4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osp/osp_precreate.c
34 * Lustre OST Proxy Device
36 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
37 * Author: Mikhail Pershin <mike.pershin@intel.com>
38 * Author: Di Wang <di.wang@intel.com>
41 #define DEBUG_SUBSYSTEM S_MDS
43 #include <linux/kthread.h>
45 #include <lustre_obdo.h>
47 #include "osp_internal.h"
50 * there are two specific states to take care about:
52 * = import is disconnected =
54 * = import is inactive =
55 * in this case osp_declare_create() returns an error
61 * Check whether statfs data is expired
63 * OSP device caches statfs data for the target, the function checks
64 * whether the data is expired or not.
66 * \param[in] d OSP device
68 * \retval 0 - not expired, 1 - expired
70 static inline int osp_statfs_need_update(struct osp_device *d)
72 return !ktime_before(ktime_get(), d->opd_statfs_fresh_till);
76 * OSP tries to maintain pool of available objects so that calls to create
77 * objects don't block most of time
79 * each time OSP gets connected to OST, we should start from precreation cleanup
81 static inline bool osp_precreate_running(struct osp_device *d)
83 return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
86 static inline bool osp_precreate_stopped(struct osp_device *d)
88 return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
91 static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data)
93 struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
96 if (osp_precreate_running(d))
97 wake_up(&d->opd_pre_waitq);
101 * RPC interpret callback for OST_STATFS RPC
103 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
104 * replied by the target. It's used to maintain statfs cache for the target.
105 * The function fills data from the reply if successful and schedules another
108 * \param[in] env LU environment provided by the caller
109 * \param[in] req RPC replied
110 * \param[in] aa callback data
111 * \param[in] rc RPC result
113 * \retval 0 on success
114 * \retval negative negated errno on error
116 static int osp_statfs_interpret(const struct lu_env *env,
117 struct ptlrpc_request *req, void *args, int rc)
119 union ptlrpc_async_args *aa = args;
120 struct obd_import *imp = req->rq_import;
121 struct obd_statfs *msfs;
122 struct osp_device *d;
127 aa = ptlrpc_req_async_args(req);
128 d = aa->pointer_arg[0];
134 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
136 GOTO(out, rc = -EPROTO);
138 d->opd_statfs = *msfs;
141 osp_pre_update_status(d, rc);
143 /* schedule next update */
144 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
145 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
146 mod_timer(&d->opd_statfs_timer,
147 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
148 d->opd_statfs_update_in_progress = 0;
150 CDEBUG(D_CACHE, "updated statfs %p\n", d);
154 /* couldn't update statfs, try again with a small delay */
155 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
156 d->opd_statfs_update_in_progress = 0;
157 if (d->opd_pre != NULL && osp_precreate_running(d))
158 wake_up(&d->opd_pre_waitq);
160 if (req->rq_import_generation == imp->imp_generation)
161 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
162 d->opd_obd->obd_name, rc);
167 * Send OST_STATFS RPC
169 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
170 * Also disables scheduled updates as times OSP may need to refresh
171 * statfs data before expiration. The function doesn't block, instead
172 * an interpretation callback osp_statfs_interpret() is used.
174 * \param[in] d OSP device
176 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
178 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
179 struct ptlrpc_request *req;
180 struct obd_import *imp;
181 union ptlrpc_async_args *aa;
186 CDEBUG(D_CACHE, "going to update statfs\n");
188 imp = d->opd_obd->u.cli.cl_import;
191 req = ptlrpc_request_alloc(imp,
192 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
196 rc = ptlrpc_request_pack(req,
197 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
198 d->opd_pre ? OST_STATFS : MDS_STATFS);
200 ptlrpc_request_free(req);
203 ptlrpc_request_set_replen(req);
205 req->rq_request_portal = OST_CREATE_PORTAL;
206 ptlrpc_at_set_req_timeout(req);
208 req->rq_interpret_reply = osp_statfs_interpret;
209 aa = ptlrpc_req_async_args(req);
210 aa->pointer_arg[0] = d;
213 * no updates till reply
215 del_timer(&d->opd_statfs_timer);
216 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
217 d->opd_statfs_update_in_progress = 1;
219 ptlrpcd_add_req(req);
221 /* we still want to sync changes if no new changes are coming */
222 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
225 if (atomic_read(&d->opd_sync_changes)) {
228 th = dt_trans_create(env, d->opd_storage);
230 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
233 rc = dt_trans_start_local(env, d->opd_storage, th);
235 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
236 d->opd_obd->obd_name,
237 atomic_read(&d->opd_sync_changes));
238 osp_sync_add_commit_cb_1s(env, d, th);
239 dt_trans_stop(env, d->opd_storage, th);
248 * Schedule an immediate update for statfs data
250 * If cached statfs data claim no free space, but OSP has got a request to
251 * destroy an object (so release some space probably), then we may need to
252 * refresh cached statfs data sooner than planned. The function checks there
253 * is no statfs update going and schedules immediate update if so.
254 * XXX: there might be a case where removed object(s) do not add free space (empty
255 * object). If the number of such deletions is high, then we can start to update
256 * statfs too often causing a RPC storm. some throttling is needed...
258 * \param[in] d OSP device where statfs data needs to be refreshed
260 void osp_statfs_need_now(struct osp_device *d)
262 if (!d->opd_statfs_update_in_progress) {
264 * if current status is -ENOSPC (lack of free space on OST)
265 * then we should poll OST immediately once object destroy
268 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
269 del_timer(&d->opd_statfs_timer);
270 wake_up(&d->opd_pre_waitq);
275 * Return number of precreated objects
277 * A simple helper to calculate the number of precreated objects on the device.
279 * \param[in] env LU environment provided by the caller
280 * \param[in] osp OSP device
282 * \retval the number of the precreated objects
284 static inline int osp_objs_precreated(const struct lu_env *env,
285 struct osp_device *osp)
287 return osp_fid_diff(&osp->opd_pre_last_created_fid,
288 &osp->opd_pre_used_fid);
292 * Check pool of precreated objects is nearly empty
294 * We should not wait till the pool of the precreated objects is exhausted,
295 * because then there will be a long period of OSP being unavailable for the
296 * new creations due to lenghty precreate RPC. Instead we ask for another
297 * precreation ahead and hopefully have it ready before the current pool is
298 * empty. Notice this function relies on an external locking.
300 * \param[in] env LU environment provided by the caller
301 * \param[in] d OSP device
303 * \retval 0 - current pool is good enough, 1 - time to precreate
305 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
306 struct osp_device *d)
308 int window = osp_objs_precreated(env, d);
310 /* don't consider new precreation till OST is healty and
312 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
313 (d->opd_pre_status == 0));
317 * Check pool of precreated objects
319 * This is protected version of osp_precreate_near_empty_nolock(), check that
322 * \param[in] env LU environment provided by the caller
323 * \param[in] d OSP device
325 * \retval 0 - current pool is good enough, 1 - time to precreate
327 static inline int osp_precreate_near_empty(const struct lu_env *env,
328 struct osp_device *d)
332 if (d->opd_pre == NULL)
335 /* XXX: do we really need locking here? */
336 spin_lock(&d->opd_pre_lock);
337 rc = osp_precreate_near_empty_nolock(env, d);
338 spin_unlock(&d->opd_pre_lock);
343 * Check given sequence is empty
345 * Returns a binary result whether the given sequence has some IDs left
346 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
347 * version of that function.
349 * \param[in] env LU environment provided by the caller
350 * \param[in] osp OSP device
352 * \retval 0 - current sequence has no IDs, 1 - otherwise
354 static inline int osp_create_end_seq(const struct lu_env *env,
355 struct osp_device *osp)
357 struct lu_fid *fid = &osp->opd_pre_used_fid;
360 spin_lock(&osp->opd_pre_lock);
361 rc = osp_fid_end_seq(env, fid);
362 spin_unlock(&osp->opd_pre_lock);
367 * Write FID into into last_oid/last_seq file
369 * The function stores the sequence and the in-sequence id into two dedicated
370 * files. The sync argument can be used to request synchronous commit, so the
371 * function won't return until the updates are committed.
373 * \param[in] env LU environment provided by the caller
374 * \param[in] osp OSP device
375 * \param[in] fid fid where sequence/id is taken
376 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
378 * \retval 0 on success
379 * \retval negative negated errno on error
381 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
382 struct lu_fid *fid, int sync)
384 struct osp_thread_info *oti = osp_env_info(env);
385 struct lu_buf *lb_oid = &oti->osi_lb;
386 struct lu_buf *lb_oseq = &oti->osi_lb2;
394 if (osp->opd_storage->dd_rdonly)
397 /* Note: through f_oid is only 32 bits, it will also write 64 bits
398 * for oid to keep compatibility with the previous version. */
400 osp_objid_buf_prep(lb_oid, &oid_off,
401 &oid, osp->opd_index);
403 osp_objseq_buf_prep(lb_oseq, &oseq_off,
404 &fid->f_seq, osp->opd_index);
406 th = dt_trans_create(env, osp->opd_storage);
411 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
412 lb_oid, oid_off, th);
416 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
417 lb_oseq, oseq_off, th);
421 rc = dt_trans_start_local(env, osp->opd_storage, th);
425 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
428 CERROR("%s: can not write to last seq file: rc = %d\n",
429 osp->opd_obd->obd_name, rc);
432 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
435 CERROR("%s: can not write to last seq file: rc = %d\n",
436 osp->opd_obd->obd_name, rc);
440 dt_trans_stop(env, osp->opd_storage, th);
445 * Switch to another sequence
447 * When a current sequence has no available IDs left, OSP has to switch to
448 * another new sequence. OSP requests it using the regular FLDB protocol
449 * and stores synchronously before that is used in precreated. This is needed
450 * to basically have the sequences referenced (not orphaned), otherwise it's
451 * possible that OST has some objects precreated and the clients have data
452 * written to it, but after MDT failover nobody refers those objects and OSP
453 * has no idea that the sequence need cleanup to be done.
454 * While this is very expensive operation, it's supposed to happen very very
455 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
457 * \param[in] env LU environment provided by the caller
458 * \param[in] osp OSP device
460 * \retval 0 on success
461 * \retval negative negated errno on error
463 static int osp_precreate_rollover_new_seq(struct lu_env *env,
464 struct osp_device *osp)
466 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
467 struct lu_fid *last_fid = &osp->opd_last_used_fid;
471 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
473 CERROR("%s: alloc fid error: rc = %d\n",
474 osp->opd_obd->obd_name, rc);
480 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
481 "fid "DFID", last_fid "DFID"\n", PFID(fid),
484 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
486 CERROR("%s: Can not update oid/seq file: rc = %d\n",
487 osp->opd_obd->obd_name, rc);
491 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
492 osp->opd_obd->obd_name, fid_seq(last_fid),
494 /* Update last_xxx to the new seq */
495 spin_lock(&osp->opd_pre_lock);
496 osp->opd_last_used_fid = *fid;
497 osp_fid_to_obdid(fid, &osp->opd_last_id);
498 osp->opd_gap_start_fid = *fid;
499 osp->opd_pre_used_fid = *fid;
500 osp->opd_pre_last_created_fid = *fid;
501 spin_unlock(&osp->opd_pre_lock);
507 * Find IDs available in current sequence
509 * The function calculates the highest possible ID and the number of IDs
510 * available in the current sequence OSP is using. The number is limited
511 * artifically by the caller (grow param) and the number of IDs available
512 * in the sequence by nature. The function doesn't require an external
515 * \param[in] env LU environment provided by the caller
516 * \param[in] osp OSP device
517 * \param[in] fid FID the caller wants to start with
518 * \param[in] grow how many the caller wants
519 * \param[out] fid the highest calculated FID
520 * \param[out] grow the number of available IDs calculated
522 * \retval 0 on success, 1 - the sequence is empty
524 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
525 struct lu_fid *fid, int *grow)
527 struct osp_thread_info *osi = osp_env_info(env);
531 if (fid_is_idif(fid)) {
532 struct lu_fid *last_fid;
533 struct ost_id *oi = &osi->osi_oi;
536 spin_lock(&osp->opd_pre_lock);
537 last_fid = &osp->opd_pre_last_created_fid;
538 fid_to_ostid(last_fid, oi);
539 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
540 *grow = end - ostid_id(oi);
541 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
542 spin_unlock(&osp->opd_pre_lock);
544 if (*grow == 0 || rc)
547 ostid_to_fid(fid, oi, osp->opd_index);
551 spin_lock(&osp->opd_pre_lock);
552 *fid = osp->opd_pre_last_created_fid;
554 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
555 *grow = end - fid->f_oid;
556 fid->f_oid += end - fid->f_oid;
557 spin_unlock(&osp->opd_pre_lock);
559 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
560 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
562 return *grow > 0 ? 0 : 1;
566 * Prepare and send precreate RPC
568 * The function finds how many objects should be precreated. Then allocates,
569 * prepares and schedules precreate RPC synchronously. Upon reply the function
570 * wake ups the threads waiting for the new objects on this target. If the
571 * target wasn't able to create all the objects requested, then the next
572 * precreate will be asking less objects (i.e. slow precreate down).
574 * \param[in] env LU environment provided by the caller
575 * \param[in] d OSP device
577 * \retval 0 on success
578 * \retval negative negated errno on error
580 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
582 struct osp_thread_info *oti = osp_env_info(env);
583 struct ptlrpc_request *req;
584 struct obd_import *imp;
585 struct ost_body *body;
587 struct lu_fid *fid = &oti->osi_fid;
590 /* don't precreate new objects till OST healthy and has free space */
591 if (unlikely(d->opd_pre_status)) {
592 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
593 d->opd_obd->obd_name, d->opd_pre_status);
598 * if not connection/initialization is compeleted, ignore
600 imp = d->opd_obd->u.cli.cl_import;
603 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
606 req->rq_request_portal = OST_CREATE_PORTAL;
607 /* we should not resend create request - anyway we will have delorphan
608 * and kill these objects */
609 req->rq_no_delay = req->rq_no_resend = 1;
611 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
613 ptlrpc_request_free(req);
617 spin_lock(&d->opd_pre_lock);
618 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
619 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
620 grow = d->opd_pre_create_count;
621 spin_unlock(&d->opd_pre_lock);
623 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
626 *fid = d->opd_pre_last_created_fid;
627 rc = osp_precreate_fids(env, d, fid, &grow);
629 /* Current seq has been used up*/
630 GOTO(out_req, rc = -ENOSPC);
632 if (!osp_is_fid_client(d)) {
633 /* Non-FID client will always send seq 0 because of
635 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
639 fid_to_ostid(fid, &body->oa.o_oi);
640 body->oa.o_valid = OBD_MD_FLGROUP;
642 ptlrpc_request_set_replen(req);
644 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
647 rc = ptlrpc_queue_wait(req);
649 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
653 LASSERT(req->rq_transno == 0);
655 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
657 GOTO(out_req, rc = -EPROTO);
659 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
662 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
663 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
664 ": rc = %d\n", d->opd_obd->obd_name,
665 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
666 GOTO(out_req, rc = -ESTALE);
669 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
671 spin_lock(&d->opd_pre_lock);
673 /* the OST has not managed to create all the
674 * objects we asked for */
675 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
676 d->opd_pre_create_slow = 1;
678 /* the OST is able to keep up with the work,
679 * we could consider increasing create_count
680 * next time if needed */
681 d->opd_pre_create_slow = 0;
684 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
685 fid_to_ostid(fid, &body->oa.o_oi);
687 d->opd_pre_last_created_fid = *fid;
688 spin_unlock(&d->opd_pre_lock);
690 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
691 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
692 PFID(&d->opd_pre_last_created_fid));
694 /* now we can wakeup all users awaiting for objects */
695 osp_pre_update_status(d, rc);
696 wake_up(&d->opd_pre_user_waitq);
698 ptlrpc_req_finished(req);
703 * Get last precreated object from target (OST)
705 * Sends synchronous RPC to the target (OST) to learn the last precreated
706 * object. This later is used to remove all unused objects (cleanup orphan
707 * procedure). Also, the next object after one we got will be used as a
708 * starting point for the new precreates.
710 * \param[in] env LU environment provided by the caller
711 * \param[in] d OSP device
713 * \retval 0 on success
714 * \retval negative negated errno on error
716 static int osp_get_lastfid_from_ost(const struct lu_env *env,
717 struct osp_device *d)
719 struct ptlrpc_request *req = NULL;
720 struct obd_import *imp;
721 struct lu_fid *last_fid;
726 imp = d->opd_obd->u.cli.cl_import;
729 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
733 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
734 sizeof(KEY_LAST_FID));
736 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
738 ptlrpc_request_free(req);
742 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
743 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
745 req->rq_no_delay = req->rq_no_resend = 1;
746 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
747 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
749 ptlrpc_request_set_replen(req);
751 rc = ptlrpc_queue_wait(req);
753 /* bad-bad OST.. let sysadm sort this out */
754 if (rc == -ENOTSUPP) {
755 CERROR("%s: server does not support FID: rc = %d\n",
756 d->opd_obd->obd_name, -ENOTSUPP);
758 ptlrpc_set_import_active(imp, 0);
762 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
763 if (last_fid == NULL) {
764 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
765 GOTO(out, rc = -EPROTO);
768 if (!fid_is_sane(last_fid)) {
769 CERROR("%s: Got insane last_fid "DFID"\n",
770 d->opd_obd->obd_name, PFID(last_fid));
771 GOTO(out, rc = -EPROTO);
774 /* Only update the last used fid, if the OST has objects for
775 * this sequence, i.e. fid_oid > 0 */
776 if (fid_oid(last_fid) > 0)
777 d->opd_last_used_fid = *last_fid;
779 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
783 ptlrpc_req_finished(req);
788 * Cleanup orphans on OST
790 * This function is called in a contex of a dedicated thread handling
791 * all the precreation suff. The function waits till local recovery
792 * is complete, then identify all the unreferenced objects (orphans)
793 * using the highest ID referenced by a local and the highest object
794 * precreated by the target. The found range is a subject to removal
795 * using specially flagged RPC. During this process OSP is marked
796 * unavailable for new objects.
798 * \param[in] env LU environment provided by the caller
799 * \param[in] d OSP device
801 * \retval 0 on success
802 * \retval negative negated errno on error
804 static int osp_precreate_cleanup_orphans(struct lu_env *env,
805 struct osp_device *d)
807 struct osp_thread_info *osi = osp_env_info(env);
808 struct lu_fid *last_fid = &osi->osi_fid;
809 struct ptlrpc_request *req = NULL;
810 struct obd_import *imp;
811 struct ost_body *body;
812 struct l_wait_info lwi = { 0 };
813 int update_status = 0;
820 * wait for local recovery to finish, so we can cleanup orphans
821 * orphans are all objects since "last used" (assigned), but
822 * there might be objects reserved and in some cases they won't
823 * be used. we can't cleanup them till we're sure they won't be
824 * used. also can't we allow new reservations because they may
825 * end up getting orphans being cleaned up below. so we block
826 * new reservations and wait till all reserved objects either
829 spin_lock(&d->opd_pre_lock);
830 d->opd_pre_recovering = 1;
831 spin_unlock(&d->opd_pre_lock);
833 * The locking above makes sure the opd_pre_reserved check below will
834 * catch all osp_precreate_reserve() calls who find
835 * "!opd_pre_recovering".
837 l_wait_event(d->opd_pre_waitq,
838 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
839 !osp_precreate_running(d) || d->opd_got_disconnected,
841 if (!osp_precreate_running(d) || d->opd_got_disconnected)
842 GOTO(out, rc = -EAGAIN);
844 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
845 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
847 *last_fid = d->opd_last_used_fid;
848 /* The OSP should already get the valid seq now */
849 LASSERT(!fid_is_zero(last_fid));
850 if (fid_oid(&d->opd_last_used_fid) < 2) {
851 /* lastfid looks strange... ask OST */
852 rc = osp_get_lastfid_from_ost(env, d);
857 imp = d->opd_obd->u.cli.cl_import;
860 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
862 GOTO(out, rc = -ENOMEM);
864 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
866 ptlrpc_request_free(req);
871 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
873 GOTO(out, rc = -EPROTO);
875 body->oa.o_flags = OBD_FL_DELORPHAN;
876 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
878 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
880 ptlrpc_request_set_replen(req);
882 /* Don't resend the delorphan req */
883 req->rq_no_resend = req->rq_no_delay = 1;
885 rc = ptlrpc_queue_wait(req);
891 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
893 GOTO(out, rc = -EPROTO);
896 * OST provides us with id new pool starts from in body->oa.o_id
898 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
900 spin_lock(&d->opd_pre_lock);
901 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
903 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
904 d->opd_pre_last_created_fid = d->opd_last_used_fid;
906 d->opd_pre_create_count = OST_MIN_PRECREATE;
907 d->opd_pre_last_created_fid = *last_fid;
910 * This empties the pre-creation pool and effectively blocks any new
913 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
914 LUSTRE_DATA_SEQ_MAX_WIDTH);
915 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
916 d->opd_pre_create_slow = 0;
917 spin_unlock(&d->opd_pre_lock);
919 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
920 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
921 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
924 ptlrpc_req_finished(req);
927 * If rc is zero, the pre-creation window should have been emptied.
928 * Since waking up the herd would be useless without pre-created
929 * objects, we defer the signal to osp_precreate_send() in that case.
933 CERROR("%s: cannot cleanup orphans: rc = %d\n",
934 d->opd_obd->obd_name, rc);
935 /* we can't proceed from here, OST seem to
936 * be in a bad shape, better to wait for
937 * a new instance of the server and repeat
938 * from the beginning. notify possible waiters
939 * this OSP isn't quite functional yet */
940 osp_pre_update_status(d, rc);
942 wake_up(&d->opd_pre_user_waitq);
945 spin_lock(&d->opd_pre_lock);
946 d->opd_pre_recovering = 0;
947 spin_unlock(&d->opd_pre_lock);
954 * Update precreate status using statfs data
956 * The function decides whether this OSP should be used for new objects.
957 * IOW, whether this OST is used up or has some free space. Cached statfs
958 * data is used to make this decision. If the latest result of statfs
959 * request (rc argument) is not success, then just mark OSP unavailable
962 * Add a bit of hysteresis so this flag isn't continually flapping,
963 * and ensure that new files don't get extremely fragmented due to
964 * only a small amount of available space in the filesystem.
965 * We want to set the ENOSPC when there is less than reserved size
966 * free and clear it when there is at least 2*reserved size free space.
967 * the function updates current precreation status used: functional or not
969 * \param[in] d OSP device
970 * \param[in] rc new precreate status for device \a d
972 * \retval 0 on success
973 * \retval negative negated errno on error
975 void osp_pre_update_status(struct osp_device *d, int rc)
977 struct obd_statfs *msfs = &d->opd_statfs;
978 int old = d->opd_pre_status;
981 d->opd_pre_status = rc;
985 if (likely(msfs->os_type)) {
986 if (unlikely(d->opd_reserved_mb_high == 0 &&
987 d->opd_reserved_mb_low == 0)) {
988 /* Use ~0.1% by default to disable object allocation,
989 * and ~0.2% to enable, size in MB, set both watermark
991 spin_lock(&d->opd_pre_lock);
992 if (d->opd_reserved_mb_high == 0 &&
993 d->opd_reserved_mb_low == 0) {
994 d->opd_reserved_mb_low =
995 ((msfs->os_bsize >> 10) *
996 msfs->os_blocks) >> 20;
997 if (d->opd_reserved_mb_low == 0)
998 d->opd_reserved_mb_low = 1;
999 d->opd_reserved_mb_high =
1000 (d->opd_reserved_mb_low << 1) + 1;
1002 spin_unlock(&d->opd_pre_lock);
1005 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1006 if (msfs->os_ffree < 32)
1007 msfs->os_state |= OS_STATE_ENOINO;
1008 else if (msfs->os_ffree > 64)
1009 msfs->os_state &= ~OS_STATE_ENOINO;
1011 if (available < d->opd_reserved_mb_low)
1012 msfs->os_state |= OS_STATE_ENOSPC;
1013 else if (available > d->opd_reserved_mb_high)
1014 msfs->os_state &= ~OS_STATE_ENOSPC;
1015 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1016 d->opd_pre_status = -ENOSPC;
1018 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
1019 "free, %llu avail, %llu MB avail, %u "
1020 "hwm -> %d: rc = %d\n",
1021 d->opd_obd->obd_name, msfs->os_blocks,
1022 msfs->os_bfree, msfs->os_bavail,
1023 available, d->opd_reserved_mb_high,
1024 d->opd_pre_status, rc);
1026 "non-committed changes: %u, in progress: %u\n",
1027 atomic_read(&d->opd_sync_changes),
1028 atomic_read(&d->opd_sync_rpcs_in_progress));
1029 } else if (unlikely(old == -ENOSPC)) {
1030 d->opd_pre_status = 0;
1031 spin_lock(&d->opd_pre_lock);
1032 d->opd_pre_create_slow = 0;
1033 d->opd_pre_create_count = OST_MIN_PRECREATE;
1034 spin_unlock(&d->opd_pre_lock);
1035 wake_up(&d->opd_pre_waitq);
1037 CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1038 " free, %llu avail, %lluMB avail, %u lwm"
1039 " -> %d: rc = %d\n", d->opd_obd->obd_name,
1040 msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1041 available, d->opd_reserved_mb_low,
1042 d->opd_pre_status, rc);
1045 /* Object precreation is skipped on the OST with
1046 * max_create_count=0. */
1047 if (d->opd_pre_max_create_count == 0)
1048 msfs->os_state |= OS_STATE_NOPRECREATE;
1050 msfs->os_state &= ~OS_STATE_NOPRECREATE;
1053 wake_up(&d->opd_pre_user_waitq);
1057 * Initialize FID for precreation
1059 * For a just created new target, a new sequence should be taken.
1060 * The function checks there is no IDIF in use (if the target was
1061 * added with the older version of Lustre), then requests a new
1062 * sequence from FLDB using the regular protocol. Then this new
1063 * sequence is stored on a persisten storage synchronously to prevent
1064 * possible object leakage (for the detail see the description for
1065 * osp_precreate_rollover_new_seq()).
1067 * \param[in] osp OSP device
1069 * \retval 0 on success
1070 * \retval negative negated errno on error
1072 int osp_init_pre_fid(struct osp_device *osp)
1075 struct osp_thread_info *osi;
1076 struct lu_client_seq *cli_seq;
1077 struct lu_fid *last_fid;
1081 LASSERT(osp->opd_pre != NULL);
1083 /* Let's check if the current last_seq/fid is valid,
1084 * otherwise request new sequence from the controller */
1085 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1086 /* Non-MDT0 can only use normal sequence for
1088 if (fid_is_norm(&osp->opd_last_used_fid))
1091 /* Initially MDT0 will start with IDIF, after
1092 * that it will request new sequence from the
1094 if (fid_is_idif(&osp->opd_last_used_fid) ||
1095 fid_is_norm(&osp->opd_last_used_fid))
1099 if (!fid_is_zero(&osp->opd_last_used_fid))
1100 CWARN("%s: invalid last used fid "DFID
1101 ", try to get new sequence.\n",
1102 osp->opd_obd->obd_name,
1103 PFID(&osp->opd_last_used_fid));
1105 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1107 CERROR("%s: init env error: rc = %d\n",
1108 osp->opd_obd->obd_name, rc);
1112 osi = osp_env_info(&env);
1113 last_fid = &osi->osi_fid;
1115 /* For a freshed fs, it will allocate a new sequence first */
1116 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1117 cli_seq = osp->opd_obd->u.cli.cl_seq;
1118 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1120 CERROR("%s: alloc fid error: rc = %d\n",
1121 osp->opd_obd->obd_name, rc);
1125 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1127 last_fid->f_oid = 1;
1128 last_fid->f_ver = 0;
1130 spin_lock(&osp->opd_pre_lock);
1131 osp->opd_last_used_fid = *last_fid;
1132 osp->opd_pre_used_fid = *last_fid;
1133 osp->opd_pre_last_created_fid = *last_fid;
1134 spin_unlock(&osp->opd_pre_lock);
1135 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1137 CERROR("%s: write fid error: rc = %d\n",
1138 osp->opd_obd->obd_name, rc);
1147 * The core of precreate functionality
1149 * The function implements the main precreation loop. Basically it
1150 * involves connecting to the target, precerate FID initialization,
1151 * identifying and removing orphans, then serving precreation. As
1152 * part of the latter, the thread is responsible for statfs data
1153 * updates. The precreation is mostly driven by another threads
1154 * asking for new OST objects - those askers wake the thread when
1155 * the number of precreated objects reach low watermark.
1156 * After a disconnect, the sequence above repeats. This is keep going
1157 * until the thread is requested to stop.
1159 * \param[in] _arg private data the thread (OSP device to handle)
1161 * \retval 0 on success
1162 * \retval negative negated errno on error
1164 static int osp_precreate_thread(void *_arg)
1166 struct osp_device *d = _arg;
1167 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1168 struct l_wait_info lwi = { 0 };
1169 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1170 back_to_sleep, NULL);
1176 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1178 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1181 spin_lock(&d->opd_pre_lock);
1182 thread->t_flags = SVC_STOPPED;
1183 spin_unlock(&d->opd_pre_lock);
1184 wake_up(&thread->t_ctl_waitq);
1189 spin_lock(&d->opd_pre_lock);
1190 thread->t_flags = SVC_RUNNING;
1191 spin_unlock(&d->opd_pre_lock);
1192 wake_up(&thread->t_ctl_waitq);
1194 while (osp_precreate_running(d)) {
1196 * need to be connected to OST
1198 while (osp_precreate_running(d)) {
1199 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1200 d->opd_imp_connected &&
1201 !d->opd_got_disconnected)
1203 l_wait_event(d->opd_pre_waitq,
1204 !osp_precreate_running(d) ||
1205 d->opd_new_connection,
1208 if (!d->opd_new_connection)
1211 d->opd_new_connection = 0;
1212 d->opd_got_disconnected = 0;
1216 if (!osp_precreate_running(d))
1220 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1221 /* Sigh, fid client is not ready yet */
1222 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1225 /* Init fid for osp_precreate if necessary */
1226 rc = osp_init_pre_fid(d);
1228 class_export_put(d->opd_exp);
1229 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1230 CERROR("%s: init pre fid error: rc = %d\n",
1231 d->opd_obd->obd_name, rc);
1236 if (osp_statfs_update(&env, d)) {
1237 l_wait_event(d->opd_pre_waitq,
1238 !osp_precreate_running(d), &lwi2);
1244 * Clean up orphans or recreate missing objects.
1246 rc = osp_precreate_cleanup_orphans(&env, d);
1248 schedule_timeout_interruptible(
1249 msecs_to_jiffies(MSEC_PER_SEC));
1255 * connected, can handle precreates now
1257 while (osp_precreate_running(d)) {
1258 l_wait_event(d->opd_pre_waitq,
1259 !osp_precreate_running(d) ||
1260 osp_precreate_near_empty(&env, d) ||
1261 osp_statfs_need_update(d) ||
1262 d->opd_got_disconnected, &lwi);
1264 if (!osp_precreate_running(d))
1267 /* something happened to the connection
1268 * have to start from the beginning */
1269 if (d->opd_got_disconnected)
1272 if (osp_statfs_need_update(d))
1273 if (osp_statfs_update(&env, d))
1276 if (d->opd_pre == NULL)
1279 /* To avoid handling different seq in precreate/orphan
1280 * cleanup, it will hold precreate until current seq is
1282 if (unlikely(osp_precreate_end_seq(&env, d) &&
1283 !osp_create_end_seq(&env, d)))
1286 if (unlikely(osp_precreate_end_seq(&env, d) &&
1287 osp_create_end_seq(&env, d))) {
1288 LCONSOLE_INFO("%s:%#llx is used up."
1289 " Update to new seq\n",
1290 d->opd_obd->obd_name,
1291 fid_seq(&d->opd_pre_last_created_fid));
1292 rc = osp_precreate_rollover_new_seq(&env, d);
1297 if (osp_precreate_near_empty(&env, d)) {
1298 rc = osp_precreate_send(&env, d);
1299 /* osp_precreate_send() sets opd_pre_status
1300 * in case of error, that prevent the using of
1302 if (rc < 0 && rc != -ENOSPC &&
1303 rc != -ETIMEDOUT && rc != -ENOTCONN)
1304 CERROR("%s: cannot precreate objects:"
1306 d->opd_obd->obd_name, rc);
1311 thread->t_flags = SVC_STOPPED;
1313 wake_up(&thread->t_ctl_waitq);
1319 * Check when to stop to wait for precreate objects.
1321 * The caller wanting a new OST object can't wait undefinitely. The
1322 * function checks for few conditions including available new OST
1323 * objects, disconnected OST, lack of space with no pending destroys,
1324 * etc. IOW, it checks whether the current OSP state is good to keep
1325 * waiting or it's better to give up.
1327 * \param[in] env LU environment provided by the caller
1328 * \param[in] d OSP device
1330 * \retval 0 - keep waiting, 1 - no luck
1332 static int osp_precreate_ready_condition(const struct lu_env *env,
1333 struct osp_device *d)
1335 if (d->opd_pre_recovering)
1338 /* ready if got enough precreated objects */
1339 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1340 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1343 /* ready if OST reported no space and no destroys in progress */
1344 if (atomic_read(&d->opd_sync_changes) +
1345 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1346 d->opd_pre_status == -ENOSPC)
1349 /* Bail out I/O fails to OST */
1350 if (d->opd_pre_status != 0 &&
1351 d->opd_pre_status != -EAGAIN &&
1352 d->opd_pre_status != -ENODEV &&
1353 d->opd_pre_status != -ENOTCONN &&
1354 d->opd_pre_status != -ENOSPC) {
1356 if (d->opd_pre_status != -EIO)
1357 CERROR("%s: precreate failed opd_pre_status %d\n",
1358 d->opd_obd->obd_name, d->opd_pre_status);
1365 static int osp_precreate_timeout_condition(void *data)
1367 struct osp_device *d = data;
1369 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1370 "reserved=%llu, sync_changes=%u, "
1371 "sync_rpcs_in_progress=%d, status=%d\n",
1372 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1373 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1374 atomic_read(&d->opd_sync_changes),
1375 atomic_read(&d->opd_sync_rpcs_in_progress),
1382 * Reserve object in precreate pool
1384 * When the caller wants to create a new object on this target (target
1385 * represented by the given OSP), it should declare this intention using
1386 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1387 * to reserve an object in the existing precreated pool or wait up to
1388 * obd_timeout for the available object to appear in the pool (a dedicated
1389 * thread will be doing real precreation in background). The object can be
1390 * consumed later with osp_precreate_get_fid() or be released with call to
1391 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1392 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1393 * If the space on the target is short and there is a pending object destroy,
1394 * then the function forces local commit to speedup space release (see
1395 * osp_sync.c for the details).
1397 * \param[in] env LU environment provided by the caller
1398 * \param[in] d OSP device
1400 * \retval 0 on success
1401 * \retval -ENOSPC when no space on OST
1402 * \retval -EAGAIN try later, slow precreation in progress
1403 * \retval -EIO when no access to OST
1405 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1407 time64_t expire = ktime_get_seconds() + obd_timeout;
1408 struct l_wait_info lwi;
1409 int precreated, rc, synced = 0;
1413 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1414 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1415 PFID(&d->opd_pre_used_fid));
1417 /* opd_pre_max_create_count 0 to not use specified OST. */
1418 if (d->opd_pre_max_create_count == 0)
1423 * - preallocation is done
1424 * - no free space expected soon
1425 * - can't connect to OST for too long (obd_timeout)
1426 * - OST can allocate fid sequence.
1428 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1429 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1432 * increase number of precreations
1434 precreated = osp_objs_precreated(env, d);
1435 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1436 d->opd_pre_create_slow == 0 &&
1437 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1438 spin_lock(&d->opd_pre_lock);
1439 d->opd_pre_create_slow = 1;
1440 d->opd_pre_create_count *= 2;
1441 spin_unlock(&d->opd_pre_lock);
1444 spin_lock(&d->opd_pre_lock);
1445 precreated = osp_objs_precreated(env, d);
1446 if (precreated > d->opd_pre_reserved &&
1447 !d->opd_pre_recovering) {
1448 d->opd_pre_reserved++;
1449 spin_unlock(&d->opd_pre_lock);
1452 /* XXX: don't wake up if precreation is in progress */
1453 if (osp_precreate_near_empty_nolock(env, d) &&
1454 !osp_precreate_end_seq_nolock(env, d))
1455 wake_up(&d->opd_pre_waitq);
1459 spin_unlock(&d->opd_pre_lock);
1462 * all precreated objects have been used and no-space
1463 * status leave us no chance to succeed very soon
1464 * but if there is destroy in progress, then we should
1465 * wait till that is done - some space might be released
1467 if (unlikely(rc == -ENOSPC)) {
1468 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1469 /* force local commit to release space */
1470 dt_commit_async(env, d->opd_storage);
1471 osp_sync_check_for_work(d);
1474 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1475 /* just wait till destroys are done */
1476 /* see l_wait_even() few lines below */
1478 if (atomic_read(&d->opd_sync_changes) +
1479 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1480 /* no hope for free space */
1485 /* XXX: don't wake up if precreation is in progress */
1486 wake_up(&d->opd_pre_waitq);
1488 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
1489 osp_precreate_timeout_condition, d);
1490 if (ktime_get_seconds() >= expire) {
1495 l_wait_event(d->opd_pre_user_waitq,
1496 osp_precreate_ready_condition(env, d), &lwi);
1503 * Get a FID from precreation pool
1505 * The function is a companion for osp_precreate_reserve() - it assigns
1506 * a specific FID from the precreate. The function should be called only
1507 * if the call to osp_precreate_reserve() was successful. The function
1508 * updates a local storage to remember the highest object ID referenced
1509 * by the node in the given sequence.
1511 * A very importan details: this is supposed to be called once the
1512 * transaction is started, so on-disk update will be atomic with the
1513 * data (like LOVEA) refering this object. Then the object won't be leaked:
1514 * either it's referenced by the committed transaction or it's a subject
1515 * to the orphan cleanup procedure.
1517 * \param[in] env LU environment provided by the caller
1518 * \param[in] d OSP device
1519 * \param[out] fid generated FID
1521 * \retval 0 on success
1522 * \retval negative negated errno on error
1524 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1527 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1528 /* grab next id from the pool */
1529 spin_lock(&d->opd_pre_lock);
1531 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1532 &d->opd_pre_last_created_fid) < 0,
1533 "next fid "DFID" last created fid "DFID"\n",
1534 PFID(&d->opd_pre_used_fid),
1535 PFID(&d->opd_pre_last_created_fid));
1538 * When sequence is used up, new one should be allocated in
1539 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1542 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1543 "next fid "DFID" last created fid "DFID"\n",
1544 PFID(&d->opd_pre_used_fid),
1545 PFID(&d->opd_pre_last_created_fid));
1546 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1547 if (fid_is_idif(pre_used_fid) &&
1548 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1549 pre_used_fid->f_seq++;
1551 d->opd_pre_used_fid.f_oid++;
1552 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1553 d->opd_pre_reserved--;
1555 * last_used_id must be changed along with getting new id otherwise
1556 * we might miscalculate gap causing object loss or leak
1558 osp_update_last_fid(d, fid);
1559 spin_unlock(&d->opd_pre_lock);
1562 * probably main thread suspended orphan cleanup till
1563 * all reservations are released, see comment in
1564 * osp_precreate_thread() just before orphan cleanup
1566 if (unlikely(d->opd_pre_reserved == 0 &&
1567 (d->opd_pre_recovering || d->opd_pre_status)))
1568 wake_up(&d->opd_pre_waitq);
1574 * Set size regular attribute on an object
1576 * When a striping is created late, it's possible that size is already
1577 * initialized on the file. Then the new striping should inherit size
1578 * from the file. The function sets size on the object using the regular
1579 * protocol (OST_PUNCH).
1580 * XXX: should be re-implemented using OUT ?
1582 * \param[in] env LU environment provided by the caller
1583 * \param[in] dt object
1584 * \param[in] size size to set.
1586 * \retval 0 on success
1587 * \retval negative negated errno on error
1589 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1592 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1593 struct ptlrpc_request *req = NULL;
1594 struct obd_import *imp;
1595 struct ost_body *body;
1596 struct obdo *oa = NULL;
1601 imp = d->opd_obd->u.cli.cl_import;
1604 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1608 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1610 ptlrpc_request_free(req);
1615 * XXX: decide how do we do here with resend
1616 * if we don't resend, then client may see wrong file size
1617 * if we do resend, then MDS thread can get stuck for quite long
1618 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1619 * (see LU-7975 and sanity/test_27F use cases)
1620 * but let's decide not to resend/delay this truncate request to OST
1621 * and allow Client to decide to resend, in a less agressive way from
1622 * after_reply(), by returning -EINPROGRESS instead of
1623 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1624 * end of this routine
1626 req->rq_no_resend = req->rq_no_delay = 1;
1628 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1629 ptlrpc_at_set_req_timeout(req);
1633 GOTO(out, rc = -ENOMEM);
1635 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1638 oa->o_blocks = OBD_OBJECT_EOF;
1639 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1640 OBD_MD_FLID | OBD_MD_FLGROUP;
1642 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1644 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1646 /* XXX: capa support? */
1647 /* osc_pack_capa(req, body, capa); */
1649 ptlrpc_request_set_replen(req);
1651 rc = ptlrpc_queue_wait(req);
1653 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1654 * since we have decided not to resend/delay, but this could
1655 * lead to wrong size to be seen at Client side and even process
1656 * trying to open to exit/fail if not itself handling -EAGAIN.
1657 * So it should be better to return -EINPROGRESS instead and
1658 * leave the decision to resend at Client side in after_reply()
1660 if (rc == -EWOULDBLOCK) {
1662 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1663 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1666 CERROR("can't punch object: %d\n", rc);
1670 ptlrpc_req_finished(req);
1677 * Initialize precreation functionality of OSP
1679 * Prepares all the internal structures and starts the precreate thread
1681 * \param[in] d OSP device
1683 * \retval 0 on success
1684 * \retval negative negated errno on error
1686 int osp_init_precreate(struct osp_device *d)
1690 OBD_ALLOC_PTR(d->opd_pre);
1691 if (d->opd_pre == NULL)
1694 /* initially precreation isn't ready */
1695 init_waitqueue_head(&d->opd_pre_user_waitq);
1696 d->opd_pre_status = -EAGAIN;
1697 fid_zero(&d->opd_pre_used_fid);
1698 d->opd_pre_used_fid.f_oid = 1;
1699 fid_zero(&d->opd_pre_last_created_fid);
1700 d->opd_pre_last_created_fid.f_oid = 1;
1702 d->opd_pre_reserved = 0;
1703 d->opd_got_disconnected = 1;
1704 d->opd_pre_create_slow = 0;
1705 d->opd_pre_create_count = OST_MIN_PRECREATE;
1706 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1707 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1708 d->opd_reserved_mb_high = 0;
1709 d->opd_reserved_mb_low = 0;
1715 * Finish precreate functionality of OSP
1718 * Asks all the activity (the thread, update timer) to stop, then
1719 * wait till that is done.
1721 * \param[in] d OSP device
1723 void osp_precreate_fini(struct osp_device *d)
1727 if (d->opd_pre == NULL)
1730 OBD_FREE_PTR(d->opd_pre);
1736 int osp_init_statfs(struct osp_device *d)
1738 struct l_wait_info lwi = { 0 };
1739 struct task_struct *task;
1743 spin_lock_init(&d->opd_pre_lock);
1744 init_waitqueue_head(&d->opd_pre_waitq);
1745 thread_set_flags(&d->opd_pre_thread, SVC_INIT);
1746 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1749 * Initialize statfs-related things
1751 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1752 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1753 1000 * NSEC_PER_SEC);
1754 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1756 ktime_to_ns(d->opd_statfs_fresh_till));
1757 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1758 (unsigned long)d, 0);
1760 if (d->opd_storage->dd_rdonly)
1764 * start thread handling precreation and statfs updates
1766 task = kthread_run(osp_precreate_thread, d,
1767 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1769 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1770 RETURN(PTR_ERR(task));
1773 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1774 osp_precreate_running(d) || osp_precreate_stopped(d),
1780 void osp_statfs_fini(struct osp_device *d)
1782 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1785 del_timer(&d->opd_statfs_timer);
1787 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1788 thread->t_flags = SVC_STOPPING;
1789 wake_up(&d->opd_pre_waitq);
1790 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));