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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osp/osp_precreate.c
34 * Lustre OST Proxy Device
36 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
37 * Author: Mikhail Pershin <mike.pershin@intel.com>
38 * Author: Di Wang <di.wang@intel.com>
41 #define DEBUG_SUBSYSTEM S_MDS
43 #include <linux/kthread.h>
45 #include <lustre_obdo.h>
47 #include "osp_internal.h"
50 * there are two specific states to take care about:
52 * = import is disconnected =
54 * = import is inactive =
55 * in this case osp_declare_create() returns an error
61 * Check whether statfs data is expired
63 * OSP device caches statfs data for the target, the function checks
64 * whether the data is expired or not.
66 * \param[in] d OSP device
68 * \retval 0 - not expired, 1 - expired
70 static inline int osp_statfs_need_update(struct osp_device *d)
72 return !ktime_before(ktime_get(), d->opd_statfs_fresh_till);
76 * OSP tries to maintain pool of available objects so that calls to create
77 * objects don't block most of time
79 * each time OSP gets connected to OST, we should start from precreation cleanup
81 static inline bool osp_precreate_running(struct osp_device *d)
83 return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
86 static inline bool osp_precreate_stopped(struct osp_device *d)
88 return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
91 static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data)
93 struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
96 if (osp_precreate_running(d))
97 wake_up(&d->opd_pre_waitq);
101 * RPC interpret callback for OST_STATFS RPC
103 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
104 * replied by the target. It's used to maintain statfs cache for the target.
105 * The function fills data from the reply if successful and schedules another
108 * \param[in] env LU environment provided by the caller
109 * \param[in] req RPC replied
110 * \param[in] aa callback data
111 * \param[in] rc RPC result
113 * \retval 0 on success
114 * \retval negative negated errno on error
116 static int osp_statfs_interpret(const struct lu_env *env,
117 struct ptlrpc_request *req,
118 union ptlrpc_async_args *aa, int rc)
120 struct obd_import *imp = req->rq_import;
121 struct obd_statfs *msfs;
122 struct osp_device *d;
127 aa = ptlrpc_req_async_args(req);
128 d = aa->pointer_arg[0];
134 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
136 GOTO(out, rc = -EPROTO);
138 d->opd_statfs = *msfs;
141 osp_pre_update_status(d, rc);
143 /* schedule next update */
144 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
145 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
146 mod_timer(&d->opd_statfs_timer,
147 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
148 d->opd_statfs_update_in_progress = 0;
150 CDEBUG(D_CACHE, "updated statfs %p\n", d);
154 /* couldn't update statfs, try again with a small delay */
155 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
156 d->opd_statfs_update_in_progress = 0;
157 if (d->opd_pre != NULL && osp_precreate_running(d))
158 wake_up(&d->opd_pre_waitq);
160 if (req->rq_import_generation == imp->imp_generation)
161 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
162 d->opd_obd->obd_name, rc);
167 * Send OST_STATFS RPC
169 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
170 * Also disables scheduled updates as times OSP may need to refresh
171 * statfs data before expiration. The function doesn't block, instead
172 * an interpretation callback osp_statfs_interpret() is used.
174 * \param[in] d OSP device
176 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
178 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
179 struct ptlrpc_request *req;
180 struct obd_import *imp;
181 union ptlrpc_async_args *aa;
186 CDEBUG(D_CACHE, "going to update statfs\n");
188 imp = d->opd_obd->u.cli.cl_import;
191 req = ptlrpc_request_alloc(imp,
192 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
196 rc = ptlrpc_request_pack(req,
197 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
198 d->opd_pre ? OST_STATFS : MDS_STATFS);
200 ptlrpc_request_free(req);
203 ptlrpc_request_set_replen(req);
205 req->rq_request_portal = OST_CREATE_PORTAL;
206 ptlrpc_at_set_req_timeout(req);
208 req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
209 aa = ptlrpc_req_async_args(req);
210 aa->pointer_arg[0] = d;
213 * no updates till reply
215 del_timer(&d->opd_statfs_timer);
216 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
217 d->opd_statfs_update_in_progress = 1;
219 ptlrpcd_add_req(req);
221 /* we still want to sync changes if no new changes are coming */
222 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
225 if (atomic_read(&d->opd_sync_changes)) {
228 th = dt_trans_create(env, d->opd_storage);
230 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
233 rc = dt_trans_start_local(env, d->opd_storage, th);
235 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
236 d->opd_obd->obd_name,
237 atomic_read(&d->opd_sync_changes));
238 osp_sync_add_commit_cb_1s(env, d, th);
239 dt_trans_stop(env, d->opd_storage, th);
248 * Schedule an immediate update for statfs data
250 * If cached statfs data claim no free space, but OSP has got a request to
251 * destroy an object (so release some space probably), then we may need to
252 * refresh cached statfs data sooner than planned. The function checks there
253 * is no statfs update going and schedules immediate update if so.
254 * XXX: there might be a case where removed object(s) do not add free space (empty
255 * object). If the number of such deletions is high, then we can start to update
256 * statfs too often causing a RPC storm. some throttling is needed...
258 * \param[in] d OSP device where statfs data needs to be refreshed
260 void osp_statfs_need_now(struct osp_device *d)
262 if (!d->opd_statfs_update_in_progress) {
264 * if current status is -ENOSPC (lack of free space on OST)
265 * then we should poll OST immediately once object destroy
268 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
269 del_timer(&d->opd_statfs_timer);
270 wake_up(&d->opd_pre_waitq);
275 * Return number of precreated objects
277 * A simple helper to calculate the number of precreated objects on the device.
279 * \param[in] env LU environment provided by the caller
280 * \param[in] osp OSP device
282 * \retval the number of the precreated objects
284 static inline int osp_objs_precreated(const struct lu_env *env,
285 struct osp_device *osp)
287 return osp_fid_diff(&osp->opd_pre_last_created_fid,
288 &osp->opd_pre_used_fid);
292 * Check pool of precreated objects is nearly empty
294 * We should not wait till the pool of the precreated objects is exhausted,
295 * because then there will be a long period of OSP being unavailable for the
296 * new creations due to lenghty precreate RPC. Instead we ask for another
297 * precreation ahead and hopefully have it ready before the current pool is
298 * empty. Notice this function relies on an external locking.
300 * \param[in] env LU environment provided by the caller
301 * \param[in] d OSP device
303 * \retval 0 - current pool is good enough, 1 - time to precreate
305 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
306 struct osp_device *d)
308 int window = osp_objs_precreated(env, d);
310 /* don't consider new precreation till OST is healty and
312 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
313 (d->opd_pre_status == 0));
317 * Check pool of precreated objects
319 * This is protected version of osp_precreate_near_empty_nolock(), check that
322 * \param[in] env LU environment provided by the caller
323 * \param[in] d OSP device
325 * \retval 0 - current pool is good enough, 1 - time to precreate
327 static inline int osp_precreate_near_empty(const struct lu_env *env,
328 struct osp_device *d)
332 if (d->opd_pre == NULL)
335 /* XXX: do we really need locking here? */
336 spin_lock(&d->opd_pre_lock);
337 rc = osp_precreate_near_empty_nolock(env, d);
338 spin_unlock(&d->opd_pre_lock);
343 * Check given sequence is empty
345 * Returns a binary result whether the given sequence has some IDs left
346 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
347 * version of that function.
349 * \param[in] env LU environment provided by the caller
350 * \param[in] osp OSP device
352 * \retval 0 - current sequence has no IDs, 1 - otherwise
354 static inline int osp_create_end_seq(const struct lu_env *env,
355 struct osp_device *osp)
357 struct lu_fid *fid = &osp->opd_pre_used_fid;
360 spin_lock(&osp->opd_pre_lock);
361 rc = osp_fid_end_seq(env, fid);
362 spin_unlock(&osp->opd_pre_lock);
367 * Write FID into into last_oid/last_seq file
369 * The function stores the sequence and the in-sequence id into two dedicated
370 * files. The sync argument can be used to request synchronous commit, so the
371 * function won't return until the updates are committed.
373 * \param[in] env LU environment provided by the caller
374 * \param[in] osp OSP device
375 * \param[in] fid fid where sequence/id is taken
376 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
378 * \retval 0 on success
379 * \retval negative negated errno on error
381 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
382 struct lu_fid *fid, int sync)
384 struct osp_thread_info *oti = osp_env_info(env);
385 struct lu_buf *lb_oid = &oti->osi_lb;
386 struct lu_buf *lb_oseq = &oti->osi_lb2;
393 if (osp->opd_storage->dd_rdonly)
396 /* Note: through f_oid is only 32 bits, it will also write 64 bits
397 * for oid to keep compatibility with the previous version. */
398 lb_oid->lb_buf = &fid->f_oid;
399 lb_oid->lb_len = sizeof(u64);
400 oid_off = sizeof(u64) * osp->opd_index;
402 lb_oseq->lb_buf = &fid->f_seq;
403 lb_oseq->lb_len = sizeof(u64);
404 oseq_off = sizeof(u64) * osp->opd_index;
406 th = dt_trans_create(env, osp->opd_storage);
411 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
412 lb_oid, oid_off, th);
416 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
417 lb_oseq, oseq_off, th);
421 rc = dt_trans_start_local(env, osp->opd_storage, th);
425 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
428 CERROR("%s: can not write to last seq file: rc = %d\n",
429 osp->opd_obd->obd_name, rc);
432 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
435 CERROR("%s: can not write to last seq file: rc = %d\n",
436 osp->opd_obd->obd_name, rc);
440 dt_trans_stop(env, osp->opd_storage, th);
445 * Switch to another sequence
447 * When a current sequence has no available IDs left, OSP has to switch to
448 * another new sequence. OSP requests it using the regular FLDB protocol
449 * and stores synchronously before that is used in precreated. This is needed
450 * to basically have the sequences referenced (not orphaned), otherwise it's
451 * possible that OST has some objects precreated and the clients have data
452 * written to it, but after MDT failover nobody refers those objects and OSP
453 * has no idea that the sequence need cleanup to be done.
454 * While this is very expensive operation, it's supposed to happen very very
455 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
457 * \param[in] env LU environment provided by the caller
458 * \param[in] osp OSP device
460 * \retval 0 on success
461 * \retval negative negated errno on error
463 static int osp_precreate_rollover_new_seq(struct lu_env *env,
464 struct osp_device *osp)
466 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
467 struct lu_fid *last_fid = &osp->opd_last_used_fid;
471 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
473 CERROR("%s: alloc fid error: rc = %d\n",
474 osp->opd_obd->obd_name, rc);
480 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
481 "fid "DFID", last_fid "DFID"\n", PFID(fid),
484 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
486 CERROR("%s: Can not update oid/seq file: rc = %d\n",
487 osp->opd_obd->obd_name, rc);
491 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
492 osp->opd_obd->obd_name, fid_seq(last_fid),
494 /* Update last_xxx to the new seq */
495 spin_lock(&osp->opd_pre_lock);
496 osp->opd_last_used_fid = *fid;
497 osp->opd_gap_start_fid = *fid;
498 osp->opd_pre_used_fid = *fid;
499 osp->opd_pre_last_created_fid = *fid;
500 spin_unlock(&osp->opd_pre_lock);
506 * Find IDs available in current sequence
508 * The function calculates the highest possible ID and the number of IDs
509 * available in the current sequence OSP is using. The number is limited
510 * artifically by the caller (grow param) and the number of IDs available
511 * in the sequence by nature. The function doesn't require an external
514 * \param[in] env LU environment provided by the caller
515 * \param[in] osp OSP device
516 * \param[in] fid FID the caller wants to start with
517 * \param[in] grow how many the caller wants
518 * \param[out] fid the highest calculated FID
519 * \param[out] grow the number of available IDs calculated
521 * \retval 0 on success, 1 - the sequence is empty
523 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
524 struct lu_fid *fid, int *grow)
526 struct osp_thread_info *osi = osp_env_info(env);
530 if (fid_is_idif(fid)) {
531 struct lu_fid *last_fid;
532 struct ost_id *oi = &osi->osi_oi;
535 spin_lock(&osp->opd_pre_lock);
536 last_fid = &osp->opd_pre_last_created_fid;
537 fid_to_ostid(last_fid, oi);
538 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
539 *grow = end - ostid_id(oi);
540 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
541 spin_unlock(&osp->opd_pre_lock);
543 if (*grow == 0 || rc)
546 ostid_to_fid(fid, oi, osp->opd_index);
550 spin_lock(&osp->opd_pre_lock);
551 *fid = osp->opd_pre_last_created_fid;
553 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
554 *grow = end - fid->f_oid;
555 fid->f_oid += end - fid->f_oid;
556 spin_unlock(&osp->opd_pre_lock);
558 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
559 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
561 return *grow > 0 ? 0 : 1;
565 * Prepare and send precreate RPC
567 * The function finds how many objects should be precreated. Then allocates,
568 * prepares and schedules precreate RPC synchronously. Upon reply the function
569 * wake ups the threads waiting for the new objects on this target. If the
570 * target wasn't able to create all the objects requested, then the next
571 * precreate will be asking less objects (i.e. slow precreate down).
573 * \param[in] env LU environment provided by the caller
574 * \param[in] d OSP device
576 * \retval 0 on success
577 * \retval negative negated errno on error
579 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
581 struct osp_thread_info *oti = osp_env_info(env);
582 struct ptlrpc_request *req;
583 struct obd_import *imp;
584 struct ost_body *body;
586 struct lu_fid *fid = &oti->osi_fid;
589 /* don't precreate new objects till OST healthy and has free space */
590 if (unlikely(d->opd_pre_status)) {
591 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
592 d->opd_obd->obd_name, d->opd_pre_status);
597 * if not connection/initialization is compeleted, ignore
599 imp = d->opd_obd->u.cli.cl_import;
602 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
605 req->rq_request_portal = OST_CREATE_PORTAL;
606 /* we should not resend create request - anyway we will have delorphan
607 * and kill these objects */
608 req->rq_no_delay = req->rq_no_resend = 1;
610 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
612 ptlrpc_request_free(req);
616 spin_lock(&d->opd_pre_lock);
617 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
618 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
619 grow = d->opd_pre_create_count;
620 spin_unlock(&d->opd_pre_lock);
622 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
625 *fid = d->opd_pre_last_created_fid;
626 rc = osp_precreate_fids(env, d, fid, &grow);
628 /* Current seq has been used up*/
629 if (!osp_is_fid_client(d)) {
630 osp_pre_update_status(d, -ENOSPC);
633 wake_up(&d->opd_pre_waitq);
637 if (!osp_is_fid_client(d)) {
638 /* Non-FID client will always send seq 0 because of
640 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
644 fid_to_ostid(fid, &body->oa.o_oi);
645 body->oa.o_valid = OBD_MD_FLGROUP;
647 ptlrpc_request_set_replen(req);
649 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
652 rc = ptlrpc_queue_wait(req);
654 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
658 LASSERT(req->rq_transno == 0);
660 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
662 GOTO(out_req, rc = -EPROTO);
664 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
667 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
668 CERROR("%s: precreate fid "DFID" < local used fid "DFID
669 ": rc = %d\n", d->opd_obd->obd_name,
670 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
671 GOTO(out_req, rc = -ESTALE);
674 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
676 spin_lock(&d->opd_pre_lock);
678 /* the OST has not managed to create all the
679 * objects we asked for */
680 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
681 d->opd_pre_create_slow = 1;
683 /* the OST is able to keep up with the work,
684 * we could consider increasing create_count
685 * next time if needed */
686 d->opd_pre_create_slow = 0;
689 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
690 fid_to_ostid(fid, &body->oa.o_oi);
692 d->opd_pre_last_created_fid = *fid;
693 spin_unlock(&d->opd_pre_lock);
695 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
696 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
697 PFID(&d->opd_pre_last_created_fid));
699 /* now we can wakeup all users awaiting for objects */
700 osp_pre_update_status(d, rc);
701 wake_up(&d->opd_pre_user_waitq);
703 ptlrpc_req_finished(req);
708 * Get last precreated object from target (OST)
710 * Sends synchronous RPC to the target (OST) to learn the last precreated
711 * object. This later is used to remove all unused objects (cleanup orphan
712 * procedure). Also, the next object after one we got will be used as a
713 * starting point for the new precreates.
715 * \param[in] env LU environment provided by the caller
716 * \param[in] d OSP device
718 * \retval 0 on success
719 * \retval negative negated errno on error
721 static int osp_get_lastfid_from_ost(const struct lu_env *env,
722 struct osp_device *d)
724 struct ptlrpc_request *req = NULL;
725 struct obd_import *imp;
726 struct lu_fid *last_fid;
731 imp = d->opd_obd->u.cli.cl_import;
734 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
738 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
739 sizeof(KEY_LAST_FID));
741 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
743 ptlrpc_request_free(req);
747 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
748 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
750 req->rq_no_delay = req->rq_no_resend = 1;
751 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
752 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
754 ptlrpc_request_set_replen(req);
756 rc = ptlrpc_queue_wait(req);
758 /* bad-bad OST.. let sysadm sort this out */
759 if (rc == -ENOTSUPP) {
760 CERROR("%s: server does not support FID: rc = %d\n",
761 d->opd_obd->obd_name, -ENOTSUPP);
763 ptlrpc_set_import_active(imp, 0);
767 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
768 if (last_fid == NULL) {
769 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
770 GOTO(out, rc = -EPROTO);
773 if (!fid_is_sane(last_fid)) {
774 CERROR("%s: Got insane last_fid "DFID"\n",
775 d->opd_obd->obd_name, PFID(last_fid));
776 GOTO(out, rc = -EPROTO);
779 /* Only update the last used fid, if the OST has objects for
780 * this sequence, i.e. fid_oid > 0 */
781 if (fid_oid(last_fid) > 0)
782 d->opd_last_used_fid = *last_fid;
784 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
788 ptlrpc_req_finished(req);
793 * Cleanup orphans on OST
795 * This function is called in a contex of a dedicated thread handling
796 * all the precreation suff. The function waits till local recovery
797 * is complete, then identify all the unreferenced objects (orphans)
798 * using the highest ID referenced by a local and the highest object
799 * precreated by the target. The found range is a subject to removal
800 * using specially flagged RPC. During this process OSP is marked
801 * unavailable for new objects.
803 * \param[in] env LU environment provided by the caller
804 * \param[in] d OSP device
806 * \retval 0 on success
807 * \retval negative negated errno on error
809 static int osp_precreate_cleanup_orphans(struct lu_env *env,
810 struct osp_device *d)
812 struct osp_thread_info *osi = osp_env_info(env);
813 struct lu_fid *last_fid = &osi->osi_fid;
814 struct ptlrpc_request *req = NULL;
815 struct obd_import *imp;
816 struct ost_body *body;
817 struct l_wait_info lwi = { 0 };
818 int update_status = 0;
825 * wait for local recovery to finish, so we can cleanup orphans
826 * orphans are all objects since "last used" (assigned), but
827 * there might be objects reserved and in some cases they won't
828 * be used. we can't cleanup them till we're sure they won't be
829 * used. also can't we allow new reservations because they may
830 * end up getting orphans being cleaned up below. so we block
831 * new reservations and wait till all reserved objects either
834 spin_lock(&d->opd_pre_lock);
835 d->opd_pre_recovering = 1;
836 spin_unlock(&d->opd_pre_lock);
838 * The locking above makes sure the opd_pre_reserved check below will
839 * catch all osp_precreate_reserve() calls who find
840 * "!opd_pre_recovering".
842 l_wait_event(d->opd_pre_waitq,
843 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
844 !osp_precreate_running(d) || d->opd_got_disconnected,
846 if (!osp_precreate_running(d) || d->opd_got_disconnected)
847 GOTO(out, rc = -EAGAIN);
849 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
850 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
852 *last_fid = d->opd_last_used_fid;
853 /* The OSP should already get the valid seq now */
854 LASSERT(!fid_is_zero(last_fid));
855 if (fid_oid(&d->opd_last_used_fid) < 2) {
856 /* lastfid looks strange... ask OST */
857 rc = osp_get_lastfid_from_ost(env, d);
862 imp = d->opd_obd->u.cli.cl_import;
865 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
867 GOTO(out, rc = -ENOMEM);
869 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
871 ptlrpc_request_free(req);
876 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
878 GOTO(out, rc = -EPROTO);
880 body->oa.o_flags = OBD_FL_DELORPHAN;
881 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
883 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
885 ptlrpc_request_set_replen(req);
887 /* Don't resend the delorphan req */
888 req->rq_no_resend = req->rq_no_delay = 1;
890 rc = ptlrpc_queue_wait(req);
896 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
898 GOTO(out, rc = -EPROTO);
901 * OST provides us with id new pool starts from in body->oa.o_id
903 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
905 spin_lock(&d->opd_pre_lock);
906 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
908 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
909 d->opd_pre_last_created_fid = d->opd_last_used_fid;
911 d->opd_pre_create_count = OST_MIN_PRECREATE;
912 d->opd_pre_last_created_fid = *last_fid;
915 * This empties the pre-creation pool and effectively blocks any new
918 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
919 LUSTRE_DATA_SEQ_MAX_WIDTH);
920 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
921 d->opd_pre_create_slow = 0;
922 spin_unlock(&d->opd_pre_lock);
924 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
925 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
926 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
929 ptlrpc_req_finished(req);
932 * If rc is zero, the pre-creation window should have been emptied.
933 * Since waking up the herd would be useless without pre-created
934 * objects, we defer the signal to osp_precreate_send() in that case.
938 CERROR("%s: cannot cleanup orphans: rc = %d\n",
939 d->opd_obd->obd_name, rc);
940 /* we can't proceed from here, OST seem to
941 * be in a bad shape, better to wait for
942 * a new instance of the server and repeat
943 * from the beginning. notify possible waiters
944 * this OSP isn't quite functional yet */
945 osp_pre_update_status(d, rc);
947 wake_up(&d->opd_pre_user_waitq);
950 spin_lock(&d->opd_pre_lock);
951 d->opd_pre_recovering = 0;
952 spin_unlock(&d->opd_pre_lock);
959 * Update precreate status using statfs data
961 * The function decides whether this OSP should be used for new objects.
962 * IOW, whether this OST is used up or has some free space. Cached statfs
963 * data is used to make this decision. If the latest result of statfs
964 * request (rc argument) is not success, then just mark OSP unavailable
967 * Add a bit of hysteresis so this flag isn't continually flapping,
968 * and ensure that new files don't get extremely fragmented due to
969 * only a small amount of available space in the filesystem.
970 * We want to set the ENOSPC when there is less than reserved size
971 * free and clear it when there is at least 2*reserved size free space.
972 * the function updates current precreation status used: functional or not
974 * \param[in] d OSP device
975 * \param[in] rc new precreate status for device \a d
977 * \retval 0 on success
978 * \retval negative negated errno on error
980 void osp_pre_update_status(struct osp_device *d, int rc)
982 struct obd_statfs *msfs = &d->opd_statfs;
983 int old = d->opd_pre_status;
986 d->opd_pre_status = rc;
990 if (likely(msfs->os_type)) {
991 if (unlikely(d->opd_reserved_mb_high == 0 &&
992 d->opd_reserved_mb_low == 0)) {
993 /* Use ~0.1% by default to disable object allocation,
994 * and ~0.2% to enable, size in MB, set both watermark
996 spin_lock(&d->opd_pre_lock);
997 if (d->opd_reserved_mb_high == 0 &&
998 d->opd_reserved_mb_low == 0) {
999 d->opd_reserved_mb_low =
1000 ((msfs->os_bsize >> 10) *
1001 msfs->os_blocks) >> 20;
1002 if (d->opd_reserved_mb_low == 0)
1003 d->opd_reserved_mb_low = 1;
1004 d->opd_reserved_mb_high =
1005 (d->opd_reserved_mb_low << 1) + 1;
1007 spin_unlock(&d->opd_pre_lock);
1010 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1011 if (msfs->os_ffree < 32)
1012 msfs->os_state |= OS_STATE_ENOINO;
1013 else if (msfs->os_ffree > 64)
1014 msfs->os_state &= ~OS_STATE_ENOINO;
1016 if (available < d->opd_reserved_mb_low)
1017 msfs->os_state |= OS_STATE_ENOSPC;
1018 else if (available > d->opd_reserved_mb_high)
1019 msfs->os_state &= ~OS_STATE_ENOSPC;
1020 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1021 d->opd_pre_status = -ENOSPC;
1023 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
1024 "free, %llu avail, %llu MB avail, %u "
1025 "hwm -> %d: rc = %d\n",
1026 d->opd_obd->obd_name, msfs->os_blocks,
1027 msfs->os_bfree, msfs->os_bavail,
1028 available, d->opd_reserved_mb_high,
1029 d->opd_pre_status, rc);
1031 "non-committed changes: %u, in progress: %u\n",
1032 atomic_read(&d->opd_sync_changes),
1033 atomic_read(&d->opd_sync_rpcs_in_progress));
1034 } else if (unlikely(old == -ENOSPC)) {
1035 d->opd_pre_status = 0;
1036 spin_lock(&d->opd_pre_lock);
1037 d->opd_pre_create_slow = 0;
1038 d->opd_pre_create_count = OST_MIN_PRECREATE;
1039 spin_unlock(&d->opd_pre_lock);
1040 wake_up(&d->opd_pre_waitq);
1042 CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1043 " free, %llu avail, %lluMB avail, %u lwm"
1044 " -> %d: rc = %d\n", d->opd_obd->obd_name,
1045 msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1046 available, d->opd_reserved_mb_low,
1047 d->opd_pre_status, rc);
1050 /* Object precreation is skipped on the OST with
1051 * max_create_count=0. */
1052 if (d->opd_pre_max_create_count == 0)
1053 msfs->os_state |= OS_STATE_NOPRECREATE;
1055 msfs->os_state &= ~OS_STATE_NOPRECREATE;
1058 wake_up(&d->opd_pre_user_waitq);
1062 * Initialize FID for precreation
1064 * For a just created new target, a new sequence should be taken.
1065 * The function checks there is no IDIF in use (if the target was
1066 * added with the older version of Lustre), then requests a new
1067 * sequence from FLDB using the regular protocol. Then this new
1068 * sequence is stored on a persisten storage synchronously to prevent
1069 * possible object leakage (for the detail see the description for
1070 * osp_precreate_rollover_new_seq()).
1072 * \param[in] osp OSP device
1074 * \retval 0 on success
1075 * \retval negative negated errno on error
1077 int osp_init_pre_fid(struct osp_device *osp)
1080 struct osp_thread_info *osi;
1081 struct lu_client_seq *cli_seq;
1082 struct lu_fid *last_fid;
1086 LASSERT(osp->opd_pre != NULL);
1088 /* Let's check if the current last_seq/fid is valid,
1089 * otherwise request new sequence from the controller */
1090 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1091 /* Non-MDT0 can only use normal sequence for
1093 if (fid_is_norm(&osp->opd_last_used_fid))
1096 /* Initially MDT0 will start with IDIF, after
1097 * that it will request new sequence from the
1099 if (fid_is_idif(&osp->opd_last_used_fid) ||
1100 fid_is_norm(&osp->opd_last_used_fid))
1104 if (!fid_is_zero(&osp->opd_last_used_fid))
1105 CWARN("%s: invalid last used fid "DFID
1106 ", try to get new sequence.\n",
1107 osp->opd_obd->obd_name,
1108 PFID(&osp->opd_last_used_fid));
1110 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1112 CERROR("%s: init env error: rc = %d\n",
1113 osp->opd_obd->obd_name, rc);
1117 osi = osp_env_info(&env);
1118 last_fid = &osi->osi_fid;
1120 /* For a freshed fs, it will allocate a new sequence first */
1121 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1122 cli_seq = osp->opd_obd->u.cli.cl_seq;
1123 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1125 CERROR("%s: alloc fid error: rc = %d\n",
1126 osp->opd_obd->obd_name, rc);
1130 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1132 last_fid->f_oid = 1;
1133 last_fid->f_ver = 0;
1135 spin_lock(&osp->opd_pre_lock);
1136 osp->opd_last_used_fid = *last_fid;
1137 osp->opd_pre_used_fid = *last_fid;
1138 osp->opd_pre_last_created_fid = *last_fid;
1139 spin_unlock(&osp->opd_pre_lock);
1140 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1142 CERROR("%s: write fid error: rc = %d\n",
1143 osp->opd_obd->obd_name, rc);
1152 * The core of precreate functionality
1154 * The function implements the main precreation loop. Basically it
1155 * involves connecting to the target, precerate FID initialization,
1156 * identifying and removing orphans, then serving precreation. As
1157 * part of the latter, the thread is responsible for statfs data
1158 * updates. The precreation is mostly driven by another threads
1159 * asking for new OST objects - those askers wake the thread when
1160 * the number of precreated objects reach low watermark.
1161 * After a disconnect, the sequence above repeats. This is keep going
1162 * until the thread is requested to stop.
1164 * \param[in] _arg private data the thread (OSP device to handle)
1166 * \retval 0 on success
1167 * \retval negative negated errno on error
1169 static int osp_precreate_thread(void *_arg)
1171 struct osp_device *d = _arg;
1172 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1173 struct l_wait_info lwi = { 0 };
1174 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1175 back_to_sleep, NULL);
1181 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1183 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1186 spin_lock(&d->opd_pre_lock);
1187 thread->t_flags = SVC_STOPPED;
1188 spin_unlock(&d->opd_pre_lock);
1189 wake_up(&thread->t_ctl_waitq);
1194 spin_lock(&d->opd_pre_lock);
1195 thread->t_flags = SVC_RUNNING;
1196 spin_unlock(&d->opd_pre_lock);
1197 wake_up(&thread->t_ctl_waitq);
1199 while (osp_precreate_running(d)) {
1201 * need to be connected to OST
1203 while (osp_precreate_running(d)) {
1204 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1205 d->opd_imp_connected &&
1206 !d->opd_got_disconnected)
1208 l_wait_event(d->opd_pre_waitq,
1209 !osp_precreate_running(d) ||
1210 d->opd_new_connection,
1213 if (!d->opd_new_connection)
1216 d->opd_new_connection = 0;
1217 d->opd_got_disconnected = 0;
1221 if (!osp_precreate_running(d))
1225 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1226 /* Sigh, fid client is not ready yet */
1227 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1230 /* Init fid for osp_precreate if necessary */
1231 rc = osp_init_pre_fid(d);
1233 class_export_put(d->opd_exp);
1234 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1235 CERROR("%s: init pre fid error: rc = %d\n",
1236 d->opd_obd->obd_name, rc);
1241 if (osp_statfs_update(&env, d)) {
1242 l_wait_event(d->opd_pre_waitq,
1243 !osp_precreate_running(d), &lwi2);
1249 * Clean up orphans or recreate missing objects.
1251 rc = osp_precreate_cleanup_orphans(&env, d);
1253 schedule_timeout_interruptible(
1254 msecs_to_jiffies(MSEC_PER_SEC));
1260 * connected, can handle precreates now
1262 while (osp_precreate_running(d)) {
1263 l_wait_event(d->opd_pre_waitq,
1264 !osp_precreate_running(d) ||
1265 osp_precreate_near_empty(&env, d) ||
1266 osp_statfs_need_update(d) ||
1267 d->opd_got_disconnected, &lwi);
1269 if (!osp_precreate_running(d))
1272 /* something happened to the connection
1273 * have to start from the beginning */
1274 if (d->opd_got_disconnected)
1277 if (osp_statfs_need_update(d))
1278 if (osp_statfs_update(&env, d))
1281 if (d->opd_pre == NULL)
1284 /* To avoid handling different seq in precreate/orphan
1285 * cleanup, it will hold precreate until current seq is
1287 if (unlikely(osp_precreate_end_seq(&env, d) &&
1288 !osp_create_end_seq(&env, d)))
1291 if (unlikely(osp_precreate_end_seq(&env, d) &&
1292 osp_create_end_seq(&env, d))) {
1293 LCONSOLE_INFO("%s:%#llx is used up."
1294 " Update to new seq\n",
1295 d->opd_obd->obd_name,
1296 fid_seq(&d->opd_pre_last_created_fid));
1297 rc = osp_precreate_rollover_new_seq(&env, d);
1302 if (osp_precreate_near_empty(&env, d)) {
1303 rc = osp_precreate_send(&env, d);
1304 /* osp_precreate_send() sets opd_pre_status
1305 * in case of error, that prevent the using of
1307 if (rc < 0 && rc != -ENOSPC &&
1308 rc != -ETIMEDOUT && rc != -ENOTCONN)
1309 CERROR("%s: cannot precreate objects:"
1311 d->opd_obd->obd_name, rc);
1316 thread->t_flags = SVC_STOPPED;
1318 wake_up(&thread->t_ctl_waitq);
1324 * Check when to stop to wait for precreate objects.
1326 * The caller wanting a new OST object can't wait undefinitely. The
1327 * function checks for few conditions including available new OST
1328 * objects, disconnected OST, lack of space with no pending destroys,
1329 * etc. IOW, it checks whether the current OSP state is good to keep
1330 * waiting or it's better to give up.
1332 * \param[in] env LU environment provided by the caller
1333 * \param[in] d OSP device
1335 * \retval 0 - keep waiting, 1 - no luck
1337 static int osp_precreate_ready_condition(const struct lu_env *env,
1338 struct osp_device *d)
1340 if (d->opd_pre_recovering)
1343 /* ready if got enough precreated objects */
1344 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1345 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1348 /* ready if OST reported no space and no destroys in progress */
1349 if (atomic_read(&d->opd_sync_changes) +
1350 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1351 d->opd_pre_status == -ENOSPC)
1354 /* Bail out I/O fails to OST */
1355 if (d->opd_pre_status != 0 &&
1356 d->opd_pre_status != -EAGAIN &&
1357 d->opd_pre_status != -ENODEV &&
1358 d->opd_pre_status != -ENOTCONN &&
1359 d->opd_pre_status != -ENOSPC) {
1361 if (d->opd_pre_status != -EIO)
1362 CERROR("%s: precreate failed opd_pre_status %d\n",
1363 d->opd_obd->obd_name, d->opd_pre_status);
1370 static int osp_precreate_timeout_condition(void *data)
1372 struct osp_device *d = data;
1374 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1375 "reserved=%llu, sync_changes=%u, "
1376 "sync_rpcs_in_progress=%d, status=%d\n",
1377 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1378 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1379 atomic_read(&d->opd_sync_changes),
1380 atomic_read(&d->opd_sync_rpcs_in_progress),
1387 * Reserve object in precreate pool
1389 * When the caller wants to create a new object on this target (target
1390 * represented by the given OSP), it should declare this intention using
1391 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1392 * to reserve an object in the existing precreated pool or wait up to
1393 * obd_timeout for the available object to appear in the pool (a dedicated
1394 * thread will be doing real precreation in background). The object can be
1395 * consumed later with osp_precreate_get_fid() or be released with call to
1396 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1397 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1398 * If the space on the target is short and there is a pending object destroy,
1399 * then the function forces local commit to speedup space release (see
1400 * osp_sync.c for the details).
1402 * \param[in] env LU environment provided by the caller
1403 * \param[in] d OSP device
1405 * \retval 0 on success
1406 * \retval -ENOSPC when no space on OST
1407 * \retval -EAGAIN try later, slow precreation in progress
1408 * \retval -EIO when no access to OST
1410 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1412 time64_t expire = ktime_get_seconds() + obd_timeout;
1413 struct l_wait_info lwi;
1414 int precreated, rc, synced = 0;
1418 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1419 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1420 PFID(&d->opd_pre_used_fid));
1422 /* opd_pre_max_create_count 0 to not use specified OST. */
1423 if (d->opd_pre_max_create_count == 0)
1428 * - preallocation is done
1429 * - no free space expected soon
1430 * - can't connect to OST for too long (obd_timeout)
1431 * - OST can allocate fid sequence.
1433 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1434 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1437 * increase number of precreations
1439 precreated = osp_objs_precreated(env, d);
1440 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1441 d->opd_pre_create_slow == 0 &&
1442 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1443 spin_lock(&d->opd_pre_lock);
1444 d->opd_pre_create_slow = 1;
1445 d->opd_pre_create_count *= 2;
1446 spin_unlock(&d->opd_pre_lock);
1449 spin_lock(&d->opd_pre_lock);
1450 precreated = osp_objs_precreated(env, d);
1451 if (precreated > d->opd_pre_reserved &&
1452 !d->opd_pre_recovering) {
1453 d->opd_pre_reserved++;
1454 spin_unlock(&d->opd_pre_lock);
1457 /* XXX: don't wake up if precreation is in progress */
1458 if (osp_precreate_near_empty_nolock(env, d) &&
1459 !osp_precreate_end_seq_nolock(env, d))
1460 wake_up(&d->opd_pre_waitq);
1464 spin_unlock(&d->opd_pre_lock);
1467 * all precreated objects have been used and no-space
1468 * status leave us no chance to succeed very soon
1469 * but if there is destroy in progress, then we should
1470 * wait till that is done - some space might be released
1472 if (unlikely(rc == -ENOSPC)) {
1473 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1474 /* force local commit to release space */
1475 dt_commit_async(env, d->opd_storage);
1476 osp_sync_force(env, d);
1479 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1480 /* just wait till destroys are done */
1481 /* see l_wait_even() few lines below */
1483 if (atomic_read(&d->opd_sync_changes) +
1484 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1485 /* no hope for free space */
1490 /* XXX: don't wake up if precreation is in progress */
1491 wake_up(&d->opd_pre_waitq);
1493 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
1494 osp_precreate_timeout_condition, d);
1495 if (ktime_get_seconds() >= expire) {
1500 l_wait_event(d->opd_pre_user_waitq,
1501 osp_precreate_ready_condition(env, d), &lwi);
1508 * Get a FID from precreation pool
1510 * The function is a companion for osp_precreate_reserve() - it assigns
1511 * a specific FID from the precreate. The function should be called only
1512 * if the call to osp_precreate_reserve() was successful. The function
1513 * updates a local storage to remember the highest object ID referenced
1514 * by the node in the given sequence.
1516 * A very importan details: this is supposed to be called once the
1517 * transaction is started, so on-disk update will be atomic with the
1518 * data (like LOVEA) refering this object. Then the object won't be leaked:
1519 * either it's referenced by the committed transaction or it's a subject
1520 * to the orphan cleanup procedure.
1522 * \param[in] env LU environment provided by the caller
1523 * \param[in] d OSP device
1524 * \param[out] fid generated FID
1526 * \retval 0 on success
1527 * \retval negative negated errno on error
1529 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1532 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1533 /* grab next id from the pool */
1534 spin_lock(&d->opd_pre_lock);
1536 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1537 &d->opd_pre_last_created_fid) < 0,
1538 "next fid "DFID" last created fid "DFID"\n",
1539 PFID(&d->opd_pre_used_fid),
1540 PFID(&d->opd_pre_last_created_fid));
1543 * When sequence is used up, new one should be allocated in
1544 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1547 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1548 "next fid "DFID" last created fid "DFID"\n",
1549 PFID(&d->opd_pre_used_fid),
1550 PFID(&d->opd_pre_last_created_fid));
1551 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1552 if (fid_is_idif(pre_used_fid) &&
1553 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1554 pre_used_fid->f_seq++;
1556 d->opd_pre_used_fid.f_oid++;
1557 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1558 d->opd_pre_reserved--;
1560 * last_used_id must be changed along with getting new id otherwise
1561 * we might miscalculate gap causing object loss or leak
1563 osp_update_last_fid(d, fid);
1564 spin_unlock(&d->opd_pre_lock);
1567 * probably main thread suspended orphan cleanup till
1568 * all reservations are released, see comment in
1569 * osp_precreate_thread() just before orphan cleanup
1571 if (unlikely(d->opd_pre_reserved == 0 &&
1572 (d->opd_pre_recovering || d->opd_pre_status)))
1573 wake_up(&d->opd_pre_waitq);
1579 * Set size regular attribute on an object
1581 * When a striping is created late, it's possible that size is already
1582 * initialized on the file. Then the new striping should inherit size
1583 * from the file. The function sets size on the object using the regular
1584 * protocol (OST_PUNCH).
1585 * XXX: should be re-implemented using OUT ?
1587 * \param[in] env LU environment provided by the caller
1588 * \param[in] dt object
1589 * \param[in] size size to set.
1591 * \retval 0 on success
1592 * \retval negative negated errno on error
1594 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1597 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1598 struct ptlrpc_request *req = NULL;
1599 struct obd_import *imp;
1600 struct ost_body *body;
1601 struct obdo *oa = NULL;
1606 imp = d->opd_obd->u.cli.cl_import;
1609 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1613 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1615 ptlrpc_request_free(req);
1620 * XXX: decide how do we do here with resend
1621 * if we don't resend, then client may see wrong file size
1622 * if we do resend, then MDS thread can get stuck for quite long
1623 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1624 * (see LU-7975 and sanity/test_27F use cases)
1625 * but let's decide not to resend/delay this truncate request to OST
1626 * and allow Client to decide to resend, in a less agressive way from
1627 * after_reply(), by returning -EINPROGRESS instead of
1628 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1629 * end of this routine
1631 req->rq_no_resend = req->rq_no_delay = 1;
1633 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1634 ptlrpc_at_set_req_timeout(req);
1638 GOTO(out, rc = -ENOMEM);
1640 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1643 oa->o_blocks = OBD_OBJECT_EOF;
1644 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1645 OBD_MD_FLID | OBD_MD_FLGROUP;
1647 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1649 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1651 /* XXX: capa support? */
1652 /* osc_pack_capa(req, body, capa); */
1654 ptlrpc_request_set_replen(req);
1656 rc = ptlrpc_queue_wait(req);
1658 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1659 * since we have decided not to resend/delay, but this could
1660 * lead to wrong size to be seen at Client side and even process
1661 * trying to open to exit/fail if not itself handling -EAGAIN.
1662 * So it should be better to return -EINPROGRESS instead and
1663 * leave the decision to resend at Client side in after_reply()
1665 if (rc == -EWOULDBLOCK) {
1667 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1668 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1671 CERROR("can't punch object: %d\n", rc);
1675 ptlrpc_req_finished(req);
1682 * Initialize precreation functionality of OSP
1684 * Prepares all the internal structures and starts the precreate thread
1686 * \param[in] d OSP device
1688 * \retval 0 on success
1689 * \retval negative negated errno on error
1691 int osp_init_precreate(struct osp_device *d)
1695 OBD_ALLOC_PTR(d->opd_pre);
1696 if (d->opd_pre == NULL)
1699 /* initially precreation isn't ready */
1700 init_waitqueue_head(&d->opd_pre_user_waitq);
1701 d->opd_pre_status = -EAGAIN;
1702 fid_zero(&d->opd_pre_used_fid);
1703 d->opd_pre_used_fid.f_oid = 1;
1704 fid_zero(&d->opd_pre_last_created_fid);
1705 d->opd_pre_last_created_fid.f_oid = 1;
1706 d->opd_pre_reserved = 0;
1707 d->opd_got_disconnected = 1;
1708 d->opd_pre_create_slow = 0;
1709 d->opd_pre_create_count = OST_MIN_PRECREATE;
1710 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1711 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1712 d->opd_reserved_mb_high = 0;
1713 d->opd_reserved_mb_low = 0;
1719 * Finish precreate functionality of OSP
1722 * Asks all the activity (the thread, update timer) to stop, then
1723 * wait till that is done.
1725 * \param[in] d OSP device
1727 void osp_precreate_fini(struct osp_device *d)
1731 if (d->opd_pre == NULL)
1734 OBD_FREE_PTR(d->opd_pre);
1740 int osp_init_statfs(struct osp_device *d)
1742 struct l_wait_info lwi = { 0 };
1743 struct task_struct *task;
1747 spin_lock_init(&d->opd_pre_lock);
1748 init_waitqueue_head(&d->opd_pre_waitq);
1749 thread_set_flags(&d->opd_pre_thread, SVC_INIT);
1750 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1753 * Initialize statfs-related things
1755 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1756 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1757 1000 * NSEC_PER_SEC);
1758 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1760 ktime_to_ns(d->opd_statfs_fresh_till));
1761 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1762 (unsigned long)d, 0);
1764 if (d->opd_storage->dd_rdonly)
1768 * start thread handling precreation and statfs updates
1770 task = kthread_run(osp_precreate_thread, d,
1771 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1773 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1774 RETURN(PTR_ERR(task));
1777 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1778 osp_precreate_running(d) || osp_precreate_stopped(d),
1784 void osp_statfs_fini(struct osp_device *d)
1786 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1789 del_timer(&d->opd_statfs_timer);
1791 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1792 thread->t_flags = SVC_STOPPING;
1793 wake_up(&d->opd_pre_waitq);
1794 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));