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 !cfs_time_before(cfs_time_current(),
73 d->opd_statfs_fresh_till);
77 * OSP tries to maintain pool of available objects so that calls to create
78 * objects don't block most of time
80 * each time OSP gets connected to OST, we should start from precreation cleanup
82 static inline bool osp_precreate_running(struct osp_device *d)
84 return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
87 static inline bool osp_precreate_stopped(struct osp_device *d)
89 return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
92 static void osp_statfs_timer_cb(unsigned long _d)
94 struct osp_device *d = (struct osp_device *) _d;
97 if (d->opd_pre != NULL && osp_precreate_running(d))
98 wake_up(&d->opd_pre_waitq);
102 * RPC interpret callback for OST_STATFS RPC
104 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
105 * replied by the target. It's used to maintain statfs cache for the target.
106 * The function fills data from the reply if successful and schedules another
109 * \param[in] env LU environment provided by the caller
110 * \param[in] req RPC replied
111 * \param[in] aa callback data
112 * \param[in] rc RPC result
114 * \retval 0 on success
115 * \retval negative negated errno on error
117 static int osp_statfs_interpret(const struct lu_env *env,
118 struct ptlrpc_request *req,
119 union ptlrpc_async_args *aa, int rc)
121 struct obd_import *imp = req->rq_import;
122 struct obd_statfs *msfs;
123 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 d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
144 mod_timer(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
145 d->opd_statfs_update_in_progress = 0;
147 CDEBUG(D_CACHE, "updated statfs %p\n", d);
151 /* couldn't update statfs, try again as soon as possible */
152 if (d->opd_pre != NULL && osp_precreate_running(d))
153 wake_up(&d->opd_pre_waitq);
155 if (req->rq_import_generation == imp->imp_generation)
156 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
157 d->opd_obd->obd_name, rc);
162 * Send OST_STATFS RPC
164 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
165 * Also disables scheduled updates as times OSP may need to refresh
166 * statfs data before expiration. The function doesn't block, instead
167 * an interpretation callback osp_statfs_interpret() is used.
169 * \param[in] d OSP device
171 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
173 struct ptlrpc_request *req;
174 struct obd_import *imp;
175 union ptlrpc_async_args *aa;
180 CDEBUG(D_CACHE, "going to update statfs\n");
182 imp = d->opd_obd->u.cli.cl_import;
185 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
189 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
191 ptlrpc_request_free(req);
194 ptlrpc_request_set_replen(req);
195 req->rq_request_portal = OST_CREATE_PORTAL;
196 ptlrpc_at_set_req_timeout(req);
198 req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
199 aa = ptlrpc_req_async_args(req);
200 aa->pointer_arg[0] = d;
203 * no updates till reply
205 del_timer(&d->opd_statfs_timer);
206 d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
207 d->opd_statfs_update_in_progress = 1;
209 ptlrpcd_add_req(req);
211 /* we still want to sync changes if no new changes are coming */
212 if (cfs_time_before(cfs_time_current(), d->opd_sync_next_commit_cb))
215 if (atomic_read(&d->opd_sync_changes)) {
218 th = dt_trans_create(env, d->opd_storage);
220 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
223 rc = dt_trans_start_local(env, d->opd_storage, th);
225 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
226 d->opd_obd->obd_name,
227 atomic_read(&d->opd_sync_changes));
228 osp_sync_add_commit_cb_1s(env, d, th);
229 dt_trans_stop(env, d->opd_storage, th);
238 * Schedule an immediate update for statfs data
240 * If cached statfs data claim no free space, but OSP has got a request to
241 * destroy an object (so release some space probably), then we may need to
242 * refresh cached statfs data sooner than planned. The function checks there
243 * is no statfs update going and schedules immediate update if so.
244 * XXX: there might be a case where removed object(s) do not add free space (empty
245 * object). If the number of such deletions is high, then we can start to update
246 * statfs too often causing a RPC storm. some throttling is needed...
248 * \param[in] d OSP device where statfs data needs to be refreshed
250 void osp_statfs_need_now(struct osp_device *d)
252 if (!d->opd_statfs_update_in_progress) {
254 * if current status is -ENOSPC (lack of free space on OST)
255 * then we should poll OST immediately once object destroy
258 d->opd_statfs_fresh_till = cfs_time_shift(-1);
259 del_timer(&d->opd_statfs_timer);
260 wake_up(&d->opd_pre_waitq);
265 * Return number of precreated objects
267 * A simple helper to calculate the number of precreated objects on the device.
269 * \param[in] env LU environment provided by the caller
270 * \param[in] osp OSP device
272 * \retval the number of the precreated objects
274 static inline int osp_objs_precreated(const struct lu_env *env,
275 struct osp_device *osp)
277 return osp_fid_diff(&osp->opd_pre_last_created_fid,
278 &osp->opd_pre_used_fid);
282 * Check pool of precreated objects is nearly empty
284 * We should not wait till the pool of the precreated objects is exhausted,
285 * because then there will be a long period of OSP being unavailable for the
286 * new creations due to lenghty precreate RPC. Instead we ask for another
287 * precreation ahead and hopefully have it ready before the current pool is
288 * empty. Notice this function relies on an external locking.
290 * \param[in] env LU environment provided by the caller
291 * \param[in] d OSP device
293 * \retval 0 - current pool is good enough, 1 - time to precreate
295 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
296 struct osp_device *d)
298 int window = osp_objs_precreated(env, d);
300 /* don't consider new precreation till OST is healty and
302 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
303 (d->opd_pre_status == 0));
307 * Check pool of precreated objects
309 * This is protected version of osp_precreate_near_empty_nolock(), check that
312 * \param[in] env LU environment provided by the caller
313 * \param[in] d OSP device
315 * \retval 0 - current pool is good enough, 1 - time to precreate
317 static inline int osp_precreate_near_empty(const struct lu_env *env,
318 struct osp_device *d)
322 /* XXX: do we really need locking here? */
323 spin_lock(&d->opd_pre_lock);
324 rc = osp_precreate_near_empty_nolock(env, d);
325 spin_unlock(&d->opd_pre_lock);
330 * Check given sequence is empty
332 * Returns a binary result whether the given sequence has some IDs left
333 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
334 * version of that function.
336 * \param[in] env LU environment provided by the caller
337 * \param[in] osp OSP device
339 * \retval 0 - current sequence has no IDs, 1 - otherwise
341 static inline int osp_create_end_seq(const struct lu_env *env,
342 struct osp_device *osp)
344 struct lu_fid *fid = &osp->opd_pre_used_fid;
347 spin_lock(&osp->opd_pre_lock);
348 rc = osp_fid_end_seq(env, fid);
349 spin_unlock(&osp->opd_pre_lock);
354 * Write FID into into last_oid/last_seq file
356 * The function stores the sequence and the in-sequence id into two dedicated
357 * files. The sync argument can be used to request synchronous commit, so the
358 * function won't return until the updates are committed.
360 * \param[in] env LU environment provided by the caller
361 * \param[in] osp OSP device
362 * \param[in] fid fid where sequence/id is taken
363 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
365 * \retval 0 on success
366 * \retval negative negated errno on error
368 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
369 struct lu_fid *fid, int sync)
371 struct osp_thread_info *oti = osp_env_info(env);
372 struct lu_buf *lb_oid = &oti->osi_lb;
373 struct lu_buf *lb_oseq = &oti->osi_lb2;
380 if (osp->opd_storage->dd_rdonly)
383 /* Note: through f_oid is only 32 bits, it will also write 64 bits
384 * for oid to keep compatibility with the previous version. */
385 lb_oid->lb_buf = &fid->f_oid;
386 lb_oid->lb_len = sizeof(u64);
387 oid_off = sizeof(u64) * osp->opd_index;
389 lb_oseq->lb_buf = &fid->f_seq;
390 lb_oseq->lb_len = sizeof(u64);
391 oseq_off = sizeof(u64) * osp->opd_index;
393 th = dt_trans_create(env, osp->opd_storage);
398 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
399 lb_oid, oid_off, th);
403 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
404 lb_oseq, oseq_off, th);
408 rc = dt_trans_start_local(env, osp->opd_storage, th);
412 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
415 CERROR("%s: can not write to last seq file: rc = %d\n",
416 osp->opd_obd->obd_name, rc);
419 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
422 CERROR("%s: can not write to last seq file: rc = %d\n",
423 osp->opd_obd->obd_name, rc);
427 dt_trans_stop(env, osp->opd_storage, th);
432 * Switch to another sequence
434 * When a current sequence has no available IDs left, OSP has to switch to
435 * another new sequence. OSP requests it using the regular FLDB protocol
436 * and stores synchronously before that is used in precreated. This is needed
437 * to basically have the sequences referenced (not orphaned), otherwise it's
438 * possible that OST has some objects precreated and the clients have data
439 * written to it, but after MDT failover nobody refers those objects and OSP
440 * has no idea that the sequence need cleanup to be done.
441 * While this is very expensive operation, it's supposed to happen very very
442 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
444 * \param[in] env LU environment provided by the caller
445 * \param[in] osp OSP device
447 * \retval 0 on success
448 * \retval negative negated errno on error
450 static int osp_precreate_rollover_new_seq(struct lu_env *env,
451 struct osp_device *osp)
453 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
454 struct lu_fid *last_fid = &osp->opd_last_used_fid;
458 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
460 CERROR("%s: alloc fid error: rc = %d\n",
461 osp->opd_obd->obd_name, rc);
467 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
468 "fid "DFID", last_fid "DFID"\n", PFID(fid),
471 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
473 CERROR("%s: Can not update oid/seq file: rc = %d\n",
474 osp->opd_obd->obd_name, rc);
478 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
479 osp->opd_obd->obd_name, fid_seq(last_fid),
481 /* Update last_xxx to the new seq */
482 spin_lock(&osp->opd_pre_lock);
483 osp->opd_last_used_fid = *fid;
484 osp->opd_gap_start_fid = *fid;
485 osp->opd_pre_used_fid = *fid;
486 osp->opd_pre_last_created_fid = *fid;
487 spin_unlock(&osp->opd_pre_lock);
493 * Find IDs available in current sequence
495 * The function calculates the highest possible ID and the number of IDs
496 * available in the current sequence OSP is using. The number is limited
497 * artifically by the caller (grow param) and the number of IDs available
498 * in the sequence by nature. The function doesn't require an external
501 * \param[in] env LU environment provided by the caller
502 * \param[in] osp OSP device
503 * \param[in] fid FID the caller wants to start with
504 * \param[in] grow how many the caller wants
505 * \param[out] fid the highest calculated FID
506 * \param[out] grow the number of available IDs calculated
508 * \retval 0 on success, 1 - the sequence is empty
510 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
511 struct lu_fid *fid, int *grow)
513 struct osp_thread_info *osi = osp_env_info(env);
517 if (fid_is_idif(fid)) {
518 struct lu_fid *last_fid;
519 struct ost_id *oi = &osi->osi_oi;
522 spin_lock(&osp->opd_pre_lock);
523 last_fid = &osp->opd_pre_last_created_fid;
524 fid_to_ostid(last_fid, oi);
525 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
526 *grow = end - ostid_id(oi);
527 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
528 spin_unlock(&osp->opd_pre_lock);
530 if (*grow == 0 || rc)
533 ostid_to_fid(fid, oi, osp->opd_index);
537 spin_lock(&osp->opd_pre_lock);
538 *fid = osp->opd_pre_last_created_fid;
540 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
541 *grow = end - fid->f_oid;
542 fid->f_oid += end - fid->f_oid;
543 spin_unlock(&osp->opd_pre_lock);
545 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
546 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
548 return *grow > 0 ? 0 : 1;
552 * Prepare and send precreate RPC
554 * The function finds how many objects should be precreated. Then allocates,
555 * prepares and schedules precreate RPC synchronously. Upon reply the function
556 * wake ups the threads waiting for the new objects on this target. If the
557 * target wasn't able to create all the objects requested, then the next
558 * precreate will be asking less objects (i.e. slow precreate down).
560 * \param[in] env LU environment provided by the caller
561 * \param[in] d OSP device
563 * \retval 0 on success
564 * \retval negative negated errno on error
566 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
568 struct osp_thread_info *oti = osp_env_info(env);
569 struct ptlrpc_request *req;
570 struct obd_import *imp;
571 struct ost_body *body;
573 struct lu_fid *fid = &oti->osi_fid;
576 /* don't precreate new objects till OST healthy and has free space */
577 if (unlikely(d->opd_pre_status)) {
578 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
579 d->opd_obd->obd_name, d->opd_pre_status);
584 * if not connection/initialization is compeleted, ignore
586 imp = d->opd_obd->u.cli.cl_import;
589 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
592 req->rq_request_portal = OST_CREATE_PORTAL;
593 /* we should not resend create request - anyway we will have delorphan
594 * and kill these objects */
595 req->rq_no_delay = req->rq_no_resend = 1;
597 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
599 ptlrpc_request_free(req);
603 spin_lock(&d->opd_pre_lock);
604 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
605 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
606 grow = d->opd_pre_create_count;
607 spin_unlock(&d->opd_pre_lock);
609 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
612 *fid = d->opd_pre_last_created_fid;
613 rc = osp_precreate_fids(env, d, fid, &grow);
615 /* Current seq has been used up*/
616 if (!osp_is_fid_client(d)) {
617 osp_pre_update_status(d, -ENOSPC);
620 wake_up(&d->opd_pre_waitq);
624 if (!osp_is_fid_client(d)) {
625 /* Non-FID client will always send seq 0 because of
627 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
631 fid_to_ostid(fid, &body->oa.o_oi);
632 body->oa.o_valid = OBD_MD_FLGROUP;
634 ptlrpc_request_set_replen(req);
636 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
639 rc = ptlrpc_queue_wait(req);
641 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
645 LASSERT(req->rq_transno == 0);
647 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
649 GOTO(out_req, rc = -EPROTO);
651 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
654 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
655 CERROR("%s: precreate fid "DFID" < local used fid "DFID
656 ": rc = %d\n", d->opd_obd->obd_name,
657 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
658 GOTO(out_req, rc = -ESTALE);
661 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
663 spin_lock(&d->opd_pre_lock);
665 /* the OST has not managed to create all the
666 * objects we asked for */
667 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
668 d->opd_pre_create_slow = 1;
670 /* the OST is able to keep up with the work,
671 * we could consider increasing create_count
672 * next time if needed */
673 d->opd_pre_create_slow = 0;
676 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
677 fid_to_ostid(fid, &body->oa.o_oi);
679 d->opd_pre_last_created_fid = *fid;
680 spin_unlock(&d->opd_pre_lock);
682 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
683 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
684 PFID(&d->opd_pre_last_created_fid));
686 /* now we can wakeup all users awaiting for objects */
687 osp_pre_update_status(d, rc);
688 wake_up(&d->opd_pre_user_waitq);
690 ptlrpc_req_finished(req);
695 * Get last precreated object from target (OST)
697 * Sends synchronous RPC to the target (OST) to learn the last precreated
698 * object. This later is used to remove all unused objects (cleanup orphan
699 * procedure). Also, the next object after one we got will be used as a
700 * starting point for the new precreates.
702 * \param[in] env LU environment provided by the caller
703 * \param[in] d OSP device
705 * \retval 0 on success
706 * \retval negative negated errno on error
708 static int osp_get_lastfid_from_ost(const struct lu_env *env,
709 struct osp_device *d)
711 struct ptlrpc_request *req = NULL;
712 struct obd_import *imp;
713 struct lu_fid *last_fid;
718 imp = d->opd_obd->u.cli.cl_import;
721 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
725 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
726 sizeof(KEY_LAST_FID));
728 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
730 ptlrpc_request_free(req);
734 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
735 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
737 req->rq_no_delay = req->rq_no_resend = 1;
738 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
739 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
741 ptlrpc_request_set_replen(req);
743 rc = ptlrpc_queue_wait(req);
745 /* bad-bad OST.. let sysadm sort this out */
746 if (rc == -ENOTSUPP) {
747 CERROR("%s: server does not support FID: rc = %d\n",
748 d->opd_obd->obd_name, -ENOTSUPP);
750 ptlrpc_set_import_active(imp, 0);
754 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
755 if (last_fid == NULL) {
756 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
757 GOTO(out, rc = -EPROTO);
760 if (!fid_is_sane(last_fid)) {
761 CERROR("%s: Got insane last_fid "DFID"\n",
762 d->opd_obd->obd_name, PFID(last_fid));
763 GOTO(out, rc = -EPROTO);
766 /* Only update the last used fid, if the OST has objects for
767 * this sequence, i.e. fid_oid > 0 */
768 if (fid_oid(last_fid) > 0)
769 d->opd_last_used_fid = *last_fid;
771 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
775 ptlrpc_req_finished(req);
780 * Cleanup orphans on OST
782 * This function is called in a contex of a dedicated thread handling
783 * all the precreation suff. The function waits till local recovery
784 * is complete, then identify all the unreferenced objects (orphans)
785 * using the highest ID referenced by a local and the highest object
786 * precreated by the target. The found range is a subject to removal
787 * using specially flagged RPC. During this process OSP is marked
788 * unavailable for new objects.
790 * \param[in] env LU environment provided by the caller
791 * \param[in] d OSP device
793 * \retval 0 on success
794 * \retval negative negated errno on error
796 static int osp_precreate_cleanup_orphans(struct lu_env *env,
797 struct osp_device *d)
799 struct osp_thread_info *osi = osp_env_info(env);
800 struct lu_fid *last_fid = &osi->osi_fid;
801 struct ptlrpc_request *req = NULL;
802 struct obd_import *imp;
803 struct ost_body *body;
804 struct l_wait_info lwi = { 0 };
805 int update_status = 0;
812 * wait for local recovery to finish, so we can cleanup orphans
813 * orphans are all objects since "last used" (assigned), but
814 * there might be objects reserved and in some cases they won't
815 * be used. we can't cleanup them till we're sure they won't be
816 * used. also can't we allow new reservations because they may
817 * end up getting orphans being cleaned up below. so we block
818 * new reservations and wait till all reserved objects either
821 spin_lock(&d->opd_pre_lock);
822 d->opd_pre_recovering = 1;
823 spin_unlock(&d->opd_pre_lock);
825 * The locking above makes sure the opd_pre_reserved check below will
826 * catch all osp_precreate_reserve() calls who find
827 * "!opd_pre_recovering".
829 l_wait_event(d->opd_pre_waitq,
830 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
831 !osp_precreate_running(d) || d->opd_got_disconnected,
833 if (!osp_precreate_running(d) || d->opd_got_disconnected)
834 GOTO(out, rc = -EAGAIN);
836 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
837 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
839 *last_fid = d->opd_last_used_fid;
840 /* The OSP should already get the valid seq now */
841 LASSERT(!fid_is_zero(last_fid));
842 if (fid_oid(&d->opd_last_used_fid) < 2) {
843 /* lastfid looks strange... ask OST */
844 rc = osp_get_lastfid_from_ost(env, d);
849 imp = d->opd_obd->u.cli.cl_import;
852 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
854 GOTO(out, rc = -ENOMEM);
856 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
858 ptlrpc_request_free(req);
863 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
865 GOTO(out, rc = -EPROTO);
867 body->oa.o_flags = OBD_FL_DELORPHAN;
868 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
870 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
872 ptlrpc_request_set_replen(req);
874 /* Don't resend the delorphan req */
875 req->rq_no_resend = req->rq_no_delay = 1;
877 rc = ptlrpc_queue_wait(req);
883 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
885 GOTO(out, rc = -EPROTO);
888 * OST provides us with id new pool starts from in body->oa.o_id
890 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
892 spin_lock(&d->opd_pre_lock);
893 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
895 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
896 d->opd_pre_last_created_fid = d->opd_last_used_fid;
898 d->opd_pre_create_count = OST_MIN_PRECREATE;
899 d->opd_pre_last_created_fid = *last_fid;
902 * This empties the pre-creation pool and effectively blocks any new
905 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
906 LUSTRE_DATA_SEQ_MAX_WIDTH);
907 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
908 d->opd_pre_create_slow = 0;
909 spin_unlock(&d->opd_pre_lock);
911 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
912 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
913 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
916 ptlrpc_req_finished(req);
919 * If rc is zero, the pre-creation window should have been emptied.
920 * Since waking up the herd would be useless without pre-created
921 * objects, we defer the signal to osp_precreate_send() in that case.
925 CERROR("%s: cannot cleanup orphans: rc = %d\n",
926 d->opd_obd->obd_name, rc);
927 /* we can't proceed from here, OST seem to
928 * be in a bad shape, better to wait for
929 * a new instance of the server and repeat
930 * from the beginning. notify possible waiters
931 * this OSP isn't quite functional yet */
932 osp_pre_update_status(d, rc);
934 wake_up(&d->opd_pre_user_waitq);
937 spin_lock(&d->opd_pre_lock);
938 d->opd_pre_recovering = 0;
939 spin_unlock(&d->opd_pre_lock);
946 * Update precreate status using statfs data
948 * The function decides whether this OSP should be used for new objects.
949 * IOW, whether this OST is used up or has some free space. Cached statfs
950 * data is used to make this decision. If the latest result of statfs
951 * request (rc argument) is not success, then just mark OSP unavailable
954 * Add a bit of hysteresis so this flag isn't continually flapping,
955 * and ensure that new files don't get extremely fragmented due to
956 * only a small amount of available space in the filesystem.
957 * We want to set the ENOSPC when there is less than reserved size
958 * free and clear it when there is at least 2*reserved size free space.
959 * the function updates current precreation status used: functional or not
961 * \param[in] d OSP device
962 * \param[in] rc new precreate status for device \a d
964 * \retval 0 on success
965 * \retval negative negated errno on error
967 void osp_pre_update_status(struct osp_device *d, int rc)
969 struct obd_statfs *msfs = &d->opd_statfs;
970 int old = d->opd_pre_status;
973 d->opd_pre_status = rc;
977 if (likely(msfs->os_type)) {
978 if (unlikely(d->opd_reserved_mb_high == 0 &&
979 d->opd_reserved_mb_low == 0)) {
980 /* Use ~0.1% by default to disable object allocation,
981 * and ~0.2% to enable, size in MB, set both watermark
983 spin_lock(&d->opd_pre_lock);
984 if (d->opd_reserved_mb_high == 0 &&
985 d->opd_reserved_mb_low == 0) {
986 d->opd_reserved_mb_low =
987 ((msfs->os_bsize >> 10) *
988 msfs->os_blocks) >> 20;
989 if (d->opd_reserved_mb_low == 0)
990 d->opd_reserved_mb_low = 1;
991 d->opd_reserved_mb_high =
992 (d->opd_reserved_mb_low << 1) + 1;
994 spin_unlock(&d->opd_pre_lock);
997 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
998 if (msfs->os_ffree < 32)
999 msfs->os_state |= OS_STATE_ENOINO;
1000 else if (msfs->os_ffree > 64)
1001 msfs->os_state &= ~OS_STATE_ENOINO;
1003 if (available < d->opd_reserved_mb_low)
1004 msfs->os_state |= OS_STATE_ENOSPC;
1005 else if (available > d->opd_reserved_mb_high)
1006 msfs->os_state &= ~OS_STATE_ENOSPC;
1007 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1008 d->opd_pre_status = -ENOSPC;
1010 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
1011 "free, %llu avail, %llu MB avail, %u "
1012 "hwm -> %d: rc = %d\n",
1013 d->opd_obd->obd_name, msfs->os_blocks,
1014 msfs->os_bfree, msfs->os_bavail,
1015 available, d->opd_reserved_mb_high,
1016 d->opd_pre_status, rc);
1018 "non-committed changes: %u, in progress: %u\n",
1019 atomic_read(&d->opd_sync_changes),
1020 atomic_read(&d->opd_sync_rpcs_in_progress));
1021 } else if (unlikely(old == -ENOSPC)) {
1022 d->opd_pre_status = 0;
1023 spin_lock(&d->opd_pre_lock);
1024 d->opd_pre_create_slow = 0;
1025 d->opd_pre_create_count = OST_MIN_PRECREATE;
1026 spin_unlock(&d->opd_pre_lock);
1027 wake_up(&d->opd_pre_waitq);
1029 CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1030 " free, %llu avail, %lluMB avail, %u lwm"
1031 " -> %d: rc = %d\n", d->opd_obd->obd_name,
1032 msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1033 available, d->opd_reserved_mb_low,
1034 d->opd_pre_status, rc);
1038 wake_up(&d->opd_pre_user_waitq);
1042 * Initialize FID for precreation
1044 * For a just created new target, a new sequence should be taken.
1045 * The function checks there is no IDIF in use (if the target was
1046 * added with the older version of Lustre), then requests a new
1047 * sequence from FLDB using the regular protocol. Then this new
1048 * sequence is stored on a persisten storage synchronously to prevent
1049 * possible object leakage (for the detail see the description for
1050 * osp_precreate_rollover_new_seq()).
1052 * \param[in] osp OSP device
1054 * \retval 0 on success
1055 * \retval negative negated errno on error
1057 int osp_init_pre_fid(struct osp_device *osp)
1060 struct osp_thread_info *osi;
1061 struct lu_client_seq *cli_seq;
1062 struct lu_fid *last_fid;
1066 LASSERT(osp->opd_pre != NULL);
1068 /* Let's check if the current last_seq/fid is valid,
1069 * otherwise request new sequence from the controller */
1070 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1071 /* Non-MDT0 can only use normal sequence for
1073 if (fid_is_norm(&osp->opd_last_used_fid))
1076 /* Initially MDT0 will start with IDIF, after
1077 * that it will request new sequence from the
1079 if (fid_is_idif(&osp->opd_last_used_fid) ||
1080 fid_is_norm(&osp->opd_last_used_fid))
1084 if (!fid_is_zero(&osp->opd_last_used_fid))
1085 CWARN("%s: invalid last used fid "DFID
1086 ", try to get new sequence.\n",
1087 osp->opd_obd->obd_name,
1088 PFID(&osp->opd_last_used_fid));
1090 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1092 CERROR("%s: init env error: rc = %d\n",
1093 osp->opd_obd->obd_name, rc);
1097 osi = osp_env_info(&env);
1098 last_fid = &osi->osi_fid;
1100 /* For a freshed fs, it will allocate a new sequence first */
1101 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1102 cli_seq = osp->opd_obd->u.cli.cl_seq;
1103 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1105 CERROR("%s: alloc fid error: rc = %d\n",
1106 osp->opd_obd->obd_name, rc);
1110 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1112 last_fid->f_oid = 1;
1113 last_fid->f_ver = 0;
1115 spin_lock(&osp->opd_pre_lock);
1116 osp->opd_last_used_fid = *last_fid;
1117 osp->opd_pre_used_fid = *last_fid;
1118 osp->opd_pre_last_created_fid = *last_fid;
1119 spin_unlock(&osp->opd_pre_lock);
1120 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1122 CERROR("%s: write fid error: rc = %d\n",
1123 osp->opd_obd->obd_name, rc);
1132 * The core of precreate functionality
1134 * The function implements the main precreation loop. Basically it
1135 * involves connecting to the target, precerate FID initialization,
1136 * identifying and removing orphans, then serving precreation. As
1137 * part of the latter, the thread is responsible for statfs data
1138 * updates. The precreation is mostly driven by another threads
1139 * asking for new OST objects - those askers wake the thread when
1140 * the number of precreated objects reach low watermark.
1141 * After a disconnect, the sequence above repeats. This is keep going
1142 * until the thread is requested to stop.
1144 * \param[in] _arg private data the thread (OSP device to handle)
1146 * \retval 0 on success
1147 * \retval negative negated errno on error
1149 static int osp_precreate_thread(void *_arg)
1151 struct osp_device *d = _arg;
1152 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1153 struct l_wait_info lwi = { 0 };
1154 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1155 back_to_sleep, NULL);
1161 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1163 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1166 spin_lock(&d->opd_pre_lock);
1167 thread->t_flags = SVC_STOPPED;
1168 spin_unlock(&d->opd_pre_lock);
1169 wake_up(&thread->t_ctl_waitq);
1174 spin_lock(&d->opd_pre_lock);
1175 thread->t_flags = SVC_RUNNING;
1176 spin_unlock(&d->opd_pre_lock);
1177 wake_up(&thread->t_ctl_waitq);
1179 while (osp_precreate_running(d)) {
1181 * need to be connected to OST
1183 while (osp_precreate_running(d)) {
1184 if (d->opd_pre_recovering &&
1185 d->opd_imp_connected &&
1186 !d->opd_got_disconnected)
1188 l_wait_event(d->opd_pre_waitq,
1189 !osp_precreate_running(d) ||
1190 d->opd_new_connection,
1193 if (!d->opd_new_connection)
1196 d->opd_new_connection = 0;
1197 d->opd_got_disconnected = 0;
1201 if (!osp_precreate_running(d))
1204 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1205 /* Sigh, fid client is not ready yet */
1206 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1209 /* Init fid for osp_precreate if necessary */
1210 rc = osp_init_pre_fid(d);
1212 class_export_put(d->opd_exp);
1213 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1214 CERROR("%s: init pre fid error: rc = %d\n",
1215 d->opd_obd->obd_name, rc);
1219 if (osp_statfs_update(&env, d)) {
1220 l_wait_event(d->opd_pre_waitq,
1221 !osp_precreate_running(d), &lwi2);
1226 * Clean up orphans or recreate missing objects.
1228 rc = osp_precreate_cleanup_orphans(&env, d);
1230 schedule_timeout_interruptible(cfs_time_seconds(1));
1234 * connected, can handle precreates now
1236 while (osp_precreate_running(d)) {
1237 l_wait_event(d->opd_pre_waitq,
1238 !osp_precreate_running(d) ||
1239 osp_precreate_near_empty(&env, d) ||
1240 osp_statfs_need_update(d) ||
1241 d->opd_got_disconnected, &lwi);
1243 if (!osp_precreate_running(d))
1246 /* something happened to the connection
1247 * have to start from the beginning */
1248 if (d->opd_got_disconnected)
1251 if (osp_statfs_need_update(d))
1252 if (osp_statfs_update(&env, d))
1255 /* To avoid handling different seq in precreate/orphan
1256 * cleanup, it will hold precreate until current seq is
1258 if (unlikely(osp_precreate_end_seq(&env, d) &&
1259 !osp_create_end_seq(&env, d)))
1262 if (unlikely(osp_precreate_end_seq(&env, d) &&
1263 osp_create_end_seq(&env, d))) {
1264 LCONSOLE_INFO("%s:%#llx is used up."
1265 " Update to new seq\n",
1266 d->opd_obd->obd_name,
1267 fid_seq(&d->opd_pre_last_created_fid));
1268 rc = osp_precreate_rollover_new_seq(&env, d);
1273 if (osp_precreate_near_empty(&env, d)) {
1274 rc = osp_precreate_send(&env, d);
1275 /* osp_precreate_send() sets opd_pre_status
1276 * in case of error, that prevent the using of
1278 if (rc < 0 && rc != -ENOSPC &&
1279 rc != -ETIMEDOUT && rc != -ENOTCONN)
1280 CERROR("%s: cannot precreate objects:"
1282 d->opd_obd->obd_name, rc);
1287 thread->t_flags = SVC_STOPPED;
1289 wake_up(&thread->t_ctl_waitq);
1295 * Check when to stop to wait for precreate objects.
1297 * The caller wanting a new OST object can't wait undefinitely. The
1298 * function checks for few conditions including available new OST
1299 * objects, disconnected OST, lack of space with no pending destroys,
1300 * etc. IOW, it checks whether the current OSP state is good to keep
1301 * waiting or it's better to give up.
1303 * \param[in] env LU environment provided by the caller
1304 * \param[in] d OSP device
1306 * \retval 0 - keep waiting, 1 - no luck
1308 static int osp_precreate_ready_condition(const struct lu_env *env,
1309 struct osp_device *d)
1311 if (d->opd_pre_recovering)
1314 /* ready if got enough precreated objects */
1315 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1316 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1319 /* ready if OST reported no space and no destroys in progress */
1320 if (atomic_read(&d->opd_sync_changes) +
1321 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1322 d->opd_pre_status == -ENOSPC)
1325 /* Bail out I/O fails to OST */
1326 if (d->opd_pre_status != 0 &&
1327 d->opd_pre_status != -EAGAIN &&
1328 d->opd_pre_status != -ENODEV &&
1329 d->opd_pre_status != -ENOTCONN &&
1330 d->opd_pre_status != -ENOSPC) {
1332 if (d->opd_pre_status != -EIO)
1333 CERROR("%s: precreate failed opd_pre_status %d\n",
1334 d->opd_obd->obd_name, d->opd_pre_status);
1341 static int osp_precreate_timeout_condition(void *data)
1343 struct osp_device *d = data;
1345 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1346 "reserved=%llu, sync_changes=%u, "
1347 "sync_rpcs_in_progress=%d, status=%d\n",
1348 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1349 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1350 atomic_read(&d->opd_sync_changes),
1351 atomic_read(&d->opd_sync_rpcs_in_progress),
1358 * Reserve object in precreate pool
1360 * When the caller wants to create a new object on this target (target
1361 * represented by the given OSP), it should declare this intention using
1362 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1363 * to reserve an object in the existing precreated pool or wait up to
1364 * obd_timeout for the available object to appear in the pool (a dedicated
1365 * thread will be doing real precreation in background). The object can be
1366 * consumed later with osp_precreate_get_fid() or be released with call to
1367 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1368 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1369 * If the space on the target is short and there is a pending object destroy,
1370 * then the function forces local commit to speedup space release (see
1371 * osp_sync.c for the details).
1373 * \param[in] env LU environment provided by the caller
1374 * \param[in] d OSP device
1376 * \retval 0 on success
1377 * \retval -ENOSPC when no space on OST
1378 * \retval -EAGAIN try later, slow precreation in progress
1379 * \retval -EIO when no access to OST
1381 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1383 struct l_wait_info lwi;
1384 cfs_time_t expire = cfs_time_shift(obd_timeout);
1385 int precreated, rc, synced = 0;
1389 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1390 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1391 PFID(&d->opd_pre_used_fid));
1393 /* opd_pre_max_create_count 0 to not use specified OST. */
1394 if (d->opd_pre_max_create_count == 0)
1399 * - preallocation is done
1400 * - no free space expected soon
1401 * - can't connect to OST for too long (obd_timeout)
1402 * - OST can allocate fid sequence.
1404 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1405 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1408 * increase number of precreations
1410 precreated = osp_objs_precreated(env, d);
1411 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1412 d->opd_pre_create_slow == 0 &&
1413 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1414 spin_lock(&d->opd_pre_lock);
1415 d->opd_pre_create_slow = 1;
1416 d->opd_pre_create_count *= 2;
1417 spin_unlock(&d->opd_pre_lock);
1420 spin_lock(&d->opd_pre_lock);
1421 precreated = osp_objs_precreated(env, d);
1422 if (precreated > d->opd_pre_reserved &&
1423 !d->opd_pre_recovering) {
1424 d->opd_pre_reserved++;
1425 spin_unlock(&d->opd_pre_lock);
1428 /* XXX: don't wake up if precreation is in progress */
1429 if (osp_precreate_near_empty_nolock(env, d) &&
1430 !osp_precreate_end_seq_nolock(env, d))
1431 wake_up(&d->opd_pre_waitq);
1435 spin_unlock(&d->opd_pre_lock);
1438 * all precreated objects have been used and no-space
1439 * status leave us no chance to succeed very soon
1440 * but if there is destroy in progress, then we should
1441 * wait till that is done - some space might be released
1443 if (unlikely(rc == -ENOSPC)) {
1444 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1445 /* force local commit to release space */
1446 dt_commit_async(env, d->opd_storage);
1447 osp_sync_force(env, d);
1450 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1451 /* just wait till destroys are done */
1452 /* see l_wait_even() few lines below */
1454 if (atomic_read(&d->opd_sync_changes) +
1455 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1456 /* no hope for free space */
1461 /* XXX: don't wake up if precreation is in progress */
1462 wake_up(&d->opd_pre_waitq);
1464 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1465 osp_precreate_timeout_condition, d);
1466 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1471 l_wait_event(d->opd_pre_user_waitq,
1472 osp_precreate_ready_condition(env, d), &lwi);
1479 * Get a FID from precreation pool
1481 * The function is a companion for osp_precreate_reserve() - it assigns
1482 * a specific FID from the precreate. The function should be called only
1483 * if the call to osp_precreate_reserve() was successful. The function
1484 * updates a local storage to remember the highest object ID referenced
1485 * by the node in the given sequence.
1487 * A very importan details: this is supposed to be called once the
1488 * transaction is started, so on-disk update will be atomic with the
1489 * data (like LOVEA) refering this object. Then the object won't be leaked:
1490 * either it's referenced by the committed transaction or it's a subject
1491 * to the orphan cleanup procedure.
1493 * \param[in] env LU environment provided by the caller
1494 * \param[in] d OSP device
1495 * \param[out] fid generated FID
1497 * \retval 0 on success
1498 * \retval negative negated errno on error
1500 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1503 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1504 /* grab next id from the pool */
1505 spin_lock(&d->opd_pre_lock);
1507 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1508 &d->opd_pre_last_created_fid) < 0,
1509 "next fid "DFID" last created fid "DFID"\n",
1510 PFID(&d->opd_pre_used_fid),
1511 PFID(&d->opd_pre_last_created_fid));
1514 * When sequence is used up, new one should be allocated in
1515 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1518 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1519 "next fid "DFID" last created fid "DFID"\n",
1520 PFID(&d->opd_pre_used_fid),
1521 PFID(&d->opd_pre_last_created_fid));
1522 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1523 if (fid_is_idif(pre_used_fid) &&
1524 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1525 pre_used_fid->f_seq++;
1527 d->opd_pre_used_fid.f_oid++;
1528 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1529 d->opd_pre_reserved--;
1531 * last_used_id must be changed along with getting new id otherwise
1532 * we might miscalculate gap causing object loss or leak
1534 osp_update_last_fid(d, fid);
1535 spin_unlock(&d->opd_pre_lock);
1538 * probably main thread suspended orphan cleanup till
1539 * all reservations are released, see comment in
1540 * osp_precreate_thread() just before orphan cleanup
1542 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1543 wake_up(&d->opd_pre_waitq);
1549 * Set size regular attribute on an object
1551 * When a striping is created late, it's possible that size is already
1552 * initialized on the file. Then the new striping should inherit size
1553 * from the file. The function sets size on the object using the regular
1554 * protocol (OST_PUNCH).
1555 * XXX: should be re-implemented using OUT ?
1557 * \param[in] env LU environment provided by the caller
1558 * \param[in] dt object
1559 * \param[in] size size to set.
1561 * \retval 0 on success
1562 * \retval negative negated errno on error
1564 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1567 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1568 struct ptlrpc_request *req = NULL;
1569 struct obd_import *imp;
1570 struct ost_body *body;
1571 struct obdo *oa = NULL;
1576 imp = d->opd_obd->u.cli.cl_import;
1579 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1583 /* XXX: capa support? */
1584 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1585 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1587 ptlrpc_request_free(req);
1592 * XXX: decide how do we do here with resend
1593 * if we don't resend, then client may see wrong file size
1594 * if we do resend, then MDS thread can get stuck for quite long
1595 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1596 * (see LU-7975 and sanity/test_27F use cases)
1597 * but let's decide not to resend/delay this truncate request to OST
1598 * and allow Client to decide to resend, in a less agressive way from
1599 * after_reply(), by returning -EINPROGRESS instead of
1600 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1601 * end of this routine
1603 req->rq_no_resend = req->rq_no_delay = 1;
1605 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1606 ptlrpc_at_set_req_timeout(req);
1610 GOTO(out, rc = -ENOMEM);
1612 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1615 oa->o_blocks = OBD_OBJECT_EOF;
1616 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1617 OBD_MD_FLID | OBD_MD_FLGROUP;
1619 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1621 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1623 /* XXX: capa support? */
1624 /* osc_pack_capa(req, body, capa); */
1626 ptlrpc_request_set_replen(req);
1628 rc = ptlrpc_queue_wait(req);
1630 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1631 * since we have decided not to resend/delay, but this could
1632 * lead to wrong size to be seen at Client side and even process
1633 * trying to open to exit/fail if not itself handling -EAGAIN.
1634 * So it should be better to return -EINPROGRESS instead and
1635 * leave the decision to resend at Client side in after_reply()
1637 if (rc == -EWOULDBLOCK) {
1639 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1640 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1643 CERROR("can't punch object: %d\n", rc);
1647 ptlrpc_req_finished(req);
1654 * Initialize precreation functionality of OSP
1656 * Prepares all the internal structures and starts the precreate thread
1658 * \param[in] d OSP device
1660 * \retval 0 on success
1661 * \retval negative negated errno on error
1663 int osp_init_precreate(struct osp_device *d)
1665 struct l_wait_info lwi = { 0 };
1666 struct task_struct *task;
1670 OBD_ALLOC_PTR(d->opd_pre);
1671 if (d->opd_pre == NULL)
1674 /* initially precreation isn't ready */
1675 d->opd_pre_status = -EAGAIN;
1676 fid_zero(&d->opd_pre_used_fid);
1677 d->opd_pre_used_fid.f_oid = 1;
1678 fid_zero(&d->opd_pre_last_created_fid);
1679 d->opd_pre_last_created_fid.f_oid = 1;
1680 d->opd_pre_reserved = 0;
1681 d->opd_got_disconnected = 1;
1682 d->opd_pre_create_slow = 0;
1683 d->opd_pre_create_count = OST_MIN_PRECREATE;
1684 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1685 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1686 d->opd_reserved_mb_high = 0;
1687 d->opd_reserved_mb_low = 0;
1689 spin_lock_init(&d->opd_pre_lock);
1690 init_waitqueue_head(&d->opd_pre_waitq);
1691 init_waitqueue_head(&d->opd_pre_user_waitq);
1692 thread_set_flags(&d->opd_pre_thread, SVC_INIT);
1693 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1696 * Initialize statfs-related things
1698 d->opd_statfs_maxage = 5; /* default update interval */
1699 d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1700 CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1701 (unsigned long long)cfs_time_current(),
1702 (unsigned long long)d->opd_statfs_fresh_till);
1703 setup_timer(&d->opd_statfs_timer, osp_statfs_timer_cb,
1706 if (d->opd_storage->dd_rdonly)
1710 * start thread handling precreation and statfs updates
1712 task = kthread_run(osp_precreate_thread, d,
1713 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1715 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1716 RETURN(PTR_ERR(task));
1719 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1720 osp_precreate_running(d) || osp_precreate_stopped(d),
1727 * Finish precreate functionality of OSP
1730 * Asks all the activity (the thread, update timer) to stop, then
1731 * wait till that is done.
1733 * \param[in] d OSP device
1735 void osp_precreate_fini(struct osp_device *d)
1737 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1740 del_timer(&d->opd_statfs_timer);
1742 if (d->opd_pre == NULL)
1745 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1746 thread->t_flags = SVC_STOPPING;
1747 wake_up(&d->opd_pre_waitq);
1748 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1751 OBD_FREE_PTR(d->opd_pre);