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>
45 # define EXPORT_SYMTAB
47 #define DEBUG_SUBSYSTEM S_MDS
49 #include <lustre_log.h>
50 #include "osp_internal.h"
52 static int osp_sync_id_traction_init(struct osp_device *d);
53 static void osp_sync_id_traction_fini(struct osp_device *d);
54 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
55 static void osp_sync_remove_from_tracker(struct osp_device *d);
58 * this is a components of OSP implementing synchronization between MDS and OST
59 * it llogs all interesting changes (currently it's uig/gid change and object
60 * destroy) atomically, then makes sure changes hit OST storage
62 * we have 4 queues of work:
64 * the first queue is llog itself, once read a change is stored in 2nd queue
65 * in form of RPC (but RPC isn't fired yet).
67 * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
68 * commit. once change is committed locally it migrates onto 3rd queue.
70 * the third queue (opd_syn_committed_here) holds changes committed locally,
71 * but not sent to OST (as the pipe can be full). once pipe becomes non-full
72 * we take a change from the queue and fire corresponded RPC.
74 * once RPC is reported committed by OST (using regular last_committed mech.)
75 * the change jumps into 4th queue (opd_syn_committed_there), now we can
76 * cancel corresponded llog record and release RPC
78 * opd_syn_changes is a number of unread llog records (to be processed).
79 * notice this number doesn't include llog records from previous boots.
80 * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
82 * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
83 * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
84 * too much memory -- how to deal with 1000th OSTs ? batching could help?
86 * opd_syn_rpc_in_flight is a number of RPC in flight.
87 * we control this with OSP_MAX_IN_FLIGHT
90 /* XXX: do math to learn reasonable threshold
91 * should it be ~ number of changes fitting bulk? */
93 #define OSP_SYN_THRESHOLD 10
94 #define OSP_MAX_IN_FLIGHT 8
95 #define OSP_MAX_IN_PROGRESS 4096
97 #define OSP_JOB_MAGIC 0x26112005
99 static inline int osp_sync_running(struct osp_device *d)
101 return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
104 static inline int osp_sync_stopped(struct osp_device *d)
106 return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
109 static inline int osp_sync_has_new_job(struct osp_device *d)
111 return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
112 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
113 || (d->opd_syn_prev_done == 0);
116 static inline int osp_sync_low_in_progress(struct osp_device *d)
118 return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
121 static inline int osp_sync_low_in_flight(struct osp_device *d)
123 return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
126 static inline int osp_sync_has_work(struct osp_device *d)
128 /* has new/old changes and low in-progress? */
129 if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
130 osp_sync_low_in_flight(d) && d->opd_imp_connected)
133 /* has remotely committed? */
134 if (!cfs_list_empty(&d->opd_syn_committed_there))
140 #define osp_sync_check_for_work(d) \
142 if (osp_sync_has_work(d)) { \
143 cfs_waitq_signal(&d->opd_syn_waitq); \
147 void __osp_sync_check_for_work(struct osp_device *d)
149 osp_sync_check_for_work(d);
152 static inline int osp_sync_can_process_new(struct osp_device *d,
153 struct llog_rec_hdr *rec)
157 if (!osp_sync_low_in_progress(d))
159 if (!osp_sync_low_in_flight(d))
161 if (!d->opd_imp_connected)
163 if (d->opd_syn_prev_done == 0)
165 if (d->opd_syn_changes == 0)
167 if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
172 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
173 llog_op_type type, struct thandle *th)
175 struct osp_thread_info *osi = osp_env_info(env);
176 struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
177 struct llog_ctxt *ctxt;
182 /* it's a layering violation, to access internals of th,
183 * but we can do this as a sanity check, for a while */
184 LASSERT(th->th_dev == d->opd_storage);
187 case MDS_UNLINK64_REC:
188 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
190 case MDS_SETATTR64_REC:
191 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
197 /* we want ->dt_trans_start() to allocate per-thandle structure */
198 th->th_tags |= LCT_OSP_THREAD;
200 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
203 rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
209 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
210 const struct lu_fid *fid, llog_op_type type,
211 int count, struct thandle *th,
212 const struct lu_attr *attr)
214 struct osp_thread_info *osi = osp_env_info(env);
215 struct llog_ctxt *ctxt;
216 struct osp_txn_info *txn;
221 /* it's a layering violation, to access internals of th,
222 * but we can do this as a sanity check, for a while */
223 LASSERT(th->th_dev == d->opd_storage);
226 case MDS_UNLINK64_REC:
227 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
228 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
229 osi->osi_unlink.lur_fid = *fid;
230 osi->osi_unlink.lur_count = count;
232 case MDS_SETATTR64_REC:
233 rc = fid_ostid_pack(fid, &osi->osi_oi);
235 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
236 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
237 osi->osi_setattr.lsr_oid = osi->osi_oi.oi_id;
238 osi->osi_setattr.lsr_oseq = osi->osi_oi.oi_seq;
240 osi->osi_setattr.lsr_uid = attr->la_uid;
241 osi->osi_setattr.lsr_gid = attr->la_gid;
247 txn = osp_txn_info(&th->th_ctx);
250 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
251 osi->osi_hdr.lrh_id = txn->oti_current_id;
253 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
256 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
260 CDEBUG(D_OTHER, "%s: new record %lu:%lu:%lu/%lu: %d\n",
261 d->opd_obd->obd_name,
262 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_oid,
263 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_oseq,
264 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
265 (unsigned long) osi->osi_cookie.lgc_index, rc);
270 if (likely(rc == 0)) {
271 spin_lock(&d->opd_syn_lock);
272 d->opd_syn_changes++;
273 spin_unlock(&d->opd_syn_lock);
279 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
280 llog_op_type type, struct thandle *th,
281 const struct lu_attr *attr)
283 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
284 lu_object_fid(&o->opo_obj.do_lu), type, 1,
288 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
289 struct lu_fid *fid, int lost, struct thandle *th)
291 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
295 * it's quite obvious we can't maintain all the structures in the memory:
296 * while OST is down, MDS can be processing thousands and thousands of unlinks
297 * filling persistent llogs and in-core respresentation
299 * this doesn't scale at all. so we need basically the following:
300 * a) destroy/setattr append llog records
301 * b) once llog has grown to X records, we process first Y committed records
303 * once record R is found via llog_process(), it becomes committed after any
304 * subsequent commit callback (at the most)
308 * called for each atomic on-disk change (not once per transaction batch)
309 * and goes over the list
310 * XXX: should be optimized?
314 * called for each RPC reported committed
316 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
318 struct osp_device *d = req->rq_cb_data;
320 CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
322 if (unlikely(req->rq_transno == 0))
325 /* do not do any opd_dyn_rpc_* accounting here
326 * it's done in osp_sync_interpret sooner or later */
329 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
330 LASSERT(cfs_list_empty(&req->rq_exp_list));
332 ptlrpc_request_addref(req);
334 spin_lock(&d->opd_syn_lock);
335 cfs_list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
336 spin_unlock(&d->opd_syn_lock);
338 /* XXX: some batching wouldn't hurt */
339 cfs_waitq_signal(&d->opd_syn_waitq);
342 static int osp_sync_interpret(const struct lu_env *env,
343 struct ptlrpc_request *req, void *aa, int rc)
345 struct osp_device *d = req->rq_cb_data;
347 if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
348 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
349 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
352 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
353 cfs_atomic_read(&req->rq_refcount),
354 rc, (unsigned) req->rq_transno);
355 LASSERT(rc || req->rq_transno);
359 * we tried to destroy object or update attributes,
360 * but object doesn't exist anymore - cancell llog record
362 LASSERT(req->rq_transno == 0);
363 LASSERT(cfs_list_empty(&req->rq_exp_list));
365 ptlrpc_request_addref(req);
367 spin_lock(&d->opd_syn_lock);
368 cfs_list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
369 spin_unlock(&d->opd_syn_lock);
371 cfs_waitq_signal(&d->opd_syn_waitq);
373 struct obd_import *imp = req->rq_import;
375 * error happened, we'll try to repeat on next boot ?
377 LASSERTF(req->rq_transno == 0 ||
378 req->rq_import_generation < imp->imp_generation,
379 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
380 req->rq_transno, rc, req->rq_import_generation,
381 imp->imp_generation);
382 if (req->rq_transno == 0) {
383 /* this is the last time we see the request
384 * if transno is not zero, then commit cb
385 * will be called at some point */
386 LASSERT(d->opd_syn_rpc_in_progress > 0);
387 spin_lock(&d->opd_syn_lock);
388 d->opd_syn_rpc_in_progress--;
389 spin_unlock(&d->opd_syn_lock);
392 cfs_waitq_signal(&d->opd_syn_waitq);
393 } else if (unlikely(d->opd_pre_status == -ENOSPC)) {
395 * if current status is -ENOSPC (lack of free space on OST)
396 * then we should poll OST immediately once object destroy
399 osp_statfs_need_now(d);
402 LASSERT(d->opd_syn_rpc_in_flight > 0);
403 spin_lock(&d->opd_syn_lock);
404 d->opd_syn_rpc_in_flight--;
405 spin_unlock(&d->opd_syn_lock);
406 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
407 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
408 d->opd_syn_rpc_in_progress);
410 osp_sync_check_for_work(d);
416 * the function walks through list of committed locally changes
417 * and send them to RPC until the pipe is full
419 static void osp_sync_send_new_rpc(struct osp_device *d,
420 struct ptlrpc_request *req)
422 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
423 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
425 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
428 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
429 struct llog_handle *llh,
430 struct llog_rec_hdr *h,
432 const struct req_format *format)
434 struct ptlrpc_request *req;
435 struct ost_body *body;
436 struct obd_import *imp;
439 /* Prepare the request */
440 imp = d->opd_obd->u.cli.cl_import;
442 req = ptlrpc_request_alloc(imp, format);
444 RETURN(ERR_PTR(-ENOMEM));
446 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
448 ptlrpc_req_finished(req);
453 * this is a trick: to save on memory allocations we put cookie
454 * into the request, but don't set corresponded flag in o_valid
455 * so that OST doesn't interpret this cookie. once the request
456 * is committed on OST we take cookie from the request and cancel
458 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
460 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
461 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
462 body->oa.o_lcookie.lgc_index = h->lrh_index;
463 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
464 req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
466 req->rq_interpret_reply = osp_sync_interpret;
467 req->rq_commit_cb = osp_sync_request_commit_cb;
470 ptlrpc_request_set_replen(req);
475 static int osp_sync_new_setattr_job(struct osp_device *d,
476 struct llog_handle *llh,
477 struct llog_rec_hdr *h)
479 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
480 struct ptlrpc_request *req;
481 struct ost_body *body;
484 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
486 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
488 RETURN(PTR_ERR(req));
490 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
492 body->oa.o_id = rec->lsr_oid;
493 body->oa.o_seq = rec->lsr_oseq;
494 body->oa.o_uid = rec->lsr_uid;
495 body->oa.o_gid = rec->lsr_gid;
496 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
497 OBD_MD_FLUID | OBD_MD_FLGID;
499 osp_sync_send_new_rpc(d, req);
503 /* Old records may be in old format, so we handle that too */
504 static int osp_sync_new_unlink_job(struct osp_device *d,
505 struct llog_handle *llh,
506 struct llog_rec_hdr *h)
508 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
509 struct ptlrpc_request *req;
510 struct ost_body *body;
513 LASSERT(h->lrh_type == MDS_UNLINK_REC);
515 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
517 RETURN(PTR_ERR(req));
519 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
521 body->oa.o_id = rec->lur_oid;
522 body->oa.o_seq = rec->lur_oseq;
523 body->oa.o_misc = rec->lur_count;
524 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
526 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
528 osp_sync_send_new_rpc(d, req);
532 static int osp_sync_new_unlink64_job(struct osp_device *d,
533 struct llog_handle *llh,
534 struct llog_rec_hdr *h)
536 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
537 struct ptlrpc_request *req;
538 struct ost_body *body;
542 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
544 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
546 RETURN(PTR_ERR(req));
548 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
551 rc = fid_ostid_pack(&rec->lur_fid, &body->oa.o_oi);
554 body->oa.o_misc = rec->lur_count;
555 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID | OBD_MD_FLOBJCOUNT;
557 osp_sync_send_new_rpc(d, req);
561 static int osp_sync_process_record(const struct lu_env *env,
562 struct osp_device *d,
563 struct llog_handle *llh,
564 struct llog_rec_hdr *rec)
566 struct llog_cookie cookie;
569 cookie.lgc_lgl = llh->lgh_id;
570 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
571 cookie.lgc_index = rec->lrh_index;
573 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
574 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
576 /* we're waiting for the record generated by this instance */
577 LASSERT(d->opd_syn_prev_done == 0);
578 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
579 sizeof(gen->lgr_gen))) {
580 CDEBUG(D_HA, "processed all old entries\n");
581 d->opd_syn_prev_done = 1;
584 /* cancel any generation record */
585 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
592 * now we prepare and fill requests to OST, put them on the queue
593 * and fire after next commit callback
596 /* notice we increment counters before sending RPC, to be consistent
597 * in RPC interpret callback which may happen very quickly */
598 spin_lock(&d->opd_syn_lock);
599 d->opd_syn_rpc_in_flight++;
600 d->opd_syn_rpc_in_progress++;
601 spin_unlock(&d->opd_syn_lock);
603 switch (rec->lrh_type) {
604 /* case MDS_UNLINK_REC is kept for compatibility */
606 rc = osp_sync_new_unlink_job(d, llh, rec);
608 case MDS_UNLINK64_REC:
609 rc = osp_sync_new_unlink64_job(d, llh, rec);
611 case MDS_SETATTR64_REC:
612 rc = osp_sync_new_setattr_job(d, llh, rec);
615 CERROR("unknown record type: %x\n", rec->lrh_type);
620 if (likely(rc == 0)) {
621 spin_lock(&d->opd_syn_lock);
622 if (d->opd_syn_prev_done) {
623 LASSERT(d->opd_syn_changes > 0);
624 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
626 * NOTE: it's possible to meet same id if
627 * OST stores few stripes of same file
629 if (rec->lrh_id > d->opd_syn_last_processed_id)
630 d->opd_syn_last_processed_id = rec->lrh_id;
632 d->opd_syn_changes--;
634 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
635 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
636 d->opd_syn_rpc_in_progress);
637 spin_unlock(&d->opd_syn_lock);
639 spin_lock(&d->opd_syn_lock);
640 d->opd_syn_rpc_in_flight--;
641 d->opd_syn_rpc_in_progress--;
642 spin_unlock(&d->opd_syn_lock);
645 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
646 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
650 static void osp_sync_process_committed(const struct lu_env *env,
651 struct osp_device *d)
653 struct obd_device *obd = d->opd_obd;
654 struct obd_import *imp = obd->u.cli.cl_import;
655 struct ost_body *body;
656 struct ptlrpc_request *req, *tmp;
657 struct llog_ctxt *ctxt;
658 struct llog_handle *llh;
664 if (cfs_list_empty(&d->opd_syn_committed_there))
668 * if current status is -ENOSPC (lack of free space on OST)
669 * then we should poll OST immediately once object destroy
671 * notice: we do this upon commit as well because some backends
672 * (like DMU) do not release space right away.
674 if (unlikely(d->opd_pre_status == -ENOSPC))
675 osp_statfs_need_now(d);
678 * now cancel them all
679 * XXX: can we improve this using some batching?
680 * with batch RPC that'll happen automatically?
681 * XXX: can we store ctxt in lod_device and save few cycles ?
683 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
686 llh = ctxt->loc_handle;
689 CFS_INIT_LIST_HEAD(&list);
690 spin_lock(&d->opd_syn_lock);
691 cfs_list_splice(&d->opd_syn_committed_there, &list);
692 CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
693 spin_unlock(&d->opd_syn_lock);
695 cfs_list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
696 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
697 cfs_list_del_init(&req->rq_exp_list);
699 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
702 /* import can be closing, thus all commit cb's are
703 * called we can check committness directly */
704 if (req->rq_transno <= imp->imp_peer_committed_transno) {
705 rc = llog_cat_cancel_records(env, llh, 1,
706 &body->oa.o_lcookie);
708 CERROR("%s: can't cancel record: %d\n",
711 DEBUG_REQ(D_HA, req, "not committed");
714 ptlrpc_req_finished(req);
720 LASSERT(d->opd_syn_rpc_in_progress >= done);
721 spin_lock(&d->opd_syn_lock);
722 d->opd_syn_rpc_in_progress -= done;
723 spin_unlock(&d->opd_syn_lock);
724 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
725 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
726 d->opd_syn_rpc_in_progress);
728 osp_sync_check_for_work(d);
730 /* wake up the thread if requested to stop:
731 * it might be waiting for in-progress to complete */
732 if (unlikely(osp_sync_running(d) == 0))
733 cfs_waitq_signal(&d->opd_syn_waitq);
739 * this is where most of queues processing happens
741 static int osp_sync_process_queues(const struct lu_env *env,
742 struct llog_handle *llh,
743 struct llog_rec_hdr *rec,
746 struct osp_device *d = data;
750 struct l_wait_info lwi = { 0 };
752 if (!osp_sync_running(d)) {
753 CDEBUG(D_HA, "stop llog processing\n");
754 return LLOG_PROC_BREAK;
757 /* process requests committed by OST */
758 osp_sync_process_committed(env, d);
760 /* if we there are changes to be processed and we have
761 * resources for this ... do now */
762 if (osp_sync_can_process_new(d, rec)) {
764 /* ask llog for another record */
765 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
767 d->opd_syn_rpc_in_progress,
768 d->opd_syn_rpc_in_flight);
773 * try to send, in case of disconnection, suspend
774 * processing till we can send this request
777 rc = osp_sync_process_record(env, d, llh, rec);
779 * XXX: probably different handling is needed
780 * for some bugs, like immediate exit or if
784 CERROR("can't send: %d\n", rc);
785 l_wait_event(d->opd_syn_waitq,
786 !osp_sync_running(d) ||
787 osp_sync_has_work(d),
790 } while (rc != 0 && osp_sync_running(d));
796 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
797 osp_sync_remove_from_tracker(d);
799 l_wait_event(d->opd_syn_waitq,
800 !osp_sync_running(d) ||
801 osp_sync_can_process_new(d, rec) ||
802 !cfs_list_empty(&d->opd_syn_committed_there),
808 * this thread runs llog_cat_process() scanner calling our callback
809 * to process llog records. in the callback we implement tricky
810 * state machine as we don't want to start scanning of the llog again
811 * and again, also we don't want to process too many records and send
812 * too many RPCs a time. so, depending on current load (num of changes
813 * being synced to OST) the callback can suspend awaiting for some
814 * new conditions, like syncs completed.
816 * in order to process llog records left by previous boots and to allow
817 * llog_process_thread() to find something (otherwise it'd just exit
818 * immediately) we add a special GENERATATION record on each boot.
820 static int osp_sync_thread(void *_arg)
822 struct osp_device *d = _arg;
823 struct ptlrpc_thread *thread = &d->opd_syn_thread;
824 struct l_wait_info lwi = { 0 };
825 struct llog_ctxt *ctxt;
826 struct obd_device *obd = d->opd_obd;
827 struct llog_handle *llh;
834 rc = lu_env_init(&env, LCT_LOCAL);
836 CERROR("%s: can't initialize env: rc = %d\n",
841 sprintf(pname, "osp-syn-%u", d->opd_index);
842 cfs_daemonize(pname);
844 spin_lock(&d->opd_syn_lock);
845 thread->t_flags = SVC_RUNNING;
846 spin_unlock(&d->opd_syn_lock);
847 cfs_waitq_signal(&thread->t_ctl_waitq);
849 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
851 CERROR("can't get appropriate context\n");
852 GOTO(out, rc = -EINVAL);
855 llh = ctxt->loc_handle;
857 CERROR("can't get llh\n");
859 GOTO(out, rc = -EINVAL);
862 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
863 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
864 "%lu changes, %u in progress, %u in flight: %d\n",
865 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
866 d->opd_syn_rpc_in_flight, rc);
868 /* we don't expect llog_process_thread() to exit till umount */
869 LASSERTF(thread->t_flags != SVC_RUNNING,
870 "%lu changes, %u in progress, %u in flight\n",
871 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
872 d->opd_syn_rpc_in_flight);
874 /* wait till all the requests are completed */
876 while (d->opd_syn_rpc_in_progress > 0) {
877 osp_sync_process_committed(&env, d);
879 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
880 rc = l_wait_event(d->opd_syn_waitq,
881 d->opd_syn_rpc_in_progress == 0,
883 if (rc == -ETIMEDOUT)
885 LASSERTF(count < 10, "%s: %d %d %sempty\n",
886 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
887 d->opd_syn_rpc_in_flight,
888 cfs_list_empty(&d->opd_syn_committed_there) ? "" :"!");
892 llog_cat_close(&env, llh);
893 rc = llog_cleanup(&env, ctxt);
895 CERROR("can't cleanup llog: %d\n", rc);
897 thread->t_flags = SVC_STOPPED;
899 cfs_waitq_signal(&thread->t_ctl_waitq);
900 LASSERTF(d->opd_syn_rpc_in_progress == 0,
901 "%s: %d %d %sempty\n",
902 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
903 d->opd_syn_rpc_in_flight,
904 cfs_list_empty(&d->opd_syn_committed_there) ? "" : "!");
911 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
913 struct osp_thread_info *osi = osp_env_info(env);
914 struct llog_handle *lgh;
915 struct obd_device *obd = d->opd_obd;
916 struct llog_ctxt *ctxt;
924 * open llog corresponding to our OST
926 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
927 obd->obd_lvfs_ctxt.dt = d->opd_storage;
929 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
932 CERROR("%s: can't get id from catalogs: rc = %d\n",
937 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
938 obd->obd_name, d->opd_index, osi->osi_cid.lci_logid.lgl_oid,
939 osi->osi_cid.lci_logid.lgl_oseq,
940 osi->osi_cid.lci_logid.lgl_ogen);
942 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
943 &osp_mds_ost_orig_logops);
947 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
950 if (likely(osi->osi_cid.lci_logid.lgl_oid != 0)) {
951 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
953 /* re-create llog if it is missing */
955 osi->osi_cid.lci_logid.lgl_oid = 0;
957 GOTO(out_cleanup, rc);
960 if (unlikely(osi->osi_cid.lci_logid.lgl_oid == 0)) {
961 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
963 GOTO(out_cleanup, rc);
964 osi->osi_cid.lci_logid = lgh->lgh_id;
967 ctxt->loc_handle = lgh;
969 rc = llog_cat_init_and_process(env, lgh);
973 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
979 * put a mark in the llog till which we'll be processing
980 * old records restless
982 d->opd_syn_generation.mnt_cnt = cfs_time_current();
983 d->opd_syn_generation.conn_cnt = cfs_time_current();
985 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
986 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
988 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
989 sizeof(osi->osi_gen.lgr_gen));
991 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie,
998 llog_cat_close(env, lgh);
1000 llog_cleanup(env, ctxt);
1004 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1006 struct llog_ctxt *ctxt;
1008 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1009 llog_cat_close(env, ctxt->loc_handle);
1010 llog_cleanup(env, ctxt);
1014 * initializes sync component of OSP
1016 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1018 struct l_wait_info lwi = { 0 };
1023 rc = osp_sync_id_traction_init(d);
1028 * initialize llog storing changes
1030 rc = osp_sync_llog_init(env, d);
1032 CERROR("%s: can't initialize llog: rc = %d\n",
1033 d->opd_obd->obd_name, rc);
1038 * Start synchronization thread
1040 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1041 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1042 spin_lock_init(&d->opd_syn_lock);
1043 cfs_waitq_init(&d->opd_syn_waitq);
1044 cfs_waitq_init(&d->opd_syn_thread.t_ctl_waitq);
1045 CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
1047 rc = cfs_create_thread(osp_sync_thread, d, 0);
1049 CERROR("%s: can't start sync thread: rc = %d\n",
1050 d->opd_obd->obd_name, rc);
1054 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1055 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1059 osp_sync_llog_fini(env, d);
1061 osp_sync_id_traction_fini(d);
1065 int osp_sync_fini(struct osp_device *d)
1067 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1071 thread->t_flags = SVC_STOPPING;
1072 cfs_waitq_signal(&d->opd_syn_waitq);
1073 cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1076 * unregister transaction callbacks only when sync thread
1077 * has finished operations with llog
1079 osp_sync_id_traction_fini(d);
1084 static DEFINE_MUTEX(osp_id_tracker_sem);
1085 static CFS_LIST_HEAD(osp_id_tracker_list);
1087 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1089 struct osp_id_tracker *tr = cookie;
1090 struct osp_device *d;
1091 struct osp_txn_info *txn;
1095 txn = osp_txn_info(&th->th_ctx);
1096 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1099 spin_lock(&tr->otr_lock);
1100 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1101 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1102 tr->otr_committed_id, txn->oti_current_id);
1103 tr->otr_committed_id = txn->oti_current_id;
1105 cfs_list_for_each_entry(d, &tr->otr_wakeup_list,
1107 d->opd_syn_last_committed_id = tr->otr_committed_id;
1108 cfs_waitq_signal(&d->opd_syn_waitq);
1111 spin_unlock(&tr->otr_lock);
1114 static int osp_sync_id_traction_init(struct osp_device *d)
1116 struct osp_id_tracker *tr, *found = NULL;
1120 LASSERT(d->opd_storage);
1121 LASSERT(d->opd_syn_tracker == NULL);
1122 CFS_INIT_LIST_HEAD(&d->opd_syn_ontrack);
1124 mutex_lock(&osp_id_tracker_sem);
1125 cfs_list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1126 if (tr->otr_dev == d->opd_storage) {
1127 LASSERT(cfs_atomic_read(&tr->otr_refcount));
1128 cfs_atomic_inc(&tr->otr_refcount);
1129 d->opd_syn_tracker = tr;
1135 if (found == NULL) {
1139 d->opd_syn_tracker = tr;
1140 spin_lock_init(&tr->otr_lock);
1141 tr->otr_dev = d->opd_storage;
1142 tr->otr_next_id = 1;
1143 tr->otr_committed_id = 0;
1144 cfs_atomic_set(&tr->otr_refcount, 1);
1145 CFS_INIT_LIST_HEAD(&tr->otr_wakeup_list);
1146 cfs_list_add(&tr->otr_list, &osp_id_tracker_list);
1147 tr->otr_tx_cb.dtc_txn_commit =
1148 osp_sync_tracker_commit_cb;
1149 tr->otr_tx_cb.dtc_cookie = tr;
1150 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1151 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1155 mutex_unlock(&osp_id_tracker_sem);
1160 static void osp_sync_id_traction_fini(struct osp_device *d)
1162 struct osp_id_tracker *tr;
1167 tr = d->opd_syn_tracker;
1173 osp_sync_remove_from_tracker(d);
1175 mutex_lock(&osp_id_tracker_sem);
1176 if (cfs_atomic_dec_and_test(&tr->otr_refcount)) {
1177 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1178 LASSERT(cfs_list_empty(&tr->otr_wakeup_list));
1179 cfs_list_del(&tr->otr_list);
1181 d->opd_syn_tracker = NULL;
1183 mutex_unlock(&osp_id_tracker_sem);
1189 * generates id for the tracker
1191 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1193 struct osp_id_tracker *tr;
1195 tr = d->opd_syn_tracker;
1198 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1199 spin_lock(&tr->otr_lock);
1200 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1201 spin_unlock(&tr->otr_lock);
1202 CERROR("%s: next %u, last synced %lu\n",
1203 d->opd_obd->obd_name, tr->otr_next_id,
1204 d->opd_syn_last_used_id);
1209 id = tr->otr_next_id++;
1210 if (id > d->opd_syn_last_used_id)
1211 d->opd_syn_last_used_id = id;
1212 if (cfs_list_empty(&d->opd_syn_ontrack))
1213 cfs_list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1214 spin_unlock(&tr->otr_lock);
1215 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1220 static void osp_sync_remove_from_tracker(struct osp_device *d)
1222 struct osp_id_tracker *tr;
1224 tr = d->opd_syn_tracker;
1227 if (cfs_list_empty(&d->opd_syn_ontrack))
1230 spin_lock(&tr->otr_lock);
1231 cfs_list_del_init(&d->opd_syn_ontrack);
1232 spin_unlock(&tr->otr_lock);