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, 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>
46 # define EXPORT_SYMTAB
48 #define DEBUG_SUBSYSTEM S_MDS
50 #include "osp_internal.h"
53 * there are two specific states to take care about:
55 * = import is disconnected =
57 * = import is inactive =
58 * in this case osp_declare_object_create() returns an error
65 static inline int osp_statfs_need_update(struct osp_device *d)
67 return !cfs_time_before(cfs_time_current(),
68 d->opd_statfs_fresh_till);
71 static void osp_statfs_timer_cb(unsigned long _d)
73 struct osp_device *d = (struct osp_device *) _d;
76 cfs_waitq_signal(&d->opd_pre_waitq);
79 static int osp_statfs_interpret(const struct lu_env *env,
80 struct ptlrpc_request *req,
81 union ptlrpc_async_args *aa, int rc)
83 struct obd_import *imp = req->rq_import;
84 struct obd_statfs *msfs;
89 aa = ptlrpc_req_async_args(req);
90 d = aa->pointer_arg[0];
96 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
98 GOTO(out, rc = -EPROTO);
100 d->opd_statfs = *msfs;
102 osp_pre_update_status(d, rc);
104 /* schedule next update */
105 d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
106 cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
107 d->opd_statfs_update_in_progress = 0;
109 CDEBUG(D_CACHE, "updated statfs %p\n", d);
113 /* couldn't update statfs, try again as soon as possible */
114 cfs_waitq_signal(&d->opd_pre_waitq);
115 if (req->rq_import_generation == imp->imp_generation)
116 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
117 d->opd_obd->obd_name, rc);
121 static int osp_statfs_update(struct osp_device *d)
123 struct ptlrpc_request *req;
124 struct obd_import *imp;
125 union ptlrpc_async_args *aa;
130 CDEBUG(D_CACHE, "going to update statfs\n");
132 imp = d->opd_obd->u.cli.cl_import;
135 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
139 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
141 ptlrpc_request_free(req);
144 ptlrpc_request_set_replen(req);
145 req->rq_request_portal = OST_CREATE_PORTAL;
146 ptlrpc_at_set_req_timeout(req);
148 req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
149 aa = ptlrpc_req_async_args(req);
150 aa->pointer_arg[0] = d;
153 * no updates till reply
155 cfs_timer_disarm(&d->opd_statfs_timer);
156 d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
157 d->opd_statfs_update_in_progress = 1;
159 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
165 * XXX: there might be a case where removed object(s) do not add free
166 * space (empty object). if the number of such deletions is high, then
167 * we can start to update statfs too often - a rpc storm
168 * TODO: some throttling is needed
170 void osp_statfs_need_now(struct osp_device *d)
172 if (!d->opd_statfs_update_in_progress) {
174 * if current status is -ENOSPC (lack of free space on OST)
175 * then we should poll OST immediately once object destroy
178 d->opd_statfs_fresh_till = cfs_time_shift(-1);
179 cfs_timer_disarm(&d->opd_statfs_timer);
180 cfs_waitq_signal(&d->opd_pre_waitq);
186 * OSP tries to maintain pool of available objects so that calls to create
187 * objects don't block most of time
189 * each time OSP gets connected to OST, we should start from precreation cleanup
191 static inline int osp_precreate_running(struct osp_device *d)
193 return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
196 static inline int osp_precreate_stopped(struct osp_device *d)
198 return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
201 static inline int osp_objs_precreated(const struct lu_env *env,
202 struct osp_device *osp)
204 struct lu_fid *fid1 = &osp->opd_pre_last_created_fid;
205 struct lu_fid *fid2 = &osp->opd_pre_used_fid;
207 LASSERTF(fid_seq(fid1) == fid_seq(fid2),
208 "Created fid"DFID" Next fid "DFID"\n", PFID(fid1), PFID(fid2));
210 if (fid_is_idif(fid1)) {
211 struct ost_id *oi1 = &osp_env_info(env)->osi_oi;
212 struct ost_id *oi2 = &osp_env_info(env)->osi_oi2;
214 LASSERT(fid_is_idif(fid1) && fid_is_idif(fid2));
215 ostid_idif_pack(fid1, oi1);
216 ostid_idif_pack(fid2, oi2);
217 LASSERT(oi1->oi_id >= oi2->oi_id);
219 return oi1->oi_id - oi2->oi_id;
222 return fid_oid(fid1) - fid_oid(fid2);
225 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
226 struct osp_device *d)
228 int window = osp_objs_precreated(env, d);
230 /* don't consider new precreation till OST is healty and
232 return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) &&
233 (d->opd_pre_status == 0));
236 static inline int osp_precreate_near_empty(const struct lu_env *env,
237 struct osp_device *d)
241 /* XXX: do we really need locking here? */
242 spin_lock(&d->opd_pre_lock);
243 rc = osp_precreate_near_empty_nolock(env, d);
244 spin_unlock(&d->opd_pre_lock);
249 * Write fid into last_oid/last_seq file.
251 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
252 struct lu_fid *fid, int sync)
254 struct osp_thread_info *oti = osp_env_info(env);
255 struct lu_buf *lb_oid = &oti->osi_lb;
256 struct lu_buf *lb_oseq = &oti->osi_lb2;
263 /* Note: through f_oid is only 32bits, it will also write
264 * 64 bits for oid to keep compatiblity with the previous
266 lb_oid->lb_buf = &fid->f_oid;
267 lb_oid->lb_len = sizeof(obd_id);
268 oid_off = sizeof(obd_id) * osp->opd_index;
270 lb_oseq->lb_buf = &fid->f_seq;
271 lb_oseq->lb_len = sizeof(obd_id);
272 oseq_off = sizeof(obd_id) * osp->opd_index;
274 th = dt_trans_create(env, osp->opd_storage);
279 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
280 lb_oid->lb_len, oid_off, th);
284 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
285 lb_oseq->lb_len, oseq_off, th);
289 rc = dt_trans_start_local(env, osp->opd_storage, th);
293 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
296 CERROR("%s: can not write to last seq file: rc = %d\n",
297 osp->opd_obd->obd_name, rc);
300 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
303 CERROR("%s: can not write to last seq file: rc = %d\n",
304 osp->opd_obd->obd_name, rc);
308 dt_trans_stop(env, osp->opd_storage, th);
313 * alloc fids for precreation.
314 * rc = 0 Success, @grow is the count of real allocation.
315 * rc = 1 Current seq is used up.
316 * rc < 0 Other error.
318 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
319 struct lu_fid *fid, int *grow)
321 struct osp_thread_info *osi = osp_env_info(env);
325 if (fid_is_idif(fid)) {
326 struct lu_fid *last_fid;
327 struct ost_id *oi = &osi->osi_oi;
329 spin_lock(&osp->opd_pre_lock);
330 last_fid = &osp->opd_pre_last_created_fid;
331 ostid_idif_pack(last_fid, oi);
332 end = min(oi->oi_id + *grow, IDIF_MAX_OID);
333 *grow = end - oi->oi_id;
335 spin_unlock(&osp->opd_pre_lock);
340 ostid_idif_unpack(oi, fid, osp->opd_index);
344 spin_lock(&osp->opd_pre_lock);
345 *fid = osp->opd_pre_last_created_fid;
347 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
348 *grow = end - fid->f_oid;
349 fid->f_oid += end - fid->f_oid;
350 spin_unlock(&osp->opd_pre_lock);
352 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
353 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
355 return *grow > 0 ? 0 : 1;
358 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
360 struct osp_thread_info *oti = osp_env_info(env);
361 struct ptlrpc_request *req;
362 struct obd_import *imp;
363 struct ost_body *body;
365 struct lu_fid *fid = &oti->osi_fid;
368 /* don't precreate new objects till OST healthy and has free space */
369 if (unlikely(d->opd_pre_status)) {
370 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
371 d->opd_obd->obd_name, d->opd_pre_status);
376 * if not connection/initialization is compeleted, ignore
378 imp = d->opd_obd->u.cli.cl_import;
381 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
384 req->rq_request_portal = OST_CREATE_PORTAL;
385 /* we should not resend create request - anyway we will have delorphan
386 * and kill these objects */
387 req->rq_no_delay = req->rq_no_resend = 1;
389 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
391 ptlrpc_request_free(req);
395 spin_lock(&d->opd_pre_lock);
396 if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
397 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
398 grow = d->opd_pre_grow_count;
399 spin_unlock(&d->opd_pre_lock);
401 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
404 rc = osp_precreate_fids(env, d, fid, &grow);
406 /* Current seq has been used up*/
407 if (!osp_is_fid_client(d)) {
408 osp_pre_update_status(d, -ENOSPC);
411 cfs_waitq_signal(&d->opd_pre_waitq);
415 if (!osp_is_fid_client(d)) {
416 /* Non-FID client will always send seq 0 because of
418 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
422 ostid_fid_pack(fid, &body->oa.o_oi);
423 body->oa.o_valid = OBD_MD_FLGROUP;
425 ptlrpc_request_set_replen(req);
427 rc = ptlrpc_queue_wait(req);
429 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
433 LASSERT(req->rq_transno == 0);
435 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
437 GOTO(out_req, rc = -EPROTO);
439 fid_ostid_unpack(fid, &body->oa.o_oi, d->opd_index);
440 LASSERTF(lu_fid_diff(fid, &d->opd_pre_used_fid) > 0,
441 "reply fid "DFID" pre used fid "DFID"\n", PFID(fid),
442 PFID(&d->opd_pre_used_fid));
444 CDEBUG(D_HA, "%s: new last_created "DFID"\n", d->opd_obd->obd_name,
447 diff = lu_fid_diff(fid, &d->opd_pre_last_created_fid);
449 spin_lock(&d->opd_pre_lock);
451 /* the OST has not managed to create all the
452 * objects we asked for */
453 d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
454 d->opd_pre_grow_slow = 1;
456 /* the OST is able to keep up with the work,
457 * we could consider increasing grow_count
458 * next time if needed */
459 d->opd_pre_grow_slow = 0;
462 d->opd_pre_last_created_fid = *fid;
463 spin_unlock(&d->opd_pre_lock);
465 CDEBUG(D_OTHER, "current precreated pool: "DFID"-"DFID"\n",
466 PFID(&d->opd_pre_used_fid), PFID(&d->opd_pre_last_created_fid));
468 /* now we can wakeup all users awaiting for objects */
469 osp_pre_update_status(d, rc);
470 cfs_waitq_signal(&d->opd_pre_user_waitq);
472 ptlrpc_req_finished(req);
476 static int osp_get_lastfid_from_ost(struct osp_device *d)
478 struct ptlrpc_request *req = NULL;
479 struct obd_import *imp;
480 struct lu_fid *last_fid = &d->opd_last_used_fid;
485 imp = d->opd_obd->u.cli.cl_import;
488 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
492 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
493 RCL_CLIENT, sizeof(KEY_LAST_FID));
495 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
496 RCL_CLIENT, sizeof(*last_fid));
498 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
500 ptlrpc_request_free(req);
504 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
505 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
507 req->rq_no_delay = req->rq_no_resend = 1;
508 fid_cpu_to_le(last_fid, last_fid);
509 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
510 memcpy(tmp, last_fid, sizeof(*last_fid));
511 ptlrpc_request_set_replen(req);
513 rc = ptlrpc_queue_wait(req);
515 /* bad-bad OST.. let sysadm sort this out */
516 if (rc == -ENOTSUPP) {
517 CERROR("%s: server does not support FID: rc = %d\n",
518 d->opd_obd->obd_name, -ENOTSUPP);
520 ptlrpc_set_import_active(imp, 0);
524 last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
525 if (last_fid == NULL || !fid_is_sane(last_fid)) {
526 CERROR("%s: Got insane last_fid "DFID"\n",
527 d->opd_obd->obd_name, PFID(last_fid));
528 GOTO(out, rc = -EPROTO);
531 /* Only update the last used fid, if the OST has objects for
532 * this sequence, i.e. fid_oid > 0 */
533 if (fid_oid(last_fid) > 0)
534 d->opd_last_used_fid = *last_fid;
536 CDEBUG(D_HA, "%s: Got insane last_fid "DFID"\n", d->opd_obd->obd_name,
540 ptlrpc_req_finished(req);
545 * asks OST to clean precreate orphans
546 * and gets next id for new objects
548 static int osp_precreate_cleanup_orphans(struct lu_env *env,
549 struct osp_device *d)
551 struct osp_thread_info *osi = osp_env_info(env);
552 struct lu_fid *last_fid = &osi->osi_fid;
553 struct ptlrpc_request *req = NULL;
554 struct obd_import *imp;
555 struct ost_body *body;
556 struct l_wait_info lwi = { 0 };
557 int update_status = 0;
564 * wait for local recovery to finish, so we can cleanup orphans
565 * orphans are all objects since "last used" (assigned), but
566 * there might be objects reserved and in some cases they won't
567 * be used. we can't cleanup them till we're sure they won't be
568 * used. also can't we allow new reservations because they may
569 * end up getting orphans being cleaned up below. so we block
570 * new reservations and wait till all reserved objects either
573 spin_lock(&d->opd_pre_lock);
574 d->opd_pre_recovering = 1;
575 spin_unlock(&d->opd_pre_lock);
577 * The locking above makes sure the opd_pre_reserved check below will
578 * catch all osp_precreate_reserve() calls who find
579 * "!opd_pre_recovering".
581 l_wait_event(d->opd_pre_waitq,
582 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
583 !osp_precreate_running(d) || d->opd_got_disconnected,
585 if (!osp_precreate_running(d) || d->opd_got_disconnected)
586 GOTO(out, rc = -EAGAIN);
588 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
589 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
591 *last_fid = d->opd_last_used_fid;
592 /* The OSP should already get the valid seq now */
593 LASSERT(!fid_is_zero(last_fid));
594 if (fid_oid(&d->opd_last_used_fid) < 2) {
595 /* lastfid looks strange... ask OST */
596 rc = osp_get_lastfid_from_ost(d);
601 imp = d->opd_obd->u.cli.cl_import;
604 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
606 GOTO(out, rc = -ENOMEM);
608 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
610 ptlrpc_request_free(req);
615 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
617 GOTO(out, rc = -EPROTO);
619 body->oa.o_flags = OBD_FL_DELORPHAN;
620 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
621 if (osp_is_fid_client(d))
622 body->oa.o_seq = fid_seq(&d->opd_last_used_fid);
625 /* remove from NEXT after used one */
626 body->oa.o_id = fid_oid(&d->opd_last_used_fid);
628 ptlrpc_request_set_replen(req);
630 /* Don't resend the delorphan req */
631 req->rq_no_resend = req->rq_no_delay = 1;
633 rc = ptlrpc_queue_wait(req);
639 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
641 GOTO(out, rc = -EPROTO);
644 * OST provides us with id new pool starts from in body->oa.o_id
646 fid_ostid_unpack(last_fid, &body->oa.o_oi, d->opd_index);
647 CDEBUG(D_INFO, "%s: last_fid "DFID" server last fid "DFID"\n",
648 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid),
651 spin_lock(&d->opd_pre_lock);
652 diff = lu_fid_diff(&d->opd_last_used_fid, last_fid);
654 d->opd_pre_grow_count = OST_MIN_PRECREATE + diff;
655 d->opd_pre_last_created_fid = d->opd_last_used_fid;
657 d->opd_pre_grow_count = OST_MIN_PRECREATE;
658 d->opd_pre_last_created_fid = *last_fid;
661 * This empties the pre-creation pool and effectively blocks any new
664 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
665 LUSTRE_DATA_SEQ_MAX_WIDTH);
666 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
667 d->opd_pre_grow_slow = 0;
668 spin_unlock(&d->opd_pre_lock);
670 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
671 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
672 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
675 ptlrpc_req_finished(req);
677 d->opd_pre_recovering = 0;
680 * If rc is zero, the pre-creation window should have been emptied.
681 * Since waking up the herd would be useless without pre-created
682 * objects, we defer the signal to osp_precreate_send() in that case.
686 CERROR("%s: cannot cleanup orphans: rc = %d\n",
687 d->opd_obd->obd_name, rc);
688 /* we can't proceed from here, OST seem to
689 * be in a bad shape, better to wait for
690 * a new instance of the server and repeat
691 * from the beginning. notify possible waiters
692 * this OSP isn't quite functional yet */
693 osp_pre_update_status(d, rc);
695 cfs_waitq_signal(&d->opd_pre_user_waitq);
703 * the function updates current precreation status used: functional or not
705 * rc is a last code from the transport, rc == 0 meaning transport works
706 * well and users of lod can use objects from this OSP
708 * the status depends on current usage of OST
710 void osp_pre_update_status(struct osp_device *d, int rc)
712 struct obd_statfs *msfs = &d->opd_statfs;
713 int old = d->opd_pre_status;
716 d->opd_pre_status = rc;
720 /* Add a bit of hysteresis so this flag isn't continually flapping,
721 * and ensure that new files don't get extremely fragmented due to
722 * only a small amount of available space in the filesystem.
723 * We want to set the NOSPC flag when there is less than ~0.1% free
724 * and clear it when there is at least ~0.2% free space, so:
725 * avail < ~0.1% max max = avail + used
726 * 1025 * avail < avail + used used = blocks - free
727 * 1024 * avail < used
728 * 1024 * avail < blocks - free
729 * avail < ((blocks - free) >> 10)
731 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
732 * lose that amount of space so in those cases we report no space left
733 * if their is less than 1 GB left. */
734 if (likely(msfs->os_type)) {
735 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
737 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
738 d->opd_pre_status = -ENOSPC;
740 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
741 LPU64" free, "LPU64" used, "LPU64" "
742 "avail -> %d: rc = %d\n",
743 d->opd_obd->obd_name, msfs->os_blocks,
744 msfs->os_bfree, used, msfs->os_bavail,
745 d->opd_pre_status, rc);
747 "non-commited changes: %lu, in progress: %u\n",
748 d->opd_syn_changes, d->opd_syn_rpc_in_progress);
749 } else if (old == -ENOSPC) {
750 d->opd_pre_status = 0;
751 d->opd_pre_grow_slow = 0;
752 d->opd_pre_grow_count = OST_MIN_PRECREATE;
753 cfs_waitq_signal(&d->opd_pre_waitq);
754 CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
755 " free, "LPU64" used, "LPU64" avail -> %d: "
756 "rc = %d\n", d->opd_obd->obd_name,
757 msfs->os_blocks, msfs->os_bfree, used,
758 msfs->os_bavail, d->opd_pre_status, rc);
763 cfs_waitq_signal(&d->opd_pre_user_waitq);
766 static int osp_init_pre_fid(struct osp_device *osp)
769 struct osp_thread_info *osi;
770 struct lu_client_seq *cli_seq;
771 struct lu_fid *last_fid;
775 /* Return if last_used fid has been initialized */
776 if (!fid_is_zero(&osp->opd_last_used_fid))
779 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
781 CERROR("%s: init env error: rc = %d\n",
782 osp->opd_obd->obd_name, rc);
786 osi = osp_env_info(&env);
787 last_fid = &osi->osi_fid;
789 /* For a freshed fs, it will allocate a new sequence first */
790 if (osp_is_fid_client(osp)) {
791 cli_seq = osp->opd_obd->u.cli.cl_seq;
792 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
794 CERROR("%s: alloc fid error: rc = %d\n",
795 osp->opd_obd->obd_name, rc);
799 last_fid->f_seq = fid_idif_seq(1, osp->opd_index);
804 spin_lock(&osp->opd_pre_lock);
805 osp->opd_last_used_fid = *last_fid;
806 osp->opd_pre_used_fid = *last_fid;
807 osp->opd_pre_last_created_fid = *last_fid;
808 spin_unlock(&osp->opd_pre_lock);
809 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
811 CERROR("%s: write fid error: rc = %d\n",
812 osp->opd_obd->obd_name, rc);
820 static int osp_precreate_thread(void *_arg)
822 struct osp_device *d = _arg;
823 struct ptlrpc_thread *thread = &d->opd_pre_thread;
824 struct l_wait_info lwi = { 0 };
831 sprintf(pname, "osp-pre-%u\n", d->opd_index);
832 cfs_daemonize(pname);
834 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
836 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
841 spin_lock(&d->opd_pre_lock);
842 thread->t_flags = SVC_RUNNING;
843 spin_unlock(&d->opd_pre_lock);
844 cfs_waitq_signal(&thread->t_ctl_waitq);
846 while (osp_precreate_running(d)) {
848 * need to be connected to OST
850 while (osp_precreate_running(d)) {
851 l_wait_event(d->opd_pre_waitq,
852 !osp_precreate_running(d) ||
853 d->opd_new_connection,
856 if (!osp_precreate_running(d))
859 if (!d->opd_new_connection)
862 d->opd_new_connection = 0;
863 d->opd_got_disconnected = 0;
867 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL) {
868 /* Get new sequence for client first */
869 LASSERT(d->opd_exp != NULL);
870 d->opd_obd->u.cli.cl_seq->lcs_exp =
871 class_export_get(d->opd_exp);
872 rc = osp_init_pre_fid(d);
874 class_export_put(d->opd_exp);
875 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
876 CERROR("%s: init pre fid error: rc = %d\n",
877 d->opd_obd->obd_name, rc);
882 osp_statfs_update(d);
885 * Clean up orphans or recreate missing objects.
887 rc = osp_precreate_cleanup_orphans(&env, d);
891 * connected, can handle precreates now
893 while (osp_precreate_running(d)) {
894 l_wait_event(d->opd_pre_waitq,
895 !osp_precreate_running(d) ||
896 osp_precreate_near_empty(&env, d) ||
897 osp_statfs_need_update(d) ||
898 d->opd_got_disconnected, &lwi);
900 if (!osp_precreate_running(d))
903 /* something happened to the connection
904 * have to start from the beginning */
905 if (d->opd_got_disconnected)
908 if (osp_statfs_need_update(d))
909 osp_statfs_update(d);
911 if (osp_precreate_near_empty(&env, d)) {
912 rc = osp_precreate_send(&env, d);
913 /* osp_precreate_send() sets opd_pre_status
914 * in case of error, that prevent the using of
916 if (rc < 0 && rc != -ENOSPC &&
917 rc != -ETIMEDOUT && rc != -ENOTCONN)
918 CERROR("%s: cannot precreate objects:"
920 d->opd_obd->obd_name, rc);
925 thread->t_flags = SVC_STOPPED;
927 cfs_waitq_signal(&thread->t_ctl_waitq);
932 static int osp_precreate_ready_condition(const struct lu_env *env,
933 struct osp_device *d)
935 if (d->opd_pre_recovering)
938 /* ready if got enough precreated objects */
939 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
940 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
943 /* ready if OST reported no space and no destoys in progress */
944 if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
945 d->opd_pre_status == -ENOSPC)
951 static int osp_precreate_timeout_condition(void *data)
953 struct osp_device *d = data;
955 LCONSOLE_WARN("%s: slow creates, last="DFID", next="DFID", "
956 "reserved="LPU64", syn_changes=%lu, "
957 "syn_rpc_in_progress=%d, status=%d\n",
958 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
959 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
960 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
967 * called to reserve object in the pool
969 * ENOSPC - no space on corresponded OST
970 * EAGAIN - precreation is in progress, try later
971 * EIO - no access to OST
973 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
975 struct l_wait_info lwi;
976 cfs_time_t expire = cfs_time_shift(obd_timeout);
982 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
983 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
984 PFID(&d->opd_pre_used_fid));
988 * - preallocation is done
989 * - no free space expected soon
990 * - can't connect to OST for too long (obd_timeout)
991 * - OST can allocate fid sequence.
993 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
994 rc == -ENODEV || rc == -EAGAIN) {
996 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 3, 90, 0)
998 * to address Andreas's concern on possible busy-loop
999 * between this thread and osp_precreate_send()
1001 if (unlikely(count++ == 1000)) {
1002 osp_precreate_timeout_condition(d);
1008 * increase number of precreations
1010 precreated = osp_objs_precreated(env, d);
1011 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
1012 d->opd_pre_grow_slow == 0 &&
1013 precreated <= (d->opd_pre_grow_count / 4 + 1)) {
1014 spin_lock(&d->opd_pre_lock);
1015 d->opd_pre_grow_slow = 1;
1016 d->opd_pre_grow_count *= 2;
1017 spin_unlock(&d->opd_pre_lock);
1020 spin_lock(&d->opd_pre_lock);
1021 precreated = osp_objs_precreated(env, d);
1022 if (precreated > d->opd_pre_reserved &&
1023 !d->opd_pre_recovering) {
1024 d->opd_pre_reserved++;
1025 spin_unlock(&d->opd_pre_lock);
1028 /* XXX: don't wake up if precreation is in progress */
1029 if (osp_precreate_near_empty_nolock(env, d))
1030 cfs_waitq_signal(&d->opd_pre_waitq);
1034 spin_unlock(&d->opd_pre_lock);
1037 * all precreated objects have been used and no-space
1038 * status leave us no chance to succeed very soon
1039 * but if there is destroy in progress, then we should
1040 * wait till that is done - some space might be released
1042 if (unlikely(rc == -ENOSPC)) {
1043 if (d->opd_syn_changes) {
1044 /* force local commit to release space */
1045 dt_commit_async(env, d->opd_storage);
1047 if (d->opd_syn_rpc_in_progress) {
1048 /* just wait till destroys are done */
1049 /* see l_wait_even() few lines below */
1051 if (d->opd_syn_changes +
1052 d->opd_syn_rpc_in_progress == 0) {
1053 /* no hope for free space */
1058 /* XXX: don't wake up if precreation is in progress */
1059 cfs_waitq_signal(&d->opd_pre_waitq);
1061 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1062 osp_precreate_timeout_condition, d);
1063 if (cfs_time_aftereq(cfs_time_current(), expire))
1066 l_wait_event(d->opd_pre_user_waitq,
1067 osp_precreate_ready_condition(env, d), &lwi);
1074 * this function relies on reservation made before
1076 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1079 /* grab next id from the pool */
1080 spin_lock(&d->opd_pre_lock);
1082 LASSERTF(lu_fid_diff(&d->opd_pre_used_fid,
1083 &d->opd_pre_last_created_fid) < 0,
1084 "next fid "DFID" last created fid "DFID"\n",
1085 PFID(&d->opd_pre_used_fid),
1086 PFID(&d->opd_pre_last_created_fid));
1088 d->opd_pre_used_fid.f_oid++;
1089 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1090 d->opd_pre_reserved--;
1092 * last_used_id must be changed along with getting new id otherwise
1093 * we might miscalculate gap causing object loss or leak
1095 osp_update_last_fid(d, fid);
1096 spin_unlock(&d->opd_pre_lock);
1099 * probably main thread suspended orphan cleanup till
1100 * all reservations are released, see comment in
1101 * osp_precreate_thread() just before orphan cleanup
1103 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1104 cfs_waitq_signal(&d->opd_pre_waitq);
1112 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1115 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1116 struct ptlrpc_request *req = NULL;
1117 struct obd_import *imp;
1118 struct ost_body *body;
1119 struct obdo *oa = NULL;
1124 imp = d->opd_obd->u.cli.cl_import;
1127 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1131 /* XXX: capa support? */
1132 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1133 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1135 ptlrpc_request_free(req);
1140 * XXX: decide how do we do here with resend
1141 * if we don't resend, then client may see wrong file size
1142 * if we do resend, then MDS thread can get stuck for quite long
1144 req->rq_no_resend = req->rq_no_delay = 1;
1146 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1147 ptlrpc_at_set_req_timeout(req);
1151 GOTO(out, rc = -ENOMEM);
1153 rc = fid_ostid_pack(lu_object_fid(&dt->do_lu), &oa->o_oi);
1156 oa->o_blocks = OBD_OBJECT_EOF;
1157 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1158 OBD_MD_FLID | OBD_MD_FLGROUP;
1160 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1162 lustre_set_wire_obdo(&body->oa, oa);
1164 /* XXX: capa support? */
1165 /* osc_pack_capa(req, body, capa); */
1167 ptlrpc_request_set_replen(req);
1169 rc = ptlrpc_queue_wait(req);
1171 CERROR("can't punch object: %d\n", rc);
1173 ptlrpc_req_finished(req);
1179 int osp_init_precreate(struct osp_device *d)
1181 struct l_wait_info lwi = { 0 };
1186 /* initially precreation isn't ready */
1187 d->opd_pre_status = -EAGAIN;
1188 fid_zero(&d->opd_pre_used_fid);
1189 d->opd_pre_used_fid.f_oid = 1;
1190 fid_zero(&d->opd_pre_last_created_fid);
1191 d->opd_pre_last_created_fid.f_oid = 1;
1192 d->opd_pre_reserved = 0;
1193 d->opd_got_disconnected = 1;
1194 d->opd_pre_grow_slow = 0;
1195 d->opd_pre_grow_count = OST_MIN_PRECREATE;
1196 d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
1197 d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
1199 spin_lock_init(&d->opd_pre_lock);
1200 cfs_waitq_init(&d->opd_pre_waitq);
1201 cfs_waitq_init(&d->opd_pre_user_waitq);
1202 cfs_waitq_init(&d->opd_pre_thread.t_ctl_waitq);
1205 * Initialize statfs-related things
1207 d->opd_statfs_maxage = 5; /* default update interval */
1208 d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1209 CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1210 (unsigned long long)cfs_time_current(),
1211 (unsigned long long)d->opd_statfs_fresh_till);
1212 cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
1215 * start thread handling precreation and statfs updates
1217 rc = cfs_create_thread(osp_precreate_thread, d, 0);
1219 CERROR("can't start precreate thread %d\n", rc);
1223 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1224 osp_precreate_running(d) || osp_precreate_stopped(d),
1230 void osp_precreate_fini(struct osp_device *d)
1232 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1236 cfs_timer_disarm(&d->opd_statfs_timer);
1238 thread->t_flags = SVC_STOPPING;
1239 cfs_waitq_signal(&d->opd_pre_waitq);
1241 cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);