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
60 * Check whether statfs data is expired
62 * OSP device caches statfs data for the target, the function checks
63 * whether the data is expired or not.
65 * \param[in] d OSP device
67 * \retval 0 - not expired, 1 - expired
69 static inline int osp_statfs_need_update(struct osp_device *d)
71 return !ktime_before(ktime_get(), d->opd_statfs_fresh_till);
75 * OSP tries to maintain pool of available objects so that calls to create
76 * objects don't block most of time
78 * each time OSP gets connected to OST, we should start from precreation cleanup
80 static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data)
82 struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
86 wake_up(&d->opd_pre_waitq);
89 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs);
92 * The function updates current precreation status if broken, and
93 * updates that cached statfs state if functional, then wakes up waiters.
94 * We don't clear opd_pre_status directly here, but rather leave this
95 * to osp_pre_update_msfs() to do if everything is OK so that we don't
96 * have a race to clear opd_pre_status and then set it to -ENOSPC again.
98 * \param[in] d OSP device
99 * \param[in] msfs statfs data
100 * \param[in] rc new precreate status for device \a d
102 static void osp_pre_update_status_msfs(struct osp_device *d,
103 struct obd_statfs *msfs, int rc)
106 d->opd_pre_status = rc;
108 osp_pre_update_msfs(d, msfs);
110 wake_up(&d->opd_pre_user_waitq);
113 /* Pass in the old statfs data in case the limits have changed */
114 void osp_pre_update_status(struct osp_device *d, int rc)
116 osp_pre_update_status_msfs(d, &d->opd_statfs, rc);
121 * RPC interpret callback for OST_STATFS RPC
123 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
124 * replied by the target. It's used to maintain statfs cache for the target.
125 * The function fills data from the reply if successful and schedules another
128 * \param[in] env LU environment provided by the caller
129 * \param[in] req RPC replied
130 * \param[in] aa callback data
131 * \param[in] rc RPC result
133 * \retval 0 on success
134 * \retval negative negated errno on error
136 static int osp_statfs_interpret(const struct lu_env *env,
137 struct ptlrpc_request *req, void *args, int rc)
139 union ptlrpc_async_args *aa = args;
140 struct obd_import *imp = req->rq_import;
141 struct obd_statfs *msfs;
142 struct osp_device *d;
147 aa = ptlrpc_req_async_args(aa, req);
148 d = aa->pointer_arg[0];
154 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
156 GOTO(out, rc = -EPROTO);
159 osp_pre_update_status_msfs(d, msfs, 0);
161 d->opd_statfs = *msfs;
163 /* schedule next update */
164 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
165 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
166 mod_timer(&d->opd_statfs_timer,
167 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
168 d->opd_statfs_update_in_progress = 0;
170 CDEBUG(D_CACHE, "updated statfs %p\n", d);
174 /* couldn't update statfs, try again with a small delay */
175 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
176 d->opd_statfs_update_in_progress = 0;
177 if (d->opd_pre && d->opd_pre_task)
178 wake_up(&d->opd_pre_waitq);
180 if (req->rq_import_generation == imp->imp_generation)
181 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
182 d->opd_obd->obd_name, rc);
187 * Send OST_STATFS RPC
189 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
190 * Also disables scheduled updates as times OSP may need to refresh
191 * statfs data before expiration. The function doesn't block, instead
192 * an interpretation callback osp_statfs_interpret() is used.
194 * \param[in] d OSP device
196 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
198 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
199 struct ptlrpc_request *req;
200 struct obd_import *imp;
201 union ptlrpc_async_args *aa;
206 CDEBUG(D_CACHE, "going to update statfs\n");
208 imp = d->opd_obd->u.cli.cl_import;
211 req = ptlrpc_request_alloc(imp,
212 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
216 rc = ptlrpc_request_pack(req,
217 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
218 d->opd_pre ? OST_STATFS : MDS_STATFS);
220 ptlrpc_request_free(req);
223 ptlrpc_request_set_replen(req);
225 req->rq_request_portal = OST_CREATE_PORTAL;
226 ptlrpc_at_set_req_timeout(req);
228 req->rq_interpret_reply = osp_statfs_interpret;
229 aa = ptlrpc_req_async_args(aa, req);
230 aa->pointer_arg[0] = d;
233 * no updates till reply
235 del_timer(&d->opd_statfs_timer);
236 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
237 d->opd_statfs_update_in_progress = 1;
239 ptlrpcd_add_req(req);
241 /* we still want to sync changes if no new changes are coming */
242 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
245 if (atomic_read(&d->opd_sync_changes)) {
248 th = dt_trans_create(env, d->opd_storage);
250 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
253 rc = dt_trans_start_local(env, d->opd_storage, th);
255 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
256 d->opd_obd->obd_name,
257 atomic_read(&d->opd_sync_changes));
258 osp_sync_add_commit_cb_1s(env, d, th);
259 dt_trans_stop(env, d->opd_storage, th);
268 * Schedule an immediate update for statfs data
270 * If cached statfs data claim no free space, but OSP has got a request to
271 * destroy an object (so release some space probably), then we may need to
272 * refresh cached statfs data sooner than planned. The function checks there
273 * is no statfs update going and schedules immediate update if so.
274 * XXX: there might be a case where removed object(s) do not add free space (empty
275 * object). If the number of such deletions is high, then we can start to update
276 * statfs too often causing a RPC storm. some throttling is needed...
278 * \param[in] d OSP device where statfs data needs to be refreshed
280 void osp_statfs_need_now(struct osp_device *d)
282 if (!d->opd_statfs_update_in_progress) {
284 * if current status is -ENOSPC (lack of free space on OST)
285 * then we should poll OST immediately once object destroy
288 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
289 del_timer(&d->opd_statfs_timer);
290 wake_up(&d->opd_pre_waitq);
295 * Return number of precreated objects
297 * A simple helper to calculate the number of precreated objects on the device.
299 * \param[in] env LU environment provided by the caller
300 * \param[in] osp OSP device
302 * \retval the number of the precreated objects
304 static inline int osp_objs_precreated(const struct lu_env *env,
305 struct osp_device *osp)
307 return osp_fid_diff(&osp->opd_pre_last_created_fid,
308 &osp->opd_pre_used_fid);
312 * Check pool of precreated objects is nearly empty
314 * We should not wait till the pool of the precreated objects is exhausted,
315 * because then there will be a long period of OSP being unavailable for the
316 * new creations due to lenghty precreate RPC. Instead we ask for another
317 * precreation ahead and hopefully have it ready before the current pool is
318 * empty. Notice this function relies on an external locking.
320 * \param[in] env LU environment provided by the caller
321 * \param[in] d OSP device
323 * \retval 0 - current pool is good enough, 1 - time to precreate
325 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
326 struct osp_device *d)
328 int window = osp_objs_precreated(env, d);
330 /* don't consider new precreation till OST is healty and
332 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
333 (d->opd_pre_status == 0));
337 * Check pool of precreated objects
339 * This is protected version of osp_precreate_near_empty_nolock(), check that
342 * \param[in] env LU environment provided by the caller
343 * \param[in] d OSP device
345 * \retval 0 - current pool is good enough, 1 - time to precreate
347 static inline int osp_precreate_near_empty(const struct lu_env *env,
348 struct osp_device *d)
352 if (d->opd_pre == NULL)
355 /* XXX: do we really need locking here? */
356 spin_lock(&d->opd_pre_lock);
357 rc = osp_precreate_near_empty_nolock(env, d);
358 spin_unlock(&d->opd_pre_lock);
363 * Check given sequence is empty
365 * Returns a binary result whether the given sequence has some IDs left
366 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
367 * version of that function.
369 * \param[in] env LU environment provided by the caller
370 * \param[in] osp OSP device
372 * \retval 0 - current sequence has no IDs, 1 - otherwise
374 static inline int osp_create_end_seq(const struct lu_env *env,
375 struct osp_device *osp)
377 struct lu_fid *fid = &osp->opd_pre_used_fid;
380 spin_lock(&osp->opd_pre_lock);
381 rc = osp_fid_end_seq(env, fid);
382 spin_unlock(&osp->opd_pre_lock);
387 * Write FID into into last_oid/last_seq file
389 * The function stores the sequence and the in-sequence id into two dedicated
390 * files. The sync argument can be used to request synchronous commit, so the
391 * function won't return until the updates are committed.
393 * \param[in] env LU environment provided by the caller
394 * \param[in] osp OSP device
395 * \param[in] fid fid where sequence/id is taken
396 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
398 * \retval 0 on success
399 * \retval negative negated errno on error
401 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
402 struct lu_fid *fid, int sync)
404 struct osp_thread_info *oti = osp_env_info(env);
405 struct lu_buf *lb_oid = &oti->osi_lb;
406 struct lu_buf *lb_oseq = &oti->osi_lb2;
414 if (osp->opd_storage->dd_rdonly)
417 /* Note: through f_oid is only 32 bits, it will also write 64 bits
418 * for oid to keep compatibility with the previous version. */
420 osp_objid_buf_prep(lb_oid, &oid_off,
421 &oid, osp->opd_index);
423 osp_objseq_buf_prep(lb_oseq, &oseq_off,
424 &fid->f_seq, osp->opd_index);
426 th = dt_trans_create(env, osp->opd_storage);
431 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
432 lb_oid, oid_off, th);
436 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
437 lb_oseq, oseq_off, th);
441 rc = dt_trans_start_local(env, osp->opd_storage, th);
445 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
448 CERROR("%s: can not write to last seq file: rc = %d\n",
449 osp->opd_obd->obd_name, rc);
452 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
455 CERROR("%s: can not write to last seq file: rc = %d\n",
456 osp->opd_obd->obd_name, rc);
460 dt_trans_stop(env, osp->opd_storage, th);
465 * Switch to another sequence
467 * When a current sequence has no available IDs left, OSP has to switch to
468 * another new sequence. OSP requests it using the regular FLDB protocol
469 * and stores synchronously before that is used in precreated. This is needed
470 * to basically have the sequences referenced (not orphaned), otherwise it's
471 * possible that OST has some objects precreated and the clients have data
472 * written to it, but after MDT failover nobody refers those objects and OSP
473 * has no idea that the sequence need cleanup to be done.
474 * While this is very expensive operation, it's supposed to happen very very
475 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
477 * \param[in] env LU environment provided by the caller
478 * \param[in] osp OSP device
480 * \retval 0 on success
481 * \retval negative negated errno on error
483 static int osp_precreate_rollover_new_seq(struct lu_env *env,
484 struct osp_device *osp)
486 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
487 struct lu_fid *last_fid = &osp->opd_last_used_fid;
491 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
493 CERROR("%s: alloc fid error: rc = %d\n",
494 osp->opd_obd->obd_name, rc);
500 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
501 "fid "DFID", last_fid "DFID"\n", PFID(fid),
504 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
506 CERROR("%s: Can not update oid/seq file: rc = %d\n",
507 osp->opd_obd->obd_name, rc);
511 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
512 osp->opd_obd->obd_name, fid_seq(last_fid),
514 /* Update last_xxx to the new seq */
515 spin_lock(&osp->opd_pre_lock);
516 osp->opd_last_used_fid = *fid;
517 osp_fid_to_obdid(fid, &osp->opd_last_id);
518 osp->opd_gap_start_fid = *fid;
519 osp->opd_pre_used_fid = *fid;
520 osp->opd_pre_last_created_fid = *fid;
521 spin_unlock(&osp->opd_pre_lock);
527 * Find IDs available in current sequence
529 * The function calculates the highest possible ID and the number of IDs
530 * available in the current sequence OSP is using. The number is limited
531 * artifically by the caller (grow param) and the number of IDs available
532 * in the sequence by nature. The function doesn't require an external
535 * \param[in] env LU environment provided by the caller
536 * \param[in] osp OSP device
537 * \param[in] fid FID the caller wants to start with
538 * \param[in] grow how many the caller wants
539 * \param[out] fid the highest calculated FID
540 * \param[out] grow the number of available IDs calculated
542 * \retval 0 on success, 1 - the sequence is empty
544 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
545 struct lu_fid *fid, int *grow)
547 struct osp_thread_info *osi = osp_env_info(env);
551 if (fid_is_idif(fid)) {
552 struct lu_fid *last_fid;
553 struct ost_id *oi = &osi->osi_oi;
556 spin_lock(&osp->opd_pre_lock);
557 last_fid = &osp->opd_pre_last_created_fid;
558 fid_to_ostid(last_fid, oi);
559 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
560 *grow = end - ostid_id(oi);
561 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
562 spin_unlock(&osp->opd_pre_lock);
564 if (*grow == 0 || rc)
567 ostid_to_fid(fid, oi, osp->opd_index);
571 spin_lock(&osp->opd_pre_lock);
572 *fid = osp->opd_pre_last_created_fid;
574 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
575 *grow = end - fid->f_oid;
576 fid->f_oid += end - fid->f_oid;
577 spin_unlock(&osp->opd_pre_lock);
579 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
580 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
582 return *grow > 0 ? 0 : 1;
586 * Prepare and send precreate RPC
588 * The function finds how many objects should be precreated. Then allocates,
589 * prepares and schedules precreate RPC synchronously. Upon reply the function
590 * wakes up the threads waiting for the new objects on this target. If the
591 * target wasn't able to create all the objects requested, then the next
592 * precreate will be asking for fewer objects (i.e. slow precreate down).
594 * \param[in] env LU environment provided by the caller
595 * \param[in] d OSP device
597 * \retval 0 on success
598 * \retval negative negated errno on error
600 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
602 struct osp_thread_info *oti = osp_env_info(env);
603 struct ptlrpc_request *req;
604 struct obd_import *imp;
605 struct ost_body *body;
607 struct lu_fid *fid = &oti->osi_fid;
610 /* don't precreate new objects till OST healthy and has free space */
611 if (unlikely(d->opd_pre_status)) {
612 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
613 d->opd_obd->obd_name, d->opd_pre_status);
618 * if not connection/initialization is compeleted, ignore
620 imp = d->opd_obd->u.cli.cl_import;
623 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
626 req->rq_request_portal = OST_CREATE_PORTAL;
627 /* we should not resend create request - anyway we will have delorphan
628 * and kill these objects */
629 req->rq_no_delay = req->rq_no_resend = 1;
631 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
633 ptlrpc_request_free(req);
637 spin_lock(&d->opd_pre_lock);
638 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
639 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
640 grow = d->opd_pre_create_count;
641 spin_unlock(&d->opd_pre_lock);
643 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
646 *fid = d->opd_pre_last_created_fid;
647 rc = osp_precreate_fids(env, d, fid, &grow);
649 /* Current seq has been used up*/
650 GOTO(out_req, rc = -ENOSPC);
652 if (!osp_is_fid_client(d)) {
653 /* Non-FID client will always send seq 0 because of
655 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
659 fid_to_ostid(fid, &body->oa.o_oi);
660 body->oa.o_valid = OBD_MD_FLGROUP;
662 ptlrpc_request_set_replen(req);
664 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
667 rc = ptlrpc_queue_wait(req);
669 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
673 LASSERT(req->rq_transno == 0);
675 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
677 GOTO(out_req, rc = -EPROTO);
679 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
682 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
683 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
684 ": rc = %d\n", d->opd_obd->obd_name,
685 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
686 GOTO(out_req, rc = -ESTALE);
689 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
691 spin_lock(&d->opd_pre_lock);
693 /* the OST has not managed to create all the
694 * objects we asked for */
695 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
696 d->opd_pre_create_slow = 1;
698 /* the OST is able to keep up with the work,
699 * we could consider increasing create_count
700 * next time if needed */
701 d->opd_pre_create_slow = 0;
704 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
705 fid_to_ostid(fid, &body->oa.o_oi);
707 d->opd_pre_last_created_fid = *fid;
708 spin_unlock(&d->opd_pre_lock);
710 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
711 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
712 PFID(&d->opd_pre_last_created_fid));
714 /* now we can wakeup all users awaiting for objects */
715 osp_pre_update_status(d, rc);
716 wake_up(&d->opd_pre_user_waitq);
718 ptlrpc_req_finished(req);
723 * Get last precreated object from target (OST)
725 * Sends synchronous RPC to the target (OST) to learn the last precreated
726 * object. This later is used to remove all unused objects (cleanup orphan
727 * procedure). Also, the next object after one we got will be used as a
728 * starting point for the new precreates.
730 * \param[in] env LU environment provided by the caller
731 * \param[in] d OSP device
733 * \retval 0 on success
734 * \retval negative negated errno on error
736 static int osp_get_lastfid_from_ost(const struct lu_env *env,
737 struct osp_device *d)
739 struct ptlrpc_request *req = NULL;
740 struct obd_import *imp;
741 struct lu_fid *last_fid;
746 imp = d->opd_obd->u.cli.cl_import;
749 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
753 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
754 sizeof(KEY_LAST_FID));
756 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
758 ptlrpc_request_free(req);
762 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
763 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
765 req->rq_no_delay = req->rq_no_resend = 1;
766 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
767 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
769 ptlrpc_request_set_replen(req);
771 rc = ptlrpc_queue_wait(req);
773 /* bad-bad OST.. let sysadm sort this out */
774 if (rc == -ENOTSUPP) {
775 CERROR("%s: server does not support FID: rc = %d\n",
776 d->opd_obd->obd_name, -ENOTSUPP);
778 ptlrpc_set_import_active(imp, 0);
782 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
783 if (last_fid == NULL) {
784 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
785 GOTO(out, rc = -EPROTO);
788 if (!fid_is_sane(last_fid)) {
789 CERROR("%s: Got insane last_fid "DFID"\n",
790 d->opd_obd->obd_name, PFID(last_fid));
791 GOTO(out, rc = -EPROTO);
794 /* Only update the last used fid, if the OST has objects for
795 * this sequence, i.e. fid_oid > 0 */
796 if (fid_oid(last_fid) > 0)
797 d->opd_last_used_fid = *last_fid;
799 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
803 ptlrpc_req_finished(req);
808 * Cleanup orphans on OST
810 * This function is called in a contex of a dedicated thread handling
811 * all the precreation suff. The function waits till local recovery
812 * is complete, then identify all the unreferenced objects (orphans)
813 * using the highest ID referenced by a local and the highest object
814 * precreated by the target. The found range is a subject to removal
815 * using specially flagged RPC. During this process OSP is marked
816 * unavailable for new objects.
818 * \param[in] env LU environment provided by the caller
819 * \param[in] d OSP device
821 * \retval 0 on success
822 * \retval negative negated errno on error
824 static int osp_precreate_cleanup_orphans(struct lu_env *env,
825 struct osp_device *d)
827 struct osp_thread_info *osi = osp_env_info(env);
828 struct lu_fid *last_fid = &osi->osi_fid;
829 struct ptlrpc_request *req = NULL;
830 struct obd_import *imp;
831 struct ost_body *body;
832 int update_status = 0;
839 * wait for local recovery to finish, so we can cleanup orphans
840 * orphans are all objects since "last used" (assigned), but
841 * there might be objects reserved and in some cases they won't
842 * be used. we can't cleanup them till we're sure they won't be
843 * used. also can't we allow new reservations because they may
844 * end up getting orphans being cleaned up below. so we block
845 * new reservations and wait till all reserved objects either
848 spin_lock(&d->opd_pre_lock);
849 d->opd_pre_recovering = 1;
850 spin_unlock(&d->opd_pre_lock);
852 * The locking above makes sure the opd_pre_reserved check below will
853 * catch all osp_precreate_reserve() calls who find
854 * "!opd_pre_recovering".
856 wait_event_idle(d->opd_pre_waitq,
857 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
858 !d->opd_pre_task || d->opd_got_disconnected);
859 if (!d->opd_pre_task || d->opd_got_disconnected)
860 GOTO(out, rc = -EAGAIN);
862 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
863 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
865 *last_fid = d->opd_last_used_fid;
866 /* The OSP should already get the valid seq now */
867 LASSERT(!fid_is_zero(last_fid));
868 if (fid_oid(&d->opd_last_used_fid) < 2) {
869 /* lastfid looks strange... ask OST */
870 rc = osp_get_lastfid_from_ost(env, d);
875 imp = d->opd_obd->u.cli.cl_import;
878 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
880 GOTO(out, rc = -ENOMEM);
882 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
884 ptlrpc_request_free(req);
889 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
891 GOTO(out, rc = -EPROTO);
893 body->oa.o_flags = OBD_FL_DELORPHAN;
894 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
896 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
898 ptlrpc_request_set_replen(req);
900 /* Don't resend the delorphan req */
901 req->rq_no_resend = req->rq_no_delay = 1;
903 rc = ptlrpc_queue_wait(req);
909 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
911 GOTO(out, rc = -EPROTO);
914 * OST provides us with id new pool starts from in body->oa.o_id
916 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
918 spin_lock(&d->opd_pre_lock);
919 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
921 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
922 d->opd_pre_last_created_fid = d->opd_last_used_fid;
924 d->opd_pre_create_count = OST_MIN_PRECREATE;
925 d->opd_pre_last_created_fid = *last_fid;
928 * This empties the pre-creation pool and effectively blocks any new
931 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
932 LUSTRE_DATA_SEQ_MAX_WIDTH);
933 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
934 d->opd_pre_create_slow = 0;
935 spin_unlock(&d->opd_pre_lock);
937 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
938 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
939 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
942 ptlrpc_req_finished(req);
945 * If rc is zero, the pre-creation window should have been emptied.
946 * Since waking up the herd would be useless without pre-created
947 * objects, we defer the signal to osp_precreate_send() in that case.
951 CERROR("%s: cannot cleanup orphans: rc = %d\n",
952 d->opd_obd->obd_name, rc);
953 /* we can't proceed from here, OST seem to
954 * be in a bad shape, better to wait for
955 * a new instance of the server and repeat
956 * from the beginning. notify possible waiters
957 * this OSP isn't quite functional yet */
958 osp_pre_update_status(d, rc);
960 wake_up(&d->opd_pre_user_waitq);
963 spin_lock(&d->opd_pre_lock);
964 d->opd_pre_recovering = 0;
965 spin_unlock(&d->opd_pre_lock);
972 * Update precreate status using statfs data
974 * The function decides whether this OSP should be used for new objects.
975 * IOW, whether this OST is used up or has some free space. Cached statfs
976 * data is used to make this decision. If the latest result of statfs
977 * request (rc argument) is not success, then just mark OSP unavailable
980 * The new statfs data is passed in \a msfs and needs to be stored into
981 * opd_statfs, but only after the various flags in os_state are set, so
982 * that the new statfs data is not visible without appropriate flags set.
983 * As such, there is no need to clear the flags here, since this is called
984 * with new statfs data, and they should not be cleared if sent from OST.
986 * Add a bit of hysteresis so this flag isn't continually flapping, and
987 * ensure that new files don't get extremely fragmented due to only a
988 * small amount of available space in the filesystem. We want to set
989 * the ENOSPC/ENOINO flags unconditionally when there is less than the
990 * reserved size free, and still copy them from the old state when there
991 * is less than 2*reserved size free space or inodes.
993 * \param[in] d OSP device
994 * \param[in] msfs statfs data
996 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
998 u32 old_state = d->opd_statfs.os_state;
999 u32 reserved_ino_low = 32; /* could be tunable in the future */
1000 u32 reserved_ino_high = reserved_ino_low * 2;
1003 /* statfs structure not initialized yet */
1004 if (unlikely(!msfs->os_type))
1007 /* if the low and high watermarks have not been initialized yet */
1008 if (unlikely(d->opd_reserved_mb_high == 0 &&
1009 d->opd_reserved_mb_low == 0)) {
1010 /* Use ~0.1% by default to disable object allocation,
1011 * and ~0.2% to enable, size in MB, set both watermark
1013 spin_lock(&d->opd_pre_lock);
1014 if (d->opd_reserved_mb_high == 0 &&
1015 d->opd_reserved_mb_low == 0) {
1016 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1017 msfs->os_blocks) >> 20;
1018 if (d->opd_reserved_mb_low == 0)
1019 d->opd_reserved_mb_low = 1;
1020 d->opd_reserved_mb_high =
1021 (d->opd_reserved_mb_low << 1) + 1;
1023 spin_unlock(&d->opd_pre_lock);
1026 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1027 if (msfs->os_ffree < reserved_ino_low)
1028 msfs->os_state |= OS_STATE_ENOINO;
1029 else if (msfs->os_ffree <= reserved_ino_high)
1030 msfs->os_state |= old_state & OS_STATE_ENOINO;
1031 /* else don't clear flags in new msfs->os_state sent from OST */
1034 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1035 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1036 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1037 msfs->os_files, msfs->os_ffree, msfs->os_state,
1039 if (available_mb < d->opd_reserved_mb_low)
1040 msfs->os_state |= OS_STATE_ENOSPC;
1041 else if (available_mb <= d->opd_reserved_mb_high)
1042 msfs->os_state |= old_state & OS_STATE_ENOSPC;
1043 /* else don't clear flags in new msfs->os_state sent from OST */
1045 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1046 d->opd_pre_status = -ENOSPC;
1047 if (!(old_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)))
1048 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1049 d->opd_obd->obd_name, msfs->os_state,
1051 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1052 atomic_read(&d->opd_sync_changes),
1053 atomic_read(&d->opd_sync_rpcs_in_progress));
1054 } else if (old_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1055 d->opd_pre_status = 0;
1056 spin_lock(&d->opd_pre_lock);
1057 d->opd_pre_create_slow = 0;
1058 d->opd_pre_create_count = OST_MIN_PRECREATE;
1059 spin_unlock(&d->opd_pre_lock);
1060 wake_up(&d->opd_pre_waitq);
1063 "%s: available: state=%x: rc = %d\n",
1064 d->opd_obd->obd_name, msfs->os_state,
1067 /* we only get here if rc == 0 in the caller */
1068 d->opd_pre_status = 0;
1071 /* Object precreation skipped on OST if manually disabled */
1072 if (d->opd_pre_max_create_count == 0)
1073 msfs->os_state |= OS_STATE_NOPRECREATE;
1074 /* else don't clear flags in new msfs->os_state sent from OST */
1076 /* copy only new statfs state to make it visible to MDS threads */
1077 if (&d->opd_statfs != msfs)
1078 d->opd_statfs = *msfs;
1082 * Initialize FID for precreation
1084 * For a just created new target, a new sequence should be taken.
1085 * The function checks there is no IDIF in use (if the target was
1086 * added with the older version of Lustre), then requests a new
1087 * sequence from FLDB using the regular protocol. Then this new
1088 * sequence is stored on a persisten storage synchronously to prevent
1089 * possible object leakage (for the detail see the description for
1090 * osp_precreate_rollover_new_seq()).
1092 * \param[in] osp OSP device
1094 * \retval 0 on success
1095 * \retval negative negated errno on error
1097 int osp_init_pre_fid(struct osp_device *osp)
1100 struct osp_thread_info *osi;
1101 struct lu_client_seq *cli_seq;
1102 struct lu_fid *last_fid;
1106 LASSERT(osp->opd_pre != NULL);
1108 /* Let's check if the current last_seq/fid is valid,
1109 * otherwise request new sequence from the controller */
1110 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1111 /* Non-MDT0 can only use normal sequence for
1113 if (fid_is_norm(&osp->opd_last_used_fid))
1116 /* Initially MDT0 will start with IDIF, after
1117 * that it will request new sequence from the
1119 if (fid_is_idif(&osp->opd_last_used_fid) ||
1120 fid_is_norm(&osp->opd_last_used_fid))
1124 if (!fid_is_zero(&osp->opd_last_used_fid))
1125 CWARN("%s: invalid last used fid "DFID
1126 ", try to get new sequence.\n",
1127 osp->opd_obd->obd_name,
1128 PFID(&osp->opd_last_used_fid));
1130 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1132 CERROR("%s: init env error: rc = %d\n",
1133 osp->opd_obd->obd_name, rc);
1137 osi = osp_env_info(&env);
1138 last_fid = &osi->osi_fid;
1140 /* For a freshed fs, it will allocate a new sequence first */
1141 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1142 cli_seq = osp->opd_obd->u.cli.cl_seq;
1143 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1145 CERROR("%s: alloc fid error: rc = %d\n",
1146 osp->opd_obd->obd_name, rc);
1150 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1152 last_fid->f_oid = 1;
1153 last_fid->f_ver = 0;
1155 spin_lock(&osp->opd_pre_lock);
1156 osp->opd_last_used_fid = *last_fid;
1157 osp->opd_pre_used_fid = *last_fid;
1158 osp->opd_pre_last_created_fid = *last_fid;
1159 spin_unlock(&osp->opd_pre_lock);
1160 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1162 CERROR("%s: write fid error: rc = %d\n",
1163 osp->opd_obd->obd_name, rc);
1172 struct osp_device *opta_dev;
1173 struct lu_env opta_env;
1174 struct completion *opta_started;
1177 * The core of precreate functionality
1179 * The function implements the main precreation loop. Basically it
1180 * involves connecting to the target, precerate FID initialization,
1181 * identifying and removing orphans, then serving precreation. As
1182 * part of the latter, the thread is responsible for statfs data
1183 * updates. The precreation is mostly driven by another threads
1184 * asking for new OST objects - those askers wake the thread when
1185 * the number of precreated objects reach low watermark.
1186 * After a disconnect, the sequence above repeats. This is keep going
1187 * until the thread is requested to stop.
1189 * \param[in] _arg private data the thread (OSP device to handle)
1191 * \retval 0 on success
1192 * \retval negative negated errno on error
1194 static int osp_precreate_thread(void *_args)
1196 struct opt_args *args = _args;
1197 struct osp_device *d = args->opta_dev;
1198 struct lu_env *env = &args->opta_env;
1203 complete(args->opta_started);
1204 while (!kthread_should_stop()) {
1206 * need to be connected to OST
1208 while (!kthread_should_stop()) {
1209 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1210 d->opd_imp_connected &&
1211 !d->opd_got_disconnected)
1213 wait_event_idle(d->opd_pre_waitq,
1214 kthread_should_stop() ||
1215 d->opd_new_connection);
1217 if (!d->opd_new_connection)
1220 d->opd_new_connection = 0;
1221 d->opd_got_disconnected = 0;
1225 if (kthread_should_stop())
1229 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1230 /* Sigh, fid client is not ready yet */
1231 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1234 /* Init fid for osp_precreate if necessary */
1235 rc = osp_init_pre_fid(d);
1237 class_export_put(d->opd_exp);
1238 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1239 CERROR("%s: init pre fid error: rc = %d\n",
1240 d->opd_obd->obd_name, rc);
1245 if (osp_statfs_update(env, d)) {
1246 if (wait_event_idle_timeout(d->opd_pre_waitq,
1247 kthread_should_stop(),
1248 cfs_time_seconds(5)) == 0)
1249 l_wait_event_abortable(
1251 kthread_should_stop());
1257 * Clean up orphans or recreate missing objects.
1259 rc = osp_precreate_cleanup_orphans(env, d);
1261 schedule_timeout_interruptible(cfs_time_seconds(1));
1267 * connected, can handle precreates now
1269 while (!kthread_should_stop()) {
1270 wait_event_idle(d->opd_pre_waitq,
1271 kthread_should_stop() ||
1272 osp_precreate_near_empty(env, d) ||
1273 osp_statfs_need_update(d) ||
1274 d->opd_got_disconnected);
1276 if (kthread_should_stop())
1279 /* something happened to the connection
1280 * have to start from the beginning */
1281 if (d->opd_got_disconnected)
1284 if (osp_statfs_need_update(d))
1285 if (osp_statfs_update(env, d))
1288 if (d->opd_pre == NULL)
1291 /* To avoid handling different seq in precreate/orphan
1292 * cleanup, it will hold precreate until current seq is
1294 if (unlikely(osp_precreate_end_seq(env, d) &&
1295 !osp_create_end_seq(env, d)))
1298 if (unlikely(osp_precreate_end_seq(env, d) &&
1299 osp_create_end_seq(env, d))) {
1300 LCONSOLE_INFO("%s:%#llx is used up."
1301 " Update to new seq\n",
1302 d->opd_obd->obd_name,
1303 fid_seq(&d->opd_pre_last_created_fid));
1304 rc = osp_precreate_rollover_new_seq(env, d);
1309 if (osp_precreate_near_empty(env, d)) {
1310 rc = osp_precreate_send(env, d);
1311 /* osp_precreate_send() sets opd_pre_status
1312 * in case of error, that prevent the using of
1314 if (rc < 0 && rc != -ENOSPC &&
1315 rc != -ETIMEDOUT && rc != -ENOTCONN)
1316 CERROR("%s: cannot precreate objects:"
1318 d->opd_obd->obd_name, rc);
1330 * Check when to stop to wait for precreate objects.
1332 * The caller wanting a new OST object can't wait undefinitely. The
1333 * function checks for few conditions including available new OST
1334 * objects, disconnected OST, lack of space with no pending destroys,
1335 * etc. IOW, it checks whether the current OSP state is good to keep
1336 * waiting or it's better to give up.
1338 * \param[in] env LU environment provided by the caller
1339 * \param[in] d OSP device
1341 * \retval 0 - keep waiting, 1 - no luck
1343 static int osp_precreate_ready_condition(const struct lu_env *env,
1344 struct osp_device *d)
1346 if (d->opd_pre_recovering)
1349 /* ready if got enough precreated objects */
1350 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1351 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1354 /* ready if OST reported no space and no destroys in progress */
1355 if (atomic_read(&d->opd_sync_changes) +
1356 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1357 d->opd_pre_status == -ENOSPC)
1360 /* Bail out I/O fails to OST */
1361 if (d->opd_pre_status != 0 &&
1362 d->opd_pre_status != -EAGAIN &&
1363 d->opd_pre_status != -ENODEV &&
1364 d->opd_pre_status != -ENOTCONN &&
1365 d->opd_pre_status != -ENOSPC) {
1367 if (d->opd_pre_status != -EIO)
1368 CERROR("%s: precreate failed opd_pre_status %d\n",
1369 d->opd_obd->obd_name, d->opd_pre_status);
1377 * Reserve object in precreate pool
1379 * When the caller wants to create a new object on this target (target
1380 * represented by the given OSP), it should declare this intention using
1381 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1382 * to reserve an object in the existing precreated pool or wait up to
1383 * obd_timeout for the available object to appear in the pool (a dedicated
1384 * thread will be doing real precreation in background). The object can be
1385 * consumed later with osp_precreate_get_fid() or be released with call to
1386 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1387 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1388 * If the space on the target is short and there is a pending object destroy,
1389 * then the function forces local commit to speedup space release (see
1390 * osp_sync.c for the details).
1392 * \param[in] env LU environment provided by the caller
1393 * \param[in] d OSP device
1395 * \retval 0 on success
1396 * \retval -ENOSPC when no space on OST
1397 * \retval -EAGAIN try later, slow precreation in progress
1398 * \retval -EIO when no access to OST
1400 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1402 time64_t expire = ktime_get_seconds() + obd_timeout;
1403 int precreated, rc, synced = 0;
1407 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1408 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1409 PFID(&d->opd_pre_used_fid));
1411 /* opd_pre_max_create_count 0 to not use specified OST. */
1412 if (d->opd_pre_max_create_count == 0)
1417 * - preallocation is done
1418 * - no free space expected soon
1419 * - can't connect to OST for too long (obd_timeout)
1420 * - OST can allocate fid sequence.
1422 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1423 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1426 * increase number of precreations
1428 precreated = osp_objs_precreated(env, d);
1429 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1430 d->opd_pre_create_slow == 0 &&
1431 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1432 spin_lock(&d->opd_pre_lock);
1433 d->opd_pre_create_slow = 1;
1434 d->opd_pre_create_count *= 2;
1435 spin_unlock(&d->opd_pre_lock);
1438 spin_lock(&d->opd_pre_lock);
1439 precreated = osp_objs_precreated(env, d);
1440 if (precreated > d->opd_pre_reserved &&
1441 !d->opd_pre_recovering) {
1442 d->opd_pre_reserved++;
1443 spin_unlock(&d->opd_pre_lock);
1446 /* XXX: don't wake up if precreation is in progress */
1447 if (osp_precreate_near_empty_nolock(env, d) &&
1448 !osp_precreate_end_seq_nolock(env, d))
1449 wake_up(&d->opd_pre_waitq);
1453 spin_unlock(&d->opd_pre_lock);
1456 * all precreated objects have been used and no-space
1457 * status leave us no chance to succeed very soon
1458 * but if there is destroy in progress, then we should
1459 * wait till that is done - some space might be released
1461 if (unlikely(rc == -ENOSPC)) {
1462 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1463 /* force local commit to release space */
1464 dt_commit_async(env, d->opd_storage);
1465 osp_sync_check_for_work(d);
1468 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1469 /* just wait till destroys are done
1470 * see wait_event_idle_timeout() below
1473 if (atomic_read(&d->opd_sync_changes) +
1474 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1475 /* no hope for free space */
1480 /* XXX: don't wake up if precreation is in progress */
1481 wake_up(&d->opd_pre_waitq);
1483 if (ktime_get_seconds() >= expire) {
1488 if (wait_event_idle_timeout(
1489 d->opd_pre_user_waitq,
1490 osp_precreate_ready_condition(env, d),
1491 cfs_time_seconds(obd_timeout)) == 0) {
1493 "%s: slow creates, last="DFID", next="DFID", "
1494 "reserved=%llu, sync_changes=%u, "
1495 "sync_rpcs_in_progress=%d, status=%d\n",
1496 d->opd_obd->obd_name,
1497 PFID(&d->opd_pre_last_created_fid),
1498 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1499 atomic_read(&d->opd_sync_changes),
1500 atomic_read(&d->opd_sync_rpcs_in_progress),
1509 * Get a FID from precreation pool
1511 * The function is a companion for osp_precreate_reserve() - it assigns
1512 * a specific FID from the precreate. The function should be called only
1513 * if the call to osp_precreate_reserve() was successful. The function
1514 * updates a local storage to remember the highest object ID referenced
1515 * by the node in the given sequence.
1517 * A very importan details: this is supposed to be called once the
1518 * transaction is started, so on-disk update will be atomic with the
1519 * data (like LOVEA) refering this object. Then the object won't be leaked:
1520 * either it's referenced by the committed transaction or it's a subject
1521 * to the orphan cleanup procedure.
1523 * \param[in] env LU environment provided by the caller
1524 * \param[in] d OSP device
1525 * \param[out] fid generated FID
1527 * \retval 0 on success
1528 * \retval negative negated errno on error
1530 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1533 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1534 /* grab next id from the pool */
1535 spin_lock(&d->opd_pre_lock);
1537 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1538 &d->opd_pre_last_created_fid) < 0,
1539 "next fid "DFID" last created fid "DFID"\n",
1540 PFID(&d->opd_pre_used_fid),
1541 PFID(&d->opd_pre_last_created_fid));
1544 * When sequence is used up, new one should be allocated in
1545 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1548 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1549 "next fid "DFID" last created fid "DFID"\n",
1550 PFID(&d->opd_pre_used_fid),
1551 PFID(&d->opd_pre_last_created_fid));
1552 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1553 if (fid_is_idif(pre_used_fid) &&
1554 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1555 pre_used_fid->f_seq++;
1557 d->opd_pre_used_fid.f_oid++;
1558 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1559 d->opd_pre_reserved--;
1561 * last_used_id must be changed along with getting new id otherwise
1562 * we might miscalculate gap causing object loss or leak
1564 osp_update_last_fid(d, fid);
1565 spin_unlock(&d->opd_pre_lock);
1568 * probably main thread suspended orphan cleanup till
1569 * all reservations are released, see comment in
1570 * osp_precreate_thread() just before orphan cleanup
1572 if (unlikely(d->opd_pre_reserved == 0 &&
1573 (d->opd_pre_recovering || d->opd_pre_status)))
1574 wake_up(&d->opd_pre_waitq);
1580 * Set size regular attribute on an object
1582 * When a striping is created late, it's possible that size is already
1583 * initialized on the file. Then the new striping should inherit size
1584 * from the file. The function sets size on the object using the regular
1585 * protocol (OST_PUNCH).
1586 * XXX: should be re-implemented using OUT ?
1588 * \param[in] env LU environment provided by the caller
1589 * \param[in] dt object
1590 * \param[in] size size to set.
1592 * \retval 0 on success
1593 * \retval negative negated errno on error
1595 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1598 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1599 struct ptlrpc_request *req = NULL;
1600 struct obd_import *imp;
1601 struct ost_body *body;
1602 struct obdo *oa = NULL;
1607 imp = d->opd_obd->u.cli.cl_import;
1610 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1614 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1616 ptlrpc_request_free(req);
1621 * XXX: decide how do we do here with resend
1622 * if we don't resend, then client may see wrong file size
1623 * if we do resend, then MDS thread can get stuck for quite long
1624 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1625 * (see LU-7975 and sanity/test_27F use cases)
1626 * but let's decide not to resend/delay this truncate request to OST
1627 * and allow Client to decide to resend, in a less agressive way from
1628 * after_reply(), by returning -EINPROGRESS instead of
1629 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1630 * end of this routine
1632 req->rq_no_resend = req->rq_no_delay = 1;
1634 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1635 ptlrpc_at_set_req_timeout(req);
1639 GOTO(out, rc = -ENOMEM);
1641 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1644 oa->o_blocks = OBD_OBJECT_EOF;
1645 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1646 OBD_MD_FLID | OBD_MD_FLGROUP;
1648 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1650 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1652 /* XXX: capa support? */
1653 /* osc_pack_capa(req, body, capa); */
1655 ptlrpc_request_set_replen(req);
1657 rc = ptlrpc_queue_wait(req);
1659 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1660 * since we have decided not to resend/delay, but this could
1661 * lead to wrong size to be seen at Client side and even process
1662 * trying to open to exit/fail if not itself handling -EAGAIN.
1663 * So it should be better to return -EINPROGRESS instead and
1664 * leave the decision to resend at Client side in after_reply()
1666 if (rc == -EWOULDBLOCK) {
1668 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1669 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1672 CERROR("can't punch object: %d\n", rc);
1676 ptlrpc_req_finished(req);
1683 * Initialize precreation functionality of OSP
1685 * Prepares all the internal structures and starts the precreate thread
1687 * \param[in] d OSP device
1689 * \retval 0 on success
1690 * \retval negative negated errno on error
1692 int osp_init_precreate(struct osp_device *d)
1696 OBD_ALLOC_PTR(d->opd_pre);
1697 if (d->opd_pre == NULL)
1700 /* initially precreation isn't ready */
1701 init_waitqueue_head(&d->opd_pre_user_waitq);
1702 d->opd_pre_status = -EAGAIN;
1703 fid_zero(&d->opd_pre_used_fid);
1704 d->opd_pre_used_fid.f_oid = 1;
1705 fid_zero(&d->opd_pre_last_created_fid);
1706 d->opd_pre_last_created_fid.f_oid = 1;
1708 d->opd_pre_reserved = 0;
1709 d->opd_got_disconnected = 1;
1710 d->opd_pre_create_slow = 0;
1711 d->opd_pre_create_count = OST_MIN_PRECREATE;
1712 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1713 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1714 d->opd_reserved_mb_high = 0;
1715 d->opd_reserved_mb_low = 0;
1721 * Finish precreate functionality of OSP
1724 * Asks all the activity (the thread, update timer) to stop, then
1725 * wait till that is done.
1727 * \param[in] d OSP device
1729 void osp_precreate_fini(struct osp_device *d)
1733 if (d->opd_pre == NULL)
1736 OBD_FREE_PTR(d->opd_pre);
1742 int osp_init_statfs(struct osp_device *d)
1744 struct task_struct *task;
1745 struct opt_args *args;
1746 DECLARE_COMPLETION_ONSTACK(started);
1751 spin_lock_init(&d->opd_pre_lock);
1752 init_waitqueue_head(&d->opd_pre_waitq);
1755 * Initialize statfs-related things
1757 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1758 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1759 1000 * NSEC_PER_SEC);
1760 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1762 ktime_to_ns(d->opd_statfs_fresh_till));
1763 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1764 (unsigned long)d, 0);
1766 if (d->opd_storage->dd_rdonly)
1769 OBD_ALLOC_PTR(args);
1773 args->opta_started = &started;
1774 rc = lu_env_init(&args->opta_env,
1775 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1777 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1784 * start thread handling precreation and statfs updates
1786 task = kthread_create(osp_precreate_thread, args,
1787 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1789 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1790 lu_env_fini(&args->opta_env);
1792 RETURN(PTR_ERR(task));
1794 d->opd_pre_task = task;
1795 wake_up_process(task);
1796 wait_for_completion(&started);
1801 void osp_statfs_fini(struct osp_device *d)
1803 struct task_struct *task = d->opd_pre_task;
1806 del_timer(&d->opd_statfs_timer);
1808 d->opd_pre_task = NULL;