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>
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_to_ostid(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_oi = osi->osi_oi;
239 osi->osi_setattr.lsr_uid = attr->la_uid;
240 osi->osi_setattr.lsr_gid = attr->la_gid;
246 txn = osp_txn_info(&th->th_ctx);
249 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
250 osi->osi_hdr.lrh_id = txn->oti_current_id;
252 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
255 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
259 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
260 d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
261 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
262 (unsigned long) osi->osi_cookie.lgc_index, rc);
267 if (likely(rc == 0)) {
268 spin_lock(&d->opd_syn_lock);
269 d->opd_syn_changes++;
270 spin_unlock(&d->opd_syn_lock);
276 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
277 llog_op_type type, struct thandle *th,
278 const struct lu_attr *attr)
280 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
281 lu_object_fid(&o->opo_obj.do_lu), type, 1,
285 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
286 struct lu_fid *fid, int lost, struct thandle *th)
288 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
292 * it's quite obvious we can't maintain all the structures in the memory:
293 * while OST is down, MDS can be processing thousands and thousands of unlinks
294 * filling persistent llogs and in-core respresentation
296 * this doesn't scale at all. so we need basically the following:
297 * a) destroy/setattr append llog records
298 * b) once llog has grown to X records, we process first Y committed records
300 * once record R is found via llog_process(), it becomes committed after any
301 * subsequent commit callback (at the most)
305 * called for each atomic on-disk change (not once per transaction batch)
306 * and goes over the list
307 * XXX: should be optimized?
311 * called for each RPC reported committed
313 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
315 struct osp_device *d = req->rq_cb_data;
317 CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
319 if (unlikely(req->rq_transno == 0))
322 /* do not do any opd_dyn_rpc_* accounting here
323 * it's done in osp_sync_interpret sooner or later */
326 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
327 LASSERT(cfs_list_empty(&req->rq_exp_list));
329 ptlrpc_request_addref(req);
331 spin_lock(&d->opd_syn_lock);
332 cfs_list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
333 spin_unlock(&d->opd_syn_lock);
335 /* XXX: some batching wouldn't hurt */
336 cfs_waitq_signal(&d->opd_syn_waitq);
339 static int osp_sync_interpret(const struct lu_env *env,
340 struct ptlrpc_request *req, void *aa, int rc)
342 struct osp_device *d = req->rq_cb_data;
344 if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
345 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
346 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
349 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
350 cfs_atomic_read(&req->rq_refcount),
351 rc, (unsigned) req->rq_transno);
352 LASSERT(rc || req->rq_transno);
356 * we tried to destroy object or update attributes,
357 * but object doesn't exist anymore - cancell llog record
359 LASSERT(req->rq_transno == 0);
360 LASSERT(cfs_list_empty(&req->rq_exp_list));
362 ptlrpc_request_addref(req);
364 spin_lock(&d->opd_syn_lock);
365 cfs_list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
366 spin_unlock(&d->opd_syn_lock);
368 cfs_waitq_signal(&d->opd_syn_waitq);
370 struct obd_import *imp = req->rq_import;
372 * error happened, we'll try to repeat on next boot ?
374 LASSERTF(req->rq_transno == 0 ||
375 req->rq_import_generation < imp->imp_generation,
376 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
377 req->rq_transno, rc, req->rq_import_generation,
378 imp->imp_generation);
379 if (req->rq_transno == 0) {
380 /* this is the last time we see the request
381 * if transno is not zero, then commit cb
382 * will be called at some point */
383 LASSERT(d->opd_syn_rpc_in_progress > 0);
384 spin_lock(&d->opd_syn_lock);
385 d->opd_syn_rpc_in_progress--;
386 spin_unlock(&d->opd_syn_lock);
389 cfs_waitq_signal(&d->opd_syn_waitq);
390 } else if (unlikely(d->opd_pre_status == -ENOSPC)) {
392 * if current status is -ENOSPC (lack of free space on OST)
393 * then we should poll OST immediately once object destroy
396 osp_statfs_need_now(d);
399 LASSERT(d->opd_syn_rpc_in_flight > 0);
400 spin_lock(&d->opd_syn_lock);
401 d->opd_syn_rpc_in_flight--;
402 spin_unlock(&d->opd_syn_lock);
403 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
404 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
405 d->opd_syn_rpc_in_progress);
407 osp_sync_check_for_work(d);
413 * the function walks through list of committed locally changes
414 * and send them to RPC until the pipe is full
416 static void osp_sync_send_new_rpc(struct osp_device *d,
417 struct ptlrpc_request *req)
419 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
420 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
422 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
425 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
426 struct llog_handle *llh,
427 struct llog_rec_hdr *h,
429 const struct req_format *format)
431 struct ptlrpc_request *req;
432 struct ost_body *body;
433 struct obd_import *imp;
436 /* Prepare the request */
437 imp = d->opd_obd->u.cli.cl_import;
439 req = ptlrpc_request_alloc(imp, format);
441 RETURN(ERR_PTR(-ENOMEM));
443 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
445 ptlrpc_req_finished(req);
450 * this is a trick: to save on memory allocations we put cookie
451 * into the request, but don't set corresponded flag in o_valid
452 * so that OST doesn't interpret this cookie. once the request
453 * is committed on OST we take cookie from the request and cancel
455 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
457 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
458 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
459 body->oa.o_lcookie.lgc_index = h->lrh_index;
460 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
461 req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
463 req->rq_interpret_reply = osp_sync_interpret;
464 req->rq_commit_cb = osp_sync_request_commit_cb;
467 ptlrpc_request_set_replen(req);
472 static int osp_sync_new_setattr_job(struct osp_device *d,
473 struct llog_handle *llh,
474 struct llog_rec_hdr *h)
476 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
477 struct ptlrpc_request *req;
478 struct ost_body *body;
481 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
483 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
485 RETURN(PTR_ERR(req));
487 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
489 body->oa.o_oi = rec->lsr_oi;
490 body->oa.o_uid = rec->lsr_uid;
491 body->oa.o_gid = rec->lsr_gid;
492 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
493 OBD_MD_FLUID | OBD_MD_FLGID;
495 osp_sync_send_new_rpc(d, req);
499 /* Old records may be in old format, so we handle that too */
500 static int osp_sync_new_unlink_job(struct osp_device *d,
501 struct llog_handle *llh,
502 struct llog_rec_hdr *h)
504 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
505 struct ptlrpc_request *req;
506 struct ost_body *body;
509 LASSERT(h->lrh_type == MDS_UNLINK_REC);
511 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
513 RETURN(PTR_ERR(req));
515 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
517 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
518 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
519 body->oa.o_misc = rec->lur_count;
520 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
522 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
524 osp_sync_send_new_rpc(d, req);
528 static int osp_sync_new_unlink64_job(struct osp_device *d,
529 struct llog_handle *llh,
530 struct llog_rec_hdr *h)
532 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
533 struct ptlrpc_request *req;
534 struct ost_body *body;
538 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
540 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
542 RETURN(PTR_ERR(req));
544 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
547 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
550 body->oa.o_misc = rec->lur_count;
551 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID | OBD_MD_FLOBJCOUNT;
553 osp_sync_send_new_rpc(d, req);
557 static int osp_sync_process_record(const struct lu_env *env,
558 struct osp_device *d,
559 struct llog_handle *llh,
560 struct llog_rec_hdr *rec)
562 struct llog_cookie cookie;
565 cookie.lgc_lgl = llh->lgh_id;
566 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
567 cookie.lgc_index = rec->lrh_index;
569 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
570 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
572 /* we're waiting for the record generated by this instance */
573 LASSERT(d->opd_syn_prev_done == 0);
574 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
575 sizeof(gen->lgr_gen))) {
576 CDEBUG(D_HA, "processed all old entries\n");
577 d->opd_syn_prev_done = 1;
580 /* cancel any generation record */
581 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
588 * now we prepare and fill requests to OST, put them on the queue
589 * and fire after next commit callback
592 /* notice we increment counters before sending RPC, to be consistent
593 * in RPC interpret callback which may happen very quickly */
594 spin_lock(&d->opd_syn_lock);
595 d->opd_syn_rpc_in_flight++;
596 d->opd_syn_rpc_in_progress++;
597 spin_unlock(&d->opd_syn_lock);
599 switch (rec->lrh_type) {
600 /* case MDS_UNLINK_REC is kept for compatibility */
602 rc = osp_sync_new_unlink_job(d, llh, rec);
604 case MDS_UNLINK64_REC:
605 rc = osp_sync_new_unlink64_job(d, llh, rec);
607 case MDS_SETATTR64_REC:
608 rc = osp_sync_new_setattr_job(d, llh, rec);
611 CERROR("unknown record type: %x\n", rec->lrh_type);
616 if (likely(rc == 0)) {
617 spin_lock(&d->opd_syn_lock);
618 if (d->opd_syn_prev_done) {
619 LASSERT(d->opd_syn_changes > 0);
620 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
622 * NOTE: it's possible to meet same id if
623 * OST stores few stripes of same file
625 if (rec->lrh_id > d->opd_syn_last_processed_id)
626 d->opd_syn_last_processed_id = rec->lrh_id;
628 d->opd_syn_changes--;
630 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
631 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
632 d->opd_syn_rpc_in_progress);
633 spin_unlock(&d->opd_syn_lock);
635 spin_lock(&d->opd_syn_lock);
636 d->opd_syn_rpc_in_flight--;
637 d->opd_syn_rpc_in_progress--;
638 spin_unlock(&d->opd_syn_lock);
641 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
642 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
646 static void osp_sync_process_committed(const struct lu_env *env,
647 struct osp_device *d)
649 struct obd_device *obd = d->opd_obd;
650 struct obd_import *imp = obd->u.cli.cl_import;
651 struct ost_body *body;
652 struct ptlrpc_request *req, *tmp;
653 struct llog_ctxt *ctxt;
654 struct llog_handle *llh;
660 if (cfs_list_empty(&d->opd_syn_committed_there))
664 * if current status is -ENOSPC (lack of free space on OST)
665 * then we should poll OST immediately once object destroy
667 * notice: we do this upon commit as well because some backends
668 * (like DMU) do not release space right away.
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 cfs_waitq_signal(&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 cfs_waitq_signal(&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 cfs_waitq_signal(&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);
1003 llog_cat_close(env, ctxt->loc_handle);
1004 llog_cleanup(env, ctxt);
1008 * initializes sync component of OSP
1010 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1012 struct l_wait_info lwi = { 0 };
1017 rc = osp_sync_id_traction_init(d);
1022 * initialize llog storing changes
1024 rc = osp_sync_llog_init(env, d);
1026 CERROR("%s: can't initialize llog: rc = %d\n",
1027 d->opd_obd->obd_name, rc);
1032 * Start synchronization thread
1034 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1035 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1036 spin_lock_init(&d->opd_syn_lock);
1037 cfs_waitq_init(&d->opd_syn_waitq);
1038 cfs_waitq_init(&d->opd_syn_thread.t_ctl_waitq);
1039 CFS_INIT_LIST_HEAD(&d->opd_syn_committed_there);
1041 rc = PTR_ERR(kthread_run(osp_sync_thread, d,
1042 "osp-syn-%u", d->opd_index));
1043 if (IS_ERR_VALUE(rc)) {
1044 CERROR("%s: can't start sync thread: rc = %d\n",
1045 d->opd_obd->obd_name, rc);
1049 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1050 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1054 osp_sync_llog_fini(env, d);
1056 osp_sync_id_traction_fini(d);
1060 int osp_sync_fini(struct osp_device *d)
1062 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1066 thread->t_flags = SVC_STOPPING;
1067 cfs_waitq_signal(&d->opd_syn_waitq);
1068 cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1071 * unregister transaction callbacks only when sync thread
1072 * has finished operations with llog
1074 osp_sync_id_traction_fini(d);
1079 static DEFINE_MUTEX(osp_id_tracker_sem);
1080 static CFS_LIST_HEAD(osp_id_tracker_list);
1082 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1084 struct osp_id_tracker *tr = cookie;
1085 struct osp_device *d;
1086 struct osp_txn_info *txn;
1090 txn = osp_txn_info(&th->th_ctx);
1091 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1094 spin_lock(&tr->otr_lock);
1095 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1096 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1097 tr->otr_committed_id, txn->oti_current_id);
1098 tr->otr_committed_id = txn->oti_current_id;
1100 cfs_list_for_each_entry(d, &tr->otr_wakeup_list,
1102 d->opd_syn_last_committed_id = tr->otr_committed_id;
1103 cfs_waitq_signal(&d->opd_syn_waitq);
1106 spin_unlock(&tr->otr_lock);
1109 static int osp_sync_id_traction_init(struct osp_device *d)
1111 struct osp_id_tracker *tr, *found = NULL;
1115 LASSERT(d->opd_storage);
1116 LASSERT(d->opd_syn_tracker == NULL);
1117 CFS_INIT_LIST_HEAD(&d->opd_syn_ontrack);
1119 mutex_lock(&osp_id_tracker_sem);
1120 cfs_list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1121 if (tr->otr_dev == d->opd_storage) {
1122 LASSERT(cfs_atomic_read(&tr->otr_refcount));
1123 cfs_atomic_inc(&tr->otr_refcount);
1124 d->opd_syn_tracker = tr;
1130 if (found == NULL) {
1134 d->opd_syn_tracker = tr;
1135 spin_lock_init(&tr->otr_lock);
1136 tr->otr_dev = d->opd_storage;
1137 tr->otr_next_id = 1;
1138 tr->otr_committed_id = 0;
1139 cfs_atomic_set(&tr->otr_refcount, 1);
1140 CFS_INIT_LIST_HEAD(&tr->otr_wakeup_list);
1141 cfs_list_add(&tr->otr_list, &osp_id_tracker_list);
1142 tr->otr_tx_cb.dtc_txn_commit =
1143 osp_sync_tracker_commit_cb;
1144 tr->otr_tx_cb.dtc_cookie = tr;
1145 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1146 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1150 mutex_unlock(&osp_id_tracker_sem);
1155 static void osp_sync_id_traction_fini(struct osp_device *d)
1157 struct osp_id_tracker *tr;
1162 tr = d->opd_syn_tracker;
1168 osp_sync_remove_from_tracker(d);
1170 mutex_lock(&osp_id_tracker_sem);
1171 if (cfs_atomic_dec_and_test(&tr->otr_refcount)) {
1172 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1173 LASSERT(cfs_list_empty(&tr->otr_wakeup_list));
1174 cfs_list_del(&tr->otr_list);
1176 d->opd_syn_tracker = NULL;
1178 mutex_unlock(&osp_id_tracker_sem);
1184 * generates id for the tracker
1186 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1188 struct osp_id_tracker *tr;
1190 tr = d->opd_syn_tracker;
1193 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1194 spin_lock(&tr->otr_lock);
1195 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1196 spin_unlock(&tr->otr_lock);
1197 CERROR("%s: next %u, last synced %lu\n",
1198 d->opd_obd->obd_name, tr->otr_next_id,
1199 d->opd_syn_last_used_id);
1204 id = tr->otr_next_id++;
1205 if (id > d->opd_syn_last_used_id)
1206 d->opd_syn_last_used_id = id;
1207 if (cfs_list_empty(&d->opd_syn_ontrack))
1208 cfs_list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1209 spin_unlock(&tr->otr_lock);
1210 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1215 static void osp_sync_remove_from_tracker(struct osp_device *d)
1217 struct osp_id_tracker *tr;
1219 tr = d->opd_syn_tracker;
1222 if (cfs_list_empty(&d->opd_syn_ontrack))
1225 spin_lock(&tr->otr_lock);
1226 cfs_list_del_init(&d->opd_syn_ontrack);
1227 spin_unlock(&tr->otr_lock);