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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osp/osp_sync.c
38 * Lustre OST Proxy Device
40 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
41 * Author: Mikhail Pershin <mike.pershin@intel.com>
42 * Author: Di Wang <di.wang@intel.com>
45 #define DEBUG_SUBSYSTEM S_MDS
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_object_create() returns an error
62 static inline int osp_statfs_need_update(struct osp_device *d)
64 return !cfs_time_before(cfs_time_current(),
65 d->opd_statfs_fresh_till);
69 * OSP tries to maintain pool of available objects so that calls to create
70 * objects don't block most of time
72 * each time OSP gets connected to OST, we should start from precreation cleanup
74 static inline bool osp_precreate_running(struct osp_device *d)
76 return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
79 static inline bool osp_precreate_stopped(struct osp_device *d)
81 return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
84 static void osp_statfs_timer_cb(unsigned long _d)
86 struct osp_device *d = (struct osp_device *) _d;
89 if (d->opd_pre != NULL && osp_precreate_running(d))
90 wake_up(&d->opd_pre_waitq);
93 static int osp_statfs_interpret(const struct lu_env *env,
94 struct ptlrpc_request *req,
95 union ptlrpc_async_args *aa, int rc)
97 struct obd_import *imp = req->rq_import;
98 struct obd_statfs *msfs;
103 aa = ptlrpc_req_async_args(req);
104 d = aa->pointer_arg[0];
110 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
112 GOTO(out, rc = -EPROTO);
114 d->opd_statfs = *msfs;
116 osp_pre_update_status(d, rc);
118 /* schedule next update */
119 d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
120 cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
121 d->opd_statfs_update_in_progress = 0;
123 CDEBUG(D_CACHE, "updated statfs %p\n", d);
127 /* couldn't update statfs, try again as soon as possible */
128 if (d->opd_pre != NULL && osp_precreate_running(d))
129 wake_up(&d->opd_pre_waitq);
131 if (req->rq_import_generation == imp->imp_generation)
132 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
133 d->opd_obd->obd_name, rc);
137 static int osp_statfs_update(struct osp_device *d)
139 struct ptlrpc_request *req;
140 struct obd_import *imp;
141 union ptlrpc_async_args *aa;
146 CDEBUG(D_CACHE, "going to update statfs\n");
148 imp = d->opd_obd->u.cli.cl_import;
151 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
155 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
157 ptlrpc_request_free(req);
160 ptlrpc_request_set_replen(req);
161 req->rq_request_portal = OST_CREATE_PORTAL;
162 ptlrpc_at_set_req_timeout(req);
164 req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
165 aa = ptlrpc_req_async_args(req);
166 aa->pointer_arg[0] = d;
169 * no updates till reply
171 cfs_timer_disarm(&d->opd_statfs_timer);
172 d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
173 d->opd_statfs_update_in_progress = 1;
175 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
181 * XXX: there might be a case where removed object(s) do not add free
182 * space (empty object). if the number of such deletions is high, then
183 * we can start to update statfs too often - a rpc storm
184 * TODO: some throttling is needed
186 void osp_statfs_need_now(struct osp_device *d)
188 if (!d->opd_statfs_update_in_progress) {
190 * if current status is -ENOSPC (lack of free space on OST)
191 * then we should poll OST immediately once object destroy
194 d->opd_statfs_fresh_till = cfs_time_shift(-1);
195 cfs_timer_disarm(&d->opd_statfs_timer);
196 wake_up(&d->opd_pre_waitq);
200 static inline int osp_objs_precreated(const struct lu_env *env,
201 struct osp_device *osp)
203 return osp_fid_diff(&osp->opd_pre_last_created_fid,
204 &osp->opd_pre_used_fid);
207 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
208 struct osp_device *d)
210 int window = osp_objs_precreated(env, d);
212 /* don't consider new precreation till OST is healty and
214 return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) &&
215 (d->opd_pre_status == 0));
218 static inline int osp_precreate_near_empty(const struct lu_env *env,
219 struct osp_device *d)
223 /* XXX: do we really need locking here? */
224 spin_lock(&d->opd_pre_lock);
225 rc = osp_precreate_near_empty_nolock(env, d);
226 spin_unlock(&d->opd_pre_lock);
230 static inline int osp_create_end_seq(const struct lu_env *env,
231 struct osp_device *osp)
233 struct lu_fid *fid = &osp->opd_pre_used_fid;
236 spin_lock(&osp->opd_pre_lock);
237 rc = osp_fid_end_seq(env, fid);
238 spin_unlock(&osp->opd_pre_lock);
243 * Write fid into last_oid/last_seq file.
245 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
246 struct lu_fid *fid, int sync)
248 struct osp_thread_info *oti = osp_env_info(env);
249 struct lu_buf *lb_oid = &oti->osi_lb;
250 struct lu_buf *lb_oseq = &oti->osi_lb2;
257 /* Note: through f_oid is only 32bits, it will also write
258 * 64 bits for oid to keep compatiblity with the previous
260 lb_oid->lb_buf = &fid->f_oid;
261 lb_oid->lb_len = sizeof(obd_id);
262 oid_off = sizeof(obd_id) * osp->opd_index;
264 lb_oseq->lb_buf = &fid->f_seq;
265 lb_oseq->lb_len = sizeof(obd_id);
266 oseq_off = sizeof(obd_id) * osp->opd_index;
268 th = dt_trans_create(env, osp->opd_storage);
273 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
274 lb_oid, oid_off, th);
278 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
279 lb_oseq, oseq_off, th);
283 rc = dt_trans_start_local(env, osp->opd_storage, th);
287 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
290 CERROR("%s: can not write to last seq file: rc = %d\n",
291 osp->opd_obd->obd_name, rc);
294 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
297 CERROR("%s: can not write to last seq file: rc = %d\n",
298 osp->opd_obd->obd_name, rc);
302 dt_trans_stop(env, osp->opd_storage, th);
306 int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp)
308 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
309 struct lu_fid *last_fid = &osp->opd_last_used_fid;
313 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
315 CERROR("%s: alloc fid error: rc = %d\n",
316 osp->opd_obd->obd_name, rc);
322 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
323 "fid "DFID", last_fid "DFID"\n", PFID(fid),
326 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
328 CERROR("%s: Can not update oid/seq file: rc = %d\n",
329 osp->opd_obd->obd_name, rc);
333 LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n",
334 osp->opd_obd->obd_name, fid_seq(last_fid),
336 /* Update last_xxx to the new seq */
337 spin_lock(&osp->opd_pre_lock);
338 osp->opd_last_used_fid = *fid;
339 osp->opd_gap_start_fid = *fid;
340 osp->opd_pre_used_fid = *fid;
341 osp->opd_pre_last_created_fid = *fid;
342 spin_unlock(&osp->opd_pre_lock);
348 * alloc fids for precreation.
349 * rc = 0 Success, @grow is the count of real allocation.
350 * rc = 1 Current seq is used up.
351 * rc < 0 Other error.
353 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
354 struct lu_fid *fid, int *grow)
356 struct osp_thread_info *osi = osp_env_info(env);
360 if (fid_is_idif(fid)) {
361 struct lu_fid *last_fid;
362 struct ost_id *oi = &osi->osi_oi;
364 spin_lock(&osp->opd_pre_lock);
365 last_fid = &osp->opd_pre_last_created_fid;
366 fid_to_ostid(last_fid, oi);
367 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
368 *grow = end - ostid_id(oi);
369 ostid_set_id(oi, ostid_id(oi) + *grow);
370 spin_unlock(&osp->opd_pre_lock);
375 ostid_to_fid(fid, oi, osp->opd_index);
379 spin_lock(&osp->opd_pre_lock);
380 *fid = osp->opd_pre_last_created_fid;
382 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
383 *grow = end - fid->f_oid;
384 fid->f_oid += end - fid->f_oid;
385 spin_unlock(&osp->opd_pre_lock);
387 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
388 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
390 return *grow > 0 ? 0 : 1;
393 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
395 struct osp_thread_info *oti = osp_env_info(env);
396 struct ptlrpc_request *req;
397 struct obd_import *imp;
398 struct ost_body *body;
400 struct lu_fid *fid = &oti->osi_fid;
403 /* don't precreate new objects till OST healthy and has free space */
404 if (unlikely(d->opd_pre_status)) {
405 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
406 d->opd_obd->obd_name, d->opd_pre_status);
411 * if not connection/initialization is compeleted, ignore
413 imp = d->opd_obd->u.cli.cl_import;
416 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
419 req->rq_request_portal = OST_CREATE_PORTAL;
420 /* we should not resend create request - anyway we will have delorphan
421 * and kill these objects */
422 req->rq_no_delay = req->rq_no_resend = 1;
424 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
426 ptlrpc_request_free(req);
430 spin_lock(&d->opd_pre_lock);
431 if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
432 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
433 grow = d->opd_pre_grow_count;
434 spin_unlock(&d->opd_pre_lock);
436 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
439 *fid = d->opd_pre_last_created_fid;
440 rc = osp_precreate_fids(env, d, fid, &grow);
442 /* Current seq has been used up*/
443 if (!osp_is_fid_client(d)) {
444 osp_pre_update_status(d, -ENOSPC);
447 wake_up(&d->opd_pre_waitq);
451 if (!osp_is_fid_client(d)) {
452 /* Non-FID client will always send seq 0 because of
454 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
458 fid_to_ostid(fid, &body->oa.o_oi);
459 body->oa.o_valid = OBD_MD_FLGROUP;
461 ptlrpc_request_set_replen(req);
463 rc = ptlrpc_queue_wait(req);
465 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
469 LASSERT(req->rq_transno == 0);
471 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
473 GOTO(out_req, rc = -EPROTO);
475 ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
476 if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
477 CERROR("%s: precreate fid "DFID" < local used fid "DFID
478 ": rc = %d\n", d->opd_obd->obd_name,
479 PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
480 GOTO(out_req, rc = -ESTALE);
483 diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
485 spin_lock(&d->opd_pre_lock);
487 /* the OST has not managed to create all the
488 * objects we asked for */
489 d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
490 d->opd_pre_grow_slow = 1;
492 /* the OST is able to keep up with the work,
493 * we could consider increasing grow_count
494 * next time if needed */
495 d->opd_pre_grow_slow = 0;
498 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
499 fid_to_ostid(fid, &body->oa.o_oi);
501 d->opd_pre_last_created_fid = *fid;
502 spin_unlock(&d->opd_pre_lock);
504 CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
505 d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
506 PFID(&d->opd_pre_last_created_fid));
508 /* now we can wakeup all users awaiting for objects */
509 osp_pre_update_status(d, rc);
510 wake_up(&d->opd_pre_user_waitq);
512 ptlrpc_req_finished(req);
516 static int osp_get_lastfid_from_ost(const struct lu_env *env,
517 struct osp_device *d)
519 struct ptlrpc_request *req = NULL;
520 struct obd_import *imp;
521 struct lu_fid *last_fid;
526 imp = d->opd_obd->u.cli.cl_import;
529 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
533 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
534 sizeof(KEY_LAST_FID));
536 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
538 ptlrpc_request_free(req);
542 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
543 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
545 req->rq_no_delay = req->rq_no_resend = 1;
546 last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
547 fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
549 ptlrpc_request_set_replen(req);
551 rc = ptlrpc_queue_wait(req);
553 /* bad-bad OST.. let sysadm sort this out */
554 if (rc == -ENOTSUPP) {
555 CERROR("%s: server does not support FID: rc = %d\n",
556 d->opd_obd->obd_name, -ENOTSUPP);
558 ptlrpc_set_import_active(imp, 0);
562 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
563 if (last_fid == NULL) {
564 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
565 GOTO(out, rc = -EPROTO);
568 if (!fid_is_sane(last_fid)) {
569 CERROR("%s: Got insane last_fid "DFID"\n",
570 d->opd_obd->obd_name, PFID(last_fid));
571 GOTO(out, rc = -EPROTO);
574 /* Only update the last used fid, if the OST has objects for
575 * this sequence, i.e. fid_oid > 0 */
576 if (fid_oid(last_fid) > 0)
577 d->opd_last_used_fid = *last_fid;
579 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
583 ptlrpc_req_finished(req);
588 * asks OST to clean precreate orphans
589 * and gets next id for new objects
591 static int osp_precreate_cleanup_orphans(struct lu_env *env,
592 struct osp_device *d)
594 struct osp_thread_info *osi = osp_env_info(env);
595 struct lu_fid *last_fid = &osi->osi_fid;
596 struct ptlrpc_request *req = NULL;
597 struct obd_import *imp;
598 struct ost_body *body;
599 struct l_wait_info lwi = { 0 };
600 int update_status = 0;
607 * wait for local recovery to finish, so we can cleanup orphans
608 * orphans are all objects since "last used" (assigned), but
609 * there might be objects reserved and in some cases they won't
610 * be used. we can't cleanup them till we're sure they won't be
611 * used. also can't we allow new reservations because they may
612 * end up getting orphans being cleaned up below. so we block
613 * new reservations and wait till all reserved objects either
616 spin_lock(&d->opd_pre_lock);
617 d->opd_pre_recovering = 1;
618 spin_unlock(&d->opd_pre_lock);
620 * The locking above makes sure the opd_pre_reserved check below will
621 * catch all osp_precreate_reserve() calls who find
622 * "!opd_pre_recovering".
624 l_wait_event(d->opd_pre_waitq,
625 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
626 !osp_precreate_running(d) || d->opd_got_disconnected,
628 if (!osp_precreate_running(d) || d->opd_got_disconnected)
629 GOTO(out, rc = -EAGAIN);
631 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
632 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
634 *last_fid = d->opd_last_used_fid;
635 /* The OSP should already get the valid seq now */
636 LASSERT(!fid_is_zero(last_fid));
637 if (fid_oid(&d->opd_last_used_fid) < 2) {
638 /* lastfid looks strange... ask OST */
639 rc = osp_get_lastfid_from_ost(env, d);
644 imp = d->opd_obd->u.cli.cl_import;
647 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
649 GOTO(out, rc = -ENOMEM);
651 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
653 ptlrpc_request_free(req);
658 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
660 GOTO(out, rc = -EPROTO);
662 body->oa.o_flags = OBD_FL_DELORPHAN;
663 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
665 fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
667 ptlrpc_request_set_replen(req);
669 /* Don't resend the delorphan req */
670 req->rq_no_resend = req->rq_no_delay = 1;
672 rc = ptlrpc_queue_wait(req);
678 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
680 GOTO(out, rc = -EPROTO);
683 * OST provides us with id new pool starts from in body->oa.o_id
685 ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
687 spin_lock(&d->opd_pre_lock);
688 diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
690 d->opd_pre_grow_count = OST_MIN_PRECREATE + diff;
691 d->opd_pre_last_created_fid = d->opd_last_used_fid;
693 d->opd_pre_grow_count = OST_MIN_PRECREATE;
694 d->opd_pre_last_created_fid = *last_fid;
697 * This empties the pre-creation pool and effectively blocks any new
700 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
701 LUSTRE_DATA_SEQ_MAX_WIDTH);
702 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
703 d->opd_pre_grow_slow = 0;
704 spin_unlock(&d->opd_pre_lock);
706 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
707 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
708 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
711 ptlrpc_req_finished(req);
713 spin_lock(&d->opd_pre_lock);
714 d->opd_pre_recovering = 0;
715 spin_unlock(&d->opd_pre_lock);
718 * If rc is zero, the pre-creation window should have been emptied.
719 * Since waking up the herd would be useless without pre-created
720 * objects, we defer the signal to osp_precreate_send() in that case.
724 CERROR("%s: cannot cleanup orphans: rc = %d\n",
725 d->opd_obd->obd_name, rc);
726 /* we can't proceed from here, OST seem to
727 * be in a bad shape, better to wait for
728 * a new instance of the server and repeat
729 * from the beginning. notify possible waiters
730 * this OSP isn't quite functional yet */
731 osp_pre_update_status(d, rc);
733 wake_up(&d->opd_pre_user_waitq);
741 * the function updates current precreation status used: functional or not
743 * rc is a last code from the transport, rc == 0 meaning transport works
744 * well and users of lod can use objects from this OSP
746 * the status depends on current usage of OST
748 void osp_pre_update_status(struct osp_device *d, int rc)
750 struct obd_statfs *msfs = &d->opd_statfs;
751 int old = d->opd_pre_status;
754 d->opd_pre_status = rc;
758 /* Add a bit of hysteresis so this flag isn't continually flapping,
759 * and ensure that new files don't get extremely fragmented due to
760 * only a small amount of available space in the filesystem.
761 * We want to set the NOSPC flag when there is less than ~0.1% free
762 * and clear it when there is at least ~0.2% free space, so:
763 * avail < ~0.1% max max = avail + used
764 * 1025 * avail < avail + used used = blocks - free
765 * 1024 * avail < used
766 * 1024 * avail < blocks - free
767 * avail < ((blocks - free) >> 10)
769 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
770 * lose that amount of space so in those cases we report no space left
771 * if their is less than 1 GB left. */
772 if (likely(msfs->os_type)) {
773 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
775 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
776 d->opd_pre_status = -ENOSPC;
778 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
779 LPU64" free, "LPU64" used, "LPU64" "
780 "avail -> %d: rc = %d\n",
781 d->opd_obd->obd_name, msfs->os_blocks,
782 msfs->os_bfree, used, msfs->os_bavail,
783 d->opd_pre_status, rc);
785 "non-commited changes: %lu, in progress: %u\n",
786 d->opd_syn_changes, d->opd_syn_rpc_in_progress);
787 } else if (old == -ENOSPC) {
788 d->opd_pre_status = 0;
789 spin_lock(&d->opd_pre_lock);
790 d->opd_pre_grow_slow = 0;
791 d->opd_pre_grow_count = OST_MIN_PRECREATE;
792 spin_unlock(&d->opd_pre_lock);
793 wake_up(&d->opd_pre_waitq);
794 CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
795 " free, "LPU64" used, "LPU64" avail -> %d: "
796 "rc = %d\n", d->opd_obd->obd_name,
797 msfs->os_blocks, msfs->os_bfree, used,
798 msfs->os_bavail, d->opd_pre_status, rc);
803 wake_up(&d->opd_pre_user_waitq);
806 int osp_init_pre_fid(struct osp_device *osp)
809 struct osp_thread_info *osi;
810 struct lu_client_seq *cli_seq;
811 struct lu_fid *last_fid;
815 LASSERT(osp->opd_pre != NULL);
817 /* Return if last_used fid has been initialized */
818 if (!fid_is_zero(&osp->opd_last_used_fid))
821 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
823 CERROR("%s: init env error: rc = %d\n",
824 osp->opd_obd->obd_name, rc);
828 osi = osp_env_info(&env);
829 last_fid = &osi->osi_fid;
831 /* For a freshed fs, it will allocate a new sequence first */
832 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
833 cli_seq = osp->opd_obd->u.cli.cl_seq;
834 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
836 CERROR("%s: alloc fid error: rc = %d\n",
837 osp->opd_obd->obd_name, rc);
841 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
846 spin_lock(&osp->opd_pre_lock);
847 osp->opd_last_used_fid = *last_fid;
848 osp->opd_pre_used_fid = *last_fid;
849 osp->opd_pre_last_created_fid = *last_fid;
850 spin_unlock(&osp->opd_pre_lock);
851 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
853 CERROR("%s: write fid error: rc = %d\n",
854 osp->opd_obd->obd_name, rc);
862 static int osp_precreate_thread(void *_arg)
864 struct osp_device *d = _arg;
865 struct ptlrpc_thread *thread = &d->opd_pre_thread;
866 struct l_wait_info lwi = { 0 };
872 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
874 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
879 spin_lock(&d->opd_pre_lock);
880 thread->t_flags = SVC_RUNNING;
881 spin_unlock(&d->opd_pre_lock);
882 wake_up(&thread->t_ctl_waitq);
884 while (osp_precreate_running(d)) {
886 * need to be connected to OST
888 while (osp_precreate_running(d)) {
889 l_wait_event(d->opd_pre_waitq,
890 !osp_precreate_running(d) ||
891 d->opd_new_connection,
894 if (!d->opd_new_connection)
897 d->opd_new_connection = 0;
898 d->opd_got_disconnected = 0;
902 if (!osp_precreate_running(d))
905 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
906 /* Sigh, fid client is not ready yet */
907 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
910 /* Init fid for osp_precreate if necessary */
911 rc = osp_init_pre_fid(d);
913 class_export_put(d->opd_exp);
914 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
915 CERROR("%s: init pre fid error: rc = %d\n",
916 d->opd_obd->obd_name, rc);
920 osp_statfs_update(d);
923 * Clean up orphans or recreate missing objects.
925 rc = osp_precreate_cleanup_orphans(&env, d);
929 * connected, can handle precreates now
931 while (osp_precreate_running(d)) {
932 l_wait_event(d->opd_pre_waitq,
933 !osp_precreate_running(d) ||
934 osp_precreate_near_empty(&env, d) ||
935 osp_statfs_need_update(d) ||
936 d->opd_got_disconnected, &lwi);
938 if (!osp_precreate_running(d))
941 /* something happened to the connection
942 * have to start from the beginning */
943 if (d->opd_got_disconnected)
946 if (osp_statfs_need_update(d))
947 osp_statfs_update(d);
949 /* To avoid handling different seq in precreate/orphan
950 * cleanup, it will hold precreate until current seq is
952 if (unlikely(osp_precreate_end_seq(&env, d) &&
953 !osp_create_end_seq(&env, d)))
956 if (unlikely(osp_precreate_end_seq(&env, d) &&
957 osp_create_end_seq(&env, d))) {
958 LCONSOLE_INFO("%s:"LPX64" is used up."
959 " Update to new seq\n",
960 d->opd_obd->obd_name,
961 fid_seq(&d->opd_pre_last_created_fid));
962 rc = osp_precreate_rollover_new_seq(&env, d);
967 if (osp_precreate_near_empty(&env, d)) {
968 rc = osp_precreate_send(&env, d);
969 /* osp_precreate_send() sets opd_pre_status
970 * in case of error, that prevent the using of
972 if (rc < 0 && rc != -ENOSPC &&
973 rc != -ETIMEDOUT && rc != -ENOTCONN)
974 CERROR("%s: cannot precreate objects:"
976 d->opd_obd->obd_name, rc);
981 thread->t_flags = SVC_STOPPED;
983 wake_up(&thread->t_ctl_waitq);
988 static int osp_precreate_ready_condition(const struct lu_env *env,
989 struct osp_device *d)
991 if (d->opd_pre_recovering)
994 /* ready if got enough precreated objects */
995 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
996 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
999 /* ready if OST reported no space and no destroys in progress */
1000 if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
1001 d->opd_pre_status == -ENOSPC)
1004 /* Bail out I/O fails to OST */
1005 if (d->opd_pre_status != 0 &&
1006 d->opd_pre_status != -EAGAIN &&
1007 d->opd_pre_status != -ENODEV &&
1008 d->opd_pre_status != -ENOSPC) {
1010 if (d->opd_pre_status != -EIO)
1011 CERROR("%s: precreate failed opd_pre_status %d\n",
1012 d->opd_obd->obd_name, d->opd_pre_status);
1019 static int osp_precreate_timeout_condition(void *data)
1021 struct osp_device *d = data;
1023 CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1024 "reserved="LPU64", syn_changes=%lu, "
1025 "syn_rpc_in_progress=%d, status=%d\n",
1026 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1027 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1028 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1035 * called to reserve object in the pool
1037 * ENOSPC - no space on corresponded OST
1038 * EAGAIN - precreation is in progress, try later
1039 * EIO - no access to OST
1041 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1043 struct l_wait_info lwi;
1044 cfs_time_t expire = cfs_time_shift(obd_timeout);
1049 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1050 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1051 PFID(&d->opd_pre_used_fid));
1055 * - preallocation is done
1056 * - no free space expected soon
1057 * - can't connect to OST for too long (obd_timeout)
1058 * - OST can allocate fid sequence.
1060 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1061 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1064 * increase number of precreations
1066 precreated = osp_objs_precreated(env, d);
1067 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
1068 d->opd_pre_grow_slow == 0 &&
1069 precreated <= (d->opd_pre_grow_count / 4 + 1)) {
1070 spin_lock(&d->opd_pre_lock);
1071 d->opd_pre_grow_slow = 1;
1072 d->opd_pre_grow_count *= 2;
1073 spin_unlock(&d->opd_pre_lock);
1076 spin_lock(&d->opd_pre_lock);
1077 precreated = osp_objs_precreated(env, d);
1078 if (precreated > d->opd_pre_reserved &&
1079 !d->opd_pre_recovering) {
1080 d->opd_pre_reserved++;
1081 spin_unlock(&d->opd_pre_lock);
1084 /* XXX: don't wake up if precreation is in progress */
1085 if (osp_precreate_near_empty_nolock(env, d) &&
1086 !osp_precreate_end_seq_nolock(env, d))
1087 wake_up(&d->opd_pre_waitq);
1091 spin_unlock(&d->opd_pre_lock);
1094 * all precreated objects have been used and no-space
1095 * status leave us no chance to succeed very soon
1096 * but if there is destroy in progress, then we should
1097 * wait till that is done - some space might be released
1099 if (unlikely(rc == -ENOSPC)) {
1100 if (d->opd_syn_changes) {
1101 /* force local commit to release space */
1102 dt_commit_async(env, d->opd_storage);
1104 if (d->opd_syn_rpc_in_progress) {
1105 /* just wait till destroys are done */
1106 /* see l_wait_even() few lines below */
1108 if (d->opd_syn_changes +
1109 d->opd_syn_rpc_in_progress == 0) {
1110 /* no hope for free space */
1115 /* XXX: don't wake up if precreation is in progress */
1116 wake_up(&d->opd_pre_waitq);
1118 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1119 osp_precreate_timeout_condition, d);
1120 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1125 l_wait_event(d->opd_pre_user_waitq,
1126 osp_precreate_ready_condition(env, d), &lwi);
1133 * this function relies on reservation made before
1135 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1138 /* grab next id from the pool */
1139 spin_lock(&d->opd_pre_lock);
1141 LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1142 &d->opd_pre_last_created_fid) < 0,
1143 "next fid "DFID" last created fid "DFID"\n",
1144 PFID(&d->opd_pre_used_fid),
1145 PFID(&d->opd_pre_last_created_fid));
1147 d->opd_pre_used_fid.f_oid++;
1148 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1149 d->opd_pre_reserved--;
1151 * last_used_id must be changed along with getting new id otherwise
1152 * we might miscalculate gap causing object loss or leak
1154 osp_update_last_fid(d, fid);
1155 spin_unlock(&d->opd_pre_lock);
1158 * probably main thread suspended orphan cleanup till
1159 * all reservations are released, see comment in
1160 * osp_precreate_thread() just before orphan cleanup
1162 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1163 wake_up(&d->opd_pre_waitq);
1171 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1174 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1175 struct ptlrpc_request *req = NULL;
1176 struct obd_import *imp;
1177 struct ost_body *body;
1178 struct obdo *oa = NULL;
1183 imp = d->opd_obd->u.cli.cl_import;
1186 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1190 /* XXX: capa support? */
1191 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1192 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1194 ptlrpc_request_free(req);
1199 * XXX: decide how do we do here with resend
1200 * if we don't resend, then client may see wrong file size
1201 * if we do resend, then MDS thread can get stuck for quite long
1203 req->rq_no_resend = req->rq_no_delay = 1;
1205 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1206 ptlrpc_at_set_req_timeout(req);
1210 GOTO(out, rc = -ENOMEM);
1212 rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1215 oa->o_blocks = OBD_OBJECT_EOF;
1216 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1217 OBD_MD_FLID | OBD_MD_FLGROUP;
1219 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1221 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1223 /* XXX: capa support? */
1224 /* osc_pack_capa(req, body, capa); */
1226 ptlrpc_request_set_replen(req);
1228 rc = ptlrpc_queue_wait(req);
1230 CERROR("can't punch object: %d\n", rc);
1232 ptlrpc_req_finished(req);
1238 int osp_init_precreate(struct osp_device *d)
1240 struct l_wait_info lwi = { 0 };
1241 struct task_struct *task;
1245 OBD_ALLOC_PTR(d->opd_pre);
1246 if (d->opd_pre == NULL)
1249 /* initially precreation isn't ready */
1250 d->opd_pre_status = -EAGAIN;
1251 fid_zero(&d->opd_pre_used_fid);
1252 d->opd_pre_used_fid.f_oid = 1;
1253 fid_zero(&d->opd_pre_last_created_fid);
1254 d->opd_pre_last_created_fid.f_oid = 1;
1255 d->opd_pre_reserved = 0;
1256 d->opd_got_disconnected = 1;
1257 d->opd_pre_grow_slow = 0;
1258 d->opd_pre_grow_count = OST_MIN_PRECREATE;
1259 d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
1260 d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
1262 spin_lock_init(&d->opd_pre_lock);
1263 init_waitqueue_head(&d->opd_pre_waitq);
1264 init_waitqueue_head(&d->opd_pre_user_waitq);
1265 init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1268 * Initialize statfs-related things
1270 d->opd_statfs_maxage = 5; /* default update interval */
1271 d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1272 CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1273 (unsigned long long)cfs_time_current(),
1274 (unsigned long long)d->opd_statfs_fresh_till);
1275 cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
1278 * start thread handling precreation and statfs updates
1280 task = kthread_run(osp_precreate_thread, d,
1281 "osp-pre-%u-%u", d->opd_index, d->opd_group);
1283 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1284 RETURN(PTR_ERR(task));
1287 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1288 osp_precreate_running(d) || osp_precreate_stopped(d),
1294 void osp_precreate_fini(struct osp_device *d)
1296 struct ptlrpc_thread *thread;
1300 cfs_timer_disarm(&d->opd_statfs_timer);
1302 if (d->opd_pre == NULL)
1305 thread = &d->opd_pre_thread;
1307 thread->t_flags = SVC_STOPPING;
1308 wake_up(&d->opd_pre_waitq);
1310 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1312 OBD_FREE_PTR(d->opd_pre);