre_status)))
1610 wake_up(&d->opd_pre_waitq);
1616 * Set size regular attribute on an object
1618 * When a striping is created late, it's possible that size is already
1619 * initialized on the file. Then the new striping should inherit size
1620 * from the file. The function sets size on the object using the regular
1621 * protocol (OST_PUNCH).
1622 * XXX: should be re-implemented using OUT ?
1624 * \param[in] env LU environment provided by the caller
1625 * \param[in] dt object
1626 * \param[in] size size to set.
1628 * \retval 0 on success
1629 * \retval negative negated errno on error
1631 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1634 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1635 struct ptlrpc_request *req = NULL;
1636 struct obd_import *imp;
1637 struct ost_body *body;
1638 struct obdo *oa = NULL;
1643 imp = d->opd_obd->u.cli.cl_import;
1646 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1650 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1652 ptlrpc_request_free(req);
1657 * XXX: decide how do we do here with resend
1658 * if we don't resend, then client may see wrong file size
1659 * if we do resend, then MDS thread can get stuck for quite long
1660 * and if we don't resend, then client will also get -EAGAIN !!
1661 * (see LU-7975 and sanity/test_27F use cases)
1662 * but let's decide not to resend/delay this truncate request to OST
1663 * and allow Client to decide to resend, in a less agressive way from
1664 * after_reply(), by returning -EINPROGRESS instead of
1665 * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
1666 * end of this routine
1668 req->rq_no_resend = req->rq_no_delay = 1;
1670 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1671 ptlrpc_at_set_req_timeout(req);
1675 GOTO(out, rc = -ENOMEM);
1677 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1680 oa->o_blocks = OBD_OBJECT_EOF;
1681 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1682 OBD_MD_FLID | OBD_MD_FLGROUP;
1684 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1686 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1688 /* XXX: capa support? */
1689 /* osc_pack_capa(req, body, capa); */
1691 ptlrpc_request_set_replen(req);
1693 rc = ptlrpc_queue_wait(req);
1695 /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
1696 * since we have decided not to resend/delay, but this could
1697 * lead to wrong size to be seen at Client side and even process
1698 * trying to open to exit/fail if not itself handling -EAGAIN.
1699 * So it should be better to return -EINPROGRESS instead and
1700 * leave the decision to resend at Client side in after_reply()
1702 if (rc == -EAGAIN) {
1704 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1705 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1708 CERROR("can't punch object: %d\n", rc);
1712 ptlrpc_req_finished(req);
1719 * Initialize precreation functionality of OSP
1721 * Prepares all the internal structures and starts the precreate thread
1723 * \param[in] d OSP device
1725 * \retval 0 on success
1726 * \retval negative negated errno on error
1728 int osp_init_precreate(struct osp_device *d)
1732 OBD_ALLOC_PTR(d->opd_pre);
1733 if (d->opd_pre == NULL)
1736 /* initially precreation isn't ready */
1737 init_waitqueue_head(&d->opd_pre_user_waitq);
1738 d->opd_pre_status = -EAGAIN;
1739 fid_zero(&d->opd_pre_used_fid);
1740 d->opd_pre_used_fid.f_oid = 1;
1741 fid_zero(&d->opd_pre_last_created_fid);
1742 d->opd_pre_last_created_fid.f_oid = 1;
1744 d->opd_pre_reserved = 0;
1745 d->opd_got_disconnected = 1;
1746 d->opd_pre_create_slow = 0;
1747 d->opd_pre_create_count = OST_MIN_PRECREATE;
1748 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1749 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1750 d->opd_reserved_mb_high = 0;
1751 d->opd_reserved_mb_low = 0;
1757 * Finish precreate functionality of OSP
1760 * Asks all the activity (the thread, update timer) to stop, then
1761 * wait till that is done.
1763 * \param[in] d OSP device
1765 void osp_precreate_fini(struct osp_device *d)
1769 if (d->opd_pre == NULL)
1772 OBD_FREE_PTR(d->opd_pre);
1778 int osp_init_statfs(struct osp_device *d)
1780 struct task_struct *task;
1781 struct opt_args *args;
1782 DECLARE_COMPLETION_ONSTACK(started);
1787 spin_lock_init(&d->opd_pre_lock);
1788 init_waitqueue_head(&d->opd_pre_waitq);
1791 * Initialize statfs-related things
1793 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1794 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1795 1000 * NSEC_PER_SEC);
1796 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1798 ktime_to_ns(d->opd_statfs_fresh_till));
1799 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1800 (unsigned long)d, 0);
1802 if (d->opd_storage->dd_rdonly)
1805 OBD_ALLOC_PTR(args);
1809 args->opta_started = &started;
1810 rc = lu_env_init(&args->opta_env,
1811 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1813 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1820 * start thread handling precreation and statfs updates
1822 task = kthread_create(osp_precreate_thread, args,
1823 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1825 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1826 lu_env_fini(&args->opta_env);
1828 RETURN(PTR_ERR(task));
1830 d->opd_pre_task = task;
1831 wake_up_process(task);
1832 wait_for_completion(&started);
1837 void osp_statfs_fini(struct osp_device *d)
1839 struct task_struct *task = d->opd_pre_task;
1842 del_timer(&d->opd_statfs_timer);
1844 d->opd_pre_task = NULL;