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 (!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;
238 osi->osi_setattr.lsr_valid = attr->la_valid;
244 txn = osp_txn_info(&th->th_ctx);
247 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
248 osi->osi_hdr.lrh_id = txn->oti_current_id;
250 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
253 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
257 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
258 d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
259 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
260 (unsigned long) osi->osi_cookie.lgc_index, rc);
265 if (likely(rc == 0)) {
266 spin_lock(&d->opd_syn_lock);
267 d->opd_syn_changes++;
268 spin_unlock(&d->opd_syn_lock);
274 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
275 llog_op_type type, struct thandle *th,
276 const struct lu_attr *attr)
278 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
279 lu_object_fid(&o->opo_obj.do_lu), type, 1,
283 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
284 struct lu_fid *fid, int lost, struct thandle *th)
286 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
290 * it's quite obvious we can't maintain all the structures in the memory:
291 * while OST is down, MDS can be processing thousands and thousands of unlinks
292 * filling persistent llogs and in-core respresentation
294 * this doesn't scale at all. so we need basically the following:
295 * a) destroy/setattr append llog records
296 * b) once llog has grown to X records, we process first Y committed records
298 * once record R is found via llog_process(), it becomes committed after any
299 * subsequent commit callback (at the most)
303 * called for each atomic on-disk change (not once per transaction batch)
304 * and goes over the list
305 * XXX: should be optimized?
309 * called for each RPC reported committed
311 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
313 struct osp_device *d = req->rq_cb_data;
315 CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
317 if (unlikely(req->rq_transno == 0))
320 /* do not do any opd_dyn_rpc_* accounting here
321 * it's done in osp_sync_interpret sooner or later */
324 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
325 LASSERT(list_empty(&req->rq_exp_list));
327 ptlrpc_request_addref(req);
329 spin_lock(&d->opd_syn_lock);
330 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
331 spin_unlock(&d->opd_syn_lock);
333 /* XXX: some batching wouldn't hurt */
334 wake_up(&d->opd_syn_waitq);
337 static int osp_sync_interpret(const struct lu_env *env,
338 struct ptlrpc_request *req, void *aa, int rc)
340 struct osp_device *d = req->rq_cb_data;
342 if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
343 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
344 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
347 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
348 atomic_read(&req->rq_refcount),
349 rc, (unsigned) req->rq_transno);
350 LASSERT(rc || req->rq_transno);
354 * we tried to destroy object or update attributes,
355 * but object doesn't exist anymore - cancell llog record
357 LASSERT(req->rq_transno == 0);
358 LASSERT(list_empty(&req->rq_exp_list));
360 ptlrpc_request_addref(req);
362 spin_lock(&d->opd_syn_lock);
363 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
364 spin_unlock(&d->opd_syn_lock);
366 wake_up(&d->opd_syn_waitq);
368 struct obd_import *imp = req->rq_import;
370 * error happened, we'll try to repeat on next boot ?
372 LASSERTF(req->rq_transno == 0 ||
373 req->rq_import_generation < imp->imp_generation,
374 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
375 req->rq_transno, rc, req->rq_import_generation,
376 imp->imp_generation);
377 if (req->rq_transno == 0) {
378 /* this is the last time we see the request
379 * if transno is not zero, then commit cb
380 * will be called at some point */
381 LASSERT(d->opd_syn_rpc_in_progress > 0);
382 spin_lock(&d->opd_syn_lock);
383 d->opd_syn_rpc_in_progress--;
384 spin_unlock(&d->opd_syn_lock);
387 wake_up(&d->opd_syn_waitq);
388 } else if (d->opd_pre != NULL &&
389 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 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 /* lsr_valid can only be 0 or LA_UID/GID set */
483 if (!rec->lsr_valid && !(rec->lsr_valid & ~(LA_UID | LA_GID))) {
484 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
485 d->opd_obd->obd_name, rec->lsr_valid);
489 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
491 RETURN(PTR_ERR(req));
493 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
495 body->oa.o_oi = rec->lsr_oi;
496 body->oa.o_uid = rec->lsr_uid;
497 body->oa.o_gid = rec->lsr_gid;
498 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
499 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
500 * we assume that both UID and GID are valid in that case. */
501 if (rec->lsr_valid == 0) {
502 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
504 if (rec->lsr_valid & LA_UID)
505 body->oa.o_valid |= OBD_MD_FLUID;
506 if (rec->lsr_valid & LA_GID)
507 body->oa.o_valid |= OBD_MD_FLGID;
510 osp_sync_send_new_rpc(d, req);
514 /* Old records may be in old format, so we handle that too */
515 static int osp_sync_new_unlink_job(struct osp_device *d,
516 struct llog_handle *llh,
517 struct llog_rec_hdr *h)
519 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
520 struct ptlrpc_request *req;
521 struct ost_body *body;
524 LASSERT(h->lrh_type == MDS_UNLINK_REC);
526 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
528 RETURN(PTR_ERR(req));
530 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
532 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
533 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
534 body->oa.o_misc = rec->lur_count;
535 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
537 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
539 osp_sync_send_new_rpc(d, req);
543 static int osp_prep_unlink_update_req(const struct lu_env *env,
544 struct osp_device *osp,
545 struct llog_handle *llh,
546 struct llog_rec_hdr *h,
547 struct ptlrpc_request **reqp)
549 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
550 struct dt_update_request *update = NULL;
551 struct ptlrpc_request *req;
553 struct llog_cookie lcookie;
558 update = out_create_update_req(&osp->opd_dt_dev);
560 RETURN(PTR_ERR(update));
562 /* This can only happens for unlink slave directory, so decrease
563 * ref for ".." and "." */
564 rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
569 rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
574 lcookie.lgc_lgl = llh->lgh_id;
575 lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
576 lcookie.lgc_index = h->lrh_index;
577 size = sizeof(lcookie);
578 buf = (const char *)&lcookie;
580 rc = out_insert_update(env, update, OUT_DESTROY, &rec->lur_fid, 1,
585 rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
586 update->dur_req, &req);
590 INIT_LIST_HEAD(&req->rq_exp_list);
591 req->rq_svc_thread = (void *)OSP_JOB_MAGIC;
593 req->rq_interpret_reply = osp_sync_interpret;
594 req->rq_commit_cb = osp_sync_request_commit_cb;
595 req->rq_cb_data = osp;
597 ptlrpc_request_set_replen(req);
601 out_destroy_update_req(update);
606 static int osp_sync_new_unlink64_job(const struct lu_env *env,
607 struct osp_device *d,
608 struct llog_handle *llh,
609 struct llog_rec_hdr *h)
611 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
612 struct ptlrpc_request *req = NULL;
613 struct ost_body *body;
617 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
619 if (d->opd_connect_mdt) {
620 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
624 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
627 RETURN(PTR_ERR(req));
629 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
632 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
635 body->oa.o_misc = rec->lur_count;
636 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
639 osp_sync_send_new_rpc(d, req);
643 static int osp_sync_process_record(const struct lu_env *env,
644 struct osp_device *d,
645 struct llog_handle *llh,
646 struct llog_rec_hdr *rec)
648 struct llog_cookie cookie;
651 cookie.lgc_lgl = llh->lgh_id;
652 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
653 cookie.lgc_index = rec->lrh_index;
655 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
656 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
658 /* we're waiting for the record generated by this instance */
659 LASSERT(d->opd_syn_prev_done == 0);
660 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
661 sizeof(gen->lgr_gen))) {
662 CDEBUG(D_HA, "processed all old entries\n");
663 d->opd_syn_prev_done = 1;
666 /* cancel any generation record */
667 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
674 * now we prepare and fill requests to OST, put them on the queue
675 * and fire after next commit callback
678 /* notice we increment counters before sending RPC, to be consistent
679 * in RPC interpret callback which may happen very quickly */
680 spin_lock(&d->opd_syn_lock);
681 d->opd_syn_rpc_in_flight++;
682 d->opd_syn_rpc_in_progress++;
683 spin_unlock(&d->opd_syn_lock);
685 switch (rec->lrh_type) {
686 /* case MDS_UNLINK_REC is kept for compatibility */
688 rc = osp_sync_new_unlink_job(d, llh, rec);
690 case MDS_UNLINK64_REC:
691 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
693 case MDS_SETATTR64_REC:
694 rc = osp_sync_new_setattr_job(d, llh, rec);
697 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
699 /* we should continue processing */
703 if (likely(rc == 0)) {
704 spin_lock(&d->opd_syn_lock);
705 if (d->opd_syn_prev_done) {
706 LASSERT(d->opd_syn_changes > 0);
707 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
709 * NOTE: it's possible to meet same id if
710 * OST stores few stripes of same file
712 if (rec->lrh_id > d->opd_syn_last_processed_id)
713 d->opd_syn_last_processed_id = rec->lrh_id;
715 d->opd_syn_changes--;
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);
720 spin_unlock(&d->opd_syn_lock);
722 spin_lock(&d->opd_syn_lock);
723 d->opd_syn_rpc_in_flight--;
724 d->opd_syn_rpc_in_progress--;
725 spin_unlock(&d->opd_syn_lock);
728 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
729 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
733 static void osp_sync_process_committed(const struct lu_env *env,
734 struct osp_device *d)
736 struct obd_device *obd = d->opd_obd;
737 struct obd_import *imp = obd->u.cli.cl_import;
738 struct ost_body *body;
739 struct ptlrpc_request *req, *tmp;
740 struct llog_ctxt *ctxt;
741 struct llog_handle *llh;
742 struct list_head list;
747 if (list_empty(&d->opd_syn_committed_there))
751 * if current status is -ENOSPC (lack of free space on OST)
752 * then we should poll OST immediately once object destroy
754 * notice: we do this upon commit as well because some backends
755 * (like DMU) do not release space right away.
757 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
758 osp_statfs_need_now(d);
761 * now cancel them all
762 * XXX: can we improve this using some batching?
763 * with batch RPC that'll happen automatically?
764 * XXX: can we store ctxt in lod_device and save few cycles ?
766 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
769 llh = ctxt->loc_handle;
772 INIT_LIST_HEAD(&list);
773 spin_lock(&d->opd_syn_lock);
774 list_splice(&d->opd_syn_committed_there, &list);
775 INIT_LIST_HEAD(&d->opd_syn_committed_there);
776 spin_unlock(&d->opd_syn_lock);
778 list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
779 struct llog_cookie *lcookie = NULL;
781 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
782 list_del_init(&req->rq_exp_list);
784 if (d->opd_connect_mdt) {
785 struct object_update_request *ureq;
786 struct object_update *update;
787 ureq = req_capsule_client_get(&req->rq_pill,
789 LASSERT(ureq != NULL &&
790 ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
792 /* 1st/2nd is for decref . and .., 3rd one is for
793 * destroy, where the log cookie is stored.
794 * See osp_prep_unlink_update_req */
795 update = object_update_request_get(ureq, 2, NULL);
796 LASSERT(update != NULL);
797 lcookie = object_update_param_get(update, 0, NULL);
798 LASSERT(lcookie != NULL);
800 body = req_capsule_client_get(&req->rq_pill,
803 lcookie = &body->oa.o_lcookie;
805 /* import can be closing, thus all commit cb's are
806 * called we can check committness directly */
807 if (req->rq_transno <= imp->imp_peer_committed_transno) {
808 rc = llog_cat_cancel_records(env, llh, 1, lcookie);
810 CERROR("%s: can't cancel record: %d\n",
813 DEBUG_REQ(D_HA, req, "not committed");
816 ptlrpc_req_finished(req);
822 LASSERT(d->opd_syn_rpc_in_progress >= done);
823 spin_lock(&d->opd_syn_lock);
824 d->opd_syn_rpc_in_progress -= done;
825 spin_unlock(&d->opd_syn_lock);
826 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
827 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
828 d->opd_syn_rpc_in_progress);
830 osp_sync_check_for_work(d);
832 /* wake up the thread if requested to stop:
833 * it might be waiting for in-progress to complete */
834 if (unlikely(osp_sync_running(d) == 0))
835 wake_up(&d->opd_syn_waitq);
841 * this is where most of queues processing happens
843 static int osp_sync_process_queues(const struct lu_env *env,
844 struct llog_handle *llh,
845 struct llog_rec_hdr *rec,
848 struct osp_device *d = data;
852 struct l_wait_info lwi = { 0 };
854 if (!osp_sync_running(d)) {
855 CDEBUG(D_HA, "stop llog processing\n");
856 return LLOG_PROC_BREAK;
859 /* process requests committed by OST */
860 osp_sync_process_committed(env, d);
862 /* if we there are changes to be processed and we have
863 * resources for this ... do now */
864 if (osp_sync_can_process_new(d, rec)) {
866 /* ask llog for another record */
867 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
869 d->opd_syn_rpc_in_progress,
870 d->opd_syn_rpc_in_flight);
875 * try to send, in case of disconnection, suspend
876 * processing till we can send this request
879 rc = osp_sync_process_record(env, d, llh, rec);
881 * XXX: probably different handling is needed
882 * for some bugs, like immediate exit or if
886 CERROR("can't send: %d\n", rc);
887 l_wait_event(d->opd_syn_waitq,
888 !osp_sync_running(d) ||
889 osp_sync_has_work(d),
892 } while (rc != 0 && osp_sync_running(d));
898 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
899 osp_sync_remove_from_tracker(d);
901 l_wait_event(d->opd_syn_waitq,
902 !osp_sync_running(d) ||
903 osp_sync_can_process_new(d, rec) ||
904 !list_empty(&d->opd_syn_committed_there),
910 * this thread runs llog_cat_process() scanner calling our callback
911 * to process llog records. in the callback we implement tricky
912 * state machine as we don't want to start scanning of the llog again
913 * and again, also we don't want to process too many records and send
914 * too many RPCs a time. so, depending on current load (num of changes
915 * being synced to OST) the callback can suspend awaiting for some
916 * new conditions, like syncs completed.
918 * in order to process llog records left by previous boots and to allow
919 * llog_process_thread() to find something (otherwise it'd just exit
920 * immediately) we add a special GENERATATION record on each boot.
922 static int osp_sync_thread(void *_arg)
924 struct osp_device *d = _arg;
925 struct ptlrpc_thread *thread = &d->opd_syn_thread;
926 struct l_wait_info lwi = { 0 };
927 struct llog_ctxt *ctxt;
928 struct obd_device *obd = d->opd_obd;
929 struct llog_handle *llh;
935 rc = lu_env_init(&env, LCT_LOCAL);
937 CERROR("%s: can't initialize env: rc = %d\n",
942 spin_lock(&d->opd_syn_lock);
943 thread->t_flags = SVC_RUNNING;
944 spin_unlock(&d->opd_syn_lock);
945 wake_up(&thread->t_ctl_waitq);
947 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
949 CERROR("can't get appropriate context\n");
950 GOTO(out, rc = -EINVAL);
953 llh = ctxt->loc_handle;
955 CERROR("can't get llh\n");
957 GOTO(out, rc = -EINVAL);
960 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
961 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
962 "%lu changes, %u in progress, %u in flight: %d\n",
963 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
964 d->opd_syn_rpc_in_flight, rc);
966 /* we don't expect llog_process_thread() to exit till umount */
967 LASSERTF(thread->t_flags != SVC_RUNNING,
968 "%lu changes, %u in progress, %u in flight\n",
969 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
970 d->opd_syn_rpc_in_flight);
972 /* wait till all the requests are completed */
974 while (d->opd_syn_rpc_in_progress > 0) {
975 osp_sync_process_committed(&env, d);
977 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
978 rc = l_wait_event(d->opd_syn_waitq,
979 d->opd_syn_rpc_in_progress == 0,
981 if (rc == -ETIMEDOUT)
983 LASSERTF(count < 10, "%s: %d %d %sempty\n",
984 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
985 d->opd_syn_rpc_in_flight,
986 list_empty(&d->opd_syn_committed_there) ? "" : "!");
990 llog_cat_close(&env, llh);
991 rc = llog_cleanup(&env, ctxt);
993 CERROR("can't cleanup llog: %d\n", rc);
995 LASSERTF(d->opd_syn_rpc_in_progress == 0,
996 "%s: %d %d %sempty\n",
997 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
998 d->opd_syn_rpc_in_flight,
999 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1001 thread->t_flags = SVC_STOPPED;
1003 wake_up(&thread->t_ctl_waitq);
1010 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1012 struct osp_thread_info *osi = osp_env_info(env);
1013 struct lu_fid *fid = &osi->osi_fid;
1014 struct llog_handle *lgh = NULL;
1015 struct obd_device *obd = d->opd_obd;
1016 struct llog_ctxt *ctxt;
1024 * open llog corresponding to our OST
1026 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1027 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1029 if (d->opd_connect_mdt)
1030 lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
1032 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1034 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1035 &osi->osi_cid, fid);
1037 CERROR("%s: can't get id from catalogs: rc = %d\n",
1042 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1043 obd->obd_name, d->opd_index,
1044 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1045 osi->osi_cid.lci_logid.lgl_ogen);
1047 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1048 &osp_mds_ost_orig_logops);
1052 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1055 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1056 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1058 /* re-create llog if it is missing */
1060 logid_set_id(&osi->osi_cid.lci_logid, 0);
1062 GOTO(out_cleanup, rc);
1065 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1066 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1068 GOTO(out_cleanup, rc);
1069 osi->osi_cid.lci_logid = lgh->lgh_id;
1072 LASSERT(lgh != NULL);
1073 ctxt->loc_handle = lgh;
1075 rc = llog_cat_init_and_process(env, lgh);
1077 GOTO(out_close, rc);
1079 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1080 &osi->osi_cid, fid);
1082 GOTO(out_close, rc);
1085 * put a mark in the llog till which we'll be processing
1086 * old records restless
1088 d->opd_syn_generation.mnt_cnt = cfs_time_current();
1089 d->opd_syn_generation.conn_cnt = cfs_time_current();
1091 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1092 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1094 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1095 sizeof(osi->osi_gen.lgr_gen));
1097 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1099 GOTO(out_close, rc);
1100 llog_ctxt_put(ctxt);
1103 llog_cat_close(env, lgh);
1105 llog_cleanup(env, ctxt);
1109 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1111 struct llog_ctxt *ctxt;
1113 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1115 llog_cat_close(env, ctxt->loc_handle);
1116 llog_cleanup(env, ctxt);
1120 * initializes sync component of OSP
1122 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1124 struct l_wait_info lwi = { 0 };
1125 struct task_struct *task;
1130 rc = osp_sync_id_traction_init(d);
1135 * initialize llog storing changes
1137 rc = osp_sync_llog_init(env, d);
1139 CERROR("%s: can't initialize llog: rc = %d\n",
1140 d->opd_obd->obd_name, rc);
1145 * Start synchronization thread
1147 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1148 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1149 spin_lock_init(&d->opd_syn_lock);
1150 init_waitqueue_head(&d->opd_syn_waitq);
1151 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1152 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1154 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1155 d->opd_index, d->opd_group);
1158 CERROR("%s: cannot start sync thread: rc = %d\n",
1159 d->opd_obd->obd_name, rc);
1163 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1164 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1168 osp_sync_llog_fini(env, d);
1170 osp_sync_id_traction_fini(d);
1174 int osp_sync_fini(struct osp_device *d)
1176 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1180 thread->t_flags = SVC_STOPPING;
1181 wake_up(&d->opd_syn_waitq);
1182 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1185 * unregister transaction callbacks only when sync thread
1186 * has finished operations with llog
1188 osp_sync_id_traction_fini(d);
1193 static DEFINE_MUTEX(osp_id_tracker_sem);
1194 static struct list_head osp_id_tracker_list =
1195 LIST_HEAD_INIT(osp_id_tracker_list);
1197 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1199 struct osp_id_tracker *tr = cookie;
1200 struct osp_device *d;
1201 struct osp_txn_info *txn;
1205 txn = osp_txn_info(&th->th_ctx);
1206 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1209 spin_lock(&tr->otr_lock);
1210 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1211 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1212 tr->otr_committed_id, txn->oti_current_id);
1213 tr->otr_committed_id = txn->oti_current_id;
1215 list_for_each_entry(d, &tr->otr_wakeup_list,
1217 d->opd_syn_last_committed_id = tr->otr_committed_id;
1218 wake_up(&d->opd_syn_waitq);
1221 spin_unlock(&tr->otr_lock);
1224 static int osp_sync_id_traction_init(struct osp_device *d)
1226 struct osp_id_tracker *tr, *found = NULL;
1230 LASSERT(d->opd_storage);
1231 LASSERT(d->opd_syn_tracker == NULL);
1232 INIT_LIST_HEAD(&d->opd_syn_ontrack);
1234 mutex_lock(&osp_id_tracker_sem);
1235 list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1236 if (tr->otr_dev == d->opd_storage) {
1237 LASSERT(atomic_read(&tr->otr_refcount));
1238 atomic_inc(&tr->otr_refcount);
1239 d->opd_syn_tracker = tr;
1245 if (found == NULL) {
1249 d->opd_syn_tracker = tr;
1250 spin_lock_init(&tr->otr_lock);
1251 tr->otr_dev = d->opd_storage;
1252 tr->otr_next_id = 1;
1253 tr->otr_committed_id = 0;
1254 atomic_set(&tr->otr_refcount, 1);
1255 INIT_LIST_HEAD(&tr->otr_wakeup_list);
1256 list_add(&tr->otr_list, &osp_id_tracker_list);
1257 tr->otr_tx_cb.dtc_txn_commit =
1258 osp_sync_tracker_commit_cb;
1259 tr->otr_tx_cb.dtc_cookie = tr;
1260 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1261 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1265 mutex_unlock(&osp_id_tracker_sem);
1270 static void osp_sync_id_traction_fini(struct osp_device *d)
1272 struct osp_id_tracker *tr;
1277 tr = d->opd_syn_tracker;
1283 osp_sync_remove_from_tracker(d);
1285 mutex_lock(&osp_id_tracker_sem);
1286 if (atomic_dec_and_test(&tr->otr_refcount)) {
1287 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1288 LASSERT(list_empty(&tr->otr_wakeup_list));
1289 list_del(&tr->otr_list);
1291 d->opd_syn_tracker = NULL;
1293 mutex_unlock(&osp_id_tracker_sem);
1299 * generates id for the tracker
1301 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1303 struct osp_id_tracker *tr;
1305 tr = d->opd_syn_tracker;
1308 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1309 spin_lock(&tr->otr_lock);
1310 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1311 spin_unlock(&tr->otr_lock);
1312 CERROR("%s: next %u, last synced %lu\n",
1313 d->opd_obd->obd_name, tr->otr_next_id,
1314 d->opd_syn_last_used_id);
1319 id = tr->otr_next_id++;
1320 if (id > d->opd_syn_last_used_id)
1321 d->opd_syn_last_used_id = id;
1322 if (list_empty(&d->opd_syn_ontrack))
1323 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1324 spin_unlock(&tr->otr_lock);
1325 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1330 static void osp_sync_remove_from_tracker(struct osp_device *d)
1332 struct osp_id_tracker *tr;
1334 tr = d->opd_syn_tracker;
1337 if (list_empty(&d->opd_syn_ontrack))
1340 spin_lock(&tr->otr_lock);
1341 list_del_init(&d->opd_syn_ontrack);
1342 spin_unlock(&tr->otr_lock);