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 (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
156 if (!osp_sync_low_in_progress(d))
158 if (!osp_sync_low_in_flight(d))
160 if (!d->opd_imp_connected)
162 if (d->opd_syn_prev_done == 0)
164 if (d->opd_syn_changes == 0)
166 if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
171 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
172 llog_op_type type, struct thandle *th)
174 struct osp_thread_info *osi = osp_env_info(env);
175 struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
176 struct llog_ctxt *ctxt;
181 /* it's a layering violation, to access internals of th,
182 * but we can do this as a sanity check, for a while */
183 LASSERT(th->th_dev == d->opd_storage);
186 case MDS_UNLINK64_REC:
187 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
189 case MDS_SETATTR64_REC:
190 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
196 /* we want ->dt_trans_start() to allocate per-thandle structure */
197 th->th_tags |= LCT_OSP_THREAD;
199 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
202 rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
208 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
209 const struct lu_fid *fid, llog_op_type type,
210 int count, struct thandle *th,
211 const struct lu_attr *attr)
213 struct osp_thread_info *osi = osp_env_info(env);
214 struct llog_ctxt *ctxt;
215 struct osp_txn_info *txn;
220 /* it's a layering violation, to access internals of th,
221 * but we can do this as a sanity check, for a while */
222 LASSERT(th->th_dev == d->opd_storage);
225 case MDS_UNLINK64_REC:
226 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
227 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
228 osi->osi_unlink.lur_fid = *fid;
229 osi->osi_unlink.lur_count = count;
231 case MDS_SETATTR64_REC:
232 rc = fid_to_ostid(fid, &osi->osi_oi);
234 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
235 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
236 osi->osi_setattr.lsr_oi = osi->osi_oi;
238 osi->osi_setattr.lsr_uid = attr->la_uid;
239 osi->osi_setattr.lsr_gid = attr->la_gid;
240 osi->osi_setattr.lsr_valid = attr->la_valid;
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(list_empty(&req->rq_exp_list));
329 ptlrpc_request_addref(req);
331 spin_lock(&d->opd_syn_lock);
332 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 wake_up(&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 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(list_empty(&req->rq_exp_list));
362 ptlrpc_request_addref(req);
364 spin_lock(&d->opd_syn_lock);
365 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
366 spin_unlock(&d->opd_syn_lock);
368 wake_up(&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 wake_up(&d->opd_syn_waitq);
390 } else if (d->opd_pre != NULL &&
391 unlikely(d->opd_pre_status == -ENOSPC)) {
393 * if current status is -ENOSPC (lack of free space on OST)
394 * then we should poll OST immediately once object destroy
397 osp_statfs_need_now(d);
400 LASSERT(d->opd_syn_rpc_in_flight > 0);
401 spin_lock(&d->opd_syn_lock);
402 d->opd_syn_rpc_in_flight--;
403 spin_unlock(&d->opd_syn_lock);
404 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
405 wake_up(&d->opd_syn_barrier_waitq);
406 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
407 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
408 d->opd_syn_rpc_in_progress);
410 osp_sync_check_for_work(d);
416 * the function walks through list of committed locally changes
417 * and send them to RPC until the pipe is full
419 static void osp_sync_send_new_rpc(struct osp_device *d,
420 struct ptlrpc_request *req)
422 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
423 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
425 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
428 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
429 struct llog_handle *llh,
430 struct llog_rec_hdr *h,
432 const struct req_format *format)
434 struct ptlrpc_request *req;
435 struct ost_body *body;
436 struct obd_import *imp;
439 /* Prepare the request */
440 imp = d->opd_obd->u.cli.cl_import;
442 req = ptlrpc_request_alloc(imp, format);
444 RETURN(ERR_PTR(-ENOMEM));
446 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
448 ptlrpc_req_finished(req);
453 * this is a trick: to save on memory allocations we put cookie
454 * into the request, but don't set corresponded flag in o_valid
455 * so that OST doesn't interpret this cookie. once the request
456 * is committed on OST we take cookie from the request and cancel
458 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
460 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
461 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
462 body->oa.o_lcookie.lgc_index = h->lrh_index;
463 INIT_LIST_HEAD(&req->rq_exp_list);
464 req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
466 req->rq_interpret_reply = osp_sync_interpret;
467 req->rq_commit_cb = osp_sync_request_commit_cb;
470 ptlrpc_request_set_replen(req);
475 static int osp_sync_new_setattr_job(struct osp_device *d,
476 struct llog_handle *llh,
477 struct llog_rec_hdr *h)
479 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
480 struct ptlrpc_request *req;
481 struct ost_body *body;
484 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
486 /* lsr_valid can only be 0 or LA_UID/GID set */
487 if ((rec->lsr_valid & ~(LA_UID | LA_GID)) != 0) {
488 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
489 d->opd_obd->obd_name, rec->lsr_valid);
490 /* return 0 so that sync thread can continue processing
495 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
497 RETURN(PTR_ERR(req));
499 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
501 body->oa.o_oi = rec->lsr_oi;
502 body->oa.o_uid = rec->lsr_uid;
503 body->oa.o_gid = rec->lsr_gid;
504 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
505 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
506 * we assume that both UID and GID are valid in that case. */
507 if (rec->lsr_valid == 0) {
508 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
510 if (rec->lsr_valid & LA_UID)
511 body->oa.o_valid |= OBD_MD_FLUID;
512 if (rec->lsr_valid & LA_GID)
513 body->oa.o_valid |= OBD_MD_FLGID;
516 osp_sync_send_new_rpc(d, req);
520 static int osp_sync_new_unlink_job(struct osp_device *d,
521 struct llog_handle *llh,
522 struct llog_rec_hdr *h)
524 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
525 struct ptlrpc_request *req;
526 struct ost_body *body;
529 LASSERT(h->lrh_type == MDS_UNLINK_REC);
531 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
533 RETURN(PTR_ERR(req));
535 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
537 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
538 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
539 body->oa.o_misc = rec->lur_count;
540 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
542 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
544 osp_sync_send_new_rpc(d, req);
548 static int osp_prep_unlink_update_req(const struct lu_env *env,
549 struct osp_device *osp,
550 struct llog_handle *llh,
551 struct llog_rec_hdr *h,
552 struct ptlrpc_request **reqp)
554 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
555 struct dt_update_request *update = NULL;
556 struct ptlrpc_request *req;
558 struct llog_cookie lcookie;
563 update = out_create_update_req(&osp->opd_dt_dev);
565 RETURN(PTR_ERR(update));
567 /* This can only happens for unlink slave directory, so decrease
568 * ref for ".." and "." */
569 rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
574 rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
579 lcookie.lgc_lgl = llh->lgh_id;
580 lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
581 lcookie.lgc_index = h->lrh_index;
582 size = sizeof(lcookie);
583 buf = (const char *)&lcookie;
585 rc = out_insert_update(env, update, OUT_DESTROY, &rec->lur_fid, 1,
590 rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
591 update->dur_req, &req);
595 INIT_LIST_HEAD(&req->rq_exp_list);
596 req->rq_svc_thread = (void *)OSP_JOB_MAGIC;
598 req->rq_interpret_reply = osp_sync_interpret;
599 req->rq_commit_cb = osp_sync_request_commit_cb;
600 req->rq_cb_data = osp;
602 ptlrpc_request_set_replen(req);
606 out_destroy_update_req(update);
611 static int osp_sync_new_unlink64_job(const struct lu_env *env,
612 struct osp_device *d,
613 struct llog_handle *llh,
614 struct llog_rec_hdr *h)
616 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
617 struct ptlrpc_request *req = NULL;
618 struct ost_body *body;
622 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
624 if (d->opd_connect_mdt) {
625 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
629 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
632 RETURN(PTR_ERR(req));
634 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
637 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
640 body->oa.o_misc = rec->lur_count;
641 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
644 osp_sync_send_new_rpc(d, req);
648 static int osp_sync_process_record(const struct lu_env *env,
649 struct osp_device *d,
650 struct llog_handle *llh,
651 struct llog_rec_hdr *rec)
653 struct llog_cookie cookie;
656 cookie.lgc_lgl = llh->lgh_id;
657 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
658 cookie.lgc_index = rec->lrh_index;
660 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
661 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
663 /* we're waiting for the record generated by this instance */
664 LASSERT(d->opd_syn_prev_done == 0);
665 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
666 sizeof(gen->lgr_gen))) {
667 CDEBUG(D_HA, "processed all old entries\n");
668 d->opd_syn_prev_done = 1;
671 /* cancel any generation record */
672 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
679 * now we prepare and fill requests to OST, put them on the queue
680 * and fire after next commit callback
683 /* notice we increment counters before sending RPC, to be consistent
684 * in RPC interpret callback which may happen very quickly */
685 spin_lock(&d->opd_syn_lock);
686 d->opd_syn_rpc_in_flight++;
687 d->opd_syn_rpc_in_progress++;
688 spin_unlock(&d->opd_syn_lock);
690 switch (rec->lrh_type) {
691 /* case MDS_UNLINK_REC is kept for compatibility */
693 rc = osp_sync_new_unlink_job(d, llh, rec);
695 case MDS_UNLINK64_REC:
696 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
698 case MDS_SETATTR64_REC:
699 rc = osp_sync_new_setattr_job(d, llh, rec);
702 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
704 /* we should continue processing */
707 /* rc > 0 means sync RPC being added to the queue */
708 if (likely(rc > 0)) {
709 spin_lock(&d->opd_syn_lock);
710 if (d->opd_syn_prev_done) {
711 LASSERT(d->opd_syn_changes > 0);
712 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
714 * NOTE: it's possible to meet same id if
715 * OST stores few stripes of same file
717 if (rec->lrh_id > d->opd_syn_last_processed_id) {
718 d->opd_syn_last_processed_id = rec->lrh_id;
719 wake_up(&d->opd_syn_barrier_waitq);
722 d->opd_syn_changes--;
724 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
725 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
726 d->opd_syn_rpc_in_progress);
727 spin_unlock(&d->opd_syn_lock);
730 spin_lock(&d->opd_syn_lock);
731 d->opd_syn_rpc_in_flight--;
732 d->opd_syn_rpc_in_progress--;
733 spin_unlock(&d->opd_syn_lock);
736 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
737 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
741 static void osp_sync_process_committed(const struct lu_env *env,
742 struct osp_device *d)
744 struct obd_device *obd = d->opd_obd;
745 struct obd_import *imp = obd->u.cli.cl_import;
746 struct ost_body *body;
747 struct ptlrpc_request *req, *tmp;
748 struct llog_ctxt *ctxt;
749 struct llog_handle *llh;
750 struct list_head list;
755 if (list_empty(&d->opd_syn_committed_there))
759 * if current status is -ENOSPC (lack of free space on OST)
760 * then we should poll OST immediately once object destroy
762 * notice: we do this upon commit as well because some backends
763 * (like DMU) do not release space right away.
765 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
766 osp_statfs_need_now(d);
769 * now cancel them all
770 * XXX: can we improve this using some batching?
771 * with batch RPC that'll happen automatically?
772 * XXX: can we store ctxt in lod_device and save few cycles ?
774 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
777 llh = ctxt->loc_handle;
780 INIT_LIST_HEAD(&list);
781 spin_lock(&d->opd_syn_lock);
782 list_splice(&d->opd_syn_committed_there, &list);
783 INIT_LIST_HEAD(&d->opd_syn_committed_there);
784 spin_unlock(&d->opd_syn_lock);
786 list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
787 struct llog_cookie *lcookie = NULL;
789 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
790 list_del_init(&req->rq_exp_list);
792 if (d->opd_connect_mdt) {
793 struct object_update_request *ureq;
794 struct object_update *update;
795 ureq = req_capsule_client_get(&req->rq_pill,
797 LASSERT(ureq != NULL &&
798 ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
800 /* 1st/2nd is for decref . and .., 3rd one is for
801 * destroy, where the log cookie is stored.
802 * See osp_prep_unlink_update_req */
803 update = object_update_request_get(ureq, 2, NULL);
804 LASSERT(update != NULL);
805 lcookie = object_update_param_get(update, 0, NULL);
806 LASSERT(lcookie != NULL);
808 body = req_capsule_client_get(&req->rq_pill,
811 lcookie = &body->oa.o_lcookie;
813 /* import can be closing, thus all commit cb's are
814 * called we can check committness directly */
815 if (req->rq_transno <= imp->imp_peer_committed_transno) {
816 rc = llog_cat_cancel_records(env, llh, 1, lcookie);
818 CERROR("%s: can't cancel record: %d\n",
821 DEBUG_REQ(D_HA, req, "not committed");
824 ptlrpc_req_finished(req);
830 LASSERT(d->opd_syn_rpc_in_progress >= done);
831 spin_lock(&d->opd_syn_lock);
832 d->opd_syn_rpc_in_progress -= done;
833 spin_unlock(&d->opd_syn_lock);
834 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
835 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
836 d->opd_syn_rpc_in_progress);
838 osp_sync_check_for_work(d);
840 /* wake up the thread if requested to stop:
841 * it might be waiting for in-progress to complete */
842 if (unlikely(osp_sync_running(d) == 0))
843 wake_up(&d->opd_syn_waitq);
849 * this is where most of queues processing happens
851 static int osp_sync_process_queues(const struct lu_env *env,
852 struct llog_handle *llh,
853 struct llog_rec_hdr *rec,
856 struct osp_device *d = data;
860 struct l_wait_info lwi = { 0 };
862 if (!osp_sync_running(d)) {
863 CDEBUG(D_HA, "stop llog processing\n");
864 return LLOG_PROC_BREAK;
867 /* process requests committed by OST */
868 osp_sync_process_committed(env, d);
870 /* if we there are changes to be processed and we have
871 * resources for this ... do now */
872 if (osp_sync_can_process_new(d, rec)) {
874 /* ask llog for another record */
875 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
877 d->opd_syn_rpc_in_progress,
878 d->opd_syn_rpc_in_flight);
883 * try to send, in case of disconnection, suspend
884 * processing till we can send this request
887 rc = osp_sync_process_record(env, d, llh, rec);
889 * XXX: probably different handling is needed
890 * for some bugs, like immediate exit or if
894 CERROR("can't send: %d\n", rc);
895 l_wait_event(d->opd_syn_waitq,
896 !osp_sync_running(d) ||
897 osp_sync_has_work(d),
900 } while (rc != 0 && osp_sync_running(d));
906 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
907 osp_sync_remove_from_tracker(d);
909 l_wait_event(d->opd_syn_waitq,
910 !osp_sync_running(d) ||
911 osp_sync_can_process_new(d, rec) ||
912 !list_empty(&d->opd_syn_committed_there),
918 * this thread runs llog_cat_process() scanner calling our callback
919 * to process llog records. in the callback we implement tricky
920 * state machine as we don't want to start scanning of the llog again
921 * and again, also we don't want to process too many records and send
922 * too many RPCs a time. so, depending on current load (num of changes
923 * being synced to OST) the callback can suspend awaiting for some
924 * new conditions, like syncs completed.
926 * in order to process llog records left by previous boots and to allow
927 * llog_process_thread() to find something (otherwise it'd just exit
928 * immediately) we add a special GENERATATION record on each boot.
930 static int osp_sync_thread(void *_arg)
932 struct osp_device *d = _arg;
933 struct ptlrpc_thread *thread = &d->opd_syn_thread;
934 struct l_wait_info lwi = { 0 };
935 struct llog_ctxt *ctxt;
936 struct obd_device *obd = d->opd_obd;
937 struct llog_handle *llh;
943 rc = lu_env_init(&env, LCT_LOCAL);
945 CERROR("%s: can't initialize env: rc = %d\n",
950 spin_lock(&d->opd_syn_lock);
951 thread->t_flags = SVC_RUNNING;
952 spin_unlock(&d->opd_syn_lock);
953 wake_up(&thread->t_ctl_waitq);
955 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
957 CERROR("can't get appropriate context\n");
958 GOTO(out, rc = -EINVAL);
961 llh = ctxt->loc_handle;
963 CERROR("can't get llh\n");
965 GOTO(out, rc = -EINVAL);
968 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
969 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
970 "%lu changes, %u in progress, %u in flight: %d\n",
971 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
972 d->opd_syn_rpc_in_flight, rc);
974 /* we don't expect llog_process_thread() to exit till umount */
975 LASSERTF(thread->t_flags != SVC_RUNNING,
976 "%lu changes, %u in progress, %u in flight\n",
977 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
978 d->opd_syn_rpc_in_flight);
980 /* wait till all the requests are completed */
982 while (d->opd_syn_rpc_in_progress > 0) {
983 osp_sync_process_committed(&env, d);
985 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
986 rc = l_wait_event(d->opd_syn_waitq,
987 d->opd_syn_rpc_in_progress == 0,
989 if (rc == -ETIMEDOUT)
991 LASSERTF(count < 10, "%s: %d %d %sempty\n",
992 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
993 d->opd_syn_rpc_in_flight,
994 list_empty(&d->opd_syn_committed_there) ? "" : "!");
998 llog_cat_close(&env, llh);
999 rc = llog_cleanup(&env, ctxt);
1001 CERROR("can't cleanup llog: %d\n", rc);
1003 LASSERTF(d->opd_syn_rpc_in_progress == 0,
1004 "%s: %d %d %sempty\n",
1005 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1006 d->opd_syn_rpc_in_flight,
1007 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1009 thread->t_flags = SVC_STOPPED;
1011 wake_up(&thread->t_ctl_waitq);
1018 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1020 struct osp_thread_info *osi = osp_env_info(env);
1021 struct lu_fid *fid = &osi->osi_fid;
1022 struct llog_handle *lgh = NULL;
1023 struct obd_device *obd = d->opd_obd;
1024 struct llog_ctxt *ctxt;
1032 * open llog corresponding to our OST
1034 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1035 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1037 if (d->opd_connect_mdt)
1038 lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
1040 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1042 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1043 &osi->osi_cid, fid);
1045 CERROR("%s: can't get id from catalogs: rc = %d\n",
1050 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1051 obd->obd_name, d->opd_index,
1052 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1053 osi->osi_cid.lci_logid.lgl_ogen);
1055 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1056 &osp_mds_ost_orig_logops);
1060 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1063 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1064 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1066 /* re-create llog if it is missing */
1068 logid_set_id(&osi->osi_cid.lci_logid, 0);
1070 GOTO(out_cleanup, rc);
1073 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1074 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1076 GOTO(out_cleanup, rc);
1077 osi->osi_cid.lci_logid = lgh->lgh_id;
1080 LASSERT(lgh != NULL);
1081 ctxt->loc_handle = lgh;
1083 rc = llog_cat_init_and_process(env, lgh);
1085 GOTO(out_close, rc);
1087 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1088 &osi->osi_cid, fid);
1090 GOTO(out_close, rc);
1093 * put a mark in the llog till which we'll be processing
1094 * old records restless
1096 d->opd_syn_generation.mnt_cnt = cfs_time_current();
1097 d->opd_syn_generation.conn_cnt = cfs_time_current();
1099 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1100 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1102 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1103 sizeof(osi->osi_gen.lgr_gen));
1105 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1107 GOTO(out_close, rc);
1108 llog_ctxt_put(ctxt);
1111 llog_cat_close(env, lgh);
1113 llog_cleanup(env, ctxt);
1117 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1119 struct llog_ctxt *ctxt;
1121 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1123 llog_cat_close(env, ctxt->loc_handle);
1124 llog_cleanup(env, ctxt);
1128 * initializes sync component of OSP
1130 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1132 struct l_wait_info lwi = { 0 };
1133 struct task_struct *task;
1138 rc = osp_sync_id_traction_init(d);
1143 * initialize llog storing changes
1145 rc = osp_sync_llog_init(env, d);
1147 CERROR("%s: can't initialize llog: rc = %d\n",
1148 d->opd_obd->obd_name, rc);
1153 * Start synchronization thread
1155 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1156 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1157 spin_lock_init(&d->opd_syn_lock);
1158 init_waitqueue_head(&d->opd_syn_waitq);
1159 init_waitqueue_head(&d->opd_syn_barrier_waitq);
1160 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1161 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1163 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1164 d->opd_index, d->opd_group);
1167 CERROR("%s: cannot start sync thread: rc = %d\n",
1168 d->opd_obd->obd_name, rc);
1172 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1173 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1177 osp_sync_llog_fini(env, d);
1179 osp_sync_id_traction_fini(d);
1183 int osp_sync_fini(struct osp_device *d)
1185 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1189 thread->t_flags = SVC_STOPPING;
1190 wake_up(&d->opd_syn_waitq);
1191 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1194 * unregister transaction callbacks only when sync thread
1195 * has finished operations with llog
1197 osp_sync_id_traction_fini(d);
1202 static DEFINE_MUTEX(osp_id_tracker_sem);
1203 static struct list_head osp_id_tracker_list =
1204 LIST_HEAD_INIT(osp_id_tracker_list);
1206 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1208 struct osp_id_tracker *tr = cookie;
1209 struct osp_device *d;
1210 struct osp_txn_info *txn;
1214 txn = osp_txn_info(&th->th_ctx);
1215 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1218 spin_lock(&tr->otr_lock);
1219 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1220 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1221 tr->otr_committed_id, txn->oti_current_id);
1222 tr->otr_committed_id = txn->oti_current_id;
1224 list_for_each_entry(d, &tr->otr_wakeup_list,
1226 d->opd_syn_last_committed_id = tr->otr_committed_id;
1227 wake_up(&d->opd_syn_waitq);
1230 spin_unlock(&tr->otr_lock);
1233 static int osp_sync_id_traction_init(struct osp_device *d)
1235 struct osp_id_tracker *tr, *found = NULL;
1239 LASSERT(d->opd_storage);
1240 LASSERT(d->opd_syn_tracker == NULL);
1241 INIT_LIST_HEAD(&d->opd_syn_ontrack);
1243 mutex_lock(&osp_id_tracker_sem);
1244 list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1245 if (tr->otr_dev == d->opd_storage) {
1246 LASSERT(atomic_read(&tr->otr_refcount));
1247 atomic_inc(&tr->otr_refcount);
1248 d->opd_syn_tracker = tr;
1254 if (found == NULL) {
1258 d->opd_syn_tracker = tr;
1259 spin_lock_init(&tr->otr_lock);
1260 tr->otr_dev = d->opd_storage;
1261 tr->otr_next_id = 1;
1262 tr->otr_committed_id = 0;
1263 atomic_set(&tr->otr_refcount, 1);
1264 INIT_LIST_HEAD(&tr->otr_wakeup_list);
1265 list_add(&tr->otr_list, &osp_id_tracker_list);
1266 tr->otr_tx_cb.dtc_txn_commit =
1267 osp_sync_tracker_commit_cb;
1268 tr->otr_tx_cb.dtc_cookie = tr;
1269 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1270 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1274 mutex_unlock(&osp_id_tracker_sem);
1279 static void osp_sync_id_traction_fini(struct osp_device *d)
1281 struct osp_id_tracker *tr;
1286 tr = d->opd_syn_tracker;
1292 osp_sync_remove_from_tracker(d);
1294 mutex_lock(&osp_id_tracker_sem);
1295 if (atomic_dec_and_test(&tr->otr_refcount)) {
1296 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1297 LASSERT(list_empty(&tr->otr_wakeup_list));
1298 list_del(&tr->otr_list);
1300 d->opd_syn_tracker = NULL;
1302 mutex_unlock(&osp_id_tracker_sem);
1308 * generates id for the tracker
1310 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1312 struct osp_id_tracker *tr;
1314 tr = d->opd_syn_tracker;
1317 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1318 spin_lock(&tr->otr_lock);
1319 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1320 spin_unlock(&tr->otr_lock);
1321 CERROR("%s: next %u, last synced %lu\n",
1322 d->opd_obd->obd_name, tr->otr_next_id,
1323 d->opd_syn_last_used_id);
1328 id = tr->otr_next_id++;
1329 if (id > d->opd_syn_last_used_id)
1330 d->opd_syn_last_used_id = id;
1331 if (list_empty(&d->opd_syn_ontrack))
1332 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1333 spin_unlock(&tr->otr_lock);
1334 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1339 static void osp_sync_remove_from_tracker(struct osp_device *d)
1341 struct osp_id_tracker *tr;
1343 tr = d->opd_syn_tracker;
1346 if (list_empty(&d->opd_syn_ontrack))
1349 spin_lock(&tr->otr_lock);
1350 list_del_init(&d->opd_syn_ontrack);
1351 spin_unlock(&tr->otr_lock);