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 wake_up(&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 wake_up(&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);
351 LASSERT(d->opd_pre != NULL);
355 * we tried to destroy object or update attributes,
356 * but object doesn't exist anymore - cancell llog record
358 LASSERT(req->rq_transno == 0);
359 LASSERT(cfs_list_empty(&req->rq_exp_list));
361 ptlrpc_request_addref(req);
363 spin_lock(&d->opd_syn_lock);
364 cfs_list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
365 spin_unlock(&d->opd_syn_lock);
367 wake_up(&d->opd_syn_waitq);
369 struct obd_import *imp = req->rq_import;
371 * error happened, we'll try to repeat on next boot ?
373 LASSERTF(req->rq_transno == 0 ||
374 req->rq_import_generation < imp->imp_generation,
375 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
376 req->rq_transno, rc, req->rq_import_generation,
377 imp->imp_generation);
378 if (req->rq_transno == 0) {
379 /* this is the last time we see the request
380 * if transno is not zero, then commit cb
381 * will be called at some point */
382 LASSERT(d->opd_syn_rpc_in_progress > 0);
383 spin_lock(&d->opd_syn_lock);
384 d->opd_syn_rpc_in_progress--;
385 spin_unlock(&d->opd_syn_lock);
388 wake_up(&d->opd_syn_waitq);
389 } else if (unlikely(d->opd_pre_status == -ENOSPC)) {
391 * if current status is -ENOSPC (lack of free space on OST)
392 * then we should poll OST immediately once object destroy
395 osp_statfs_need_now(d);
398 LASSERT(d->opd_syn_rpc_in_flight > 0);
399 spin_lock(&d->opd_syn_lock);
400 d->opd_syn_rpc_in_flight--;
401 spin_unlock(&d->opd_syn_lock);
402 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
403 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
404 d->opd_syn_rpc_in_progress);
406 osp_sync_check_for_work(d);
412 * the function walks through list of committed locally changes
413 * and send them to RPC until the pipe is full
415 static void osp_sync_send_new_rpc(struct osp_device *d,
416 struct ptlrpc_request *req)
418 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
419 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
421 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
424 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
425 struct llog_handle *llh,
426 struct llog_rec_hdr *h,
428 const struct req_format *format)
430 struct ptlrpc_request *req;
431 struct ost_body *body;
432 struct obd_import *imp;
435 /* Prepare the request */
436 imp = d->opd_obd->u.cli.cl_import;
438 req = ptlrpc_request_alloc(imp, format);
440 RETURN(ERR_PTR(-ENOMEM));
442 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
444 ptlrpc_req_finished(req);
449 * this is a trick: to save on memory allocations we put cookie
450 * into the request, but don't set corresponded flag in o_valid
451 * so that OST doesn't interpret this cookie. once the request
452 * is committed on OST we take cookie from the request and cancel
454 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
456 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
457 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
458 body->oa.o_lcookie.lgc_index = h->lrh_index;
459 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
460 req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
462 req->rq_interpret_reply = osp_sync_interpret;
463 req->rq_commit_cb = osp_sync_request_commit_cb;
466 ptlrpc_request_set_replen(req);
471 static int osp_sync_new_setattr_job(struct osp_device *d,
472 struct llog_handle *llh,
473 struct llog_rec_hdr *h)
475 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
476 struct ptlrpc_request *req;
477 struct ost_body *body;
480 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
482 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
484 RETURN(PTR_ERR(req));
486 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
488 body->oa.o_oi = rec->lsr_oi;
489 body->oa.o_uid = rec->lsr_uid;
490 body->oa.o_gid = rec->lsr_gid;
491 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
492 OBD_MD_FLUID | OBD_MD_FLGID;
494 osp_sync_send_new_rpc(d, req);
498 /* Old records may be in old format, so we handle that too */
499 static int osp_sync_new_unlink_job(struct osp_device *d,
500 struct llog_handle *llh,
501 struct llog_rec_hdr *h)
503 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
504 struct ptlrpc_request *req;
505 struct ost_body *body;
508 LASSERT(h->lrh_type == MDS_UNLINK_REC);
510 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
512 RETURN(PTR_ERR(req));
514 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
516 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
517 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
518 body->oa.o_misc = rec->lur_count;
519 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
521 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
523 osp_sync_send_new_rpc(d, req);
527 static int osp_sync_new_unlink64_job(struct osp_device *d,
528 struct llog_handle *llh,
529 struct llog_rec_hdr *h)
531 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
532 struct ptlrpc_request *req;
533 struct ost_body *body;
537 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
539 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
541 RETURN(PTR_ERR(req));
543 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
546 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
549 body->oa.o_misc = rec->lur_count;
550 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID | OBD_MD_FLOBJCOUNT;
552 osp_sync_send_new_rpc(d, req);
556 static int osp_sync_process_record(const struct lu_env *env,
557 struct osp_device *d,
558 struct llog_handle *llh,
559 struct llog_rec_hdr *rec)
561 struct llog_cookie cookie;
564 cookie.lgc_lgl = llh->lgh_id;
565 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
566 cookie.lgc_index = rec->lrh_index;
568 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
569 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
571 /* we're waiting for the record generated by this instance */
572 LASSERT(d->opd_syn_prev_done == 0);
573 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
574 sizeof(gen->lgr_gen))) {
575 CDEBUG(D_HA, "processed all old entries\n");
576 d->opd_syn_prev_done = 1;
579 /* cancel any generation record */
580 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
587 * now we prepare and fill requests to OST, put them on the queue
588 * and fire after next commit callback
591 /* notice we increment counters before sending RPC, to be consistent
592 * in RPC interpret callback which may happen very quickly */
593 spin_lock(&d->opd_syn_lock);
594 d->opd_syn_rpc_in_flight++;
595 d->opd_syn_rpc_in_progress++;
596 spin_unlock(&d->opd_syn_lock);
598 switch (rec->lrh_type) {
599 /* case MDS_UNLINK_REC is kept for compatibility */
601 rc = osp_sync_new_unlink_job(d, llh, rec);
603 case MDS_UNLINK64_REC:
604 rc = osp_sync_new_unlink64_job(d, llh, rec);
606 case MDS_SETATTR64_REC:
607 rc = osp_sync_new_setattr_job(d, llh, rec);
610 CERROR("unknown record type: %x\n", rec->lrh_type);
615 if (likely(rc == 0)) {
616 spin_lock(&d->opd_syn_lock);
617 if (d->opd_syn_prev_done) {
618 LASSERT(d->opd_syn_changes > 0);
619 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
621 * NOTE: it's possible to meet same id if
622 * OST stores few stripes of same file
624 if (rec->lrh_id > d->opd_syn_last_processed_id)
625 d->opd_syn_last_processed_id = rec->lrh_id;
627 d->opd_syn_changes--;
629 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
630 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
631 d->opd_syn_rpc_in_progress);
632 spin_unlock(&d->opd_syn_lock);
634 spin_lock(&d->opd_syn_lock);
635 d->opd_syn_rpc_in_flight--;
636 d->opd_syn_rpc_in_progress--;
637 spin_unlock(&d->opd_syn_lock);
640 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
641 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
645 static void osp_sync_process_committed(const struct lu_env *env,
646 struct osp_device *d)
648 struct obd_device *obd = d->opd_obd;
649 struct obd_import *imp = obd->u.cli.cl_import;
650 struct ost_body *body;
651 struct ptlrpc_request *req, *tmp;
652 struct llog_ctxt *ctxt;
653 struct llog_handle *llh;
659 if (cfs_list_empty(&d->opd_syn_committed_there))
663 * if current status is -ENOSPC (lack of free space on OST)
664 * then we should poll OST immediately once object destroy
666 * notice: we do this upon commit as well because some backends
667 * (like DMU) do not release space right away.
669 LASSERT(d->opd_pre != NULL);
670 if (unlikely(d->opd_pre_status == -ENOSPC))
671 osp_statfs_need_now(d);
674 * now cancel them all
675 * XXX: can we improve this using some batching?
676 * with batch RPC that'll happen automatically?
677 * XXX: can we store ctxt in lod_device and save few cycles ?
679 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
682 llh = ctxt->loc_handle;
685 CFS_INIT_LIST_HEAD(&list);
686 spin_lock(&d->opd_syn_lock);
687 cfs_list_splice(&d->opd_syn_committed_there, &list);
688 CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
689 spin_unlock(&d->opd_syn_lock);
691 cfs_list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
692 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
693 cfs_list_del_init(&req->rq_exp_list);
695 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
698 /* import can be closing, thus all commit cb's are
699 * called we can check committness directly */
700 if (req->rq_transno <= imp->imp_peer_committed_transno) {
701 rc = llog_cat_cancel_records(env, llh, 1,
702 &body->oa.o_lcookie);
704 CERROR("%s: can't cancel record: %d\n",
707 DEBUG_REQ(D_HA, req, "not committed");
710 ptlrpc_req_finished(req);
716 LASSERT(d->opd_syn_rpc_in_progress >= done);
717 spin_lock(&d->opd_syn_lock);
718 d->opd_syn_rpc_in_progress -= done;
719 spin_unlock(&d->opd_syn_lock);
720 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
721 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
722 d->opd_syn_rpc_in_progress);
724 osp_sync_check_for_work(d);
726 /* wake up the thread if requested to stop:
727 * it might be waiting for in-progress to complete */
728 if (unlikely(osp_sync_running(d) == 0))
729 wake_up(&d->opd_syn_waitq);
735 * this is where most of queues processing happens
737 static int osp_sync_process_queues(const struct lu_env *env,
738 struct llog_handle *llh,
739 struct llog_rec_hdr *rec,
742 struct osp_device *d = data;
746 struct l_wait_info lwi = { 0 };
748 if (!osp_sync_running(d)) {
749 CDEBUG(D_HA, "stop llog processing\n");
750 return LLOG_PROC_BREAK;
753 /* process requests committed by OST */
754 osp_sync_process_committed(env, d);
756 /* if we there are changes to be processed and we have
757 * resources for this ... do now */
758 if (osp_sync_can_process_new(d, rec)) {
760 /* ask llog for another record */
761 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
763 d->opd_syn_rpc_in_progress,
764 d->opd_syn_rpc_in_flight);
769 * try to send, in case of disconnection, suspend
770 * processing till we can send this request
773 rc = osp_sync_process_record(env, d, llh, rec);
775 * XXX: probably different handling is needed
776 * for some bugs, like immediate exit or if
780 CERROR("can't send: %d\n", rc);
781 l_wait_event(d->opd_syn_waitq,
782 !osp_sync_running(d) ||
783 osp_sync_has_work(d),
786 } while (rc != 0 && osp_sync_running(d));
792 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
793 osp_sync_remove_from_tracker(d);
795 l_wait_event(d->opd_syn_waitq,
796 !osp_sync_running(d) ||
797 osp_sync_can_process_new(d, rec) ||
798 !cfs_list_empty(&d->opd_syn_committed_there),
804 * this thread runs llog_cat_process() scanner calling our callback
805 * to process llog records. in the callback we implement tricky
806 * state machine as we don't want to start scanning of the llog again
807 * and again, also we don't want to process too many records and send
808 * too many RPCs a time. so, depending on current load (num of changes
809 * being synced to OST) the callback can suspend awaiting for some
810 * new conditions, like syncs completed.
812 * in order to process llog records left by previous boots and to allow
813 * llog_process_thread() to find something (otherwise it'd just exit
814 * immediately) we add a special GENERATATION record on each boot.
816 static int osp_sync_thread(void *_arg)
818 struct osp_device *d = _arg;
819 struct ptlrpc_thread *thread = &d->opd_syn_thread;
820 struct l_wait_info lwi = { 0 };
821 struct llog_ctxt *ctxt;
822 struct obd_device *obd = d->opd_obd;
823 struct llog_handle *llh;
829 rc = lu_env_init(&env, LCT_LOCAL);
831 CERROR("%s: can't initialize env: rc = %d\n",
836 spin_lock(&d->opd_syn_lock);
837 thread->t_flags = SVC_RUNNING;
838 spin_unlock(&d->opd_syn_lock);
839 wake_up(&thread->t_ctl_waitq);
841 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
843 CERROR("can't get appropriate context\n");
844 GOTO(out, rc = -EINVAL);
847 llh = ctxt->loc_handle;
849 CERROR("can't get llh\n");
851 GOTO(out, rc = -EINVAL);
854 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
855 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
856 "%lu changes, %u in progress, %u in flight: %d\n",
857 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
858 d->opd_syn_rpc_in_flight, rc);
860 /* we don't expect llog_process_thread() to exit till umount */
861 LASSERTF(thread->t_flags != SVC_RUNNING,
862 "%lu changes, %u in progress, %u in flight\n",
863 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
864 d->opd_syn_rpc_in_flight);
866 /* wait till all the requests are completed */
868 while (d->opd_syn_rpc_in_progress > 0) {
869 osp_sync_process_committed(&env, d);
871 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
872 rc = l_wait_event(d->opd_syn_waitq,
873 d->opd_syn_rpc_in_progress == 0,
875 if (rc == -ETIMEDOUT)
877 LASSERTF(count < 10, "%s: %d %d %sempty\n",
878 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
879 d->opd_syn_rpc_in_flight,
880 cfs_list_empty(&d->opd_syn_committed_there) ? "" :"!");
884 llog_cat_close(&env, llh);
885 rc = llog_cleanup(&env, ctxt);
887 CERROR("can't cleanup llog: %d\n", rc);
889 LASSERTF(d->opd_syn_rpc_in_progress == 0,
890 "%s: %d %d %sempty\n",
891 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
892 d->opd_syn_rpc_in_flight,
893 cfs_list_empty(&d->opd_syn_committed_there) ? "" : "!");
895 thread->t_flags = SVC_STOPPED;
897 wake_up(&thread->t_ctl_waitq);
904 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
906 struct osp_thread_info *osi = osp_env_info(env);
907 struct llog_handle *lgh = NULL;
908 struct obd_device *obd = d->opd_obd;
909 struct llog_ctxt *ctxt;
917 * open llog corresponding to our OST
919 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
920 obd->obd_lvfs_ctxt.dt = d->opd_storage;
922 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
925 CERROR("%s: can't get id from catalogs: rc = %d\n",
930 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
931 obd->obd_name, d->opd_index,
932 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
933 osi->osi_cid.lci_logid.lgl_ogen);
935 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
936 &osp_mds_ost_orig_logops);
940 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
943 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
944 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
946 /* re-create llog if it is missing */
948 logid_set_id(&osi->osi_cid.lci_logid, 0);
950 GOTO(out_cleanup, rc);
953 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
954 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
956 GOTO(out_cleanup, rc);
957 osi->osi_cid.lci_logid = lgh->lgh_id;
960 LASSERT(lgh != NULL);
961 ctxt->loc_handle = lgh;
963 rc = llog_cat_init_and_process(env, lgh);
967 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
973 * put a mark in the llog till which we'll be processing
974 * old records restless
976 d->opd_syn_generation.mnt_cnt = cfs_time_current();
977 d->opd_syn_generation.conn_cnt = cfs_time_current();
979 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
980 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
982 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
983 sizeof(osi->osi_gen.lgr_gen));
985 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie,
992 llog_cat_close(env, lgh);
994 llog_cleanup(env, ctxt);
998 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1000 struct llog_ctxt *ctxt;
1002 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1004 llog_cat_close(env, ctxt->loc_handle);
1005 llog_cleanup(env, ctxt);
1009 * initializes sync component of OSP
1011 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1013 struct l_wait_info lwi = { 0 };
1018 rc = osp_sync_id_traction_init(d);
1023 * initialize llog storing changes
1025 rc = osp_sync_llog_init(env, d);
1027 CERROR("%s: can't initialize llog: rc = %d\n",
1028 d->opd_obd->obd_name, rc);
1033 * Start synchronization thread
1035 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1036 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1037 spin_lock_init(&d->opd_syn_lock);
1038 init_waitqueue_head(&d->opd_syn_waitq);
1039 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1040 CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
1042 rc = PTR_ERR(kthread_run(osp_sync_thread, d,
1043 "osp-syn-%u", d->opd_index));
1044 if (IS_ERR_VALUE(rc)) {
1045 CERROR("%s: can't start sync thread: rc = %d\n",
1046 d->opd_obd->obd_name, rc);
1050 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1051 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1055 osp_sync_llog_fini(env, d);
1057 osp_sync_id_traction_fini(d);
1061 int osp_sync_fini(struct osp_device *d)
1063 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1067 thread->t_flags = SVC_STOPPING;
1068 wake_up(&d->opd_syn_waitq);
1069 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1072 * unregister transaction callbacks only when sync thread
1073 * has finished operations with llog
1075 osp_sync_id_traction_fini(d);
1080 static DEFINE_MUTEX(osp_id_tracker_sem);
1081 static CFS_LIST_HEAD(osp_id_tracker_list);
1083 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1085 struct osp_id_tracker *tr = cookie;
1086 struct osp_device *d;
1087 struct osp_txn_info *txn;
1091 txn = osp_txn_info(&th->th_ctx);
1092 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1095 spin_lock(&tr->otr_lock);
1096 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1097 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1098 tr->otr_committed_id, txn->oti_current_id);
1099 tr->otr_committed_id = txn->oti_current_id;
1101 cfs_list_for_each_entry(d, &tr->otr_wakeup_list,
1103 d->opd_syn_last_committed_id = tr->otr_committed_id;
1104 wake_up(&d->opd_syn_waitq);
1107 spin_unlock(&tr->otr_lock);
1110 static int osp_sync_id_traction_init(struct osp_device *d)
1112 struct osp_id_tracker *tr, *found = NULL;
1116 LASSERT(d->opd_storage);
1117 LASSERT(d->opd_syn_tracker == NULL);
1118 CFS_INIT_LIST_HEAD(&d->opd_syn_ontrack);
1120 mutex_lock(&osp_id_tracker_sem);
1121 cfs_list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1122 if (tr->otr_dev == d->opd_storage) {
1123 LASSERT(cfs_atomic_read(&tr->otr_refcount));
1124 cfs_atomic_inc(&tr->otr_refcount);
1125 d->opd_syn_tracker = tr;
1131 if (found == NULL) {
1135 d->opd_syn_tracker = tr;
1136 spin_lock_init(&tr->otr_lock);
1137 tr->otr_dev = d->opd_storage;
1138 tr->otr_next_id = 1;
1139 tr->otr_committed_id = 0;
1140 cfs_atomic_set(&tr->otr_refcount, 1);
1141 CFS_INIT_LIST_HEAD(&tr->otr_wakeup_list);
1142 cfs_list_add(&tr->otr_list, &osp_id_tracker_list);
1143 tr->otr_tx_cb.dtc_txn_commit =
1144 osp_sync_tracker_commit_cb;
1145 tr->otr_tx_cb.dtc_cookie = tr;
1146 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1147 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1151 mutex_unlock(&osp_id_tracker_sem);
1156 static void osp_sync_id_traction_fini(struct osp_device *d)
1158 struct osp_id_tracker *tr;
1163 tr = d->opd_syn_tracker;
1169 osp_sync_remove_from_tracker(d);
1171 mutex_lock(&osp_id_tracker_sem);
1172 if (cfs_atomic_dec_and_test(&tr->otr_refcount)) {
1173 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1174 LASSERT(cfs_list_empty(&tr->otr_wakeup_list));
1175 cfs_list_del(&tr->otr_list);
1177 d->opd_syn_tracker = NULL;
1179 mutex_unlock(&osp_id_tracker_sem);
1185 * generates id for the tracker
1187 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1189 struct osp_id_tracker *tr;
1191 tr = d->opd_syn_tracker;
1194 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1195 spin_lock(&tr->otr_lock);
1196 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1197 spin_unlock(&tr->otr_lock);
1198 CERROR("%s: next %u, last synced %lu\n",
1199 d->opd_obd->obd_name, tr->otr_next_id,
1200 d->opd_syn_last_used_id);
1205 id = tr->otr_next_id++;
1206 if (id > d->opd_syn_last_used_id)
1207 d->opd_syn_last_used_id = id;
1208 if (cfs_list_empty(&d->opd_syn_ontrack))
1209 cfs_list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1210 spin_unlock(&tr->otr_lock);
1211 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1216 static void osp_sync_remove_from_tracker(struct osp_device *d)
1218 struct osp_id_tracker *tr;
1220 tr = d->opd_syn_tracker;
1223 if (cfs_list_empty(&d->opd_syn_ontrack))
1226 spin_lock(&tr->otr_lock);
1227 cfs_list_del_init(&d->opd_syn_ontrack);
1228 spin_unlock(&tr->otr_lock);