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 obd_statfs *sfs;
143 struct osp_device *d;
148 aa = ptlrpc_req_async_args(aa, req);
149 d = aa->pointer_arg[0];
155 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
157 GOTO(out, rc = -EPROTO);
160 osp_pre_update_status_msfs(d, msfs, 0);
162 d->opd_statfs = *msfs;
164 /* schedule next update */
165 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
166 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
167 mod_timer(&d->opd_statfs_timer,
168 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
169 d->opd_statfs_update_in_progress = 0;
171 sfs = &d->opd_statfs;
172 CDEBUG(D_CACHE, "%s (%p): %llu blocks, %llu free, %llu avail, "
173 "%u bsize, %u reserved mb low, %u reserved mb high,"
174 "%llu files, %llu free files\n", d->opd_obd->obd_name, d,
175 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail, sfs->os_bsize,
176 d->opd_reserved_mb_low, d->opd_reserved_mb_high,
177 sfs->os_files, sfs->os_ffree);
181 /* couldn't update statfs, try again with a small delay */
182 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
183 d->opd_statfs_update_in_progress = 0;
184 if (d->opd_pre && d->opd_pre_task)
185 wake_up(&d->opd_pre_waitq);
187 if (req->rq_import_generation == imp->imp_generation)
188 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
189 d->opd_obd->obd_name, rc);
194 * Send OST_STATFS RPC
196 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
197 * Also disables scheduled updates as times OSP may need to refresh
198 * statfs data before expiration. The function doesn't block, instead
199 * an interpretation callback osp_statfs_interpret() is used.
201 * \param[in] d OSP device
203 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
205 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
206 struct ptlrpc_request *req;
207 struct obd_import *imp;
208 union ptlrpc_async_args *aa;
213 CDEBUG(D_CACHE, "going to update statfs\n");
215 imp = d->opd_obd->u.cli.cl_import;
218 req = ptlrpc_request_alloc(imp,
219 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
223 rc = ptlrpc_request_pack(req,
224 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
225 d->opd_pre ? OST_STATFS : MDS_STATFS);
227 ptlrpc_request_free(req);
230 ptlrpc_request_set_replen(req);
232 req->rq_request_portal = OST_CREATE_PORTAL;
233 ptlrpc_at_set_req_timeout(req);
235 req->rq_interpret_reply = osp_statfs_interpret;
236 aa = ptlrpc_req_async_args(aa, req);
237 aa->pointer_arg[0] = d;
240 * no updates till reply
242 del_timer(&d->opd_statfs_timer);
243 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
244 d->opd_statfs_update_in_progress = 1;
246 ptlrpcd_add_req(req);
248 /* we still want to sync changes if no new changes are coming */
249 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
252 if (atomic_read(&d->opd_sync_changes)) {
255 th = dt_trans_create(env, d->opd_storage);
257 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
260 rc = dt_trans_start_local(env, d->opd_storage, th);
262 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
263 d->opd_obd->obd_name,
264 atomic_read(&d->opd_sync_changes));
265 osp_sync_add_commit_cb_1s(env, d, th);
266 dt_trans_stop(env, d->opd_storage, th);
275 * Schedule an immediate update for statfs data
277 * If cached statfs data claim no free space, but OSP has got a request to
278 * destroy an object (so release some space probably), then we may need to
279 * refresh cached statfs data sooner than planned. The function checks there
280 * is no statfs update going and schedules immediate update if so.
281 * XXX: there might be a case where removed object(s) do not add free space (empty
282 * object). If the number of such deletions is high, then we can start to update
283 * statfs too often causing a RPC storm. some throttling is needed...
285 * \param[in] d OSP device where statfs data needs to be refreshed
287 void osp_statfs_need_now(struct osp_device *d)
289 if (!d->opd_statfs_update_in_progress) {
291 * if current status is -ENOSPC (lack of free space on OST)
292 * then we should poll OST immediately once object destroy
295 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
296 del_timer(&d->opd_statfs_timer);
297 wake_up(&d->opd_pre_waitq);
302 * Return number of precreated objects
304 * A simple helper to calculate the number of precreated objects on the device.
306 * \param[in] env LU environment provided by the caller
307 * \param[in] osp OSP device
309 * \retval the number of the precreated objects
311 static inline int osp_objs_precreated(const struct lu_env *env,
312 struct osp_device *osp)
314 return osp_fid_diff(&osp->opd_pre_last_created_fid,
315 &osp->opd_pre_used_fid);
319 * Check pool of precreated objects is nearly empty
321 * We should not wait till the pool of the precreated objects is exhausted,
322 * because then there will be a long period of OSP being unavailable for the
323 * new creations due to lenghty precreate RPC. Instead we ask for another
324 * precreation ahead and hopefully have it ready before the current pool is
325 * empty. Notice this function relies on an external locking.
327 * \param[in] env LU environment provided by the caller
328 * \param[in] d OSP device
330 * \retval 0 - current pool is good enough, 1 - time to precreate
332 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
333 struct osp_device *d)
335 int window = osp_objs_precreated(env, d);
337 /* don't consider new precreation till OST is healty and
339 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
340 (d->opd_pre_status == 0));
344 * Check pool of precreated objects
346 * This is protected version of osp_precreate_near_empty_nolock(), check that
349 * \param[in] env LU environment provided by the caller
350 * \param[in] d OSP device
352 * \retval 0 - current pool is good enough, 1 - time to precreate
354 static inline int osp_precreate_near_empty(const struct lu_env *env,
355 struct osp_device *d)
359 if (d->opd_pre == NULL)
362 /* XXX: do we really need locking here? */
363 spin_lock(&d->opd_pre_lock);
364 rc = osp_precreate_near_empty_nolock(env, d);
365 spin_unlock(&d->opd_pre_lock);
370 * Check given sequence is empty
372 * Returns a binary result whether the given sequence has some IDs left
373 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
374 * version of that function.
376 * \param[in] env LU environment provided by the caller
377 * \param[in] osp OSP device
379 * \retval 0 - current sequence has no IDs, 1 - otherwise
381 static inline int osp_create_end_seq(const struct lu_env *env,
382 struct osp_device *osp)
384 struct lu_fid *fid = &osp->opd_pre_used_fid;
387 spin_lock(&osp->opd_pre_lock);
388 rc = osp_fid_end_seq(env, fid);
389 spin_unlock(&osp->opd_pre_lock);
394 * Write FID into into last_oid/last_seq file
396 * The function stores the sequence and the in-sequence id into two dedicated
397 * files. The sync argument can be used to request synchronous commit, so the
398 * function won't return until the updates are committed.
400 * \param[in] env LU environment provided by the caller
401 * \param[in] osp OSP device
402 * \param[in] fid fid where sequence/id is taken
403 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
405 * \retval 0 on success
406 * \retval negative negated errno on error
408 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
409 struct lu_fid *fid, int sync)
411 struct osp_thread_info *oti = osp_env_info(env);
412 struct lu_buf *lb_oid = &oti->osi_lb;
413 struct lu_buf *lb_oseq = &oti->osi_lb2;
421 if (osp->opd_storage->dd_rdonly)
424 /* Note: through f_oid is only 32 bits, it will also write 64 bits
425 * for oid to keep compatibility with the previous version. */
427 osp_objid_buf_prep(lb_oid, &oid_off,
428 &oid, osp->opd_index);
430 osp_objseq_buf_prep(lb_oseq, &oseq_off,
431 &fid->f_seq, osp->opd_index);
433 th = dt_trans_create(env, osp->opd_storage);
438 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
439 lb_oid, oid_off, th);
443 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
444 lb_oseq, oseq_off, th);
448 rc = dt_trans_start_local(env, osp->opd_storage, th);
452 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
455 CERROR("%s: can not write to last seq file: rc = %d\n",
456 osp->opd_obd->obd_name, rc);
459 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
462 CERROR("%s: can not write to last seq file: rc = %d\n",
463 osp->opd_obd->obd_name, rc);
467 dt_trans_stop(env, osp->opd_storage, th);
472 * Switch to another sequence
474 * When a current sequence has no available IDs left, OSP has to switch to
475 * another new sequence. OSP requests it using the regular FLDB protocol
476 * and stores synchronously before that is used in precreated. This is needed
477 * to basically have the sequences referenced (not orphaned), otherwise it's
478 * possible that OST has some objects precreated and the clients have data
479 * written to it, but after MDT failover nobody refers those objects and OSP
480 * has no idea that the sequence need cleanup to be done.
481 * While this is very expensive operation, it's supposed to happen very very
482 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
484 * \param[in] env LU environment provided by the caller
485 * \param[in] osp OSP device
487 * \retval 0 on success
488 * \retval negative negated errno on error
490 static int osp_precreate_rollover_new_seq(struct lu_env *env,
491 struct osp_device *osp)
493 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
494 struct lu_fid *last_fid = &osp->opd_last_used_fid;
498 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
500 CERROR("%s: alloc fid error: rc = %d\n",
501 osp->opd_obd->obd_name, rc);
507 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
508 "fid "DFID", last_fid "DFID"\n", PFID(fid),
511 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
513 CERROR("%s: Can not update oid/seq file: rc = %d\n",
514 osp->opd_obd->obd_name, rc);
518 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
519 osp->opd_obd->obd_name, fid_seq(last_fid),
521 /* Update last_xxx to the new seq */
522 spin_lock(&osp->opd_pre_lock);
523 osp->opd_last_used_fid = *fid;
524 osp_fid_to_obdid(fid, &osp->opd_last_id);
525 osp->opd_gap_start_fid = *fid;
526 osp->opd_pre_used_fid = *fid;
527 osp->opd_pre_last_created_fid = *fid;
528 spin_unlock(&osp->opd_pre_lock);
534 * Find IDs available in current sequence
536 * The function calculates the highest possible ID and the number of IDs
537 * available in the current sequence OSP is using. The number is limited
538 * artifically by the caller (grow param) and the number of IDs available
539 * in the sequence by nature. The function doesn't require an external
542 * \param[in] env LU environment provided by the caller
543 * \param[in] osp OSP device
544 * \param[in] fid FID the caller wants to start with
545 * \param[in] grow how many the caller wants
546 * \param[out] fid the highest calculated FID
547 * \param[out] grow the number of available IDs calculated
549 * \retval 0 on success, 1 - the sequence is empty
551 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
552 struct lu_fid *fid, int *grow)
554 struct osp_thread_info *osi = osp_env_info(env);
558 if (fid_is_idif(fid)) {
559 struct lu_fid *last_fid;
560 struct ost_id *oi = &osi->osi_oi;
563 spin_lock(&osp->opd_pre_lock);
564 last_fid = &osp->opd_pre_last_created_fid;
565 fid_to_ostid(last_fid, oi);
566 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
567 *grow = end - ostid_id(oi);
568 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
569 spin_unlock(&osp->opd_pre_lock);
571 if (*grow == 0 || rc)
574 ostid_to_fid(fid, oi, osp->opd_index);
578 spin_lock(&osp->opd_pre_lock);
579 *fid = osp->opd_pre_last_created_fid;
581 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
582 *grow = end - fid->f_oid;
583 fid->f_oid += end - fid->f_oid;
584 spin_unlock(&osp->opd_pre_lock);
586 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
587 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
589 return *grow > 0 ? 0 : 1;
593 * Prepare and send precreate RPC
595 * The function finds how many objects should be precreated. Then allocates,
596 * prepares and schedules precreate RPC synchronously. Upon reply the function
597 * wakes up the threads waiting for the new objects on this target. If the
598 * target wasn't able to create all the objects requested, then the next
599 * precreate will be asking for fewer objects (i.e. slow precreate down).
601 * \param[in] env LU environment provided by the caller
602 * \param[in] d OSP device
604 * \retval 0 on success
605 * \retval negative negated errno on error
607 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
609 struct osp_thread_info *oti = osp_env_info(env);
610 struct ptlrpc_request *req;
611 struct obd_import *imp;
612 struct ost_body *body;
614 struct lu_fid *fid = &oti->osi_fid;
617 /* don't precreate new objects till OST healthy and has free space */
618 if (unlikely(d->opd_pre_status)) {
619 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
620 d->opd_obd->obd_name, d->opd_pre_status);
625 * if not connection/initialization is compeleted, ignore
627 imp = d->opd_obd->u.cli.cl_import;
630 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
633 req->rq_request_portal = OST_CREATE_PORTAL;
634 /* we should not resend create request - anyway we will have delorphan
635 * and kill these objects */
636 req->rq_no_delay = req->rq_no_resend = 1;
638 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
640 ptlrpc_request_free(req);
644 spin_lock(&d->opd_pre_lock);
645 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
646 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
647 grow = d->opd_pre_create_count;
648 spin_unlock(&d->opd_pre_lock);
650 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
653 *fid = d->opd_pre_last_created_fid;
654 rc = osp_precreate_fids(env, d, fid, &grow);
656 /* Current seq has been used up*/
657 GOTO(out_req, rc = -ENOSPC);
659 if (!osp_is_fid_client(d)) {
660 /* Non-FID client will always send seq 0 because of
662 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
666 fid_to_ostid(fid, &body->oa.o_oi);
667 body->oa.o_valid = OBD_MD_FLGROUP;
669 ptlrpc_request_set_replen(req);
671 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
674 rc = ptlrpc_queue_wait(req);
676 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
680 LASSERT(req->rq_transno == 0);
682 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
684 GOTO(out_req, rc = -EPROTO);
686 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
689 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
690 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
691 ": rc = %d\n", d->opd_obd->obd_name,
692 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
693 GOTO(out_req, rc = -ESTALE);
696 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
698 spin_lock(&d->opd_pre_lock);
700 /* the OST has not managed to create all the
701 * objects we asked for */
702 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
703 d->opd_pre_create_slow = 1;
705 /* the OST is able to keep up with the work,
706 * we could consider increasing create_count
707 * next time if needed */
708 d->opd_pre_create_slow = 0;
711 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
712 fid_to_ostid(fid, &body->oa.o_oi);
714 d->opd_pre_last_created_fid = *fid;
715 spin_unlock(&d->opd_pre_lock);
717 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
718 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
719 PFID(&d->opd_pre_last_created_fid));
721 /* now we can wakeup all users awaiting for objects */
722 osp_pre_update_status(d, rc);
723 wake_up(&d->opd_pre_user_waitq);
725 ptlrpc_req_finished(req);
730 * Get last precreated object from target (OST)
732 * Sends synchronous RPC to the target (OST) to learn the last precreated
733 * object. This later is used to remove all unused objects (cleanup orphan
734 * procedure). Also, the next object after one we got will be used as a
735 * starting point for the new precreates.
737 * \param[in] env LU environment provided by the caller
738 * \param[in] d OSP device
740 * \retval 0 on success
741 * \retval negative negated errno on error
743 static int osp_get_lastfid_from_ost(const struct lu_env *env,
744 struct osp_device *d)
746 struct ptlrpc_request *req = NULL;
747 struct obd_import *imp;
748 struct lu_fid *last_fid;
753 imp = d->opd_obd->u.cli.cl_import;
756 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
760 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
761 sizeof(KEY_LAST_FID));
763 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
765 ptlrpc_request_free(req);
769 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
770 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
772 req->rq_no_delay = req->rq_no_resend = 1;
773 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
774 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
776 ptlrpc_request_set_replen(req);
778 rc = ptlrpc_queue_wait(req);
780 /* bad-bad OST.. let sysadm sort this out */
781 if (rc == -ENOTSUPP) {
782 CERROR("%s: server does not support FID: rc = %d\n",
783 d->opd_obd->obd_name, -ENOTSUPP);
785 ptlrpc_set_import_active(imp, 0);
789 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
790 if (last_fid == NULL) {
791 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
792 GOTO(out, rc = -EPROTO);
795 if (!fid_is_sane(last_fid)) {
796 CERROR("%s: Got insane last_fid "DFID"\n",
797 d->opd_obd->obd_name, PFID(last_fid));
798 GOTO(out, rc = -EPROTO);
801 /* Only update the last used fid, if the OST has objects for
802 * this sequence, i.e. fid_oid > 0 */
803 if (fid_oid(last_fid) > 0)
804 d->opd_last_used_fid = *last_fid;
806 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
810 ptlrpc_req_finished(req);
815 * Cleanup orphans on OST
817 * This function is called in a contex of a dedicated thread handling
818 * all the precreation suff. The function waits till local recovery
819 * is complete, then identify all the unreferenced objects (orphans)
820 * using the highest ID referenced by a local and the highest object
821 * precreated by the target. The found range is a subject to removal
822 * using specially flagged RPC. During this process OSP is marked
823 * unavailable for new objects.
825 * \param[in] env LU environment provided by the caller
826 * \param[in] d OSP device
828 * \retval 0 on success
829 * \retval negative negated errno on error
831 static int osp_precreate_cleanup_orphans(struct lu_env *env,
832 struct osp_device *d)
834 struct osp_thread_info *osi = osp_env_info(env);
835 struct lu_fid *last_fid = &osi->osi_fid;
836 struct ptlrpc_request *req = NULL;
837 struct obd_import *imp;
838 struct ost_body *body;
839 int update_status = 0;
846 * wait for local recovery to finish, so we can cleanup orphans
847 * orphans are all objects since "last used" (assigned), but
848 * there might be objects reserved and in some cases they won't
849 * be used. we can't cleanup them till we're sure they won't be
850 * used. also can't we allow new reservations because they may
851 * end up getting orphans being cleaned up below. so we block
852 * new reservations and wait till all reserved objects either
855 spin_lock(&d->opd_pre_lock);
856 d->opd_pre_recovering = 1;
857 spin_unlock(&d->opd_pre_lock);
859 * The locking above makes sure the opd_pre_reserved check below will
860 * catch all osp_precreate_reserve() calls who find
861 * "!opd_pre_recovering".
863 wait_event_idle(d->opd_pre_waitq,
864 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
865 !d->opd_pre_task || d->opd_got_disconnected);
866 if (!d->opd_pre_task || d->opd_got_disconnected)
867 GOTO(out, rc = -EAGAIN);
869 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
870 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
872 *last_fid = d->opd_last_used_fid;
873 /* The OSP should already get the valid seq now */
874 LASSERT(!fid_is_zero(last_fid));
875 if (fid_oid(&d->opd_last_used_fid) < 2) {
876 /* lastfid looks strange... ask OST */
877 rc = osp_get_lastfid_from_ost(env, d);
882 imp = d->opd_obd->u.cli.cl_import;
885 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
887 GOTO(out, rc = -ENOMEM);
889 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
891 ptlrpc_request_free(req);
896 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
898 GOTO(out, rc = -EPROTO);
900 body->oa.o_flags = OBD_FL_DELORPHAN;
901 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
903 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
905 ptlrpc_request_set_replen(req);
907 /* Don't resend the delorphan req */
908 req->rq_no_resend = req->rq_no_delay = 1;
910 rc = ptlrpc_queue_wait(req);
916 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
918 GOTO(out, rc = -EPROTO);
921 * OST provides us with id new pool starts from in body->oa.o_id
923 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
925 spin_lock(&d->opd_pre_lock);
926 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
928 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
929 d->opd_pre_last_created_fid = d->opd_last_used_fid;
931 d->opd_pre_create_count = OST_MIN_PRECREATE;
932 d->opd_pre_last_created_fid = *last_fid;
935 * This empties the pre-creation pool and effectively blocks any new
938 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
939 LUSTRE_DATA_SEQ_MAX_WIDTH);
940 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
941 d->opd_pre_create_slow = 0;
942 spin_unlock(&d->opd_pre_lock);
944 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
945 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
946 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
949 ptlrpc_req_finished(req);
952 * If rc is zero, the pre-creation window should have been emptied.
953 * Since waking up the herd would be useless without pre-created
954 * objects, we defer the signal to osp_precreate_send() in that case.
958 CERROR("%s: cannot cleanup orphans: rc = %d\n",
959 d->opd_obd->obd_name, rc);
960 /* we can't proceed from here, OST seem to
961 * be in a bad shape, better to wait for
962 * a new instance of the server and repeat
963 * from the beginning. notify possible waiters
964 * this OSP isn't quite functional yet */
965 osp_pre_update_status(d, rc);
967 wake_up(&d->opd_pre_user_waitq);
970 spin_lock(&d->opd_pre_lock);
971 d->opd_pre_recovering = 0;
972 spin_unlock(&d->opd_pre_lock);
979 * Update precreate status using statfs data
981 * The function decides whether this OSP should be used for new objects.
982 * IOW, whether this OST is used up or has some free space. Cached statfs
983 * data is used to make this decision. If the latest result of statfs
984 * request (rc argument) is not success, then just mark OSP unavailable
987 * The new statfs data is passed in \a msfs and needs to be stored into
988 * opd_statfs, but only after the various flags in os_state are set, so
989 * that the new statfs data is not visible without appropriate flags set.
990 * As such, there is no need to clear the flags here, since this is called
991 * with new statfs data, and they should not be cleared if sent from OST.
993 * Add a bit of hysteresis so this flag isn't continually flapping, and
994 * ensure that new files don't get extremely fragmented due to only a
995 * small amount of available space in the filesystem. We want to set
996 * the ENOSPC/ENOINO flags unconditionally when there is less than the
997 * reserved size free, and still copy them from the old state when there
998 * is less than 2*reserved size free space or inodes.
1000 * \param[in] d OSP device
1001 * \param[in] msfs statfs data
1003 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1005 u32 old_state = d->opd_statfs.os_state;
1006 u32 reserved_ino_low = 32; /* could be tunable in the future */
1007 u32 reserved_ino_high = reserved_ino_low * 2;
1010 /* statfs structure not initialized yet */
1011 if (unlikely(!msfs->os_type))
1014 /* if the low and high watermarks have not been initialized yet */
1015 if (unlikely(d->opd_reserved_mb_high == 0 &&
1016 d->opd_reserved_mb_low == 0)) {
1017 /* Use ~0.1% by default to disable object allocation,
1018 * and ~0.2% to enable, size in MB, set both watermark
1020 spin_lock(&d->opd_pre_lock);
1021 if (d->opd_reserved_mb_high == 0 &&
1022 d->opd_reserved_mb_low == 0) {
1023 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1024 msfs->os_blocks) >> 20;
1025 if (d->opd_reserved_mb_low == 0)
1026 d->opd_reserved_mb_low = 1;
1027 d->opd_reserved_mb_high =
1028 (d->opd_reserved_mb_low << 1) + 1;
1030 spin_unlock(&d->opd_pre_lock);
1033 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1034 if (msfs->os_ffree < reserved_ino_low)
1035 msfs->os_state |= OS_STATFS_ENOINO;
1036 else if (msfs->os_ffree <= reserved_ino_high)
1037 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1038 /* else don't clear flags in new msfs->os_state sent from OST */
1041 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1042 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1043 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1044 msfs->os_files, msfs->os_ffree, msfs->os_state,
1046 if (available_mb < d->opd_reserved_mb_low)
1047 msfs->os_state |= OS_STATFS_ENOSPC;
1048 else if (available_mb <= d->opd_reserved_mb_high)
1049 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1050 /* else don't clear flags in new msfs->os_state sent from OST */
1052 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1053 d->opd_pre_status = -ENOSPC;
1054 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1055 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1056 d->opd_obd->obd_name, msfs->os_state,
1058 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1059 atomic_read(&d->opd_sync_changes),
1060 atomic_read(&d->opd_sync_rpcs_in_progress));
1061 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1062 d->opd_pre_status = 0;
1063 spin_lock(&d->opd_pre_lock);
1064 d->opd_pre_create_slow = 0;
1065 d->opd_pre_create_count = OST_MIN_PRECREATE;
1066 spin_unlock(&d->opd_pre_lock);
1067 wake_up(&d->opd_pre_waitq);
1070 "%s: available: state=%x: rc = %d\n",
1071 d->opd_obd->obd_name, msfs->os_state,
1074 /* we only get here if rc == 0 in the caller */
1075 d->opd_pre_status = 0;
1078 /* Object precreation skipped on OST if manually disabled */
1079 if (d->opd_pre_max_create_count == 0)
1080 msfs->os_state |= OS_STATFS_NOPRECREATE;
1081 /* else don't clear flags in new msfs->os_state sent from OST */
1083 /* copy only new statfs state to make it visible to MDS threads */
1084 if (&d->opd_statfs != msfs)
1085 d->opd_statfs = *msfs;
1089 * Initialize FID for precreation
1091 * For a just created new target, a new sequence should be taken.
1092 * The function checks there is no IDIF in use (if the target was
1093 * added with the older version of Lustre), then requests a new
1094 * sequence from FLDB using the regular protocol. Then this new
1095 * sequence is stored on a persisten storage synchronously to prevent
1096 * possible object leakage (for the detail see the description for
1097 * osp_precreate_rollover_new_seq()).
1099 * \param[in] osp OSP device
1101 * \retval 0 on success
1102 * \retval negative negated errno on error
1104 int osp_init_pre_fid(struct osp_device *osp)
1107 struct osp_thread_info *osi;
1108 struct lu_client_seq *cli_seq;
1109 struct lu_fid *last_fid;
1113 LASSERT(osp->opd_pre != NULL);
1115 /* Let's check if the current last_seq/fid is valid,
1116 * otherwise request new sequence from the controller */
1117 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1118 /* Non-MDT0 can only use normal sequence for
1120 if (fid_is_norm(&osp->opd_last_used_fid))
1123 /* Initially MDT0 will start with IDIF, after
1124 * that it will request new sequence from the
1126 if (fid_is_idif(&osp->opd_last_used_fid) ||
1127 fid_is_norm(&osp->opd_last_used_fid))
1131 if (!fid_is_zero(&osp->opd_last_used_fid))
1132 CWARN("%s: invalid last used fid "DFID
1133 ", try to get new sequence.\n",
1134 osp->opd_obd->obd_name,
1135 PFID(&osp->opd_last_used_fid));
1137 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1139 CERROR("%s: init env error: rc = %d\n",
1140 osp->opd_obd->obd_name, rc);
1144 osi = osp_env_info(&env);
1145 last_fid = &osi->osi_fid;
1147 /* For a freshed fs, it will allocate a new sequence first */
1148 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1149 cli_seq = osp->opd_obd->u.cli.cl_seq;
1150 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1152 CERROR("%s: alloc fid error: rc = %d\n",
1153 osp->opd_obd->obd_name, rc);
1157 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1159 last_fid->f_oid = 1;
1160 last_fid->f_ver = 0;
1162 spin_lock(&osp->opd_pre_lock);
1163 osp->opd_last_used_fid = *last_fid;
1164 osp->opd_pre_used_fid = *last_fid;
1165 osp->opd_pre_last_created_fid = *last_fid;
1166 spin_unlock(&osp->opd_pre_lock);
1167 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1169 CERROR("%s: write fid error: rc = %d\n",
1170 osp->opd_obd->obd_name, rc);
1179 struct osp_device *opta_dev;
1180 struct lu_env opta_env;
1181 struct completion *opta_started;
1184 * The core of precreate functionality
1186 * The function implements the main precreation loop. Basically it
1187 * involves connecting to the target, precerate FID initialization,
1188 * identifying and removing orphans, then serving precreation. As
1189 * part of the latter, the thread is responsible for statfs data
1190 * updates. The precreation is mostly driven by another threads
1191 * asking for new OST objects - those askers wake the thread when
1192 * the number of precreated objects reach low watermark.
1193 * After a disconnect, the sequence above repeats. This is keep going
1194 * until the thread is requested to stop.
1196 * \param[in] _arg private data the thread (OSP device to handle)
1198 * \retval 0 on success
1199 * \retval negative negated errno on error
1201 static int osp_precreate_thread(void *_args)
1203 struct opt_args *args = _args;
1204 struct osp_device *d = args->opta_dev;
1205 struct lu_env *env = &args->opta_env;
1210 complete(args->opta_started);
1211 while (!kthread_should_stop()) {
1213 * need to be connected to OST
1215 while (!kthread_should_stop()) {
1216 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1217 d->opd_imp_connected &&
1218 !d->opd_got_disconnected)
1220 wait_event_idle(d->opd_pre_waitq,
1221 kthread_should_stop() ||
1222 d->opd_new_connection);
1224 if (!d->opd_new_connection)
1227 d->opd_new_connection = 0;
1228 d->opd_got_disconnected = 0;
1232 if (kthread_should_stop())
1236 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1237 /* Sigh, fid client is not ready yet */
1238 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1241 /* Init fid for osp_precreate if necessary */
1242 rc = osp_init_pre_fid(d);
1244 class_export_put(d->opd_exp);
1245 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1246 CERROR("%s: init pre fid error: rc = %d\n",
1247 d->opd_obd->obd_name, rc);
1252 if (osp_statfs_update(env, d)) {
1253 if (wait_event_idle_timeout(d->opd_pre_waitq,
1254 kthread_should_stop(),
1255 cfs_time_seconds(5)) == 0)
1256 l_wait_event_abortable(
1258 kthread_should_stop());
1264 * Clean up orphans or recreate missing objects.
1266 rc = osp_precreate_cleanup_orphans(env, d);
1268 schedule_timeout_interruptible(cfs_time_seconds(1));
1274 * connected, can handle precreates now
1276 while (!kthread_should_stop()) {
1277 wait_event_idle(d->opd_pre_waitq,
1278 kthread_should_stop() ||
1279 osp_precreate_near_empty(env, d) ||
1280 osp_statfs_need_update(d) ||
1281 d->opd_got_disconnected);
1283 if (kthread_should_stop())
1286 /* something happened to the connection
1287 * have to start from the beginning */
1288 if (d->opd_got_disconnected)
1291 if (osp_statfs_need_update(d))
1292 if (osp_statfs_update(env, d))
1295 if (d->opd_pre == NULL)
1298 /* To avoid handling different seq in precreate/orphan
1299 * cleanup, it will hold precreate until current seq is
1301 if (unlikely(osp_precreate_end_seq(env, d) &&
1302 !osp_create_end_seq(env, d)))
1305 if (unlikely(osp_precreate_end_seq(env, d) &&
1306 osp_create_end_seq(env, d))) {
1307 LCONSOLE_INFO("%s:%#llx is used up."
1308 " Update to new seq\n",
1309 d->opd_obd->obd_name,
1310 fid_seq(&d->opd_pre_last_created_fid));
1311 rc = osp_precreate_rollover_new_seq(env, d);
1316 if (osp_precreate_near_empty(env, d)) {
1317 rc = osp_precreate_send(env, d);
1318 /* osp_precreate_send() sets opd_pre_status
1319 * in case of error, that prevent the using of
1321 if (rc < 0 && rc != -ENOSPC &&
1322 rc != -ETIMEDOUT && rc != -ENOTCONN)
1323 CERROR("%s: cannot precreate objects:"
1325 d->opd_obd->obd_name, rc);
1337 * Check when to stop to wait for precreate objects.
1339 * The caller wanting a new OST object can't wait undefinitely. The
1340 * function checks for few conditions including available new OST
1341 * objects, disconnected OST, lack of space with no pending destroys,
1342 * etc. IOW, it checks whether the current OSP state is good to keep
1343 * waiting or it's better to give up.
1345 * \param[in] env LU environment provided by the caller
1346 * \param[in] d OSP device
1348 * \retval 0 - keep waiting, 1 - no luck
1350 static int osp_precreate_ready_condition(const struct lu_env *env,
1351 struct osp_device *d)
1353 if (d->opd_pre_recovering)
1356 /* ready if got enough precreated objects */
1357 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1358 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1361 /* ready if OST reported no space and no destroys in progress */
1362 if (atomic_read(&d->opd_sync_changes) +
1363 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1364 d->opd_pre_status == -ENOSPC)
1367 /* Bail out I/O fails to OST */
1368 if (d->opd_pre_status != 0 &&
1369 d->opd_pre_status != -EAGAIN &&
1370 d->opd_pre_status != -ENODEV &&
1371 d->opd_pre_status != -ENOTCONN &&
1372 d->opd_pre_status != -ENOSPC) {
1374 if (d->opd_pre_status != -EIO)
1375 CERROR("%s: precreate failed opd_pre_status %d\n",
1376 d->opd_obd->obd_name, d->opd_pre_status);
1384 * Reserve object in precreate pool
1386 * When the caller wants to create a new object on this target (target
1387 * represented by the given OSP), it should declare this intention using
1388 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1389 * to reserve an object in the existing precreated pool or wait up to
1390 * obd_timeout for the available object to appear in the pool (a dedicated
1391 * thread will be doing real precreation in background). The object can be
1392 * consumed later with osp_precreate_get_fid() or be released with call to
1393 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1394 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1395 * If the space on the target is short and there is a pending object destroy,
1396 * then the function forces local commit to speedup space release (see
1397 * osp_sync.c for the details).
1399 * \param[in] env LU environment provided by the caller
1400 * \param[in] d OSP device
1402 * \retval 0 on success
1403 * \retval -ENOSPC when no space on OST
1404 * \retval -EAGAIN try later, slow precreation in progress
1405 * \retval -EIO when no access to OST
1407 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1409 time64_t expire = ktime_get_seconds() + obd_timeout;
1410 int precreated, rc, synced = 0;
1414 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1415 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1416 PFID(&d->opd_pre_used_fid));
1418 /* opd_pre_max_create_count 0 to not use specified OST. */
1419 if (d->opd_pre_max_create_count == 0)
1424 * - preallocation is done
1425 * - no free space expected soon
1426 * - can't connect to OST for too long (obd_timeout)
1427 * - OST can allocate fid sequence.
1429 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1430 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1433 * increase number of precreations
1435 precreated = osp_objs_precreated(env, d);
1436 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1437 d->opd_pre_create_slow == 0 &&
1438 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1439 spin_lock(&d->opd_pre_lock);
1440 d->opd_pre_create_slow = 1;
1441 d->opd_pre_create_count *= 2;
1442 spin_unlock(&d->opd_pre_lock);
1445 spin_lock(&d->opd_pre_lock);
1446 precreated = osp_objs_precreated(env, d);
1447 if (precreated > d->opd_pre_reserved &&
1448 !d->opd_pre_recovering) {
1449 d->opd_pre_reserved++;
1450 spin_unlock(&d->opd_pre_lock);
1453 /* XXX: don't wake up if precreation is in progress */
1454 if (osp_precreate_near_empty_nolock(env, d) &&
1455 !osp_precreate_end_seq_nolock(env, d))
1456 wake_up(&d->opd_pre_waitq);
1460 spin_unlock(&d->opd_pre_lock);
1463 * all precreated objects have been used and no-space
1464 * status leave us no chance to succeed very soon
1465 * but if there is destroy in progress, then we should
1466 * wait till that is done - some space might be released
1468 if (unlikely(rc == -ENOSPC)) {
1469 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1470 /* force local commit to release space */
1471 dt_commit_async(env, d->opd_storage);
1472 osp_sync_check_for_work(d);
1475 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1476 /* just wait till destroys are done
1477 * see wait_event_idle_timeout() below
1480 if (atomic_read(&d->opd_sync_changes) +
1481 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1482 /* no hope for free space */
1487 /* XXX: don't wake up if precreation is in progress */
1488 wake_up(&d->opd_pre_waitq);
1490 if (ktime_get_seconds() >= expire) {
1495 if (wait_event_idle_timeout(
1496 d->opd_pre_user_waitq,
1497 osp_precreate_ready_condition(env, d),
1498 cfs_time_seconds(obd_timeout)) == 0) {
1500 "%s: slow creates, last="DFID", next="DFID", "
1501 "reserved=%llu, sync_changes=%u, "
1502 "sync_rpcs_in_progress=%d, status=%d\n",
1503 d->opd_obd->obd_name,
1504 PFID(&d->opd_pre_last_created_fid),
1505 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1506 atomic_read(&d->opd_sync_changes),
1507 atomic_read(&d->opd_sync_rpcs_in_progress),
1516 * Get a FID from precreation pool
1518 * The function is a companion for osp_precreate_reserve() - it assigns
1519 * a specific FID from the precreate. The function should be called only
1520 * if the call to osp_precreate_reserve() was successful. The function
1521 * updates a local storage to remember the highest object ID referenced
1522 * by the node in the given sequence.
1524 * A very importan details: this is supposed to be called once the
1525 * transaction is started, so on-disk update will be atomic with the
1526 * data (like LOVEA) refering this object. Then the object won't be leaked:
1527 * either it's referenced by the committed transaction or it's a subject
1528 * to the orphan cleanup procedure.
1530 * \param[in] env LU environment provided by the caller
1531 * \param[in] d OSP device
1532 * \param[out] fid generated FID
1534 * \retval 0 on success
1535 * \retval negative negated errno on error
1537 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1540 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1541 /* grab next id from the pool */
1542 spin_lock(&d->opd_pre_lock);
1544 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1545 &d->opd_pre_last_created_fid) < 0,
1546 "next fid "DFID" last created fid "DFID"\n",
1547 PFID(&d->opd_pre_used_fid),
1548 PFID(&d->opd_pre_last_created_fid));
1551 * When sequence is used up, new one should be allocated in
1552 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1555 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1556 "next fid "DFID" last created fid "DFID"\n",
1557 PFID(&d->opd_pre_used_fid),
1558 PFID(&d->opd_pre_last_created_fid));
1559 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1560 if (fid_is_idif(pre_used_fid) &&
1561 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1562 pre_used_fid->f_seq++;
1564 d->opd_pre_used_fid.f_oid++;
1565 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1566 d->opd_pre_reserved--;
1568 * last_used_id must be changed along with getting new id otherwise
1569 * we might miscalculate gap causing object loss or leak
1571 osp_update_last_fid(d, fid);
1572 spin_unlock(&d->opd_pre_lock);
1575 * probably main thread suspended orphan cleanup till
1576 * all reservations are released, see comment in
1577 * osp_precreate_thread() just before orphan cleanup
1579 if (unlikely(d->opd_pre_reserved == 0 &&
1580 (d->opd_pre_recovering || d->opd_pre_status)))
1581 wake_up(&d->opd_pre_waitq);
1587 * Set size regular attribute on an object
1589 * When a striping is created late, it's possible that size is already
1590 * initialized on the file. Then the new striping should inherit size
1591 * from the file. The function sets size on the object using the regular
1592 * protocol (OST_PUNCH).
1593 * XXX: should be re-implemented using OUT ?
1595 * \param[in] env LU environment provided by the caller
1596 * \param[in] dt object
1597 * \param[in] size size to set.
1599 * \retval 0 on success
1600 * \retval negative negated errno on error
1602 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1605 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1606 struct ptlrpc_request *req = NULL;
1607 struct obd_import *imp;
1608 struct ost_body *body;
1609 struct obdo *oa = NULL;
1614 imp = d->opd_obd->u.cli.cl_import;
1617 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1621 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1623 ptlrpc_request_free(req);
1628 * XXX: decide how do we do here with resend
1629 * if we don't resend, then client may see wrong file size
1630 * if we do resend, then MDS thread can get stuck for quite long
1631 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1632 * (see LU-7975 and sanity/test_27F use cases)
1633 * but let's decide not to resend/delay this truncate request to OST
1634 * and allow Client to decide to resend, in a less agressive way from
1635 * after_reply(), by returning -EINPROGRESS instead of
1636 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1637 * end of this routine
1639 req->rq_no_resend = req->rq_no_delay = 1;
1641 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1642 ptlrpc_at_set_req_timeout(req);
1646 GOTO(out, rc = -ENOMEM);
1648 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1651 oa->o_blocks = OBD_OBJECT_EOF;
1652 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1653 OBD_MD_FLID | OBD_MD_FLGROUP;
1655 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1657 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1659 /* XXX: capa support? */
1660 /* osc_pack_capa(req, body, capa); */
1662 ptlrpc_request_set_replen(req);
1664 rc = ptlrpc_queue_wait(req);
1666 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1667 * since we have decided not to resend/delay, but this could
1668 * lead to wrong size to be seen at Client side and even process
1669 * trying to open to exit/fail if not itself handling -EAGAIN.
1670 * So it should be better to return -EINPROGRESS instead and
1671 * leave the decision to resend at Client side in after_reply()
1673 if (rc == -EWOULDBLOCK) {
1675 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1676 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1679 CERROR("can't punch object: %d\n", rc);
1683 ptlrpc_req_finished(req);
1690 * Initialize precreation functionality of OSP
1692 * Prepares all the internal structures and starts the precreate thread
1694 * \param[in] d OSP device
1696 * \retval 0 on success
1697 * \retval negative negated errno on error
1699 int osp_init_precreate(struct osp_device *d)
1703 OBD_ALLOC_PTR(d->opd_pre);
1704 if (d->opd_pre == NULL)
1707 /* initially precreation isn't ready */
1708 init_waitqueue_head(&d->opd_pre_user_waitq);
1709 d->opd_pre_status = -EAGAIN;
1710 fid_zero(&d->opd_pre_used_fid);
1711 d->opd_pre_used_fid.f_oid = 1;
1712 fid_zero(&d->opd_pre_last_created_fid);
1713 d->opd_pre_last_created_fid.f_oid = 1;
1715 d->opd_pre_reserved = 0;
1716 d->opd_got_disconnected = 1;
1717 d->opd_pre_create_slow = 0;
1718 d->opd_pre_create_count = OST_MIN_PRECREATE;
1719 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1720 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1721 d->opd_reserved_mb_high = 0;
1722 d->opd_reserved_mb_low = 0;
1728 * Finish precreate functionality of OSP
1731 * Asks all the activity (the thread, update timer) to stop, then
1732 * wait till that is done.
1734 * \param[in] d OSP device
1736 void osp_precreate_fini(struct osp_device *d)
1740 if (d->opd_pre == NULL)
1743 OBD_FREE_PTR(d->opd_pre);
1749 int osp_init_statfs(struct osp_device *d)
1751 struct task_struct *task;
1752 struct opt_args *args;
1753 DECLARE_COMPLETION_ONSTACK(started);
1758 spin_lock_init(&d->opd_pre_lock);
1759 init_waitqueue_head(&d->opd_pre_waitq);
1762 * Initialize statfs-related things
1764 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1765 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1766 1000 * NSEC_PER_SEC);
1767 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1769 ktime_to_ns(d->opd_statfs_fresh_till));
1770 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1771 (unsigned long)d, 0);
1773 if (d->opd_storage->dd_rdonly)
1776 OBD_ALLOC_PTR(args);
1780 args->opta_started = &started;
1781 rc = lu_env_init(&args->opta_env,
1782 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1784 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1791 * start thread handling precreation and statfs updates
1793 task = kthread_create(osp_precreate_thread, args,
1794 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1796 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1797 lu_env_fini(&args->opta_env);
1799 RETURN(PTR_ERR(task));
1801 d->opd_pre_task = task;
1802 wake_up_process(task);
1803 wait_for_completion(&started);
1808 void osp_statfs_fini(struct osp_device *d)
1810 struct task_struct *task = d->opd_pre_task;
1813 del_timer(&d->opd_statfs_timer);
1815 d->opd_pre_task = NULL;