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);
84 /* invalidate statfs data so osp_precreate_thread() can refresh */
85 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
87 wake_up(&d->opd_pre_waitq);
90 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs);
93 * The function updates current precreation status if broken, and
94 * updates that cached statfs state if functional, then wakes up waiters.
95 * We don't clear opd_pre_status directly here, but rather leave this
96 * to osp_pre_update_msfs() to do if everything is OK so that we don't
97 * have a race to clear opd_pre_status and then set it to -ENOSPC again.
99 * \param[in] d OSP device
100 * \param[in] msfs statfs data
101 * \param[in] rc new precreate status for device \a d
103 static void osp_pre_update_status_msfs(struct osp_device *d,
104 struct obd_statfs *msfs, int rc)
106 CDEBUG(D_INFO, "%s: Updating status = %d\n", d->opd_obd->obd_name, rc);
108 d->opd_pre_status = rc;
110 osp_pre_update_msfs(d, msfs);
112 wake_up_all(&d->opd_pre_user_waitq);
115 /* Pass in the old statfs data in case the limits have changed */
116 void osp_pre_update_status(struct osp_device *d, int rc)
118 osp_pre_update_status_msfs(d, &d->opd_statfs, rc);
123 * RPC interpret callback for OST_STATFS RPC
125 * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
126 * replied by the target. It's used to maintain statfs cache for the target.
127 * The function fills data from the reply if successful and schedules another
130 * \param[in] env LU environment provided by the caller
131 * \param[in] req RPC replied
132 * \param[in] aa callback data
133 * \param[in] rc RPC result
135 * \retval 0 on success
136 * \retval negative negated errno on error
138 static int osp_statfs_interpret(const struct lu_env *env,
139 struct ptlrpc_request *req, void *args, int rc)
141 union ptlrpc_async_args *aa = args;
142 struct obd_import *imp = req->rq_import;
143 struct obd_statfs *msfs;
144 struct obd_statfs *sfs;
145 struct osp_device *d;
150 aa = ptlrpc_req_async_args(aa, req);
151 d = aa->pointer_arg[0];
157 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
159 GOTO(out, rc = -EPROTO);
162 osp_pre_update_status_msfs(d, msfs, 0);
164 osp_pre_update_msfs(d, msfs);
166 /* schedule next update */
167 maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
168 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns);
169 mod_timer(&d->opd_statfs_timer,
170 jiffies + cfs_time_seconds(d->opd_statfs_maxage));
171 d->opd_statfs_update_in_progress = 0;
173 sfs = &d->opd_statfs;
175 "%s (%p): %llu blocks, %llu free, %llu avail, %u bsize, %u reserved mb low, %u reserved mb high, %u reserved ino low, %u reserved ino high, %llu files, %llu free files %#x\n",
176 d->opd_obd->obd_name, d, sfs->os_blocks, sfs->os_bfree,
177 sfs->os_bavail, sfs->os_bsize, d->opd_reserved_mb_low,
178 d->opd_reserved_mb_high, d->opd_reserved_ino_low,
179 d->opd_reserved_ino_high, sfs->os_files, sfs->os_ffree,
184 /* couldn't update statfs, try again with a small delay */
185 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC);
186 d->opd_statfs_update_in_progress = 0;
187 if (d->opd_pre && d->opd_pre_task)
188 wake_up(&d->opd_pre_waitq);
190 if (req->rq_import_generation == imp->imp_generation)
191 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
192 d->opd_obd->obd_name, rc);
197 * Send OST_STATFS RPC
199 * Sends OST_STATFS RPC to refresh cached statfs data for the target.
200 * Also disables scheduled updates as times OSP may need to refresh
201 * statfs data before expiration. The function doesn't block, instead
202 * an interpretation callback osp_statfs_interpret() is used.
204 * \param[in] d OSP device
206 static int osp_statfs_update(const struct lu_env *env, struct osp_device *d)
208 u64 expire = obd_timeout * 1000 * NSEC_PER_SEC;
209 struct ptlrpc_request *req;
210 struct obd_import *imp;
211 union ptlrpc_async_args *aa;
216 CDEBUG(D_CACHE, "going to update statfs\n");
218 imp = d->opd_obd->u.cli.cl_import;
221 req = ptlrpc_request_alloc(imp,
222 d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
226 rc = ptlrpc_request_pack(req,
227 d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
228 d->opd_pre ? OST_STATFS : MDS_STATFS);
230 ptlrpc_request_free(req);
233 ptlrpc_request_set_replen(req);
235 req->rq_request_portal = OST_CREATE_PORTAL;
236 ptlrpc_at_set_req_timeout(req);
238 req->rq_interpret_reply = osp_statfs_interpret;
239 aa = ptlrpc_req_async_args(aa, req);
240 aa->pointer_arg[0] = d;
243 * no updates till reply
245 del_timer(&d->opd_statfs_timer);
246 d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire);
247 d->opd_statfs_update_in_progress = 1;
249 ptlrpcd_add_req(req);
251 /* we still want to sync changes if no new changes are coming */
252 if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb))
255 if (atomic_read(&d->opd_sync_changes)) {
258 th = dt_trans_create(env, d->opd_storage);
260 CERROR("%s: can't sync\n", d->opd_obd->obd_name);
263 rc = dt_trans_start_local(env, d->opd_storage, th);
265 CDEBUG(D_OTHER, "%s: sync forced, %d changes\n",
266 d->opd_obd->obd_name,
267 atomic_read(&d->opd_sync_changes));
268 osp_sync_add_commit_cb_1s(env, d, th);
270 dt_trans_stop(env, d->opd_storage, th);
278 * Schedule an immediate update for statfs data
280 * If cached statfs data claim no free space, but OSP has got a request to
281 * destroy an object (so release some space probably), then we may need to
282 * refresh cached statfs data sooner than planned. The function checks there
283 * is no statfs update going and schedules immediate update if so.
284 * XXX: there might be a case where removed object(s) do not add free space (empty
285 * object). If the number of such deletions is high, then we can start to update
286 * statfs too often causing a RPC storm. some throttling is needed...
288 * \param[in] d OSP device where statfs data needs to be refreshed
290 void osp_statfs_need_now(struct osp_device *d)
292 if (!d->opd_statfs_update_in_progress) {
294 * if current status is -ENOSPC (lack of free space on OST)
295 * then we should poll OST immediately once object destroy
298 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
299 del_timer(&d->opd_statfs_timer);
300 wake_up(&d->opd_pre_waitq);
305 * Return number of precreated objects
307 * A simple helper to calculate the number of precreated objects on the device.
309 * \param[in] env LU environment provided by the caller
310 * \param[in] osp OSP device
312 * \retval the number of the precreated objects
314 static inline int osp_objs_precreated(const struct lu_env *env,
315 struct osp_device *osp)
317 return osp_fid_diff(&osp->opd_pre_last_created_fid,
318 &osp->opd_pre_used_fid);
322 * Check pool of precreated objects is nearly empty
324 * We should not wait till the pool of the precreated objects is exhausted,
325 * because then there will be a long period of OSP being unavailable for the
326 * new creations due to lenghty precreate RPC. Instead we ask for another
327 * precreation ahead and hopefully have it ready before the current pool is
328 * empty. Notice this function relies on an external locking.
330 * \param[in] env LU environment provided by the caller
331 * \param[in] d OSP device
333 * \retval 0 - current pool is good enough, 1 - time to precreate
335 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
336 struct osp_device *d)
338 int window = osp_objs_precreated(env, d);
340 /* don't consider new precreation till OST is healty and
342 return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2 ||
343 d->opd_force_creation) && (d->opd_pre_status == 0));
347 * Check pool of precreated objects
349 * This is protected version of osp_precreate_near_empty_nolock(), check that
352 * \param[in] env LU environment provided by the caller
353 * \param[in] d OSP device
355 * \retval 0 - current pool is good enough, 1 - time to precreate
357 static inline int osp_precreate_near_empty(const struct lu_env *env,
358 struct osp_device *d)
362 if (d->opd_pre == NULL)
365 /* XXX: do we really need locking here? */
366 spin_lock(&d->opd_pre_lock);
367 rc = osp_precreate_near_empty_nolock(env, d);
368 spin_unlock(&d->opd_pre_lock);
373 * Check given sequence is empty
375 * Returns a binary result whether the given sequence has some IDs left
376 * or not. Find the details in osp_fid_end_seq(). This is a lock protected
377 * version of that function.
379 * \param[in] env LU environment provided by the caller
380 * \param[in] osp OSP device
382 * \retval 0 - current sequence has no IDs, 1 - otherwise
384 static inline int osp_create_end_seq(const struct lu_env *env,
385 struct osp_device *osp)
387 struct lu_fid *fid = &osp->opd_pre_used_fid;
390 spin_lock(&osp->opd_pre_lock);
391 rc = osp_fid_end_seq(env, fid);
392 spin_unlock(&osp->opd_pre_lock);
397 * Write FID into into last_oid/last_seq file
399 * The function stores the sequence and the in-sequence id into two dedicated
400 * files. The sync argument can be used to request synchronous commit, so the
401 * function won't return until the updates are committed.
403 * \param[in] env LU environment provided by the caller
404 * \param[in] osp OSP device
405 * \param[in] fid fid where sequence/id is taken
406 * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously
408 * \retval 0 on success
409 * \retval negative negated errno on error
411 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
412 struct lu_fid *fid, int sync)
414 struct osp_thread_info *oti = osp_env_info(env);
415 struct lu_buf *lb_oid = &oti->osi_lb;
416 struct lu_buf *lb_oseq = &oti->osi_lb2;
424 if (osp->opd_storage->dd_rdonly)
427 /* Note: through f_oid is only 32 bits, it will also write 64 bits
428 * for oid to keep compatibility with the previous version. */
430 osp_objid_buf_prep(lb_oid, &oid_off,
431 &oid, osp->opd_index);
433 osp_objseq_buf_prep(lb_oseq, &oseq_off,
434 &fid->f_seq, osp->opd_index);
436 th = dt_trans_create(env, osp->opd_storage);
441 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
442 lb_oid, oid_off, th);
446 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
447 lb_oseq, oseq_off, th);
451 rc = dt_trans_start_local(env, osp->opd_storage, th);
455 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
458 CERROR("%s: can not write to last seq file: rc = %d\n",
459 osp->opd_obd->obd_name, rc);
462 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
465 CERROR("%s: can not write to last seq file: rc = %d\n",
466 osp->opd_obd->obd_name, rc);
470 dt_trans_stop(env, osp->opd_storage, th);
475 * Switch to another sequence
477 * When a current sequence has no available IDs left, OSP has to switch to
478 * another new sequence. OSP requests it using the regular FLDB protocol
479 * and stores synchronously before that is used in precreated. This is needed
480 * to basically have the sequences referenced (not orphaned), otherwise it's
481 * possible that OST has some objects precreated and the clients have data
482 * written to it, but after MDT failover nobody refers those objects and OSP
483 * has no idea that the sequence need cleanup to be done.
484 * While this is very expensive operation, it's supposed to happen very very
485 * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
487 * \param[in] env LU environment provided by the caller
488 * \param[in] osp OSP device
490 * \retval 0 on success
491 * \retval negative negated errno on error
493 static int osp_precreate_rollover_new_seq(struct lu_env *env,
494 struct osp_device *osp)
496 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
497 struct lu_fid *last_fid = &osp->opd_last_used_fid;
501 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
503 CERROR("%s: alloc fid error: rc = %d\n",
504 osp->opd_obd->obd_name, rc);
510 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
511 "fid "DFID", last_fid "DFID"\n", PFID(fid),
514 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
516 CERROR("%s: Can not update oid/seq file: rc = %d\n",
517 osp->opd_obd->obd_name, rc);
521 LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
522 osp->opd_obd->obd_name, fid_seq(last_fid),
524 /* Update last_xxx to the new seq */
525 spin_lock(&osp->opd_pre_lock);
526 osp->opd_last_used_fid = *fid;
527 osp_fid_to_obdid(fid, &osp->opd_last_id);
528 osp->opd_gap_start_fid = *fid;
529 osp->opd_pre_used_fid = *fid;
530 osp->opd_pre_last_created_fid = *fid;
531 spin_unlock(&osp->opd_pre_lock);
537 * Find IDs available in current sequence
539 * The function calculates the highest possible ID and the number of IDs
540 * available in the current sequence OSP is using. The number is limited
541 * artifically by the caller (grow param) and the number of IDs available
542 * in the sequence by nature. The function doesn't require an external
545 * \param[in] env LU environment provided by the caller
546 * \param[in] osp OSP device
547 * \param[in] fid FID the caller wants to start with
548 * \param[in] grow how many the caller wants
549 * \param[out] fid the highest calculated FID
550 * \param[out] grow the number of available IDs calculated
552 * \retval 0 on success, 1 - the sequence is empty
554 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
555 struct lu_fid *fid, int *grow)
557 struct osp_thread_info *osi = osp_env_info(env);
561 if (fid_is_idif(fid)) {
562 struct lu_fid *last_fid;
563 struct ost_id *oi = &osi->osi_oi;
566 spin_lock(&osp->opd_pre_lock);
567 last_fid = &osp->opd_pre_last_created_fid;
568 fid_to_ostid(last_fid, oi);
569 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
570 *grow = end - ostid_id(oi);
571 rc = ostid_set_id(oi, ostid_id(oi) + *grow);
572 spin_unlock(&osp->opd_pre_lock);
574 if (*grow == 0 || rc)
577 ostid_to_fid(fid, oi, osp->opd_index);
581 spin_lock(&osp->opd_pre_lock);
582 *fid = osp->opd_pre_last_created_fid;
584 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
585 *grow = end - fid->f_oid;
586 fid->f_oid += end - fid->f_oid;
587 spin_unlock(&osp->opd_pre_lock);
589 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
590 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
592 return *grow > 0 ? 0 : 1;
596 * Prepare and send precreate RPC
598 * The function finds how many objects should be precreated. Then allocates,
599 * prepares and schedules precreate RPC synchronously. Upon reply the function
600 * wakes up the threads waiting for the new objects on this target. If the
601 * target wasn't able to create all the objects requested, then the next
602 * precreate will be asking for fewer objects (i.e. slow precreate down).
604 * \param[in] env LU environment provided by the caller
605 * \param[in] d OSP device
607 * \retval 0 on success
608 * \retval negative negated errno on error
610 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
612 struct osp_thread_info *oti = osp_env_info(env);
613 struct ptlrpc_request *req;
614 struct obd_import *imp;
615 struct ost_body *body;
617 struct lu_fid *fid = &oti->osi_fid;
620 /* don't precreate new objects till OST healthy and has free space */
621 if (unlikely(d->opd_pre_status)) {
622 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
623 d->opd_obd->obd_name, d->opd_pre_status);
628 * if not connection/initialization is compeleted, ignore
630 imp = d->opd_obd->u.cli.cl_import;
633 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
636 req->rq_request_portal = OST_CREATE_PORTAL;
638 /* Delorphan happens only with a first MDT-OST connect. resend/replay
639 * handles objects creation on reconnects, no need to do delorhpan
643 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
645 ptlrpc_request_free(req);
649 spin_lock(&d->opd_pre_lock);
650 if (d->opd_force_creation)
651 d->opd_pre_create_count = OST_MIN_PRECREATE;
652 else if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
653 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
654 grow = d->opd_pre_create_count;
655 spin_unlock(&d->opd_pre_lock);
657 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
660 *fid = d->opd_pre_last_created_fid;
661 rc = osp_precreate_fids(env, d, fid, &grow);
663 /* Current seq has been used up*/
664 GOTO(out_req, rc = -ENOSPC);
666 if (!osp_is_fid_client(d)) {
667 /* Non-FID client will always send seq 0 because of
669 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
673 fid_to_ostid(fid, &body->oa.o_oi);
674 body->oa.o_valid = OBD_MD_FLGROUP;
676 ptlrpc_request_set_replen(req);
678 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
681 rc = ptlrpc_queue_wait(req);
683 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
686 /* have osp_precreate_reserve() to wait for repeat */
691 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
693 GOTO(out_req, rc = -EPROTO);
695 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
698 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
699 CERROR("%s: precreate fid "DFID" <= local used fid "DFID
700 ": rc = %d\n", d->opd_obd->obd_name,
701 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
702 GOTO(out_req, rc = -ESTALE);
705 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
707 spin_lock(&d->opd_pre_lock);
709 /* the OST has not managed to create all the
710 * objects we asked for */
711 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
712 d->opd_pre_create_slow = 1;
714 /* the OST is able to keep up with the work,
715 * we could consider increasing create_count
716 * next time if needed */
717 d->opd_pre_create_slow = 0;
720 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
721 fid_to_ostid(fid, &body->oa.o_oi);
723 d->opd_pre_last_created_fid = *fid;
724 d->opd_force_creation = false;
725 spin_unlock(&d->opd_pre_lock);
727 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
728 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
729 PFID(&d->opd_pre_last_created_fid));
731 /* now we can wakeup all users awaiting for objects */
732 osp_pre_update_status(d, rc);
734 /* pause to let osp_precreate_reserve to go first */
735 CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_PRECREATE_PAUSE, 2);
737 ptlrpc_req_finished(req);
742 * Get last precreated object from target (OST)
744 * Sends synchronous RPC to the target (OST) to learn the last precreated
745 * object. This later is used to remove all unused objects (cleanup orphan
746 * procedure). Also, the next object after one we got will be used as a
747 * starting point for the new precreates.
749 * \param[in] env LU environment provided by the caller
750 * \param[in] d OSP device
751 * \param[in] update update or not update last used fid
753 * \retval 0 on success
754 * \retval negative negated errno on error
756 static int osp_get_lastfid_from_ost(const struct lu_env *env,
757 struct osp_device *d, bool update)
759 struct ptlrpc_request *req = NULL;
760 struct obd_import *imp;
761 struct lu_fid *last_fid;
766 imp = d->opd_obd->u.cli.cl_import;
769 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
773 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
774 sizeof(KEY_LAST_FID));
776 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
778 ptlrpc_request_free(req);
782 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
783 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
785 req->rq_no_delay = req->rq_no_resend = 1;
786 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
787 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
789 ptlrpc_request_set_replen(req);
791 rc = ptlrpc_queue_wait(req);
793 /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld),
794 * let sysadm sort this * out.
797 ptlrpc_set_import_active(imp, 0);
801 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
802 if (last_fid == NULL) {
803 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
804 GOTO(out, rc = -EPROTO);
807 if (!fid_is_sane(last_fid)) {
808 CERROR("%s: Got insane last_fid "DFID"\n",
809 d->opd_obd->obd_name, PFID(last_fid));
810 GOTO(out, rc = -EPROTO);
813 /* Only update the last used fid, if the OST has objects for
814 * this sequence, i.e. fid_oid > 0 */
815 if (fid_oid(last_fid) > 0 && update)
816 d->opd_last_used_fid = *last_fid;
818 if (fid_oid(last_fid) == 0 &&
819 fid_seq(last_fid) == fid_seq(&d->opd_last_used_fid)) {
820 /* reformatted OST, it requires creation request
821 * to recreate objects
823 d->opd_force_creation = true;
825 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
829 ptlrpc_req_finished(req);
834 * Cleanup orphans on OST
836 * This function is called in a contex of a dedicated thread handling
837 * all the precreation suff. The function waits till local recovery
838 * is complete, then identify all the unreferenced objects (orphans)
839 * using the highest ID referenced by a local and the highest object
840 * precreated by the target. The found range is a subject to removal
841 * using specially flagged RPC. During this process OSP is marked
842 * unavailable for new objects.
844 * \param[in] env LU environment provided by the caller
845 * \param[in] d OSP device
847 * \retval 0 on success
848 * \retval negative negated errno on error
850 static int osp_precreate_cleanup_orphans(struct lu_env *env,
851 struct osp_device *d)
853 struct osp_thread_info *osi = osp_env_info(env);
854 struct lu_fid *last_fid = &osi->osi_fid;
855 struct ptlrpc_request *req = NULL;
856 struct obd_import *imp;
857 struct ost_body *body;
858 int update_status = 0;
865 * Do cleanup orphans only with a first connection, after that
866 * all precreate requests uses resend/replay flags to support OST
867 * failover/reconnect.
869 if (d->opd_cleanup_orphans_done) {
870 rc = osp_get_lastfid_from_ost(env, d, false);
874 * wait for local recovery to finish, so we can cleanup orphans
875 * orphans are all objects since "last used" (assigned), but
876 * there might be objects reserved and in some cases they won't
877 * be used. we can't cleanup them till we're sure they won't be
878 * used. also can't we allow new reservations because they may
879 * end up getting orphans being cleaned up below. so we block
880 * new reservations and wait till all reserved objects either
883 spin_lock(&d->opd_pre_lock);
884 d->opd_pre_recovering = 1;
885 spin_unlock(&d->opd_pre_lock);
887 * The locking above makes sure the opd_pre_reserved check below will
888 * catch all osp_precreate_reserve() calls who find
889 * "!opd_pre_recovering".
891 wait_event_idle(d->opd_pre_waitq,
892 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
893 !d->opd_pre_task || d->opd_got_disconnected);
894 if (!d->opd_pre_task || d->opd_got_disconnected)
895 GOTO(out, rc = -EAGAIN);
897 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
898 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
900 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val);
902 *last_fid = d->opd_last_used_fid;
903 /* The OSP should already get the valid seq now */
904 LASSERT(!fid_is_zero(last_fid));
905 if (fid_oid(&d->opd_last_used_fid) < 2 ||
906 OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
907 /* lastfid looks strange... ask OST */
908 rc = osp_get_lastfid_from_ost(env, d, true);
913 imp = d->opd_obd->u.cli.cl_import;
916 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
918 GOTO(out, rc = -ENOMEM);
920 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
922 ptlrpc_request_free(req);
927 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
929 GOTO(out, rc = -EPROTO);
931 body->oa.o_flags = OBD_FL_DELORPHAN;
932 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
934 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
936 ptlrpc_request_set_replen(req);
938 /* Don't resend the delorphan req */
939 req->rq_no_resend = req->rq_no_delay = 1;
941 rc = ptlrpc_queue_wait(req);
947 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
949 GOTO(out, rc = -EPROTO);
952 * OST provides us with id new pool starts from in body->oa.o_id
954 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
956 spin_lock(&d->opd_pre_lock);
957 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
959 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
960 d->opd_pre_last_created_fid = d->opd_last_used_fid;
962 d->opd_pre_create_count = OST_MIN_PRECREATE;
963 d->opd_pre_last_created_fid = *last_fid;
966 * This empties the pre-creation pool and effectively blocks any new
969 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
970 LUSTRE_DATA_SEQ_MAX_WIDTH);
971 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
972 d->opd_pre_create_slow = 0;
973 spin_unlock(&d->opd_pre_lock);
975 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
976 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
977 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
980 ptlrpc_req_finished(req);
984 * If rc is zero, the pre-creation window should have been emptied.
985 * Since waking up the herd would be useless without pre-created
986 * objects, we defer the signal to osp_precreate_send() in that case.
990 CERROR("%s: cannot cleanup orphans: rc = %d\n",
991 d->opd_obd->obd_name, rc);
992 /* we can't proceed from here, OST seem to
993 * be in a bad shape, better to wait for
994 * a new instance of the server and repeat
995 * from the beginning. notify possible waiters
996 * this OSP isn't quite functional yet */
997 osp_pre_update_status(d, rc);
999 wake_up_all(&d->opd_pre_user_waitq);
1002 spin_lock(&d->opd_pre_lock);
1003 d->opd_pre_recovering = 0;
1004 spin_unlock(&d->opd_pre_lock);
1005 d->opd_cleanup_orphans_done = true;
1012 * Update precreate status using statfs data
1014 * The function decides whether this OSP should be used for new objects.
1015 * IOW, whether this OST is used up or has some free space. Cached statfs
1016 * data is used to make this decision. If the latest result of statfs
1017 * request (rc argument) is not success, then just mark OSP unavailable
1020 * The new statfs data is passed in \a msfs and needs to be stored into
1021 * opd_statfs, but only after the various flags in os_state are set, so
1022 * that the new statfs data is not visible without appropriate flags set.
1023 * As such, there is no need to clear the flags here, since this is called
1024 * with new statfs data, and they should not be cleared if sent from OST.
1026 * Add a bit of hysteresis so this flag isn't continually flapping, and
1027 * ensure that new files don't get extremely fragmented due to only a
1028 * small amount of available space in the filesystem. We want to set
1029 * the ENOSPC/ENOINO flags unconditionally when there is less than the
1030 * reserved size free, and still copy them from the old state when there
1031 * is less than 2*reserved size free space or inodes.
1033 * \param[in] d OSP device
1034 * \param[in] msfs statfs data
1036 static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
1038 u32 old_state = d->opd_statfs.os_state;
1041 /* statfs structure not initialized yet */
1042 if (unlikely(!msfs->os_type))
1045 /* if the low and high watermarks have not been initialized yet */
1046 if (unlikely(d->opd_reserved_mb_high == 0 &&
1047 d->opd_reserved_mb_low == 0)) {
1048 /* Use ~0.1% by default to disable object allocation,
1049 * and ~0.2% to enable, size in MB, set both watermark
1051 spin_lock(&d->opd_pre_lock);
1052 if (d->opd_reserved_mb_high == 0 &&
1053 d->opd_reserved_mb_low == 0) {
1054 d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
1055 msfs->os_blocks) >> 20;
1056 if (d->opd_reserved_mb_low < 1)
1057 d->opd_reserved_mb_low = 1;
1058 d->opd_reserved_mb_high =
1059 (d->opd_reserved_mb_low << 1) + 1;
1061 spin_unlock(&d->opd_pre_lock);
1064 if (unlikely(d->opd_reserved_ino_high == 0 &&
1065 d->opd_reserved_ino_low == 0)) {
1066 /* Use ~0.0001% by default to disallow distributed transactions,
1067 * and ~0.0002% to allow, set both watermark
1069 spin_lock(&d->opd_pre_lock);
1070 if (d->opd_reserved_ino_high == 0 &&
1071 d->opd_reserved_ino_low == 0) {
1072 d->opd_reserved_ino_low = msfs->os_ffree >> 20;
1073 if (d->opd_reserved_ino_low < 32)
1074 d->opd_reserved_ino_low = 32;
1075 d->opd_reserved_ino_high =
1076 (d->opd_reserved_ino_low << 1) + 1;
1078 spin_unlock(&d->opd_pre_lock);
1081 available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
1082 if (msfs->os_ffree < d->opd_reserved_ino_low)
1083 msfs->os_state |= OS_STATFS_ENOINO;
1084 else if (msfs->os_ffree <= d->opd_reserved_ino_high)
1085 msfs->os_state |= old_state & OS_STATFS_ENOINO;
1086 /* else don't clear flags in new msfs->os_state sent from OST */
1088 if (available_mb < d->opd_reserved_mb_low)
1089 msfs->os_state |= OS_STATFS_ENOSPC;
1090 else if (available_mb <= d->opd_reserved_mb_high)
1091 msfs->os_state |= old_state & OS_STATFS_ENOSPC;
1092 /* else don't clear flags in new msfs->os_state sent from OST */
1095 "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
1096 d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
1097 msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
1098 msfs->os_files, msfs->os_ffree, msfs->os_state,
1099 d->opd_pre ? d->opd_pre_status : 0);
1104 if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1105 d->opd_pre_status = -ENOSPC;
1106 if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)))
1107 CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
1108 d->opd_obd->obd_name, msfs->os_state,
1110 CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
1111 atomic_read(&d->opd_sync_changes),
1112 atomic_read(&d->opd_sync_rpcs_in_progress));
1113 } else if (old_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) {
1114 d->opd_pre_status = 0;
1115 spin_lock(&d->opd_pre_lock);
1116 d->opd_pre_create_slow = 0;
1117 d->opd_pre_create_count = OST_MIN_PRECREATE;
1118 spin_unlock(&d->opd_pre_lock);
1119 wake_up(&d->opd_pre_waitq);
1122 "%s: available: state=%x: rc = %d\n",
1123 d->opd_obd->obd_name, msfs->os_state,
1126 /* we only get here if rc == 0 in the caller */
1127 d->opd_pre_status = 0;
1130 /* Object precreation skipped on OST if manually disabled */
1131 if (d->opd_pre_max_create_count == 0)
1132 msfs->os_state |= OS_STATFS_NOPRECREATE;
1133 /* else don't clear flags in new msfs->os_state sent from OST */
1136 /* copy only new statfs state to make it visible to MDS threads */
1137 if (&d->opd_statfs != msfs)
1138 d->opd_statfs = *msfs;
1142 * Initialize FID for precreation
1144 * For a just created new target, a new sequence should be taken.
1145 * The function checks there is no IDIF in use (if the target was
1146 * added with the older version of Lustre), then requests a new
1147 * sequence from FLDB using the regular protocol. Then this new
1148 * sequence is stored on a persisten storage synchronously to prevent
1149 * possible object leakage (for the detail see the description for
1150 * osp_precreate_rollover_new_seq()).
1152 * \param[in] osp OSP device
1154 * \retval 0 on success
1155 * \retval negative negated errno on error
1157 int osp_init_pre_fid(struct osp_device *osp)
1160 struct osp_thread_info *osi;
1161 struct lu_client_seq *cli_seq;
1162 struct lu_fid *last_fid;
1166 LASSERT(osp->opd_pre != NULL);
1168 /* Let's check if the current last_seq/fid is valid,
1169 * otherwise request new sequence from the controller */
1170 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1171 /* Non-MDT0 can only use normal sequence for
1173 if (fid_is_norm(&osp->opd_last_used_fid))
1176 /* Initially MDT0 will start with IDIF, after
1177 * that it will request new sequence from the
1179 if (fid_is_idif(&osp->opd_last_used_fid) ||
1180 fid_is_norm(&osp->opd_last_used_fid))
1184 if (!fid_is_zero(&osp->opd_last_used_fid))
1185 CWARN("%s: invalid last used fid "DFID
1186 ", try to get new sequence.\n",
1187 osp->opd_obd->obd_name,
1188 PFID(&osp->opd_last_used_fid));
1190 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1192 CERROR("%s: init env error: rc = %d\n",
1193 osp->opd_obd->obd_name, rc);
1197 osi = osp_env_info(&env);
1198 last_fid = &osi->osi_fid;
1200 /* For a freshed fs, it will allocate a new sequence first */
1201 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1202 cli_seq = osp->opd_obd->u.cli.cl_seq;
1203 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1205 CERROR("%s: alloc fid error: rc = %d\n",
1206 osp->opd_obd->obd_name, rc);
1210 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1212 last_fid->f_oid = 1;
1213 last_fid->f_ver = 0;
1215 spin_lock(&osp->opd_pre_lock);
1216 osp->opd_last_used_fid = *last_fid;
1217 osp->opd_pre_used_fid = *last_fid;
1218 osp->opd_pre_last_created_fid = *last_fid;
1219 spin_unlock(&osp->opd_pre_lock);
1220 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1222 CERROR("%s: write fid error: rc = %d\n",
1223 osp->opd_obd->obd_name, rc);
1232 struct osp_device *opta_dev;
1233 struct lu_env opta_env;
1234 struct completion *opta_started;
1237 * The core of precreate functionality
1239 * The function implements the main precreation loop. Basically it
1240 * involves connecting to the target, precerate FID initialization,
1241 * identifying and removing orphans, then serving precreation. As
1242 * part of the latter, the thread is responsible for statfs data
1243 * updates. The precreation is mostly driven by another threads
1244 * asking for new OST objects - those askers wake the thread when
1245 * the number of precreated objects reach low watermark.
1246 * After a disconnect, the sequence above repeats. This is keep going
1247 * until the thread is requested to stop.
1249 * \param[in] _arg private data the thread (OSP device to handle)
1251 * \retval 0 on success
1252 * \retval negative negated errno on error
1254 static int osp_precreate_thread(void *_args)
1256 struct opt_args *args = _args;
1257 struct osp_device *d = args->opta_dev;
1258 struct lu_env *env = &args->opta_env;
1263 complete(args->opta_started);
1264 while (!kthread_should_stop()) {
1266 * need to be connected to OST
1268 while (!kthread_should_stop()) {
1269 if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
1270 d->opd_imp_connected &&
1271 !d->opd_got_disconnected)
1273 wait_event_idle(d->opd_pre_waitq,
1274 kthread_should_stop() ||
1275 d->opd_new_connection);
1277 if (!d->opd_new_connection)
1280 OBD_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
1282 d->opd_new_connection = 0;
1283 d->opd_got_disconnected = 0;
1287 if (kthread_should_stop())
1291 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1292 /* Sigh, fid client is not ready yet */
1293 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1296 /* Init fid for osp_precreate if necessary */
1297 rc = osp_init_pre_fid(d);
1299 class_export_put(d->opd_exp);
1300 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1301 CERROR("%s: init pre fid error: rc = %d\n",
1302 d->opd_obd->obd_name, rc);
1307 if (osp_statfs_update(env, d)) {
1308 if (wait_event_idle_timeout(d->opd_pre_waitq,
1309 kthread_should_stop(),
1310 cfs_time_seconds(5)) == 0)
1311 l_wait_event_abortable(
1313 kthread_should_stop());
1319 * Clean up orphans or recreate missing objects.
1321 rc = osp_precreate_cleanup_orphans(env, d);
1323 schedule_timeout_interruptible(cfs_time_seconds(1));
1329 * connected, can handle precreates now
1331 while (!kthread_should_stop()) {
1332 wait_event_idle(d->opd_pre_waitq,
1333 kthread_should_stop() ||
1334 osp_precreate_near_empty(env, d) ||
1335 osp_statfs_need_update(d) ||
1336 d->opd_got_disconnected);
1338 if (kthread_should_stop())
1341 /* something happened to the connection
1342 * have to start from the beginning */
1343 if (d->opd_got_disconnected)
1346 if (osp_statfs_need_update(d))
1347 if (osp_statfs_update(env, d))
1350 if (d->opd_pre == NULL)
1353 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
1354 d->opd_pre_recovering = 1;
1358 /* To avoid handling different seq in precreate/orphan
1359 * cleanup, it will hold precreate until current seq is
1361 if (unlikely(osp_precreate_end_seq(env, d) &&
1362 !osp_create_end_seq(env, d)))
1365 if (unlikely(osp_precreate_end_seq(env, d) &&
1366 osp_create_end_seq(env, d))) {
1367 LCONSOLE_INFO("%s:%#llx is used up."
1368 " Update to new seq\n",
1369 d->opd_obd->obd_name,
1370 fid_seq(&d->opd_pre_last_created_fid));
1371 rc = osp_precreate_rollover_new_seq(env, d);
1376 if (osp_precreate_near_empty(env, d)) {
1377 rc = osp_precreate_send(env, d);
1378 /* osp_precreate_send() sets opd_pre_status
1379 * in case of error, that prevent the using of
1381 if (rc < 0 && rc != -ENOSPC &&
1382 rc != -ETIMEDOUT && rc != -ENOTCONN)
1383 CERROR("%s: cannot precreate objects:"
1385 d->opd_obd->obd_name, rc);
1397 * Check when to stop to wait for precreate objects.
1399 * The caller wanting a new OST object can't wait undefinitely. The
1400 * function checks for few conditions including available new OST
1401 * objects, disconnected OST, lack of space with no pending destroys,
1402 * etc. IOW, it checks whether the current OSP state is good to keep
1403 * waiting or it's better to give up.
1405 * \param[in] env LU environment provided by the caller
1406 * \param[in] d OSP device
1408 * \retval 0 - keep waiting, 1 - no luck
1410 static int osp_precreate_ready_condition(const struct lu_env *env,
1411 struct osp_device *d)
1413 /* Bail out I/O fails to OST */
1414 if (d->opd_pre_status != 0 &&
1415 d->opd_pre_status != -EAGAIN &&
1416 d->opd_pre_status != -ENODEV &&
1417 d->opd_pre_status != -ENOTCONN &&
1418 d->opd_pre_status != -ENOSPC) {
1420 if (d->opd_pre_status != -EIO)
1421 CERROR("%s: precreate failed opd_pre_status %d\n",
1422 d->opd_obd->obd_name, d->opd_pre_status);
1426 if (d->opd_pre_recovering)
1429 /* ready if got enough precreated objects */
1430 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1431 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1434 /* ready if OST reported no space and no destroys in progress */
1435 if (atomic_read(&d->opd_sync_changes) +
1436 atomic_read(&d->opd_sync_rpcs_in_progress) == 0 &&
1437 d->opd_pre_status == -ENOSPC)
1444 * Reserve object in precreate pool
1446 * When the caller wants to create a new object on this target (target
1447 * represented by the given OSP), it should declare this intention using
1448 * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1449 * to reserve an object in the existing precreated pool or wait up to
1450 * obd_timeout for the available object to appear in the pool (a dedicated
1451 * thread will be doing real precreation in background). The object can be
1452 * consumed later with osp_precreate_get_fid() or be released with call to
1453 * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1454 * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1455 * If the space on the target is short and there is a pending object destroy,
1456 * then the function forces local commit to speedup space release (see
1457 * osp_sync.c for the details).
1459 * \param[in] env LU environment provided by the caller
1460 * \param[in] d OSP device
1462 * \retval 0 on success
1463 * \retval -ENOSPC when no space on OST
1464 * \retval -EAGAIN try later, slow precreation in progress
1465 * \retval -EIO when no access to OST
1467 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
1470 time64_t expire = ktime_get_seconds() + obd_timeout;
1471 int precreated, rc, synced = 0;
1475 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1476 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1477 PFID(&d->opd_pre_used_fid));
1479 /* opd_pre_max_create_count 0 to not use specified OST. */
1480 if (d->opd_pre_max_create_count == 0)
1485 * - preallocation is done
1486 * - no free space expected soon
1487 * - can't connect to OST for too long (obd_timeout)
1488 * - OST can allocate fid sequence.
1490 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1491 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1494 * increase number of precreations
1496 precreated = osp_objs_precreated(env, d);
1497 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1498 d->opd_pre_create_slow == 0 &&
1499 precreated <= (d->opd_pre_create_count / 4 + 1)) {
1500 spin_lock(&d->opd_pre_lock);
1501 d->opd_pre_create_slow = 1;
1502 d->opd_pre_create_count *= 2;
1503 spin_unlock(&d->opd_pre_lock);
1506 spin_lock(&d->opd_pre_lock);
1507 precreated = osp_objs_precreated(env, d);
1508 if (precreated > d->opd_pre_reserved &&
1509 !d->opd_pre_recovering) {
1510 d->opd_pre_reserved++;
1511 spin_unlock(&d->opd_pre_lock);
1514 /* XXX: don't wake up if precreation is in progress */
1515 if (osp_precreate_near_empty_nolock(env, d) &&
1516 !osp_precreate_end_seq_nolock(env, d))
1517 wake_up(&d->opd_pre_waitq);
1521 spin_unlock(&d->opd_pre_lock);
1524 * all precreated objects have been used and no-space
1525 * status leave us no chance to succeed very soon
1526 * but if there is destroy in progress, then we should
1527 * wait till that is done - some space might be released
1529 if (unlikely(rc == -ENOSPC)) {
1530 if (atomic_read(&d->opd_sync_changes) && synced == 0) {
1531 /* force local commit to release space */
1532 dt_commit_async(env, d->opd_storage);
1533 osp_sync_check_for_work(d);
1536 if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
1537 /* just wait till destroys are done
1538 * see wait_event_idle_timeout() below
1541 if (atomic_read(&d->opd_sync_changes) +
1542 atomic_read(&d->opd_sync_rpcs_in_progress) == 0) {
1543 /* no hope for free space */
1548 /* XXX: don't wake up if precreation is in progress */
1549 wake_up(&d->opd_pre_waitq);
1551 if (ktime_get_seconds() >= expire) {
1557 LASSERT(d->opd_pre);
1562 CDEBUG(D_INFO, "%s: Sleeping on objects\n",
1563 d->opd_obd->obd_name);
1564 if (wait_event_idle_timeout(
1565 d->opd_pre_user_waitq,
1566 osp_precreate_ready_condition(env, d),
1567 cfs_time_seconds(obd_timeout)) == 0) {
1569 "%s: slow creates, last="DFID", next="DFID", "
1570 "reserved=%llu, sync_changes=%u, "
1571 "sync_rpcs_in_progress=%d, status=%d\n",
1572 d->opd_obd->obd_name,
1573 PFID(&d->opd_pre_last_created_fid),
1574 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1575 atomic_read(&d->opd_sync_changes),
1576 atomic_read(&d->opd_sync_rpcs_in_progress),
1579 CDEBUG(D_INFO, "%s: Waked up, status=%d\n",
1580 d->opd_obd->obd_name, d->opd_pre_status);
1588 * Get a FID from precreation pool
1590 * The function is a companion for osp_precreate_reserve() - it assigns
1591 * a specific FID from the precreate. The function should be called only
1592 * if the call to osp_precreate_reserve() was successful. The function
1593 * updates a local storage to remember the highest object ID referenced
1594 * by the node in the given sequence.
1596 * A very importan details: this is supposed to be called once the
1597 * transaction is started, so on-disk update will be atomic with the
1598 * data (like LOVEA) refering this object. Then the object won't be leaked:
1599 * either it's referenced by the committed transaction or it's a subject
1600 * to the orphan cleanup procedure.
1602 * \param[in] env LU environment provided by the caller
1603 * \param[in] d OSP device
1604 * \param[out] fid generated FID
1606 * \retval 0 on success
1607 * \retval negative negated errno on error
1609 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1612 struct lu_fid *pre_used_fid = &d->opd_pre_used_fid;
1613 /* grab next id from the pool */
1614 spin_lock(&d->opd_pre_lock);
1616 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1617 &d->opd_pre_last_created_fid) < 0,
1618 "next fid "DFID" last created fid "DFID"\n",
1619 PFID(&d->opd_pre_used_fid),
1620 PFID(&d->opd_pre_last_created_fid));
1623 * When sequence is used up, new one should be allocated in
1624 * osp_precreate_rollover_new_seq. So ASSERT here to avoid
1627 LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0,
1628 "next fid "DFID" last created fid "DFID"\n",
1629 PFID(&d->opd_pre_used_fid),
1630 PFID(&d->opd_pre_last_created_fid));
1631 /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */
1632 if (fid_is_idif(pre_used_fid) &&
1633 unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH))
1634 pre_used_fid->f_seq++;
1636 d->opd_pre_used_fid.f_oid++;
1637 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1638 d->opd_pre_reserved--;
1640 * last_used_id must be changed along with getting new id otherwise
1641 * we might miscalculate gap causing object loss or leak
1643 osp_update_last_fid(d, fid);
1644 spin_unlock(&d->opd_pre_lock);
1647 * probably main thread suspended orphan cleanup till
1648 * all reservations are released, see comment in
1649 * osp_precreate_thread() just before orphan cleanup
1651 if (unlikely(d->opd_pre_reserved == 0 &&
1652 (d->opd_pre_recovering || d->opd_pre_status)))
1653 wake_up(&d->opd_pre_waitq);
1659 * Set size regular attribute on an object
1661 * When a striping is created late, it's possible that size is already
1662 * initialized on the file. Then the new striping should inherit size
1663 * from the file. The function sets size on the object using the regular
1664 * protocol (OST_PUNCH).
1665 * XXX: should be re-implemented using OUT ?
1667 * \param[in] env LU environment provided by the caller
1668 * \param[in] dt object
1669 * \param[in] size size to set.
1671 * \retval 0 on success
1672 * \retval negative negated errno on error
1674 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1677 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1678 struct ptlrpc_request *req = NULL;
1679 struct obd_import *imp;
1680 struct ost_body *body;
1681 struct obdo *oa = NULL;
1686 imp = d->opd_obd->u.cli.cl_import;
1689 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1693 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1695 ptlrpc_request_free(req);
1700 * XXX: decide how do we do here with resend
1701 * if we don't resend, then client may see wrong file size
1702 * if we do resend, then MDS thread can get stuck for quite long
1703 * and if we don't resend, then client will also get -EAGAIN !!
1704 * (see LU-7975 and sanity/test_27F use cases)
1705 * but let's decide not to resend/delay this truncate request to OST
1706 * and allow Client to decide to resend, in a less agressive way from
1707 * after_reply(), by returning -EINPROGRESS instead of
1708 * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
1709 * end of this routine
1711 req->rq_no_resend = req->rq_no_delay = 1;
1713 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1714 ptlrpc_at_set_req_timeout(req);
1718 GOTO(out, rc = -ENOMEM);
1720 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1723 oa->o_blocks = OBD_OBJECT_EOF;
1724 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1725 OBD_MD_FLID | OBD_MD_FLGROUP;
1727 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1729 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1731 /* XXX: capa support? */
1732 /* osc_pack_capa(req, body, capa); */
1734 ptlrpc_request_set_replen(req);
1736 rc = ptlrpc_queue_wait(req);
1738 /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
1739 * since we have decided not to resend/delay, but this could
1740 * lead to wrong size to be seen at Client side and even process
1741 * trying to open to exit/fail if not itself handling -EAGAIN.
1742 * So it should be better to return -EINPROGRESS instead and
1743 * leave the decision to resend at Client side in after_reply()
1745 if (rc == -EAGAIN) {
1747 CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1748 "-EWOULDBLOCK/-EAGAIN to allow Client to "
1751 CERROR("can't punch object: %d\n", rc);
1755 ptlrpc_req_finished(req);
1762 * Initialize precreation functionality of OSP
1764 * Prepares all the internal structures and starts the precreate thread
1766 * \param[in] d OSP device
1768 * \retval 0 on success
1769 * \retval negative negated errno on error
1771 int osp_init_precreate(struct osp_device *d)
1775 OBD_ALLOC_PTR(d->opd_pre);
1776 if (d->opd_pre == NULL)
1779 /* initially precreation isn't ready */
1780 init_waitqueue_head(&d->opd_pre_user_waitq);
1781 d->opd_pre_status = -EAGAIN;
1782 fid_zero(&d->opd_pre_used_fid);
1783 d->opd_pre_used_fid.f_oid = 1;
1784 fid_zero(&d->opd_pre_last_created_fid);
1785 d->opd_pre_last_created_fid.f_oid = 1;
1787 d->opd_pre_reserved = 0;
1788 d->opd_got_disconnected = 1;
1789 d->opd_pre_create_slow = 0;
1790 d->opd_pre_create_count = OST_MIN_PRECREATE;
1791 d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1792 d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1793 d->opd_reserved_mb_high = 0;
1794 d->opd_reserved_mb_low = 0;
1795 d->opd_cleanup_orphans_done = false;
1796 d->opd_force_creation = false;
1802 * Finish precreate functionality of OSP
1805 * Asks all the activity (the thread, update timer) to stop, then
1806 * wait till that is done.
1808 * \param[in] d OSP device
1810 void osp_precreate_fini(struct osp_device *d)
1814 if (d->opd_pre == NULL)
1817 OBD_FREE_PTR(d->opd_pre);
1823 int osp_init_statfs(struct osp_device *d)
1825 struct task_struct *task;
1826 struct opt_args *args;
1827 DECLARE_COMPLETION_ONSTACK(started);
1832 spin_lock_init(&d->opd_pre_lock);
1833 init_waitqueue_head(&d->opd_pre_waitq);
1836 * Initialize statfs-related things
1838 d->opd_statfs_maxage = 5; /* defaultupdate interval */
1839 d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(),
1840 1000 * NSEC_PER_SEC);
1841 CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n",
1843 ktime_to_ns(d->opd_statfs_fresh_till));
1844 cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb,
1845 (unsigned long)d, 0);
1847 if (d->opd_storage->dd_rdonly)
1850 OBD_ALLOC_PTR(args);
1854 args->opta_started = &started;
1855 rc = lu_env_init(&args->opta_env,
1856 d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1858 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1865 * start thread handling precreation and statfs updates
1867 task = kthread_create(osp_precreate_thread, args,
1868 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1870 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1871 lu_env_fini(&args->opta_env);
1873 RETURN(PTR_ERR(task));
1875 d->opd_pre_task = task;
1876 wake_up_process(task);
1877 wait_for_completion(&started);
1882 void osp_statfs_fini(struct osp_device *d)
1884 struct task_struct *task = d->opd_pre_task;
1887 del_timer(&d->opd_statfs_timer);
1889 d->opd_pre_task = NULL;