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/
31 * lustre/osp/osp_precreate.c
33 * Lustre OST Proxy Device
35 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
36 * Author: Mikhail Pershin <mike.pershin@intel.com>
37 * Author: Di Wang <di.wang@intel.com>
40 #define DEBUG_SUBSYSTEM S_MDS
42 #include <linux/kthread.h>
44 #include <lustre_obdo.h>
46 #include "osp_internal.h"
49 * there are two specific states to take care about:
51 * = import is disconnected =
53 * = import is inactive =
54 * in this case osp_declare_create() returns an error
59 * Check whether statfs data is expired
61 * OSP device caches statfs data for the target, the function checks
62 * whether the data is expired or not.
64 * \param[in] d OSP device
66 * \retval 0 - not expired, 1 - expired
68 static inline int osp_statfs_need_update(struct osp_device *d)
70 return !ktime_before(ktime_get(), d->opd_statfs_fresh_till);
74 * OSP tries to maintain pool of available objects so that calls to create
75 * objects don't block most of time
77 * each time OSP gets connected to OST, we should start from precreation cleanup
79 static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data)
81 struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
85 wake_up(&d->opd_pre_waitq);
88 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs);
91 * The function updates current precreation status if broken, and
92 * updates that cached statfs state if functional, then wakes up waiters.
93 * We don't clear opd_pre_status directly here, but rather leave this
94 * to osp_pre_update_msfs() to do if everything is OK so that we don't
95 * have a race to clear opd_pre_status and then set it to -ENOSPC again.
97 * \param[in] d OSP device
98 * \param[in] msfs statfs data
99 * \param[in] rc new precreate status for device \a d
101 static void osp_pre_update_status_msfs(struct osp_device *d,
102 struct obd_statfs *msfs, int rc)
105 d->opd_pre_status = rc;
107 osp_pre_update_msfs(d, msfs);
109 wake_up(&d->opd_pre_user_waitq);
112 /* Pass in the old statfs data in case the limits have changed */
113 void osp_pre_update_status(struct osp_device *d, int rc)
115 osp_pre_update_status_msfs(d, &d->opd_statfs, rc);
120 * RPC interpret callback for OST_STATFS RPC
122 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
123 * replied by the target. It's used to maintain statfs cache for the target.
124 * The function fills data from the reply if successful and schedules another
127 * \param[in] env LU environment provided by the caller
128 * \param[in] req RPC replied
129 * \param[in] aa callback data
130 * \param[in] rc RPC result
132 * \retval 0 on success
133 * \retval negative negated errno on error
135 static int osp_statfs_interpret(const struct lu_env *env,
136 struct ptlrpc_request *req, void *args, int rc)
138 union ptlrpc_async_args *aa = args;
139 struct obd_import *imp = req->rq_import;
140 struct obd_statfs *msfs;
141 struct obd_statfs *sfs;
142 struct osp_device *d;
147 aa = ptlrpc_req_async_args(aa, req);
148 d = aa->pointer_arg[0];
154 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
156 GOTO(out, rc = -EPROTO);
159 osp_pre_update_status_msfs(d, msfs, 0);
161 d->opd_statfs = *msfs;
163 /* schedule next update */
164 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
165 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
166 mod_timer(&d->opd_statfs_timer,
167 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
168 d->opd_statfs_update_in_progress = 0;
170 sfs = &d->opd_statfs;
171 CDEBUG(D_CACHE, "%s (%p): %llu blocks, %llu free, %llu avail, "
172 "%u bsize, %u reserved mb low, %u reserved mb high,"
173 "%llu files, %llu free files\n", d->opd_obd->obd_name, d,
174 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail, sfs->os_bsize,
175 d->opd_reserved_mb_low, d->opd_reserved_mb_high,
176 sfs->os_files, sfs->os_ffree);
180 /* couldn't update statfs, try again with a small delay */
181 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
182 d->opd_statfs_update_in_progress = 0;
183 if (d->opd_pre && d->opd_pre_task)
184 wake_up(&d->opd_pre_waitq);
186 if (req->rq_import_generation == imp->imp_generation)
187 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
188 d->opd_obd->obd_name, rc);
193 * Send OST_STATFS RPC
195 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
196 * Also disables scheduled updates as times OSP may need to refresh
197 * statfs data before expiration. The function doesn't block, instead
198 * an interpretation callback osp_statfs_interpret() is used.
200 * \param[in] d OSP device
202 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
204 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
205 struct ptlrpc_request *req;
206 struct obd_import *imp;
207 union ptlrpc_async_args *aa;
212 CDEBUG(D_CACHE, "going to update statfs\n");
214 imp = d->opd_obd->u.cli.cl_import;
217 req = ptlrpc_request_alloc(imp,
218 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
222 rc = ptlrpc_request_pack(req,
223 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
224 d->opd_pre ? OST_STATFS : MDS_STATFS);
226 ptlrpc_request_free(req);
229 ptlrpc_request_set_replen(req);
231 req->rq_request_portal = OST_CREATE_PORTAL;
232 ptlrpc_at_set_req_timeout(req);
234 req->rq_interpret_reply = osp_statfs_interpret;
235 aa = ptlrpc_req_async_args(aa, req);
236 aa->pointer_arg[0] = d;
239 * no updates till reply
241 del_timer(&d->opd_statfs_timer);
242 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
243 d->opd_statfs_update_in_progress = 1;
245 ptlrpcd_add_req(req);
247 /* we still want to sync changes if no new changes are coming */
248 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
251 if (atomic_read(&d->opd_sync_changes)) {
254 th = dt_trans_create(env, d->opd_storage);
256 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
259 rc = dt_trans_start_local(env, d->opd_storage, th);
261 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
262 d->opd_obd->obd_name,
263 atomic_read(&d->opd_sync_changes));
264 osp_sync_add_commit_cb_1s(env, d, th);
266 dt_trans_stop(env, d->opd_storage, th);
274 * Schedule an immediate update for statfs data
276 * If cached statfs data claim no free space, but OSP has got a request to
277 * destroy an object (so release some space probably), then we may need to
278 * refresh cached statfs data sooner than planned. The function checks there
279 * is no statfs update going and schedules immediate update if so.
280 * XXX: there might be a case where removed object(s) do not add free space (empty
281 * object). If the number of such deletions is high, then we can start to update
282 * statfs too often causing a RPC storm. some throttling is needed...
284 * \param[in] d OSP device where statfs data needs to be refreshed
286 void osp_statfs_need_now(struct osp_device *d)
288 if (!d->opd_statfs_update_in_progress) {
290 * if current status is -ENOSPC (lack of free space on OST)
291 * then we should poll OST immediately once object destroy
294 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
295 del_timer(&d->opd_statfs_timer);
296 wake_up(&d->opd_pre_waitq);
301 * Return number of precreated objects
303 * A simple helper to calculate the number of precreated objects on the device.
305 * \param[in] env LU environment provided by the caller
306 * \param[in] osp OSP device
308 * \retval the number of the precreated objects
310 static inline int osp_objs_precreated(const struct lu_env *env,
311 struct osp_device *osp)
313 return osp_fid_diff(&osp->opd_pre_last_created_fid,
314 &osp->opd_pre_used_fid);
318 * Check pool of precreated objects is nearly empty
320 * We should not wait till the pool of the precreated objects is exhausted,
321 * because then there will be a long period of OSP being unavailable for the
322 * new creations due to lenghty precreate RPC. Instead we ask for another
323 * precreation ahead and hopefully have it ready before the current pool is
324 * empty. Notice this function relies on an external locking.
326 * \param[in] env LU environment provided by the caller
327 * \param[in] d OSP device
329 * \retval 0 - current pool is good enough, 1 - time to precreate
331 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
332 struct osp_device *d)
334 int window = osp_objs_precreated(env, d);
336 /* don't consider new precreation till OST is healty and
338 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
339 (d->opd_pre_status == 0));
343 * Check pool of precreated objects
345 * This is protected version of osp_precreate_near_empty_nolock(), check that
348 * \param[in] env LU environment provided by the caller
349 * \param[in] d OSP device
351 * \retval 0 - current pool is good enough, 1 - time to precreate
353 static inline int osp_precreate_near_empty(const struct lu_env *env,
354 struct osp_device *d)
358 if (d->opd_pre == NULL)
361 /* XXX: do we really need locking here? */
362 spin_lock(&d->opd_pre_lock);
363 rc = osp_precreate_near_empty_nolock(env, d);
364 spin_unlock(&d->opd_pre_lock);
369 * Check given sequence is empty
371 * Returns a binary result whether the given sequence has some IDs left
372 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
373 * version of that function.
375 * \param[in] env LU environment provided by the caller
376 * \param[in] osp OSP device
378 * \retval 0 - current sequence has no IDs, 1 - otherwise
380 static inline int osp_create_end_seq(const struct lu_env *env,
381 struct osp_device *osp)
383 struct lu_fid *fid = &osp->opd_pre_used_fid;
386 spin_lock(&osp->opd_pre_lock);
387 rc = osp_fid_end_seq(env, fid);
388 spin_unlock(&osp->opd_pre_lock);
393 * Write FID into into last_oid/last_seq file
395 * The function stores the sequence and the in-sequence id into two dedicated
396 * files. The sync argument can be used to request synchronous commit, so the
397 * function won't return until the updates are committed.
399 * \param[in] env LU environment provided by the caller
400 * \param[in] osp OSP device
401 * \param[in] fid fid where sequence/id is taken
402 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
404 * \retval 0 on success
405 * \retval negative negated errno on error
407 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
408 struct lu_fid *fid, int sync)
410 struct osp_thread_info *oti = osp_env_info(env);
411 struct lu_buf *lb_oid = &oti->osi_lb;
412 struct lu_buf *lb_oseq = &oti->osi_lb2;
420 if (osp->opd_storage->dd_rdonly)
423 /* Note: through f_oid is only 32 bits, it will also write 64 bits
424 * for oid to keep compatibility with the previous version. */
426 osp_objid_buf_prep(lb_oid, &oid_off,
427 &oid, osp->opd_index);
429 osp_objseq_buf_prep(lb_oseq, &oseq_off,
430 &fid->f_seq, osp->opd_index);
432 th = dt_trans_create(env, osp->opd_storage);
437 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
438 lb_oid, oid_off, th);
442 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
443 lb_oseq, oseq_off, th);
447 rc = dt_trans_start_local(env, osp->opd_storage, th);
451 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
454 CERROR("%s: can not write to last seq file: rc = %d\n",
455 osp->opd_obd->obd_name, rc);
458 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
461 CERROR("%s: can not write to last seq file: rc = %d\n",
462 osp->opd_obd->obd_name, rc);
466 dt_trans_stop(env, osp->opd_storage, th);
471 * Switch to another sequence
473 * When a current sequence has no available IDs left, OSP has to switch to
474 * another new sequence. OSP requests it using the regular FLDB protocol
475 * and stores synchronously before that is used in precreated. This is needed
476 * to basically have the sequences referenced (not orphaned), otherwise it's
477 * possible that OST has some objects precreated and the clients have data
478 * written to it, but after MDT failover nobody refers those objects and OSP
479 * has no idea that the sequence need cleanup to be done.
480 * While this is very expensive operation, it's supposed to happen very very
481 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
483 * \param[in] env LU environment provided by the caller
484 * \param[in] osp OSP device
486 * \retval 0 on success
487 * \retval negative negated errno on error
489 static int osp_precreate_rollover_new_seq(struct lu_env *env,
490 struct osp_device *osp)
492 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
493 struct lu_fid *last_fid = &osp->opd_last_used_fid;
497 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
499 CERROR("%s: alloc fid error: rc = %d\n",
500 osp->opd_obd->obd_name, rc);
506 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
507 "fid "DFID", last_fid "DFID"\n", PFID(fid),
510 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
512 CERROR("%s: Can not update oid/seq file: rc = %d\n",
513 osp->opd_obd->obd_name, rc);
517 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
518 osp->opd_obd->obd_name, fid_seq(last_fid),
520 /* Update last_xxx to the new seq */
521 spin_lock(&osp->opd_pre_lock);
522 osp->opd_last_used_fid = *fid;
523 osp_fid_to_obdid(fid, &osp->opd_last_id);
524 osp->opd_gap_start_fid = *fid;
525 osp->opd_pre_used_fid = *fid;
526 osp->opd_pre_last_created_fid = *fid;
527 spin_unlock(&osp->opd_pre_lock);
533 * Find IDs available in current sequence
535 * The function calculates the highest possible ID and the number of IDs
536 * available in the current sequence OSP is using. The number is limited
537 * artifically by the caller (grow param) and the number of IDs available
538 * in the sequence by nature. The function doesn't require an external
541 * \param[in] env LU environment provided by the caller
542 * \param[in] osp OSP device
543 * \param[in] fid FID the caller wants to start with
544 * \param[in] grow how many the caller wants
545 * \param[out] fid the highest calculated FID
546 * \param[out] grow the number of available IDs calculated
548 * \retval 0 on success, 1 - the sequence is empty
550 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
551 struct lu_fid *fid, int *grow)
553 struct osp_thread_info *osi = osp_env_info(env);
557 if (fid_is_idif(fid)) {
558 struct lu_fid *last_fid;
559 struct ost_id *oi = &osi->osi_oi;
562 spin_lock(&osp->opd_pre_lock);
563 last_fid = &osp->opd_pre_last_created_fid;
564 fid_to_ostid(last_fid, oi);
565 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
566 *grow = end - ostid_id(oi);
567 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
568 spin_unlock(&osp->opd_pre_lock);
570 if (*grow == 0 || rc)
573 ostid_to_fid(fid, oi, osp->opd_index);
577 spin_lock(&osp->opd_pre_lock);
578 *fid = osp->opd_pre_last_created_fid;
580 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
581 *grow = end - fid->f_oid;
582 fid->f_oid += end - fid->f_oid;
583 spin_unlock(&osp->opd_pre_lock);
585 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
586 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
588 return *grow > 0 ? 0 : 1;
592 * Prepare and send precreate RPC
594 * The function finds how many objects should be precreated. Then allocates,
595 * prepares and schedules precreate RPC synchronously. Upon reply the function
596 * wakes up the threads waiting for the new objects on this target. If the
597 * target wasn't able to create all the objects requested, then the next
598 * precreate will be asking for fewer objects (i.e. slow precreate down).
600 * \param[in] env LU environment provided by the caller
601 * \param[in] d OSP device
603 * \retval 0 on success
604 * \retval negative negated errno on error
606 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
608 struct osp_thread_info *oti = osp_env_info(env);
609 struct ptlrpc_request *req;
610 struct obd_import *imp;
611 struct ost_body *body;
613 struct lu_fid *fid = &oti->osi_fid;
616 /* don't precreate new objects till OST healthy and has free space */
617 if (unlikely(d->opd_pre_status)) {
618 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
619 d->opd_obd->obd_name, d->opd_pre_status);
624 * if not connection/initialization is compeleted, ignore
626 imp = d->opd_obd->u.cli.cl_import;
629 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
632 req->rq_request_portal = OST_CREATE_PORTAL;
633 /* we should not resend create request - anyway we will have delorphan
634 * and kill these objects */
635 req->rq_no_delay = req->rq_no_resend = 1;
637 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
639 ptlrpc_request_free(req);
643 spin_lock(&d->opd_pre_lock);
644 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
645 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
646 grow = d->opd_pre_create_count;
647 spin_unlock(&d->opd_pre_lock);
649 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
652 *fid = d->opd_pre_last_created_fid;
653 rc = osp_precreate_fids(env, d, fid, &grow);
655 /* Current seq has been used up*/
656 GOTO(out_req, rc = -ENOSPC);
658 if (!osp_is_fid_client(d)) {
659 /* Non-FID client will always send seq 0 because of
661 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
665 fid_to_ostid(fid, &body->oa.o_oi);
666 body->oa.o_valid = OBD_MD_FLGROUP;
668 ptlrpc_request_set_replen(req);
670 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
673 rc = ptlrpc_queue_wait(req);
675 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
678 /* have osp_precreate_reserve() to wait for repeat */
682 LASSERT(req->rq_transno == 0);
684 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
686 GOTO(out_req, rc = -EPROTO);
688 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
691 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
692 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
693 ": rc = %d\n", d->opd_obd->obd_name,
694 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
695 GOTO(out_req, rc = -ESTALE);
698 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
700 spin_lock(&d->opd_pre_lock);
702 /* the OST has not managed to create all the
703 * objects we asked for */
704 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
705 d->opd_pre_create_slow = 1;
707 /* the OST is able to keep up with the work,
708 * we could consider increasing create_count
709 * next time if needed */
710 d->opd_pre_create_slow = 0;
713 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
714 fid_to_ostid(fid, &body->oa.o_oi);
716 d->opd_pre_last_created_fid = *fid;
717 spin_unlock(&d->opd_pre_lock);
719 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
720 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
721 PFID(&d->opd_pre_last_created_fid));
723 /* now we can wakeup all users awaiting for objects */
724 osp_pre_update_status(d, rc);
725 wake_up(&d->opd_pre_user_waitq);
727 /* pause to let osp_precreate_reserve to go first */
728 CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_PRECREATE_PAUSE, 2);
730 ptlrpc_req_finished(req);
735 * Get last precreated object from target (OST)
737 * Sends synchronous RPC to the target (OST) to learn the last precreated
738 * object. This later is used to remove all unused objects (cleanup orphan
739 * procedure). Also, the next object after one we got will be used as a
740 * starting point for the new precreates.
742 * \param[in] env LU environment provided by the caller
743 * \param[in] d OSP device
745 * \retval 0 on success
746 * \retval negative negated errno on error
748 static int osp_get_lastfid_from_ost(const struct lu_env *env,
749 struct osp_device *d)
751 struct ptlrpc_request *req = NULL;
752 struct obd_import *imp;
753 struct lu_fid *last_fid;
758 imp = d->opd_obd->u.cli.cl_import;
761 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
765 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
766 sizeof(KEY_LAST_FID));
768 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
770 ptlrpc_request_free(req);
774 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
775 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
777 req->rq_no_delay = req->rq_no_resend = 1;
778 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
779 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
781 ptlrpc_request_set_replen(req);
783 rc = ptlrpc_queue_wait(req);
785 /* bad-bad OST.. let sysadm sort this out */
786 if (rc == -ENOTSUPP) {
787 CERROR("%s: server does not support FID: rc = %d\n",
788 d->opd_obd->obd_name, -ENOTSUPP);
790 ptlrpc_set_import_active(imp, 0);
794 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
795 if (last_fid == NULL) {
796 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
797 GOTO(out, rc = -EPROTO);
800 if (!fid_is_sane(last_fid)) {
801 CERROR("%s: Got insane last_fid "DFID"\n",
802 d->opd_obd->obd_name, PFID(last_fid));
803 GOTO(out, rc = -EPROTO);
806 /* Only update the last used fid, if the OST has objects for
807 * this sequence, i.e. fid_oid > 0 */
808 if (fid_oid(last_fid) > 0)
809 d->opd_last_used_fid = *last_fid;
811 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
815 ptlrpc_req_finished(req);
820 * Cleanup orphans on OST
822 * This function is called in a contex of a dedicated thread handling
823 * all the precreation suff. The function waits till local recovery
824 * is complete, then identify all the unreferenced objects (orphans)
825 * using the highest ID referenced by a local and the highest object
826 * precreated by the target. The found range is a subject to removal
827 * using specially flagged RPC. During this process OSP is marked
828 * unavailable for new objects.
830 * \param[in] env LU environment provided by the caller
831 * \param[in] d OSP device
833 * \retval 0 on success
834 * \retval negative negated errno on error
836 static int osp_precreate_cleanup_orphans(struct lu_env *env,
837 struct osp_device *d)
839 struct osp_thread_info *osi = osp_env_info(env);
840 struct lu_fid *last_fid = &osi->osi_fid;
841 struct ptlrpc_request *req = NULL;
842 struct obd_import *imp;
843 struct ost_body *body;
844 int update_status = 0;
851 * wait for local recovery to finish, so we can cleanup orphans
852 * orphans are all objects since "last used" (assigned), but
853 * there might be objects reserved and in some cases they won't
854 * be used. we can't cleanup them till we're sure they won't be
855 * used. also can't we allow new reservations because they may
856 * end up getting orphans being cleaned up below. so we block
857 * new reservations and wait till all reserved objects either
860 spin_lock(&d->opd_pre_lock);
861 d->opd_pre_recovering = 1;
862 spin_unlock(&d->opd_pre_lock);
864 * The locking above makes sure the opd_pre_reserved check below will
865 * catch all osp_precreate_reserve() calls who find
866 * "!opd_pre_recovering".
868 wait_event_idle(d->opd_pre_waitq,
869 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
870 !d->opd_pre_task || d->opd_got_disconnected);
871 if (!d->opd_pre_task || d->opd_got_disconnected)
872 GOTO(out, rc = -EAGAIN);
874 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
875 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
877 *last_fid = d->opd_last_used_fid;
878 /* The OSP should already get the valid seq now */
879 LASSERT(!fid_is_zero(last_fid));
880 if (fid_oid(&d->opd_last_used_fid) < 2) {
881 /* lastfid looks strange... ask OST */
882 rc = osp_get_lastfid_from_ost(env, d);
887 imp = d->opd_obd->u.cli.cl_import;
890 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
892 GOTO(out, rc = -ENOMEM);
894 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
896 ptlrpc_request_free(req);
901 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
903 GOTO(out, rc = -EPROTO);
905 body->oa.o_flags = OBD_FL_DELORPHAN;
906 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
908 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
910 ptlrpc_request_set_replen(req);
912 /* Don't resend the delorphan req */
913 req->rq_no_resend = req->rq_no_delay = 1;
915 rc = ptlrpc_queue_wait(req);
921 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
923 GOTO(out, rc = -EPROTO);
926 * OST provides us with id new pool starts from in body->oa.o_id
928 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
930 spin_lock(&d->opd_pre_lock);
931 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
933 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
934 d->opd_pre_last_created_fid = d->opd_last_used_fid;
936 d->opd_pre_create_count = OST_MIN_PRECREATE;
937 d->opd_pre_last_created_fid = *last_fid;
940 * This empties the pre-creation pool and effectively blocks any new
943 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
944 LUSTRE_DATA_SEQ_MAX_WIDTH);
945 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
946 d->opd_pre_create_slow = 0;
947 spin_unlock(&d->opd_pre_lock);
949 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
950 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
951 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
954 ptlrpc_req_finished(req);
957 * If rc is zero, the pre-creation window should have been emptied.
958 * Since waking up the herd would be useless without pre-created
959 * objects, we defer the signal to osp_precreate_send() in that case.
963 CERROR("%s: cannot cleanup orphans: rc = %d\n",
964 d->opd_obd->obd_name, rc);
965 /* we can't proceed from here, OST seem to
966 * be in a bad shape, better to wait for
967 * a new instance of the server and repeat
968 * from the beginning. notify possible waiters
969 * this OSP isn't quite functional yet */
970 osp_pre_update_status(d, rc);
972 wake_up(&d->opd_pre_user_waitq);
975 spin_lock(&d->opd_pre_lock);
976 d->opd_pre_recovering = 0;
977 spin_unlock(&d->opd_pre_lock);
984 * Update precreate status using statfs data
986 * The function decides whether this OSP should be used for new objects.
987 * IOW, whether this OST is used up or has some free space. Cached statfs
988 * data is used to make this decision. If the latest result of statfs
989 * request (rc argument) is not success, then just mark OSP unavailable
992 * The new statfs data is passed in \a msfs and needs to be stored into
993 * opd_statfs, but only after the various flags in os_state are set, so
994 * that the new statfs data is not visible without appropriate flags set.
995 * As such, there is no need to clear the flags here, since this is called
996 * with new statfs data, and they should not be cleared if sent from OST.
998 * Add a bit of hysteresis so this flag isn't continually flapping, and
999 * ensure that new files don't get extremely fragmented due to only a
1000 * small amount of available space in the filesystem. We want to set
1001 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1002 * reserved size free, and still copy them from the old state when there
1003 * is less than 2*reserved size free space or inodes.
1005 * \param[in] d OSP device
1006 * \param[in] msfs statfs data
1008 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1010 u32 old_state = d->opd_statfs.os_state;
1011 u32 reserved_ino_low = 32; /* could be tunable in the future */
1012 u32 reserved_ino_high = reserved_ino_low * 2;
1015 /* statfs structure not initialized yet */
1016 if (unlikely(!msfs->os_type))
1019 /* if the low and high watermarks have not been initialized yet */
1020 if (unlikely(d->opd_reserved_mb_high == 0 &&
1021 d->opd_reserved_mb_low == 0)) {
1022 /* Use ~0.1% by default to disable object allocation,
1023 * and ~0.2% to enable, size in MB, set both watermark
1025 spin_lock(&d->opd_pre_lock);
1026 if (d->opd_reserved_mb_high == 0 &&
1027 d->opd_reserved_mb_low == 0) {
1028 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1029 msfs->os_blocks) >> 20;
1030 if (d->opd_reserved_mb_low == 0)
1031 d->opd_reserved_mb_low = 1;
1032 d->opd_reserved_mb_high =
1033 (d->opd_reserved_mb_low << 1) + 1;
1035 spin_unlock(&d->opd_pre_lock);
1038 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1039 if (msfs->os_ffree < reserved_ino_low)
1040 msfs->os_state |= OS_STATFS_ENOINO;
1041 else if (msfs->os_ffree <= reserved_ino_high)
1042 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1043 /* else don't clear flags in new msfs->os_state sent from OST */
1046 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1047 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1048 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1049 msfs->os_files, msfs->os_ffree, msfs->os_state,
1051 if (available_mb < d->opd_reserved_mb_low)
1052 msfs->os_state |= OS_STATFS_ENOSPC;
1053 else if (available_mb <= d->opd_reserved_mb_high)
1054 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1055 /* else don't clear flags in new msfs->os_state sent from OST */
1057 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1058 d->opd_pre_status = -ENOSPC;
1059 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1060 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1061 d->opd_obd->obd_name, msfs->os_state,
1063 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1064 atomic_read(&d->opd_sync_changes),
1065 atomic_read(&d->opd_sync_rpcs_in_progress));
1066 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1067 d->opd_pre_status = 0;
1068 spin_lock(&d->opd_pre_lock);
1069 d->opd_pre_create_slow = 0;
1070 d->opd_pre_create_count = OST_MIN_PRECREATE;
1071 spin_unlock(&d->opd_pre_lock);
1072 wake_up(&d->opd_pre_waitq);
1075 "%s: available: state=%x: rc = %d\n",
1076 d->opd_obd->obd_name, msfs->os_state,
1079 /* we only get here if rc == 0 in the caller */
1080 d->opd_pre_status = 0;
1083 /* Object precreation skipped on OST if manually disabled */
1084 if (d->opd_pre_max_create_count == 0)
1085 msfs->os_state |= OS_STATFS_NOPRECREATE;
1086 /* else don't clear flags in new msfs->os_state sent from OST */
1088 /* copy only new statfs state to make it visible to MDS threads */
1089 if (&d->opd_statfs != msfs)
1090 d->opd_statfs = *msfs;
1094 * Initialize FID for precreation
1096 * For a just created new target, a new sequence should be taken.
1097 * The function checks there is no IDIF in use (if the target was
1098 * added with the older version of Lustre), then requests a new
1099 * sequence from FLDB using the regular protocol. Then this new
1100 * sequence is stored on a persisten storage synchronously to prevent
1101 * possible object leakage (for the detail see the description for
1102 * osp_precreate_rollover_new_seq()).
1104 * \param[in] osp OSP device
1106 * \retval 0 on success
1107 * \retval negative negated errno on error
1109 int osp_init_pre_fid(struct osp_device *osp)
1112 struct osp_thread_info *osi;
1113 struct lu_client_seq *cli_seq;
1114 struct lu_fid *last_fid;
1118 LASSERT(osp->opd_pre != NULL);
1120 /* Let's check if the current last_seq/fid is valid,
1121 * otherwise request new sequence from the controller */
1122 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1123 /* Non-MDT0 can only use normal sequence for
1125 if (fid_is_norm(&osp->opd_last_used_fid))
1128 /* Initially MDT0 will start with IDIF, after
1129 * that it will request new sequence from the
1131 if (fid_is_idif(&osp->opd_last_used_fid) ||
1132 fid_is_norm(&osp->opd_last_used_fid))
1136 if (!fid_is_zero(&osp->opd_last_used_fid))
1137 CWARN("%s: invalid last used fid "DFID
1138 ", try to get new sequence.\n",
1139 osp->opd_obd->obd_name,
1140 PFID(&osp->opd_last_used_fid));
1142 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1144 CERROR("%s: init env error: rc = %d\n",
1145 osp->opd_obd->obd_name, rc);
1149 osi = osp_env_info(&env);
1150 last_fid = &osi->osi_fid;
1152 /* For a freshed fs, it will allocate a new sequence first */
1153 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1154 cli_seq = osp->opd_obd->u.cli.cl_seq;
1155 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1157 CERROR("%s: alloc fid error: rc = %d\n",
1158 osp->opd_obd->obd_name, rc);
1162 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1164 last_fid->f_oid = 1;
1165 last_fid->f_ver = 0;
1167 spin_lock(&osp->opd_pre_lock);
1168 osp->opd_last_used_fid = *last_fid;
1169 osp->opd_pre_used_fid = *last_fid;
1170 osp->opd_pre_last_created_fid = *last_fid;
1171 spin_unlock(&osp->opd_pre_lock);
1172 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1174 CERROR("%s: write fid error: rc = %d\n",
1175 osp->opd_obd->obd_name, rc);
1184 struct osp_device *opta_dev;
1185 struct lu_env opta_env;
1186 struct completion *opta_started;
1189 * The core of precreate functionality
1191 * The function implements the main precreation loop. Basically it
1192 * involves connecting to the target, precerate FID initialization,
1193 * identifying and removing orphans, then serving precreation. As
1194 * part of the latter, the thread is responsible for statfs data
1195 * updates. The precreation is mostly driven by another threads
1196 * asking for new OST objects - those askers wake the thread when
1197 * the number of precreated objects reach low watermark.
1198 * After a disconnect, the sequence above repeats. This is keep going
1199 * until the thread is requested to stop.
1201 * \param[in] _arg private data the thread (OSP device to handle)
1203 * \retval 0 on success
1204 * \retval negative negated errno on error
1206 static int osp_precreate_thread(void *_args)
1208 struct opt_args *args = _args;
1209 struct osp_device *d = args->opta_dev;
1210 struct lu_env *env = &args->opta_env;
1215 complete(args->opta_started);
1216 while (!kthread_should_stop()) {
1218 * need to be connected to OST
1220 while (!kthread_should_stop()) {
1221 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1222 d->opd_imp_connected &&
1223 !d->opd_got_disconnected)
1225 wait_event_idle(d->opd_pre_waitq,
1226 kthread_should_stop() ||
1227 d->opd_new_connection);
1229 if (!d->opd_new_connection)
1232 OBD_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
1234 d->opd_new_connection = 0;
1235 d->opd_got_disconnected = 0;
1239 if (kthread_should_stop())
1243 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1244 /* Sigh, fid client is not ready yet */
1245 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1248 /* Init fid for osp_precreate if necessary */
1249 rc = osp_init_pre_fid(d);
1251 class_export_put(d->opd_exp);
1252 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1253 CERROR("%s: init pre fid error: rc = %d\n",
1254 d->opd_obd->obd_name, rc);
1259 if (osp_statfs_update(env, d)) {
1260 if (wait_event_idle_timeout(d->opd_pre_waitq,
1261 kthread_should_stop(),
1262 cfs_time_seconds(5)) == 0)
1263 l_wait_event_abortable(
1265 kthread_should_stop());
1271 * Clean up orphans or recreate missing objects.
1273 rc = osp_precreate_cleanup_orphans(env, d);
1275 schedule_timeout_interruptible(cfs_time_seconds(1));
1281 * connected, can handle precreates now
1283 while (!kthread_should_stop()) {
1284 wait_event_idle(d->opd_pre_waitq,
1285 kthread_should_stop() ||
1286 osp_precreate_near_empty(env, d) ||
1287 osp_statfs_need_update(d) ||
1288 d->opd_got_disconnected);
1290 if (kthread_should_stop())
1293 /* something happened to the connection
1294 * have to start from the beginning */
1295 if (d->opd_got_disconnected)
1298 if (osp_statfs_need_update(d))
1299 if (osp_statfs_update(env, d))
1302 if (d->opd_pre == NULL)
1305 /* To avoid handling different seq in precreate/orphan
1306 * cleanup, it will hold precreate until current seq is
1308 if (unlikely(osp_precreate_end_seq(env, d) &&
1309 !osp_create_end_seq(env, d)))
1312 if (unlikely(osp_precreate_end_seq(env, d) &&
1313 osp_create_end_seq(env, d))) {
1314 LCONSOLE_INFO("%s:%#llx is used up."
1315 " Update to new seq\n",
1316 d->opd_obd->obd_name,
1317 fid_seq(&d->opd_pre_last_created_fid));
1318 rc = osp_precreate_rollover_new_seq(env, d);
1323 if (osp_precreate_near_empty(env, d)) {
1324 rc = osp_precreate_send(env, d);
1325 /* osp_precreate_send() sets opd_pre_status
1326 * in case of error, that prevent the using of
1328 if (rc < 0 && rc != -ENOSPC &&
1329 rc != -ETIMEDOUT && rc != -ENOTCONN)
1330 CERROR("%s: cannot precreate objects:"
1332 d->opd_obd->obd_name, rc);
1344 * Check when to stop to wait for precreate objects.
1346 * The caller wanting a new OST object can't wait undefinitely. The
1347 * function checks for few conditions including available new OST
1348 * objects, disconnected OST, lack of space with no pending destroys,
1349 * etc. IOW, it checks whether the current OSP state is good to keep
1350 * waiting or it's better to give up.
1352 * \param[in] env LU environment provided by the caller
1353 * \param[in] d OSP device
1355 * \retval 0 - keep waiting, 1 - no luck
1357 static int osp_precreate_ready_condition(const struct lu_env *env,
1358 struct osp_device *d)
1360 if (d->opd_pre_recovering)
1363 /* ready if got enough precreated objects */
1364 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1365 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1368 /* ready if OST reported no space and no destroys in progress */
1369 if (atomic_read(&d->opd_sync_changes) +
1370 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1371 d->opd_pre_status == -ENOSPC)
1374 /* Bail out I/O fails to OST */
1375 if (d->opd_pre_status != 0 &&
1376 d->opd_pre_status != -EAGAIN &&
1377 d->opd_pre_status != -ENODEV &&
1378 d->opd_pre_status != -ENOTCONN &&
1379 d->opd_pre_status != -ENOSPC) {
1381 if (d->opd_pre_status != -EIO)
1382 CERROR("%s: precreate failed opd_pre_status %d\n",
1383 d->opd_obd->obd_name, d->opd_pre_status);
1391 * Reserve object in precreate pool
1393 * When the caller wants to create a new object on this target (target
1394 * represented by the given OSP), it should declare this intention using
1395 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1396 * to reserve an object in the existing precreated pool or wait up to
1397 * obd_timeout for the available object to appear in the pool (a dedicated
1398 * thread will be doing real precreation in background). The object can be
1399 * consumed later with osp_precreate_get_fid() or be released with call to
1400 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1401 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1402 * If the space on the target is short and there is a pending object destroy,
1403 * then the function forces local commit to speedup space release (see
1404 * osp_sync.c for the details).
1406 * \param[in] env LU environment provided by the caller
1407 * \param[in] d OSP device
1409 * \retval 0 on success
1410 * \retval -ENOSPC when no space on OST
1411 * \retval -EAGAIN try later, slow precreation in progress
1412 * \retval -EIO when no access to OST
1414 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
1417 time64_t expire = ktime_get_seconds() + obd_timeout;
1418 int precreated, rc, synced = 0;
1422 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1423 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1424 PFID(&d->opd_pre_used_fid));
1426 /* opd_pre_max_create_count 0 to not use specified OST. */
1427 if (d->opd_pre_max_create_count == 0)
1432 * - preallocation is done
1433 * - no free space expected soon
1434 * - can't connect to OST for too long (obd_timeout)
1435 * - OST can allocate fid sequence.
1437 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1438 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1441 * increase number of precreations
1443 precreated = osp_objs_precreated(env, d);
1444 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1445 d->opd_pre_create_slow == 0 &&
1446 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1447 spin_lock(&d->opd_pre_lock);
1448 d->opd_pre_create_slow = 1;
1449 d->opd_pre_create_count *= 2;
1450 spin_unlock(&d->opd_pre_lock);
1453 spin_lock(&d->opd_pre_lock);
1454 precreated = osp_objs_precreated(env, d);
1455 if (precreated > d->opd_pre_reserved &&
1456 !d->opd_pre_recovering) {
1457 d->opd_pre_reserved++;
1458 spin_unlock(&d->opd_pre_lock);
1461 /* XXX: don't wake up if precreation is in progress */
1462 if (osp_precreate_near_empty_nolock(env, d) &&
1463 !osp_precreate_end_seq_nolock(env, d))
1464 wake_up(&d->opd_pre_waitq);
1468 spin_unlock(&d->opd_pre_lock);
1471 * all precreated objects have been used and no-space
1472 * status leave us no chance to succeed very soon
1473 * but if there is destroy in progress, then we should
1474 * wait till that is done - some space might be released
1476 if (unlikely(rc == -ENOSPC)) {
1477 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1478 /* force local commit to release space */
1479 dt_commit_async(env, d->opd_storage);
1480 osp_sync_check_for_work(d);
1483 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1484 /* just wait till destroys are done
1485 * see wait_event_idle_timeout() below
1488 if (atomic_read(&d->opd_sync_changes) +
1489 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1490 /* no hope for free space */
1495 /* XXX: don't wake up if precreation is in progress */
1496 wake_up(&d->opd_pre_waitq);
1498 if (ktime_get_seconds() >= expire) {
1504 LASSERT(d->opd_pre);
1509 if (wait_event_idle_timeout(
1510 d->opd_pre_user_waitq,
1511 osp_precreate_ready_condition(env, d),
1512 cfs_time_seconds(obd_timeout)) == 0) {
1514 "%s: slow creates, last="DFID", next="DFID", "
1515 "reserved=%llu, sync_changes=%u, "
1516 "sync_rpcs_in_progress=%d, status=%d\n",
1517 d->opd_obd->obd_name,
1518 PFID(&d->opd_pre_last_created_fid),
1519 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1520 atomic_read(&d->opd_sync_changes),
1521 atomic_read(&d->opd_sync_rpcs_in_progress),
1530 * Get a FID from precreation pool
1532 * The function is a companion for osp_precreate_reserve() - it assigns
1533 * a specific FID from the precreate. The function should be called only
1534 * if the call to osp_precreate_reserve() was successful. The function
1535 * updates a local storage to remember the highest object ID referenced
1536 * by the node in the given sequence.
1538 * A very importan details: this is supposed to be called once the
1539 * transaction is started, so on-disk update will be atomic with the
1540 * data (like LOVEA) refering this object. Then the object won't be leaked:
1541 * either it's referenced by the committed transaction or it's a subject
1542 * to the orphan cleanup procedure.
1544 * \param[in] env LU environment provided by the caller
1545 * \param[in] d OSP device
1546 * \param[out] fid generated FID
1548 * \retval 0 on success
1549 * \retval negative negated errno on error
1551 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1554 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1555 /* grab next id from the pool */
1556 spin_lock(&d->opd_pre_lock);
1558 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1559 &d->opd_pre_last_created_fid) < 0,
1560 "next fid "DFID" last created fid "DFID"\n",
1561 PFID(&d->opd_pre_used_fid),
1562 PFID(&d->opd_pre_last_created_fid));
1565 * When sequence is used up, new one should be allocated in
1566 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1569 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1570 "next fid "DFID" last created fid "DFID"\n",
1571 PFID(&d->opd_pre_used_fid),
1572 PFID(&d->opd_pre_last_created_fid));
1573 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1574 if (fid_is_idif(pre_used_fid) &&
1575 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1576 pre_used_fid->f_seq++;
1578 d->opd_pre_used_fid.f_oid++;
1579 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1580 d->opd_pre_reserved--;
1582 * last_used_id must be changed along with getting new id otherwise
1583 * we might miscalculate gap causing object loss or leak
1585 osp_update_last_fid(d, fid);
1586 spin_unlock(&d->opd_pre_lock);
1589 * probably main thread suspended orphan cleanup till
1590 * all reservations are released, see comment in
1591 * osp_precreate_thread() just before orphan cleanup
1593 if (unlikely(d->opd_pre_reserved == 0 &&
1594 (d->opd_pre_recovering || d->opd_pre_status)))
1595 wake_up(&d->opd_pre_waitq);
1601 * Set size regular attribute on an object
1603 * When a striping is created late, it's possible that size is already
1604 * initialized on the file. Then the new striping should inherit size
1605 * from the file. The function sets size on the object using the regular
1606 * protocol (OST_PUNCH).
1607 * XXX: should be re-implemented using OUT ?
1609 * \param[in] env LU environment provided by the caller
1610 * \param[in] dt object
1611 * \param[in] size size to set.
1613 * \retval 0 on success
1614 * \retval negative negated errno on error
1616 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1619 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1620 struct ptlrpc_request *req = NULL;
1621 struct obd_import *imp;
1622 struct ost_body *body;
1623 struct obdo *oa = NULL;
1628 imp = d->opd_obd->u.cli.cl_import;
1631 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1635 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1637 ptlrpc_request_free(req);
1642 * XXX: decide how do we do here with resend
1643 * if we don't resend, then client may see wrong file size
1644 * if we do resend, then MDS thread can get stuck for quite long
1645 * and if we don't resend, then client will also get -EAGAIN !!
1646 * (see LU-7975 and sanity/test_27F use cases)
1647 * but let's decide not to resend/delay this truncate request to OST
1648 * and allow Client to decide to resend, in a less agressive way from
1649 * after_reply(), by returning -EINPROGRESS instead of
1650 * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
1651 * end of this routine
1653 req->rq_no_resend = req->rq_no_delay = 1;
1655 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1656 ptlrpc_at_set_req_timeout(req);
1660 GOTO(out, rc = -ENOMEM);
1662 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1665 oa->o_blocks = OBD_OBJECT_EOF;
1666 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1667 OBD_MD_FLID | OBD_MD_FLGROUP;
1669 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1671 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1673 /* XXX: capa support? */
1674 /* osc_pack_capa(req, body, capa); */
1676 ptlrpc_request_set_replen(req);
1678 rc = ptlrpc_queue_wait(req);
1680 /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
1681 * since we have decided not to resend/delay, but this could
1682 * lead to wrong size to be seen at Client side and even process
1683 * trying to open to exit/fail if not itself handling -EAGAIN.
1684 * So it should be better to return -EINPROGRESS instead and
1685 * leave the decision to resend at Client side in after_reply()
1687 if (rc == -EAGAIN) {
1689 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1690 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1693 CERROR("can't punch object: %d\n", rc);
1697 ptlrpc_req_finished(req);
1704 * Initialize precreation functionality of OSP
1706 * Prepares all the internal structures and starts the precreate thread
1708 * \param[in] d OSP device
1710 * \retval 0 on success
1711 * \retval negative negated errno on error
1713 int osp_init_precreate(struct osp_device *d)
1717 OBD_ALLOC_PTR(d->opd_pre);
1718 if (d->opd_pre == NULL)
1721 /* initially precreation isn't ready */
1722 init_waitqueue_head(&d->opd_pre_user_waitq);
1723 d->opd_pre_status = -EAGAIN;
1724 fid_zero(&d->opd_pre_used_fid);
1725 d->opd_pre_used_fid.f_oid = 1;
1726 fid_zero(&d->opd_pre_last_created_fid);
1727 d->opd_pre_last_created_fid.f_oid = 1;
1729 d->opd_pre_reserved = 0;
1730 d->opd_got_disconnected = 1;
1731 d->opd_pre_create_slow = 0;
1732 d->opd_pre_create_count = OST_MIN_PRECREATE;
1733 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1734 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1735 d->opd_reserved_mb_high = 0;
1736 d->opd_reserved_mb_low = 0;
1742 * Finish precreate functionality of OSP
1745 * Asks all the activity (the thread, update timer) to stop, then
1746 * wait till that is done.
1748 * \param[in] d OSP device
1750 void osp_precreate_fini(struct osp_device *d)
1754 if (d->opd_pre == NULL)
1757 OBD_FREE_PTR(d->opd_pre);
1763 int osp_init_statfs(struct osp_device *d)
1765 struct task_struct *task;
1766 struct opt_args *args;
1767 DECLARE_COMPLETION_ONSTACK(started);
1772 spin_lock_init(&d->opd_pre_lock);
1773 init_waitqueue_head(&d->opd_pre_waitq);
1776 * Initialize statfs-related things
1778 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1779 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1780 1000 * NSEC_PER_SEC);
1781 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1783 ktime_to_ns(d->opd_statfs_fresh_till));
1784 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1785 (unsigned long)d, 0);
1787 if (d->opd_storage->dd_rdonly)
1790 OBD_ALLOC_PTR(args);
1794 args->opta_started = &started;
1795 rc = lu_env_init(&args->opta_env,
1796 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1798 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1805 * start thread handling precreation and statfs updates
1807 task = kthread_create(osp_precreate_thread, args,
1808 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1810 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1811 lu_env_fini(&args->opta_env);
1813 RETURN(PTR_ERR(task));
1815 d->opd_pre_task = task;
1816 wake_up_process(task);
1817 wait_for_completion(&started);
1822 void osp_statfs_fini(struct osp_device *d)
1824 struct task_struct *task = d->opd_pre_task;
1827 del_timer(&d->opd_statfs_timer);
1829 d->opd_pre_task = NULL;