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);
85 wake_up(&d->opd_pre_waitq);
88 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs);
91 * The function updates current precreation status if broken, and
92 * updates that cached statfs state if functional, then wakes up waiters.
93 * We don't clear opd_pre_status directly here, but rather leave this
94 * to osp_pre_update_msfs() to do if everything is OK so that we don't
95 * have a race to clear opd_pre_status and then set it to -ENOSPC again.
97 * \param[in] d OSP device
98 * \param[in] msfs statfs data
99 * \param[in] rc new precreate status for device \a d
101 static void osp_pre_update_status_msfs(struct osp_device *d,
102 struct obd_statfs *msfs, int rc)
104 CDEBUG(D_INFO, "%s: Updating status = %d\n", d->opd_obd->obd_name, rc);
106 d->opd_pre_status = rc;
108 osp_pre_update_msfs(d, msfs);
110 wake_up_all(&d->opd_pre_user_waitq);
113 /* Pass in the old statfs data in case the limits have changed */
114 void osp_pre_update_status(struct osp_device *d, int rc)
116 osp_pre_update_status_msfs(d, &d->opd_statfs, rc);
121 * RPC interpret callback for OST_STATFS RPC
123 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
124 * replied by the target. It's used to maintain statfs cache for the target.
125 * The function fills data from the reply if successful and schedules another
128 * \param[in] env LU environment provided by the caller
129 * \param[in] req RPC replied
130 * \param[in] aa callback data
131 * \param[in] rc RPC result
133 * \retval 0 on success
134 * \retval negative negated errno on error
136 static int osp_statfs_interpret(const struct lu_env *env,
137 struct ptlrpc_request *req, void *args, int rc)
139 union ptlrpc_async_args *aa = args;
140 struct obd_import *imp = req->rq_import;
141 struct obd_statfs *msfs;
142 struct obd_statfs *sfs;
143 struct osp_device *d;
148 aa = ptlrpc_req_async_args(aa, req);
149 d = aa->pointer_arg[0];
155 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
157 GOTO(out, rc = -EPROTO);
160 osp_pre_update_status_msfs(d, msfs, 0);
162 d->opd_statfs = *msfs;
164 /* schedule next update */
165 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
166 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
167 mod_timer(&d->opd_statfs_timer,
168 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
169 d->opd_statfs_update_in_progress = 0;
171 sfs = &d->opd_statfs;
172 CDEBUG(D_CACHE, "%s (%p): %llu blocks, %llu free, %llu avail, "
173 "%u bsize, %u reserved mb low, %u reserved mb high,"
174 "%llu files, %llu free files\n", d->opd_obd->obd_name, d,
175 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail, sfs->os_bsize,
176 d->opd_reserved_mb_low, d->opd_reserved_mb_high,
177 sfs->os_files, sfs->os_ffree);
181 /* couldn't update statfs, try again with a small delay */
182 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
183 d->opd_statfs_update_in_progress = 0;
184 if (d->opd_pre && d->opd_pre_task)
185 wake_up(&d->opd_pre_waitq);
187 if (req->rq_import_generation == imp->imp_generation)
188 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
189 d->opd_obd->obd_name, rc);
194 * Send OST_STATFS RPC
196 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
197 * Also disables scheduled updates as times OSP may need to refresh
198 * statfs data before expiration. The function doesn't block, instead
199 * an interpretation callback osp_statfs_interpret() is used.
201 * \param[in] d OSP device
203 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
205 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
206 struct ptlrpc_request *req;
207 struct obd_import *imp;
208 union ptlrpc_async_args *aa;
213 CDEBUG(D_CACHE, "going to update statfs\n");
215 imp = d->opd_obd->u.cli.cl_import;
218 req = ptlrpc_request_alloc(imp,
219 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
223 rc = ptlrpc_request_pack(req,
224 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
225 d->opd_pre ? OST_STATFS : MDS_STATFS);
227 ptlrpc_request_free(req);
230 ptlrpc_request_set_replen(req);
232 req->rq_request_portal = OST_CREATE_PORTAL;
233 ptlrpc_at_set_req_timeout(req);
235 req->rq_interpret_reply = osp_statfs_interpret;
236 aa = ptlrpc_req_async_args(aa, req);
237 aa->pointer_arg[0] = d;
240 * no updates till reply
242 del_timer(&d->opd_statfs_timer);
243 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
244 d->opd_statfs_update_in_progress = 1;
246 ptlrpcd_add_req(req);
248 /* we still want to sync changes if no new changes are coming */
249 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
252 if (atomic_read(&d->opd_sync_changes)) {
255 th = dt_trans_create(env, d->opd_storage);
257 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
260 rc = dt_trans_start_local(env, d->opd_storage, th);
262 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
263 d->opd_obd->obd_name,
264 atomic_read(&d->opd_sync_changes));
265 osp_sync_add_commit_cb_1s(env, d, th);
267 dt_trans_stop(env, d->opd_storage, th);
275 * Schedule an immediate update for statfs data
277 * If cached statfs data claim no free space, but OSP has got a request to
278 * destroy an object (so release some space probably), then we may need to
279 * refresh cached statfs data sooner than planned. The function checks there
280 * is no statfs update going and schedules immediate update if so.
281 * XXX: there might be a case where removed object(s) do not add free space (empty
282 * object). If the number of such deletions is high, then we can start to update
283 * statfs too often causing a RPC storm. some throttling is needed...
285 * \param[in] d OSP device where statfs data needs to be refreshed
287 void osp_statfs_need_now(struct osp_device *d)
289 if (!d->opd_statfs_update_in_progress) {
291 * if current status is -ENOSPC (lack of free space on OST)
292 * then we should poll OST immediately once object destroy
295 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
296 del_timer(&d->opd_statfs_timer);
297 wake_up(&d->opd_pre_waitq);
302 * Return number of precreated objects
304 * A simple helper to calculate the number of precreated objects on the device.
306 * \param[in] env LU environment provided by the caller
307 * \param[in] osp OSP device
309 * \retval the number of the precreated objects
311 static inline int osp_objs_precreated(const struct lu_env *env,
312 struct osp_device *osp)
314 return osp_fid_diff(&osp->opd_pre_last_created_fid,
315 &osp->opd_pre_used_fid);
319 * Check pool of precreated objects is nearly empty
321 * We should not wait till the pool of the precreated objects is exhausted,
322 * because then there will be a long period of OSP being unavailable for the
323 * new creations due to lenghty precreate RPC. Instead we ask for another
324 * precreation ahead and hopefully have it ready before the current pool is
325 * empty. Notice this function relies on an external locking.
327 * \param[in] env LU environment provided by the caller
328 * \param[in] d OSP device
330 * \retval 0 - current pool is good enough, 1 - time to precreate
332 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
333 struct osp_device *d)
335 int window = osp_objs_precreated(env, d);
337 /* don't consider new precreation till OST is healty and
339 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
340 (d->opd_pre_status == 0));
344 * Check pool of precreated objects
346 * This is protected version of osp_precreate_near_empty_nolock(), check that
349 * \param[in] env LU environment provided by the caller
350 * \param[in] d OSP device
352 * \retval 0 - current pool is good enough, 1 - time to precreate
354 static inline int osp_precreate_near_empty(const struct lu_env *env,
355 struct osp_device *d)
359 if (d->opd_pre == NULL)
362 /* XXX: do we really need locking here? */
363 spin_lock(&d->opd_pre_lock);
364 rc = osp_precreate_near_empty_nolock(env, d);
365 spin_unlock(&d->opd_pre_lock);
370 * Check given sequence is empty
372 * Returns a binary result whether the given sequence has some IDs left
373 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
374 * version of that function.
376 * \param[in] env LU environment provided by the caller
377 * \param[in] osp OSP device
379 * \retval 0 - current sequence has no IDs, 1 - otherwise
381 static inline int osp_create_end_seq(const struct lu_env *env,
382 struct osp_device *osp)
384 struct lu_fid *fid = &osp->opd_pre_used_fid;
387 spin_lock(&osp->opd_pre_lock);
388 rc = osp_fid_end_seq(env, fid);
389 spin_unlock(&osp->opd_pre_lock);
394 * Write FID into into last_oid/last_seq file
396 * The function stores the sequence and the in-sequence id into two dedicated
397 * files. The sync argument can be used to request synchronous commit, so the
398 * function won't return until the updates are committed.
400 * \param[in] env LU environment provided by the caller
401 * \param[in] osp OSP device
402 * \param[in] fid fid where sequence/id is taken
403 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
405 * \retval 0 on success
406 * \retval negative negated errno on error
408 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
409 struct lu_fid *fid, int sync)
411 struct osp_thread_info *oti = osp_env_info(env);
412 struct lu_buf *lb_oid = &oti->osi_lb;
413 struct lu_buf *lb_oseq = &oti->osi_lb2;
421 if (osp->opd_storage->dd_rdonly)
424 /* Note: through f_oid is only 32 bits, it will also write 64 bits
425 * for oid to keep compatibility with the previous version. */
427 osp_objid_buf_prep(lb_oid, &oid_off,
428 &oid, osp->opd_index);
430 osp_objseq_buf_prep(lb_oseq, &oseq_off,
431 &fid->f_seq, osp->opd_index);
433 th = dt_trans_create(env, osp->opd_storage);
438 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
439 lb_oid, oid_off, th);
443 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
444 lb_oseq, oseq_off, th);
448 rc = dt_trans_start_local(env, osp->opd_storage, th);
452 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
455 CERROR("%s: can not write to last seq file: rc = %d\n",
456 osp->opd_obd->obd_name, rc);
459 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
462 CERROR("%s: can not write to last seq file: rc = %d\n",
463 osp->opd_obd->obd_name, rc);
467 dt_trans_stop(env, osp->opd_storage, th);
472 * Switch to another sequence
474 * When a current sequence has no available IDs left, OSP has to switch to
475 * another new sequence. OSP requests it using the regular FLDB protocol
476 * and stores synchronously before that is used in precreated. This is needed
477 * to basically have the sequences referenced (not orphaned), otherwise it's
478 * possible that OST has some objects precreated and the clients have data
479 * written to it, but after MDT failover nobody refers those objects and OSP
480 * has no idea that the sequence need cleanup to be done.
481 * While this is very expensive operation, it's supposed to happen very very
482 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
484 * \param[in] env LU environment provided by the caller
485 * \param[in] osp OSP device
487 * \retval 0 on success
488 * \retval negative negated errno on error
490 static int osp_precreate_rollover_new_seq(struct lu_env *env,
491 struct osp_device *osp)
493 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
494 struct lu_fid *last_fid = &osp->opd_last_used_fid;
498 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
500 CERROR("%s: alloc fid error: rc = %d\n",
501 osp->opd_obd->obd_name, rc);
507 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
508 "fid "DFID", last_fid "DFID"\n", PFID(fid),
511 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
513 CERROR("%s: Can not update oid/seq file: rc = %d\n",
514 osp->opd_obd->obd_name, rc);
518 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
519 osp->opd_obd->obd_name, fid_seq(last_fid),
521 /* Update last_xxx to the new seq */
522 spin_lock(&osp->opd_pre_lock);
523 osp->opd_last_used_fid = *fid;
524 osp_fid_to_obdid(fid, &osp->opd_last_id);
525 osp->opd_gap_start_fid = *fid;
526 osp->opd_pre_used_fid = *fid;
527 osp->opd_pre_last_created_fid = *fid;
528 spin_unlock(&osp->opd_pre_lock);
534 * Find IDs available in current sequence
536 * The function calculates the highest possible ID and the number of IDs
537 * available in the current sequence OSP is using. The number is limited
538 * artifically by the caller (grow param) and the number of IDs available
539 * in the sequence by nature. The function doesn't require an external
542 * \param[in] env LU environment provided by the caller
543 * \param[in] osp OSP device
544 * \param[in] fid FID the caller wants to start with
545 * \param[in] grow how many the caller wants
546 * \param[out] fid the highest calculated FID
547 * \param[out] grow the number of available IDs calculated
549 * \retval 0 on success, 1 - the sequence is empty
551 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
552 struct lu_fid *fid, int *grow)
554 struct osp_thread_info *osi = osp_env_info(env);
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, IDIF_MAX_OID);
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), (__u64)LUSTRE_DATA_SEQ_MAX_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;
634 /* we should not resend create request - anyway we will have delorphan
635 * and kill these objects */
636 req->rq_no_delay = req->rq_no_resend = 1;
638 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
640 ptlrpc_request_free(req);
644 spin_lock(&d->opd_pre_lock);
645 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
646 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
647 grow = d->opd_pre_create_count;
648 spin_unlock(&d->opd_pre_lock);
650 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
653 *fid = d->opd_pre_last_created_fid;
654 rc = osp_precreate_fids(env, d, fid, &grow);
656 /* Current seq has been used up*/
657 GOTO(out_req, rc = -ENOSPC);
659 if (!osp_is_fid_client(d)) {
660 /* Non-FID client will always send seq 0 because of
662 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
666 fid_to_ostid(fid, &body->oa.o_oi);
667 body->oa.o_valid = OBD_MD_FLGROUP;
669 ptlrpc_request_set_replen(req);
671 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
674 rc = ptlrpc_queue_wait(req);
676 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
679 /* have osp_precreate_reserve() to wait for repeat */
683 LASSERT(req->rq_transno == 0);
685 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
687 GOTO(out_req, rc = -EPROTO);
689 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
692 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
693 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
694 ": rc = %d\n", d->opd_obd->obd_name,
695 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
696 GOTO(out_req, rc = -ESTALE);
699 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
701 spin_lock(&d->opd_pre_lock);
703 /* the OST has not managed to create all the
704 * objects we asked for */
705 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
706 d->opd_pre_create_slow = 1;
708 /* the OST is able to keep up with the work,
709 * we could consider increasing create_count
710 * next time if needed */
711 d->opd_pre_create_slow = 0;
714 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
715 fid_to_ostid(fid, &body->oa.o_oi);
717 d->opd_pre_last_created_fid = *fid;
718 spin_unlock(&d->opd_pre_lock);
720 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
721 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
722 PFID(&d->opd_pre_last_created_fid));
724 /* now we can wakeup all users awaiting for objects */
725 osp_pre_update_status(d, rc);
727 /* pause to let osp_precreate_reserve to go first */
728 CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_PRECREATE_PAUSE, 2);
730 ptlrpc_req_finished(req);
735 * Get last precreated object from target (OST)
737 * Sends synchronous RPC to the target (OST) to learn the last precreated
738 * object. This later is used to remove all unused objects (cleanup orphan
739 * procedure). Also, the next object after one we got will be used as a
740 * starting point for the new precreates.
742 * \param[in] env LU environment provided by the caller
743 * \param[in] d OSP device
745 * \retval 0 on success
746 * \retval negative negated errno on error
748 static int osp_get_lastfid_from_ost(const struct lu_env *env,
749 struct osp_device *d)
751 struct ptlrpc_request *req = NULL;
752 struct obd_import *imp;
753 struct lu_fid *last_fid;
758 imp = d->opd_obd->u.cli.cl_import;
761 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
765 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
766 sizeof(KEY_LAST_FID));
768 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
770 ptlrpc_request_free(req);
774 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
775 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
777 req->rq_no_delay = req->rq_no_resend = 1;
778 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
779 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
781 ptlrpc_request_set_replen(req);
783 rc = ptlrpc_queue_wait(req);
785 /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld),
786 * let sysadm sort this * out.
789 ptlrpc_set_import_active(imp, 0);
793 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
794 if (last_fid == NULL) {
795 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
796 GOTO(out, rc = -EPROTO);
799 if (!fid_is_sane(last_fid)) {
800 CERROR("%s: Got insane last_fid "DFID"\n",
801 d->opd_obd->obd_name, PFID(last_fid));
802 GOTO(out, rc = -EPROTO);
805 /* Only update the last used fid, if the OST has objects for
806 * this sequence, i.e. fid_oid > 0 */
807 if (fid_oid(last_fid) > 0)
808 d->opd_last_used_fid = *last_fid;
810 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
814 ptlrpc_req_finished(req);
819 * Cleanup orphans on OST
821 * This function is called in a contex of a dedicated thread handling
822 * all the precreation suff. The function waits till local recovery
823 * is complete, then identify all the unreferenced objects (orphans)
824 * using the highest ID referenced by a local and the highest object
825 * precreated by the target. The found range is a subject to removal
826 * using specially flagged RPC. During this process OSP is marked
827 * unavailable for new objects.
829 * \param[in] env LU environment provided by the caller
830 * \param[in] d OSP device
832 * \retval 0 on success
833 * \retval negative negated errno on error
835 static int osp_precreate_cleanup_orphans(struct lu_env *env,
836 struct osp_device *d)
838 struct osp_thread_info *osi = osp_env_info(env);
839 struct lu_fid *last_fid = &osi->osi_fid;
840 struct ptlrpc_request *req = NULL;
841 struct obd_import *imp;
842 struct ost_body *body;
843 int update_status = 0;
850 * wait for local recovery to finish, so we can cleanup orphans
851 * orphans are all objects since "last used" (assigned), but
852 * there might be objects reserved and in some cases they won't
853 * be used. we can't cleanup them till we're sure they won't be
854 * used. also can't we allow new reservations because they may
855 * end up getting orphans being cleaned up below. so we block
856 * new reservations and wait till all reserved objects either
859 spin_lock(&d->opd_pre_lock);
860 d->opd_pre_recovering = 1;
861 spin_unlock(&d->opd_pre_lock);
863 * The locking above makes sure the opd_pre_reserved check below will
864 * catch all osp_precreate_reserve() calls who find
865 * "!opd_pre_recovering".
867 wait_event_idle(d->opd_pre_waitq,
868 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
869 !d->opd_pre_task || d->opd_got_disconnected);
870 if (!d->opd_pre_task || d->opd_got_disconnected)
871 GOTO(out, rc = -EAGAIN);
873 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
874 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
876 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val);
878 *last_fid = d->opd_last_used_fid;
879 /* The OSP should already get the valid seq now */
880 LASSERT(!fid_is_zero(last_fid));
881 if (fid_oid(&d->opd_last_used_fid) < 2 ||
882 OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
883 /* lastfid looks strange... ask OST */
884 rc = osp_get_lastfid_from_ost(env, d);
889 imp = d->opd_obd->u.cli.cl_import;
892 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
894 GOTO(out, rc = -ENOMEM);
896 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
898 ptlrpc_request_free(req);
903 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
905 GOTO(out, rc = -EPROTO);
907 body->oa.o_flags = OBD_FL_DELORPHAN;
908 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
910 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
912 ptlrpc_request_set_replen(req);
914 /* Don't resend the delorphan req */
915 req->rq_no_resend = req->rq_no_delay = 1;
917 rc = ptlrpc_queue_wait(req);
923 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
925 GOTO(out, rc = -EPROTO);
928 * OST provides us with id new pool starts from in body->oa.o_id
930 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
932 spin_lock(&d->opd_pre_lock);
933 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
935 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
936 d->opd_pre_last_created_fid = d->opd_last_used_fid;
938 d->opd_pre_create_count = OST_MIN_PRECREATE;
939 d->opd_pre_last_created_fid = *last_fid;
942 * This empties the pre-creation pool and effectively blocks any new
945 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
946 LUSTRE_DATA_SEQ_MAX_WIDTH);
947 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
948 d->opd_pre_create_slow = 0;
949 spin_unlock(&d->opd_pre_lock);
951 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
952 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
953 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
956 ptlrpc_req_finished(req);
960 * If rc is zero, the pre-creation window should have been emptied.
961 * Since waking up the herd would be useless without pre-created
962 * objects, we defer the signal to osp_precreate_send() in that case.
966 CERROR("%s: cannot cleanup orphans: rc = %d\n",
967 d->opd_obd->obd_name, rc);
968 /* we can't proceed from here, OST seem to
969 * be in a bad shape, better to wait for
970 * a new instance of the server and repeat
971 * from the beginning. notify possible waiters
972 * this OSP isn't quite functional yet */
973 osp_pre_update_status(d, rc);
975 wake_up_all(&d->opd_pre_user_waitq);
978 spin_lock(&d->opd_pre_lock);
979 d->opd_pre_recovering = 0;
980 spin_unlock(&d->opd_pre_lock);
987 * Update precreate status using statfs data
989 * The function decides whether this OSP should be used for new objects.
990 * IOW, whether this OST is used up or has some free space. Cached statfs
991 * data is used to make this decision. If the latest result of statfs
992 * request (rc argument) is not success, then just mark OSP unavailable
995 * The new statfs data is passed in \a msfs and needs to be stored into
996 * opd_statfs, but only after the various flags in os_state are set, so
997 * that the new statfs data is not visible without appropriate flags set.
998 * As such, there is no need to clear the flags here, since this is called
999 * with new statfs data, and they should not be cleared if sent from OST.
1001 * Add a bit of hysteresis so this flag isn't continually flapping, and
1002 * ensure that new files don't get extremely fragmented due to only a
1003 * small amount of available space in the filesystem. We want to set
1004 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1005 * reserved size free, and still copy them from the old state when there
1006 * is less than 2*reserved size free space or inodes.
1008 * \param[in] d OSP device
1009 * \param[in] msfs statfs data
1011 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1013 u32 old_state = d->opd_statfs.os_state;
1014 u32 reserved_ino_low = 32; /* could be tunable in the future */
1015 u32 reserved_ino_high = reserved_ino_low * 2;
1018 /* statfs structure not initialized yet */
1019 if (unlikely(!msfs->os_type))
1022 /* if the low and high watermarks have not been initialized yet */
1023 if (unlikely(d->opd_reserved_mb_high == 0 &&
1024 d->opd_reserved_mb_low == 0)) {
1025 /* Use ~0.1% by default to disable object allocation,
1026 * and ~0.2% to enable, size in MB, set both watermark
1028 spin_lock(&d->opd_pre_lock);
1029 if (d->opd_reserved_mb_high == 0 &&
1030 d->opd_reserved_mb_low == 0) {
1031 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1032 msfs->os_blocks) >> 20;
1033 if (d->opd_reserved_mb_low == 0)
1034 d->opd_reserved_mb_low = 1;
1035 d->opd_reserved_mb_high =
1036 (d->opd_reserved_mb_low << 1) + 1;
1038 spin_unlock(&d->opd_pre_lock);
1041 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1042 if (msfs->os_ffree < reserved_ino_low)
1043 msfs->os_state |= OS_STATFS_ENOINO;
1044 else if (msfs->os_ffree <= reserved_ino_high)
1045 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1046 /* else don't clear flags in new msfs->os_state sent from OST */
1049 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1050 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1051 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1052 msfs->os_files, msfs->os_ffree, msfs->os_state,
1054 if (available_mb < d->opd_reserved_mb_low)
1055 msfs->os_state |= OS_STATFS_ENOSPC;
1056 else if (available_mb <= d->opd_reserved_mb_high)
1057 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1058 /* else don't clear flags in new msfs->os_state sent from OST */
1060 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1061 d->opd_pre_status = -ENOSPC;
1062 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1063 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1064 d->opd_obd->obd_name, msfs->os_state,
1066 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1067 atomic_read(&d->opd_sync_changes),
1068 atomic_read(&d->opd_sync_rpcs_in_progress));
1069 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1070 d->opd_pre_status = 0;
1071 spin_lock(&d->opd_pre_lock);
1072 d->opd_pre_create_slow = 0;
1073 d->opd_pre_create_count = OST_MIN_PRECREATE;
1074 spin_unlock(&d->opd_pre_lock);
1075 wake_up(&d->opd_pre_waitq);
1078 "%s: available: state=%x: rc = %d\n",
1079 d->opd_obd->obd_name, msfs->os_state,
1082 /* we only get here if rc == 0 in the caller */
1083 d->opd_pre_status = 0;
1086 /* Object precreation skipped on OST if manually disabled */
1087 if (d->opd_pre_max_create_count == 0)
1088 msfs->os_state |= OS_STATFS_NOPRECREATE;
1089 /* else don't clear flags in new msfs->os_state sent from OST */
1091 /* copy only new statfs state to make it visible to MDS threads */
1092 if (&d->opd_statfs != msfs)
1093 d->opd_statfs = *msfs;
1097 * Initialize FID for precreation
1099 * For a just created new target, a new sequence should be taken.
1100 * The function checks there is no IDIF in use (if the target was
1101 * added with the older version of Lustre), then requests a new
1102 * sequence from FLDB using the regular protocol. Then this new
1103 * sequence is stored on a persisten storage synchronously to prevent
1104 * possible object leakage (for the detail see the description for
1105 * osp_precreate_rollover_new_seq()).
1107 * \param[in] osp OSP device
1109 * \retval 0 on success
1110 * \retval negative negated errno on error
1112 int osp_init_pre_fid(struct osp_device *osp)
1115 struct osp_thread_info *osi;
1116 struct lu_client_seq *cli_seq;
1117 struct lu_fid *last_fid;
1121 LASSERT(osp->opd_pre != NULL);
1123 /* Let's check if the current last_seq/fid is valid,
1124 * otherwise request new sequence from the controller */
1125 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1126 /* Non-MDT0 can only use normal sequence for
1128 if (fid_is_norm(&osp->opd_last_used_fid))
1131 /* Initially MDT0 will start with IDIF, after
1132 * that it will request new sequence from the
1134 if (fid_is_idif(&osp->opd_last_used_fid) ||
1135 fid_is_norm(&osp->opd_last_used_fid))
1139 if (!fid_is_zero(&osp->opd_last_used_fid))
1140 CWARN("%s: invalid last used fid "DFID
1141 ", try to get new sequence.\n",
1142 osp->opd_obd->obd_name,
1143 PFID(&osp->opd_last_used_fid));
1145 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1147 CERROR("%s: init env error: rc = %d\n",
1148 osp->opd_obd->obd_name, rc);
1152 osi = osp_env_info(&env);
1153 last_fid = &osi->osi_fid;
1155 /* For a freshed fs, it will allocate a new sequence first */
1156 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1157 cli_seq = osp->opd_obd->u.cli.cl_seq;
1158 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1160 CERROR("%s: alloc fid error: rc = %d\n",
1161 osp->opd_obd->obd_name, rc);
1165 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1167 last_fid->f_oid = 1;
1168 last_fid->f_ver = 0;
1170 spin_lock(&osp->opd_pre_lock);
1171 osp->opd_last_used_fid = *last_fid;
1172 osp->opd_pre_used_fid = *last_fid;
1173 osp->opd_pre_last_created_fid = *last_fid;
1174 spin_unlock(&osp->opd_pre_lock);
1175 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1177 CERROR("%s: write fid error: rc = %d\n",
1178 osp->opd_obd->obd_name, rc);
1187 struct osp_device *opta_dev;
1188 struct lu_env opta_env;
1189 struct completion *opta_started;
1192 * The core of precreate functionality
1194 * The function implements the main precreation loop. Basically it
1195 * involves connecting to the target, precerate FID initialization,
1196 * identifying and removing orphans, then serving precreation. As
1197 * part of the latter, the thread is responsible for statfs data
1198 * updates. The precreation is mostly driven by another threads
1199 * asking for new OST objects - those askers wake the thread when
1200 * the number of precreated objects reach low watermark.
1201 * After a disconnect, the sequence above repeats. This is keep going
1202 * until the thread is requested to stop.
1204 * \param[in] _arg private data the thread (OSP device to handle)
1206 * \retval 0 on success
1207 * \retval negative negated errno on error
1209 static int osp_precreate_thread(void *_args)
1211 struct opt_args *args = _args;
1212 struct osp_device *d = args->opta_dev;
1213 struct lu_env *env = &args->opta_env;
1218 complete(args->opta_started);
1219 while (!kthread_should_stop()) {
1221 * need to be connected to OST
1223 while (!kthread_should_stop()) {
1224 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1225 d->opd_imp_connected &&
1226 !d->opd_got_disconnected)
1228 wait_event_idle(d->opd_pre_waitq,
1229 kthread_should_stop() ||
1230 d->opd_new_connection);
1232 if (!d->opd_new_connection)
1235 OBD_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
1237 d->opd_new_connection = 0;
1238 d->opd_got_disconnected = 0;
1242 if (kthread_should_stop())
1246 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1247 /* Sigh, fid client is not ready yet */
1248 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1251 /* Init fid for osp_precreate if necessary */
1252 rc = osp_init_pre_fid(d);
1254 class_export_put(d->opd_exp);
1255 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1256 CERROR("%s: init pre fid error: rc = %d\n",
1257 d->opd_obd->obd_name, rc);
1262 if (osp_statfs_update(env, d)) {
1263 if (wait_event_idle_timeout(d->opd_pre_waitq,
1264 kthread_should_stop(),
1265 cfs_time_seconds(5)) == 0)
1266 l_wait_event_abortable(
1268 kthread_should_stop());
1274 * Clean up orphans or recreate missing objects.
1276 rc = osp_precreate_cleanup_orphans(env, d);
1278 schedule_timeout_interruptible(cfs_time_seconds(1));
1284 * connected, can handle precreates now
1286 while (!kthread_should_stop()) {
1287 wait_event_idle(d->opd_pre_waitq,
1288 kthread_should_stop() ||
1289 osp_precreate_near_empty(env, d) ||
1290 osp_statfs_need_update(d) ||
1291 d->opd_got_disconnected);
1293 if (kthread_should_stop())
1296 /* something happened to the connection
1297 * have to start from the beginning */
1298 if (d->opd_got_disconnected)
1301 if (osp_statfs_need_update(d))
1302 if (osp_statfs_update(env, d))
1305 if (d->opd_pre == NULL)
1308 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
1309 d->opd_pre_recovering = 1;
1313 /* To avoid handling different seq in precreate/orphan
1314 * cleanup, it will hold precreate until current seq is
1316 if (unlikely(osp_precreate_end_seq(env, d) &&
1317 !osp_create_end_seq(env, d)))
1320 if (unlikely(osp_precreate_end_seq(env, d) &&
1321 osp_create_end_seq(env, d))) {
1322 LCONSOLE_INFO("%s:%#llx is used up."
1323 " Update to new seq\n",
1324 d->opd_obd->obd_name,
1325 fid_seq(&d->opd_pre_last_created_fid));
1326 rc = osp_precreate_rollover_new_seq(env, d);
1331 if (osp_precreate_near_empty(env, d)) {
1332 rc = osp_precreate_send(env, d);
1333 /* osp_precreate_send() sets opd_pre_status
1334 * in case of error, that prevent the using of
1336 if (rc < 0 && rc != -ENOSPC &&
1337 rc != -ETIMEDOUT && rc != -ENOTCONN)
1338 CERROR("%s: cannot precreate objects:"
1340 d->opd_obd->obd_name, rc);
1352 * Check when to stop to wait for precreate objects.
1354 * The caller wanting a new OST object can't wait undefinitely. The
1355 * function checks for few conditions including available new OST
1356 * objects, disconnected OST, lack of space with no pending destroys,
1357 * etc. IOW, it checks whether the current OSP state is good to keep
1358 * waiting or it's better to give up.
1360 * \param[in] env LU environment provided by the caller
1361 * \param[in] d OSP device
1363 * \retval 0 - keep waiting, 1 - no luck
1365 static int osp_precreate_ready_condition(const struct lu_env *env,
1366 struct osp_device *d)
1368 /* Bail out I/O fails to OST */
1369 if (d->opd_pre_status != 0 &&
1370 d->opd_pre_status != -EAGAIN &&
1371 d->opd_pre_status != -ENODEV &&
1372 d->opd_pre_status != -ENOTCONN &&
1373 d->opd_pre_status != -ENOSPC) {
1375 if (d->opd_pre_status != -EIO)
1376 CERROR("%s: precreate failed opd_pre_status %d\n",
1377 d->opd_obd->obd_name, d->opd_pre_status);
1381 if (d->opd_pre_recovering)
1384 /* ready if got enough precreated objects */
1385 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1386 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1389 /* ready if OST reported no space and no destroys in progress */
1390 if (atomic_read(&d->opd_sync_changes) +
1391 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1392 d->opd_pre_status == -ENOSPC)
1399 * Reserve object in precreate pool
1401 * When the caller wants to create a new object on this target (target
1402 * represented by the given OSP), it should declare this intention using
1403 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1404 * to reserve an object in the existing precreated pool or wait up to
1405 * obd_timeout for the available object to appear in the pool (a dedicated
1406 * thread will be doing real precreation in background). The object can be
1407 * consumed later with osp_precreate_get_fid() or be released with call to
1408 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1409 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1410 * If the space on the target is short and there is a pending object destroy,
1411 * then the function forces local commit to speedup space release (see
1412 * osp_sync.c for the details).
1414 * \param[in] env LU environment provided by the caller
1415 * \param[in] d OSP device
1417 * \retval 0 on success
1418 * \retval -ENOSPC when no space on OST
1419 * \retval -EAGAIN try later, slow precreation in progress
1420 * \retval -EIO when no access to OST
1422 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
1425 time64_t expire = ktime_get_seconds() + obd_timeout;
1426 int precreated, rc, synced = 0;
1430 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1431 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1432 PFID(&d->opd_pre_used_fid));
1434 /* opd_pre_max_create_count 0 to not use specified OST. */
1435 if (d->opd_pre_max_create_count == 0)
1440 * - preallocation is done
1441 * - no free space expected soon
1442 * - can't connect to OST for too long (obd_timeout)
1443 * - OST can allocate fid sequence.
1445 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1446 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1449 * increase number of precreations
1451 precreated = osp_objs_precreated(env, d);
1452 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1453 d->opd_pre_create_slow == 0 &&
1454 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1455 spin_lock(&d->opd_pre_lock);
1456 d->opd_pre_create_slow = 1;
1457 d->opd_pre_create_count *= 2;
1458 spin_unlock(&d->opd_pre_lock);
1461 spin_lock(&d->opd_pre_lock);
1462 precreated = osp_objs_precreated(env, d);
1463 if (precreated > d->opd_pre_reserved &&
1464 !d->opd_pre_recovering) {
1465 d->opd_pre_reserved++;
1466 spin_unlock(&d->opd_pre_lock);
1469 /* XXX: don't wake up if precreation is in progress */
1470 if (osp_precreate_near_empty_nolock(env, d) &&
1471 !osp_precreate_end_seq_nolock(env, d))
1472 wake_up(&d->opd_pre_waitq);
1476 spin_unlock(&d->opd_pre_lock);
1479 * all precreated objects have been used and no-space
1480 * status leave us no chance to succeed very soon
1481 * but if there is destroy in progress, then we should
1482 * wait till that is done - some space might be released
1484 if (unlikely(rc == -ENOSPC)) {
1485 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1486 /* force local commit to release space */
1487 dt_commit_async(env, d->opd_storage);
1488 osp_sync_check_for_work(d);
1491 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1492 /* just wait till destroys are done
1493 * see wait_event_idle_timeout() below
1496 if (atomic_read(&d->opd_sync_changes) +
1497 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1498 /* no hope for free space */
1503 /* XXX: don't wake up if precreation is in progress */
1504 wake_up(&d->opd_pre_waitq);
1506 if (ktime_get_seconds() >= expire) {
1512 LASSERT(d->opd_pre);
1517 CDEBUG(D_INFO, "%s: Sleeping on objects\n",
1518 d->opd_obd->obd_name);
1519 if (wait_event_idle_timeout(
1520 d->opd_pre_user_waitq,
1521 osp_precreate_ready_condition(env, d),
1522 cfs_time_seconds(obd_timeout)) == 0) {
1524 "%s: slow creates, last="DFID", next="DFID", "
1525 "reserved=%llu, sync_changes=%u, "
1526 "sync_rpcs_in_progress=%d, status=%d\n",
1527 d->opd_obd->obd_name,
1528 PFID(&d->opd_pre_last_created_fid),
1529 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1530 atomic_read(&d->opd_sync_changes),
1531 atomic_read(&d->opd_sync_rpcs_in_progress),
1534 CDEBUG(D_INFO, "%s: Waked up, status=%d\n",
1535 d->opd_obd->obd_name, d->opd_pre_status);
1543 * Get a FID from precreation pool
1545 * The function is a companion for osp_precreate_reserve() - it assigns
1546 * a specific FID from the precreate. The function should be called only
1547 * if the call to osp_precreate_reserve() was successful. The function
1548 * updates a local storage to remember the highest object ID referenced
1549 * by the node in the given sequence.
1551 * A very importan details: this is supposed to be called once the
1552 * transaction is started, so on-disk update will be atomic with the
1553 * data (like LOVEA) refering this object. Then the object won't be leaked:
1554 * either it's referenced by the committed transaction or it's a subject
1555 * to the orphan cleanup procedure.
1557 * \param[in] env LU environment provided by the caller
1558 * \param[in] d OSP device
1559 * \param[out] fid generated FID
1561 * \retval 0 on success
1562 * \retval negative negated errno on error
1564 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1567 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1568 /* grab next id from the pool */
1569 spin_lock(&d->opd_pre_lock);
1571 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1572 &d->opd_pre_last_created_fid) < 0,
1573 "next fid "DFID" last created fid "DFID"\n",
1574 PFID(&d->opd_pre_used_fid),
1575 PFID(&d->opd_pre_last_created_fid));
1578 * When sequence is used up, new one should be allocated in
1579 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1582 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1583 "next fid "DFID" last created fid "DFID"\n",
1584 PFID(&d->opd_pre_used_fid),
1585 PFID(&d->opd_pre_last_created_fid));
1586 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1587 if (fid_is_idif(pre_used_fid) &&
1588 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1589 pre_used_fid->f_seq++;
1591 d->opd_pre_used_fid.f_oid++;
1592 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1593 d->opd_pre_reserved--;
1595 * last_used_id must be changed along with getting new id otherwise
1596 * we might miscalculate gap causing object loss or leak
1598 osp_update_last_fid(d, fid);
1599 spin_unlock(&d->opd_pre_lock);
1602 * probably main thread suspended orphan cleanup till
1603 * all reservations are released, see comment in
1604 * osp_precreate_thread() just before orphan cleanup
1606 if (unlikely(d->opd_pre_reserved == 0 &&
1607 (d->opd_pre_recovering || d->opd_pre_status)))
1608 wake_up(&d->opd_pre_waitq);
1614 * Set size regular attribute on an object
1616 * When a striping is created late, it's possible that size is already
1617 * initialized on the file. Then the new striping should inherit size
1618 * from the file. The function sets size on the object using the regular
1619 * protocol (OST_PUNCH).
1620 * XXX: should be re-implemented using OUT ?
1622 * \param[in] env LU environment provided by the caller
1623 * \param[in] dt object
1624 * \param[in] size size to set.
1626 * \retval 0 on success
1627 * \retval negative negated errno on error
1629 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1632 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1633 struct ptlrpc_request *req = NULL;
1634 struct obd_import *imp;
1635 struct ost_body *body;
1636 struct obdo *oa = NULL;
1641 imp = d->opd_obd->u.cli.cl_import;
1644 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1648 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1650 ptlrpc_request_free(req);
1655 * XXX: decide how do we do here with resend
1656 * if we don't resend, then client may see wrong file size
1657 * if we do resend, then MDS thread can get stuck for quite long
1658 * and if we don't resend, then client will also get -EAGAIN !!
1659 * (see LU-7975 and sanity/test_27F use cases)
1660 * but let's decide not to resend/delay this truncate request to OST
1661 * and allow Client to decide to resend, in a less agressive way from
1662 * after_reply(), by returning -EINPROGRESS instead of
1663 * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
1664 * end of this routine
1666 req->rq_no_resend = req->rq_no_delay = 1;
1668 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1669 ptlrpc_at_set_req_timeout(req);
1673 GOTO(out, rc = -ENOMEM);
1675 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1678 oa->o_blocks = OBD_OBJECT_EOF;
1679 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1680 OBD_MD_FLID | OBD_MD_FLGROUP;
1682 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1684 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1686 /* XXX: capa support? */
1687 /* osc_pack_capa(req, body, capa); */
1689 ptlrpc_request_set_replen(req);
1691 rc = ptlrpc_queue_wait(req);
1693 /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
1694 * since we have decided not to resend/delay, but this could
1695 * lead to wrong size to be seen at Client side and even process
1696 * trying to open to exit/fail if not itself handling -EAGAIN.
1697 * So it should be better to return -EINPROGRESS instead and
1698 * leave the decision to resend at Client side in after_reply()
1700 if (rc == -EAGAIN) {
1702 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1703 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1706 CERROR("can't punch object: %d\n", rc);
1710 ptlrpc_req_finished(req);
1717 * Initialize precreation functionality of OSP
1719 * Prepares all the internal structures and starts the precreate thread
1721 * \param[in] d OSP device
1723 * \retval 0 on success
1724 * \retval negative negated errno on error
1726 int osp_init_precreate(struct osp_device *d)
1730 OBD_ALLOC_PTR(d->opd_pre);
1731 if (d->opd_pre == NULL)
1734 /* initially precreation isn't ready */
1735 init_waitqueue_head(&d->opd_pre_user_waitq);
1736 d->opd_pre_status = -EAGAIN;
1737 fid_zero(&d->opd_pre_used_fid);
1738 d->opd_pre_used_fid.f_oid = 1;
1739 fid_zero(&d->opd_pre_last_created_fid);
1740 d->opd_pre_last_created_fid.f_oid = 1;
1742 d->opd_pre_reserved = 0;
1743 d->opd_got_disconnected = 1;
1744 d->opd_pre_create_slow = 0;
1745 d->opd_pre_create_count = OST_MIN_PRECREATE;
1746 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1747 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1748 d->opd_reserved_mb_high = 0;
1749 d->opd_reserved_mb_low = 0;
1755 * Finish precreate functionality of OSP
1758 * Asks all the activity (the thread, update timer) to stop, then
1759 * wait till that is done.
1761 * \param[in] d OSP device
1763 void osp_precreate_fini(struct osp_device *d)
1767 if (d->opd_pre == NULL)
1770 OBD_FREE_PTR(d->opd_pre);
1776 int osp_init_statfs(struct osp_device *d)
1778 struct task_struct *task;
1779 struct opt_args *args;
1780 DECLARE_COMPLETION_ONSTACK(started);
1785 spin_lock_init(&d->opd_pre_lock);
1786 init_waitqueue_head(&d->opd_pre_waitq);
1789 * Initialize statfs-related things
1791 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1792 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1793 1000 * NSEC_PER_SEC);
1794 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1796 ktime_to_ns(d->opd_statfs_fresh_till));
1797 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1798 (unsigned long)d, 0);
1800 if (d->opd_storage->dd_rdonly)
1803 OBD_ALLOC_PTR(args);
1807 args->opta_started = &started;
1808 rc = lu_env_init(&args->opta_env,
1809 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1811 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1818 * start thread handling precreation and statfs updates
1820 task = kthread_create(osp_precreate_thread, args,
1821 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1823 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1824 lu_env_fini(&args->opta_env);
1826 RETURN(PTR_ERR(task));
1828 d->opd_pre_task = task;
1829 wake_up_process(task);
1830 wait_for_completion(&started);
1835 void osp_statfs_fini(struct osp_device *d)
1837 struct task_struct *task = d->opd_pre_task;
1840 del_timer(&d->opd_statfs_timer);
1842 d->opd_pre_task = NULL;