4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osp/osp_sync.c
38 * Lustre OST Proxy Device
40 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
41 * Author: Mikhail Pershin <mike.pershin@intel.com>
44 #define DEBUG_SUBSYSTEM S_MDS
46 #include <lustre_log.h>
47 #include "osp_internal.h"
49 static int osp_sync_id_traction_init(struct osp_device *d);
50 static void osp_sync_id_traction_fini(struct osp_device *d);
51 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
52 static void osp_sync_remove_from_tracker(struct osp_device *d);
55 * this is a components of OSP implementing synchronization between MDS and OST
56 * it llogs all interesting changes (currently it's uig/gid change and object
57 * destroy) atomically, then makes sure changes hit OST storage
59 * we have 4 queues of work:
61 * the first queue is llog itself, once read a change is stored in 2nd queue
62 * in form of RPC (but RPC isn't fired yet).
64 * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
65 * commit. once change is committed locally it migrates onto 3rd queue.
67 * the third queue (opd_syn_committed_here) holds changes committed locally,
68 * but not sent to OST (as the pipe can be full). once pipe becomes non-full
69 * we take a change from the queue and fire corresponded RPC.
71 * once RPC is reported committed by OST (using regular last_committed mech.)
72 * the change jumps into 4th queue (opd_syn_committed_there), now we can
73 * cancel corresponded llog record and release RPC
75 * opd_syn_changes is a number of unread llog records (to be processed).
76 * notice this number doesn't include llog records from previous boots.
77 * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
79 * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
80 * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
81 * too much memory -- how to deal with 1000th OSTs ? batching could help?
83 * opd_syn_rpc_in_flight is a number of RPC in flight.
84 * we control this with OSP_MAX_IN_FLIGHT
87 /* XXX: do math to learn reasonable threshold
88 * should it be ~ number of changes fitting bulk? */
90 #define OSP_SYN_THRESHOLD 10
91 #define OSP_MAX_IN_FLIGHT 8
92 #define OSP_MAX_IN_PROGRESS 4096
94 #define OSP_JOB_MAGIC 0x26112005
96 static inline int osp_sync_running(struct osp_device *d)
98 return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
101 static inline int osp_sync_stopped(struct osp_device *d)
103 return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
106 static inline int osp_sync_has_new_job(struct osp_device *d)
108 return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
109 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
110 || (d->opd_syn_prev_done == 0);
113 static inline int osp_sync_low_in_progress(struct osp_device *d)
115 return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
118 static inline int osp_sync_low_in_flight(struct osp_device *d)
120 return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
123 static inline int osp_sync_has_work(struct osp_device *d)
125 /* has new/old changes and low in-progress? */
126 if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
127 osp_sync_low_in_flight(d) && d->opd_imp_connected)
130 /* has remotely committed? */
131 if (!cfs_list_empty(&d->opd_syn_committed_there))
137 #define osp_sync_check_for_work(d) \
139 if (osp_sync_has_work(d)) { \
140 cfs_waitq_signal(&d->opd_syn_waitq); \
144 void __osp_sync_check_for_work(struct osp_device *d)
146 osp_sync_check_for_work(d);
149 static inline int osp_sync_can_process_new(struct osp_device *d,
150 struct llog_rec_hdr *rec)
154 if (!osp_sync_low_in_progress(d))
156 if (!osp_sync_low_in_flight(d))
158 if (!d->opd_imp_connected)
160 if (d->opd_syn_prev_done == 0)
162 if (d->opd_syn_changes == 0)
164 if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
169 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
170 llog_op_type type, struct thandle *th)
172 struct osp_thread_info *osi = osp_env_info(env);
173 struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
174 struct llog_ctxt *ctxt;
179 /* it's a layering violation, to access internals of th,
180 * but we can do this as a sanity check, for a while */
181 LASSERT(th->th_dev == d->opd_storage);
184 case MDS_UNLINK64_REC:
185 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
187 case MDS_SETATTR64_REC:
188 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
194 /* we want ->dt_trans_start() to allocate per-thandle structure */
195 th->th_tags |= LCT_OSP_THREAD;
197 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
200 rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
206 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
207 const struct lu_fid *fid, llog_op_type type,
208 int count, struct thandle *th,
209 const struct lu_attr *attr)
211 struct osp_thread_info *osi = osp_env_info(env);
212 struct llog_ctxt *ctxt;
213 struct osp_txn_info *txn;
218 /* it's a layering violation, to access internals of th,
219 * but we can do this as a sanity check, for a while */
220 LASSERT(th->th_dev == d->opd_storage);
223 case MDS_UNLINK64_REC:
224 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
225 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
226 osi->osi_unlink.lur_fid = *fid;
227 osi->osi_unlink.lur_count = count;
229 case MDS_SETATTR64_REC:
230 rc = fid_to_ostid(fid, &osi->osi_oi);
232 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
233 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
234 osi->osi_setattr.lsr_oi = osi->osi_oi;
236 osi->osi_setattr.lsr_uid = attr->la_uid;
237 osi->osi_setattr.lsr_gid = attr->la_gid;
243 txn = osp_txn_info(&th->th_ctx);
246 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
247 osi->osi_hdr.lrh_id = txn->oti_current_id;
249 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
252 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
256 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
257 d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
258 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
259 (unsigned long) osi->osi_cookie.lgc_index, rc);
264 if (likely(rc == 0)) {
265 spin_lock(&d->opd_syn_lock);
266 d->opd_syn_changes++;
267 spin_unlock(&d->opd_syn_lock);
273 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
274 llog_op_type type, struct thandle *th,
275 const struct lu_attr *attr)
277 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
278 lu_object_fid(&o->opo_obj.do_lu), type, 1,
282 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
283 struct lu_fid *fid, int lost, struct thandle *th)
285 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
289 * it's quite obvious we can't maintain all the structures in the memory:
290 * while OST is down, MDS can be processing thousands and thousands of unlinks
291 * filling persistent llogs and in-core respresentation
293 * this doesn't scale at all. so we need basically the following:
294 * a) destroy/setattr append llog records
295 * b) once llog has grown to X records, we process first Y committed records
297 * once record R is found via llog_process(), it becomes committed after any
298 * subsequent commit callback (at the most)
302 * called for each atomic on-disk change (not once per transaction batch)
303 * and goes over the list
304 * XXX: should be optimized?
308 * called for each RPC reported committed
310 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
312 struct osp_device *d = req->rq_cb_data;
314 CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
316 if (unlikely(req->rq_transno == 0))
319 /* do not do any opd_dyn_rpc_* accounting here
320 * it's done in osp_sync_interpret sooner or later */
323 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
324 LASSERT(cfs_list_empty(&req->rq_exp_list));
326 ptlrpc_request_addref(req);
328 spin_lock(&d->opd_syn_lock);
329 cfs_list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
330 spin_unlock(&d->opd_syn_lock);
332 /* XXX: some batching wouldn't hurt */
333 cfs_waitq_signal(&d->opd_syn_waitq);
336 static int osp_sync_interpret(const struct lu_env *env,
337 struct ptlrpc_request *req, void *aa, int rc)
339 struct osp_device *d = req->rq_cb_data;
341 if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
342 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
343 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
346 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
347 cfs_atomic_read(&req->rq_refcount),
348 rc, (unsigned) req->rq_transno);
349 LASSERT(rc || req->rq_transno);
353 * we tried to destroy object or update attributes,
354 * but object doesn't exist anymore - cancell llog record
356 LASSERT(req->rq_transno == 0);
357 LASSERT(cfs_list_empty(&req->rq_exp_list));
359 ptlrpc_request_addref(req);
361 spin_lock(&d->opd_syn_lock);
362 cfs_list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
363 spin_unlock(&d->opd_syn_lock);
365 cfs_waitq_signal(&d->opd_syn_waitq);
367 struct obd_import *imp = req->rq_import;
369 * error happened, we'll try to repeat on next boot ?
371 LASSERTF(req->rq_transno == 0 ||
372 req->rq_import_generation < imp->imp_generation,
373 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
374 req->rq_transno, rc, req->rq_import_generation,
375 imp->imp_generation);
376 if (req->rq_transno == 0) {
377 /* this is the last time we see the request
378 * if transno is not zero, then commit cb
379 * will be called at some point */
380 LASSERT(d->opd_syn_rpc_in_progress > 0);
381 spin_lock(&d->opd_syn_lock);
382 d->opd_syn_rpc_in_progress--;
383 spin_unlock(&d->opd_syn_lock);
386 cfs_waitq_signal(&d->opd_syn_waitq);
387 } else if (unlikely(d->opd_pre_status == -ENOSPC)) {
389 * if current status is -ENOSPC (lack of free space on OST)
390 * then we should poll OST immediately once object destroy
393 osp_statfs_need_now(d);
396 LASSERT(d->opd_syn_rpc_in_flight > 0);
397 spin_lock(&d->opd_syn_lock);
398 d->opd_syn_rpc_in_flight--;
399 spin_unlock(&d->opd_syn_lock);
400 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
401 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
402 d->opd_syn_rpc_in_progress);
404 osp_sync_check_for_work(d);
410 * the function walks through list of committed locally changes
411 * and send them to RPC until the pipe is full
413 static void osp_sync_send_new_rpc(struct osp_device *d,
414 struct ptlrpc_request *req)
416 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
417 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
419 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
422 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
423 struct llog_handle *llh,
424 struct llog_rec_hdr *h,
426 const struct req_format *format)
428 struct ptlrpc_request *req;
429 struct ost_body *body;
430 struct obd_import *imp;
433 /* Prepare the request */
434 imp = d->opd_obd->u.cli.cl_import;
436 req = ptlrpc_request_alloc(imp, format);
438 RETURN(ERR_PTR(-ENOMEM));
440 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
442 ptlrpc_req_finished(req);
447 * this is a trick: to save on memory allocations we put cookie
448 * into the request, but don't set corresponded flag in o_valid
449 * so that OST doesn't interpret this cookie. once the request
450 * is committed on OST we take cookie from the request and cancel
452 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
454 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
455 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
456 body->oa.o_lcookie.lgc_index = h->lrh_index;
457 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
458 req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
460 req->rq_interpret_reply = osp_sync_interpret;
461 req->rq_commit_cb = osp_sync_request_commit_cb;
464 ptlrpc_request_set_replen(req);
469 static int osp_sync_new_setattr_job(struct osp_device *d,
470 struct llog_handle *llh,
471 struct llog_rec_hdr *h)
473 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
474 struct ptlrpc_request *req;
475 struct ost_body *body;
478 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
480 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
482 RETURN(PTR_ERR(req));
484 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
486 body->oa.o_oi = rec->lsr_oi;
487 body->oa.o_uid = rec->lsr_uid;
488 body->oa.o_gid = rec->lsr_gid;
489 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
490 OBD_MD_FLUID | OBD_MD_FLGID;
492 osp_sync_send_new_rpc(d, req);
496 /* Old records may be in old format, so we handle that too */
497 static int osp_sync_new_unlink_job(struct osp_device *d,
498 struct llog_handle *llh,
499 struct llog_rec_hdr *h)
501 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
502 struct ptlrpc_request *req;
503 struct ost_body *body;
506 LASSERT(h->lrh_type == MDS_UNLINK_REC);
508 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
510 RETURN(PTR_ERR(req));
512 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
514 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
515 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
516 body->oa.o_misc = rec->lur_count;
517 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
519 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
521 osp_sync_send_new_rpc(d, req);
525 static int osp_sync_new_unlink64_job(struct osp_device *d,
526 struct llog_handle *llh,
527 struct llog_rec_hdr *h)
529 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
530 struct ptlrpc_request *req;
531 struct ost_body *body;
535 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
537 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
539 RETURN(PTR_ERR(req));
541 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
544 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
547 body->oa.o_misc = rec->lur_count;
548 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID | OBD_MD_FLOBJCOUNT;
550 osp_sync_send_new_rpc(d, req);
554 static int osp_sync_process_record(const struct lu_env *env,
555 struct osp_device *d,
556 struct llog_handle *llh,
557 struct llog_rec_hdr *rec)
559 struct llog_cookie cookie;
562 cookie.lgc_lgl = llh->lgh_id;
563 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
564 cookie.lgc_index = rec->lrh_index;
566 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
567 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
569 /* we're waiting for the record generated by this instance */
570 LASSERT(d->opd_syn_prev_done == 0);
571 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
572 sizeof(gen->lgr_gen))) {
573 CDEBUG(D_HA, "processed all old entries\n");
574 d->opd_syn_prev_done = 1;
577 /* cancel any generation record */
578 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
585 * now we prepare and fill requests to OST, put them on the queue
586 * and fire after next commit callback
589 /* notice we increment counters before sending RPC, to be consistent
590 * in RPC interpret callback which may happen very quickly */
591 spin_lock(&d->opd_syn_lock);
592 d->opd_syn_rpc_in_flight++;
593 d->opd_syn_rpc_in_progress++;
594 spin_unlock(&d->opd_syn_lock);
596 switch (rec->lrh_type) {
597 /* case MDS_UNLINK_REC is kept for compatibility */
599 rc = osp_sync_new_unlink_job(d, llh, rec);
601 case MDS_UNLINK64_REC:
602 rc = osp_sync_new_unlink64_job(d, llh, rec);
604 case MDS_SETATTR64_REC:
605 rc = osp_sync_new_setattr_job(d, llh, rec);
608 CERROR("unknown record type: %x\n", rec->lrh_type);
613 if (likely(rc == 0)) {
614 spin_lock(&d->opd_syn_lock);
615 if (d->opd_syn_prev_done) {
616 LASSERT(d->opd_syn_changes > 0);
617 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
619 * NOTE: it's possible to meet same id if
620 * OST stores few stripes of same file
622 if (rec->lrh_id > d->opd_syn_last_processed_id)
623 d->opd_syn_last_processed_id = rec->lrh_id;
625 d->opd_syn_changes--;
627 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
628 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
629 d->opd_syn_rpc_in_progress);
630 spin_unlock(&d->opd_syn_lock);
632 spin_lock(&d->opd_syn_lock);
633 d->opd_syn_rpc_in_flight--;
634 d->opd_syn_rpc_in_progress--;
635 spin_unlock(&d->opd_syn_lock);
638 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
639 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
643 static void osp_sync_process_committed(const struct lu_env *env,
644 struct osp_device *d)
646 struct obd_device *obd = d->opd_obd;
647 struct obd_import *imp = obd->u.cli.cl_import;
648 struct ost_body *body;
649 struct ptlrpc_request *req, *tmp;
650 struct llog_ctxt *ctxt;
651 struct llog_handle *llh;
657 if (cfs_list_empty(&d->opd_syn_committed_there))
661 * if current status is -ENOSPC (lack of free space on OST)
662 * then we should poll OST immediately once object destroy
664 * notice: we do this upon commit as well because some backends
665 * (like DMU) do not release space right away.
667 if (unlikely(d->opd_pre_status == -ENOSPC))
668 osp_statfs_need_now(d);
671 * now cancel them all
672 * XXX: can we improve this using some batching?
673 * with batch RPC that'll happen automatically?
674 * XXX: can we store ctxt in lod_device and save few cycles ?
676 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
679 llh = ctxt->loc_handle;
682 CFS_INIT_LIST_HEAD(&list);
683 spin_lock(&d->opd_syn_lock);
684 cfs_list_splice(&d->opd_syn_committed_there, &list);
685 CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
686 spin_unlock(&d->opd_syn_lock);
688 cfs_list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
689 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
690 cfs_list_del_init(&req->rq_exp_list);
692 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
695 /* import can be closing, thus all commit cb's are
696 * called we can check committness directly */
697 if (req->rq_transno <= imp->imp_peer_committed_transno) {
698 rc = llog_cat_cancel_records(env, llh, 1,
699 &body->oa.o_lcookie);
701 CERROR("%s: can't cancel record: %d\n",
704 DEBUG_REQ(D_HA, req, "not committed");
707 ptlrpc_req_finished(req);
713 LASSERT(d->opd_syn_rpc_in_progress >= done);
714 spin_lock(&d->opd_syn_lock);
715 d->opd_syn_rpc_in_progress -= done;
716 spin_unlock(&d->opd_syn_lock);
717 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
718 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
719 d->opd_syn_rpc_in_progress);
721 osp_sync_check_for_work(d);
723 /* wake up the thread if requested to stop:
724 * it might be waiting for in-progress to complete */
725 if (unlikely(osp_sync_running(d) == 0))
726 cfs_waitq_signal(&d->opd_syn_waitq);
732 * this is where most of queues processing happens
734 static int osp_sync_process_queues(const struct lu_env *env,
735 struct llog_handle *llh,
736 struct llog_rec_hdr *rec,
739 struct osp_device *d = data;
743 struct l_wait_info lwi = { 0 };
745 if (!osp_sync_running(d)) {
746 CDEBUG(D_HA, "stop llog processing\n");
747 return LLOG_PROC_BREAK;
750 /* process requests committed by OST */
751 osp_sync_process_committed(env, d);
753 /* if we there are changes to be processed and we have
754 * resources for this ... do now */
755 if (osp_sync_can_process_new(d, rec)) {
757 /* ask llog for another record */
758 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
760 d->opd_syn_rpc_in_progress,
761 d->opd_syn_rpc_in_flight);
766 * try to send, in case of disconnection, suspend
767 * processing till we can send this request
770 rc = osp_sync_process_record(env, d, llh, rec);
772 * XXX: probably different handling is needed
773 * for some bugs, like immediate exit or if
777 CERROR("can't send: %d\n", rc);
778 l_wait_event(d->opd_syn_waitq,
779 !osp_sync_running(d) ||
780 osp_sync_has_work(d),
783 } while (rc != 0 && osp_sync_running(d));
789 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
790 osp_sync_remove_from_tracker(d);
792 l_wait_event(d->opd_syn_waitq,
793 !osp_sync_running(d) ||
794 osp_sync_can_process_new(d, rec) ||
795 !cfs_list_empty(&d->opd_syn_committed_there),
801 * this thread runs llog_cat_process() scanner calling our callback
802 * to process llog records. in the callback we implement tricky
803 * state machine as we don't want to start scanning of the llog again
804 * and again, also we don't want to process too many records and send
805 * too many RPCs a time. so, depending on current load (num of changes
806 * being synced to OST) the callback can suspend awaiting for some
807 * new conditions, like syncs completed.
809 * in order to process llog records left by previous boots and to allow
810 * llog_process_thread() to find something (otherwise it'd just exit
811 * immediately) we add a special GENERATATION record on each boot.
813 static int osp_sync_thread(void *_arg)
815 struct osp_device *d = _arg;
816 struct ptlrpc_thread *thread = &d->opd_syn_thread;
817 struct l_wait_info lwi = { 0 };
818 struct llog_ctxt *ctxt;
819 struct obd_device *obd = d->opd_obd;
820 struct llog_handle *llh;
826 rc = lu_env_init(&env, LCT_LOCAL);
828 CERROR("%s: can't initialize env: rc = %d\n",
833 spin_lock(&d->opd_syn_lock);
834 thread->t_flags = SVC_RUNNING;
835 spin_unlock(&d->opd_syn_lock);
836 cfs_waitq_signal(&thread->t_ctl_waitq);
838 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
840 CERROR("can't get appropriate context\n");
841 GOTO(out, rc = -EINVAL);
844 llh = ctxt->loc_handle;
846 CERROR("can't get llh\n");
848 GOTO(out, rc = -EINVAL);
851 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
852 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
853 "%lu changes, %u in progress, %u in flight: %d\n",
854 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
855 d->opd_syn_rpc_in_flight, rc);
857 /* we don't expect llog_process_thread() to exit till umount */
858 LASSERTF(thread->t_flags != SVC_RUNNING,
859 "%lu changes, %u in progress, %u in flight\n",
860 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
861 d->opd_syn_rpc_in_flight);
863 /* wait till all the requests are completed */
865 while (d->opd_syn_rpc_in_progress > 0) {
866 osp_sync_process_committed(&env, d);
868 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
869 rc = l_wait_event(d->opd_syn_waitq,
870 d->opd_syn_rpc_in_progress == 0,
872 if (rc == -ETIMEDOUT)
874 LASSERTF(count < 10, "%s: %d %d %sempty\n",
875 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
876 d->opd_syn_rpc_in_flight,
877 cfs_list_empty(&d->opd_syn_committed_there) ? "" :"!");
881 llog_cat_close(&env, llh);
882 rc = llog_cleanup(&env, ctxt);
884 CERROR("can't cleanup llog: %d\n", rc);
886 LASSERTF(d->opd_syn_rpc_in_progress == 0,
887 "%s: %d %d %sempty\n",
888 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
889 d->opd_syn_rpc_in_flight,
890 cfs_list_empty(&d->opd_syn_committed_there) ? "" : "!");
892 thread->t_flags = SVC_STOPPED;
894 cfs_waitq_signal(&thread->t_ctl_waitq);
901 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
903 struct osp_thread_info *osi = osp_env_info(env);
904 struct llog_handle *lgh = NULL;
905 struct obd_device *obd = d->opd_obd;
906 struct llog_ctxt *ctxt;
914 * open llog corresponding to our OST
916 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
917 obd->obd_lvfs_ctxt.dt = d->opd_storage;
919 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
922 CERROR("%s: can't get id from catalogs: rc = %d\n",
927 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
928 obd->obd_name, d->opd_index,
929 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
930 osi->osi_cid.lci_logid.lgl_ogen);
932 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
933 &osp_mds_ost_orig_logops);
937 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
940 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
941 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
943 /* re-create llog if it is missing */
945 logid_set_id(&osi->osi_cid.lci_logid, 0);
947 GOTO(out_cleanup, rc);
950 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
951 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
953 GOTO(out_cleanup, rc);
954 osi->osi_cid.lci_logid = lgh->lgh_id;
957 LASSERT(lgh != NULL);
958 ctxt->loc_handle = lgh;
960 rc = llog_cat_init_and_process(env, lgh);
964 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
970 * put a mark in the llog till which we'll be processing
971 * old records restless
973 d->opd_syn_generation.mnt_cnt = cfs_time_current();
974 d->opd_syn_generation.conn_cnt = cfs_time_current();
976 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
977 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
979 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
980 sizeof(osi->osi_gen.lgr_gen));
982 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie,
989 llog_cat_close(env, lgh);
991 llog_cleanup(env, ctxt);
995 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
997 struct llog_ctxt *ctxt;
999 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1000 llog_cat_close(env, ctxt->loc_handle);
1001 llog_cleanup(env, ctxt);
1005 * initializes sync component of OSP
1007 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1009 struct l_wait_info lwi = { 0 };
1014 rc = osp_sync_id_traction_init(d);
1019 * initialize llog storing changes
1021 rc = osp_sync_llog_init(env, d);
1023 CERROR("%s: can't initialize llog: rc = %d\n",
1024 d->opd_obd->obd_name, rc);
1029 * Start synchronization thread
1031 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1032 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1033 spin_lock_init(&d->opd_syn_lock);
1034 cfs_waitq_init(&d->opd_syn_waitq);
1035 cfs_waitq_init(&d->opd_syn_thread.t_ctl_waitq);
1036 CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
1038 rc = PTR_ERR(kthread_run(osp_sync_thread, d,
1039 "osp-syn-%u", d->opd_index));
1040 if (IS_ERR_VALUE(rc)) {
1041 CERROR("%s: can't start sync thread: rc = %d\n",
1042 d->opd_obd->obd_name, rc);
1046 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1047 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1051 osp_sync_llog_fini(env, d);
1053 osp_sync_id_traction_fini(d);
1057 int osp_sync_fini(struct osp_device *d)
1059 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1063 thread->t_flags = SVC_STOPPING;
1064 cfs_waitq_signal(&d->opd_syn_waitq);
1065 cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1068 * unregister transaction callbacks only when sync thread
1069 * has finished operations with llog
1071 osp_sync_id_traction_fini(d);
1076 static DEFINE_MUTEX(osp_id_tracker_sem);
1077 static CFS_LIST_HEAD(osp_id_tracker_list);
1079 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1081 struct osp_id_tracker *tr = cookie;
1082 struct osp_device *d;
1083 struct osp_txn_info *txn;
1087 txn = osp_txn_info(&th->th_ctx);
1088 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1091 spin_lock(&tr->otr_lock);
1092 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1093 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1094 tr->otr_committed_id, txn->oti_current_id);
1095 tr->otr_committed_id = txn->oti_current_id;
1097 cfs_list_for_each_entry(d, &tr->otr_wakeup_list,
1099 d->opd_syn_last_committed_id = tr->otr_committed_id;
1100 cfs_waitq_signal(&d->opd_syn_waitq);
1103 spin_unlock(&tr->otr_lock);
1106 static int osp_sync_id_traction_init(struct osp_device *d)
1108 struct osp_id_tracker *tr, *found = NULL;
1112 LASSERT(d->opd_storage);
1113 LASSERT(d->opd_syn_tracker == NULL);
1114 CFS_INIT_LIST_HEAD(&d->opd_syn_ontrack);
1116 mutex_lock(&osp_id_tracker_sem);
1117 cfs_list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1118 if (tr->otr_dev == d->opd_storage) {
1119 LASSERT(cfs_atomic_read(&tr->otr_refcount));
1120 cfs_atomic_inc(&tr->otr_refcount);
1121 d->opd_syn_tracker = tr;
1127 if (found == NULL) {
1131 d->opd_syn_tracker = tr;
1132 spin_lock_init(&tr->otr_lock);
1133 tr->otr_dev = d->opd_storage;
1134 tr->otr_next_id = 1;
1135 tr->otr_committed_id = 0;
1136 cfs_atomic_set(&tr->otr_refcount, 1);
1137 CFS_INIT_LIST_HEAD(&tr->otr_wakeup_list);
1138 cfs_list_add(&tr->otr_list, &osp_id_tracker_list);
1139 tr->otr_tx_cb.dtc_txn_commit =
1140 osp_sync_tracker_commit_cb;
1141 tr->otr_tx_cb.dtc_cookie = tr;
1142 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1143 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1147 mutex_unlock(&osp_id_tracker_sem);
1152 static void osp_sync_id_traction_fini(struct osp_device *d)
1154 struct osp_id_tracker *tr;
1159 tr = d->opd_syn_tracker;
1165 osp_sync_remove_from_tracker(d);
1167 mutex_lock(&osp_id_tracker_sem);
1168 if (cfs_atomic_dec_and_test(&tr->otr_refcount)) {
1169 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1170 LASSERT(cfs_list_empty(&tr->otr_wakeup_list));
1171 cfs_list_del(&tr->otr_list);
1173 d->opd_syn_tracker = NULL;
1175 mutex_unlock(&osp_id_tracker_sem);
1181 * generates id for the tracker
1183 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1185 struct osp_id_tracker *tr;
1187 tr = d->opd_syn_tracker;
1190 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1191 spin_lock(&tr->otr_lock);
1192 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1193 spin_unlock(&tr->otr_lock);
1194 CERROR("%s: next %u, last synced %lu\n",
1195 d->opd_obd->obd_name, tr->otr_next_id,
1196 d->opd_syn_last_used_id);
1201 id = tr->otr_next_id++;
1202 if (id > d->opd_syn_last_used_id)
1203 d->opd_syn_last_used_id = id;
1204 if (cfs_list_empty(&d->opd_syn_ontrack))
1205 cfs_list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1206 spin_unlock(&tr->otr_lock);
1207 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1212 static void osp_sync_remove_from_tracker(struct osp_device *d)
1214 struct osp_id_tracker *tr;
1216 tr = d->opd_syn_tracker;
1219 if (cfs_list_empty(&d->opd_syn_ontrack))
1222 spin_lock(&tr->otr_lock);
1223 cfs_list_del_init(&d->opd_syn_ontrack);
1224 spin_unlock(&tr->otr_lock);