4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/osp/osp_precreate.c
33 * Lustre OST Proxy Device
35 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
36 * Author: Mikhail Pershin <mike.pershin@intel.com>
37 * Author: Di Wang <di.wang@intel.com>
40 #define DEBUG_SUBSYSTEM S_MDS
42 #include <linux/kthread.h>
44 #include <lustre_obdo.h>
46 #include "osp_internal.h"
49 * there are two specific states to take care about:
51 * = import is disconnected =
53 * = import is inactive =
54 * in this case osp_declare_create() returns an error
59 * Check whether statfs data is expired
61 * OSP device caches statfs data for the target, the function checks
62 * whether the data is expired or not.
64 * \param[in] d OSP device
66 * \retval 0 - not expired, 1 - expired
68 static inline int osp_statfs_need_update(struct osp_device *d)
70 return !ktime_before(ktime_get(), d->opd_statfs_fresh_till);
74 * OSP tries to maintain pool of available objects so that calls to create
75 * objects don't block most of time
77 * each time OSP gets connected to OST, we should start from precreation cleanup
79 static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data)
81 struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
84 /* invalidate statfs data so osp_precreate_thread() can refresh */
85 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
87 wake_up(&d->opd_pre_waitq);
90 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs);
93 * The function updates current precreation status if broken, and
94 * updates that cached statfs state if functional, then wakes up waiters.
95 * We don't clear opd_pre_status directly here, but rather leave this
96 * to osp_pre_update_msfs() to do if everything is OK so that we don't
97 * have a race to clear opd_pre_status and then set it to -ENOSPC again.
99 * \param[in] d OSP device
100 * \param[in] msfs statfs data
101 * \param[in] rc new precreate status for device \a d
103 static void osp_pre_update_status_msfs(struct osp_device *d,
104 struct obd_statfs *msfs, int rc)
106 CDEBUG(D_INFO, "%s: Updating status = %d\n", d->opd_obd->obd_name, rc);
108 d->opd_pre_status = rc;
110 osp_pre_update_msfs(d, msfs);
112 wake_up_all(&d->opd_pre_user_waitq);
115 /* Pass in the old statfs data in case the limits have changed */
116 void osp_pre_update_status(struct osp_device *d, int rc)
118 osp_pre_update_status_msfs(d, &d->opd_statfs, rc);
123 * RPC interpret callback for OST_STATFS RPC
125 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
126 * replied by the target. It's used to maintain statfs cache for the target.
127 * The function fills data from the reply if successful and schedules another
130 * \param[in] env LU environment provided by the caller
131 * \param[in] req RPC replied
132 * \param[in] aa callback data
133 * \param[in] rc RPC result
135 * \retval 0 on success
136 * \retval negative negated errno on error
138 static int osp_statfs_interpret(const struct lu_env *env,
139 struct ptlrpc_request *req, void *args, int rc)
141 union ptlrpc_async_args *aa = args;
142 struct obd_import *imp = req->rq_import;
143 struct obd_statfs *msfs;
144 struct obd_statfs *sfs;
145 struct osp_device *d;
150 aa = ptlrpc_req_async_args(aa, req);
151 d = aa->pointer_arg[0];
157 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
159 GOTO(out, rc = -EPROTO);
162 osp_pre_update_status_msfs(d, msfs, 0);
164 osp_pre_update_msfs(d, msfs);
166 /* schedule next update */
167 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
168 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
169 mod_timer(&d->opd_statfs_timer,
170 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
171 d->opd_statfs_update_in_progress = 0;
173 sfs = &d->opd_statfs;
175 "%s (%p): %llu blocks, %llu free, %llu avail, %u bsize, %u reserved mb low, %u reserved mb high, %u reserved ino low, %u reserved ino high, %llu files, %llu free files %#x\n",
176 d->opd_obd->obd_name, d, sfs->os_blocks, sfs->os_bfree,
177 sfs->os_bavail, sfs->os_bsize, d->opd_reserved_mb_low,
178 d->opd_reserved_mb_high, d->opd_reserved_ino_low,
179 d->opd_reserved_ino_high, sfs->os_files, sfs->os_ffree,
184 /* couldn't update statfs, try again with a small delay */
185 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
186 d->opd_statfs_update_in_progress = 0;
187 if (d->opd_pre && d->opd_pre_task)
188 wake_up(&d->opd_pre_waitq);
190 if (req->rq_import_generation == imp->imp_generation)
191 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
192 d->opd_obd->obd_name, rc);
197 * Send OST_STATFS RPC
199 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
200 * Also disables scheduled updates as times OSP may need to refresh
201 * statfs data before expiration. The function doesn't block, instead
202 * an interpretation callback osp_statfs_interpret() is used.
204 * \param[in] env LU environment provided by the caller
205 * \param[in] d OSP device
207 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
209 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
210 struct ptlrpc_request *req;
211 struct obd_import *imp;
212 union ptlrpc_async_args *aa;
217 CDEBUG(D_CACHE, "going to update statfs\n");
219 imp = d->opd_obd->u.cli.cl_import;
222 req = ptlrpc_request_alloc(imp,
223 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
227 rc = ptlrpc_request_pack(req,
228 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
229 d->opd_pre ? OST_STATFS : MDS_STATFS);
231 ptlrpc_request_free(req);
234 ptlrpc_request_set_replen(req);
236 req->rq_request_portal = OST_CREATE_PORTAL;
237 ptlrpc_at_set_req_timeout(req);
239 req->rq_interpret_reply = osp_statfs_interpret;
240 aa = ptlrpc_req_async_args(aa, req);
241 aa->pointer_arg[0] = d;
244 * no updates till reply
246 timer_delete(&d->opd_statfs_timer);
247 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
248 d->opd_statfs_update_in_progress = 1;
250 ptlrpcd_add_req(req);
252 /* we still want to sync changes if no new changes are coming */
253 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
256 if (atomic_read(&d->opd_sync_changes)) {
259 th = dt_trans_create(env, d->opd_storage);
261 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
264 rc = dt_trans_start_local(env, d->opd_storage, th);
266 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
267 d->opd_obd->obd_name,
268 atomic_read(&d->opd_sync_changes));
269 osp_sync_add_commit_cb_1s(env, d, th);
271 dt_trans_stop(env, d->opd_storage, th);
279 * Schedule an immediate update for statfs data
281 * If cached statfs data claim no free space, but OSP has got a request to
282 * destroy an object (so release some space probably), then we may need to
283 * refresh cached statfs data sooner than planned. The function checks there
284 * is no statfs update going and schedules immediate update if so.
285 * XXX: there might be a case where removed object(s) do not add free space (empty
286 * object). If the number of such deletions is high, then we can start to update
287 * statfs too often causing a RPC storm. some throttling is needed...
289 * \param[in] d OSP device where statfs data needs to be refreshed
291 void osp_statfs_need_now(struct osp_device *d)
293 if (!d->opd_statfs_update_in_progress) {
295 * if current status is -ENOSPC (lack of free space on OST)
296 * then we should poll OST immediately once object destroy
299 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
300 timer_delete(&d->opd_statfs_timer);
301 wake_up(&d->opd_pre_waitq);
306 * Check pool of precreated objects is getting low.
308 * We should not wait till the pool of the precreated objects is too low,
309 * because then there will be a long period of OSP being unavailable for the
310 * new creations due to lengthy precreate RPC. Instead we ask for another
311 * precreation ahead and hopefully have it ready before the current pool is
312 * empty. Notice this function relies on external locking by opd_pre_lock.
314 * \param[in] d OSP device
316 * \retval 0 - current pool is good enough, 1 - time to precreate
318 static inline int osp_precreate_is_low_nolock(struct osp_device *d)
320 int available = osp_objs_precreated_nolock(d) - d->opd_pre_reserved;
321 int precreate_needed = d->opd_pre_create_count > 1024 ?
322 d->opd_pre_create_count / 4 : d->opd_pre_create_count / 2;
324 if (precreate_needed > 1024)
325 precreate_needed = 1024;
327 /* no new precreation until OST is healthy and has free space */
328 return ((d->opd_pre_create_count - available > precreate_needed ||
329 d->opd_force_creation) && (d->opd_pre_status == 0));
333 * Check pool of precreated objects
335 * This is protected version of osp_precreate_is_low_nolock(), check that
338 * \param[in] d OSP device
340 * \retval 0 - current pool is good enough, 1 - time to precreate
342 static inline int osp_precreate_is_low(struct osp_device *d)
346 if (d->opd_pre == NULL)
349 /* XXX: do we really need locking here? */
350 spin_lock(&d->opd_pre_lock);
351 rc = osp_precreate_is_low_nolock(d);
352 spin_unlock(&d->opd_pre_lock);
357 * Write FID into into last_oid/last_seq file
359 * The function stores the sequence and the in-sequence id into two dedicated
360 * files. The sync argument can be used to request synchronous commit, so the
361 * function won't return until the updates are committed.
363 * \param[in] env LU environment provided by the caller
364 * \param[in] osp OSP device
365 * \param[in] fid fid where sequence/id is taken
366 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
368 * \retval 0 on success
369 * \retval negative negated errno on error
371 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
372 struct lu_fid *fid, int sync)
374 struct osp_thread_info *oti = osp_env_info(env);
375 struct lu_buf *lb_oid = &oti->osi_lb;
376 struct lu_buf *lb_oseq = &oti->osi_lb2;
384 if (osp->opd_storage->dd_rdonly)
387 /* Note: through f_oid is only 32 bits, it will also write 64 bits
388 * for oid to keep compatibility with the previous version. */
390 osp_objid_buf_prep(lb_oid, &oid_off,
391 &oid, osp->opd_index);
393 osp_objseq_buf_prep(lb_oseq, &oseq_off,
394 &fid->f_seq, osp->opd_index);
396 th = dt_trans_create(env, osp->opd_storage);
401 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
402 lb_oid, oid_off, th);
406 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
407 lb_oseq, oseq_off, th);
411 rc = dt_trans_start_local(env, osp->opd_storage, th);
415 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
418 CERROR("%s: can not write to last seq file: rc = %d\n",
419 osp->opd_obd->obd_name, rc);
422 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
425 CERROR("%s: can not write to last seq file: rc = %d\n",
426 osp->opd_obd->obd_name, rc);
430 dt_trans_stop(env, osp->opd_storage, th);
434 static void osp_update_fldb_cache(const struct lu_env *env,
435 struct osp_device *osp,
438 struct lu_seq_range range = { 0 };
439 struct lu_server_fld *server_fld;
440 struct lu_site *site;
442 site = osp->opd_storage->dd_lu_dev.ld_site;
443 server_fld = lu_site2seq(site)->ss_server_fld;
447 fld_range_set_type(&range, LU_SEQ_RANGE_ANY);
448 fld_server_lookup(env, server_fld, fid_seq(fid), &range);
452 * Switch to another sequence
454 * When a current sequence has no available IDs left, OSP has to switch to
455 * another new sequence. OSP requests it using the regular FLDB protocol
456 * and stores synchronously before that is used in precreate. This is needed
457 * to basically have the sequences referenced (not orphaned), otherwise it's
458 * possible that OST has some objects precreated and the clients have data
459 * written to it, but after MDT failover nobody refers those objects and OSP
460 * has no idea that the sequence need cleanup to be done.
461 * While this is very expensive operation, it's supposed to happen infrequently
462 * because sequence has LUSTRE_DATA_SEQ_MAX_WIDTH=32M objects by default.
464 * \param[in] env LU environment provided by the caller
465 * \param[in] osp OSP device
467 * \retval 0 on success
468 * \retval negative negated errno on error
470 static int osp_precreate_rollover_new_seq(struct lu_env *env,
471 struct osp_device *osp)
473 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
474 struct lu_fid *last_fid = &osp->opd_last_used_fid;
478 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
480 CERROR("%s: alloc fid error: rc = %d\n",
481 osp->opd_obd->obd_name, rc);
485 if (fid_seq(fid) <= fid_seq(last_fid)) {
487 CERROR("%s: not a new sequence: fid "DFID", last_used_fid "DFID": rc = %d\n",
488 osp->opd_obd->obd_name, PFID(fid), PFID(last_fid), rc);
495 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
497 CERROR("%s: Can not update oid/seq file: rc = %d\n",
498 osp->opd_obd->obd_name, rc);
502 LCONSOLE(D_INFO, "%s: update sequence from %#llx to %#llx\n",
503 osp->opd_obd->obd_name, fid_seq(last_fid),
505 /* Update last_xxx to the new seq */
506 spin_lock(&osp->opd_pre_lock);
507 osp->opd_last_used_fid = *fid;
508 osp_fid_to_obdid(fid, &osp->opd_last_id);
509 osp->opd_gap_start_fid = *fid;
510 osp->opd_pre_used_fid = *fid;
511 osp->opd_pre_last_created_fid = *fid;
512 spin_unlock(&osp->opd_pre_lock);
515 osp_update_fldb_cache(env, osp, fid);
521 * Find IDs available in current sequence
523 * The function calculates the highest possible ID and the number of IDs
524 * available in the current sequence OSP is using. The number is limited
525 * artifically by the caller (grow param) and the number of IDs available
526 * in the sequence by nature. The function doesn't require an external
529 * \param[in] env LU environment provided by the caller
530 * \param[in] osp OSP device
531 * \param[in] fid FID the caller wants to start with
532 * \param[in] grow how many the caller wants
533 * \param[out] fid the highest calculated FID
534 * \param[out] grow the number of available IDs calculated
536 * \retval 0 on success, 1 - the sequence is empty
538 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
539 struct lu_fid *fid, int *grow)
541 struct osp_thread_info *osi = osp_env_info(env);
542 __u64 seq_width = osp->opd_pre_seq_width;
546 if (fid_is_idif(fid)) {
547 struct lu_fid *last_fid;
548 struct ost_id *oi = &osi->osi_oi;
551 spin_lock(&osp->opd_pre_lock);
552 last_fid = &osp->opd_pre_last_created_fid;
553 fid_to_ostid(last_fid, oi);
554 end = min(ostid_id(oi) + *grow, min(IDIF_MAX_OID, seq_width));
555 *grow = end - ostid_id(oi);
556 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
557 spin_unlock(&osp->opd_pre_lock);
559 if (*grow == 0 || rc)
562 ostid_to_fid(fid, oi, osp->opd_index);
566 spin_lock(&osp->opd_pre_lock);
567 *fid = osp->opd_pre_last_created_fid;
569 end = min((end + *grow), min(OBIF_MAX_OID, seq_width));
570 *grow = end - fid->f_oid;
571 fid->f_oid += end - fid->f_oid;
572 spin_unlock(&osp->opd_pre_lock);
574 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
575 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
577 return *grow > 0 ? 0 : 1;
581 * Prepare and send precreate RPC
583 * The function finds how many objects should be precreated. Then allocates,
584 * prepares and schedules precreate RPC synchronously. Upon reply the function
585 * wakes up the threads waiting for the new objects on this target. If the
586 * target wasn't able to create all the objects requested, then the next
587 * precreate will be asking for fewer objects (i.e. slow precreate down).
589 * \param[in] env LU environment provided by the caller
590 * \param[in] d OSP device
592 * \retval 0 on success
593 * \retval negative negated errno on error
595 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
597 struct osp_thread_info *oti = osp_env_info(env);
598 struct ptlrpc_request *req;
599 struct obd_import *imp;
600 struct ost_body *body;
602 struct lu_fid *fid = &oti->osi_fid;
605 /* don't precreate new objects till OST healthy and has free space */
606 if (unlikely(d->opd_pre_status)) {
607 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
608 d->opd_obd->obd_name, d->opd_pre_status);
613 * if not connection/initialization is compeleted, ignore
615 imp = d->opd_obd->u.cli.cl_import;
618 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
621 req->rq_request_portal = OST_CREATE_PORTAL;
623 /* We should not resend create request - anyway we will have delorphan
624 * and kill these objects.
625 * Only needed for MDS+OSS rolling upgrade interop with 2.16+older.
627 if (unlikely(!imp_connect_replay_create(imp)))
628 req->rq_no_delay = req->rq_no_resend = 1;
630 /* Delorphan happens only with a first MDT-OST connect. resend/replay
631 * handles objects creation on reconnects, no need to do delorhpan
635 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
637 ptlrpc_request_free(req);
641 spin_lock(&d->opd_pre_lock);
642 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
643 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
644 grow = d->opd_pre_create_count;
645 spin_unlock(&d->opd_pre_lock);
647 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
650 *fid = d->opd_pre_last_created_fid;
651 rc = osp_precreate_fids(env, d, fid, &grow);
653 /* Current seq has been used up*/
654 GOTO(out_req, rc = -ENOSPC);
656 if (!osp_is_fid_client(d)) {
657 /* Non-FID client will always send seq 0 because of
659 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
663 fid_to_ostid(fid, &body->oa.o_oi);
664 body->oa.o_valid = OBD_MD_FLGROUP;
666 ptlrpc_request_set_replen(req);
668 if (CFS_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
671 rc = ptlrpc_queue_wait(req);
673 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
676 /* have osp_precreate_reserve() to wait for repeat */
681 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
683 GOTO(out_req, rc = -EPROTO);
685 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
688 spin_lock(&d->opd_pre_lock);
690 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
691 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
692 ": rc = %d\n", d->opd_obd->obd_name,
693 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
694 spin_unlock(&d->opd_pre_lock);
695 GOTO(out_req, rc = -ESTALE);
698 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
701 /* the OST has not managed to create all the
702 * objects we asked for */
703 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
704 d->opd_pre_create_slow = 1;
706 /* the OST is able to keep up with the work,
707 * we could consider increasing create_count
708 * next time if needed */
709 d->opd_pre_create_slow = 0;
712 if ((body->oa.o_valid & OBD_MD_FLSIZE) && body->oa.o_size)
713 d->opd_pre_seq_width = body->oa.o_size;
715 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
716 fid_to_ostid(fid, &body->oa.o_oi);
718 d->opd_pre_last_created_fid = *fid;
719 d->opd_force_creation = false;
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 ptlrpc_req_finished(req);
732 osp_update_fldb_cache(env, d, fid);
738 * Get last precreated object from target (OST)
740 * Sends synchronous RPC to the target (OST) to learn the last precreated
741 * object. This later is used to remove all unused objects (cleanup orphan
742 * procedure). Also, the next object after one we got will be used as a
743 * starting point for the new precreates.
745 * \param[in] env LU environment provided by the caller
746 * \param[in] d OSP device
747 * \param[in] update update or not update last used fid
749 * \retval 0 on success
750 * \retval negative negated errno on error
752 static int osp_get_lastfid_from_ost(const struct lu_env *env,
753 struct osp_device *d, bool update)
755 struct ptlrpc_request *req = NULL;
756 struct obd_import *imp;
757 struct lu_fid *last_fid;
762 imp = d->opd_obd->u.cli.cl_import;
765 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
769 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
770 sizeof(KEY_LAST_FID));
772 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
774 ptlrpc_request_free(req);
778 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
779 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
781 req->rq_no_delay = req->rq_no_resend = 1;
782 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
783 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
785 ptlrpc_request_set_replen(req);
787 rc = ptlrpc_queue_wait(req);
789 /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld),
790 * let sysadm sort this * out.
793 ptlrpc_set_import_active(imp, 0);
797 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
798 if (last_fid == NULL) {
799 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
800 GOTO(out, rc = -EPROTO);
803 if (!fid_is_sane(last_fid)) {
804 CERROR("%s: Got insane last_fid "DFID"\n",
805 d->opd_obd->obd_name, PFID(last_fid));
806 GOTO(out, rc = -EPROTO);
809 /* Only update the last used fid, if the OST has objects for
810 * this sequence, i.e. fid_oid > 0 */
811 if (fid_oid(last_fid) > 0 && update)
812 d->opd_last_used_fid = *last_fid;
814 if (fid_seq(last_fid) == fid_seq(&d->opd_last_used_fid)) {
815 if (fid_oid(last_fid) == 0 ||
816 (fid_seq_is_norm(fid_seq(last_fid)) &&
817 fid_oid(last_fid) == LUSTRE_FID_INIT_OID)) {
818 /* reformatted OST, it requires creation request
819 * to recreate objects
821 spin_lock(&d->opd_pre_lock);
822 d->opd_force_creation = true;
823 d->opd_pre_create_count = OST_MIN_PRECREATE;
824 spin_unlock(&d->opd_pre_lock);
827 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
831 ptlrpc_req_finished(req);
836 * Cleanup orphans on OST
838 * This function is called in a contex of a dedicated thread handling
839 * all the precreation suff. The function waits till local recovery
840 * is complete, then identify all the unreferenced objects (orphans)
841 * using the highest ID referenced by a local and the highest object
842 * precreated by the target. The found range is a subject to removal
843 * using specially flagged RPC. During this process OSP is marked
844 * unavailable for new objects.
846 * \param[in] env LU environment provided by the caller
847 * \param[in] d OSP device
849 * \retval 0 on success
850 * \retval negative negated errno on error
852 static int osp_precreate_cleanup_orphans(struct lu_env *env,
853 struct osp_device *d)
855 struct osp_thread_info *osi = osp_env_info(env);
856 struct lu_fid *last_fid = &osi->osi_fid;
857 struct ptlrpc_request *req = NULL;
858 struct obd_import *imp = d->opd_obd->u.cli.cl_import;
859 struct ost_body *body;
860 int update_status = 0;
867 * Do cleanup orphans only with a first connection, after that
868 * all precreate requests uses resend/replay flags to support OST
869 * failover/reconnect.
871 if (d->opd_cleanup_orphans_done && imp_connect_replay_create(imp)) {
872 rc = osp_get_lastfid_from_ost(env, d, false);
876 * wait for local recovery to finish, so we can cleanup orphans
877 * orphans are all objects since "last used" (assigned), but
878 * there might be objects reserved and in some cases they won't
879 * be used. we can't cleanup them till we're sure they won't be
880 * used. also can't we allow new reservations because they may
881 * end up getting orphans being cleaned up below. so we block
882 * new reservations and wait till all reserved objects either
885 spin_lock(&d->opd_pre_lock);
886 d->opd_pre_recovering = 1;
887 spin_unlock(&d->opd_pre_lock);
889 * The locking above makes sure the opd_pre_reserved check below will
890 * catch all osp_precreate_reserve() calls who find
891 * "!opd_pre_recovering".
893 wait_event_idle(d->opd_pre_waitq,
894 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
895 !d->opd_pre_task || d->opd_got_disconnected);
896 if (!d->opd_pre_task || d->opd_got_disconnected)
897 GOTO(out, rc = -EAGAIN);
899 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
900 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
902 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val);
904 *last_fid = d->opd_last_used_fid;
905 /* The OSP should already get the valid seq now */
906 LASSERT(!fid_is_zero(last_fid));
907 if (fid_oid(&d->opd_last_used_fid) < 2) {
908 /* lastfid looks strange... ask OST */
909 rc = osp_get_lastfid_from_ost(env, d, true);
914 imp = d->opd_obd->u.cli.cl_import;
917 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
919 GOTO(out, rc = -ENOMEM);
921 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
923 ptlrpc_request_free(req);
928 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
930 GOTO(out, rc = -EPROTO);
932 body->oa.o_flags = OBD_FL_DELORPHAN;
933 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
935 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
937 ptlrpc_request_set_replen(req);
939 /* Don't resend the delorphan req */
940 req->rq_no_resend = req->rq_no_delay = 1;
942 rc = ptlrpc_queue_wait(req);
948 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
950 GOTO(out, rc = -EPROTO);
953 * OST provides us with id new pool starts from in body->oa.o_id
955 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
957 spin_lock(&d->opd_pre_lock);
958 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
960 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
961 d->opd_pre_last_created_fid = d->opd_last_used_fid;
963 d->opd_pre_create_count = OST_MIN_PRECREATE;
964 d->opd_pre_last_created_fid = *last_fid;
967 * This empties the pre-creation pool and effectively blocks any new
970 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
971 d->opd_pre_create_slow = 0;
972 if ((body->oa.o_valid & OBD_MD_FLSIZE) && body->oa.o_size)
973 d->opd_pre_seq_width = body->oa.o_size;
974 spin_unlock(&d->opd_pre_lock);
976 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
977 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
978 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
981 ptlrpc_req_finished(req);
985 * If rc is zero, the pre-creation window should have been emptied.
986 * Since waking up the herd would be useless without pre-created
987 * objects, we defer the signal to osp_precreate_send() in that case.
991 CERROR("%s: cannot cleanup orphans: rc = %d\n",
992 d->opd_obd->obd_name, rc);
993 /* we can't proceed from here, OST seem to
994 * be in a bad shape, better to wait for
995 * a new instance of the server and repeat
996 * from the beginning. notify possible waiters
997 * this OSP isn't quite functional yet */
998 osp_pre_update_status(d, rc);
1000 wake_up_all(&d->opd_pre_user_waitq);
1003 spin_lock(&d->opd_pre_lock);
1004 d->opd_pre_recovering = 0;
1005 spin_unlock(&d->opd_pre_lock);
1006 d->opd_cleanup_orphans_done = true;
1013 * Update precreate status using statfs data
1015 * The function decides whether this OSP should be used for new objects.
1016 * IOW, whether this OST is used up or has some free space. Cached statfs
1017 * data is used to make this decision. If the latest result of statfs
1018 * request (rc argument) is not success, then just mark OSP unavailable
1021 * The new statfs data is passed in \a msfs and needs to be stored into
1022 * opd_statfs, but only after the various flags in os_state are set, so
1023 * that the new statfs data is not visible without appropriate flags set.
1024 * As such, there is no need to clear the flags here, since this is called
1025 * with new statfs data, and they should not be cleared if sent from OST.
1027 * Add a bit of hysteresis so this flag isn't continually flapping, and
1028 * ensure that new files don't get extremely fragmented due to only a
1029 * small amount of available space in the filesystem. We want to set
1030 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1031 * reserved size free, and still copy them from the old state when there
1032 * is less than 2*reserved size free space or inodes.
1034 * \param[in] d OSP device
1035 * \param[in] msfs statfs data
1037 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1039 u32 old_state = d->opd_statfs.os_state;
1042 /* statfs structure not initialized yet */
1043 if (unlikely(!msfs->os_type))
1046 /* if the low and high watermarks have not been initialized yet */
1047 if (unlikely(d->opd_reserved_mb_high == 0 &&
1048 d->opd_reserved_mb_low == 0)) {
1049 /* Use ~0.1% by default to disable object allocation,
1050 * and ~0.2% to enable, size in MB, set both watermark
1052 spin_lock(&d->opd_pre_lock);
1053 if (d->opd_reserved_mb_high == 0 &&
1054 d->opd_reserved_mb_low == 0) {
1055 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1056 msfs->os_blocks) >> 20;
1057 if (d->opd_reserved_mb_low < 1)
1058 d->opd_reserved_mb_low = 1;
1059 d->opd_reserved_mb_high =
1060 (d->opd_reserved_mb_low << 1) + 1;
1062 spin_unlock(&d->opd_pre_lock);
1065 if (unlikely(d->opd_reserved_ino_high == 0 &&
1066 d->opd_reserved_ino_low == 0)) {
1067 /* Use ~0.0001% by default to disallow distributed transactions,
1068 * and ~0.0002% to allow, set both watermark
1070 spin_lock(&d->opd_pre_lock);
1071 if (d->opd_reserved_ino_high == 0 &&
1072 d->opd_reserved_ino_low == 0) {
1073 d->opd_reserved_ino_low = msfs->os_ffree >> 20;
1074 if (d->opd_reserved_ino_low < 32)
1075 d->opd_reserved_ino_low = 32;
1076 d->opd_reserved_ino_high =
1077 (d->opd_reserved_ino_low << 1) + 1;
1079 spin_unlock(&d->opd_pre_lock);
1082 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1083 if (msfs->os_ffree < d->opd_reserved_ino_low)
1084 msfs->os_state |= OS_STATFS_ENOINO;
1085 else if (msfs->os_ffree <= d->opd_reserved_ino_high)
1086 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1087 /* else don't clear flags in new msfs->os_state sent from OST */
1089 if (available_mb < d->opd_reserved_mb_low)
1090 msfs->os_state |= OS_STATFS_ENOSPC;
1091 else if (available_mb <= d->opd_reserved_mb_high)
1092 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1093 /* else don't clear flags in new msfs->os_state sent from OST */
1096 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1097 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1098 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1099 msfs->os_files, msfs->os_ffree, msfs->os_state,
1100 d->opd_pre ? d->opd_pre_status : 0);
1105 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1106 d->opd_pre_status = -ENOSPC;
1107 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1108 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1109 d->opd_obd->obd_name, msfs->os_state,
1111 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1112 atomic_read(&d->opd_sync_changes),
1113 atomic_read(&d->opd_sync_rpcs_in_progress));
1114 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1115 d->opd_pre_status = 0;
1116 spin_lock(&d->opd_pre_lock);
1117 d->opd_pre_create_slow = 0;
1118 d->opd_pre_create_count = OST_MIN_PRECREATE;
1119 spin_unlock(&d->opd_pre_lock);
1120 wake_up(&d->opd_pre_waitq);
1123 "%s: available: state=%x: rc = %d\n",
1124 d->opd_obd->obd_name, msfs->os_state,
1127 /* we only get here if rc == 0 in the caller */
1128 d->opd_pre_status = 0;
1131 /* Object precreation skipped on OST if manually disabled */
1132 if (d->opd_pre_max_create_count == 0)
1133 msfs->os_state |= OS_STATFS_NOCREATE;
1134 /* else don't clear flags in new msfs->os_state sent from OST */
1137 /* copy only new statfs state to make it visible to MDS threads */
1138 if (&d->opd_statfs != msfs)
1139 d->opd_statfs = *msfs;
1143 * Initialize FID for precreation
1145 * For a just created new target, a new sequence should be taken.
1146 * The function checks there is no IDIF in use (if the target was
1147 * added with the older version of Lustre), then requests a new
1148 * sequence from FLDB using the regular protocol. Then this new
1149 * sequence is stored on a persisten storage synchronously to prevent
1150 * possible object leakage (for the detail see the description for
1151 * osp_precreate_rollover_new_seq()).
1153 * \param[in] osp OSP device
1155 * \retval 0 on success
1156 * \retval negative negated errno on error
1158 static int osp_init_pre_fid(struct lu_env *env, struct osp_device *osp)
1160 struct osp_thread_info *osi;
1161 struct lu_client_seq *cli_seq;
1162 struct lu_fid *last_fid;
1166 LASSERT(osp->opd_pre != NULL);
1168 if (CFS_FAIL_CHECK(OBD_FAIL_OSP_FAIL_SEQ_ALLOC)) {
1169 unsigned int timeout = cfs_fail_val ?: 1;
1171 schedule_timeout_uninterruptible(cfs_time_seconds(timeout));
1175 /* Let's check if the current last_seq/fid is valid,
1176 * otherwise request new sequence from the controller */
1177 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1178 /* Non-MDT0 can only use normal sequence for
1180 if (fid_is_norm(&osp->opd_last_used_fid))
1183 /* Initially MDT0 will start with IDIF, after
1184 * that it will request new sequence from the
1186 if (fid_is_idif(&osp->opd_last_used_fid) ||
1187 fid_is_norm(&osp->opd_last_used_fid))
1191 if (!fid_is_zero(&osp->opd_last_used_fid))
1192 CWARN("%s: invalid last used fid "DFID
1193 ", try to get new sequence.\n",
1194 osp->opd_obd->obd_name,
1195 PFID(&osp->opd_last_used_fid));
1197 osi = osp_env_info(env);
1198 last_fid = &osi->osi_fid;
1200 /* For a freshed fs, it will allocate a new sequence first */
1201 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1202 cli_seq = osp->opd_obd->u.cli.cl_seq;
1203 rc = seq_client_get_seq(env, cli_seq, &last_fid->f_seq);
1205 CERROR("%s: alloc fid error: rc = %d\n",
1206 osp->opd_obd->obd_name, rc);
1210 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1212 last_fid->f_oid = 1;
1213 last_fid->f_ver = 0;
1215 spin_lock(&osp->opd_pre_lock);
1216 osp->opd_last_used_fid = *last_fid;
1217 osp->opd_pre_used_fid = *last_fid;
1218 osp->opd_pre_last_created_fid = *last_fid;
1219 spin_unlock(&osp->opd_pre_lock);
1220 rc = osp_write_last_oid_seq_files(env, osp, last_fid, 1);
1222 CERROR("%s: write fid error: rc = %d\n",
1223 osp->opd_obd->obd_name, rc);
1231 struct osp_device *opta_dev;
1232 struct lu_env opta_env;
1233 struct completion *opta_started;
1236 * The core of precreate functionality
1238 * The function implements the main precreation loop. Basically it
1239 * involves connecting to the target, precerate FID initialization,
1240 * identifying and removing orphans, then serving precreation. As
1241 * part of the latter, the thread is responsible for statfs data
1242 * updates. The precreation is mostly driven by another threads
1243 * asking for new OST objects - those askers wake the thread when
1244 * the number of precreated objects reach low watermark.
1245 * After a disconnect, the sequence above repeats. This is keep going
1246 * until the thread is requested to stop.
1248 * \param[in] _arg private data the thread (OSP device to handle)
1250 * \retval 0 on success
1251 * \retval negative negated errno on error
1253 static int osp_precreate_thread(void *_args)
1255 struct opt_args *args = _args;
1256 struct osp_device *d = args->opta_dev;
1257 struct lu_env *env = &args->opta_env;
1262 complete(args->opta_started);
1264 /* wait for connection from the layers above */
1265 wait_event_idle(d->opd_pre_waitq,
1266 kthread_should_stop() ||
1267 d->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1269 while (!kthread_should_stop()) {
1271 * need to be connected to OST
1273 while (!kthread_should_stop()) {
1274 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1275 d->opd_imp_connected &&
1276 !d->opd_got_disconnected)
1278 wait_event_idle(d->opd_pre_waitq,
1279 kthread_should_stop() ||
1280 d->opd_new_connection);
1282 if (!d->opd_new_connection)
1285 CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
1287 d->opd_new_connection = 0;
1288 d->opd_got_disconnected = 0;
1292 if (kthread_should_stop())
1296 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1297 LASSERT(d->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1299 /* Init fid for osp_precreate if necessary */
1300 rc = osp_init_pre_fid(env, d);
1302 CERROR("%s: init pre fid error: rc = %d\n",
1303 d->opd_obd->obd_name, rc);
1308 if (osp_statfs_update(env, d)) {
1309 if (wait_event_idle_timeout(d->opd_pre_waitq,
1310 kthread_should_stop(),
1311 cfs_time_seconds(5)) == 0)
1312 l_wait_event_abortable(
1314 kthread_should_stop());
1320 * Clean up orphans or recreate missing objects.
1322 rc = osp_precreate_cleanup_orphans(env, d);
1324 schedule_timeout_interruptible(cfs_time_seconds(1));
1330 * connected, can handle precreates now
1332 while (!kthread_should_stop()) {
1333 wait_event_idle(d->opd_pre_waitq,
1334 kthread_should_stop() ||
1335 (osp_precreate_is_low(d) &&
1336 !(osp_precreate_end_seq(d) &&
1337 osp_objs_precreated(d) != 0)) ||
1338 osp_statfs_need_update(d) ||
1339 d->opd_got_disconnected);
1341 if (kthread_should_stop())
1344 /* something happened to the connection
1345 * have to start from the beginning */
1346 if (d->opd_got_disconnected)
1349 if (osp_statfs_need_update(d))
1350 if (osp_statfs_update(env, d))
1353 if (d->opd_pre == NULL)
1356 /* To avoid handling different seq in precreate/orphan
1357 * cleanup, it will hold precreate until current seq is
1359 if (unlikely(osp_precreate_end_seq(d))) {
1360 if (osp_objs_precreated(d) == 0) {
1361 rc = osp_precreate_rollover_new_seq(env, d);
1369 if (osp_precreate_is_low(d)) {
1370 rc = osp_precreate_send(env, d);
1371 /* osp_precreate_send() sets opd_pre_status
1372 * in case of error, that prevent the using of
1374 if (rc < 0 && rc != -ENOSPC &&
1375 rc != -ETIMEDOUT && rc != -ENOTCONN)
1376 CERROR("%s: cannot precreate objects:"
1378 d->opd_obd->obd_name, rc);
1390 * Check when to stop to wait for precreate objects.
1392 * The caller wanting a new OST object can't wait undefinitely. The
1393 * function checks for few conditions including available new OST
1394 * objects, disconnected OST, lack of space with no pending destroys,
1395 * etc. IOW, it checks whether the current OSP state is good to keep
1396 * waiting or it's better to give up.
1398 * \param[in] env LU environment provided by the caller
1399 * \param[in] d OSP device
1401 * \retval 0 - keep waiting, 1 - no luck
1403 static int osp_precreate_ready_condition(const struct lu_env *env,
1404 struct osp_device *d)
1406 /* Bail out I/O fails to OST */
1407 if (d->opd_pre_status != 0 &&
1408 d->opd_pre_status != -EAGAIN &&
1409 d->opd_pre_status != -ENODEV &&
1410 d->opd_pre_status != -ENOTCONN &&
1411 d->opd_pre_status != -ENOSPC) {
1413 if (d->opd_pre_status != -EIO)
1414 CERROR("%s: precreate failed opd_pre_status %d\n",
1415 d->opd_obd->obd_name, d->opd_pre_status);
1419 if (d->opd_pre_recovering || d->opd_force_creation)
1422 /* ready if got enough precreated objects */
1423 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1424 if (d->opd_pre_reserved + 1 < osp_objs_precreated(d))
1427 /* ready if OST reported no space and no destroys in progress */
1428 if (atomic_read(&d->opd_sync_changes) +
1429 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1430 d->opd_pre_status == -ENOSPC)
1437 * Reserve object in precreate pool
1439 * When the caller wants to create a new object on this target (target
1440 * represented by the given OSP), it should declare this intention using
1441 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1442 * to reserve an object in the existing precreated pool or wait up to
1443 * obd_timeout for the available object to appear in the pool (a dedicated
1444 * thread will be doing real precreation in background). The object can be
1445 * consumed later with osp_precreate_get_fid() or be released with call to
1446 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1447 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1448 * If the space on the target is short and there is a pending object destroy,
1449 * then the function forces local commit to speedup space release (see
1450 * osp_sync.c for the details).
1452 * \param[in] env LU environment provided by the caller
1453 * \param[in] d OSP device
1455 * \retval 0 on success
1456 * \retval -ENOSPC when no space on OST
1457 * \retval -EAGAIN try later, slow precreation in progress
1458 * \retval -EIO when no access to OST
1460 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
1463 time64_t expire = ktime_get_seconds() + obd_timeout;
1464 int precreated, rc, synced = 0;
1468 LASSERTF(osp_objs_precreated(d) >= 0, "Last created FID "DFID
1469 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1470 PFID(&d->opd_pre_used_fid));
1472 /* opd_pre_max_create_count 0 to not use specified OST. */
1473 if (d->opd_pre_max_create_count == 0)
1478 * - preallocation is done
1479 * - no free space expected soon
1480 * - can't connect to OST for too long (obd_timeout)
1481 * - OST can allocate fid sequence.
1483 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1484 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1486 spin_lock(&d->opd_pre_lock);
1487 precreated = osp_objs_precreated_nolock(d);
1489 * increase number of precreations
1491 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1492 d->opd_pre_create_slow == 0 &&
1493 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1494 d->opd_pre_create_slow = 1;
1495 d->opd_pre_create_count *= 2;
1498 if (!d->opd_pre_recovering && !d->opd_force_creation) {
1499 if (precreated > d->opd_pre_reserved) {
1500 d->opd_pre_reserved++;
1501 spin_unlock(&d->opd_pre_lock);
1505 * XXX: don't wake up if precreation
1508 if (osp_precreate_is_low_nolock(d) &&
1509 !osp_precreate_end_seq_nolock(d))
1510 wake_up(&d->opd_pre_waitq);
1513 } else if (unlikely(precreated &&
1514 osp_precreate_end_seq_nolock(d))) {
1516 * precreate pool is reaching the end of the
1517 * current seq, and doesn't have enough objects
1520 spin_unlock(&d->opd_pre_lock);
1524 spin_unlock(&d->opd_pre_lock);
1527 * all precreated objects have been used and no-space
1528 * status leave us no chance to succeed very soon
1529 * but if there is destroy in progress, then we should
1530 * wait till that is done - some space might be released
1532 if (unlikely(rc == -ENOSPC)) {
1533 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1534 /* force local commit to release space */
1535 dt_commit_async(env, d->opd_storage);
1536 osp_sync_check_for_work(d);
1539 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1540 /* just wait till destroys are done
1541 * see wait_event_idle_timeout() below
1544 if (atomic_read(&d->opd_sync_changes) +
1545 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1546 /* no hope for free space */
1551 /* XXX: don't wake up if precreation is in progress */
1552 wake_up(&d->opd_pre_waitq);
1554 if (ktime_get_seconds() >= expire) {
1560 LASSERT(d->opd_pre);
1565 CDEBUG(D_INFO, "%s: Sleeping on objects\n",
1566 d->opd_obd->obd_name);
1567 if (wait_event_idle_timeout(
1568 d->opd_pre_user_waitq,
1569 osp_precreate_ready_condition(env, d),
1570 cfs_time_seconds(obd_timeout)) == 0) {
1572 "%s: slow creates, last="DFID", next="DFID", "
1573 "reserved=%llu, sync_changes=%u, "
1574 "sync_rpcs_in_progress=%d, status=%d\n",
1575 d->opd_obd->obd_name,
1576 PFID(&d->opd_pre_last_created_fid),
1577 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1578 atomic_read(&d->opd_sync_changes),
1579 atomic_read(&d->opd_sync_rpcs_in_progress),
1582 CDEBUG(D_INFO, "%s: Waked up, status=%d\n",
1583 d->opd_obd->obd_name, d->opd_pre_status);
1591 * Get a FID from precreation pool
1593 * The function is a companion for osp_precreate_reserve() - it assigns
1594 * a specific FID from the precreate. The function should be called only
1595 * if the call to osp_precreate_reserve() was successful. The function
1596 * updates a local storage to remember the highest object ID referenced
1597 * by the node in the given sequence.
1599 * A very importan details: this is supposed to be called once the
1600 * transaction is started, so on-disk update will be atomic with the
1601 * data (like LOVEA) refering this object. Then the object won't be leaked:
1602 * either it's referenced by the committed transaction or it's a subject
1603 * to the orphan cleanup procedure.
1605 * \param[in] env LU environment provided by the caller
1606 * \param[in] d OSP device
1607 * \param[out] fid generated FID
1609 * \retval 0 on success
1610 * \retval negative negated errno on error
1612 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1615 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1617 /* grab next id from the pool */
1618 spin_lock(&d->opd_pre_lock);
1620 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1621 &d->opd_pre_last_created_fid) < 0,
1622 "next fid "DFID" > last created fid "DFID"\n",
1623 PFID(&d->opd_pre_used_fid),
1624 PFID(&d->opd_pre_last_created_fid));
1626 /* Non-IDIF FIDs shouldn't get here with OID == OBIF_MAX_OID. For IDIF,
1627 * f_oid wraps and "f_seq" (holding high 16 bits of ID) needs increment
1629 if (fid_is_idif(pre_used_fid) &&
1630 unlikely(fid_oid(pre_used_fid) == OBIF_MAX_OID)) {
1632 __u32 idx = fid_idif_ost_idx(pre_used_fid);
1634 fid_to_ostid(pre_used_fid, &oi);
1636 ostid_to_fid(pre_used_fid, &oi, idx);
1638 pre_used_fid->f_oid++;
1641 memcpy(fid, pre_used_fid, sizeof(*fid));
1642 d->opd_pre_reserved--;
1644 * last_used_id must be changed along with getting new id otherwise
1645 * we might miscalculate gap causing object loss or leak
1647 osp_update_last_fid(d, fid, false);
1648 spin_unlock(&d->opd_pre_lock);
1651 * probably main thread suspended orphan cleanup till
1652 * all reservations are released, see comment in
1653 * osp_precreate_thread() just before orphan cleanup
1655 if (unlikely(d->opd_pre_reserved == 0 &&
1656 (d->opd_pre_recovering || d->opd_pre_status)))
1657 wake_up(&d->opd_pre_waitq);
1663 * Set size regular attribute on an object
1665 * When a striping is created late, it's possible that size is already
1666 * initialized on the file. Then the new striping should inherit size
1667 * from the file. The function sets size on the object using the regular
1668 * protocol (OST_PUNCH).
1669 * XXX: should be re-implemented using OUT ?
1671 * \param[in] env LU environment provided by the caller
1672 * \param[in] dt object
1673 * \param[in] size size to set.
1675 * \retval 0 on success
1676 * \retval negative negated errno on error
1678 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1681 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1682 struct ptlrpc_request *req = NULL;
1683 struct obd_import *imp;
1684 struct ost_body *body;
1685 struct obdo *oa = NULL;
1690 imp = d->opd_obd->u.cli.cl_import;
1693 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1697 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1699 ptlrpc_request_free(req);
1704 * XXX: decide how do we do here with resend
1705 * if we don't resend, then client may see wrong file size
1706 * if we do resend, then MDS thread can get stuck for quite long
1707 * and if we don't resend, then client will also get -EAGAIN !!
1708 * (see LU-7975 and sanity/test_27F use cases)
1709 * but let's decide not to resend/delay this truncate request to OST
1710 * and allow Client to decide to resend, in a less agressive way from
1711 * after_reply(), by returning -EINPROGRESS instead of
1712 * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
1713 * end of this routine
1715 req->rq_no_resend = req->rq_no_delay = 1;
1717 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1718 ptlrpc_at_set_req_timeout(req);
1722 GOTO(out, rc = -ENOMEM);
1724 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1727 oa->o_blocks = OBD_OBJECT_EOF;
1728 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1729 OBD_MD_FLID | OBD_MD_FLGROUP;
1731 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1733 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1735 /* XXX: capa support? */
1736 /* osc_pack_capa(req, body, capa); */
1738 ptlrpc_request_set_replen(req);
1740 rc = ptlrpc_queue_wait(req);
1742 /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
1743 * since we have decided not to resend/delay, but this could
1744 * lead to wrong size to be seen at Client side and even process
1745 * trying to open to exit/fail if not itself handling -EAGAIN.
1746 * So it should be better to return -EINPROGRESS instead and
1747 * leave the decision to resend at Client side in after_reply()
1749 if (rc == -EAGAIN) {
1751 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1752 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1755 CERROR("can't punch object: %d\n", rc);
1759 ptlrpc_req_finished(req);
1766 * Initialize precreation functionality of OSP
1768 * Prepares all the internal structures and starts the precreate thread
1770 * \param[in] d OSP device
1772 * \retval 0 on success
1773 * \retval negative negated errno on error
1775 int osp_init_precreate(struct osp_device *d)
1779 OBD_ALLOC_PTR(d->opd_pre);
1780 if (d->opd_pre == NULL)
1783 /* initially precreation isn't ready */
1784 init_waitqueue_head(&d->opd_pre_user_waitq);
1785 d->opd_pre_status = -EAGAIN;
1786 fid_zero(&d->opd_pre_used_fid);
1787 d->opd_pre_used_fid.f_oid = 1;
1788 fid_zero(&d->opd_pre_last_created_fid);
1789 d->opd_pre_last_created_fid.f_oid = 1;
1791 d->opd_pre_reserved = 0;
1792 d->opd_pre_seq_width = LUSTRE_DATA_SEQ_MAX_WIDTH;
1793 d->opd_got_disconnected = 1;
1794 d->opd_pre_create_slow = 0;
1795 d->opd_pre_create_count = OST_MIN_PRECREATE;
1796 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1797 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1798 d->opd_reserved_mb_high = 0;
1799 d->opd_reserved_mb_low = 0;
1800 d->opd_cleanup_orphans_done = false;
1801 d->opd_force_creation = false;
1807 * Finish precreate functionality of OSP
1810 * Asks all the activity (the thread, update timer) to stop, then
1811 * wait till that is done.
1813 * \param[in] d OSP device
1815 void osp_precreate_fini(struct osp_device *d)
1819 if (d->opd_pre == NULL)
1822 OBD_FREE_PTR(d->opd_pre);
1828 int osp_init_statfs(struct osp_device *d)
1830 struct task_struct *task;
1831 struct opt_args *args;
1832 DECLARE_COMPLETION_ONSTACK(started);
1837 spin_lock_init(&d->opd_pre_lock);
1838 init_waitqueue_head(&d->opd_pre_waitq);
1841 * Initialize statfs-related things
1843 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1844 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1845 1000 * NSEC_PER_SEC);
1846 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1848 ktime_to_ns(d->opd_statfs_fresh_till));
1849 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1850 (unsigned long)d, 0);
1852 if (d->opd_storage->dd_rdonly)
1855 OBD_ALLOC_PTR(args);
1859 args->opta_started = &started;
1860 rc = lu_env_init(&args->opta_env,
1861 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1863 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1870 * start thread handling precreation and statfs updates
1872 task = kthread_create(osp_precreate_thread, args,
1873 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1875 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1876 lu_env_fini(&args->opta_env);
1878 RETURN(PTR_ERR(task));
1880 d->opd_pre_task = task;
1881 wake_up_process(task);
1882 wait_for_completion(&started);
1887 void osp_statfs_fini(struct osp_device *d)
1889 struct task_struct *task = d->opd_pre_task;
1892 timer_delete(&d->opd_statfs_timer);
1894 d->opd_pre_task = NULL;