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 (d->opd_pre != NULL && 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;
140 osp_pre_update_status(d, rc);
142 /* schedule next update */
143 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
144 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
145 mod_timer(&d->opd_statfs_timer,
146 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
147 d->opd_statfs_update_in_progress = 0;
149 CDEBUG(D_CACHE, "updated statfs %p\n", d);
153 /* couldn't update statfs, try again with a small delay */
154 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
155 d->opd_statfs_update_in_progress = 0;
156 if (d->opd_pre != NULL && osp_precreate_running(d))
157 wake_up(&d->opd_pre_waitq);
159 if (req->rq_import_generation == imp->imp_generation)
160 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
161 d->opd_obd->obd_name, rc);
166 * Send OST_STATFS RPC
168 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
169 * Also disables scheduled updates as times OSP may need to refresh
170 * statfs data before expiration. The function doesn't block, instead
171 * an interpretation callback osp_statfs_interpret() is used.
173 * \param[in] d OSP device
175 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
177 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
178 struct ptlrpc_request *req;
179 struct obd_import *imp;
180 union ptlrpc_async_args *aa;
185 CDEBUG(D_CACHE, "going to update statfs\n");
187 imp = d->opd_obd->u.cli.cl_import;
190 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
194 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
196 ptlrpc_request_free(req);
199 ptlrpc_request_set_replen(req);
200 req->rq_request_portal = OST_CREATE_PORTAL;
201 ptlrpc_at_set_req_timeout(req);
203 req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
204 aa = ptlrpc_req_async_args(req);
205 aa->pointer_arg[0] = d;
208 * no updates till reply
210 del_timer(&d->opd_statfs_timer);
211 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
212 d->opd_statfs_update_in_progress = 1;
214 ptlrpcd_add_req(req);
216 /* we still want to sync changes if no new changes are coming */
217 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
220 if (atomic_read(&d->opd_sync_changes)) {
223 th = dt_trans_create(env, d->opd_storage);
225 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
228 rc = dt_trans_start_local(env, d->opd_storage, th);
230 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
231 d->opd_obd->obd_name,
232 atomic_read(&d->opd_sync_changes));
233 osp_sync_add_commit_cb_1s(env, d, th);
234 dt_trans_stop(env, d->opd_storage, th);
243 * Schedule an immediate update for statfs data
245 * If cached statfs data claim no free space, but OSP has got a request to
246 * destroy an object (so release some space probably), then we may need to
247 * refresh cached statfs data sooner than planned. The function checks there
248 * is no statfs update going and schedules immediate update if so.
249 * XXX: there might be a case where removed object(s) do not add free space (empty
250 * object). If the number of such deletions is high, then we can start to update
251 * statfs too often causing a RPC storm. some throttling is needed...
253 * \param[in] d OSP device where statfs data needs to be refreshed
255 void osp_statfs_need_now(struct osp_device *d)
257 if (!d->opd_statfs_update_in_progress) {
259 * if current status is -ENOSPC (lack of free space on OST)
260 * then we should poll OST immediately once object destroy
263 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
264 del_timer(&d->opd_statfs_timer);
265 wake_up(&d->opd_pre_waitq);
270 * Return number of precreated objects
272 * A simple helper to calculate the number of precreated objects on the device.
274 * \param[in] env LU environment provided by the caller
275 * \param[in] osp OSP device
277 * \retval the number of the precreated objects
279 static inline int osp_objs_precreated(const struct lu_env *env,
280 struct osp_device *osp)
282 return osp_fid_diff(&osp->opd_pre_last_created_fid,
283 &osp->opd_pre_used_fid);
287 * Check pool of precreated objects is nearly empty
289 * We should not wait till the pool of the precreated objects is exhausted,
290 * because then there will be a long period of OSP being unavailable for the
291 * new creations due to lenghty precreate RPC. Instead we ask for another
292 * precreation ahead and hopefully have it ready before the current pool is
293 * empty. Notice this function relies on an external locking.
295 * \param[in] env LU environment provided by the caller
296 * \param[in] d OSP device
298 * \retval 0 - current pool is good enough, 1 - time to precreate
300 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
301 struct osp_device *d)
303 int window = osp_objs_precreated(env, d);
305 /* don't consider new precreation till OST is healty and
307 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
308 (d->opd_pre_status == 0));
312 * Check pool of precreated objects
314 * This is protected version of osp_precreate_near_empty_nolock(), check that
317 * \param[in] env LU environment provided by the caller
318 * \param[in] d OSP device
320 * \retval 0 - current pool is good enough, 1 - time to precreate
322 static inline int osp_precreate_near_empty(const struct lu_env *env,
323 struct osp_device *d)
327 /* XXX: do we really need locking here? */
328 spin_lock(&d->opd_pre_lock);
329 rc = osp_precreate_near_empty_nolock(env, d);
330 spin_unlock(&d->opd_pre_lock);
335 * Check given sequence is empty
337 * Returns a binary result whether the given sequence has some IDs left
338 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
339 * version of that function.
341 * \param[in] env LU environment provided by the caller
342 * \param[in] osp OSP device
344 * \retval 0 - current sequence has no IDs, 1 - otherwise
346 static inline int osp_create_end_seq(const struct lu_env *env,
347 struct osp_device *osp)
349 struct lu_fid *fid = &osp->opd_pre_used_fid;
352 spin_lock(&osp->opd_pre_lock);
353 rc = osp_fid_end_seq(env, fid);
354 spin_unlock(&osp->opd_pre_lock);
359 * Write FID into into last_oid/last_seq file
361 * The function stores the sequence and the in-sequence id into two dedicated
362 * files. The sync argument can be used to request synchronous commit, so the
363 * function won't return until the updates are committed.
365 * \param[in] env LU environment provided by the caller
366 * \param[in] osp OSP device
367 * \param[in] fid fid where sequence/id is taken
368 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
370 * \retval 0 on success
371 * \retval negative negated errno on error
373 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
374 struct lu_fid *fid, int sync)
376 struct osp_thread_info *oti = osp_env_info(env);
377 struct lu_buf *lb_oid = &oti->osi_lb;
378 struct lu_buf *lb_oseq = &oti->osi_lb2;
385 if (osp->opd_storage->dd_rdonly)
388 /* Note: through f_oid is only 32 bits, it will also write 64 bits
389 * for oid to keep compatibility with the previous version. */
390 lb_oid->lb_buf = &fid->f_oid;
391 lb_oid->lb_len = sizeof(u64);
392 oid_off = sizeof(u64) * osp->opd_index;
394 lb_oseq->lb_buf = &fid->f_seq;
395 lb_oseq->lb_len = sizeof(u64);
396 oseq_off = sizeof(u64) * osp->opd_index;
398 th = dt_trans_create(env, osp->opd_storage);
403 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
404 lb_oid, oid_off, th);
408 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
409 lb_oseq, oseq_off, th);
413 rc = dt_trans_start_local(env, osp->opd_storage, th);
417 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
420 CERROR("%s: can not write to last seq file: rc = %d\n",
421 osp->opd_obd->obd_name, rc);
424 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
427 CERROR("%s: can not write to last seq file: rc = %d\n",
428 osp->opd_obd->obd_name, rc);
432 dt_trans_stop(env, osp->opd_storage, th);
437 * Switch to another sequence
439 * When a current sequence has no available IDs left, OSP has to switch to
440 * another new sequence. OSP requests it using the regular FLDB protocol
441 * and stores synchronously before that is used in precreated. This is needed
442 * to basically have the sequences referenced (not orphaned), otherwise it's
443 * possible that OST has some objects precreated and the clients have data
444 * written to it, but after MDT failover nobody refers those objects and OSP
445 * has no idea that the sequence need cleanup to be done.
446 * While this is very expensive operation, it's supposed to happen very very
447 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
449 * \param[in] env LU environment provided by the caller
450 * \param[in] osp OSP device
452 * \retval 0 on success
453 * \retval negative negated errno on error
455 static int osp_precreate_rollover_new_seq(struct lu_env *env,
456 struct osp_device *osp)
458 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
459 struct lu_fid *last_fid = &osp->opd_last_used_fid;
463 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
465 CERROR("%s: alloc fid error: rc = %d\n",
466 osp->opd_obd->obd_name, rc);
472 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
473 "fid "DFID", last_fid "DFID"\n", PFID(fid),
476 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
478 CERROR("%s: Can not update oid/seq file: rc = %d\n",
479 osp->opd_obd->obd_name, rc);
483 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
484 osp->opd_obd->obd_name, fid_seq(last_fid),
486 /* Update last_xxx to the new seq */
487 spin_lock(&osp->opd_pre_lock);
488 osp->opd_last_used_fid = *fid;
489 osp->opd_gap_start_fid = *fid;
490 osp->opd_pre_used_fid = *fid;
491 osp->opd_pre_last_created_fid = *fid;
492 spin_unlock(&osp->opd_pre_lock);
498 * Find IDs available in current sequence
500 * The function calculates the highest possible ID and the number of IDs
501 * available in the current sequence OSP is using. The number is limited
502 * artifically by the caller (grow param) and the number of IDs available
503 * in the sequence by nature. The function doesn't require an external
506 * \param[in] env LU environment provided by the caller
507 * \param[in] osp OSP device
508 * \param[in] fid FID the caller wants to start with
509 * \param[in] grow how many the caller wants
510 * \param[out] fid the highest calculated FID
511 * \param[out] grow the number of available IDs calculated
513 * \retval 0 on success, 1 - the sequence is empty
515 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
516 struct lu_fid *fid, int *grow)
518 struct osp_thread_info *osi = osp_env_info(env);
522 if (fid_is_idif(fid)) {
523 struct lu_fid *last_fid;
524 struct ost_id *oi = &osi->osi_oi;
527 spin_lock(&osp->opd_pre_lock);
528 last_fid = &osp->opd_pre_last_created_fid;
529 fid_to_ostid(last_fid, oi);
530 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
531 *grow = end - ostid_id(oi);
532 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
533 spin_unlock(&osp->opd_pre_lock);
535 if (*grow == 0 || rc)
538 ostid_to_fid(fid, oi, osp->opd_index);
542 spin_lock(&osp->opd_pre_lock);
543 *fid = osp->opd_pre_last_created_fid;
545 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
546 *grow = end - fid->f_oid;
547 fid->f_oid += end - fid->f_oid;
548 spin_unlock(&osp->opd_pre_lock);
550 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
551 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
553 return *grow > 0 ? 0 : 1;
557 * Prepare and send precreate RPC
559 * The function finds how many objects should be precreated. Then allocates,
560 * prepares and schedules precreate RPC synchronously. Upon reply the function
561 * wake ups the threads waiting for the new objects on this target. If the
562 * target wasn't able to create all the objects requested, then the next
563 * precreate will be asking less objects (i.e. slow precreate down).
565 * \param[in] env LU environment provided by the caller
566 * \param[in] d OSP device
568 * \retval 0 on success
569 * \retval negative negated errno on error
571 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
573 struct osp_thread_info *oti = osp_env_info(env);
574 struct ptlrpc_request *req;
575 struct obd_import *imp;
576 struct ost_body *body;
578 struct lu_fid *fid = &oti->osi_fid;
581 /* don't precreate new objects till OST healthy and has free space */
582 if (unlikely(d->opd_pre_status)) {
583 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
584 d->opd_obd->obd_name, d->opd_pre_status);
589 * if not connection/initialization is compeleted, ignore
591 imp = d->opd_obd->u.cli.cl_import;
594 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
597 req->rq_request_portal = OST_CREATE_PORTAL;
598 /* we should not resend create request - anyway we will have delorphan
599 * and kill these objects */
600 req->rq_no_delay = req->rq_no_resend = 1;
602 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
604 ptlrpc_request_free(req);
608 spin_lock(&d->opd_pre_lock);
609 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
610 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
611 grow = d->opd_pre_create_count;
612 spin_unlock(&d->opd_pre_lock);
614 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
617 *fid = d->opd_pre_last_created_fid;
618 rc = osp_precreate_fids(env, d, fid, &grow);
620 /* Current seq has been used up*/
621 if (!osp_is_fid_client(d)) {
622 osp_pre_update_status(d, -ENOSPC);
625 wake_up(&d->opd_pre_waitq);
629 if (!osp_is_fid_client(d)) {
630 /* Non-FID client will always send seq 0 because of
632 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
636 fid_to_ostid(fid, &body->oa.o_oi);
637 body->oa.o_valid = OBD_MD_FLGROUP;
639 ptlrpc_request_set_replen(req);
641 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
644 rc = ptlrpc_queue_wait(req);
646 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
650 LASSERT(req->rq_transno == 0);
652 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
654 GOTO(out_req, rc = -EPROTO);
656 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
659 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
660 CERROR("%s: precreate fid "DFID" < local used fid "DFID
661 ": rc = %d\n", d->opd_obd->obd_name,
662 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
663 GOTO(out_req, rc = -ESTALE);
666 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
668 spin_lock(&d->opd_pre_lock);
670 /* the OST has not managed to create all the
671 * objects we asked for */
672 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
673 d->opd_pre_create_slow = 1;
675 /* the OST is able to keep up with the work,
676 * we could consider increasing create_count
677 * next time if needed */
678 d->opd_pre_create_slow = 0;
681 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
682 fid_to_ostid(fid, &body->oa.o_oi);
684 d->opd_pre_last_created_fid = *fid;
685 spin_unlock(&d->opd_pre_lock);
687 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
688 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
689 PFID(&d->opd_pre_last_created_fid));
691 /* now we can wakeup all users awaiting for objects */
692 osp_pre_update_status(d, rc);
693 wake_up(&d->opd_pre_user_waitq);
695 ptlrpc_req_finished(req);
700 * Get last precreated object from target (OST)
702 * Sends synchronous RPC to the target (OST) to learn the last precreated
703 * object. This later is used to remove all unused objects (cleanup orphan
704 * procedure). Also, the next object after one we got will be used as a
705 * starting point for the new precreates.
707 * \param[in] env LU environment provided by the caller
708 * \param[in] d OSP device
710 * \retval 0 on success
711 * \retval negative negated errno on error
713 static int osp_get_lastfid_from_ost(const struct lu_env *env,
714 struct osp_device *d)
716 struct ptlrpc_request *req = NULL;
717 struct obd_import *imp;
718 struct lu_fid *last_fid;
723 imp = d->opd_obd->u.cli.cl_import;
726 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
730 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
731 sizeof(KEY_LAST_FID));
733 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
735 ptlrpc_request_free(req);
739 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
740 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
742 req->rq_no_delay = req->rq_no_resend = 1;
743 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
744 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
746 ptlrpc_request_set_replen(req);
748 rc = ptlrpc_queue_wait(req);
750 /* bad-bad OST.. let sysadm sort this out */
751 if (rc == -ENOTSUPP) {
752 CERROR("%s: server does not support FID: rc = %d\n",
753 d->opd_obd->obd_name, -ENOTSUPP);
755 ptlrpc_set_import_active(imp, 0);
759 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
760 if (last_fid == NULL) {
761 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
762 GOTO(out, rc = -EPROTO);
765 if (!fid_is_sane(last_fid)) {
766 CERROR("%s: Got insane last_fid "DFID"\n",
767 d->opd_obd->obd_name, PFID(last_fid));
768 GOTO(out, rc = -EPROTO);
771 /* Only update the last used fid, if the OST has objects for
772 * this sequence, i.e. fid_oid > 0 */
773 if (fid_oid(last_fid) > 0)
774 d->opd_last_used_fid = *last_fid;
776 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
780 ptlrpc_req_finished(req);
785 * Cleanup orphans on OST
787 * This function is called in a contex of a dedicated thread handling
788 * all the precreation suff. The function waits till local recovery
789 * is complete, then identify all the unreferenced objects (orphans)
790 * using the highest ID referenced by a local and the highest object
791 * precreated by the target. The found range is a subject to removal
792 * using specially flagged RPC. During this process OSP is marked
793 * unavailable for new objects.
795 * \param[in] env LU environment provided by the caller
796 * \param[in] d OSP device
798 * \retval 0 on success
799 * \retval negative negated errno on error
801 static int osp_precreate_cleanup_orphans(struct lu_env *env,
802 struct osp_device *d)
804 struct osp_thread_info *osi = osp_env_info(env);
805 struct lu_fid *last_fid = &osi->osi_fid;
806 struct ptlrpc_request *req = NULL;
807 struct obd_import *imp;
808 struct ost_body *body;
809 struct l_wait_info lwi = { 0 };
810 int update_status = 0;
817 * wait for local recovery to finish, so we can cleanup orphans
818 * orphans are all objects since "last used" (assigned), but
819 * there might be objects reserved and in some cases they won't
820 * be used. we can't cleanup them till we're sure they won't be
821 * used. also can't we allow new reservations because they may
822 * end up getting orphans being cleaned up below. so we block
823 * new reservations and wait till all reserved objects either
826 spin_lock(&d->opd_pre_lock);
827 d->opd_pre_recovering = 1;
828 spin_unlock(&d->opd_pre_lock);
830 * The locking above makes sure the opd_pre_reserved check below will
831 * catch all osp_precreate_reserve() calls who find
832 * "!opd_pre_recovering".
834 l_wait_event(d->opd_pre_waitq,
835 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
836 !osp_precreate_running(d) || d->opd_got_disconnected,
838 if (!osp_precreate_running(d) || d->opd_got_disconnected)
839 GOTO(out, rc = -EAGAIN);
841 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
842 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
844 *last_fid = d->opd_last_used_fid;
845 /* The OSP should already get the valid seq now */
846 LASSERT(!fid_is_zero(last_fid));
847 if (fid_oid(&d->opd_last_used_fid) < 2) {
848 /* lastfid looks strange... ask OST */
849 rc = osp_get_lastfid_from_ost(env, d);
854 imp = d->opd_obd->u.cli.cl_import;
857 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
859 GOTO(out, rc = -ENOMEM);
861 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
863 ptlrpc_request_free(req);
868 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
870 GOTO(out, rc = -EPROTO);
872 body->oa.o_flags = OBD_FL_DELORPHAN;
873 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
875 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
877 ptlrpc_request_set_replen(req);
879 /* Don't resend the delorphan req */
880 req->rq_no_resend = req->rq_no_delay = 1;
882 rc = ptlrpc_queue_wait(req);
888 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
890 GOTO(out, rc = -EPROTO);
893 * OST provides us with id new pool starts from in body->oa.o_id
895 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
897 spin_lock(&d->opd_pre_lock);
898 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
900 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
901 d->opd_pre_last_created_fid = d->opd_last_used_fid;
903 d->opd_pre_create_count = OST_MIN_PRECREATE;
904 d->opd_pre_last_created_fid = *last_fid;
907 * This empties the pre-creation pool and effectively blocks any new
910 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
911 LUSTRE_DATA_SEQ_MAX_WIDTH);
912 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
913 d->opd_pre_create_slow = 0;
914 spin_unlock(&d->opd_pre_lock);
916 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
917 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
918 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
921 ptlrpc_req_finished(req);
924 * If rc is zero, the pre-creation window should have been emptied.
925 * Since waking up the herd would be useless without pre-created
926 * objects, we defer the signal to osp_precreate_send() in that case.
930 CERROR("%s: cannot cleanup orphans: rc = %d\n",
931 d->opd_obd->obd_name, rc);
932 /* we can't proceed from here, OST seem to
933 * be in a bad shape, better to wait for
934 * a new instance of the server and repeat
935 * from the beginning. notify possible waiters
936 * this OSP isn't quite functional yet */
937 osp_pre_update_status(d, rc);
939 wake_up(&d->opd_pre_user_waitq);
942 spin_lock(&d->opd_pre_lock);
943 d->opd_pre_recovering = 0;
944 spin_unlock(&d->opd_pre_lock);
951 * Update precreate status using statfs data
953 * The function decides whether this OSP should be used for new objects.
954 * IOW, whether this OST is used up or has some free space. Cached statfs
955 * data is used to make this decision. If the latest result of statfs
956 * request (rc argument) is not success, then just mark OSP unavailable
959 * Add a bit of hysteresis so this flag isn't continually flapping,
960 * and ensure that new files don't get extremely fragmented due to
961 * only a small amount of available space in the filesystem.
962 * We want to set the ENOSPC when there is less than reserved size
963 * free and clear it when there is at least 2*reserved size free space.
964 * the function updates current precreation status used: functional or not
966 * \param[in] d OSP device
967 * \param[in] rc new precreate status for device \a d
969 * \retval 0 on success
970 * \retval negative negated errno on error
972 void osp_pre_update_status(struct osp_device *d, int rc)
974 struct obd_statfs *msfs = &d->opd_statfs;
975 int old = d->opd_pre_status;
978 d->opd_pre_status = rc;
982 if (likely(msfs->os_type)) {
983 if (unlikely(d->opd_reserved_mb_high == 0 &&
984 d->opd_reserved_mb_low == 0)) {
985 /* Use ~0.1% by default to disable object allocation,
986 * and ~0.2% to enable, size in MB, set both watermark
988 spin_lock(&d->opd_pre_lock);
989 if (d->opd_reserved_mb_high == 0 &&
990 d->opd_reserved_mb_low == 0) {
991 d->opd_reserved_mb_low =
992 ((msfs->os_bsize >> 10) *
993 msfs->os_blocks) >> 20;
994 if (d->opd_reserved_mb_low == 0)
995 d->opd_reserved_mb_low = 1;
996 d->opd_reserved_mb_high =
997 (d->opd_reserved_mb_low << 1) + 1;
999 spin_unlock(&d->opd_pre_lock);
1002 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1003 if (msfs->os_ffree < 32)
1004 msfs->os_state |= OS_STATE_ENOINO;
1005 else if (msfs->os_ffree > 64)
1006 msfs->os_state &= ~OS_STATE_ENOINO;
1008 if (available < d->opd_reserved_mb_low)
1009 msfs->os_state |= OS_STATE_ENOSPC;
1010 else if (available > d->opd_reserved_mb_high)
1011 msfs->os_state &= ~OS_STATE_ENOSPC;
1012 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1013 d->opd_pre_status = -ENOSPC;
1015 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
1016 "free, %llu avail, %llu MB avail, %u "
1017 "hwm -> %d: rc = %d\n",
1018 d->opd_obd->obd_name, msfs->os_blocks,
1019 msfs->os_bfree, msfs->os_bavail,
1020 available, d->opd_reserved_mb_high,
1021 d->opd_pre_status, rc);
1023 "non-committed changes: %u, in progress: %u\n",
1024 atomic_read(&d->opd_sync_changes),
1025 atomic_read(&d->opd_sync_rpcs_in_progress));
1026 } else if (unlikely(old == -ENOSPC)) {
1027 d->opd_pre_status = 0;
1028 spin_lock(&d->opd_pre_lock);
1029 d->opd_pre_create_slow = 0;
1030 d->opd_pre_create_count = OST_MIN_PRECREATE;
1031 spin_unlock(&d->opd_pre_lock);
1032 wake_up(&d->opd_pre_waitq);
1034 CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1035 " free, %llu avail, %lluMB avail, %u lwm"
1036 " -> %d: rc = %d\n", d->opd_obd->obd_name,
1037 msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1038 available, d->opd_reserved_mb_low,
1039 d->opd_pre_status, rc);
1043 wake_up(&d->opd_pre_user_waitq);
1047 * Initialize FID for precreation
1049 * For a just created new target, a new sequence should be taken.
1050 * The function checks there is no IDIF in use (if the target was
1051 * added with the older version of Lustre), then requests a new
1052 * sequence from FLDB using the regular protocol. Then this new
1053 * sequence is stored on a persisten storage synchronously to prevent
1054 * possible object leakage (for the detail see the description for
1055 * osp_precreate_rollover_new_seq()).
1057 * \param[in] osp OSP device
1059 * \retval 0 on success
1060 * \retval negative negated errno on error
1062 int osp_init_pre_fid(struct osp_device *osp)
1065 struct osp_thread_info *osi;
1066 struct lu_client_seq *cli_seq;
1067 struct lu_fid *last_fid;
1071 LASSERT(osp->opd_pre != NULL);
1073 /* Let's check if the current last_seq/fid is valid,
1074 * otherwise request new sequence from the controller */
1075 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1076 /* Non-MDT0 can only use normal sequence for
1078 if (fid_is_norm(&osp->opd_last_used_fid))
1081 /* Initially MDT0 will start with IDIF, after
1082 * that it will request new sequence from the
1084 if (fid_is_idif(&osp->opd_last_used_fid) ||
1085 fid_is_norm(&osp->opd_last_used_fid))
1089 if (!fid_is_zero(&osp->opd_last_used_fid))
1090 CWARN("%s: invalid last used fid "DFID
1091 ", try to get new sequence.\n",
1092 osp->opd_obd->obd_name,
1093 PFID(&osp->opd_last_used_fid));
1095 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1097 CERROR("%s: init env error: rc = %d\n",
1098 osp->opd_obd->obd_name, rc);
1102 osi = osp_env_info(&env);
1103 last_fid = &osi->osi_fid;
1105 /* For a freshed fs, it will allocate a new sequence first */
1106 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1107 cli_seq = osp->opd_obd->u.cli.cl_seq;
1108 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1110 CERROR("%s: alloc fid error: rc = %d\n",
1111 osp->opd_obd->obd_name, rc);
1115 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1117 last_fid->f_oid = 1;
1118 last_fid->f_ver = 0;
1120 spin_lock(&osp->opd_pre_lock);
1121 osp->opd_last_used_fid = *last_fid;
1122 osp->opd_pre_used_fid = *last_fid;
1123 osp->opd_pre_last_created_fid = *last_fid;
1124 spin_unlock(&osp->opd_pre_lock);
1125 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1127 CERROR("%s: write fid error: rc = %d\n",
1128 osp->opd_obd->obd_name, rc);
1137 * The core of precreate functionality
1139 * The function implements the main precreation loop. Basically it
1140 * involves connecting to the target, precerate FID initialization,
1141 * identifying and removing orphans, then serving precreation. As
1142 * part of the latter, the thread is responsible for statfs data
1143 * updates. The precreation is mostly driven by another threads
1144 * asking for new OST objects - those askers wake the thread when
1145 * the number of precreated objects reach low watermark.
1146 * After a disconnect, the sequence above repeats. This is keep going
1147 * until the thread is requested to stop.
1149 * \param[in] _arg private data the thread (OSP device to handle)
1151 * \retval 0 on success
1152 * \retval negative negated errno on error
1154 static int osp_precreate_thread(void *_arg)
1156 struct osp_device *d = _arg;
1157 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1158 struct l_wait_info lwi = { 0 };
1159 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1160 back_to_sleep, NULL);
1166 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1168 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1171 spin_lock(&d->opd_pre_lock);
1172 thread->t_flags = SVC_STOPPED;
1173 spin_unlock(&d->opd_pre_lock);
1174 wake_up(&thread->t_ctl_waitq);
1179 spin_lock(&d->opd_pre_lock);
1180 thread->t_flags = SVC_RUNNING;
1181 spin_unlock(&d->opd_pre_lock);
1182 wake_up(&thread->t_ctl_waitq);
1184 while (osp_precreate_running(d)) {
1186 * need to be connected to OST
1188 while (osp_precreate_running(d)) {
1189 if (d->opd_pre_recovering &&
1190 d->opd_imp_connected &&
1191 !d->opd_got_disconnected)
1193 l_wait_event(d->opd_pre_waitq,
1194 !osp_precreate_running(d) ||
1195 d->opd_new_connection,
1198 if (!d->opd_new_connection)
1201 d->opd_new_connection = 0;
1202 d->opd_got_disconnected = 0;
1206 if (!osp_precreate_running(d))
1209 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1210 /* Sigh, fid client is not ready yet */
1211 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1214 /* Init fid for osp_precreate if necessary */
1215 rc = osp_init_pre_fid(d);
1217 class_export_put(d->opd_exp);
1218 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1219 CERROR("%s: init pre fid error: rc = %d\n",
1220 d->opd_obd->obd_name, rc);
1224 if (osp_statfs_update(&env, d)) {
1225 l_wait_event(d->opd_pre_waitq,
1226 !osp_precreate_running(d), &lwi2);
1231 * Clean up orphans or recreate missing objects.
1233 rc = osp_precreate_cleanup_orphans(&env, d);
1235 schedule_timeout_interruptible(cfs_time_seconds(1));
1239 * connected, can handle precreates now
1241 while (osp_precreate_running(d)) {
1242 l_wait_event(d->opd_pre_waitq,
1243 !osp_precreate_running(d) ||
1244 osp_precreate_near_empty(&env, d) ||
1245 osp_statfs_need_update(d) ||
1246 d->opd_got_disconnected, &lwi);
1248 if (!osp_precreate_running(d))
1251 /* something happened to the connection
1252 * have to start from the beginning */
1253 if (d->opd_got_disconnected)
1256 if (osp_statfs_need_update(d))
1257 if (osp_statfs_update(&env, d))
1260 /* To avoid handling different seq in precreate/orphan
1261 * cleanup, it will hold precreate until current seq is
1263 if (unlikely(osp_precreate_end_seq(&env, d) &&
1264 !osp_create_end_seq(&env, d)))
1267 if (unlikely(osp_precreate_end_seq(&env, d) &&
1268 osp_create_end_seq(&env, d))) {
1269 LCONSOLE_INFO("%s:%#llx is used up."
1270 " Update to new seq\n",
1271 d->opd_obd->obd_name,
1272 fid_seq(&d->opd_pre_last_created_fid));
1273 rc = osp_precreate_rollover_new_seq(&env, d);
1278 if (osp_precreate_near_empty(&env, d)) {
1279 rc = osp_precreate_send(&env, d);
1280 /* osp_precreate_send() sets opd_pre_status
1281 * in case of error, that prevent the using of
1283 if (rc < 0 && rc != -ENOSPC &&
1284 rc != -ETIMEDOUT && rc != -ENOTCONN)
1285 CERROR("%s: cannot precreate objects:"
1287 d->opd_obd->obd_name, rc);
1292 thread->t_flags = SVC_STOPPED;
1294 wake_up(&thread->t_ctl_waitq);
1300 * Check when to stop to wait for precreate objects.
1302 * The caller wanting a new OST object can't wait undefinitely. The
1303 * function checks for few conditions including available new OST
1304 * objects, disconnected OST, lack of space with no pending destroys,
1305 * etc. IOW, it checks whether the current OSP state is good to keep
1306 * waiting or it's better to give up.
1308 * \param[in] env LU environment provided by the caller
1309 * \param[in] d OSP device
1311 * \retval 0 - keep waiting, 1 - no luck
1313 static int osp_precreate_ready_condition(const struct lu_env *env,
1314 struct osp_device *d)
1316 if (d->opd_pre_recovering)
1319 /* ready if got enough precreated objects */
1320 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1321 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1324 /* ready if OST reported no space and no destroys in progress */
1325 if (atomic_read(&d->opd_sync_changes) +
1326 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1327 d->opd_pre_status == -ENOSPC)
1330 /* Bail out I/O fails to OST */
1331 if (d->opd_pre_status != 0 &&
1332 d->opd_pre_status != -EAGAIN &&
1333 d->opd_pre_status != -ENODEV &&
1334 d->opd_pre_status != -ENOTCONN &&
1335 d->opd_pre_status != -ENOSPC) {
1337 if (d->opd_pre_status != -EIO)
1338 CERROR("%s: precreate failed opd_pre_status %d\n",
1339 d->opd_obd->obd_name, d->opd_pre_status);
1346 static int osp_precreate_timeout_condition(void *data)
1348 struct osp_device *d = data;
1350 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1351 "reserved=%llu, sync_changes=%u, "
1352 "sync_rpcs_in_progress=%d, status=%d\n",
1353 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1354 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1355 atomic_read(&d->opd_sync_changes),
1356 atomic_read(&d->opd_sync_rpcs_in_progress),
1363 * Reserve object in precreate pool
1365 * When the caller wants to create a new object on this target (target
1366 * represented by the given OSP), it should declare this intention using
1367 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1368 * to reserve an object in the existing precreated pool or wait up to
1369 * obd_timeout for the available object to appear in the pool (a dedicated
1370 * thread will be doing real precreation in background). The object can be
1371 * consumed later with osp_precreate_get_fid() or be released with call to
1372 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1373 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1374 * If the space on the target is short and there is a pending object destroy,
1375 * then the function forces local commit to speedup space release (see
1376 * osp_sync.c for the details).
1378 * \param[in] env LU environment provided by the caller
1379 * \param[in] d OSP device
1381 * \retval 0 on success
1382 * \retval -ENOSPC when no space on OST
1383 * \retval -EAGAIN try later, slow precreation in progress
1384 * \retval -EIO when no access to OST
1386 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1388 time64_t expire = ktime_get_seconds() + obd_timeout;
1389 struct l_wait_info lwi;
1390 int precreated, rc, synced = 0;
1394 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1395 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1396 PFID(&d->opd_pre_used_fid));
1398 /* opd_pre_max_create_count 0 to not use specified OST. */
1399 if (d->opd_pre_max_create_count == 0)
1404 * - preallocation is done
1405 * - no free space expected soon
1406 * - can't connect to OST for too long (obd_timeout)
1407 * - OST can allocate fid sequence.
1409 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1410 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1413 * increase number of precreations
1415 precreated = osp_objs_precreated(env, d);
1416 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1417 d->opd_pre_create_slow == 0 &&
1418 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1419 spin_lock(&d->opd_pre_lock);
1420 d->opd_pre_create_slow = 1;
1421 d->opd_pre_create_count *= 2;
1422 spin_unlock(&d->opd_pre_lock);
1425 spin_lock(&d->opd_pre_lock);
1426 precreated = osp_objs_precreated(env, d);
1427 if (precreated > d->opd_pre_reserved &&
1428 !d->opd_pre_recovering) {
1429 d->opd_pre_reserved++;
1430 spin_unlock(&d->opd_pre_lock);
1433 /* XXX: don't wake up if precreation is in progress */
1434 if (osp_precreate_near_empty_nolock(env, d) &&
1435 !osp_precreate_end_seq_nolock(env, d))
1436 wake_up(&d->opd_pre_waitq);
1440 spin_unlock(&d->opd_pre_lock);
1443 * all precreated objects have been used and no-space
1444 * status leave us no chance to succeed very soon
1445 * but if there is destroy in progress, then we should
1446 * wait till that is done - some space might be released
1448 if (unlikely(rc == -ENOSPC)) {
1449 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1450 /* force local commit to release space */
1451 dt_commit_async(env, d->opd_storage);
1452 osp_sync_force(env, d);
1455 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1456 /* just wait till destroys are done */
1457 /* see l_wait_even() few lines below */
1459 if (atomic_read(&d->opd_sync_changes) +
1460 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1461 /* no hope for free space */
1466 /* XXX: don't wake up if precreation is in progress */
1467 wake_up(&d->opd_pre_waitq);
1469 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
1470 osp_precreate_timeout_condition, d);
1471 if (ktime_get_seconds() >= expire) {
1476 l_wait_event(d->opd_pre_user_waitq,
1477 osp_precreate_ready_condition(env, d), &lwi);
1484 * Get a FID from precreation pool
1486 * The function is a companion for osp_precreate_reserve() - it assigns
1487 * a specific FID from the precreate. The function should be called only
1488 * if the call to osp_precreate_reserve() was successful. The function
1489 * updates a local storage to remember the highest object ID referenced
1490 * by the node in the given sequence.
1492 * A very importan details: this is supposed to be called once the
1493 * transaction is started, so on-disk update will be atomic with the
1494 * data (like LOVEA) refering this object. Then the object won't be leaked:
1495 * either it's referenced by the committed transaction or it's a subject
1496 * to the orphan cleanup procedure.
1498 * \param[in] env LU environment provided by the caller
1499 * \param[in] d OSP device
1500 * \param[out] fid generated FID
1502 * \retval 0 on success
1503 * \retval negative negated errno on error
1505 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1508 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1509 /* grab next id from the pool */
1510 spin_lock(&d->opd_pre_lock);
1512 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1513 &d->opd_pre_last_created_fid) < 0,
1514 "next fid "DFID" last created fid "DFID"\n",
1515 PFID(&d->opd_pre_used_fid),
1516 PFID(&d->opd_pre_last_created_fid));
1519 * When sequence is used up, new one should be allocated in
1520 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1523 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1524 "next fid "DFID" last created fid "DFID"\n",
1525 PFID(&d->opd_pre_used_fid),
1526 PFID(&d->opd_pre_last_created_fid));
1527 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1528 if (fid_is_idif(pre_used_fid) &&
1529 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1530 pre_used_fid->f_seq++;
1532 d->opd_pre_used_fid.f_oid++;
1533 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1534 d->opd_pre_reserved--;
1536 * last_used_id must be changed along with getting new id otherwise
1537 * we might miscalculate gap causing object loss or leak
1539 osp_update_last_fid(d, fid);
1540 spin_unlock(&d->opd_pre_lock);
1543 * probably main thread suspended orphan cleanup till
1544 * all reservations are released, see comment in
1545 * osp_precreate_thread() just before orphan cleanup
1547 if (unlikely(d->opd_pre_reserved == 0 &&
1548 (d->opd_pre_recovering || d->opd_pre_status)))
1549 wake_up(&d->opd_pre_waitq);
1555 * Set size regular attribute on an object
1557 * When a striping is created late, it's possible that size is already
1558 * initialized on the file. Then the new striping should inherit size
1559 * from the file. The function sets size on the object using the regular
1560 * protocol (OST_PUNCH).
1561 * XXX: should be re-implemented using OUT ?
1563 * \param[in] env LU environment provided by the caller
1564 * \param[in] dt object
1565 * \param[in] size size to set.
1567 * \retval 0 on success
1568 * \retval negative negated errno on error
1570 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1573 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1574 struct ptlrpc_request *req = NULL;
1575 struct obd_import *imp;
1576 struct ost_body *body;
1577 struct obdo *oa = NULL;
1582 imp = d->opd_obd->u.cli.cl_import;
1585 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1589 /* XXX: capa support? */
1590 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1591 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1593 ptlrpc_request_free(req);
1598 * XXX: decide how do we do here with resend
1599 * if we don't resend, then client may see wrong file size
1600 * if we do resend, then MDS thread can get stuck for quite long
1601 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1602 * (see LU-7975 and sanity/test_27F use cases)
1603 * but let's decide not to resend/delay this truncate request to OST
1604 * and allow Client to decide to resend, in a less agressive way from
1605 * after_reply(), by returning -EINPROGRESS instead of
1606 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1607 * end of this routine
1609 req->rq_no_resend = req->rq_no_delay = 1;
1611 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1612 ptlrpc_at_set_req_timeout(req);
1616 GOTO(out, rc = -ENOMEM);
1618 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1621 oa->o_blocks = OBD_OBJECT_EOF;
1622 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1623 OBD_MD_FLID | OBD_MD_FLGROUP;
1625 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1627 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1629 /* XXX: capa support? */
1630 /* osc_pack_capa(req, body, capa); */
1632 ptlrpc_request_set_replen(req);
1634 rc = ptlrpc_queue_wait(req);
1636 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1637 * since we have decided not to resend/delay, but this could
1638 * lead to wrong size to be seen at Client side and even process
1639 * trying to open to exit/fail if not itself handling -EAGAIN.
1640 * So it should be better to return -EINPROGRESS instead and
1641 * leave the decision to resend at Client side in after_reply()
1643 if (rc == -EWOULDBLOCK) {
1645 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1646 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1649 CERROR("can't punch object: %d\n", rc);
1653 ptlrpc_req_finished(req);
1660 * Initialize precreation functionality of OSP
1662 * Prepares all the internal structures and starts the precreate thread
1664 * \param[in] d OSP device
1666 * \retval 0 on success
1667 * \retval negative negated errno on error
1669 int osp_init_precreate(struct osp_device *d)
1671 struct l_wait_info lwi = { 0 };
1672 struct task_struct *task;
1676 OBD_ALLOC_PTR(d->opd_pre);
1677 if (d->opd_pre == NULL)
1680 /* initially precreation isn't ready */
1681 d->opd_pre_status = -EAGAIN;
1682 fid_zero(&d->opd_pre_used_fid);
1683 d->opd_pre_used_fid.f_oid = 1;
1684 fid_zero(&d->opd_pre_last_created_fid);
1685 d->opd_pre_last_created_fid.f_oid = 1;
1686 d->opd_pre_reserved = 0;
1687 d->opd_got_disconnected = 1;
1688 d->opd_pre_create_slow = 0;
1689 d->opd_pre_create_count = OST_MIN_PRECREATE;
1690 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1691 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1692 d->opd_reserved_mb_high = 0;
1693 d->opd_reserved_mb_low = 0;
1695 spin_lock_init(&d->opd_pre_lock);
1696 init_waitqueue_head(&d->opd_pre_waitq);
1697 init_waitqueue_head(&d->opd_pre_user_waitq);
1698 thread_set_flags(&d->opd_pre_thread, SVC_INIT);
1699 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1702 * Initialize statfs-related things
1704 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1705 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1706 1000 * NSEC_PER_SEC);
1707 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1709 ktime_to_ns(d->opd_statfs_fresh_till));
1710 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1711 (unsigned long)d, 0);
1713 if (d->opd_storage->dd_rdonly)
1717 * start thread handling precreation and statfs updates
1719 task = kthread_run(osp_precreate_thread, d,
1720 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1722 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1723 RETURN(PTR_ERR(task));
1726 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1727 osp_precreate_running(d) || osp_precreate_stopped(d),
1734 * Finish precreate functionality of OSP
1737 * Asks all the activity (the thread, update timer) to stop, then
1738 * wait till that is done.
1740 * \param[in] d OSP device
1742 void osp_precreate_fini(struct osp_device *d)
1744 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1747 del_timer(&d->opd_statfs_timer);
1749 if (d->opd_pre == NULL)
1752 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1753 thread->t_flags = SVC_STOPPING;
1754 wake_up(&d->opd_pre_waitq);
1755 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1758 OBD_FREE_PTR(d->opd_pre);