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] d OSP device
206 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
208 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
209 struct ptlrpc_request *req;
210 struct obd_import *imp;
211 union ptlrpc_async_args *aa;
216 CDEBUG(D_CACHE, "going to update statfs\n");
218 imp = d->opd_obd->u.cli.cl_import;
221 req = ptlrpc_request_alloc(imp,
222 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
226 rc = ptlrpc_request_pack(req,
227 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
228 d->opd_pre ? OST_STATFS : MDS_STATFS);
230 ptlrpc_request_free(req);
233 ptlrpc_request_set_replen(req);
235 req->rq_request_portal = OST_CREATE_PORTAL;
236 ptlrpc_at_set_req_timeout(req);
238 req->rq_interpret_reply = osp_statfs_interpret;
239 aa = ptlrpc_req_async_args(aa, req);
240 aa->pointer_arg[0] = d;
243 * no updates till reply
245 timer_delete(&d->opd_statfs_timer);
246 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
247 d->opd_statfs_update_in_progress = 1;
249 ptlrpcd_add_req(req);
251 /* we still want to sync changes if no new changes are coming */
252 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
255 if (atomic_read(&d->opd_sync_changes)) {
258 th = dt_trans_create(env, d->opd_storage);
260 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
263 rc = dt_trans_start_local(env, d->opd_storage, th);
265 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
266 d->opd_obd->obd_name,
267 atomic_read(&d->opd_sync_changes));
268 osp_sync_add_commit_cb_1s(env, d, th);
270 dt_trans_stop(env, d->opd_storage, th);
278 * Schedule an immediate update for statfs data
280 * If cached statfs data claim no free space, but OSP has got a request to
281 * destroy an object (so release some space probably), then we may need to
282 * refresh cached statfs data sooner than planned. The function checks there
283 * is no statfs update going and schedules immediate update if so.
284 * XXX: there might be a case where removed object(s) do not add free space (empty
285 * object). If the number of such deletions is high, then we can start to update
286 * statfs too often causing a RPC storm. some throttling is needed...
288 * \param[in] d OSP device where statfs data needs to be refreshed
290 void osp_statfs_need_now(struct osp_device *d)
292 if (!d->opd_statfs_update_in_progress) {
294 * if current status is -ENOSPC (lack of free space on OST)
295 * then we should poll OST immediately once object destroy
298 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
299 timer_delete(&d->opd_statfs_timer);
300 wake_up(&d->opd_pre_waitq);
305 * Return number of precreated objects
307 * A simple helper to calculate the number of precreated objects on the device.
309 * \param[in] env LU environment provided by the caller
310 * \param[in] osp OSP device
312 * \retval the number of the precreated objects
314 static inline int osp_objs_precreated(const struct lu_env *env,
315 struct osp_device *osp)
317 return osp_fid_diff(&osp->opd_pre_last_created_fid,
318 &osp->opd_pre_used_fid);
322 * Check pool of precreated objects is nearly empty
324 * We should not wait till the pool of the precreated objects is exhausted,
325 * because then there will be a long period of OSP being unavailable for the
326 * new creations due to lenghty precreate RPC. Instead we ask for another
327 * precreation ahead and hopefully have it ready before the current pool is
328 * empty. Notice this function relies on an external locking.
330 * \param[in] env LU environment provided by the caller
331 * \param[in] d OSP device
333 * \retval 0 - current pool is good enough, 1 - time to precreate
335 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
336 struct osp_device *d)
338 int window = osp_objs_precreated(env, d);
340 /* don't consider new precreation till OST is healty and
342 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2 ||
343 d->opd_force_creation) && (d->opd_pre_status == 0));
347 * Check pool of precreated objects
349 * This is protected version of osp_precreate_near_empty_nolock(), check that
352 * \param[in] env LU environment provided by the caller
353 * \param[in] d OSP device
355 * \retval 0 - current pool is good enough, 1 - time to precreate
357 static inline int osp_precreate_near_empty(const struct lu_env *env,
358 struct osp_device *d)
362 if (d->opd_pre == NULL)
365 /* XXX: do we really need locking here? */
366 spin_lock(&d->opd_pre_lock);
367 rc = osp_precreate_near_empty_nolock(env, d);
368 spin_unlock(&d->opd_pre_lock);
373 * Write FID into into last_oid/last_seq file
375 * The function stores the sequence and the in-sequence id into two dedicated
376 * files. The sync argument can be used to request synchronous commit, so the
377 * function won't return until the updates are committed.
379 * \param[in] env LU environment provided by the caller
380 * \param[in] osp OSP device
381 * \param[in] fid fid where sequence/id is taken
382 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
384 * \retval 0 on success
385 * \retval negative negated errno on error
387 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
388 struct lu_fid *fid, int sync)
390 struct osp_thread_info *oti = osp_env_info(env);
391 struct lu_buf *lb_oid = &oti->osi_lb;
392 struct lu_buf *lb_oseq = &oti->osi_lb2;
400 if (osp->opd_storage->dd_rdonly)
403 /* Note: through f_oid is only 32 bits, it will also write 64 bits
404 * for oid to keep compatibility with the previous version. */
406 osp_objid_buf_prep(lb_oid, &oid_off,
407 &oid, osp->opd_index);
409 osp_objseq_buf_prep(lb_oseq, &oseq_off,
410 &fid->f_seq, osp->opd_index);
412 th = dt_trans_create(env, osp->opd_storage);
417 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
418 lb_oid, oid_off, th);
422 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
423 lb_oseq, oseq_off, th);
427 rc = dt_trans_start_local(env, osp->opd_storage, th);
431 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
434 CERROR("%s: can not write to last seq file: rc = %d\n",
435 osp->opd_obd->obd_name, rc);
438 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
441 CERROR("%s: can not write to last seq file: rc = %d\n",
442 osp->opd_obd->obd_name, rc);
446 dt_trans_stop(env, osp->opd_storage, th);
450 static void osp_update_fldb_cache(const struct lu_env *env,
451 struct osp_device *osp,
454 struct lu_seq_range range = { 0 };
455 struct lu_server_fld *server_fld;
456 struct lu_site *site;
458 site = osp->opd_storage->dd_lu_dev.ld_site;
459 server_fld = lu_site2seq(site)->ss_server_fld;
463 fld_range_set_type(&range, LU_SEQ_RANGE_ANY);
464 fld_server_lookup(env, server_fld, fid_seq(fid), &range);
468 * Switch to another sequence
470 * When a current sequence has no available IDs left, OSP has to switch to
471 * another new sequence. OSP requests it using the regular FLDB protocol
472 * and stores synchronously before that is used in precreate. This is needed
473 * to basically have the sequences referenced (not orphaned), otherwise it's
474 * possible that OST has some objects precreated and the clients have data
475 * written to it, but after MDT failover nobody refers those objects and OSP
476 * has no idea that the sequence need cleanup to be done.
477 * While this is very expensive operation, it's supposed to happen infrequently
478 * because sequence has LUSTRE_DATA_SEQ_MAX_WIDTH=32M objects by default.
480 * \param[in] env LU environment provided by the caller
481 * \param[in] osp OSP device
483 * \retval 0 on success
484 * \retval negative negated errno on error
486 static int osp_precreate_rollover_new_seq(struct lu_env *env,
487 struct osp_device *osp)
489 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
490 struct lu_fid *last_fid = &osp->opd_last_used_fid;
494 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
496 CERROR("%s: alloc fid error: rc = %d\n",
497 osp->opd_obd->obd_name, rc);
503 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
504 "fid "DFID", last_fid "DFID"\n", PFID(fid),
507 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
509 CERROR("%s: Can not update oid/seq file: rc = %d\n",
510 osp->opd_obd->obd_name, rc);
514 LCONSOLE(D_INFO, "%s: update sequence from %#llx to %#llx\n",
515 osp->opd_obd->obd_name, fid_seq(last_fid),
517 /* Update last_xxx to the new seq */
518 spin_lock(&osp->opd_pre_lock);
519 osp->opd_last_used_fid = *fid;
520 osp_fid_to_obdid(fid, &osp->opd_last_id);
521 osp->opd_gap_start_fid = *fid;
522 osp->opd_pre_used_fid = *fid;
523 osp->opd_pre_last_created_fid = *fid;
524 spin_unlock(&osp->opd_pre_lock);
527 osp_update_fldb_cache(env, osp, fid);
533 * Find IDs available in current sequence
535 * The function calculates the highest possible ID and the number of IDs
536 * available in the current sequence OSP is using. The number is limited
537 * artifically by the caller (grow param) and the number of IDs available
538 * in the sequence by nature. The function doesn't require an external
541 * \param[in] env LU environment provided by the caller
542 * \param[in] osp OSP device
543 * \param[in] fid FID the caller wants to start with
544 * \param[in] grow how many the caller wants
545 * \param[out] fid the highest calculated FID
546 * \param[out] grow the number of available IDs calculated
548 * \retval 0 on success, 1 - the sequence is empty
550 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
551 struct lu_fid *fid, int *grow)
553 struct osp_thread_info *osi = osp_env_info(env);
554 __u64 seq_width = osp->opd_pre_seq_width;
558 if (fid_is_idif(fid)) {
559 struct lu_fid *last_fid;
560 struct ost_id *oi = &osi->osi_oi;
563 spin_lock(&osp->opd_pre_lock);
564 last_fid = &osp->opd_pre_last_created_fid;
565 fid_to_ostid(last_fid, oi);
566 end = min(ostid_id(oi) + *grow, min(IDIF_MAX_OID, seq_width));
567 *grow = end - ostid_id(oi);
568 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
569 spin_unlock(&osp->opd_pre_lock);
571 if (*grow == 0 || rc)
574 ostid_to_fid(fid, oi, osp->opd_index);
578 spin_lock(&osp->opd_pre_lock);
579 *fid = osp->opd_pre_last_created_fid;
581 end = min((end + *grow), min(OBIF_MAX_OID, seq_width));
582 *grow = end - fid->f_oid;
583 fid->f_oid += end - fid->f_oid;
584 spin_unlock(&osp->opd_pre_lock);
586 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
587 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
589 return *grow > 0 ? 0 : 1;
593 * Prepare and send precreate RPC
595 * The function finds how many objects should be precreated. Then allocates,
596 * prepares and schedules precreate RPC synchronously. Upon reply the function
597 * wakes up the threads waiting for the new objects on this target. If the
598 * target wasn't able to create all the objects requested, then the next
599 * precreate will be asking for fewer objects (i.e. slow precreate down).
601 * \param[in] env LU environment provided by the caller
602 * \param[in] d OSP device
604 * \retval 0 on success
605 * \retval negative negated errno on error
607 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
609 struct osp_thread_info *oti = osp_env_info(env);
610 struct ptlrpc_request *req;
611 struct obd_import *imp;
612 struct ost_body *body;
614 struct lu_fid *fid = &oti->osi_fid;
617 /* don't precreate new objects till OST healthy and has free space */
618 if (unlikely(d->opd_pre_status)) {
619 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
620 d->opd_obd->obd_name, d->opd_pre_status);
625 * if not connection/initialization is compeleted, ignore
627 imp = d->opd_obd->u.cli.cl_import;
630 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
633 req->rq_request_portal = OST_CREATE_PORTAL;
635 /* We should not resend create request - anyway we will have delorphan
636 * and kill these objects.
637 * Only needed for MDS+OSS rolling upgrade interop with 2.16+older.
639 if (unlikely(!imp_connect_replay_create(imp)))
640 req->rq_no_delay = req->rq_no_resend = 1;
642 /* Delorphan happens only with a first MDT-OST connect. resend/replay
643 * handles objects creation on reconnects, no need to do delorhpan
647 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
649 ptlrpc_request_free(req);
653 spin_lock(&d->opd_pre_lock);
654 if (d->opd_force_creation)
655 d->opd_pre_create_count = OST_MIN_PRECREATE;
656 else if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
657 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
658 grow = d->opd_pre_create_count;
659 spin_unlock(&d->opd_pre_lock);
661 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
664 *fid = d->opd_pre_last_created_fid;
665 rc = osp_precreate_fids(env, d, fid, &grow);
667 /* Current seq has been used up*/
668 GOTO(out_req, rc = -ENOSPC);
670 if (!osp_is_fid_client(d)) {
671 /* Non-FID client will always send seq 0 because of
673 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
677 fid_to_ostid(fid, &body->oa.o_oi);
678 body->oa.o_valid = OBD_MD_FLGROUP;
680 ptlrpc_request_set_replen(req);
682 if (CFS_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
685 rc = ptlrpc_queue_wait(req);
687 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
690 /* have osp_precreate_reserve() to wait for repeat */
695 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
697 GOTO(out_req, rc = -EPROTO);
699 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
702 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
703 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
704 ": rc = %d\n", d->opd_obd->obd_name,
705 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
706 GOTO(out_req, rc = -ESTALE);
709 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
711 spin_lock(&d->opd_pre_lock);
713 /* the OST has not managed to create all the
714 * objects we asked for */
715 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
716 d->opd_pre_create_slow = 1;
718 /* the OST is able to keep up with the work,
719 * we could consider increasing create_count
720 * next time if needed */
721 d->opd_pre_create_slow = 0;
724 if ((body->oa.o_valid & OBD_MD_FLSIZE) && body->oa.o_size)
725 d->opd_pre_seq_width = body->oa.o_size;
727 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
728 fid_to_ostid(fid, &body->oa.o_oi);
730 d->opd_pre_last_created_fid = *fid;
731 d->opd_force_creation = false;
732 spin_unlock(&d->opd_pre_lock);
734 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
735 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
736 PFID(&d->opd_pre_last_created_fid));
738 /* now we can wakeup all users awaiting for objects */
739 osp_pre_update_status(d, rc);
741 ptlrpc_req_finished(req);
744 osp_update_fldb_cache(env, d, fid);
750 * Get last precreated object from target (OST)
752 * Sends synchronous RPC to the target (OST) to learn the last precreated
753 * object. This later is used to remove all unused objects (cleanup orphan
754 * procedure). Also, the next object after one we got will be used as a
755 * starting point for the new precreates.
757 * \param[in] env LU environment provided by the caller
758 * \param[in] d OSP device
759 * \param[in] update update or not update last used fid
761 * \retval 0 on success
762 * \retval negative negated errno on error
764 static int osp_get_lastfid_from_ost(const struct lu_env *env,
765 struct osp_device *d, bool update)
767 struct ptlrpc_request *req = NULL;
768 struct obd_import *imp;
769 struct lu_fid *last_fid;
774 imp = d->opd_obd->u.cli.cl_import;
777 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
781 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
782 sizeof(KEY_LAST_FID));
784 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
786 ptlrpc_request_free(req);
790 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
791 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
793 req->rq_no_delay = req->rq_no_resend = 1;
794 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
795 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
797 ptlrpc_request_set_replen(req);
799 rc = ptlrpc_queue_wait(req);
801 /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld),
802 * let sysadm sort this * out.
805 ptlrpc_set_import_active(imp, 0);
809 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
810 if (last_fid == NULL) {
811 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
812 GOTO(out, rc = -EPROTO);
815 if (!fid_is_sane(last_fid)) {
816 CERROR("%s: Got insane last_fid "DFID"\n",
817 d->opd_obd->obd_name, PFID(last_fid));
818 GOTO(out, rc = -EPROTO);
821 /* Only update the last used fid, if the OST has objects for
822 * this sequence, i.e. fid_oid > 0 */
823 if (fid_oid(last_fid) > 0 && update)
824 d->opd_last_used_fid = *last_fid;
826 if (fid_seq(last_fid) == fid_seq(&d->opd_last_used_fid)) {
827 if (fid_oid(last_fid) == 0 ||
828 (fid_seq_is_norm(fid_seq(last_fid)) &&
829 fid_oid(last_fid) == LUSTRE_FID_INIT_OID)) {
830 /* reformatted OST, it requires creation request
831 * to recreate objects
833 spin_lock(&d->opd_pre_lock);
834 d->opd_force_creation = true;
835 spin_unlock(&d->opd_pre_lock);
838 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
842 ptlrpc_req_finished(req);
847 * Cleanup orphans on OST
849 * This function is called in a contex of a dedicated thread handling
850 * all the precreation suff. The function waits till local recovery
851 * is complete, then identify all the unreferenced objects (orphans)
852 * using the highest ID referenced by a local and the highest object
853 * precreated by the target. The found range is a subject to removal
854 * using specially flagged RPC. During this process OSP is marked
855 * unavailable for new objects.
857 * \param[in] env LU environment provided by the caller
858 * \param[in] d OSP device
860 * \retval 0 on success
861 * \retval negative negated errno on error
863 static int osp_precreate_cleanup_orphans(struct lu_env *env,
864 struct osp_device *d)
866 struct osp_thread_info *osi = osp_env_info(env);
867 struct lu_fid *last_fid = &osi->osi_fid;
868 struct ptlrpc_request *req = NULL;
869 struct obd_import *imp = d->opd_obd->u.cli.cl_import;
870 struct ost_body *body;
871 int update_status = 0;
878 * Do cleanup orphans only with a first connection, after that
879 * all precreate requests uses resend/replay flags to support OST
880 * failover/reconnect.
882 if (d->opd_cleanup_orphans_done && imp_connect_replay_create(imp)) {
883 rc = osp_get_lastfid_from_ost(env, d, false);
887 * wait for local recovery to finish, so we can cleanup orphans
888 * orphans are all objects since "last used" (assigned), but
889 * there might be objects reserved and in some cases they won't
890 * be used. we can't cleanup them till we're sure they won't be
891 * used. also can't we allow new reservations because they may
892 * end up getting orphans being cleaned up below. so we block
893 * new reservations and wait till all reserved objects either
896 spin_lock(&d->opd_pre_lock);
897 d->opd_pre_recovering = 1;
898 spin_unlock(&d->opd_pre_lock);
900 * The locking above makes sure the opd_pre_reserved check below will
901 * catch all osp_precreate_reserve() calls who find
902 * "!opd_pre_recovering".
904 wait_event_idle(d->opd_pre_waitq,
905 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
906 !d->opd_pre_task || d->opd_got_disconnected);
907 if (!d->opd_pre_task || d->opd_got_disconnected)
908 GOTO(out, rc = -EAGAIN);
910 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
911 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
913 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val);
915 *last_fid = d->opd_last_used_fid;
916 /* The OSP should already get the valid seq now */
917 LASSERT(!fid_is_zero(last_fid));
918 if (fid_oid(&d->opd_last_used_fid) < 2) {
919 /* lastfid looks strange... ask OST */
920 rc = osp_get_lastfid_from_ost(env, d, true);
925 imp = d->opd_obd->u.cli.cl_import;
928 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
930 GOTO(out, rc = -ENOMEM);
932 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
934 ptlrpc_request_free(req);
939 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
941 GOTO(out, rc = -EPROTO);
943 body->oa.o_flags = OBD_FL_DELORPHAN;
944 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
946 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
948 ptlrpc_request_set_replen(req);
950 /* Don't resend the delorphan req */
951 req->rq_no_resend = req->rq_no_delay = 1;
953 rc = ptlrpc_queue_wait(req);
959 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
961 GOTO(out, rc = -EPROTO);
964 * OST provides us with id new pool starts from in body->oa.o_id
966 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
968 spin_lock(&d->opd_pre_lock);
969 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
971 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
972 d->opd_pre_last_created_fid = d->opd_last_used_fid;
974 d->opd_pre_create_count = OST_MIN_PRECREATE;
975 d->opd_pre_last_created_fid = *last_fid;
978 * This empties the pre-creation pool and effectively blocks any new
981 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
982 d->opd_pre_create_slow = 0;
983 if ((body->oa.o_valid & OBD_MD_FLSIZE) && body->oa.o_size)
984 d->opd_pre_seq_width = body->oa.o_size;
985 spin_unlock(&d->opd_pre_lock);
987 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
988 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
989 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
992 ptlrpc_req_finished(req);
996 * If rc is zero, the pre-creation window should have been emptied.
997 * Since waking up the herd would be useless without pre-created
998 * objects, we defer the signal to osp_precreate_send() in that case.
1001 if (update_status) {
1002 CERROR("%s: cannot cleanup orphans: rc = %d\n",
1003 d->opd_obd->obd_name, rc);
1004 /* we can't proceed from here, OST seem to
1005 * be in a bad shape, better to wait for
1006 * a new instance of the server and repeat
1007 * from the beginning. notify possible waiters
1008 * this OSP isn't quite functional yet */
1009 osp_pre_update_status(d, rc);
1011 wake_up_all(&d->opd_pre_user_waitq);
1014 spin_lock(&d->opd_pre_lock);
1015 d->opd_pre_recovering = 0;
1016 spin_unlock(&d->opd_pre_lock);
1017 d->opd_cleanup_orphans_done = true;
1024 * Update precreate status using statfs data
1026 * The function decides whether this OSP should be used for new objects.
1027 * IOW, whether this OST is used up or has some free space. Cached statfs
1028 * data is used to make this decision. If the latest result of statfs
1029 * request (rc argument) is not success, then just mark OSP unavailable
1032 * The new statfs data is passed in \a msfs and needs to be stored into
1033 * opd_statfs, but only after the various flags in os_state are set, so
1034 * that the new statfs data is not visible without appropriate flags set.
1035 * As such, there is no need to clear the flags here, since this is called
1036 * with new statfs data, and they should not be cleared if sent from OST.
1038 * Add a bit of hysteresis so this flag isn't continually flapping, and
1039 * ensure that new files don't get extremely fragmented due to only a
1040 * small amount of available space in the filesystem. We want to set
1041 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1042 * reserved size free, and still copy them from the old state when there
1043 * is less than 2*reserved size free space or inodes.
1045 * \param[in] d OSP device
1046 * \param[in] msfs statfs data
1048 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1050 u32 old_state = d->opd_statfs.os_state;
1053 /* statfs structure not initialized yet */
1054 if (unlikely(!msfs->os_type))
1057 /* if the low and high watermarks have not been initialized yet */
1058 if (unlikely(d->opd_reserved_mb_high == 0 &&
1059 d->opd_reserved_mb_low == 0)) {
1060 /* Use ~0.1% by default to disable object allocation,
1061 * and ~0.2% to enable, size in MB, set both watermark
1063 spin_lock(&d->opd_pre_lock);
1064 if (d->opd_reserved_mb_high == 0 &&
1065 d->opd_reserved_mb_low == 0) {
1066 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1067 msfs->os_blocks) >> 20;
1068 if (d->opd_reserved_mb_low < 1)
1069 d->opd_reserved_mb_low = 1;
1070 d->opd_reserved_mb_high =
1071 (d->opd_reserved_mb_low << 1) + 1;
1073 spin_unlock(&d->opd_pre_lock);
1076 if (unlikely(d->opd_reserved_ino_high == 0 &&
1077 d->opd_reserved_ino_low == 0)) {
1078 /* Use ~0.0001% by default to disallow distributed transactions,
1079 * and ~0.0002% to allow, set both watermark
1081 spin_lock(&d->opd_pre_lock);
1082 if (d->opd_reserved_ino_high == 0 &&
1083 d->opd_reserved_ino_low == 0) {
1084 d->opd_reserved_ino_low = msfs->os_ffree >> 20;
1085 if (d->opd_reserved_ino_low < 32)
1086 d->opd_reserved_ino_low = 32;
1087 d->opd_reserved_ino_high =
1088 (d->opd_reserved_ino_low << 1) + 1;
1090 spin_unlock(&d->opd_pre_lock);
1093 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1094 if (msfs->os_ffree < d->opd_reserved_ino_low)
1095 msfs->os_state |= OS_STATFS_ENOINO;
1096 else if (msfs->os_ffree <= d->opd_reserved_ino_high)
1097 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1098 /* else don't clear flags in new msfs->os_state sent from OST */
1100 if (available_mb < d->opd_reserved_mb_low)
1101 msfs->os_state |= OS_STATFS_ENOSPC;
1102 else if (available_mb <= d->opd_reserved_mb_high)
1103 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1104 /* else don't clear flags in new msfs->os_state sent from OST */
1107 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1108 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1109 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1110 msfs->os_files, msfs->os_ffree, msfs->os_state,
1111 d->opd_pre ? d->opd_pre_status : 0);
1116 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1117 d->opd_pre_status = -ENOSPC;
1118 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1119 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1120 d->opd_obd->obd_name, msfs->os_state,
1122 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1123 atomic_read(&d->opd_sync_changes),
1124 atomic_read(&d->opd_sync_rpcs_in_progress));
1125 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1126 d->opd_pre_status = 0;
1127 spin_lock(&d->opd_pre_lock);
1128 d->opd_pre_create_slow = 0;
1129 d->opd_pre_create_count = OST_MIN_PRECREATE;
1130 spin_unlock(&d->opd_pre_lock);
1131 wake_up(&d->opd_pre_waitq);
1134 "%s: available: state=%x: rc = %d\n",
1135 d->opd_obd->obd_name, msfs->os_state,
1138 /* we only get here if rc == 0 in the caller */
1139 d->opd_pre_status = 0;
1142 /* Object precreation skipped on OST if manually disabled */
1143 if (d->opd_pre_max_create_count == 0)
1144 msfs->os_state |= OS_STATFS_NOPRECREATE;
1145 /* else don't clear flags in new msfs->os_state sent from OST */
1148 /* copy only new statfs state to make it visible to MDS threads */
1149 if (&d->opd_statfs != msfs)
1150 d->opd_statfs = *msfs;
1154 * Initialize FID for precreation
1156 * For a just created new target, a new sequence should be taken.
1157 * The function checks there is no IDIF in use (if the target was
1158 * added with the older version of Lustre), then requests a new
1159 * sequence from FLDB using the regular protocol. Then this new
1160 * sequence is stored on a persisten storage synchronously to prevent
1161 * possible object leakage (for the detail see the description for
1162 * osp_precreate_rollover_new_seq()).
1164 * \param[in] osp OSP device
1166 * \retval 0 on success
1167 * \retval negative negated errno on error
1169 static int osp_init_pre_fid(struct lu_env *env, struct osp_device *osp)
1171 struct osp_thread_info *osi;
1172 struct lu_client_seq *cli_seq;
1173 struct lu_fid *last_fid;
1177 LASSERT(osp->opd_pre != NULL);
1179 /* Let's check if the current last_seq/fid is valid,
1180 * otherwise request new sequence from the controller */
1181 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1182 /* Non-MDT0 can only use normal sequence for
1184 if (fid_is_norm(&osp->opd_last_used_fid))
1187 /* Initially MDT0 will start with IDIF, after
1188 * that it will request new sequence from the
1190 if (fid_is_idif(&osp->opd_last_used_fid) ||
1191 fid_is_norm(&osp->opd_last_used_fid))
1195 if (!fid_is_zero(&osp->opd_last_used_fid))
1196 CWARN("%s: invalid last used fid "DFID
1197 ", try to get new sequence.\n",
1198 osp->opd_obd->obd_name,
1199 PFID(&osp->opd_last_used_fid));
1201 osi = osp_env_info(env);
1202 last_fid = &osi->osi_fid;
1204 /* For a freshed fs, it will allocate a new sequence first */
1205 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1206 cli_seq = osp->opd_obd->u.cli.cl_seq;
1207 rc = seq_client_get_seq(env, cli_seq, &last_fid->f_seq);
1209 CERROR("%s: alloc fid error: rc = %d\n",
1210 osp->opd_obd->obd_name, rc);
1214 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1216 last_fid->f_oid = 1;
1217 last_fid->f_ver = 0;
1219 spin_lock(&osp->opd_pre_lock);
1220 osp->opd_last_used_fid = *last_fid;
1221 osp->opd_pre_used_fid = *last_fid;
1222 osp->opd_pre_last_created_fid = *last_fid;
1223 spin_unlock(&osp->opd_pre_lock);
1224 rc = osp_write_last_oid_seq_files(env, osp, last_fid, 1);
1226 CERROR("%s: write fid error: rc = %d\n",
1227 osp->opd_obd->obd_name, rc);
1235 struct osp_device *opta_dev;
1236 struct lu_env opta_env;
1237 struct completion *opta_started;
1240 * The core of precreate functionality
1242 * The function implements the main precreation loop. Basically it
1243 * involves connecting to the target, precerate FID initialization,
1244 * identifying and removing orphans, then serving precreation. As
1245 * part of the latter, the thread is responsible for statfs data
1246 * updates. The precreation is mostly driven by another threads
1247 * asking for new OST objects - those askers wake the thread when
1248 * the number of precreated objects reach low watermark.
1249 * After a disconnect, the sequence above repeats. This is keep going
1250 * until the thread is requested to stop.
1252 * \param[in] _arg private data the thread (OSP device to handle)
1254 * \retval 0 on success
1255 * \retval negative negated errno on error
1257 static int osp_precreate_thread(void *_args)
1259 struct opt_args *args = _args;
1260 struct osp_device *d = args->opta_dev;
1261 struct lu_env *env = &args->opta_env;
1266 complete(args->opta_started);
1268 /* wait for connection from the layers above */
1269 wait_event_idle(d->opd_pre_waitq,
1270 kthread_should_stop() ||
1271 d->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1273 while (!kthread_should_stop()) {
1275 * need to be connected to OST
1277 while (!kthread_should_stop()) {
1278 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1279 d->opd_imp_connected &&
1280 !d->opd_got_disconnected)
1282 wait_event_idle(d->opd_pre_waitq,
1283 kthread_should_stop() ||
1284 d->opd_new_connection);
1286 if (!d->opd_new_connection)
1289 CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
1291 d->opd_new_connection = 0;
1292 d->opd_got_disconnected = 0;
1296 if (kthread_should_stop())
1300 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1301 LASSERT(d->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1303 /* Init fid for osp_precreate if necessary */
1304 rc = osp_init_pre_fid(env, d);
1306 class_export_put(d->opd_exp);
1307 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1308 CERROR("%s: init pre fid error: rc = %d\n",
1309 d->opd_obd->obd_name, rc);
1314 if (osp_statfs_update(env, d)) {
1315 if (wait_event_idle_timeout(d->opd_pre_waitq,
1316 kthread_should_stop(),
1317 cfs_time_seconds(5)) == 0)
1318 l_wait_event_abortable(
1320 kthread_should_stop());
1326 * Clean up orphans or recreate missing objects.
1328 rc = osp_precreate_cleanup_orphans(env, d);
1330 schedule_timeout_interruptible(cfs_time_seconds(1));
1336 * connected, can handle precreates now
1338 while (!kthread_should_stop()) {
1339 wait_event_idle(d->opd_pre_waitq,
1340 kthread_should_stop() ||
1341 (osp_precreate_near_empty(env, d) &&
1342 !(osp_precreate_end_seq(env, d) &&
1343 osp_objs_precreated(env, d) != 0)) ||
1344 osp_statfs_need_update(d) ||
1345 d->opd_got_disconnected);
1347 if (kthread_should_stop())
1350 /* something happened to the connection
1351 * have to start from the beginning */
1352 if (d->opd_got_disconnected)
1355 if (osp_statfs_need_update(d))
1356 if (osp_statfs_update(env, d))
1359 if (d->opd_pre == NULL)
1362 /* To avoid handling different seq in precreate/orphan
1363 * cleanup, it will hold precreate until current seq is
1365 if (unlikely(osp_precreate_end_seq(env, d))) {
1366 if (osp_objs_precreated(env, d) == 0) {
1367 rc = osp_precreate_rollover_new_seq(env, d);
1375 if (osp_precreate_near_empty(env, d)) {
1376 rc = osp_precreate_send(env, d);
1377 /* osp_precreate_send() sets opd_pre_status
1378 * in case of error, that prevent the using of
1380 if (rc < 0 && rc != -ENOSPC &&
1381 rc != -ETIMEDOUT && rc != -ENOTCONN)
1382 CERROR("%s: cannot precreate objects:"
1384 d->opd_obd->obd_name, rc);
1396 * Check when to stop to wait for precreate objects.
1398 * The caller wanting a new OST object can't wait undefinitely. The
1399 * function checks for few conditions including available new OST
1400 * objects, disconnected OST, lack of space with no pending destroys,
1401 * etc. IOW, it checks whether the current OSP state is good to keep
1402 * waiting or it's better to give up.
1404 * \param[in] env LU environment provided by the caller
1405 * \param[in] d OSP device
1407 * \retval 0 - keep waiting, 1 - no luck
1409 static int osp_precreate_ready_condition(const struct lu_env *env,
1410 struct osp_device *d)
1412 /* Bail out I/O fails to OST */
1413 if (d->opd_pre_status != 0 &&
1414 d->opd_pre_status != -EAGAIN &&
1415 d->opd_pre_status != -ENODEV &&
1416 d->opd_pre_status != -ENOTCONN &&
1417 d->opd_pre_status != -ENOSPC) {
1419 if (d->opd_pre_status != -EIO)
1420 CERROR("%s: precreate failed opd_pre_status %d\n",
1421 d->opd_obd->obd_name, d->opd_pre_status);
1425 if (d->opd_pre_recovering || d->opd_force_creation)
1428 /* ready if got enough precreated objects */
1429 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1430 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1433 /* ready if OST reported no space and no destroys in progress */
1434 if (atomic_read(&d->opd_sync_changes) +
1435 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1436 d->opd_pre_status == -ENOSPC)
1443 * Reserve object in precreate pool
1445 * When the caller wants to create a new object on this target (target
1446 * represented by the given OSP), it should declare this intention using
1447 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1448 * to reserve an object in the existing precreated pool or wait up to
1449 * obd_timeout for the available object to appear in the pool (a dedicated
1450 * thread will be doing real precreation in background). The object can be
1451 * consumed later with osp_precreate_get_fid() or be released with call to
1452 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1453 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1454 * If the space on the target is short and there is a pending object destroy,
1455 * then the function forces local commit to speedup space release (see
1456 * osp_sync.c for the details).
1458 * \param[in] env LU environment provided by the caller
1459 * \param[in] d OSP device
1461 * \retval 0 on success
1462 * \retval -ENOSPC when no space on OST
1463 * \retval -EAGAIN try later, slow precreation in progress
1464 * \retval -EIO when no access to OST
1466 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
1469 time64_t expire = ktime_get_seconds() + obd_timeout;
1470 int precreated, rc, synced = 0;
1474 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1475 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1476 PFID(&d->opd_pre_used_fid));
1478 /* opd_pre_max_create_count 0 to not use specified OST. */
1479 if (d->opd_pre_max_create_count == 0)
1484 * - preallocation is done
1485 * - no free space expected soon
1486 * - can't connect to OST for too long (obd_timeout)
1487 * - OST can allocate fid sequence.
1489 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1490 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1493 * increase number of precreations
1495 precreated = osp_objs_precreated(env, d);
1496 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1497 d->opd_pre_create_slow == 0 &&
1498 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1499 spin_lock(&d->opd_pre_lock);
1500 d->opd_pre_create_slow = 1;
1501 d->opd_pre_create_count *= 2;
1502 spin_unlock(&d->opd_pre_lock);
1505 spin_lock(&d->opd_pre_lock);
1506 precreated = osp_objs_precreated(env, d);
1507 if (!d->opd_pre_recovering && !d->opd_force_creation) {
1508 if (precreated > d->opd_pre_reserved) {
1509 d->opd_pre_reserved++;
1510 spin_unlock(&d->opd_pre_lock);
1514 * XXX: don't wake up if precreation
1517 if (osp_precreate_near_empty_nolock(env, d) &&
1518 !osp_precreate_end_seq_nolock(env, d))
1519 wake_up(&d->opd_pre_waitq);
1522 } else if (unlikely(precreated &&
1523 osp_precreate_end_seq_nolock(env, d))) {
1525 * precreate pool is reaching the end of the
1526 * current seq, and doesn't have enough objects
1529 spin_unlock(&d->opd_pre_lock);
1533 spin_unlock(&d->opd_pre_lock);
1536 * all precreated objects have been used and no-space
1537 * status leave us no chance to succeed very soon
1538 * but if there is destroy in progress, then we should
1539 * wait till that is done - some space might be released
1541 if (unlikely(rc == -ENOSPC)) {
1542 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1543 /* force local commit to release space */
1544 dt_commit_async(env, d->opd_storage);
1545 osp_sync_check_for_work(d);
1548 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1549 /* just wait till destroys are done
1550 * see wait_event_idle_timeout() below
1553 if (atomic_read(&d->opd_sync_changes) +
1554 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1555 /* no hope for free space */
1560 /* XXX: don't wake up if precreation is in progress */
1561 wake_up(&d->opd_pre_waitq);
1563 if (ktime_get_seconds() >= expire) {
1569 LASSERT(d->opd_pre);
1574 CDEBUG(D_INFO, "%s: Sleeping on objects\n",
1575 d->opd_obd->obd_name);
1576 if (wait_event_idle_timeout(
1577 d->opd_pre_user_waitq,
1578 osp_precreate_ready_condition(env, d),
1579 cfs_time_seconds(obd_timeout)) == 0) {
1581 "%s: slow creates, last="DFID", next="DFID", "
1582 "reserved=%llu, sync_changes=%u, "
1583 "sync_rpcs_in_progress=%d, status=%d\n",
1584 d->opd_obd->obd_name,
1585 PFID(&d->opd_pre_last_created_fid),
1586 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1587 atomic_read(&d->opd_sync_changes),
1588 atomic_read(&d->opd_sync_rpcs_in_progress),
1591 CDEBUG(D_INFO, "%s: Waked up, status=%d\n",
1592 d->opd_obd->obd_name, d->opd_pre_status);
1600 * Get a FID from precreation pool
1602 * The function is a companion for osp_precreate_reserve() - it assigns
1603 * a specific FID from the precreate. The function should be called only
1604 * if the call to osp_precreate_reserve() was successful. The function
1605 * updates a local storage to remember the highest object ID referenced
1606 * by the node in the given sequence.
1608 * A very importan details: this is supposed to be called once the
1609 * transaction is started, so on-disk update will be atomic with the
1610 * data (like LOVEA) refering this object. Then the object won't be leaked:
1611 * either it's referenced by the committed transaction or it's a subject
1612 * to the orphan cleanup procedure.
1614 * \param[in] env LU environment provided by the caller
1615 * \param[in] d OSP device
1616 * \param[out] fid generated FID
1618 * \retval 0 on success
1619 * \retval negative negated errno on error
1621 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1624 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1626 /* grab next id from the pool */
1627 spin_lock(&d->opd_pre_lock);
1629 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1630 &d->opd_pre_last_created_fid) < 0,
1631 "next fid "DFID" > last created fid "DFID"\n",
1632 PFID(&d->opd_pre_used_fid),
1633 PFID(&d->opd_pre_last_created_fid));
1635 /* Non-IDIF FIDs shouldn't get here with OID == OBIF_MAX_OID. For IDIF,
1636 * f_oid wraps and "f_seq" (holding high 16 bits of ID) needs increment
1638 if (fid_is_idif(pre_used_fid) &&
1639 unlikely(fid_oid(pre_used_fid) == OBIF_MAX_OID)) {
1641 __u32 idx = fid_idif_ost_idx(pre_used_fid);
1643 fid_to_ostid(pre_used_fid, &oi);
1645 ostid_to_fid(pre_used_fid, &oi, idx);
1647 pre_used_fid->f_oid++;
1650 memcpy(fid, pre_used_fid, sizeof(*fid));
1651 d->opd_pre_reserved--;
1653 * last_used_id must be changed along with getting new id otherwise
1654 * we might miscalculate gap causing object loss or leak
1656 osp_update_last_fid(d, fid);
1657 spin_unlock(&d->opd_pre_lock);
1660 * probably main thread suspended orphan cleanup till
1661 * all reservations are released, see comment in
1662 * osp_precreate_thread() just before orphan cleanup
1664 if (unlikely(d->opd_pre_reserved == 0 &&
1665 (d->opd_pre_recovering || d->opd_pre_status)))
1666 wake_up(&d->opd_pre_waitq);
1672 * Set size regular attribute on an object
1674 * When a striping is created late, it's possible that size is already
1675 * initialized on the file. Then the new striping should inherit size
1676 * from the file. The function sets size on the object using the regular
1677 * protocol (OST_PUNCH).
1678 * XXX: should be re-implemented using OUT ?
1680 * \param[in] env LU environment provided by the caller
1681 * \param[in] dt object
1682 * \param[in] size size to set.
1684 * \retval 0 on success
1685 * \retval negative negated errno on error
1687 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1690 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1691 struct ptlrpc_request *req = NULL;
1692 struct obd_import *imp;
1693 struct ost_body *body;
1694 struct obdo *oa = NULL;
1699 imp = d->opd_obd->u.cli.cl_import;
1702 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1706 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1708 ptlrpc_request_free(req);
1713 * XXX: decide how do we do here with resend
1714 * if we don't resend, then client may see wrong file size
1715 * if we do resend, then MDS thread can get stuck for quite long
1716 * and if we don't resend, then client will also get -EAGAIN !!
1717 * (see LU-7975 and sanity/test_27F use cases)
1718 * but let's decide not to resend/delay this truncate request to OST
1719 * and allow Client to decide to resend, in a less agressive way from
1720 * after_reply(), by returning -EINPROGRESS instead of
1721 * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
1722 * end of this routine
1724 req->rq_no_resend = req->rq_no_delay = 1;
1726 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1727 ptlrpc_at_set_req_timeout(req);
1731 GOTO(out, rc = -ENOMEM);
1733 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1736 oa->o_blocks = OBD_OBJECT_EOF;
1737 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1738 OBD_MD_FLID | OBD_MD_FLGROUP;
1740 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1742 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1744 /* XXX: capa support? */
1745 /* osc_pack_capa(req, body, capa); */
1747 ptlrpc_request_set_replen(req);
1749 rc = ptlrpc_queue_wait(req);
1751 /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
1752 * since we have decided not to resend/delay, but this could
1753 * lead to wrong size to be seen at Client side and even process
1754 * trying to open to exit/fail if not itself handling -EAGAIN.
1755 * So it should be better to return -EINPROGRESS instead and
1756 * leave the decision to resend at Client side in after_reply()
1758 if (rc == -EAGAIN) {
1760 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1761 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1764 CERROR("can't punch object: %d\n", rc);
1768 ptlrpc_req_finished(req);
1775 * Initialize precreation functionality of OSP
1777 * Prepares all the internal structures and starts the precreate thread
1779 * \param[in] d OSP device
1781 * \retval 0 on success
1782 * \retval negative negated errno on error
1784 int osp_init_precreate(struct osp_device *d)
1788 OBD_ALLOC_PTR(d->opd_pre);
1789 if (d->opd_pre == NULL)
1792 /* initially precreation isn't ready */
1793 init_waitqueue_head(&d->opd_pre_user_waitq);
1794 d->opd_pre_status = -EAGAIN;
1795 fid_zero(&d->opd_pre_used_fid);
1796 d->opd_pre_used_fid.f_oid = 1;
1797 fid_zero(&d->opd_pre_last_created_fid);
1798 d->opd_pre_last_created_fid.f_oid = 1;
1800 d->opd_pre_reserved = 0;
1801 d->opd_pre_seq_width = LUSTRE_DATA_SEQ_MAX_WIDTH;
1802 d->opd_got_disconnected = 1;
1803 d->opd_pre_create_slow = 0;
1804 d->opd_pre_create_count = OST_MIN_PRECREATE;
1805 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1806 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1807 d->opd_reserved_mb_high = 0;
1808 d->opd_reserved_mb_low = 0;
1809 d->opd_cleanup_orphans_done = false;
1810 d->opd_force_creation = false;
1816 * Finish precreate functionality of OSP
1819 * Asks all the activity (the thread, update timer) to stop, then
1820 * wait till that is done.
1822 * \param[in] d OSP device
1824 void osp_precreate_fini(struct osp_device *d)
1828 if (d->opd_pre == NULL)
1831 OBD_FREE_PTR(d->opd_pre);
1837 int osp_init_statfs(struct osp_device *d)
1839 struct task_struct *task;
1840 struct opt_args *args;
1841 DECLARE_COMPLETION_ONSTACK(started);
1846 spin_lock_init(&d->opd_pre_lock);
1847 init_waitqueue_head(&d->opd_pre_waitq);
1850 * Initialize statfs-related things
1852 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1853 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1854 1000 * NSEC_PER_SEC);
1855 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1857 ktime_to_ns(d->opd_statfs_fresh_till));
1858 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1859 (unsigned long)d, 0);
1861 if (d->opd_storage->dd_rdonly)
1864 OBD_ALLOC_PTR(args);
1868 args->opta_started = &started;
1869 rc = lu_env_init(&args->opta_env,
1870 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1872 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1879 * start thread handling precreation and statfs updates
1881 task = kthread_create(osp_precreate_thread, args,
1882 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1884 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1885 lu_env_fini(&args->opta_env);
1887 RETURN(PTR_ERR(task));
1889 d->opd_pre_task = task;
1890 wake_up_process(task);
1891 wait_for_completion(&started);
1896 void osp_statfs_fini(struct osp_device *d)
1898 struct task_struct *task = d->opd_pre_task;
1901 timer_delete(&d->opd_statfs_timer);
1903 d->opd_pre_task = NULL;