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);
248 static inline int osp_create_end_seq(const struct lu_env *env,
249 struct osp_device *osp)
251 struct lu_fid *fid = &osp->opd_pre_used_fid;
254 spin_lock(&osp->opd_pre_lock);
255 rc = osp_fid_end_seq(env, fid);
256 spin_unlock(&osp->opd_pre_lock);
261 * Write fid into last_oid/last_seq file.
263 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
264 struct lu_fid *fid, int sync)
266 struct osp_thread_info *oti = osp_env_info(env);
267 struct lu_buf *lb_oid = &oti->osi_lb;
268 struct lu_buf *lb_oseq = &oti->osi_lb2;
275 /* Note: through f_oid is only 32bits, it will also write
276 * 64 bits for oid to keep compatiblity with the previous
278 lb_oid->lb_buf = &fid->f_oid;
279 lb_oid->lb_len = sizeof(obd_id);
280 oid_off = sizeof(obd_id) * osp->opd_index;
282 lb_oseq->lb_buf = &fid->f_seq;
283 lb_oseq->lb_len = sizeof(obd_id);
284 oseq_off = sizeof(obd_id) * osp->opd_index;
286 th = dt_trans_create(env, osp->opd_storage);
291 rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
292 lb_oid->lb_len, oid_off, th);
296 rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
297 lb_oseq->lb_len, oseq_off, th);
301 rc = dt_trans_start_local(env, osp->opd_storage, th);
305 rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
308 CERROR("%s: can not write to last seq file: rc = %d\n",
309 osp->opd_obd->obd_name, rc);
312 rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
315 CERROR("%s: can not write to last seq file: rc = %d\n",
316 osp->opd_obd->obd_name, rc);
320 dt_trans_stop(env, osp->opd_storage, th);
324 int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp)
326 struct lu_fid *fid = &osp_env_info(env)->osi_fid;
327 struct lu_fid *last_fid = &osp->opd_last_used_fid;
331 rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
333 CERROR("%s: alloc fid error: rc = %d\n",
334 osp->opd_obd->obd_name, rc);
340 LASSERTF(fid_seq(fid) != fid_seq(last_fid),
341 "fid "DFID", last_fid "DFID"\n", PFID(fid),
344 rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
346 CERROR("%s: Can not update oid/seq file: rc = %d\n",
347 osp->opd_obd->obd_name, rc);
351 LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n",
352 osp->opd_obd->obd_name, fid_seq(last_fid),
354 /* Update last_xxx to the new seq */
355 spin_lock(&osp->opd_pre_lock);
356 osp->opd_last_used_fid = *fid;
357 osp->opd_gap_start_fid = *fid;
358 osp->opd_pre_used_fid = *fid;
359 osp->opd_pre_last_created_fid = *fid;
360 spin_unlock(&osp->opd_pre_lock);
366 * alloc fids for precreation.
367 * rc = 0 Success, @grow is the count of real allocation.
368 * rc = 1 Current seq is used up.
369 * rc < 0 Other error.
371 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
372 struct lu_fid *fid, int *grow)
374 struct osp_thread_info *osi = osp_env_info(env);
378 if (fid_is_idif(fid)) {
379 struct lu_fid *last_fid;
380 struct ost_id *oi = &osi->osi_oi;
382 spin_lock(&osp->opd_pre_lock);
383 last_fid = &osp->opd_pre_last_created_fid;
384 ostid_idif_pack(last_fid, oi);
385 end = min(oi->oi_id + *grow, IDIF_MAX_OID);
386 *grow = end - oi->oi_id;
388 spin_unlock(&osp->opd_pre_lock);
393 ostid_idif_unpack(oi, fid, osp->opd_index);
397 spin_lock(&osp->opd_pre_lock);
398 *fid = osp->opd_pre_last_created_fid;
400 end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
401 *grow = end - fid->f_oid;
402 fid->f_oid += end - fid->f_oid;
403 spin_unlock(&osp->opd_pre_lock);
405 CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
406 *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
408 return *grow > 0 ? 0 : 1;
411 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
413 struct osp_thread_info *oti = osp_env_info(env);
414 struct ptlrpc_request *req;
415 struct obd_import *imp;
416 struct ost_body *body;
418 struct lu_fid *fid = &oti->osi_fid;
421 /* don't precreate new objects till OST healthy and has free space */
422 if (unlikely(d->opd_pre_status)) {
423 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
424 d->opd_obd->obd_name, d->opd_pre_status);
429 * if not connection/initialization is compeleted, ignore
431 imp = d->opd_obd->u.cli.cl_import;
434 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
437 req->rq_request_portal = OST_CREATE_PORTAL;
438 /* we should not resend create request - anyway we will have delorphan
439 * and kill these objects */
440 req->rq_no_delay = req->rq_no_resend = 1;
442 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
444 ptlrpc_request_free(req);
448 spin_lock(&d->opd_pre_lock);
449 if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
450 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
451 grow = d->opd_pre_grow_count;
452 spin_unlock(&d->opd_pre_lock);
454 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
457 *fid = d->opd_pre_last_created_fid;
458 rc = osp_precreate_fids(env, d, fid, &grow);
460 /* Current seq has been used up*/
461 if (!osp_is_fid_client(d)) {
462 osp_pre_update_status(d, -ENOSPC);
465 cfs_waitq_signal(&d->opd_pre_waitq);
469 if (!osp_is_fid_client(d)) {
470 /* Non-FID client will always send seq 0 because of
472 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
476 fid_ostid_pack(fid, &body->oa.o_oi);
477 body->oa.o_valid = OBD_MD_FLGROUP;
479 ptlrpc_request_set_replen(req);
481 rc = ptlrpc_queue_wait(req);
483 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
487 LASSERT(req->rq_transno == 0);
489 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
491 GOTO(out_req, rc = -EPROTO);
493 fid_ostid_unpack(fid, &body->oa.o_oi, d->opd_index);
494 LASSERTF(lu_fid_diff(fid, &d->opd_pre_used_fid) > 0,
495 "reply fid "DFID" pre used fid "DFID"\n", PFID(fid),
496 PFID(&d->opd_pre_used_fid));
498 CDEBUG(D_HA, "%s: new last_created "DFID"\n", d->opd_obd->obd_name,
501 diff = lu_fid_diff(fid, &d->opd_pre_last_created_fid);
503 spin_lock(&d->opd_pre_lock);
505 /* the OST has not managed to create all the
506 * objects we asked for */
507 d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
508 d->opd_pre_grow_slow = 1;
510 /* the OST is able to keep up with the work,
511 * we could consider increasing grow_count
512 * next time if needed */
513 d->opd_pre_grow_slow = 0;
516 d->opd_pre_last_created_fid = *fid;
517 spin_unlock(&d->opd_pre_lock);
519 CDEBUG(D_OTHER, "current precreated pool: "DFID"-"DFID"\n",
520 PFID(&d->opd_pre_used_fid), PFID(&d->opd_pre_last_created_fid));
522 /* now we can wakeup all users awaiting for objects */
523 osp_pre_update_status(d, rc);
524 cfs_waitq_signal(&d->opd_pre_user_waitq);
526 ptlrpc_req_finished(req);
530 static int osp_get_lastfid_from_ost(const struct lu_env *env,
531 struct osp_device *d)
533 struct ost_id *oi = &osp_env_info(env)->osi_oi;
534 struct ptlrpc_request *req = NULL;
535 struct obd_import *imp;
536 struct lu_fid *last_fid = &d->opd_last_used_fid;
541 imp = d->opd_obd->u.cli.cl_import;
544 req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
548 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY, RCL_CLIENT,
549 sizeof(KEY_LAST_FID));
551 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL, RCL_CLIENT,
554 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
556 ptlrpc_request_free(req);
560 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
561 memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
563 req->rq_no_delay = req->rq_no_resend = 1;
564 fid_ostid_pack(last_fid, oi);
565 ostid_cpu_to_le(oi, oi);
566 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
567 memcpy(tmp, oi, sizeof(*oi));
568 ptlrpc_request_set_replen(req);
570 rc = ptlrpc_queue_wait(req);
572 /* bad-bad OST.. let sysadm sort this out */
573 if (rc == -ENOTSUPP) {
574 CERROR("%s: server does not support FID: rc = %d\n",
575 d->opd_obd->obd_name, -ENOTSUPP);
577 ptlrpc_set_import_active(imp, 0);
581 oi = req_capsule_server_get(&req->rq_pill, &RMF_OST_ID);
583 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
584 GOTO(out, rc = -EPROTO);
587 rc = fid_ostid_unpack(last_fid, oi, d->opd_index);
588 if (rc != 0 || !fid_is_sane(last_fid)) {
589 CERROR("%s: Got insane last_fid "DFID"\n",
590 d->opd_obd->obd_name, PFID(last_fid));
591 GOTO(out, rc = -EPROTO);
594 /* Only update the last used fid, if the OST has objects for
595 * this sequence, i.e. fid_oid > 0 */
596 if (fid_oid(last_fid) > 0)
597 d->opd_last_used_fid = *last_fid;
599 CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
603 ptlrpc_req_finished(req);
608 * asks OST to clean precreate orphans
609 * and gets next id for new objects
611 static int osp_precreate_cleanup_orphans(struct lu_env *env,
612 struct osp_device *d)
614 struct osp_thread_info *osi = osp_env_info(env);
615 struct lu_fid *last_fid = &osi->osi_fid;
616 struct ptlrpc_request *req = NULL;
617 struct obd_import *imp;
618 struct ost_body *body;
619 struct l_wait_info lwi = { 0 };
620 int update_status = 0;
627 * wait for local recovery to finish, so we can cleanup orphans
628 * orphans are all objects since "last used" (assigned), but
629 * there might be objects reserved and in some cases they won't
630 * be used. we can't cleanup them till we're sure they won't be
631 * used. also can't we allow new reservations because they may
632 * end up getting orphans being cleaned up below. so we block
633 * new reservations and wait till all reserved objects either
636 spin_lock(&d->opd_pre_lock);
637 d->opd_pre_recovering = 1;
638 spin_unlock(&d->opd_pre_lock);
640 * The locking above makes sure the opd_pre_reserved check below will
641 * catch all osp_precreate_reserve() calls who find
642 * "!opd_pre_recovering".
644 l_wait_event(d->opd_pre_waitq,
645 (!d->opd_pre_reserved && d->opd_recovery_completed) ||
646 !osp_precreate_running(d) || d->opd_got_disconnected,
648 if (!osp_precreate_running(d) || d->opd_got_disconnected)
649 GOTO(out, rc = -EAGAIN);
651 CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
652 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
654 *last_fid = d->opd_last_used_fid;
655 /* The OSP should already get the valid seq now */
656 LASSERT(!fid_is_zero(last_fid));
657 if (fid_oid(&d->opd_last_used_fid) < 2) {
658 /* lastfid looks strange... ask OST */
659 rc = osp_get_lastfid_from_ost(env, d);
664 imp = d->opd_obd->u.cli.cl_import;
667 req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
669 GOTO(out, rc = -ENOMEM);
671 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
673 ptlrpc_request_free(req);
678 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
680 GOTO(out, rc = -EPROTO);
682 body->oa.o_flags = OBD_FL_DELORPHAN;
683 body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
685 fid_ostid_pack(&d->opd_last_used_fid, &body->oa.o_oi);
687 ptlrpc_request_set_replen(req);
689 /* Don't resend the delorphan req */
690 req->rq_no_resend = req->rq_no_delay = 1;
692 rc = ptlrpc_queue_wait(req);
698 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
700 GOTO(out, rc = -EPROTO);
703 * OST provides us with id new pool starts from in body->oa.o_id
705 fid_ostid_unpack(last_fid, &body->oa.o_oi, d->opd_index);
706 CDEBUG(D_INFO, "%s: last_fid "DFID" server last fid "DFID"\n",
707 d->opd_obd->obd_name, PFID(&d->opd_last_used_fid),
710 spin_lock(&d->opd_pre_lock);
711 diff = lu_fid_diff(&d->opd_last_used_fid, last_fid);
713 d->opd_pre_grow_count = OST_MIN_PRECREATE + diff;
714 d->opd_pre_last_created_fid = d->opd_last_used_fid;
716 d->opd_pre_grow_count = OST_MIN_PRECREATE;
717 d->opd_pre_last_created_fid = *last_fid;
720 * This empties the pre-creation pool and effectively blocks any new
723 LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
724 LUSTRE_DATA_SEQ_MAX_WIDTH);
725 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
726 d->opd_pre_grow_slow = 0;
727 spin_unlock(&d->opd_pre_lock);
729 CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
730 "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
731 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
734 ptlrpc_req_finished(req);
736 d->opd_pre_recovering = 0;
739 * If rc is zero, the pre-creation window should have been emptied.
740 * Since waking up the herd would be useless without pre-created
741 * objects, we defer the signal to osp_precreate_send() in that case.
745 CERROR("%s: cannot cleanup orphans: rc = %d\n",
746 d->opd_obd->obd_name, rc);
747 /* we can't proceed from here, OST seem to
748 * be in a bad shape, better to wait for
749 * a new instance of the server and repeat
750 * from the beginning. notify possible waiters
751 * this OSP isn't quite functional yet */
752 osp_pre_update_status(d, rc);
754 cfs_waitq_signal(&d->opd_pre_user_waitq);
762 * the function updates current precreation status used: functional or not
764 * rc is a last code from the transport, rc == 0 meaning transport works
765 * well and users of lod can use objects from this OSP
767 * the status depends on current usage of OST
769 void osp_pre_update_status(struct osp_device *d, int rc)
771 struct obd_statfs *msfs = &d->opd_statfs;
772 int old = d->opd_pre_status;
775 d->opd_pre_status = rc;
779 /* Add a bit of hysteresis so this flag isn't continually flapping,
780 * and ensure that new files don't get extremely fragmented due to
781 * only a small amount of available space in the filesystem.
782 * We want to set the NOSPC flag when there is less than ~0.1% free
783 * and clear it when there is at least ~0.2% free space, so:
784 * avail < ~0.1% max max = avail + used
785 * 1025 * avail < avail + used used = blocks - free
786 * 1024 * avail < used
787 * 1024 * avail < blocks - free
788 * avail < ((blocks - free) >> 10)
790 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
791 * lose that amount of space so in those cases we report no space left
792 * if their is less than 1 GB left. */
793 if (likely(msfs->os_type)) {
794 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
796 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
797 d->opd_pre_status = -ENOSPC;
799 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
800 LPU64" free, "LPU64" used, "LPU64" "
801 "avail -> %d: rc = %d\n",
802 d->opd_obd->obd_name, msfs->os_blocks,
803 msfs->os_bfree, used, msfs->os_bavail,
804 d->opd_pre_status, rc);
806 "non-commited changes: %lu, in progress: %u\n",
807 d->opd_syn_changes, d->opd_syn_rpc_in_progress);
808 } else if (old == -ENOSPC) {
809 d->opd_pre_status = 0;
810 d->opd_pre_grow_slow = 0;
811 d->opd_pre_grow_count = OST_MIN_PRECREATE;
812 cfs_waitq_signal(&d->opd_pre_waitq);
813 CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
814 " free, "LPU64" used, "LPU64" avail -> %d: "
815 "rc = %d\n", d->opd_obd->obd_name,
816 msfs->os_blocks, msfs->os_bfree, used,
817 msfs->os_bavail, d->opd_pre_status, rc);
822 cfs_waitq_signal(&d->opd_pre_user_waitq);
825 static int osp_init_pre_fid(struct osp_device *osp)
828 struct osp_thread_info *osi;
829 struct lu_client_seq *cli_seq;
830 struct lu_fid *last_fid;
834 /* Return if last_used fid has been initialized */
835 if (!fid_is_zero(&osp->opd_last_used_fid))
838 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
840 CERROR("%s: init env error: rc = %d\n",
841 osp->opd_obd->obd_name, rc);
845 osi = osp_env_info(&env);
846 last_fid = &osi->osi_fid;
848 /* For a freshed fs, it will allocate a new sequence first */
849 if (osp_is_fid_client(osp) && osp->opd_group != 0) {
850 cli_seq = osp->opd_obd->u.cli.cl_seq;
851 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
853 CERROR("%s: alloc fid error: rc = %d\n",
854 osp->opd_obd->obd_name, rc);
858 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
863 spin_lock(&osp->opd_pre_lock);
864 osp->opd_last_used_fid = *last_fid;
865 osp->opd_pre_used_fid = *last_fid;
866 osp->opd_pre_last_created_fid = *last_fid;
867 spin_unlock(&osp->opd_pre_lock);
868 rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
870 CERROR("%s: write fid error: rc = %d\n",
871 osp->opd_obd->obd_name, rc);
879 static int osp_precreate_thread(void *_arg)
881 struct osp_device *d = _arg;
882 struct ptlrpc_thread *thread = &d->opd_pre_thread;
883 struct l_wait_info lwi = { 0 };
890 sprintf(pname, "osp-pre-%u", d->opd_index);
891 cfs_daemonize(pname);
893 rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
895 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
900 spin_lock(&d->opd_pre_lock);
901 thread->t_flags = SVC_RUNNING;
902 spin_unlock(&d->opd_pre_lock);
903 cfs_waitq_signal(&thread->t_ctl_waitq);
905 while (osp_precreate_running(d)) {
907 * need to be connected to OST
909 while (osp_precreate_running(d)) {
910 l_wait_event(d->opd_pre_waitq,
911 !osp_precreate_running(d) ||
912 d->opd_new_connection,
915 if (!d->opd_new_connection)
918 d->opd_new_connection = 0;
919 d->opd_got_disconnected = 0;
923 if (!osp_precreate_running(d))
926 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
927 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL) {
928 /* Get new sequence for client first */
929 LASSERT(d->opd_exp != NULL);
930 d->opd_obd->u.cli.cl_seq->lcs_exp =
931 class_export_get(d->opd_exp);
932 rc = osp_init_pre_fid(d);
934 class_export_put(d->opd_exp);
935 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
936 CERROR("%s: init pre fid error: rc = %d\n",
937 d->opd_obd->obd_name, rc);
942 osp_statfs_update(d);
945 * Clean up orphans or recreate missing objects.
947 rc = osp_precreate_cleanup_orphans(&env, d);
951 * connected, can handle precreates now
953 while (osp_precreate_running(d)) {
954 l_wait_event(d->opd_pre_waitq,
955 !osp_precreate_running(d) ||
956 osp_precreate_near_empty(&env, d) ||
957 osp_statfs_need_update(d) ||
958 d->opd_got_disconnected, &lwi);
960 if (!osp_precreate_running(d))
963 /* something happened to the connection
964 * have to start from the beginning */
965 if (d->opd_got_disconnected)
968 if (osp_statfs_need_update(d))
969 osp_statfs_update(d);
971 /* To avoid handling different seq in precreate/orphan
972 * cleanup, it will hold precreate until current seq is
974 if (unlikely(osp_precreate_end_seq(&env, d) &&
975 !osp_create_end_seq(&env, d)))
978 if (unlikely(osp_precreate_end_seq(&env, d) &&
979 osp_create_end_seq(&env, d))) {
980 LCONSOLE_INFO("%s:"LPX64" is used up."
981 " Update to new seq\n",
982 d->opd_obd->obd_name,
983 fid_seq(&d->opd_pre_last_created_fid));
984 rc = osp_precreate_rollover_new_seq(&env, d);
989 if (osp_precreate_near_empty(&env, d)) {
990 rc = osp_precreate_send(&env, d);
991 /* osp_precreate_send() sets opd_pre_status
992 * in case of error, that prevent the using of
994 if (rc < 0 && rc != -ENOSPC &&
995 rc != -ETIMEDOUT && rc != -ENOTCONN)
996 CERROR("%s: cannot precreate objects:"
998 d->opd_obd->obd_name, rc);
1003 thread->t_flags = SVC_STOPPED;
1005 cfs_waitq_signal(&thread->t_ctl_waitq);
1010 static int osp_precreate_ready_condition(const struct lu_env *env,
1011 struct osp_device *d)
1013 if (d->opd_pre_recovering)
1016 /* ready if got enough precreated objects */
1017 /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1018 if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1021 /* ready if OST reported no space and no destoys in progress */
1022 if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
1023 d->opd_pre_status == -ENOSPC)
1029 static int osp_precreate_timeout_condition(void *data)
1031 struct osp_device *d = data;
1033 LCONSOLE_WARN("%s: slow creates, last="DFID", next="DFID", "
1034 "reserved="LPU64", syn_changes=%lu, "
1035 "syn_rpc_in_progress=%d, status=%d\n",
1036 d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1037 PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1038 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1045 * called to reserve object in the pool
1047 * ENOSPC - no space on corresponded OST
1048 * EAGAIN - precreation is in progress, try later
1049 * EIO - no access to OST
1051 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1053 struct l_wait_info lwi;
1054 cfs_time_t expire = cfs_time_shift(obd_timeout);
1060 LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1061 "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1062 PFID(&d->opd_pre_used_fid));
1066 * - preallocation is done
1067 * - no free space expected soon
1068 * - can't connect to OST for too long (obd_timeout)
1069 * - OST can allocate fid sequence.
1071 while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1072 rc == -ENODEV || rc == -EAGAIN) {
1074 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 3, 90, 0)
1076 * to address Andreas's concern on possible busy-loop
1077 * between this thread and osp_precreate_send()
1079 if (unlikely(count++ == 1000)) {
1080 osp_precreate_timeout_condition(d);
1086 * increase number of precreations
1088 precreated = osp_objs_precreated(env, d);
1089 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
1090 d->opd_pre_grow_slow == 0 &&
1091 precreated <= (d->opd_pre_grow_count / 4 + 1)) {
1092 spin_lock(&d->opd_pre_lock);
1093 d->opd_pre_grow_slow = 1;
1094 d->opd_pre_grow_count *= 2;
1095 spin_unlock(&d->opd_pre_lock);
1098 spin_lock(&d->opd_pre_lock);
1099 precreated = osp_objs_precreated(env, d);
1100 if (precreated > d->opd_pre_reserved &&
1101 !d->opd_pre_recovering) {
1102 d->opd_pre_reserved++;
1103 spin_unlock(&d->opd_pre_lock);
1106 /* XXX: don't wake up if precreation is in progress */
1107 if (osp_precreate_near_empty_nolock(env, d) &&
1108 !osp_precreate_end_seq_nolock(env, d))
1109 cfs_waitq_signal(&d->opd_pre_waitq);
1113 spin_unlock(&d->opd_pre_lock);
1116 * all precreated objects have been used and no-space
1117 * status leave us no chance to succeed very soon
1118 * but if there is destroy in progress, then we should
1119 * wait till that is done - some space might be released
1121 if (unlikely(rc == -ENOSPC)) {
1122 if (d->opd_syn_changes) {
1123 /* force local commit to release space */
1124 dt_commit_async(env, d->opd_storage);
1126 if (d->opd_syn_rpc_in_progress) {
1127 /* just wait till destroys are done */
1128 /* see l_wait_even() few lines below */
1130 if (d->opd_syn_changes +
1131 d->opd_syn_rpc_in_progress == 0) {
1132 /* no hope for free space */
1137 /* XXX: don't wake up if precreation is in progress */
1138 cfs_waitq_signal(&d->opd_pre_waitq);
1140 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1141 osp_precreate_timeout_condition, d);
1142 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1147 l_wait_event(d->opd_pre_user_waitq,
1148 osp_precreate_ready_condition(env, d), &lwi);
1155 * this function relies on reservation made before
1157 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1160 /* grab next id from the pool */
1161 spin_lock(&d->opd_pre_lock);
1163 LASSERTF(lu_fid_diff(&d->opd_pre_used_fid,
1164 &d->opd_pre_last_created_fid) < 0,
1165 "next fid "DFID" last created fid "DFID"\n",
1166 PFID(&d->opd_pre_used_fid),
1167 PFID(&d->opd_pre_last_created_fid));
1169 d->opd_pre_used_fid.f_oid++;
1170 memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1171 d->opd_pre_reserved--;
1173 * last_used_id must be changed along with getting new id otherwise
1174 * we might miscalculate gap causing object loss or leak
1176 osp_update_last_fid(d, fid);
1177 spin_unlock(&d->opd_pre_lock);
1180 * probably main thread suspended orphan cleanup till
1181 * all reservations are released, see comment in
1182 * osp_precreate_thread() just before orphan cleanup
1184 if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1185 cfs_waitq_signal(&d->opd_pre_waitq);
1193 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1196 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1197 struct ptlrpc_request *req = NULL;
1198 struct obd_import *imp;
1199 struct ost_body *body;
1200 struct obdo *oa = NULL;
1205 imp = d->opd_obd->u.cli.cl_import;
1208 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1212 /* XXX: capa support? */
1213 /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1214 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1216 ptlrpc_request_free(req);
1221 * XXX: decide how do we do here with resend
1222 * if we don't resend, then client may see wrong file size
1223 * if we do resend, then MDS thread can get stuck for quite long
1225 req->rq_no_resend = req->rq_no_delay = 1;
1227 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1228 ptlrpc_at_set_req_timeout(req);
1232 GOTO(out, rc = -ENOMEM);
1234 rc = fid_ostid_pack(lu_object_fid(&dt->do_lu), &oa->o_oi);
1237 oa->o_blocks = OBD_OBJECT_EOF;
1238 oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1239 OBD_MD_FLID | OBD_MD_FLGROUP;
1241 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1243 lustre_set_wire_obdo(&body->oa, oa);
1245 /* XXX: capa support? */
1246 /* osc_pack_capa(req, body, capa); */
1248 ptlrpc_request_set_replen(req);
1250 rc = ptlrpc_queue_wait(req);
1252 CERROR("can't punch object: %d\n", rc);
1254 ptlrpc_req_finished(req);
1260 int osp_init_precreate(struct osp_device *d)
1262 struct l_wait_info lwi = { 0 };
1267 /* initially precreation isn't ready */
1268 d->opd_pre_status = -EAGAIN;
1269 fid_zero(&d->opd_pre_used_fid);
1270 d->opd_pre_used_fid.f_oid = 1;
1271 fid_zero(&d->opd_pre_last_created_fid);
1272 d->opd_pre_last_created_fid.f_oid = 1;
1273 d->opd_pre_reserved = 0;
1274 d->opd_got_disconnected = 1;
1275 d->opd_pre_grow_slow = 0;
1276 d->opd_pre_grow_count = OST_MIN_PRECREATE;
1277 d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
1278 d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
1280 spin_lock_init(&d->opd_pre_lock);
1281 cfs_waitq_init(&d->opd_pre_waitq);
1282 cfs_waitq_init(&d->opd_pre_user_waitq);
1283 cfs_waitq_init(&d->opd_pre_thread.t_ctl_waitq);
1286 * Initialize statfs-related things
1288 d->opd_statfs_maxage = 5; /* default update interval */
1289 d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1290 CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1291 (unsigned long long)cfs_time_current(),
1292 (unsigned long long)d->opd_statfs_fresh_till);
1293 cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
1296 * start thread handling precreation and statfs updates
1298 rc = cfs_create_thread(osp_precreate_thread, d, 0);
1300 CERROR("can't start precreate thread %d\n", rc);
1304 l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1305 osp_precreate_running(d) || osp_precreate_stopped(d),
1311 void osp_precreate_fini(struct osp_device *d)
1313 struct ptlrpc_thread *thread = &d->opd_pre_thread;
1317 cfs_timer_disarm(&d->opd_statfs_timer);
1319 thread->t_flags = SVC_STOPPING;
1320 cfs_waitq_signal(&d->opd_pre_waitq);
1322 cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);