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_force_creation) && (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;
637 /* Delorphan happens only with a first MDT-OST connect. resend/replay
638 * handles objects creation on reconnects, no need to do delorhpan
642 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
644 ptlrpc_request_free(req);
648 spin_lock(&d->opd_pre_lock);
649 if (d->opd_force_creation)
650 d->opd_pre_create_count = OST_MIN_PRECREATE;
651 else if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
652 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
653 grow = d->opd_pre_create_count;
654 spin_unlock(&d->opd_pre_lock);
656 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
659 *fid = d->opd_pre_last_created_fid;
660 rc = osp_precreate_fids(env, d, fid, &grow);
662 /* Current seq has been used up*/
663 GOTO(out_req, rc = -ENOSPC);
665 if (!osp_is_fid_client(d)) {
666 /* Non-FID client will always send seq 0 because of
668 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
672 fid_to_ostid(fid, &body->oa.o_oi);
673 body->oa.o_valid = OBD_MD_FLGROUP;
675 ptlrpc_request_set_replen(req);
677 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
680 rc = ptlrpc_queue_wait(req);
682 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
685 /* have osp_precreate_reserve() to wait for repeat */
690 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
692 GOTO(out_req, rc = -EPROTO);
694 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
697 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
698 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
699 ": rc = %d\n", d->opd_obd->obd_name,
700 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
701 GOTO(out_req, rc = -ESTALE);
704 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
706 spin_lock(&d->opd_pre_lock);
708 /* the OST has not managed to create all the
709 * objects we asked for */
710 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
711 d->opd_pre_create_slow = 1;
713 /* the OST is able to keep up with the work,
714 * we could consider increasing create_count
715 * next time if needed */
716 d->opd_pre_create_slow = 0;
719 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
720 fid_to_ostid(fid, &body->oa.o_oi);
722 d->opd_pre_last_created_fid = *fid;
723 d->opd_force_creation = false;
724 spin_unlock(&d->opd_pre_lock);
726 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
727 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
728 PFID(&d->opd_pre_last_created_fid));
730 /* now we can wakeup all users awaiting for objects */
731 osp_pre_update_status(d, rc);
733 /* pause to let osp_precreate_reserve to go first */
734 CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_PRECREATE_PAUSE, 2);
736 ptlrpc_req_finished(req);
741 * Get last precreated object from target (OST)
743 * Sends synchronous RPC to the target (OST) to learn the last precreated
744 * object. This later is used to remove all unused objects (cleanup orphan
745 * procedure). Also, the next object after one we got will be used as a
746 * starting point for the new precreates.
748 * \param[in] env LU environment provided by the caller
749 * \param[in] d OSP device
750 * \param[in] update update or not update last used fid
752 * \retval 0 on success
753 * \retval negative negated errno on error
755 static int osp_get_lastfid_from_ost(const struct lu_env *env,
756 struct osp_device *d, bool update)
758 struct ptlrpc_request *req = NULL;
759 struct obd_import *imp;
760 struct lu_fid *last_fid;
765 imp = d->opd_obd->u.cli.cl_import;
768 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
772 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
773 sizeof(KEY_LAST_FID));
775 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
777 ptlrpc_request_free(req);
781 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
782 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
784 req->rq_no_delay = req->rq_no_resend = 1;
785 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
786 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
788 ptlrpc_request_set_replen(req);
790 rc = ptlrpc_queue_wait(req);
792 /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld),
793 * let sysadm sort this * out.
796 ptlrpc_set_import_active(imp, 0);
800 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
801 if (last_fid == NULL) {
802 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
803 GOTO(out, rc = -EPROTO);
806 if (!fid_is_sane(last_fid)) {
807 CERROR("%s: Got insane last_fid "DFID"\n",
808 d->opd_obd->obd_name, PFID(last_fid));
809 GOTO(out, rc = -EPROTO);
812 /* Only update the last used fid, if the OST has objects for
813 * this sequence, i.e. fid_oid > 0 */
814 if (fid_oid(last_fid) > 0 && update)
815 d->opd_last_used_fid = *last_fid;
817 if (fid_oid(last_fid) == 0 &&
818 fid_seq(last_fid) == fid_seq(&d->opd_last_used_fid)) {
819 /* reformatted OST, it requires creation request
820 * to recreate objects
822 d->opd_force_creation = true;
824 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
828 ptlrpc_req_finished(req);
833 * Cleanup orphans on OST
835 * This function is called in a contex of a dedicated thread handling
836 * all the precreation suff. The function waits till local recovery
837 * is complete, then identify all the unreferenced objects (orphans)
838 * using the highest ID referenced by a local and the highest object
839 * precreated by the target. The found range is a subject to removal
840 * using specially flagged RPC. During this process OSP is marked
841 * unavailable for new objects.
843 * \param[in] env LU environment provided by the caller
844 * \param[in] d OSP device
846 * \retval 0 on success
847 * \retval negative negated errno on error
849 static int osp_precreate_cleanup_orphans(struct lu_env *env,
850 struct osp_device *d)
852 struct osp_thread_info *osi = osp_env_info(env);
853 struct lu_fid *last_fid = &osi->osi_fid;
854 struct ptlrpc_request *req = NULL;
855 struct obd_import *imp;
856 struct ost_body *body;
857 int update_status = 0;
864 * Do cleanup orphans only with a first connection, after that
865 * all precreate requests uses resend/replay flags to support OST
866 * failover/reconnect.
868 if (d->opd_cleanup_orphans_done) {
869 rc = osp_get_lastfid_from_ost(env, d, false);
873 * wait for local recovery to finish, so we can cleanup orphans
874 * orphans are all objects since "last used" (assigned), but
875 * there might be objects reserved and in some cases they won't
876 * be used. we can't cleanup them till we're sure they won't be
877 * used. also can't we allow new reservations because they may
878 * end up getting orphans being cleaned up below. so we block
879 * new reservations and wait till all reserved objects either
882 spin_lock(&d->opd_pre_lock);
883 d->opd_pre_recovering = 1;
884 spin_unlock(&d->opd_pre_lock);
886 * The locking above makes sure the opd_pre_reserved check below will
887 * catch all osp_precreate_reserve() calls who find
888 * "!opd_pre_recovering".
890 wait_event_idle(d->opd_pre_waitq,
891 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
892 !d->opd_pre_task || d->opd_got_disconnected);
893 if (!d->opd_pre_task || d->opd_got_disconnected)
894 GOTO(out, rc = -EAGAIN);
896 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
897 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
899 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val);
901 *last_fid = d->opd_last_used_fid;
902 /* The OSP should already get the valid seq now */
903 LASSERT(!fid_is_zero(last_fid));
904 if (fid_oid(&d->opd_last_used_fid) < 2 ||
905 OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
906 /* lastfid looks strange... ask OST */
907 rc = osp_get_lastfid_from_ost(env, d, true);
912 imp = d->opd_obd->u.cli.cl_import;
915 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
917 GOTO(out, rc = -ENOMEM);
919 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
921 ptlrpc_request_free(req);
926 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
928 GOTO(out, rc = -EPROTO);
930 body->oa.o_flags = OBD_FL_DELORPHAN;
931 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
933 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
935 ptlrpc_request_set_replen(req);
937 /* Don't resend the delorphan req */
938 req->rq_no_resend = req->rq_no_delay = 1;
940 rc = ptlrpc_queue_wait(req);
946 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
948 GOTO(out, rc = -EPROTO);
951 * OST provides us with id new pool starts from in body->oa.o_id
953 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
955 spin_lock(&d->opd_pre_lock);
956 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
958 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
959 d->opd_pre_last_created_fid = d->opd_last_used_fid;
961 d->opd_pre_create_count = OST_MIN_PRECREATE;
962 d->opd_pre_last_created_fid = *last_fid;
965 * This empties the pre-creation pool and effectively blocks any new
968 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
969 LUSTRE_DATA_SEQ_MAX_WIDTH);
970 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
971 d->opd_pre_create_slow = 0;
972 spin_unlock(&d->opd_pre_lock);
974 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
975 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
976 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
979 ptlrpc_req_finished(req);
983 * If rc is zero, the pre-creation window should have been emptied.
984 * Since waking up the herd would be useless without pre-created
985 * objects, we defer the signal to osp_precreate_send() in that case.
989 CERROR("%s: cannot cleanup orphans: rc = %d\n",
990 d->opd_obd->obd_name, rc);
991 /* we can't proceed from here, OST seem to
992 * be in a bad shape, better to wait for
993 * a new instance of the server and repeat
994 * from the beginning. notify possible waiters
995 * this OSP isn't quite functional yet */
996 osp_pre_update_status(d, rc);
998 wake_up_all(&d->opd_pre_user_waitq);
1001 spin_lock(&d->opd_pre_lock);
1002 d->opd_pre_recovering = 0;
1003 spin_unlock(&d->opd_pre_lock);
1004 d->opd_cleanup_orphans_done = true;
1011 * Update precreate status using statfs data
1013 * The function decides whether this OSP should be used for new objects.
1014 * IOW, whether this OST is used up or has some free space. Cached statfs
1015 * data is used to make this decision. If the latest result of statfs
1016 * request (rc argument) is not success, then just mark OSP unavailable
1019 * The new statfs data is passed in \a msfs and needs to be stored into
1020 * opd_statfs, but only after the various flags in os_state are set, so
1021 * that the new statfs data is not visible without appropriate flags set.
1022 * As such, there is no need to clear the flags here, since this is called
1023 * with new statfs data, and they should not be cleared if sent from OST.
1025 * Add a bit of hysteresis so this flag isn't continually flapping, and
1026 * ensure that new files don't get extremely fragmented due to only a
1027 * small amount of available space in the filesystem. We want to set
1028 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1029 * reserved size free, and still copy them from the old state when there
1030 * is less than 2*reserved size free space or inodes.
1032 * \param[in] d OSP device
1033 * \param[in] msfs statfs data
1035 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1037 u32 old_state = d->opd_statfs.os_state;
1040 /* statfs structure not initialized yet */
1041 if (unlikely(!msfs->os_type))
1044 /* if the low and high watermarks have not been initialized yet */
1045 if (unlikely(d->opd_reserved_mb_high == 0 &&
1046 d->opd_reserved_mb_low == 0)) {
1047 /* Use ~0.1% by default to disable object allocation,
1048 * and ~0.2% to enable, size in MB, set both watermark
1050 spin_lock(&d->opd_pre_lock);
1051 if (d->opd_reserved_mb_high == 0 &&
1052 d->opd_reserved_mb_low == 0) {
1053 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1054 msfs->os_blocks) >> 20;
1055 if (d->opd_reserved_mb_low < 1)
1056 d->opd_reserved_mb_low = 1;
1057 d->opd_reserved_mb_high =
1058 (d->opd_reserved_mb_low << 1) + 1;
1060 spin_unlock(&d->opd_pre_lock);
1063 if (unlikely(d->opd_reserved_ino_high == 0 &&
1064 d->opd_reserved_ino_low == 0)) {
1065 /* Use ~0.1% by default to disallow distributed transactions,
1066 * and ~0.2% to allow, set both watermark
1068 spin_lock(&d->opd_pre_lock);
1069 if (d->opd_reserved_ino_high == 0 &&
1070 d->opd_reserved_ino_low == 0) {
1071 d->opd_reserved_ino_low = msfs->os_ffree >> 20;
1072 if (d->opd_reserved_ino_low < 32)
1073 d->opd_reserved_ino_low = 32;
1074 d->opd_reserved_ino_high =
1075 (d->opd_reserved_ino_low << 1) + 1;
1077 spin_unlock(&d->opd_pre_lock);
1080 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1081 if (msfs->os_ffree < d->opd_reserved_ino_low)
1082 msfs->os_state |= OS_STATFS_ENOINO;
1083 else if (msfs->os_ffree <= d->opd_reserved_ino_high)
1084 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1085 /* else don't clear flags in new msfs->os_state sent from OST */
1087 if (available_mb < d->opd_reserved_mb_low)
1088 msfs->os_state |= OS_STATFS_ENOSPC;
1089 else if (available_mb <= d->opd_reserved_mb_high)
1090 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1091 /* else don't clear flags in new msfs->os_state sent from OST */
1094 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1095 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1096 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1097 msfs->os_files, msfs->os_ffree, msfs->os_state,
1100 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1101 d->opd_pre_status = -ENOSPC;
1102 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1103 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1104 d->opd_obd->obd_name, msfs->os_state,
1106 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1107 atomic_read(&d->opd_sync_changes),
1108 atomic_read(&d->opd_sync_rpcs_in_progress));
1109 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1110 d->opd_pre_status = 0;
1111 spin_lock(&d->opd_pre_lock);
1112 d->opd_pre_create_slow = 0;
1113 d->opd_pre_create_count = OST_MIN_PRECREATE;
1114 spin_unlock(&d->opd_pre_lock);
1115 wake_up(&d->opd_pre_waitq);
1118 "%s: available: state=%x: rc = %d\n",
1119 d->opd_obd->obd_name, msfs->os_state,
1122 /* we only get here if rc == 0 in the caller */
1123 d->opd_pre_status = 0;
1126 /* Object precreation skipped on OST if manually disabled */
1127 if (d->opd_pre_max_create_count == 0)
1128 msfs->os_state |= OS_STATFS_NOPRECREATE;
1129 /* else don't clear flags in new msfs->os_state sent from OST */
1131 /* copy only new statfs state to make it visible to MDS threads */
1132 if (&d->opd_statfs != msfs)
1133 d->opd_statfs = *msfs;
1137 * Initialize FID for precreation
1139 * For a just created new target, a new sequence should be taken.
1140 * The function checks there is no IDIF in use (if the target was
1141 * added with the older version of Lustre), then requests a new
1142 * sequence from FLDB using the regular protocol. Then this new
1143 * sequence is stored on a persisten storage synchronously to prevent
1144 * possible object leakage (for the detail see the description for
1145 * osp_precreate_rollover_new_seq()).
1147 * \param[in] osp OSP device
1149 * \retval 0 on success
1150 * \retval negative negated errno on error
1152 int osp_init_pre_fid(struct osp_device *osp)
1155 struct osp_thread_info *osi;
1156 struct lu_client_seq *cli_seq;
1157 struct lu_fid *last_fid;
1161 LASSERT(osp->opd_pre != NULL);
1163 /* Let's check if the current last_seq/fid is valid,
1164 * otherwise request new sequence from the controller */
1165 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1166 /* Non-MDT0 can only use normal sequence for
1168 if (fid_is_norm(&osp->opd_last_used_fid))
1171 /* Initially MDT0 will start with IDIF, after
1172 * that it will request new sequence from the
1174 if (fid_is_idif(&osp->opd_last_used_fid) ||
1175 fid_is_norm(&osp->opd_last_used_fid))
1179 if (!fid_is_zero(&osp->opd_last_used_fid))
1180 CWARN("%s: invalid last used fid "DFID
1181 ", try to get new sequence.\n",
1182 osp->opd_obd->obd_name,
1183 PFID(&osp->opd_last_used_fid));
1185 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1187 CERROR("%s: init env error: rc = %d\n",
1188 osp->opd_obd->obd_name, rc);
1192 osi = osp_env_info(&env);
1193 last_fid = &osi->osi_fid;
1195 /* For a freshed fs, it will allocate a new sequence first */
1196 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1197 cli_seq = osp->opd_obd->u.cli.cl_seq;
1198 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1200 CERROR("%s: alloc fid error: rc = %d\n",
1201 osp->opd_obd->obd_name, rc);
1205 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1207 last_fid->f_oid = 1;
1208 last_fid->f_ver = 0;
1210 spin_lock(&osp->opd_pre_lock);
1211 osp->opd_last_used_fid = *last_fid;
1212 osp->opd_pre_used_fid = *last_fid;
1213 osp->opd_pre_last_created_fid = *last_fid;
1214 spin_unlock(&osp->opd_pre_lock);
1215 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1217 CERROR("%s: write fid error: rc = %d\n",
1218 osp->opd_obd->obd_name, rc);
1227 struct osp_device *opta_dev;
1228 struct lu_env opta_env;
1229 struct completion *opta_started;
1232 * The core of precreate functionality
1234 * The function implements the main precreation loop. Basically it
1235 * involves connecting to the target, precerate FID initialization,
1236 * identifying and removing orphans, then serving precreation. As
1237 * part of the latter, the thread is responsible for statfs data
1238 * updates. The precreation is mostly driven by another threads
1239 * asking for new OST objects - those askers wake the thread when
1240 * the number of precreated objects reach low watermark.
1241 * After a disconnect, the sequence above repeats. This is keep going
1242 * until the thread is requested to stop.
1244 * \param[in] _arg private data the thread (OSP device to handle)
1246 * \retval 0 on success
1247 * \retval negative negated errno on error
1249 static int osp_precreate_thread(void *_args)
1251 struct opt_args *args = _args;
1252 struct osp_device *d = args->opta_dev;
1253 struct lu_env *env = &args->opta_env;
1258 complete(args->opta_started);
1259 while (!kthread_should_stop()) {
1261 * need to be connected to OST
1263 while (!kthread_should_stop()) {
1264 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1265 d->opd_imp_connected &&
1266 !d->opd_got_disconnected)
1268 wait_event_idle(d->opd_pre_waitq,
1269 kthread_should_stop() ||
1270 d->opd_new_connection);
1272 if (!d->opd_new_connection)
1275 OBD_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
1277 d->opd_new_connection = 0;
1278 d->opd_got_disconnected = 0;
1282 if (kthread_should_stop())
1286 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1287 /* Sigh, fid client is not ready yet */
1288 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1291 /* Init fid for osp_precreate if necessary */
1292 rc = osp_init_pre_fid(d);
1294 class_export_put(d->opd_exp);
1295 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1296 CERROR("%s: init pre fid error: rc = %d\n",
1297 d->opd_obd->obd_name, rc);
1302 if (osp_statfs_update(env, d)) {
1303 if (wait_event_idle_timeout(d->opd_pre_waitq,
1304 kthread_should_stop(),
1305 cfs_time_seconds(5)) == 0)
1306 l_wait_event_abortable(
1308 kthread_should_stop());
1314 * Clean up orphans or recreate missing objects.
1316 rc = osp_precreate_cleanup_orphans(env, d);
1318 schedule_timeout_interruptible(cfs_time_seconds(1));
1324 * connected, can handle precreates now
1326 while (!kthread_should_stop()) {
1327 wait_event_idle(d->opd_pre_waitq,
1328 kthread_should_stop() ||
1329 osp_precreate_near_empty(env, d) ||
1330 osp_statfs_need_update(d) ||
1331 d->opd_got_disconnected);
1333 if (kthread_should_stop())
1336 /* something happened to the connection
1337 * have to start from the beginning */
1338 if (d->opd_got_disconnected)
1341 if (osp_statfs_need_update(d))
1342 if (osp_statfs_update(env, d))
1345 if (d->opd_pre == NULL)
1348 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
1349 d->opd_pre_recovering = 1;
1353 /* To avoid handling different seq in precreate/orphan
1354 * cleanup, it will hold precreate until current seq is
1356 if (unlikely(osp_precreate_end_seq(env, d) &&
1357 !osp_create_end_seq(env, d)))
1360 if (unlikely(osp_precreate_end_seq(env, d) &&
1361 osp_create_end_seq(env, d))) {
1362 LCONSOLE_INFO("%s:%#llx is used up."
1363 " Update to new seq\n",
1364 d->opd_obd->obd_name,
1365 fid_seq(&d->opd_pre_last_created_fid));
1366 rc = osp_precreate_rollover_new_seq(env, d);
1371 if (osp_precreate_near_empty(env, d)) {
1372 rc = osp_precreate_send(env, d);
1373 /* osp_precreate_send() sets opd_pre_status
1374 * in case of error, that prevent the using of
1376 if (rc < 0 && rc != -ENOSPC &&
1377 rc != -ETIMEDOUT && rc != -ENOTCONN)
1378 CERROR("%s: cannot precreate objects:"
1380 d->opd_obd->obd_name, rc);
1392 * Check when to stop to wait for precreate objects.
1394 * The caller wanting a new OST object can't wait undefinitely. The
1395 * function checks for few conditions including available new OST
1396 * objects, disconnected OST, lack of space with no pending destroys,
1397 * etc. IOW, it checks whether the current OSP state is good to keep
1398 * waiting or it's better to give up.
1400 * \param[in] env LU environment provided by the caller
1401 * \param[in] d OSP device
1403 * \retval 0 - keep waiting, 1 - no luck
1405 static int osp_precreate_ready_condition(const struct lu_env *env,
1406 struct osp_device *d)
1408 /* Bail out I/O fails to OST */
1409 if (d->opd_pre_status != 0 &&
1410 d->opd_pre_status != -EAGAIN &&
1411 d->opd_pre_status != -ENODEV &&
1412 d->opd_pre_status != -ENOTCONN &&
1413 d->opd_pre_status != -ENOSPC) {
1415 if (d->opd_pre_status != -EIO)
1416 CERROR("%s: precreate failed opd_pre_status %d\n",
1417 d->opd_obd->obd_name, d->opd_pre_status);
1421 if (d->opd_pre_recovering)
1424 /* ready if got enough precreated objects */
1425 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1426 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1429 /* ready if OST reported no space and no destroys in progress */
1430 if (atomic_read(&d->opd_sync_changes) +
1431 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1432 d->opd_pre_status == -ENOSPC)
1439 * Reserve object in precreate pool
1441 * When the caller wants to create a new object on this target (target
1442 * represented by the given OSP), it should declare this intention using
1443 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1444 * to reserve an object in the existing precreated pool or wait up to
1445 * obd_timeout for the available object to appear in the pool (a dedicated
1446 * thread will be doing real precreation in background). The object can be
1447 * consumed later with osp_precreate_get_fid() or be released with call to
1448 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1449 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1450 * If the space on the target is short and there is a pending object destroy,
1451 * then the function forces local commit to speedup space release (see
1452 * osp_sync.c for the details).
1454 * \param[in] env LU environment provided by the caller
1455 * \param[in] d OSP device
1457 * \retval 0 on success
1458 * \retval -ENOSPC when no space on OST
1459 * \retval -EAGAIN try later, slow precreation in progress
1460 * \retval -EIO when no access to OST
1462 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
1465 time64_t expire = ktime_get_seconds() + obd_timeout;
1466 int precreated, rc, synced = 0;
1470 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1471 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1472 PFID(&d->opd_pre_used_fid));
1474 /* opd_pre_max_create_count 0 to not use specified OST. */
1475 if (d->opd_pre_max_create_count == 0)
1480 * - preallocation is done
1481 * - no free space expected soon
1482 * - can't connect to OST for too long (obd_timeout)
1483 * - OST can allocate fid sequence.
1485 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1486 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1489 * increase number of precreations
1491 precreated = osp_objs_precreated(env, d);
1492 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1493 d->opd_pre_create_slow == 0 &&
1494 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1495 spin_lock(&d->opd_pre_lock);
1496 d->opd_pre_create_slow = 1;
1497 d->opd_pre_create_count *= 2;
1498 spin_unlock(&d->opd_pre_lock);
1501 spin_lock(&d->opd_pre_lock);
1502 precreated = osp_objs_precreated(env, d);
1503 if (precreated > d->opd_pre_reserved &&
1504 !d->opd_pre_recovering) {
1505 d->opd_pre_reserved++;
1506 spin_unlock(&d->opd_pre_lock);
1509 /* XXX: don't wake up if precreation is in progress */
1510 if (osp_precreate_near_empty_nolock(env, d) &&
1511 !osp_precreate_end_seq_nolock(env, d))
1512 wake_up(&d->opd_pre_waitq);
1516 spin_unlock(&d->opd_pre_lock);
1519 * all precreated objects have been used and no-space
1520 * status leave us no chance to succeed very soon
1521 * but if there is destroy in progress, then we should
1522 * wait till that is done - some space might be released
1524 if (unlikely(rc == -ENOSPC)) {
1525 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1526 /* force local commit to release space */
1527 dt_commit_async(env, d->opd_storage);
1528 osp_sync_check_for_work(d);
1531 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1532 /* just wait till destroys are done
1533 * see wait_event_idle_timeout() below
1536 if (atomic_read(&d->opd_sync_changes) +
1537 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1538 /* no hope for free space */
1543 /* XXX: don't wake up if precreation is in progress */
1544 wake_up(&d->opd_pre_waitq);
1546 if (ktime_get_seconds() >= expire) {
1552 LASSERT(d->opd_pre);
1557 CDEBUG(D_INFO, "%s: Sleeping on objects\n",
1558 d->opd_obd->obd_name);
1559 if (wait_event_idle_timeout(
1560 d->opd_pre_user_waitq,
1561 osp_precreate_ready_condition(env, d),
1562 cfs_time_seconds(obd_timeout)) == 0) {
1564 "%s: slow creates, last="DFID", next="DFID", "
1565 "reserved=%llu, sync_changes=%u, "
1566 "sync_rpcs_in_progress=%d, status=%d\n",
1567 d->opd_obd->obd_name,
1568 PFID(&d->opd_pre_last_created_fid),
1569 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1570 atomic_read(&d->opd_sync_changes),
1571 atomic_read(&d->opd_sync_rpcs_in_progress),
1574 CDEBUG(D_INFO, "%s: Waked up, status=%d\n",
1575 d->opd_obd->obd_name, d->opd_pre_status);
1583 * Get a FID from precreation pool
1585 * The function is a companion for osp_precreate_reserve() - it assigns
1586 * a specific FID from the precreate. The function should be called only
1587 * if the call to osp_precreate_reserve() was successful. The function
1588 * updates a local storage to remember the highest object ID referenced
1589 * by the node in the given sequence.
1591 * A very importan details: this is supposed to be called once the
1592 * transaction is started, so on-disk update will be atomic with the
1593 * data (like LOVEA) refering this object. Then the object won't be leaked:
1594 * either it's referenced by the committed transaction or it's a subject
1595 * to the orphan cleanup procedure.
1597 * \param[in] env LU environment provided by the caller
1598 * \param[in] d OSP device
1599 * \param[out] fid generated FID
1601 * \retval 0 on success
1602 * \retval negative negated errno on error
1604 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1607 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1608 /* grab next id from the pool */
1609 spin_lock(&d->opd_pre_lock);
1611 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1612 &d->opd_pre_last_created_fid) < 0,
1613 "next fid "DFID" last created fid "DFID"\n",
1614 PFID(&d->opd_pre_used_fid),
1615 PFID(&d->opd_pre_last_created_fid));
1618 * When sequence is used up, new one should be allocated in
1619 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1622 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1623 "next fid "DFID" last created fid "DFID"\n",
1624 PFID(&d->opd_pre_used_fid),
1625 PFID(&d->opd_pre_last_created_fid));
1626 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1627 if (fid_is_idif(pre_used_fid) &&
1628 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1629 pre_used_fid->f_seq++;
1631 d->opd_pre_used_fid.f_oid++;
1632 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1633 d->opd_pre_reserved--;
1635 * last_used_id must be changed along with getting new id otherwise
1636 * we might miscalculate gap causing object loss or leak
1638 osp_update_last_fid(d, fid);
1639 spin_unlock(&d->opd_pre_lock);
1642 * probably main thread suspended orphan cleanup till
1643 * all reservations are released, see comment in
1644 * osp_precreate_thread() just before orphan cleanup
1646 if (unlikely(d->opd_pre_reserved == 0 &&
1647 (d->opd_pre_recovering || d->opd_pre_status)))
1648 wake_up(&d->opd_pre_waitq);
1654 * Set size regular attribute on an object
1656 * When a striping is created late, it's possible that size is already
1657 * initialized on the file. Then the new striping should inherit size
1658 * from the file. The function sets size on the object using the regular
1659 * protocol (OST_PUNCH).
1660 * XXX: should be re-implemented using OUT ?
1662 * \param[in] env LU environment provided by the caller
1663 * \param[in] dt object
1664 * \param[in] size size to set.
1666 * \retval 0 on success
1667 * \retval negative negated errno on error
1669 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1672 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1673 struct ptlrpc_request *req = NULL;
1674 struct obd_import *imp;
1675 struct ost_body *body;
1676 struct obdo *oa = NULL;
1681 imp = d->opd_obd->u.cli.cl_import;
1684 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1688 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1690 ptlrpc_request_free(req);
1695 * XXX: decide how do we do here with resend
1696 * if we don't resend, then client may see wrong file size
1697 * if we do resend, then MDS thread can get stuck for quite long
1698 * and if we don't resend, then client will also get -EAGAIN !!
1699 * (see LU-7975 and sanity/test_27F use cases)
1700 * but let's decide not to resend/delay this truncate request to OST
1701 * and allow Client to decide to resend, in a less agressive way from
1702 * after_reply(), by returning -EINPROGRESS instead of
1703 * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
1704 * end of this routine
1706 req->rq_no_resend = req->rq_no_delay = 1;
1708 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1709 ptlrpc_at_set_req_timeout(req);
1713 GOTO(out, rc = -ENOMEM);
1715 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1718 oa->o_blocks = OBD_OBJECT_EOF;
1719 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1720 OBD_MD_FLID | OBD_MD_FLGROUP;
1722 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1724 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1726 /* XXX: capa support? */
1727 /* osc_pack_capa(req, body, capa); */
1729 ptlrpc_request_set_replen(req);
1731 rc = ptlrpc_queue_wait(req);
1733 /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
1734 * since we have decided not to resend/delay, but this could
1735 * lead to wrong size to be seen at Client side and even process
1736 * trying to open to exit/fail if not itself handling -EAGAIN.
1737 * So it should be better to return -EINPROGRESS instead and
1738 * leave the decision to resend at Client side in after_reply()
1740 if (rc == -EAGAIN) {
1742 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1743 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1746 CERROR("can't punch object: %d\n", rc);
1750 ptlrpc_req_finished(req);
1757 * Initialize precreation functionality of OSP
1759 * Prepares all the internal structures and starts the precreate thread
1761 * \param[in] d OSP device
1763 * \retval 0 on success
1764 * \retval negative negated errno on error
1766 int osp_init_precreate(struct osp_device *d)
1770 OBD_ALLOC_PTR(d->opd_pre);
1771 if (d->opd_pre == NULL)
1774 /* initially precreation isn't ready */
1775 init_waitqueue_head(&d->opd_pre_user_waitq);
1776 d->opd_pre_status = -EAGAIN;
1777 fid_zero(&d->opd_pre_used_fid);
1778 d->opd_pre_used_fid.f_oid = 1;
1779 fid_zero(&d->opd_pre_last_created_fid);
1780 d->opd_pre_last_created_fid.f_oid = 1;
1782 d->opd_pre_reserved = 0;
1783 d->opd_got_disconnected = 1;
1784 d->opd_pre_create_slow = 0;
1785 d->opd_pre_create_count = OST_MIN_PRECREATE;
1786 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1787 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1788 d->opd_reserved_mb_high = 0;
1789 d->opd_reserved_mb_low = 0;
1790 d->opd_cleanup_orphans_done = false;
1791 d->opd_force_creation = false;
1797 * Finish precreate functionality of OSP
1800 * Asks all the activity (the thread, update timer) to stop, then
1801 * wait till that is done.
1803 * \param[in] d OSP device
1805 void osp_precreate_fini(struct osp_device *d)
1809 if (d->opd_pre == NULL)
1812 OBD_FREE_PTR(d->opd_pre);
1818 int osp_init_statfs(struct osp_device *d)
1820 struct task_struct *task;
1821 struct opt_args *args;
1822 DECLARE_COMPLETION_ONSTACK(started);
1827 spin_lock_init(&d->opd_pre_lock);
1828 init_waitqueue_head(&d->opd_pre_waitq);
1831 * Initialize statfs-related things
1833 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1834 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1835 1000 * NSEC_PER_SEC);
1836 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1838 ktime_to_ns(d->opd_statfs_fresh_till));
1839 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1840 (unsigned long)d, 0);
1842 if (d->opd_storage->dd_rdonly)
1845 OBD_ALLOC_PTR(args);
1849 args->opta_started = &started;
1850 rc = lu_env_init(&args->opta_env,
1851 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1853 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1860 * start thread handling precreation and statfs updates
1862 task = kthread_create(osp_precreate_thread, args,
1863 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1865 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1866 lu_env_fini(&args->opta_env);
1868 RETURN(PTR_ERR(task));
1870 d->opd_pre_task = task;
1871 wake_up_process(task);
1872 wait_for_completion(&started);
1877 void osp_statfs_fini(struct osp_device *d)
1879 struct task_struct *task = d->opd_pre_task;
1882 del_timer(&d->opd_statfs_timer);
1884 d->opd_pre_task = NULL;