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 inline bool osp_precreate_running(struct osp_device *d)
82 return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
85 static inline bool osp_precreate_stopped(struct osp_device *d)
87 return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
90 static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data)
92 struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
95 if (osp_precreate_running(d))
96 wake_up(&d->opd_pre_waitq);
99 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs);
102 * The function updates current precreation status if broken, and
103 * updates that cached statfs state if functional, then wakes up waiters.
104 * We don't clear opd_pre_status directly here, but rather leave this
105 * to osp_pre_update_msfs() to do if everything is OK so that we don't
106 * have a race to clear opd_pre_status and then set it to -ENOSPC again.
108 * \param[in] d OSP device
109 * \param[in] msfs statfs data
110 * \param[in] rc new precreate status for device \a d
112 static void osp_pre_update_status_msfs(struct osp_device *d,
113 struct obd_statfs *msfs, int rc)
116 d->opd_pre_status = rc;
118 osp_pre_update_msfs(d, msfs);
120 wake_up(&d->opd_pre_user_waitq);
123 /* Pass in the old statfs data in case the limits have changed */
124 void osp_pre_update_status(struct osp_device *d, int rc)
126 osp_pre_update_status_msfs(d, &d->opd_statfs, rc);
131 * RPC interpret callback for OST_STATFS RPC
133 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
134 * replied by the target. It's used to maintain statfs cache for the target.
135 * The function fills data from the reply if successful and schedules another
138 * \param[in] env LU environment provided by the caller
139 * \param[in] req RPC replied
140 * \param[in] aa callback data
141 * \param[in] rc RPC result
143 * \retval 0 on success
144 * \retval negative negated errno on error
146 static int osp_statfs_interpret(const struct lu_env *env,
147 struct ptlrpc_request *req, void *args, int rc)
149 union ptlrpc_async_args *aa = args;
150 struct obd_import *imp = req->rq_import;
151 struct obd_statfs *msfs;
152 struct osp_device *d;
157 aa = ptlrpc_req_async_args(aa, req);
158 d = aa->pointer_arg[0];
164 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
166 GOTO(out, rc = -EPROTO);
169 osp_pre_update_status_msfs(d, msfs, 0);
171 d->opd_statfs = *msfs;
173 /* schedule next update */
174 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
175 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
176 mod_timer(&d->opd_statfs_timer,
177 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
178 d->opd_statfs_update_in_progress = 0;
180 CDEBUG(D_CACHE, "updated statfs %p\n", d);
184 /* couldn't update statfs, try again with a small delay */
185 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
186 d->opd_statfs_update_in_progress = 0;
187 if (d->opd_pre != NULL && osp_precreate_running(d))
188 wake_up(&d->opd_pre_waitq);
190 if (req->rq_import_generation == imp->imp_generation)
191 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
192 d->opd_obd->obd_name, rc);
197 * Send OST_STATFS RPC
199 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
200 * Also disables scheduled updates as times OSP may need to refresh
201 * statfs data before expiration. The function doesn't block, instead
202 * an interpretation callback osp_statfs_interpret() is used.
204 * \param[in] d OSP device
206 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
208 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
209 struct ptlrpc_request *req;
210 struct obd_import *imp;
211 union ptlrpc_async_args *aa;
216 CDEBUG(D_CACHE, "going to update statfs\n");
218 imp = d->opd_obd->u.cli.cl_import;
221 req = ptlrpc_request_alloc(imp,
222 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
226 rc = ptlrpc_request_pack(req,
227 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
228 d->opd_pre ? OST_STATFS : MDS_STATFS);
230 ptlrpc_request_free(req);
233 ptlrpc_request_set_replen(req);
235 req->rq_request_portal = OST_CREATE_PORTAL;
236 ptlrpc_at_set_req_timeout(req);
238 req->rq_interpret_reply = osp_statfs_interpret;
239 aa = ptlrpc_req_async_args(aa, req);
240 aa->pointer_arg[0] = d;
243 * no updates till reply
245 del_timer(&d->opd_statfs_timer);
246 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
247 d->opd_statfs_update_in_progress = 1;
249 ptlrpcd_add_req(req);
251 /* we still want to sync changes if no new changes are coming */
252 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
255 if (atomic_read(&d->opd_sync_changes)) {
258 th = dt_trans_create(env, d->opd_storage);
260 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
263 rc = dt_trans_start_local(env, d->opd_storage, th);
265 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
266 d->opd_obd->obd_name,
267 atomic_read(&d->opd_sync_changes));
268 osp_sync_add_commit_cb_1s(env, d, th);
269 dt_trans_stop(env, d->opd_storage, th);
278 * Schedule an immediate update for statfs data
280 * If cached statfs data claim no free space, but OSP has got a request to
281 * destroy an object (so release some space probably), then we may need to
282 * refresh cached statfs data sooner than planned. The function checks there
283 * is no statfs update going and schedules immediate update if so.
284 * XXX: there might be a case where removed object(s) do not add free space (empty
285 * object). If the number of such deletions is high, then we can start to update
286 * statfs too often causing a RPC storm. some throttling is needed...
288 * \param[in] d OSP device where statfs data needs to be refreshed
290 void osp_statfs_need_now(struct osp_device *d)
292 if (!d->opd_statfs_update_in_progress) {
294 * if current status is -ENOSPC (lack of free space on OST)
295 * then we should poll OST immediately once object destroy
298 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
299 del_timer(&d->opd_statfs_timer);
300 wake_up(&d->opd_pre_waitq);
305 * Return number of precreated objects
307 * A simple helper to calculate the number of precreated objects on the device.
309 * \param[in] env LU environment provided by the caller
310 * \param[in] osp OSP device
312 * \retval the number of the precreated objects
314 static inline int osp_objs_precreated(const struct lu_env *env,
315 struct osp_device *osp)
317 return osp_fid_diff(&osp->opd_pre_last_created_fid,
318 &osp->opd_pre_used_fid);
322 * Check pool of precreated objects is nearly empty
324 * We should not wait till the pool of the precreated objects is exhausted,
325 * because then there will be a long period of OSP being unavailable for the
326 * new creations due to lenghty precreate RPC. Instead we ask for another
327 * precreation ahead and hopefully have it ready before the current pool is
328 * empty. Notice this function relies on an external locking.
330 * \param[in] env LU environment provided by the caller
331 * \param[in] d OSP device
333 * \retval 0 - current pool is good enough, 1 - time to precreate
335 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
336 struct osp_device *d)
338 int window = osp_objs_precreated(env, d);
340 /* don't consider new precreation till OST is healty and
342 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
343 (d->opd_pre_status == 0));
347 * Check pool of precreated objects
349 * This is protected version of osp_precreate_near_empty_nolock(), check that
352 * \param[in] env LU environment provided by the caller
353 * \param[in] d OSP device
355 * \retval 0 - current pool is good enough, 1 - time to precreate
357 static inline int osp_precreate_near_empty(const struct lu_env *env,
358 struct osp_device *d)
362 if (d->opd_pre == NULL)
365 /* XXX: do we really need locking here? */
366 spin_lock(&d->opd_pre_lock);
367 rc = osp_precreate_near_empty_nolock(env, d);
368 spin_unlock(&d->opd_pre_lock);
373 * Check given sequence is empty
375 * Returns a binary result whether the given sequence has some IDs left
376 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
377 * version of that function.
379 * \param[in] env LU environment provided by the caller
380 * \param[in] osp OSP device
382 * \retval 0 - current sequence has no IDs, 1 - otherwise
384 static inline int osp_create_end_seq(const struct lu_env *env,
385 struct osp_device *osp)
387 struct lu_fid *fid = &osp->opd_pre_used_fid;
390 spin_lock(&osp->opd_pre_lock);
391 rc = osp_fid_end_seq(env, fid);
392 spin_unlock(&osp->opd_pre_lock);
397 * Write FID into into last_oid/last_seq file
399 * The function stores the sequence and the in-sequence id into two dedicated
400 * files. The sync argument can be used to request synchronous commit, so the
401 * function won't return until the updates are committed.
403 * \param[in] env LU environment provided by the caller
404 * \param[in] osp OSP device
405 * \param[in] fid fid where sequence/id is taken
406 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
408 * \retval 0 on success
409 * \retval negative negated errno on error
411 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
412 struct lu_fid *fid, int sync)
414 struct osp_thread_info *oti = osp_env_info(env);
415 struct lu_buf *lb_oid = &oti->osi_lb;
416 struct lu_buf *lb_oseq = &oti->osi_lb2;
424 if (osp->opd_storage->dd_rdonly)
427 /* Note: through f_oid is only 32 bits, it will also write 64 bits
428 * for oid to keep compatibility with the previous version. */
430 osp_objid_buf_prep(lb_oid, &oid_off,
431 &oid, osp->opd_index);
433 osp_objseq_buf_prep(lb_oseq, &oseq_off,
434 &fid->f_seq, osp->opd_index);
436 th = dt_trans_create(env, osp->opd_storage);
441 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
442 lb_oid, oid_off, th);
446 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
447 lb_oseq, oseq_off, th);
451 rc = dt_trans_start_local(env, osp->opd_storage, th);
455 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
458 CERROR("%s: can not write to last seq file: rc = %d\n",
459 osp->opd_obd->obd_name, rc);
462 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
465 CERROR("%s: can not write to last seq file: rc = %d\n",
466 osp->opd_obd->obd_name, rc);
470 dt_trans_stop(env, osp->opd_storage, th);
475 * Switch to another sequence
477 * When a current sequence has no available IDs left, OSP has to switch to
478 * another new sequence. OSP requests it using the regular FLDB protocol
479 * and stores synchronously before that is used in precreated. This is needed
480 * to basically have the sequences referenced (not orphaned), otherwise it's
481 * possible that OST has some objects precreated and the clients have data
482 * written to it, but after MDT failover nobody refers those objects and OSP
483 * has no idea that the sequence need cleanup to be done.
484 * While this is very expensive operation, it's supposed to happen very very
485 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
487 * \param[in] env LU environment provided by the caller
488 * \param[in] osp OSP device
490 * \retval 0 on success
491 * \retval negative negated errno on error
493 static int osp_precreate_rollover_new_seq(struct lu_env *env,
494 struct osp_device *osp)
496 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
497 struct lu_fid *last_fid = &osp->opd_last_used_fid;
501 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
503 CERROR("%s: alloc fid error: rc = %d\n",
504 osp->opd_obd->obd_name, rc);
510 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
511 "fid "DFID", last_fid "DFID"\n", PFID(fid),
514 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
516 CERROR("%s: Can not update oid/seq file: rc = %d\n",
517 osp->opd_obd->obd_name, rc);
521 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
522 osp->opd_obd->obd_name, fid_seq(last_fid),
524 /* Update last_xxx to the new seq */
525 spin_lock(&osp->opd_pre_lock);
526 osp->opd_last_used_fid = *fid;
527 osp_fid_to_obdid(fid, &osp->opd_last_id);
528 osp->opd_gap_start_fid = *fid;
529 osp->opd_pre_used_fid = *fid;
530 osp->opd_pre_last_created_fid = *fid;
531 spin_unlock(&osp->opd_pre_lock);
537 * Find IDs available in current sequence
539 * The function calculates the highest possible ID and the number of IDs
540 * available in the current sequence OSP is using. The number is limited
541 * artifically by the caller (grow param) and the number of IDs available
542 * in the sequence by nature. The function doesn't require an external
545 * \param[in] env LU environment provided by the caller
546 * \param[in] osp OSP device
547 * \param[in] fid FID the caller wants to start with
548 * \param[in] grow how many the caller wants
549 * \param[out] fid the highest calculated FID
550 * \param[out] grow the number of available IDs calculated
552 * \retval 0 on success, 1 - the sequence is empty
554 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
555 struct lu_fid *fid, int *grow)
557 struct osp_thread_info *osi = osp_env_info(env);
561 if (fid_is_idif(fid)) {
562 struct lu_fid *last_fid;
563 struct ost_id *oi = &osi->osi_oi;
566 spin_lock(&osp->opd_pre_lock);
567 last_fid = &osp->opd_pre_last_created_fid;
568 fid_to_ostid(last_fid, oi);
569 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
570 *grow = end - ostid_id(oi);
571 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
572 spin_unlock(&osp->opd_pre_lock);
574 if (*grow == 0 || rc)
577 ostid_to_fid(fid, oi, osp->opd_index);
581 spin_lock(&osp->opd_pre_lock);
582 *fid = osp->opd_pre_last_created_fid;
584 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
585 *grow = end - fid->f_oid;
586 fid->f_oid += end - fid->f_oid;
587 spin_unlock(&osp->opd_pre_lock);
589 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
590 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
592 return *grow > 0 ? 0 : 1;
596 * Prepare and send precreate RPC
598 * The function finds how many objects should be precreated. Then allocates,
599 * prepares and schedules precreate RPC synchronously. Upon reply the function
600 * wakes up the threads waiting for the new objects on this target. If the
601 * target wasn't able to create all the objects requested, then the next
602 * precreate will be asking for fewer objects (i.e. slow precreate down).
604 * \param[in] env LU environment provided by the caller
605 * \param[in] d OSP device
607 * \retval 0 on success
608 * \retval negative negated errno on error
610 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
612 struct osp_thread_info *oti = osp_env_info(env);
613 struct ptlrpc_request *req;
614 struct obd_import *imp;
615 struct ost_body *body;
617 struct lu_fid *fid = &oti->osi_fid;
620 /* don't precreate new objects till OST healthy and has free space */
621 if (unlikely(d->opd_pre_status)) {
622 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
623 d->opd_obd->obd_name, d->opd_pre_status);
628 * if not connection/initialization is compeleted, ignore
630 imp = d->opd_obd->u.cli.cl_import;
633 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
636 req->rq_request_portal = OST_CREATE_PORTAL;
637 /* we should not resend create request - anyway we will have delorphan
638 * and kill these objects */
639 req->rq_no_delay = req->rq_no_resend = 1;
641 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
643 ptlrpc_request_free(req);
647 spin_lock(&d->opd_pre_lock);
648 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
649 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
650 grow = d->opd_pre_create_count;
651 spin_unlock(&d->opd_pre_lock);
653 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
656 *fid = d->opd_pre_last_created_fid;
657 rc = osp_precreate_fids(env, d, fid, &grow);
659 /* Current seq has been used up*/
660 GOTO(out_req, rc = -ENOSPC);
662 if (!osp_is_fid_client(d)) {
663 /* Non-FID client will always send seq 0 because of
665 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
669 fid_to_ostid(fid, &body->oa.o_oi);
670 body->oa.o_valid = OBD_MD_FLGROUP;
672 ptlrpc_request_set_replen(req);
674 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
677 rc = ptlrpc_queue_wait(req);
679 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
683 LASSERT(req->rq_transno == 0);
685 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
687 GOTO(out_req, rc = -EPROTO);
689 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
692 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
693 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
694 ": rc = %d\n", d->opd_obd->obd_name,
695 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
696 GOTO(out_req, rc = -ESTALE);
699 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
701 spin_lock(&d->opd_pre_lock);
703 /* the OST has not managed to create all the
704 * objects we asked for */
705 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
706 d->opd_pre_create_slow = 1;
708 /* the OST is able to keep up with the work,
709 * we could consider increasing create_count
710 * next time if needed */
711 d->opd_pre_create_slow = 0;
714 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
715 fid_to_ostid(fid, &body->oa.o_oi);
717 d->opd_pre_last_created_fid = *fid;
718 spin_unlock(&d->opd_pre_lock);
720 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
721 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
722 PFID(&d->opd_pre_last_created_fid));
724 /* now we can wakeup all users awaiting for objects */
725 osp_pre_update_status(d, rc);
726 wake_up(&d->opd_pre_user_waitq);
728 ptlrpc_req_finished(req);
733 * Get last precreated object from target (OST)
735 * Sends synchronous RPC to the target (OST) to learn the last precreated
736 * object. This later is used to remove all unused objects (cleanup orphan
737 * procedure). Also, the next object after one we got will be used as a
738 * starting point for the new precreates.
740 * \param[in] env LU environment provided by the caller
741 * \param[in] d OSP device
743 * \retval 0 on success
744 * \retval negative negated errno on error
746 static int osp_get_lastfid_from_ost(const struct lu_env *env,
747 struct osp_device *d)
749 struct ptlrpc_request *req = NULL;
750 struct obd_import *imp;
751 struct lu_fid *last_fid;
756 imp = d->opd_obd->u.cli.cl_import;
759 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
763 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
764 sizeof(KEY_LAST_FID));
766 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
768 ptlrpc_request_free(req);
772 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
773 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
775 req->rq_no_delay = req->rq_no_resend = 1;
776 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
777 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
779 ptlrpc_request_set_replen(req);
781 rc = ptlrpc_queue_wait(req);
783 /* bad-bad OST.. let sysadm sort this out */
784 if (rc == -ENOTSUPP) {
785 CERROR("%s: server does not support FID: rc = %d\n",
786 d->opd_obd->obd_name, -ENOTSUPP);
788 ptlrpc_set_import_active(imp, 0);
792 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
793 if (last_fid == NULL) {
794 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
795 GOTO(out, rc = -EPROTO);
798 if (!fid_is_sane(last_fid)) {
799 CERROR("%s: Got insane last_fid "DFID"\n",
800 d->opd_obd->obd_name, PFID(last_fid));
801 GOTO(out, rc = -EPROTO);
804 /* Only update the last used fid, if the OST has objects for
805 * this sequence, i.e. fid_oid > 0 */
806 if (fid_oid(last_fid) > 0)
807 d->opd_last_used_fid = *last_fid;
809 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
813 ptlrpc_req_finished(req);
818 * Cleanup orphans on OST
820 * This function is called in a contex of a dedicated thread handling
821 * all the precreation suff. The function waits till local recovery
822 * is complete, then identify all the unreferenced objects (orphans)
823 * using the highest ID referenced by a local and the highest object
824 * precreated by the target. The found range is a subject to removal
825 * using specially flagged RPC. During this process OSP is marked
826 * unavailable for new objects.
828 * \param[in] env LU environment provided by the caller
829 * \param[in] d OSP device
831 * \retval 0 on success
832 * \retval negative negated errno on error
834 static int osp_precreate_cleanup_orphans(struct lu_env *env,
835 struct osp_device *d)
837 struct osp_thread_info *osi = osp_env_info(env);
838 struct lu_fid *last_fid = &osi->osi_fid;
839 struct ptlrpc_request *req = NULL;
840 struct obd_import *imp;
841 struct ost_body *body;
842 int update_status = 0;
849 * wait for local recovery to finish, so we can cleanup orphans
850 * orphans are all objects since "last used" (assigned), but
851 * there might be objects reserved and in some cases they won't
852 * be used. we can't cleanup them till we're sure they won't be
853 * used. also can't we allow new reservations because they may
854 * end up getting orphans being cleaned up below. so we block
855 * new reservations and wait till all reserved objects either
858 spin_lock(&d->opd_pre_lock);
859 d->opd_pre_recovering = 1;
860 spin_unlock(&d->opd_pre_lock);
862 * The locking above makes sure the opd_pre_reserved check below will
863 * catch all osp_precreate_reserve() calls who find
864 * "!opd_pre_recovering".
866 wait_event_idle(d->opd_pre_waitq,
867 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
868 !osp_precreate_running(d) || d->opd_got_disconnected);
869 if (!osp_precreate_running(d) || d->opd_got_disconnected)
870 GOTO(out, rc = -EAGAIN);
872 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
873 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
875 *last_fid = d->opd_last_used_fid;
876 /* The OSP should already get the valid seq now */
877 LASSERT(!fid_is_zero(last_fid));
878 if (fid_oid(&d->opd_last_used_fid) < 2) {
879 /* lastfid looks strange... ask OST */
880 rc = osp_get_lastfid_from_ost(env, d);
885 imp = d->opd_obd->u.cli.cl_import;
888 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
890 GOTO(out, rc = -ENOMEM);
892 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
894 ptlrpc_request_free(req);
899 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
901 GOTO(out, rc = -EPROTO);
903 body->oa.o_flags = OBD_FL_DELORPHAN;
904 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
906 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
908 ptlrpc_request_set_replen(req);
910 /* Don't resend the delorphan req */
911 req->rq_no_resend = req->rq_no_delay = 1;
913 rc = ptlrpc_queue_wait(req);
919 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
921 GOTO(out, rc = -EPROTO);
924 * OST provides us with id new pool starts from in body->oa.o_id
926 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
928 spin_lock(&d->opd_pre_lock);
929 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
931 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
932 d->opd_pre_last_created_fid = d->opd_last_used_fid;
934 d->opd_pre_create_count = OST_MIN_PRECREATE;
935 d->opd_pre_last_created_fid = *last_fid;
938 * This empties the pre-creation pool and effectively blocks any new
941 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
942 LUSTRE_DATA_SEQ_MAX_WIDTH);
943 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
944 d->opd_pre_create_slow = 0;
945 spin_unlock(&d->opd_pre_lock);
947 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
948 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
949 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
952 ptlrpc_req_finished(req);
955 * If rc is zero, the pre-creation window should have been emptied.
956 * Since waking up the herd would be useless without pre-created
957 * objects, we defer the signal to osp_precreate_send() in that case.
961 CERROR("%s: cannot cleanup orphans: rc = %d\n",
962 d->opd_obd->obd_name, rc);
963 /* we can't proceed from here, OST seem to
964 * be in a bad shape, better to wait for
965 * a new instance of the server and repeat
966 * from the beginning. notify possible waiters
967 * this OSP isn't quite functional yet */
968 osp_pre_update_status(d, rc);
970 wake_up(&d->opd_pre_user_waitq);
973 spin_lock(&d->opd_pre_lock);
974 d->opd_pre_recovering = 0;
975 spin_unlock(&d->opd_pre_lock);
982 * Update precreate status using statfs data
984 * The function decides whether this OSP should be used for new objects.
985 * IOW, whether this OST is used up or has some free space. Cached statfs
986 * data is used to make this decision. If the latest result of statfs
987 * request (rc argument) is not success, then just mark OSP unavailable
990 * The new statfs data is passed in \a msfs and needs to be stored into
991 * opd_statfs, but only after the various flags in os_state are set, so
992 * that the new statfs data is not visible without appropriate flags set.
993 * As such, there is no need to clear the flags here, since this is called
994 * with new statfs data, and they should not be cleared if sent from OST.
996 * Add a bit of hysteresis so this flag isn't continually flapping, and
997 * ensure that new files don't get extremely fragmented due to only a
998 * small amount of available space in the filesystem. We want to set
999 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1000 * reserved size free, and still copy them from the old state when there
1001 * is less than 2*reserved size free space or inodes.
1003 * \param[in] d OSP device
1004 * \param[in] msfs statfs data
1006 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1008 u32 old_state = d->opd_statfs.os_state;
1009 u32 reserved_ino_low = 32; /* could be tunable in the future */
1010 u32 reserved_ino_high = reserved_ino_low * 2;
1013 /* statfs structure not initialized yet */
1014 if (unlikely(!msfs->os_type))
1017 /* if the low and high watermarks have not been initialized yet */
1018 if (unlikely(d->opd_reserved_mb_high == 0 &&
1019 d->opd_reserved_mb_low == 0)) {
1020 /* Use ~0.1% by default to disable object allocation,
1021 * and ~0.2% to enable, size in MB, set both watermark
1023 spin_lock(&d->opd_pre_lock);
1024 if (d->opd_reserved_mb_high == 0 &&
1025 d->opd_reserved_mb_low == 0) {
1026 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1027 msfs->os_blocks) >> 20;
1028 if (d->opd_reserved_mb_low == 0)
1029 d->opd_reserved_mb_low = 1;
1030 d->opd_reserved_mb_high =
1031 (d->opd_reserved_mb_low << 1) + 1;
1033 spin_unlock(&d->opd_pre_lock);
1036 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1037 if (msfs->os_ffree < reserved_ino_low)
1038 msfs->os_state |= OS_STATE_ENOINO;
1039 else if (msfs->os_ffree <= reserved_ino_high)
1040 msfs->os_state |= old_state & OS_STATE_ENOINO;
1041 /* else don't clear flags in new msfs->os_state sent from OST */
1044 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1045 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1046 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1047 msfs->os_files, msfs->os_ffree, msfs->os_state,
1049 if (available_mb < d->opd_reserved_mb_low)
1050 msfs->os_state |= OS_STATE_ENOSPC;
1051 else if (available_mb <= d->opd_reserved_mb_high)
1052 msfs->os_state |= old_state & OS_STATE_ENOSPC;
1053 /* else don't clear flags in new msfs->os_state sent from OST */
1055 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1056 d->opd_pre_status = -ENOSPC;
1057 if (!(old_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)))
1058 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1059 d->opd_obd->obd_name, msfs->os_state,
1061 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1062 atomic_read(&d->opd_sync_changes),
1063 atomic_read(&d->opd_sync_rpcs_in_progress));
1064 } else if (old_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1065 d->opd_pre_status = 0;
1066 spin_lock(&d->opd_pre_lock);
1067 d->opd_pre_create_slow = 0;
1068 d->opd_pre_create_count = OST_MIN_PRECREATE;
1069 spin_unlock(&d->opd_pre_lock);
1070 wake_up(&d->opd_pre_waitq);
1073 "%s: available: state=%x: rc = %d\n",
1074 d->opd_obd->obd_name, msfs->os_state,
1077 /* we only get here if rc == 0 in the caller */
1078 d->opd_pre_status = 0;
1081 /* Object precreation skipped on OST if manually disabled */
1082 if (d->opd_pre_max_create_count == 0)
1083 msfs->os_state |= OS_STATE_NOPRECREATE;
1084 /* else don't clear flags in new msfs->os_state sent from OST */
1086 /* copy only new statfs state to make it visible to MDS threads */
1087 if (&d->opd_statfs != msfs)
1088 d->opd_statfs = *msfs;
1092 * Initialize FID for precreation
1094 * For a just created new target, a new sequence should be taken.
1095 * The function checks there is no IDIF in use (if the target was
1096 * added with the older version of Lustre), then requests a new
1097 * sequence from FLDB using the regular protocol. Then this new
1098 * sequence is stored on a persisten storage synchronously to prevent
1099 * possible object leakage (for the detail see the description for
1100 * osp_precreate_rollover_new_seq()).
1102 * \param[in] osp OSP device
1104 * \retval 0 on success
1105 * \retval negative negated errno on error
1107 int osp_init_pre_fid(struct osp_device *osp)
1110 struct osp_thread_info *osi;
1111 struct lu_client_seq *cli_seq;
1112 struct lu_fid *last_fid;
1116 LASSERT(osp->opd_pre != NULL);
1118 /* Let's check if the current last_seq/fid is valid,
1119 * otherwise request new sequence from the controller */
1120 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1121 /* Non-MDT0 can only use normal sequence for
1123 if (fid_is_norm(&osp->opd_last_used_fid))
1126 /* Initially MDT0 will start with IDIF, after
1127 * that it will request new sequence from the
1129 if (fid_is_idif(&osp->opd_last_used_fid) ||
1130 fid_is_norm(&osp->opd_last_used_fid))
1134 if (!fid_is_zero(&osp->opd_last_used_fid))
1135 CWARN("%s: invalid last used fid "DFID
1136 ", try to get new sequence.\n",
1137 osp->opd_obd->obd_name,
1138 PFID(&osp->opd_last_used_fid));
1140 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1142 CERROR("%s: init env error: rc = %d\n",
1143 osp->opd_obd->obd_name, rc);
1147 osi = osp_env_info(&env);
1148 last_fid = &osi->osi_fid;
1150 /* For a freshed fs, it will allocate a new sequence first */
1151 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1152 cli_seq = osp->opd_obd->u.cli.cl_seq;
1153 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1155 CERROR("%s: alloc fid error: rc = %d\n",
1156 osp->opd_obd->obd_name, rc);
1160 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1162 last_fid->f_oid = 1;
1163 last_fid->f_ver = 0;
1165 spin_lock(&osp->opd_pre_lock);
1166 osp->opd_last_used_fid = *last_fid;
1167 osp->opd_pre_used_fid = *last_fid;
1168 osp->opd_pre_last_created_fid = *last_fid;
1169 spin_unlock(&osp->opd_pre_lock);
1170 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1172 CERROR("%s: write fid error: rc = %d\n",
1173 osp->opd_obd->obd_name, rc);
1182 * The core of precreate functionality
1184 * The function implements the main precreation loop. Basically it
1185 * involves connecting to the target, precerate FID initialization,
1186 * identifying and removing orphans, then serving precreation. As
1187 * part of the latter, the thread is responsible for statfs data
1188 * updates. The precreation is mostly driven by another threads
1189 * asking for new OST objects - those askers wake the thread when
1190 * the number of precreated objects reach low watermark.
1191 * After a disconnect, the sequence above repeats. This is keep going
1192 * until the thread is requested to stop.
1194 * \param[in] _arg private data the thread (OSP device to handle)
1196 * \retval 0 on success
1197 * \retval negative negated errno on error
1199 static int osp_precreate_thread(void *_arg)
1201 struct osp_device *d = _arg;
1202 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1203 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1204 back_to_sleep, NULL);
1210 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1212 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1215 spin_lock(&d->opd_pre_lock);
1216 thread->t_flags = SVC_STOPPED;
1217 spin_unlock(&d->opd_pre_lock);
1218 wake_up(&thread->t_ctl_waitq);
1223 spin_lock(&d->opd_pre_lock);
1224 thread->t_flags = SVC_RUNNING;
1225 spin_unlock(&d->opd_pre_lock);
1226 wake_up(&thread->t_ctl_waitq);
1228 while (osp_precreate_running(d)) {
1230 * need to be connected to OST
1232 while (osp_precreate_running(d)) {
1233 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1234 d->opd_imp_connected &&
1235 !d->opd_got_disconnected)
1237 wait_event_idle(d->opd_pre_waitq,
1238 !osp_precreate_running(d) ||
1239 d->opd_new_connection);
1241 if (!d->opd_new_connection)
1244 d->opd_new_connection = 0;
1245 d->opd_got_disconnected = 0;
1249 if (!osp_precreate_running(d))
1253 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1254 /* Sigh, fid client is not ready yet */
1255 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1258 /* Init fid for osp_precreate if necessary */
1259 rc = osp_init_pre_fid(d);
1261 class_export_put(d->opd_exp);
1262 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1263 CERROR("%s: init pre fid error: rc = %d\n",
1264 d->opd_obd->obd_name, rc);
1269 if (osp_statfs_update(&env, d)) {
1270 l_wait_event(d->opd_pre_waitq,
1271 !osp_precreate_running(d), &lwi2);
1277 * Clean up orphans or recreate missing objects.
1279 rc = osp_precreate_cleanup_orphans(&env, d);
1281 schedule_timeout_interruptible(cfs_time_seconds(1));
1287 * connected, can handle precreates now
1289 while (osp_precreate_running(d)) {
1290 wait_event_idle(d->opd_pre_waitq,
1291 !osp_precreate_running(d) ||
1292 osp_precreate_near_empty(&env, d) ||
1293 osp_statfs_need_update(d) ||
1294 d->opd_got_disconnected);
1296 if (!osp_precreate_running(d))
1299 /* something happened to the connection
1300 * have to start from the beginning */
1301 if (d->opd_got_disconnected)
1304 if (osp_statfs_need_update(d))
1305 if (osp_statfs_update(&env, d))
1308 if (d->opd_pre == NULL)
1311 /* To avoid handling different seq in precreate/orphan
1312 * cleanup, it will hold precreate until current seq is
1314 if (unlikely(osp_precreate_end_seq(&env, d) &&
1315 !osp_create_end_seq(&env, d)))
1318 if (unlikely(osp_precreate_end_seq(&env, d) &&
1319 osp_create_end_seq(&env, d))) {
1320 LCONSOLE_INFO("%s:%#llx is used up."
1321 " Update to new seq\n",
1322 d->opd_obd->obd_name,
1323 fid_seq(&d->opd_pre_last_created_fid));
1324 rc = osp_precreate_rollover_new_seq(&env, d);
1329 if (osp_precreate_near_empty(&env, d)) {
1330 rc = osp_precreate_send(&env, d);
1331 /* osp_precreate_send() sets opd_pre_status
1332 * in case of error, that prevent the using of
1334 if (rc < 0 && rc != -ENOSPC &&
1335 rc != -ETIMEDOUT && rc != -ENOTCONN)
1336 CERROR("%s: cannot precreate objects:"
1338 d->opd_obd->obd_name, rc);
1343 thread->t_flags = SVC_STOPPED;
1345 wake_up(&thread->t_ctl_waitq);
1351 * Check when to stop to wait for precreate objects.
1353 * The caller wanting a new OST object can't wait undefinitely. The
1354 * function checks for few conditions including available new OST
1355 * objects, disconnected OST, lack of space with no pending destroys,
1356 * etc. IOW, it checks whether the current OSP state is good to keep
1357 * waiting or it's better to give up.
1359 * \param[in] env LU environment provided by the caller
1360 * \param[in] d OSP device
1362 * \retval 0 - keep waiting, 1 - no luck
1364 static int osp_precreate_ready_condition(const struct lu_env *env,
1365 struct osp_device *d)
1367 if (d->opd_pre_recovering)
1370 /* ready if got enough precreated objects */
1371 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1372 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1375 /* ready if OST reported no space and no destroys in progress */
1376 if (atomic_read(&d->opd_sync_changes) +
1377 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1378 d->opd_pre_status == -ENOSPC)
1381 /* Bail out I/O fails to OST */
1382 if (d->opd_pre_status != 0 &&
1383 d->opd_pre_status != -EAGAIN &&
1384 d->opd_pre_status != -ENODEV &&
1385 d->opd_pre_status != -ENOTCONN &&
1386 d->opd_pre_status != -ENOSPC) {
1388 if (d->opd_pre_status != -EIO)
1389 CERROR("%s: precreate failed opd_pre_status %d\n",
1390 d->opd_obd->obd_name, d->opd_pre_status);
1397 static int osp_precreate_timeout_condition(void *data)
1399 struct osp_device *d = data;
1401 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1402 "reserved=%llu, sync_changes=%u, "
1403 "sync_rpcs_in_progress=%d, status=%d\n",
1404 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1405 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1406 atomic_read(&d->opd_sync_changes),
1407 atomic_read(&d->opd_sync_rpcs_in_progress),
1414 * Reserve object in precreate pool
1416 * When the caller wants to create a new object on this target (target
1417 * represented by the given OSP), it should declare this intention using
1418 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1419 * to reserve an object in the existing precreated pool or wait up to
1420 * obd_timeout for the available object to appear in the pool (a dedicated
1421 * thread will be doing real precreation in background). The object can be
1422 * consumed later with osp_precreate_get_fid() or be released with call to
1423 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1424 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1425 * If the space on the target is short and there is a pending object destroy,
1426 * then the function forces local commit to speedup space release (see
1427 * osp_sync.c for the details).
1429 * \param[in] env LU environment provided by the caller
1430 * \param[in] d OSP device
1432 * \retval 0 on success
1433 * \retval -ENOSPC when no space on OST
1434 * \retval -EAGAIN try later, slow precreation in progress
1435 * \retval -EIO when no access to OST
1437 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1439 time64_t expire = ktime_get_seconds() + obd_timeout;
1440 struct l_wait_info lwi;
1441 int precreated, rc, synced = 0;
1445 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1446 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1447 PFID(&d->opd_pre_used_fid));
1449 /* opd_pre_max_create_count 0 to not use specified OST. */
1450 if (d->opd_pre_max_create_count == 0)
1455 * - preallocation is done
1456 * - no free space expected soon
1457 * - can't connect to OST for too long (obd_timeout)
1458 * - OST can allocate fid sequence.
1460 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1461 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1464 * increase number of precreations
1466 precreated = osp_objs_precreated(env, d);
1467 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1468 d->opd_pre_create_slow == 0 &&
1469 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1470 spin_lock(&d->opd_pre_lock);
1471 d->opd_pre_create_slow = 1;
1472 d->opd_pre_create_count *= 2;
1473 spin_unlock(&d->opd_pre_lock);
1476 spin_lock(&d->opd_pre_lock);
1477 precreated = osp_objs_precreated(env, d);
1478 if (precreated > d->opd_pre_reserved &&
1479 !d->opd_pre_recovering) {
1480 d->opd_pre_reserved++;
1481 spin_unlock(&d->opd_pre_lock);
1484 /* XXX: don't wake up if precreation is in progress */
1485 if (osp_precreate_near_empty_nolock(env, d) &&
1486 !osp_precreate_end_seq_nolock(env, d))
1487 wake_up(&d->opd_pre_waitq);
1491 spin_unlock(&d->opd_pre_lock);
1494 * all precreated objects have been used and no-space
1495 * status leave us no chance to succeed very soon
1496 * but if there is destroy in progress, then we should
1497 * wait till that is done - some space might be released
1499 if (unlikely(rc == -ENOSPC)) {
1500 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1501 /* force local commit to release space */
1502 dt_commit_async(env, d->opd_storage);
1503 osp_sync_check_for_work(d);
1506 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1507 /* just wait till destroys are done */
1508 /* see l_wait_even() few lines below */
1510 if (atomic_read(&d->opd_sync_changes) +
1511 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1512 /* no hope for free space */
1517 /* XXX: don't wake up if precreation is in progress */
1518 wake_up(&d->opd_pre_waitq);
1520 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
1521 osp_precreate_timeout_condition, d);
1522 if (ktime_get_seconds() >= expire) {
1527 l_wait_event(d->opd_pre_user_waitq,
1528 osp_precreate_ready_condition(env, d), &lwi);
1535 * Get a FID from precreation pool
1537 * The function is a companion for osp_precreate_reserve() - it assigns
1538 * a specific FID from the precreate. The function should be called only
1539 * if the call to osp_precreate_reserve() was successful. The function
1540 * updates a local storage to remember the highest object ID referenced
1541 * by the node in the given sequence.
1543 * A very importan details: this is supposed to be called once the
1544 * transaction is started, so on-disk update will be atomic with the
1545 * data (like LOVEA) refering this object. Then the object won't be leaked:
1546 * either it's referenced by the committed transaction or it's a subject
1547 * to the orphan cleanup procedure.
1549 * \param[in] env LU environment provided by the caller
1550 * \param[in] d OSP device
1551 * \param[out] fid generated FID
1553 * \retval 0 on success
1554 * \retval negative negated errno on error
1556 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1559 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1560 /* grab next id from the pool */
1561 spin_lock(&d->opd_pre_lock);
1563 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1564 &d->opd_pre_last_created_fid) < 0,
1565 "next fid "DFID" last created fid "DFID"\n",
1566 PFID(&d->opd_pre_used_fid),
1567 PFID(&d->opd_pre_last_created_fid));
1570 * When sequence is used up, new one should be allocated in
1571 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1574 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1575 "next fid "DFID" last created fid "DFID"\n",
1576 PFID(&d->opd_pre_used_fid),
1577 PFID(&d->opd_pre_last_created_fid));
1578 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1579 if (fid_is_idif(pre_used_fid) &&
1580 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1581 pre_used_fid->f_seq++;
1583 d->opd_pre_used_fid.f_oid++;
1584 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1585 d->opd_pre_reserved--;
1587 * last_used_id must be changed along with getting new id otherwise
1588 * we might miscalculate gap causing object loss or leak
1590 osp_update_last_fid(d, fid);
1591 spin_unlock(&d->opd_pre_lock);
1594 * probably main thread suspended orphan cleanup till
1595 * all reservations are released, see comment in
1596 * osp_precreate_thread() just before orphan cleanup
1598 if (unlikely(d->opd_pre_reserved == 0 &&
1599 (d->opd_pre_recovering || d->opd_pre_status)))
1600 wake_up(&d->opd_pre_waitq);
1606 * Set size regular attribute on an object
1608 * When a striping is created late, it's possible that size is already
1609 * initialized on the file. Then the new striping should inherit size
1610 * from the file. The function sets size on the object using the regular
1611 * protocol (OST_PUNCH).
1612 * XXX: should be re-implemented using OUT ?
1614 * \param[in] env LU environment provided by the caller
1615 * \param[in] dt object
1616 * \param[in] size size to set.
1618 * \retval 0 on success
1619 * \retval negative negated errno on error
1621 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1624 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1625 struct ptlrpc_request *req = NULL;
1626 struct obd_import *imp;
1627 struct ost_body *body;
1628 struct obdo *oa = NULL;
1633 imp = d->opd_obd->u.cli.cl_import;
1636 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1640 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1642 ptlrpc_request_free(req);
1647 * XXX: decide how do we do here with resend
1648 * if we don't resend, then client may see wrong file size
1649 * if we do resend, then MDS thread can get stuck for quite long
1650 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1651 * (see LU-7975 and sanity/test_27F use cases)
1652 * but let's decide not to resend/delay this truncate request to OST
1653 * and allow Client to decide to resend, in a less agressive way from
1654 * after_reply(), by returning -EINPROGRESS instead of
1655 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1656 * end of this routine
1658 req->rq_no_resend = req->rq_no_delay = 1;
1660 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1661 ptlrpc_at_set_req_timeout(req);
1665 GOTO(out, rc = -ENOMEM);
1667 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1670 oa->o_blocks = OBD_OBJECT_EOF;
1671 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1672 OBD_MD_FLID | OBD_MD_FLGROUP;
1674 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1676 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1678 /* XXX: capa support? */
1679 /* osc_pack_capa(req, body, capa); */
1681 ptlrpc_request_set_replen(req);
1683 rc = ptlrpc_queue_wait(req);
1685 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1686 * since we have decided not to resend/delay, but this could
1687 * lead to wrong size to be seen at Client side and even process
1688 * trying to open to exit/fail if not itself handling -EAGAIN.
1689 * So it should be better to return -EINPROGRESS instead and
1690 * leave the decision to resend at Client side in after_reply()
1692 if (rc == -EWOULDBLOCK) {
1694 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1695 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1698 CERROR("can't punch object: %d\n", rc);
1702 ptlrpc_req_finished(req);
1709 * Initialize precreation functionality of OSP
1711 * Prepares all the internal structures and starts the precreate thread
1713 * \param[in] d OSP device
1715 * \retval 0 on success
1716 * \retval negative negated errno on error
1718 int osp_init_precreate(struct osp_device *d)
1722 OBD_ALLOC_PTR(d->opd_pre);
1723 if (d->opd_pre == NULL)
1726 /* initially precreation isn't ready */
1727 init_waitqueue_head(&d->opd_pre_user_waitq);
1728 d->opd_pre_status = -EAGAIN;
1729 fid_zero(&d->opd_pre_used_fid);
1730 d->opd_pre_used_fid.f_oid = 1;
1731 fid_zero(&d->opd_pre_last_created_fid);
1732 d->opd_pre_last_created_fid.f_oid = 1;
1734 d->opd_pre_reserved = 0;
1735 d->opd_got_disconnected = 1;
1736 d->opd_pre_create_slow = 0;
1737 d->opd_pre_create_count = OST_MIN_PRECREATE;
1738 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1739 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1740 d->opd_reserved_mb_high = 0;
1741 d->opd_reserved_mb_low = 0;
1747 * Finish precreate functionality of OSP
1750 * Asks all the activity (the thread, update timer) to stop, then
1751 * wait till that is done.
1753 * \param[in] d OSP device
1755 void osp_precreate_fini(struct osp_device *d)
1759 if (d->opd_pre == NULL)
1762 OBD_FREE_PTR(d->opd_pre);
1768 int osp_init_statfs(struct osp_device *d)
1770 struct task_struct *task;
1774 spin_lock_init(&d->opd_pre_lock);
1775 init_waitqueue_head(&d->opd_pre_waitq);
1776 thread_set_flags(&d->opd_pre_thread, SVC_INIT);
1777 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1780 * Initialize statfs-related things
1782 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1783 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1784 1000 * NSEC_PER_SEC);
1785 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1787 ktime_to_ns(d->opd_statfs_fresh_till));
1788 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1789 (unsigned long)d, 0);
1791 if (d->opd_storage->dd_rdonly)
1795 * start thread handling precreation and statfs updates
1797 task = kthread_run(osp_precreate_thread, d,
1798 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1800 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1801 RETURN(PTR_ERR(task));
1804 wait_event_idle(d->opd_pre_thread.t_ctl_waitq,
1805 osp_precreate_running(d) || osp_precreate_stopped(d));
1811 void osp_statfs_fini(struct osp_device *d)
1813 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1816 del_timer(&d->opd_statfs_timer);
1818 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1819 thread->t_flags = SVC_STOPPING;
1820 wake_up(&d->opd_pre_waitq);
1821 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));