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, 2016, 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_sync.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(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);
215 * Schedule an immediate update for statfs data
217 * If cached statfs data claim no free space, but OSP has got a request to
218 * destroy an object (so release some space probably), then we may need to
219 * refresh cached statfs data sooner than planned. The function checks there
220 * is no statfs update going and schedules immediate update if so.
221 * XXX: there might be a case where removed object(s) do not add free space (empty
222 * object). If the number of such deletions is high, then we can start to update
223 * statfs too often causing a RPC storm. some throttling is needed...
225 * \param[in] d OSP device where statfs data needs to be refreshed
227 void osp_statfs_need_now(struct osp_device *d)
229 if (!d->opd_statfs_update_in_progress) {
231 * if current status is -ENOSPC (lack of free space on OST)
232 * then we should poll OST immediately once object destroy
235 d->opd_statfs_fresh_till = cfs_time_shift(-1);
236 del_timer(&d->opd_statfs_timer);
237 wake_up(&d->opd_pre_waitq);
242 * Return number of precreated objects
244 * A simple helper to calculate the number of precreated objects on the device.
246 * \param[in] env LU environment provided by the caller
247 * \param[in] osp OSP device
249 * \retval the number of the precreated objects
251 static inline int osp_objs_precreated(const struct lu_env *env,
252 struct osp_device *osp)
254 return osp_fid_diff(&osp->opd_pre_last_created_fid,
255 &osp->opd_pre_used_fid);
259 * Check pool of precreated objects is nearly empty
261 * We should not wait till the pool of the precreated objects is exhausted,
262 * because then there will be a long period of OSP being unavailable for the
263 * new creations due to lenghty precreate RPC. Instead we ask for another
264 * precreation ahead and hopefully have it ready before the current pool is
265 * empty. Notice this function relies on an external locking.
267 * \param[in] env LU environment provided by the caller
268 * \param[in] d OSP device
270 * \retval 0 - current pool is good enough, 1 - time to precreate
272 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
273 struct osp_device *d)
275 int window = osp_objs_precreated(env, d);
277 /* don't consider new precreation till OST is healty and
279 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
280 (d->opd_pre_status == 0));
284 * Check pool of precreated objects
286 * This is protected version of osp_precreate_near_empty_nolock(), check that
289 * \param[in] env LU environment provided by the caller
290 * \param[in] d OSP device
292 * \retval 0 - current pool is good enough, 1 - time to precreate
294 static inline int osp_precreate_near_empty(const struct lu_env *env,
295 struct osp_device *d)
299 /* XXX: do we really need locking here? */
300 spin_lock(&d->opd_pre_lock);
301 rc = osp_precreate_near_empty_nolock(env, d);
302 spin_unlock(&d->opd_pre_lock);
307 * Check given sequence is empty
309 * Returns a binary result whether the given sequence has some IDs left
310 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
311 * version of that function.
313 * \param[in] env LU environment provided by the caller
314 * \param[in] osp OSP device
316 * \retval 0 - current sequence has no IDs, 1 - otherwise
318 static inline int osp_create_end_seq(const struct lu_env *env,
319 struct osp_device *osp)
321 struct lu_fid *fid = &osp->opd_pre_used_fid;
324 spin_lock(&osp->opd_pre_lock);
325 rc = osp_fid_end_seq(env, fid);
326 spin_unlock(&osp->opd_pre_lock);
331 * Write FID into into last_oid/last_seq file
333 * The function stores the sequence and the in-sequence id into two dedicated
334 * files. The sync argument can be used to request synchronous commit, so the
335 * function won't return until the updates are committed.
337 * \param[in] env LU environment provided by the caller
338 * \param[in] osp OSP device
339 * \param[in] fid fid where sequence/id is taken
340 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
342 * \retval 0 on success
343 * \retval negative negated errno on error
345 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
346 struct lu_fid *fid, int sync)
348 struct osp_thread_info *oti = osp_env_info(env);
349 struct lu_buf *lb_oid = &oti->osi_lb;
350 struct lu_buf *lb_oseq = &oti->osi_lb2;
357 if (osp->opd_storage->dd_rdonly)
360 /* Note: through f_oid is only 32 bits, it will also write 64 bits
361 * for oid to keep compatibility with the previous version. */
362 lb_oid->lb_buf = &fid->f_oid;
363 lb_oid->lb_len = sizeof(u64);
364 oid_off = sizeof(u64) * osp->opd_index;
366 lb_oseq->lb_buf = &fid->f_seq;
367 lb_oseq->lb_len = sizeof(u64);
368 oseq_off = sizeof(u64) * osp->opd_index;
370 th = dt_trans_create(env, osp->opd_storage);
375 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
376 lb_oid, oid_off, th);
380 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
381 lb_oseq, oseq_off, th);
385 rc = dt_trans_start_local(env, osp->opd_storage, th);
389 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
392 CERROR("%s: can not write to last seq file: rc = %d\n",
393 osp->opd_obd->obd_name, rc);
396 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
399 CERROR("%s: can not write to last seq file: rc = %d\n",
400 osp->opd_obd->obd_name, rc);
404 dt_trans_stop(env, osp->opd_storage, th);
409 * Switch to another sequence
411 * When a current sequence has no available IDs left, OSP has to switch to
412 * another new sequence. OSP requests it using the regular FLDB protocol
413 * and stores synchronously before that is used in precreated. This is needed
414 * to basically have the sequences referenced (not orphaned), otherwise it's
415 * possible that OST has some objects precreated and the clients have data
416 * written to it, but after MDT failover nobody refers those objects and OSP
417 * has no idea that the sequence need cleanup to be done.
418 * While this is very expensive operation, it's supposed to happen very very
419 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
421 * \param[in] env LU environment provided by the caller
422 * \param[in] osp OSP device
424 * \retval 0 on success
425 * \retval negative negated errno on error
427 static int osp_precreate_rollover_new_seq(struct lu_env *env,
428 struct osp_device *osp)
430 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
431 struct lu_fid *last_fid = &osp->opd_last_used_fid;
435 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
437 CERROR("%s: alloc fid error: rc = %d\n",
438 osp->opd_obd->obd_name, rc);
444 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
445 "fid "DFID", last_fid "DFID"\n", PFID(fid),
448 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
450 CERROR("%s: Can not update oid/seq file: rc = %d\n",
451 osp->opd_obd->obd_name, rc);
455 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
456 osp->opd_obd->obd_name, fid_seq(last_fid),
458 /* Update last_xxx to the new seq */
459 spin_lock(&osp->opd_pre_lock);
460 osp->opd_last_used_fid = *fid;
461 osp->opd_gap_start_fid = *fid;
462 osp->opd_pre_used_fid = *fid;
463 osp->opd_pre_last_created_fid = *fid;
464 spin_unlock(&osp->opd_pre_lock);
470 * Find IDs available in current sequence
472 * The function calculates the highest possible ID and the number of IDs
473 * available in the current sequence OSP is using. The number is limited
474 * artifically by the caller (grow param) and the number of IDs available
475 * in the sequence by nature. The function doesn't require an external
478 * \param[in] env LU environment provided by the caller
479 * \param[in] osp OSP device
480 * \param[in] fid FID the caller wants to start with
481 * \param[in] grow how many the caller wants
482 * \param[out] fid the highest calculated FID
483 * \param[out] grow the number of available IDs calculated
485 * \retval 0 on success, 1 - the sequence is empty
487 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
488 struct lu_fid *fid, int *grow)
490 struct osp_thread_info *osi = osp_env_info(env);
494 if (fid_is_idif(fid)) {
495 struct lu_fid *last_fid;
496 struct ost_id *oi = &osi->osi_oi;
499 spin_lock(&osp->opd_pre_lock);
500 last_fid = &osp->opd_pre_last_created_fid;
501 fid_to_ostid(last_fid, oi);
502 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
503 *grow = end - ostid_id(oi);
504 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
505 spin_unlock(&osp->opd_pre_lock);
507 if (*grow == 0 || rc)
510 ostid_to_fid(fid, oi, osp->opd_index);
514 spin_lock(&osp->opd_pre_lock);
515 *fid = osp->opd_pre_last_created_fid;
517 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
518 *grow = end - fid->f_oid;
519 fid->f_oid += end - fid->f_oid;
520 spin_unlock(&osp->opd_pre_lock);
522 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
523 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
525 return *grow > 0 ? 0 : 1;
529 * Prepare and send precreate RPC
531 * The function finds how many objects should be precreated. Then allocates,
532 * prepares and schedules precreate RPC synchronously. Upon reply the function
533 * wake ups the threads waiting for the new objects on this target. If the
534 * target wasn't able to create all the objects requested, then the next
535 * precreate will be asking less objects (i.e. slow precreate down).
537 * \param[in] env LU environment provided by the caller
538 * \param[in] d OSP device
540 * \retval 0 on success
541 * \retval negative negated errno on error
543 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
545 struct osp_thread_info *oti = osp_env_info(env);
546 struct ptlrpc_request *req;
547 struct obd_import *imp;
548 struct ost_body *body;
550 struct lu_fid *fid = &oti->osi_fid;
553 /* don't precreate new objects till OST healthy and has free space */
554 if (unlikely(d->opd_pre_status)) {
555 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
556 d->opd_obd->obd_name, d->opd_pre_status);
561 * if not connection/initialization is compeleted, ignore
563 imp = d->opd_obd->u.cli.cl_import;
566 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
569 req->rq_request_portal = OST_CREATE_PORTAL;
570 /* we should not resend create request - anyway we will have delorphan
571 * and kill these objects */
572 req->rq_no_delay = req->rq_no_resend = 1;
574 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
576 ptlrpc_request_free(req);
580 spin_lock(&d->opd_pre_lock);
581 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
582 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
583 grow = d->opd_pre_create_count;
584 spin_unlock(&d->opd_pre_lock);
586 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
589 *fid = d->opd_pre_last_created_fid;
590 rc = osp_precreate_fids(env, d, fid, &grow);
592 /* Current seq has been used up*/
593 if (!osp_is_fid_client(d)) {
594 osp_pre_update_status(d, -ENOSPC);
597 wake_up(&d->opd_pre_waitq);
601 if (!osp_is_fid_client(d)) {
602 /* Non-FID client will always send seq 0 because of
604 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
608 fid_to_ostid(fid, &body->oa.o_oi);
609 body->oa.o_valid = OBD_MD_FLGROUP;
611 ptlrpc_request_set_replen(req);
613 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
616 rc = ptlrpc_queue_wait(req);
618 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
622 LASSERT(req->rq_transno == 0);
624 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
626 GOTO(out_req, rc = -EPROTO);
628 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
631 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
632 CERROR("%s: precreate fid "DFID" < local used fid "DFID
633 ": rc = %d\n", d->opd_obd->obd_name,
634 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
635 GOTO(out_req, rc = -ESTALE);
638 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
640 spin_lock(&d->opd_pre_lock);
642 /* the OST has not managed to create all the
643 * objects we asked for */
644 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
645 d->opd_pre_create_slow = 1;
647 /* the OST is able to keep up with the work,
648 * we could consider increasing create_count
649 * next time if needed */
650 d->opd_pre_create_slow = 0;
653 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
654 fid_to_ostid(fid, &body->oa.o_oi);
656 d->opd_pre_last_created_fid = *fid;
657 spin_unlock(&d->opd_pre_lock);
659 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
660 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
661 PFID(&d->opd_pre_last_created_fid));
663 /* now we can wakeup all users awaiting for objects */
664 osp_pre_update_status(d, rc);
665 wake_up(&d->opd_pre_user_waitq);
667 ptlrpc_req_finished(req);
672 * Get last precreated object from target (OST)
674 * Sends synchronous RPC to the target (OST) to learn the last precreated
675 * object. This later is used to remove all unused objects (cleanup orphan
676 * procedure). Also, the next object after one we got will be used as a
677 * starting point for the new precreates.
679 * \param[in] env LU environment provided by the caller
680 * \param[in] d OSP device
682 * \retval 0 on success
683 * \retval negative negated errno on error
685 static int osp_get_lastfid_from_ost(const struct lu_env *env,
686 struct osp_device *d)
688 struct ptlrpc_request *req = NULL;
689 struct obd_import *imp;
690 struct lu_fid *last_fid;
695 imp = d->opd_obd->u.cli.cl_import;
698 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
702 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
703 sizeof(KEY_LAST_FID));
705 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
707 ptlrpc_request_free(req);
711 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
712 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
714 req->rq_no_delay = req->rq_no_resend = 1;
715 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
716 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
718 ptlrpc_request_set_replen(req);
720 rc = ptlrpc_queue_wait(req);
722 /* bad-bad OST.. let sysadm sort this out */
723 if (rc == -ENOTSUPP) {
724 CERROR("%s: server does not support FID: rc = %d\n",
725 d->opd_obd->obd_name, -ENOTSUPP);
727 ptlrpc_set_import_active(imp, 0);
731 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
732 if (last_fid == NULL) {
733 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
734 GOTO(out, rc = -EPROTO);
737 if (!fid_is_sane(last_fid)) {
738 CERROR("%s: Got insane last_fid "DFID"\n",
739 d->opd_obd->obd_name, PFID(last_fid));
740 GOTO(out, rc = -EPROTO);
743 /* Only update the last used fid, if the OST has objects for
744 * this sequence, i.e. fid_oid > 0 */
745 if (fid_oid(last_fid) > 0)
746 d->opd_last_used_fid = *last_fid;
748 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
752 ptlrpc_req_finished(req);
757 * Cleanup orphans on OST
759 * This function is called in a contex of a dedicated thread handling
760 * all the precreation suff. The function waits till local recovery
761 * is complete, then identify all the unreferenced objects (orphans)
762 * using the highest ID referenced by a local and the highest object
763 * precreated by the target. The found range is a subject to removal
764 * using specially flagged RPC. During this process OSP is marked
765 * unavailable for new objects.
767 * \param[in] env LU environment provided by the caller
768 * \param[in] d OSP device
770 * \retval 0 on success
771 * \retval negative negated errno on error
773 static int osp_precreate_cleanup_orphans(struct lu_env *env,
774 struct osp_device *d)
776 struct osp_thread_info *osi = osp_env_info(env);
777 struct lu_fid *last_fid = &osi->osi_fid;
778 struct ptlrpc_request *req = NULL;
779 struct obd_import *imp;
780 struct ost_body *body;
781 struct l_wait_info lwi = { 0 };
782 int update_status = 0;
789 * wait for local recovery to finish, so we can cleanup orphans
790 * orphans are all objects since "last used" (assigned), but
791 * there might be objects reserved and in some cases they won't
792 * be used. we can't cleanup them till we're sure they won't be
793 * used. also can't we allow new reservations because they may
794 * end up getting orphans being cleaned up below. so we block
795 * new reservations and wait till all reserved objects either
798 spin_lock(&d->opd_pre_lock);
799 d->opd_pre_recovering = 1;
800 spin_unlock(&d->opd_pre_lock);
802 * The locking above makes sure the opd_pre_reserved check below will
803 * catch all osp_precreate_reserve() calls who find
804 * "!opd_pre_recovering".
806 l_wait_event(d->opd_pre_waitq,
807 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
808 !osp_precreate_running(d) || d->opd_got_disconnected,
810 if (!osp_precreate_running(d) || d->opd_got_disconnected)
811 GOTO(out, rc = -EAGAIN);
813 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
814 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
816 *last_fid = d->opd_last_used_fid;
817 /* The OSP should already get the valid seq now */
818 LASSERT(!fid_is_zero(last_fid));
819 if (fid_oid(&d->opd_last_used_fid) < 2) {
820 /* lastfid looks strange... ask OST */
821 rc = osp_get_lastfid_from_ost(env, d);
826 imp = d->opd_obd->u.cli.cl_import;
829 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
831 GOTO(out, rc = -ENOMEM);
833 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
835 ptlrpc_request_free(req);
840 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
842 GOTO(out, rc = -EPROTO);
844 body->oa.o_flags = OBD_FL_DELORPHAN;
845 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
847 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
849 ptlrpc_request_set_replen(req);
851 /* Don't resend the delorphan req */
852 req->rq_no_resend = req->rq_no_delay = 1;
854 rc = ptlrpc_queue_wait(req);
860 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
862 GOTO(out, rc = -EPROTO);
865 * OST provides us with id new pool starts from in body->oa.o_id
867 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
869 spin_lock(&d->opd_pre_lock);
870 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
872 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
873 d->opd_pre_last_created_fid = d->opd_last_used_fid;
875 d->opd_pre_create_count = OST_MIN_PRECREATE;
876 d->opd_pre_last_created_fid = *last_fid;
879 * This empties the pre-creation pool and effectively blocks any new
882 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
883 LUSTRE_DATA_SEQ_MAX_WIDTH);
884 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
885 d->opd_pre_create_slow = 0;
886 spin_unlock(&d->opd_pre_lock);
888 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
889 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
890 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
893 ptlrpc_req_finished(req);
896 * If rc is zero, the pre-creation window should have been emptied.
897 * Since waking up the herd would be useless without pre-created
898 * objects, we defer the signal to osp_precreate_send() in that case.
902 CERROR("%s: cannot cleanup orphans: rc = %d\n",
903 d->opd_obd->obd_name, rc);
904 /* we can't proceed from here, OST seem to
905 * be in a bad shape, better to wait for
906 * a new instance of the server and repeat
907 * from the beginning. notify possible waiters
908 * this OSP isn't quite functional yet */
909 osp_pre_update_status(d, rc);
911 wake_up(&d->opd_pre_user_waitq);
914 spin_lock(&d->opd_pre_lock);
915 d->opd_pre_recovering = 0;
916 spin_unlock(&d->opd_pre_lock);
923 * Update precreate status using statfs data
925 * The function decides whether this OSP should be used for new objects.
926 * IOW, whether this OST is used up or has some free space. Cached statfs
927 * data is used to make this decision. If the latest result of statfs
928 * request (rc argument) is not success, then just mark OSP unavailable
931 * Add a bit of hysteresis so this flag isn't continually flapping,
932 * and ensure that new files don't get extremely fragmented due to
933 * only a small amount of available space in the filesystem.
934 * We want to set the ENOSPC when there is less than reserved size
935 * free and clear it when there is at least 2*reserved size free space.
936 * the function updates current precreation status used: functional or not
938 * \param[in] d OSP device
939 * \param[in] rc new precreate status for device \a d
941 * \retval 0 on success
942 * \retval negative negated errno on error
944 void osp_pre_update_status(struct osp_device *d, int rc)
946 struct obd_statfs *msfs = &d->opd_statfs;
947 int old = d->opd_pre_status;
950 d->opd_pre_status = rc;
954 if (likely(msfs->os_type)) {
955 if (unlikely(d->opd_reserved_mb_high == 0 &&
956 d->opd_reserved_mb_low == 0)) {
957 /* Use ~0.1% by default to disable object allocation,
958 * and ~0.2% to enable, size in MB, set both watermark
960 spin_lock(&d->opd_pre_lock);
961 if (d->opd_reserved_mb_high == 0 &&
962 d->opd_reserved_mb_low == 0) {
963 d->opd_reserved_mb_low =
964 ((msfs->os_bsize >> 10) *
965 msfs->os_blocks) >> 20;
966 if (d->opd_reserved_mb_low == 0)
967 d->opd_reserved_mb_low = 1;
968 d->opd_reserved_mb_high =
969 (d->opd_reserved_mb_low << 1) + 1;
971 spin_unlock(&d->opd_pre_lock);
974 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
975 if (msfs->os_ffree < 32)
976 msfs->os_state |= OS_STATE_ENOINO;
977 else if (msfs->os_ffree > 64)
978 msfs->os_state &= ~OS_STATE_ENOINO;
980 if (available < d->opd_reserved_mb_low)
981 msfs->os_state |= OS_STATE_ENOSPC;
982 else if (available > d->opd_reserved_mb_high)
983 msfs->os_state &= ~OS_STATE_ENOSPC;
984 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
985 d->opd_pre_status = -ENOSPC;
987 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
988 "free, %llu avail, %llu MB avail, %u "
989 "hwm -> %d: rc = %d\n",
990 d->opd_obd->obd_name, msfs->os_blocks,
991 msfs->os_bfree, msfs->os_bavail,
992 available, d->opd_reserved_mb_high,
993 d->opd_pre_status, rc);
995 "non-committed changes: %u, in progress: %u\n",
996 atomic_read(&d->opd_sync_changes),
997 atomic_read(&d->opd_sync_rpcs_in_progress));
998 } else if (unlikely(old == -ENOSPC)) {
999 d->opd_pre_status = 0;
1000 spin_lock(&d->opd_pre_lock);
1001 d->opd_pre_create_slow = 0;
1002 d->opd_pre_create_count = OST_MIN_PRECREATE;
1003 spin_unlock(&d->opd_pre_lock);
1004 wake_up(&d->opd_pre_waitq);
1006 CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1007 " free, %llu avail, %lluMB avail, %u lwm"
1008 " -> %d: rc = %d\n", d->opd_obd->obd_name,
1009 msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1010 available, d->opd_reserved_mb_low,
1011 d->opd_pre_status, rc);
1015 wake_up(&d->opd_pre_user_waitq);
1019 * Initialize FID for precreation
1021 * For a just created new target, a new sequence should be taken.
1022 * The function checks there is no IDIF in use (if the target was
1023 * added with the older version of Lustre), then requests a new
1024 * sequence from FLDB using the regular protocol. Then this new
1025 * sequence is stored on a persisten storage synchronously to prevent
1026 * possible object leakage (for the detail see the description for
1027 * osp_precreate_rollover_new_seq()).
1029 * \param[in] osp OSP device
1031 * \retval 0 on success
1032 * \retval negative negated errno on error
1034 int osp_init_pre_fid(struct osp_device *osp)
1037 struct osp_thread_info *osi;
1038 struct lu_client_seq *cli_seq;
1039 struct lu_fid *last_fid;
1043 LASSERT(osp->opd_pre != NULL);
1045 /* Let's check if the current last_seq/fid is valid,
1046 * otherwise request new sequence from the controller */
1047 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1048 /* Non-MDT0 can only use normal sequence for
1050 if (fid_is_norm(&osp->opd_last_used_fid))
1053 /* Initially MDT0 will start with IDIF, after
1054 * that it will request new sequence from the
1056 if (fid_is_idif(&osp->opd_last_used_fid) ||
1057 fid_is_norm(&osp->opd_last_used_fid))
1061 if (!fid_is_zero(&osp->opd_last_used_fid))
1062 CWARN("%s: invalid last used fid "DFID
1063 ", try to get new sequence.\n",
1064 osp->opd_obd->obd_name,
1065 PFID(&osp->opd_last_used_fid));
1067 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1069 CERROR("%s: init env error: rc = %d\n",
1070 osp->opd_obd->obd_name, rc);
1074 osi = osp_env_info(&env);
1075 last_fid = &osi->osi_fid;
1077 /* For a freshed fs, it will allocate a new sequence first */
1078 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1079 cli_seq = osp->opd_obd->u.cli.cl_seq;
1080 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1082 CERROR("%s: alloc fid error: rc = %d\n",
1083 osp->opd_obd->obd_name, rc);
1087 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1089 last_fid->f_oid = 1;
1090 last_fid->f_ver = 0;
1092 spin_lock(&osp->opd_pre_lock);
1093 osp->opd_last_used_fid = *last_fid;
1094 osp->opd_pre_used_fid = *last_fid;
1095 osp->opd_pre_last_created_fid = *last_fid;
1096 spin_unlock(&osp->opd_pre_lock);
1097 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1099 CERROR("%s: write fid error: rc = %d\n",
1100 osp->opd_obd->obd_name, rc);
1109 * The core of precreate functionality
1111 * The function implements the main precreation loop. Basically it
1112 * involves connecting to the target, precerate FID initialization,
1113 * identifying and removing orphans, then serving precreation. As
1114 * part of the latter, the thread is responsible for statfs data
1115 * updates. The precreation is mostly driven by another threads
1116 * asking for new OST objects - those askers wake the thread when
1117 * the number of precreated objects reach low watermark.
1118 * After a disconnect, the sequence above repeats. This is keep going
1119 * until the thread is requested to stop.
1121 * \param[in] _arg private data the thread (OSP device to handle)
1123 * \retval 0 on success
1124 * \retval negative negated errno on error
1126 static int osp_precreate_thread(void *_arg)
1128 struct osp_device *d = _arg;
1129 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1130 struct l_wait_info lwi = { 0 };
1131 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1132 back_to_sleep, NULL);
1138 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1140 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1143 spin_lock(&d->opd_pre_lock);
1144 thread->t_flags = SVC_STOPPED;
1145 spin_unlock(&d->opd_pre_lock);
1146 wake_up(&thread->t_ctl_waitq);
1151 spin_lock(&d->opd_pre_lock);
1152 thread->t_flags = SVC_RUNNING;
1153 spin_unlock(&d->opd_pre_lock);
1154 wake_up(&thread->t_ctl_waitq);
1156 while (osp_precreate_running(d)) {
1158 * need to be connected to OST
1160 while (osp_precreate_running(d)) {
1161 if (d->opd_pre_recovering &&
1162 d->opd_imp_connected &&
1163 !d->opd_got_disconnected)
1165 l_wait_event(d->opd_pre_waitq,
1166 !osp_precreate_running(d) ||
1167 d->opd_new_connection,
1170 if (!d->opd_new_connection)
1173 d->opd_new_connection = 0;
1174 d->opd_got_disconnected = 0;
1178 if (!osp_precreate_running(d))
1181 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1182 /* Sigh, fid client is not ready yet */
1183 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1186 /* Init fid for osp_precreate if necessary */
1187 rc = osp_init_pre_fid(d);
1189 class_export_put(d->opd_exp);
1190 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1191 CERROR("%s: init pre fid error: rc = %d\n",
1192 d->opd_obd->obd_name, rc);
1196 if (osp_statfs_update(d)) {
1197 l_wait_event(d->opd_pre_waitq,
1198 !osp_precreate_running(d), &lwi2);
1203 * Clean up orphans or recreate missing objects.
1205 rc = osp_precreate_cleanup_orphans(&env, d);
1207 schedule_timeout_interruptible(cfs_time_seconds(1));
1211 * connected, can handle precreates now
1213 while (osp_precreate_running(d)) {
1214 l_wait_event(d->opd_pre_waitq,
1215 !osp_precreate_running(d) ||
1216 osp_precreate_near_empty(&env, d) ||
1217 osp_statfs_need_update(d) ||
1218 d->opd_got_disconnected, &lwi);
1220 if (!osp_precreate_running(d))
1223 /* something happened to the connection
1224 * have to start from the beginning */
1225 if (d->opd_got_disconnected)
1228 if (osp_statfs_need_update(d))
1229 if (osp_statfs_update(d))
1232 /* To avoid handling different seq in precreate/orphan
1233 * cleanup, it will hold precreate until current seq is
1235 if (unlikely(osp_precreate_end_seq(&env, d) &&
1236 !osp_create_end_seq(&env, d)))
1239 if (unlikely(osp_precreate_end_seq(&env, d) &&
1240 osp_create_end_seq(&env, d))) {
1241 LCONSOLE_INFO("%s:%#llx is used up."
1242 " Update to new seq\n",
1243 d->opd_obd->obd_name,
1244 fid_seq(&d->opd_pre_last_created_fid));
1245 rc = osp_precreate_rollover_new_seq(&env, d);
1250 if (osp_precreate_near_empty(&env, d)) {
1251 rc = osp_precreate_send(&env, d);
1252 /* osp_precreate_send() sets opd_pre_status
1253 * in case of error, that prevent the using of
1255 if (rc < 0 && rc != -ENOSPC &&
1256 rc != -ETIMEDOUT && rc != -ENOTCONN)
1257 CERROR("%s: cannot precreate objects:"
1259 d->opd_obd->obd_name, rc);
1264 thread->t_flags = SVC_STOPPED;
1266 wake_up(&thread->t_ctl_waitq);
1272 * Check when to stop to wait for precreate objects.
1274 * The caller wanting a new OST object can't wait undefinitely. The
1275 * function checks for few conditions including available new OST
1276 * objects, disconnected OST, lack of space with no pending destroys,
1277 * etc. IOW, it checks whether the current OSP state is good to keep
1278 * waiting or it's better to give up.
1280 * \param[in] env LU environment provided by the caller
1281 * \param[in] d OSP device
1283 * \retval 0 - keep waiting, 1 - no luck
1285 static int osp_precreate_ready_condition(const struct lu_env *env,
1286 struct osp_device *d)
1288 if (d->opd_pre_recovering)
1291 /* ready if got enough precreated objects */
1292 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1293 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1296 /* ready if OST reported no space and no destroys in progress */
1297 if (atomic_read(&d->opd_sync_changes) +
1298 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1299 d->opd_pre_status == -ENOSPC)
1302 /* Bail out I/O fails to OST */
1303 if (d->opd_pre_status != 0 &&
1304 d->opd_pre_status != -EAGAIN &&
1305 d->opd_pre_status != -ENODEV &&
1306 d->opd_pre_status != -ENOTCONN &&
1307 d->opd_pre_status != -ENOSPC) {
1309 if (d->opd_pre_status != -EIO)
1310 CERROR("%s: precreate failed opd_pre_status %d\n",
1311 d->opd_obd->obd_name, d->opd_pre_status);
1318 static int osp_precreate_timeout_condition(void *data)
1320 struct osp_device *d = data;
1322 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1323 "reserved=%llu, sync_changes=%u, "
1324 "sync_rpcs_in_progress=%d, status=%d\n",
1325 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1326 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1327 atomic_read(&d->opd_sync_changes),
1328 atomic_read(&d->opd_sync_rpcs_in_progress),
1335 * Reserve object in precreate pool
1337 * When the caller wants to create a new object on this target (target
1338 * represented by the given OSP), it should declare this intention using
1339 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1340 * to reserve an object in the existing precreated pool or wait up to
1341 * obd_timeout for the available object to appear in the pool (a dedicated
1342 * thread will be doing real precreation in background). The object can be
1343 * consumed later with osp_precreate_get_fid() or be released with call to
1344 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1345 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1346 * If the space on the target is short and there is a pending object destroy,
1347 * then the function forces local commit to speedup space release (see
1348 * osp_sync.c for the details).
1350 * \param[in] env LU environment provided by the caller
1351 * \param[in] d OSP device
1353 * \retval 0 on success
1354 * \retval -ENOSPC when no space on OST
1355 * \retval -EAGAIN try later, slow precreation in progress
1356 * \retval -EIO when no access to OST
1358 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1360 struct l_wait_info lwi;
1361 cfs_time_t expire = cfs_time_shift(obd_timeout);
1366 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1367 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1368 PFID(&d->opd_pre_used_fid));
1370 /* opd_pre_max_create_count 0 to not use specified OST. */
1371 if (d->opd_pre_max_create_count == 0)
1376 * - preallocation is done
1377 * - no free space expected soon
1378 * - can't connect to OST for too long (obd_timeout)
1379 * - OST can allocate fid sequence.
1381 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1382 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1385 * increase number of precreations
1387 precreated = osp_objs_precreated(env, d);
1388 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1389 d->opd_pre_create_slow == 0 &&
1390 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1391 spin_lock(&d->opd_pre_lock);
1392 d->opd_pre_create_slow = 1;
1393 d->opd_pre_create_count *= 2;
1394 spin_unlock(&d->opd_pre_lock);
1397 spin_lock(&d->opd_pre_lock);
1398 precreated = osp_objs_precreated(env, d);
1399 if (precreated > d->opd_pre_reserved &&
1400 !d->opd_pre_recovering) {
1401 d->opd_pre_reserved++;
1402 spin_unlock(&d->opd_pre_lock);
1405 /* XXX: don't wake up if precreation is in progress */
1406 if (osp_precreate_near_empty_nolock(env, d) &&
1407 !osp_precreate_end_seq_nolock(env, d))
1408 wake_up(&d->opd_pre_waitq);
1412 spin_unlock(&d->opd_pre_lock);
1415 * all precreated objects have been used and no-space
1416 * status leave us no chance to succeed very soon
1417 * but if there is destroy in progress, then we should
1418 * wait till that is done - some space might be released
1420 if (unlikely(rc == -ENOSPC)) {
1421 if (atomic_read(&d->opd_sync_changes)) {
1422 /* force local commit to release space */
1423 dt_commit_async(env, d->opd_storage);
1425 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1426 /* just wait till destroys are done */
1427 /* see l_wait_even() few lines below */
1429 if (atomic_read(&d->opd_sync_changes) +
1430 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1431 /* no hope for free space */
1436 /* XXX: don't wake up if precreation is in progress */
1437 wake_up(&d->opd_pre_waitq);
1439 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1440 osp_precreate_timeout_condition, d);
1441 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1446 l_wait_event(d->opd_pre_user_waitq,
1447 osp_precreate_ready_condition(env, d), &lwi);
1454 * Get a FID from precreation pool
1456 * The function is a companion for osp_precreate_reserve() - it assigns
1457 * a specific FID from the precreate. The function should be called only
1458 * if the call to osp_precreate_reserve() was successful. The function
1459 * updates a local storage to remember the highest object ID referenced
1460 * by the node in the given sequence.
1462 * A very importan details: this is supposed to be called once the
1463 * transaction is started, so on-disk update will be atomic with the
1464 * data (like LOVEA) refering this object. Then the object won't be leaked:
1465 * either it's referenced by the committed transaction or it's a subject
1466 * to the orphan cleanup procedure.
1468 * \param[in] env LU environment provided by the caller
1469 * \param[in] d OSP device
1470 * \param[out] fid generated FID
1472 * \retval 0 on success
1473 * \retval negative negated errno on error
1475 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1478 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1479 /* grab next id from the pool */
1480 spin_lock(&d->opd_pre_lock);
1482 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1483 &d->opd_pre_last_created_fid) < 0,
1484 "next fid "DFID" last created fid "DFID"\n",
1485 PFID(&d->opd_pre_used_fid),
1486 PFID(&d->opd_pre_last_created_fid));
1489 * When sequence is used up, new one should be allocated in
1490 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1493 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1494 "next fid "DFID" last created fid "DFID"\n",
1495 PFID(&d->opd_pre_used_fid),
1496 PFID(&d->opd_pre_last_created_fid));
1497 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1498 if (fid_is_idif(pre_used_fid) &&
1499 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1500 pre_used_fid->f_seq++;
1502 d->opd_pre_used_fid.f_oid++;
1503 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1504 d->opd_pre_reserved--;
1506 * last_used_id must be changed along with getting new id otherwise
1507 * we might miscalculate gap causing object loss or leak
1509 osp_update_last_fid(d, fid);
1510 spin_unlock(&d->opd_pre_lock);
1513 * probably main thread suspended orphan cleanup till
1514 * all reservations are released, see comment in
1515 * osp_precreate_thread() just before orphan cleanup
1517 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1518 wake_up(&d->opd_pre_waitq);
1524 * Set size regular attribute on an object
1526 * When a striping is created late, it's possible that size is already
1527 * initialized on the file. Then the new striping should inherit size
1528 * from the file. The function sets size on the object using the regular
1529 * protocol (OST_PUNCH).
1530 * XXX: should be re-implemented using OUT ?
1532 * \param[in] env LU environment provided by the caller
1533 * \param[in] dt object
1534 * \param[in] size size to set.
1536 * \retval 0 on success
1537 * \retval negative negated errno on error
1539 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1542 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1543 struct ptlrpc_request *req = NULL;
1544 struct obd_import *imp;
1545 struct ost_body *body;
1546 struct obdo *oa = NULL;
1551 imp = d->opd_obd->u.cli.cl_import;
1554 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1558 /* XXX: capa support? */
1559 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1560 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1562 ptlrpc_request_free(req);
1567 * XXX: decide how do we do here with resend
1568 * if we don't resend, then client may see wrong file size
1569 * if we do resend, then MDS thread can get stuck for quite long
1570 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1571 * (see LU-7975 and sanity/test_27F use cases)
1572 * but let's decide not to resend/delay this truncate request to OST
1573 * and allow Client to decide to resend, in a less agressive way from
1574 * after_reply(), by returning -EINPROGRESS instead of
1575 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1576 * end of this routine
1578 req->rq_no_resend = req->rq_no_delay = 1;
1580 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1581 ptlrpc_at_set_req_timeout(req);
1585 GOTO(out, rc = -ENOMEM);
1587 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1590 oa->o_blocks = OBD_OBJECT_EOF;
1591 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1592 OBD_MD_FLID | OBD_MD_FLGROUP;
1594 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1596 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1598 /* XXX: capa support? */
1599 /* osc_pack_capa(req, body, capa); */
1601 ptlrpc_request_set_replen(req);
1603 rc = ptlrpc_queue_wait(req);
1605 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1606 * since we have decided not to resend/delay, but this could
1607 * lead to wrong size to be seen at Client side and even process
1608 * trying to open to exit/fail if not itself handling -EAGAIN.
1609 * So it should be better to return -EINPROGRESS instead and
1610 * leave the decision to resend at Client side in after_reply()
1612 if (rc == -EWOULDBLOCK) {
1614 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1615 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1618 CERROR("can't punch object: %d\n", rc);
1622 ptlrpc_req_finished(req);
1629 * Initialize precreation functionality of OSP
1631 * Prepares all the internal structures and starts the precreate thread
1633 * \param[in] d OSP device
1635 * \retval 0 on success
1636 * \retval negative negated errno on error
1638 int osp_init_precreate(struct osp_device *d)
1640 struct l_wait_info lwi = { 0 };
1641 struct task_struct *task;
1645 OBD_ALLOC_PTR(d->opd_pre);
1646 if (d->opd_pre == NULL)
1649 /* initially precreation isn't ready */
1650 d->opd_pre_status = -EAGAIN;
1651 fid_zero(&d->opd_pre_used_fid);
1652 d->opd_pre_used_fid.f_oid = 1;
1653 fid_zero(&d->opd_pre_last_created_fid);
1654 d->opd_pre_last_created_fid.f_oid = 1;
1655 d->opd_pre_reserved = 0;
1656 d->opd_got_disconnected = 1;
1657 d->opd_pre_create_slow = 0;
1658 d->opd_pre_create_count = OST_MIN_PRECREATE;
1659 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1660 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1661 d->opd_reserved_mb_high = 0;
1662 d->opd_reserved_mb_low = 0;
1664 spin_lock_init(&d->opd_pre_lock);
1665 init_waitqueue_head(&d->opd_pre_waitq);
1666 init_waitqueue_head(&d->opd_pre_user_waitq);
1667 thread_set_flags(&d->opd_pre_thread, SVC_INIT);
1668 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1671 * Initialize statfs-related things
1673 d->opd_statfs_maxage = 5; /* default update interval */
1674 d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1675 CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1676 (unsigned long long)cfs_time_current(),
1677 (unsigned long long)d->opd_statfs_fresh_till);
1678 setup_timer(&d->opd_statfs_timer, osp_statfs_timer_cb,
1681 if (d->opd_storage->dd_rdonly)
1685 * start thread handling precreation and statfs updates
1687 task = kthread_run(osp_precreate_thread, d,
1688 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1690 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1691 RETURN(PTR_ERR(task));
1694 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1695 osp_precreate_running(d) || osp_precreate_stopped(d),
1702 * Finish precreate functionality of OSP
1705 * Asks all the activity (the thread, update timer) to stop, then
1706 * wait till that is done.
1708 * \param[in] d OSP device
1710 void osp_precreate_fini(struct osp_device *d)
1712 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1715 del_timer(&d->opd_statfs_timer);
1717 if (d->opd_pre == NULL)
1720 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1721 thread->t_flags = SVC_STOPPING;
1722 wake_up(&d->opd_pre_waitq);
1723 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1726 OBD_FREE_PTR(d->opd_pre);