4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osp/osp_precreate.c
34 * Lustre OST Proxy Device
36 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
37 * Author: Mikhail Pershin <mike.pershin@intel.com>
38 * Author: Di Wang <di.wang@intel.com>
41 #define DEBUG_SUBSYSTEM S_MDS
43 #include <linux/kthread.h>
45 #include <lustre_obdo.h>
47 #include "osp_internal.h"
50 * there are two specific states to take care about:
52 * = import is disconnected =
54 * = import is inactive =
55 * in this case osp_declare_create() returns an error
61 * Check whether statfs data is expired
63 * OSP device caches statfs data for the target, the function checks
64 * whether the data is expired or not.
66 * \param[in] d OSP device
68 * \retval 0 - not expired, 1 - expired
70 static inline int osp_statfs_need_update(struct osp_device *d)
72 return !ktime_before(ktime_get(), d->opd_statfs_fresh_till);
76 * OSP tries to maintain pool of available objects so that calls to create
77 * objects don't block most of time
79 * each time OSP gets connected to OST, we should start from precreation cleanup
81 static inline bool osp_precreate_running(struct osp_device *d)
83 return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
86 static inline bool osp_precreate_stopped(struct osp_device *d)
88 return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
91 static void osp_statfs_timer_cb(unsigned long _d)
93 struct osp_device *d = (struct osp_device *) _d;
96 if (d->opd_pre != NULL && osp_precreate_running(d))
97 wake_up(&d->opd_pre_waitq);
101 * RPC interpret callback for OST_STATFS RPC
103 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
104 * replied by the target. It's used to maintain statfs cache for the target.
105 * The function fills data from the reply if successful and schedules another
108 * \param[in] env LU environment provided by the caller
109 * \param[in] req RPC replied
110 * \param[in] aa callback data
111 * \param[in] rc RPC result
113 * \retval 0 on success
114 * \retval negative negated errno on error
116 static int osp_statfs_interpret(const struct lu_env *env,
117 struct ptlrpc_request *req,
118 union ptlrpc_async_args *aa, int rc)
120 struct obd_import *imp = req->rq_import;
121 struct obd_statfs *msfs;
122 struct osp_device *d;
127 aa = ptlrpc_req_async_args(req);
128 d = aa->pointer_arg[0];
134 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
136 GOTO(out, rc = -EPROTO);
138 d->opd_statfs = *msfs;
140 osp_pre_update_status(d, rc);
142 /* schedule next update */
143 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
144 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
145 mod_timer(&d->opd_statfs_timer,
146 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
147 d->opd_statfs_update_in_progress = 0;
149 CDEBUG(D_CACHE, "updated statfs %p\n", d);
153 /* couldn't update statfs, try again as soon as possible */
154 if (d->opd_pre != NULL && osp_precreate_running(d))
155 wake_up(&d->opd_pre_waitq);
157 if (req->rq_import_generation == imp->imp_generation)
158 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
159 d->opd_obd->obd_name, rc);
164 * Send OST_STATFS RPC
166 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
167 * Also disables scheduled updates as times OSP may need to refresh
168 * statfs data before expiration. The function doesn't block, instead
169 * an interpretation callback osp_statfs_interpret() is used.
171 * \param[in] d OSP device
173 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
175 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
176 struct ptlrpc_request *req;
177 struct obd_import *imp;
178 union ptlrpc_async_args *aa;
183 CDEBUG(D_CACHE, "going to update statfs\n");
185 imp = d->opd_obd->u.cli.cl_import;
188 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
192 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
194 ptlrpc_request_free(req);
197 ptlrpc_request_set_replen(req);
198 req->rq_request_portal = OST_CREATE_PORTAL;
199 ptlrpc_at_set_req_timeout(req);
201 req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
202 aa = ptlrpc_req_async_args(req);
203 aa->pointer_arg[0] = d;
206 * no updates till reply
208 del_timer(&d->opd_statfs_timer);
209 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
210 d->opd_statfs_update_in_progress = 1;
212 ptlrpcd_add_req(req);
214 /* we still want to sync changes if no new changes are coming */
215 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
218 if (atomic_read(&d->opd_sync_changes)) {
221 th = dt_trans_create(env, d->opd_storage);
223 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
226 rc = dt_trans_start_local(env, d->opd_storage, th);
228 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
229 d->opd_obd->obd_name,
230 atomic_read(&d->opd_sync_changes));
231 osp_sync_add_commit_cb_1s(env, d, th);
232 dt_trans_stop(env, d->opd_storage, th);
241 * Schedule an immediate update for statfs data
243 * If cached statfs data claim no free space, but OSP has got a request to
244 * destroy an object (so release some space probably), then we may need to
245 * refresh cached statfs data sooner than planned. The function checks there
246 * is no statfs update going and schedules immediate update if so.
247 * XXX: there might be a case where removed object(s) do not add free space (empty
248 * object). If the number of such deletions is high, then we can start to update
249 * statfs too often causing a RPC storm. some throttling is needed...
251 * \param[in] d OSP device where statfs data needs to be refreshed
253 void osp_statfs_need_now(struct osp_device *d)
255 if (!d->opd_statfs_update_in_progress) {
257 * if current status is -ENOSPC (lack of free space on OST)
258 * then we should poll OST immediately once object destroy
261 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
262 del_timer(&d->opd_statfs_timer);
263 wake_up(&d->opd_pre_waitq);
268 * Return number of precreated objects
270 * A simple helper to calculate the number of precreated objects on the device.
272 * \param[in] env LU environment provided by the caller
273 * \param[in] osp OSP device
275 * \retval the number of the precreated objects
277 static inline int osp_objs_precreated(const struct lu_env *env,
278 struct osp_device *osp)
280 return osp_fid_diff(&osp->opd_pre_last_created_fid,
281 &osp->opd_pre_used_fid);
285 * Check pool of precreated objects is nearly empty
287 * We should not wait till the pool of the precreated objects is exhausted,
288 * because then there will be a long period of OSP being unavailable for the
289 * new creations due to lenghty precreate RPC. Instead we ask for another
290 * precreation ahead and hopefully have it ready before the current pool is
291 * empty. Notice this function relies on an external locking.
293 * \param[in] env LU environment provided by the caller
294 * \param[in] d OSP device
296 * \retval 0 - current pool is good enough, 1 - time to precreate
298 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
299 struct osp_device *d)
301 int window = osp_objs_precreated(env, d);
303 /* don't consider new precreation till OST is healty and
305 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
306 (d->opd_pre_status == 0));
310 * Check pool of precreated objects
312 * This is protected version of osp_precreate_near_empty_nolock(), check that
315 * \param[in] env LU environment provided by the caller
316 * \param[in] d OSP device
318 * \retval 0 - current pool is good enough, 1 - time to precreate
320 static inline int osp_precreate_near_empty(const struct lu_env *env,
321 struct osp_device *d)
325 /* XXX: do we really need locking here? */
326 spin_lock(&d->opd_pre_lock);
327 rc = osp_precreate_near_empty_nolock(env, d);
328 spin_unlock(&d->opd_pre_lock);
333 * Check given sequence is empty
335 * Returns a binary result whether the given sequence has some IDs left
336 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
337 * version of that function.
339 * \param[in] env LU environment provided by the caller
340 * \param[in] osp OSP device
342 * \retval 0 - current sequence has no IDs, 1 - otherwise
344 static inline int osp_create_end_seq(const struct lu_env *env,
345 struct osp_device *osp)
347 struct lu_fid *fid = &osp->opd_pre_used_fid;
350 spin_lock(&osp->opd_pre_lock);
351 rc = osp_fid_end_seq(env, fid);
352 spin_unlock(&osp->opd_pre_lock);
357 * Write FID into into last_oid/last_seq file
359 * The function stores the sequence and the in-sequence id into two dedicated
360 * files. The sync argument can be used to request synchronous commit, so the
361 * function won't return until the updates are committed.
363 * \param[in] env LU environment provided by the caller
364 * \param[in] osp OSP device
365 * \param[in] fid fid where sequence/id is taken
366 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
368 * \retval 0 on success
369 * \retval negative negated errno on error
371 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
372 struct lu_fid *fid, int sync)
374 struct osp_thread_info *oti = osp_env_info(env);
375 struct lu_buf *lb_oid = &oti->osi_lb;
376 struct lu_buf *lb_oseq = &oti->osi_lb2;
383 if (osp->opd_storage->dd_rdonly)
386 /* Note: through f_oid is only 32 bits, it will also write 64 bits
387 * for oid to keep compatibility with the previous version. */
388 lb_oid->lb_buf = &fid->f_oid;
389 lb_oid->lb_len = sizeof(u64);
390 oid_off = sizeof(u64) * osp->opd_index;
392 lb_oseq->lb_buf = &fid->f_seq;
393 lb_oseq->lb_len = sizeof(u64);
394 oseq_off = sizeof(u64) * osp->opd_index;
396 th = dt_trans_create(env, osp->opd_storage);
401 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
402 lb_oid, oid_off, th);
406 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
407 lb_oseq, oseq_off, th);
411 rc = dt_trans_start_local(env, osp->opd_storage, th);
415 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
418 CERROR("%s: can not write to last seq file: rc = %d\n",
419 osp->opd_obd->obd_name, rc);
422 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
425 CERROR("%s: can not write to last seq file: rc = %d\n",
426 osp->opd_obd->obd_name, rc);
430 dt_trans_stop(env, osp->opd_storage, th);
435 * Switch to another sequence
437 * When a current sequence has no available IDs left, OSP has to switch to
438 * another new sequence. OSP requests it using the regular FLDB protocol
439 * and stores synchronously before that is used in precreated. This is needed
440 * to basically have the sequences referenced (not orphaned), otherwise it's
441 * possible that OST has some objects precreated and the clients have data
442 * written to it, but after MDT failover nobody refers those objects and OSP
443 * has no idea that the sequence need cleanup to be done.
444 * While this is very expensive operation, it's supposed to happen very very
445 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
447 * \param[in] env LU environment provided by the caller
448 * \param[in] osp OSP device
450 * \retval 0 on success
451 * \retval negative negated errno on error
453 static int osp_precreate_rollover_new_seq(struct lu_env *env,
454 struct osp_device *osp)
456 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
457 struct lu_fid *last_fid = &osp->opd_last_used_fid;
461 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
463 CERROR("%s: alloc fid error: rc = %d\n",
464 osp->opd_obd->obd_name, rc);
470 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
471 "fid "DFID", last_fid "DFID"\n", PFID(fid),
474 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
476 CERROR("%s: Can not update oid/seq file: rc = %d\n",
477 osp->opd_obd->obd_name, rc);
481 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
482 osp->opd_obd->obd_name, fid_seq(last_fid),
484 /* Update last_xxx to the new seq */
485 spin_lock(&osp->opd_pre_lock);
486 osp->opd_last_used_fid = *fid;
487 osp->opd_gap_start_fid = *fid;
488 osp->opd_pre_used_fid = *fid;
489 osp->opd_pre_last_created_fid = *fid;
490 spin_unlock(&osp->opd_pre_lock);
496 * Find IDs available in current sequence
498 * The function calculates the highest possible ID and the number of IDs
499 * available in the current sequence OSP is using. The number is limited
500 * artifically by the caller (grow param) and the number of IDs available
501 * in the sequence by nature. The function doesn't require an external
504 * \param[in] env LU environment provided by the caller
505 * \param[in] osp OSP device
506 * \param[in] fid FID the caller wants to start with
507 * \param[in] grow how many the caller wants
508 * \param[out] fid the highest calculated FID
509 * \param[out] grow the number of available IDs calculated
511 * \retval 0 on success, 1 - the sequence is empty
513 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
514 struct lu_fid *fid, int *grow)
516 struct osp_thread_info *osi = osp_env_info(env);
520 if (fid_is_idif(fid)) {
521 struct lu_fid *last_fid;
522 struct ost_id *oi = &osi->osi_oi;
525 spin_lock(&osp->opd_pre_lock);
526 last_fid = &osp->opd_pre_last_created_fid;
527 fid_to_ostid(last_fid, oi);
528 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
529 *grow = end - ostid_id(oi);
530 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
531 spin_unlock(&osp->opd_pre_lock);
533 if (*grow == 0 || rc)
536 ostid_to_fid(fid, oi, osp->opd_index);
540 spin_lock(&osp->opd_pre_lock);
541 *fid = osp->opd_pre_last_created_fid;
543 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
544 *grow = end - fid->f_oid;
545 fid->f_oid += end - fid->f_oid;
546 spin_unlock(&osp->opd_pre_lock);
548 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
549 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
551 return *grow > 0 ? 0 : 1;
555 * Prepare and send precreate RPC
557 * The function finds how many objects should be precreated. Then allocates,
558 * prepares and schedules precreate RPC synchronously. Upon reply the function
559 * wake ups the threads waiting for the new objects on this target. If the
560 * target wasn't able to create all the objects requested, then the next
561 * precreate will be asking less objects (i.e. slow precreate down).
563 * \param[in] env LU environment provided by the caller
564 * \param[in] d OSP device
566 * \retval 0 on success
567 * \retval negative negated errno on error
569 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
571 struct osp_thread_info *oti = osp_env_info(env);
572 struct ptlrpc_request *req;
573 struct obd_import *imp;
574 struct ost_body *body;
576 struct lu_fid *fid = &oti->osi_fid;
579 /* don't precreate new objects till OST healthy and has free space */
580 if (unlikely(d->opd_pre_status)) {
581 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
582 d->opd_obd->obd_name, d->opd_pre_status);
587 * if not connection/initialization is compeleted, ignore
589 imp = d->opd_obd->u.cli.cl_import;
592 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
595 req->rq_request_portal = OST_CREATE_PORTAL;
596 /* we should not resend create request - anyway we will have delorphan
597 * and kill these objects */
598 req->rq_no_delay = req->rq_no_resend = 1;
600 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
602 ptlrpc_request_free(req);
606 spin_lock(&d->opd_pre_lock);
607 if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
608 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
609 grow = d->opd_pre_create_count;
610 spin_unlock(&d->opd_pre_lock);
612 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
615 *fid = d->opd_pre_last_created_fid;
616 rc = osp_precreate_fids(env, d, fid, &grow);
618 /* Current seq has been used up*/
619 if (!osp_is_fid_client(d)) {
620 osp_pre_update_status(d, -ENOSPC);
623 wake_up(&d->opd_pre_waitq);
627 if (!osp_is_fid_client(d)) {
628 /* Non-FID client will always send seq 0 because of
630 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
634 fid_to_ostid(fid, &body->oa.o_oi);
635 body->oa.o_valid = OBD_MD_FLGROUP;
637 ptlrpc_request_set_replen(req);
639 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
642 rc = ptlrpc_queue_wait(req);
644 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
648 LASSERT(req->rq_transno == 0);
650 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
652 GOTO(out_req, rc = -EPROTO);
654 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
657 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
658 CERROR("%s: precreate fid "DFID" < local used fid "DFID
659 ": rc = %d\n", d->opd_obd->obd_name,
660 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
661 GOTO(out_req, rc = -ESTALE);
664 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
666 spin_lock(&d->opd_pre_lock);
668 /* the OST has not managed to create all the
669 * objects we asked for */
670 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
671 d->opd_pre_create_slow = 1;
673 /* the OST is able to keep up with the work,
674 * we could consider increasing create_count
675 * next time if needed */
676 d->opd_pre_create_slow = 0;
679 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
680 fid_to_ostid(fid, &body->oa.o_oi);
682 d->opd_pre_last_created_fid = *fid;
683 spin_unlock(&d->opd_pre_lock);
685 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
686 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
687 PFID(&d->opd_pre_last_created_fid));
689 /* now we can wakeup all users awaiting for objects */
690 osp_pre_update_status(d, rc);
691 wake_up(&d->opd_pre_user_waitq);
693 ptlrpc_req_finished(req);
698 * Get last precreated object from target (OST)
700 * Sends synchronous RPC to the target (OST) to learn the last precreated
701 * object. This later is used to remove all unused objects (cleanup orphan
702 * procedure). Also, the next object after one we got will be used as a
703 * starting point for the new precreates.
705 * \param[in] env LU environment provided by the caller
706 * \param[in] d OSP device
708 * \retval 0 on success
709 * \retval negative negated errno on error
711 static int osp_get_lastfid_from_ost(const struct lu_env *env,
712 struct osp_device *d)
714 struct ptlrpc_request *req = NULL;
715 struct obd_import *imp;
716 struct lu_fid *last_fid;
721 imp = d->opd_obd->u.cli.cl_import;
724 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
728 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
729 sizeof(KEY_LAST_FID));
731 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
733 ptlrpc_request_free(req);
737 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
738 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
740 req->rq_no_delay = req->rq_no_resend = 1;
741 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
742 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
744 ptlrpc_request_set_replen(req);
746 rc = ptlrpc_queue_wait(req);
748 /* bad-bad OST.. let sysadm sort this out */
749 if (rc == -ENOTSUPP) {
750 CERROR("%s: server does not support FID: rc = %d\n",
751 d->opd_obd->obd_name, -ENOTSUPP);
753 ptlrpc_set_import_active(imp, 0);
757 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
758 if (last_fid == NULL) {
759 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
760 GOTO(out, rc = -EPROTO);
763 if (!fid_is_sane(last_fid)) {
764 CERROR("%s: Got insane last_fid "DFID"\n",
765 d->opd_obd->obd_name, PFID(last_fid));
766 GOTO(out, rc = -EPROTO);
769 /* Only update the last used fid, if the OST has objects for
770 * this sequence, i.e. fid_oid > 0 */
771 if (fid_oid(last_fid) > 0)
772 d->opd_last_used_fid = *last_fid;
774 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
778 ptlrpc_req_finished(req);
783 * Cleanup orphans on OST
785 * This function is called in a contex of a dedicated thread handling
786 * all the precreation suff. The function waits till local recovery
787 * is complete, then identify all the unreferenced objects (orphans)
788 * using the highest ID referenced by a local and the highest object
789 * precreated by the target. The found range is a subject to removal
790 * using specially flagged RPC. During this process OSP is marked
791 * unavailable for new objects.
793 * \param[in] env LU environment provided by the caller
794 * \param[in] d OSP device
796 * \retval 0 on success
797 * \retval negative negated errno on error
799 static int osp_precreate_cleanup_orphans(struct lu_env *env,
800 struct osp_device *d)
802 struct osp_thread_info *osi = osp_env_info(env);
803 struct lu_fid *last_fid = &osi->osi_fid;
804 struct ptlrpc_request *req = NULL;
805 struct obd_import *imp;
806 struct ost_body *body;
807 struct l_wait_info lwi = { 0 };
808 int update_status = 0;
815 * wait for local recovery to finish, so we can cleanup orphans
816 * orphans are all objects since "last used" (assigned), but
817 * there might be objects reserved and in some cases they won't
818 * be used. we can't cleanup them till we're sure they won't be
819 * used. also can't we allow new reservations because they may
820 * end up getting orphans being cleaned up below. so we block
821 * new reservations and wait till all reserved objects either
824 spin_lock(&d->opd_pre_lock);
825 d->opd_pre_recovering = 1;
826 spin_unlock(&d->opd_pre_lock);
828 * The locking above makes sure the opd_pre_reserved check below will
829 * catch all osp_precreate_reserve() calls who find
830 * "!opd_pre_recovering".
832 l_wait_event(d->opd_pre_waitq,
833 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
834 !osp_precreate_running(d) || d->opd_got_disconnected,
836 if (!osp_precreate_running(d) || d->opd_got_disconnected)
837 GOTO(out, rc = -EAGAIN);
839 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
840 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
842 *last_fid = d->opd_last_used_fid;
843 /* The OSP should already get the valid seq now */
844 LASSERT(!fid_is_zero(last_fid));
845 if (fid_oid(&d->opd_last_used_fid) < 2) {
846 /* lastfid looks strange... ask OST */
847 rc = osp_get_lastfid_from_ost(env, d);
852 imp = d->opd_obd->u.cli.cl_import;
855 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
857 GOTO(out, rc = -ENOMEM);
859 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
861 ptlrpc_request_free(req);
866 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
868 GOTO(out, rc = -EPROTO);
870 body->oa.o_flags = OBD_FL_DELORPHAN;
871 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
873 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
875 ptlrpc_request_set_replen(req);
877 /* Don't resend the delorphan req */
878 req->rq_no_resend = req->rq_no_delay = 1;
880 rc = ptlrpc_queue_wait(req);
886 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
888 GOTO(out, rc = -EPROTO);
891 * OST provides us with id new pool starts from in body->oa.o_id
893 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
895 spin_lock(&d->opd_pre_lock);
896 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
898 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
899 d->opd_pre_last_created_fid = d->opd_last_used_fid;
901 d->opd_pre_create_count = OST_MIN_PRECREATE;
902 d->opd_pre_last_created_fid = *last_fid;
905 * This empties the pre-creation pool and effectively blocks any new
908 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
909 LUSTRE_DATA_SEQ_MAX_WIDTH);
910 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
911 d->opd_pre_create_slow = 0;
912 spin_unlock(&d->opd_pre_lock);
914 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
915 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
916 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
919 ptlrpc_req_finished(req);
922 * If rc is zero, the pre-creation window should have been emptied.
923 * Since waking up the herd would be useless without pre-created
924 * objects, we defer the signal to osp_precreate_send() in that case.
928 CERROR("%s: cannot cleanup orphans: rc = %d\n",
929 d->opd_obd->obd_name, rc);
930 /* we can't proceed from here, OST seem to
931 * be in a bad shape, better to wait for
932 * a new instance of the server and repeat
933 * from the beginning. notify possible waiters
934 * this OSP isn't quite functional yet */
935 osp_pre_update_status(d, rc);
937 wake_up(&d->opd_pre_user_waitq);
940 spin_lock(&d->opd_pre_lock);
941 d->opd_pre_recovering = 0;
942 spin_unlock(&d->opd_pre_lock);
949 * Update precreate status using statfs data
951 * The function decides whether this OSP should be used for new objects.
952 * IOW, whether this OST is used up or has some free space. Cached statfs
953 * data is used to make this decision. If the latest result of statfs
954 * request (rc argument) is not success, then just mark OSP unavailable
957 * Add a bit of hysteresis so this flag isn't continually flapping,
958 * and ensure that new files don't get extremely fragmented due to
959 * only a small amount of available space in the filesystem.
960 * We want to set the ENOSPC when there is less than reserved size
961 * free and clear it when there is at least 2*reserved size free space.
962 * the function updates current precreation status used: functional or not
964 * \param[in] d OSP device
965 * \param[in] rc new precreate status for device \a d
967 * \retval 0 on success
968 * \retval negative negated errno on error
970 void osp_pre_update_status(struct osp_device *d, int rc)
972 struct obd_statfs *msfs = &d->opd_statfs;
973 int old = d->opd_pre_status;
976 d->opd_pre_status = rc;
980 if (likely(msfs->os_type)) {
981 if (unlikely(d->opd_reserved_mb_high == 0 &&
982 d->opd_reserved_mb_low == 0)) {
983 /* Use ~0.1% by default to disable object allocation,
984 * and ~0.2% to enable, size in MB, set both watermark
986 spin_lock(&d->opd_pre_lock);
987 if (d->opd_reserved_mb_high == 0 &&
988 d->opd_reserved_mb_low == 0) {
989 d->opd_reserved_mb_low =
990 ((msfs->os_bsize >> 10) *
991 msfs->os_blocks) >> 20;
992 if (d->opd_reserved_mb_low == 0)
993 d->opd_reserved_mb_low = 1;
994 d->opd_reserved_mb_high =
995 (d->opd_reserved_mb_low << 1) + 1;
997 spin_unlock(&d->opd_pre_lock);
1000 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1001 if (msfs->os_ffree < 32)
1002 msfs->os_state |= OS_STATE_ENOINO;
1003 else if (msfs->os_ffree > 64)
1004 msfs->os_state &= ~OS_STATE_ENOINO;
1006 if (available < d->opd_reserved_mb_low)
1007 msfs->os_state |= OS_STATE_ENOSPC;
1008 else if (available > d->opd_reserved_mb_high)
1009 msfs->os_state &= ~OS_STATE_ENOSPC;
1010 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
1011 d->opd_pre_status = -ENOSPC;
1013 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
1014 "free, %llu avail, %llu MB avail, %u "
1015 "hwm -> %d: rc = %d\n",
1016 d->opd_obd->obd_name, msfs->os_blocks,
1017 msfs->os_bfree, msfs->os_bavail,
1018 available, d->opd_reserved_mb_high,
1019 d->opd_pre_status, rc);
1021 "non-committed changes: %u, in progress: %u\n",
1022 atomic_read(&d->opd_sync_changes),
1023 atomic_read(&d->opd_sync_rpcs_in_progress));
1024 } else if (unlikely(old == -ENOSPC)) {
1025 d->opd_pre_status = 0;
1026 spin_lock(&d->opd_pre_lock);
1027 d->opd_pre_create_slow = 0;
1028 d->opd_pre_create_count = OST_MIN_PRECREATE;
1029 spin_unlock(&d->opd_pre_lock);
1030 wake_up(&d->opd_pre_waitq);
1032 CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1033 " free, %llu avail, %lluMB avail, %u lwm"
1034 " -> %d: rc = %d\n", d->opd_obd->obd_name,
1035 msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1036 available, d->opd_reserved_mb_low,
1037 d->opd_pre_status, rc);
1041 wake_up(&d->opd_pre_user_waitq);
1045 * Initialize FID for precreation
1047 * For a just created new target, a new sequence should be taken.
1048 * The function checks there is no IDIF in use (if the target was
1049 * added with the older version of Lustre), then requests a new
1050 * sequence from FLDB using the regular protocol. Then this new
1051 * sequence is stored on a persisten storage synchronously to prevent
1052 * possible object leakage (for the detail see the description for
1053 * osp_precreate_rollover_new_seq()).
1055 * \param[in] osp OSP device
1057 * \retval 0 on success
1058 * \retval negative negated errno on error
1060 int osp_init_pre_fid(struct osp_device *osp)
1063 struct osp_thread_info *osi;
1064 struct lu_client_seq *cli_seq;
1065 struct lu_fid *last_fid;
1069 LASSERT(osp->opd_pre != NULL);
1071 /* Let's check if the current last_seq/fid is valid,
1072 * otherwise request new sequence from the controller */
1073 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1074 /* Non-MDT0 can only use normal sequence for
1076 if (fid_is_norm(&osp->opd_last_used_fid))
1079 /* Initially MDT0 will start with IDIF, after
1080 * that it will request new sequence from the
1082 if (fid_is_idif(&osp->opd_last_used_fid) ||
1083 fid_is_norm(&osp->opd_last_used_fid))
1087 if (!fid_is_zero(&osp->opd_last_used_fid))
1088 CWARN("%s: invalid last used fid "DFID
1089 ", try to get new sequence.\n",
1090 osp->opd_obd->obd_name,
1091 PFID(&osp->opd_last_used_fid));
1093 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1095 CERROR("%s: init env error: rc = %d\n",
1096 osp->opd_obd->obd_name, rc);
1100 osi = osp_env_info(&env);
1101 last_fid = &osi->osi_fid;
1103 /* For a freshed fs, it will allocate a new sequence first */
1104 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1105 cli_seq = osp->opd_obd->u.cli.cl_seq;
1106 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1108 CERROR("%s: alloc fid error: rc = %d\n",
1109 osp->opd_obd->obd_name, rc);
1113 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1115 last_fid->f_oid = 1;
1116 last_fid->f_ver = 0;
1118 spin_lock(&osp->opd_pre_lock);
1119 osp->opd_last_used_fid = *last_fid;
1120 osp->opd_pre_used_fid = *last_fid;
1121 osp->opd_pre_last_created_fid = *last_fid;
1122 spin_unlock(&osp->opd_pre_lock);
1123 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1125 CERROR("%s: write fid error: rc = %d\n",
1126 osp->opd_obd->obd_name, rc);
1135 * The core of precreate functionality
1137 * The function implements the main precreation loop. Basically it
1138 * involves connecting to the target, precerate FID initialization,
1139 * identifying and removing orphans, then serving precreation. As
1140 * part of the latter, the thread is responsible for statfs data
1141 * updates. The precreation is mostly driven by another threads
1142 * asking for new OST objects - those askers wake the thread when
1143 * the number of precreated objects reach low watermark.
1144 * After a disconnect, the sequence above repeats. This is keep going
1145 * until the thread is requested to stop.
1147 * \param[in] _arg private data the thread (OSP device to handle)
1149 * \retval 0 on success
1150 * \retval negative negated errno on error
1152 static int osp_precreate_thread(void *_arg)
1154 struct osp_device *d = _arg;
1155 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1156 struct l_wait_info lwi = { 0 };
1157 struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1158 back_to_sleep, NULL);
1164 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1166 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1169 spin_lock(&d->opd_pre_lock);
1170 thread->t_flags = SVC_STOPPED;
1171 spin_unlock(&d->opd_pre_lock);
1172 wake_up(&thread->t_ctl_waitq);
1177 spin_lock(&d->opd_pre_lock);
1178 thread->t_flags = SVC_RUNNING;
1179 spin_unlock(&d->opd_pre_lock);
1180 wake_up(&thread->t_ctl_waitq);
1182 while (osp_precreate_running(d)) {
1184 * need to be connected to OST
1186 while (osp_precreate_running(d)) {
1187 if (d->opd_pre_recovering &&
1188 d->opd_imp_connected &&
1189 !d->opd_got_disconnected)
1191 l_wait_event(d->opd_pre_waitq,
1192 !osp_precreate_running(d) ||
1193 d->opd_new_connection,
1196 if (!d->opd_new_connection)
1199 d->opd_new_connection = 0;
1200 d->opd_got_disconnected = 0;
1204 if (!osp_precreate_running(d))
1207 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1208 /* Sigh, fid client is not ready yet */
1209 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1212 /* Init fid for osp_precreate if necessary */
1213 rc = osp_init_pre_fid(d);
1215 class_export_put(d->opd_exp);
1216 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1217 CERROR("%s: init pre fid error: rc = %d\n",
1218 d->opd_obd->obd_name, rc);
1222 if (osp_statfs_update(&env, d)) {
1223 l_wait_event(d->opd_pre_waitq,
1224 !osp_precreate_running(d), &lwi2);
1229 * Clean up orphans or recreate missing objects.
1231 rc = osp_precreate_cleanup_orphans(&env, d);
1233 schedule_timeout_interruptible(cfs_time_seconds(1));
1237 * connected, can handle precreates now
1239 while (osp_precreate_running(d)) {
1240 l_wait_event(d->opd_pre_waitq,
1241 !osp_precreate_running(d) ||
1242 osp_precreate_near_empty(&env, d) ||
1243 osp_statfs_need_update(d) ||
1244 d->opd_got_disconnected, &lwi);
1246 if (!osp_precreate_running(d))
1249 /* something happened to the connection
1250 * have to start from the beginning */
1251 if (d->opd_got_disconnected)
1254 if (osp_statfs_need_update(d))
1255 if (osp_statfs_update(&env, d))
1258 /* To avoid handling different seq in precreate/orphan
1259 * cleanup, it will hold precreate until current seq is
1261 if (unlikely(osp_precreate_end_seq(&env, d) &&
1262 !osp_create_end_seq(&env, d)))
1265 if (unlikely(osp_precreate_end_seq(&env, d) &&
1266 osp_create_end_seq(&env, d))) {
1267 LCONSOLE_INFO("%s:%#llx is used up."
1268 " Update to new seq\n",
1269 d->opd_obd->obd_name,
1270 fid_seq(&d->opd_pre_last_created_fid));
1271 rc = osp_precreate_rollover_new_seq(&env, d);
1276 if (osp_precreate_near_empty(&env, d)) {
1277 rc = osp_precreate_send(&env, d);
1278 /* osp_precreate_send() sets opd_pre_status
1279 * in case of error, that prevent the using of
1281 if (rc < 0 && rc != -ENOSPC &&
1282 rc != -ETIMEDOUT && rc != -ENOTCONN)
1283 CERROR("%s: cannot precreate objects:"
1285 d->opd_obd->obd_name, rc);
1290 thread->t_flags = SVC_STOPPED;
1292 wake_up(&thread->t_ctl_waitq);
1298 * Check when to stop to wait for precreate objects.
1300 * The caller wanting a new OST object can't wait undefinitely. The
1301 * function checks for few conditions including available new OST
1302 * objects, disconnected OST, lack of space with no pending destroys,
1303 * etc. IOW, it checks whether the current OSP state is good to keep
1304 * waiting or it's better to give up.
1306 * \param[in] env LU environment provided by the caller
1307 * \param[in] d OSP device
1309 * \retval 0 - keep waiting, 1 - no luck
1311 static int osp_precreate_ready_condition(const struct lu_env *env,
1312 struct osp_device *d)
1314 if (d->opd_pre_recovering)
1317 /* ready if got enough precreated objects */
1318 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1319 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1322 /* ready if OST reported no space and no destroys in progress */
1323 if (atomic_read(&d->opd_sync_changes) +
1324 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1325 d->opd_pre_status == -ENOSPC)
1328 /* Bail out I/O fails to OST */
1329 if (d->opd_pre_status != 0 &&
1330 d->opd_pre_status != -EAGAIN &&
1331 d->opd_pre_status != -ENODEV &&
1332 d->opd_pre_status != -ENOTCONN &&
1333 d->opd_pre_status != -ENOSPC) {
1335 if (d->opd_pre_status != -EIO)
1336 CERROR("%s: precreate failed opd_pre_status %d\n",
1337 d->opd_obd->obd_name, d->opd_pre_status);
1344 static int osp_precreate_timeout_condition(void *data)
1346 struct osp_device *d = data;
1348 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1349 "reserved=%llu, sync_changes=%u, "
1350 "sync_rpcs_in_progress=%d, status=%d\n",
1351 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1352 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1353 atomic_read(&d->opd_sync_changes),
1354 atomic_read(&d->opd_sync_rpcs_in_progress),
1361 * Reserve object in precreate pool
1363 * When the caller wants to create a new object on this target (target
1364 * represented by the given OSP), it should declare this intention using
1365 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1366 * to reserve an object in the existing precreated pool or wait up to
1367 * obd_timeout for the available object to appear in the pool (a dedicated
1368 * thread will be doing real precreation in background). The object can be
1369 * consumed later with osp_precreate_get_fid() or be released with call to
1370 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1371 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1372 * If the space on the target is short and there is a pending object destroy,
1373 * then the function forces local commit to speedup space release (see
1374 * osp_sync.c for the details).
1376 * \param[in] env LU environment provided by the caller
1377 * \param[in] d OSP device
1379 * \retval 0 on success
1380 * \retval -ENOSPC when no space on OST
1381 * \retval -EAGAIN try later, slow precreation in progress
1382 * \retval -EIO when no access to OST
1384 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1386 time64_t expire = ktime_get_seconds() + obd_timeout;
1387 struct l_wait_info lwi;
1388 int precreated, rc, synced = 0;
1392 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1393 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1394 PFID(&d->opd_pre_used_fid));
1396 /* opd_pre_max_create_count 0 to not use specified OST. */
1397 if (d->opd_pre_max_create_count == 0)
1402 * - preallocation is done
1403 * - no free space expected soon
1404 * - can't connect to OST for too long (obd_timeout)
1405 * - OST can allocate fid sequence.
1407 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1408 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1411 * increase number of precreations
1413 precreated = osp_objs_precreated(env, d);
1414 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1415 d->opd_pre_create_slow == 0 &&
1416 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1417 spin_lock(&d->opd_pre_lock);
1418 d->opd_pre_create_slow = 1;
1419 d->opd_pre_create_count *= 2;
1420 spin_unlock(&d->opd_pre_lock);
1423 spin_lock(&d->opd_pre_lock);
1424 precreated = osp_objs_precreated(env, d);
1425 if (precreated > d->opd_pre_reserved &&
1426 !d->opd_pre_recovering) {
1427 d->opd_pre_reserved++;
1428 spin_unlock(&d->opd_pre_lock);
1431 /* XXX: don't wake up if precreation is in progress */
1432 if (osp_precreate_near_empty_nolock(env, d) &&
1433 !osp_precreate_end_seq_nolock(env, d))
1434 wake_up(&d->opd_pre_waitq);
1438 spin_unlock(&d->opd_pre_lock);
1441 * all precreated objects have been used and no-space
1442 * status leave us no chance to succeed very soon
1443 * but if there is destroy in progress, then we should
1444 * wait till that is done - some space might be released
1446 if (unlikely(rc == -ENOSPC)) {
1447 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1448 /* force local commit to release space */
1449 dt_commit_async(env, d->opd_storage);
1450 osp_sync_force(env, d);
1453 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1454 /* just wait till destroys are done */
1455 /* see l_wait_even() few lines below */
1457 if (atomic_read(&d->opd_sync_changes) +
1458 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1459 /* no hope for free space */
1464 /* XXX: don't wake up if precreation is in progress */
1465 wake_up(&d->opd_pre_waitq);
1467 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
1468 osp_precreate_timeout_condition, d);
1469 if (ktime_get_seconds() >= expire) {
1474 l_wait_event(d->opd_pre_user_waitq,
1475 osp_precreate_ready_condition(env, d), &lwi);
1482 * Get a FID from precreation pool
1484 * The function is a companion for osp_precreate_reserve() - it assigns
1485 * a specific FID from the precreate. The function should be called only
1486 * if the call to osp_precreate_reserve() was successful. The function
1487 * updates a local storage to remember the highest object ID referenced
1488 * by the node in the given sequence.
1490 * A very importan details: this is supposed to be called once the
1491 * transaction is started, so on-disk update will be atomic with the
1492 * data (like LOVEA) refering this object. Then the object won't be leaked:
1493 * either it's referenced by the committed transaction or it's a subject
1494 * to the orphan cleanup procedure.
1496 * \param[in] env LU environment provided by the caller
1497 * \param[in] d OSP device
1498 * \param[out] fid generated FID
1500 * \retval 0 on success
1501 * \retval negative negated errno on error
1503 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1506 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1507 /* grab next id from the pool */
1508 spin_lock(&d->opd_pre_lock);
1510 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1511 &d->opd_pre_last_created_fid) < 0,
1512 "next fid "DFID" last created fid "DFID"\n",
1513 PFID(&d->opd_pre_used_fid),
1514 PFID(&d->opd_pre_last_created_fid));
1517 * When sequence is used up, new one should be allocated in
1518 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1521 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1522 "next fid "DFID" last created fid "DFID"\n",
1523 PFID(&d->opd_pre_used_fid),
1524 PFID(&d->opd_pre_last_created_fid));
1525 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1526 if (fid_is_idif(pre_used_fid) &&
1527 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1528 pre_used_fid->f_seq++;
1530 d->opd_pre_used_fid.f_oid++;
1531 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1532 d->opd_pre_reserved--;
1534 * last_used_id must be changed along with getting new id otherwise
1535 * we might miscalculate gap causing object loss or leak
1537 osp_update_last_fid(d, fid);
1538 spin_unlock(&d->opd_pre_lock);
1541 * probably main thread suspended orphan cleanup till
1542 * all reservations are released, see comment in
1543 * osp_precreate_thread() just before orphan cleanup
1545 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1546 wake_up(&d->opd_pre_waitq);
1552 * Set size regular attribute on an object
1554 * When a striping is created late, it's possible that size is already
1555 * initialized on the file. Then the new striping should inherit size
1556 * from the file. The function sets size on the object using the regular
1557 * protocol (OST_PUNCH).
1558 * XXX: should be re-implemented using OUT ?
1560 * \param[in] env LU environment provided by the caller
1561 * \param[in] dt object
1562 * \param[in] size size to set.
1564 * \retval 0 on success
1565 * \retval negative negated errno on error
1567 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1570 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1571 struct ptlrpc_request *req = NULL;
1572 struct obd_import *imp;
1573 struct ost_body *body;
1574 struct obdo *oa = NULL;
1579 imp = d->opd_obd->u.cli.cl_import;
1582 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1586 /* XXX: capa support? */
1587 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1588 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1590 ptlrpc_request_free(req);
1595 * XXX: decide how do we do here with resend
1596 * if we don't resend, then client may see wrong file size
1597 * if we do resend, then MDS thread can get stuck for quite long
1598 * and if we don't resend, then client will also get -EWOULDBLOCK !!
1599 * (see LU-7975 and sanity/test_27F use cases)
1600 * but let's decide not to resend/delay this truncate request to OST
1601 * and allow Client to decide to resend, in a less agressive way from
1602 * after_reply(), by returning -EINPROGRESS instead of
1603 * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1604 * end of this routine
1606 req->rq_no_resend = req->rq_no_delay = 1;
1608 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1609 ptlrpc_at_set_req_timeout(req);
1613 GOTO(out, rc = -ENOMEM);
1615 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1618 oa->o_blocks = OBD_OBJECT_EOF;
1619 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1620 OBD_MD_FLID | OBD_MD_FLGROUP;
1622 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1624 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1626 /* XXX: capa support? */
1627 /* osc_pack_capa(req, body, capa); */
1629 ptlrpc_request_set_replen(req);
1631 rc = ptlrpc_queue_wait(req);
1633 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1634 * since we have decided not to resend/delay, but this could
1635 * lead to wrong size to be seen at Client side and even process
1636 * trying to open to exit/fail if not itself handling -EAGAIN.
1637 * So it should be better to return -EINPROGRESS instead and
1638 * leave the decision to resend at Client side in after_reply()
1640 if (rc == -EWOULDBLOCK) {
1642 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1643 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1646 CERROR("can't punch object: %d\n", rc);
1650 ptlrpc_req_finished(req);
1657 * Initialize precreation functionality of OSP
1659 * Prepares all the internal structures and starts the precreate thread
1661 * \param[in] d OSP device
1663 * \retval 0 on success
1664 * \retval negative negated errno on error
1666 int osp_init_precreate(struct osp_device *d)
1668 struct l_wait_info lwi = { 0 };
1669 struct task_struct *task;
1673 OBD_ALLOC_PTR(d->opd_pre);
1674 if (d->opd_pre == NULL)
1677 /* initially precreation isn't ready */
1678 d->opd_pre_status = -EAGAIN;
1679 fid_zero(&d->opd_pre_used_fid);
1680 d->opd_pre_used_fid.f_oid = 1;
1681 fid_zero(&d->opd_pre_last_created_fid);
1682 d->opd_pre_last_created_fid.f_oid = 1;
1683 d->opd_pre_reserved = 0;
1684 d->opd_got_disconnected = 1;
1685 d->opd_pre_create_slow = 0;
1686 d->opd_pre_create_count = OST_MIN_PRECREATE;
1687 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1688 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1689 d->opd_reserved_mb_high = 0;
1690 d->opd_reserved_mb_low = 0;
1692 spin_lock_init(&d->opd_pre_lock);
1693 init_waitqueue_head(&d->opd_pre_waitq);
1694 init_waitqueue_head(&d->opd_pre_user_waitq);
1695 thread_set_flags(&d->opd_pre_thread, SVC_INIT);
1696 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1699 * Initialize statfs-related things
1701 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1702 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1703 1000 * NSEC_PER_SEC);
1704 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1706 ktime_to_ns(d->opd_statfs_fresh_till));
1707 setup_timer(&d->opd_statfs_timer, osp_statfs_timer_cb,
1710 if (d->opd_storage->dd_rdonly)
1714 * start thread handling precreation and statfs updates
1716 task = kthread_run(osp_precreate_thread, d,
1717 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1719 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1720 RETURN(PTR_ERR(task));
1723 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1724 osp_precreate_running(d) || osp_precreate_stopped(d),
1731 * Finish precreate functionality of OSP
1734 * Asks all the activity (the thread, update timer) to stop, then
1735 * wait till that is done.
1737 * \param[in] d OSP device
1739 void osp_precreate_fini(struct osp_device *d)
1741 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1744 del_timer(&d->opd_statfs_timer);
1746 if (d->opd_pre == NULL)
1749 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1750 thread->t_flags = SVC_STOPPING;
1751 wake_up(&d->opd_pre_waitq);
1752 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1755 OBD_FREE_PTR(d->opd_pre);