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 =
241 ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
242 ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0);
248 txn = osp_txn_info(&th->th_ctx);
251 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
252 osi->osi_hdr.lrh_id = txn->oti_current_id;
254 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
257 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
261 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
262 d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
263 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
264 (unsigned long) osi->osi_cookie.lgc_index, rc);
269 if (likely(rc == 0)) {
270 spin_lock(&d->opd_syn_lock);
271 d->opd_syn_changes++;
272 spin_unlock(&d->opd_syn_lock);
278 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
279 llog_op_type type, struct thandle *th,
280 const struct lu_attr *attr)
282 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
283 lu_object_fid(&o->opo_obj.do_lu), type, 1,
287 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
288 struct lu_fid *fid, int lost, struct thandle *th)
290 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
294 * it's quite obvious we can't maintain all the structures in the memory:
295 * while OST is down, MDS can be processing thousands and thousands of unlinks
296 * filling persistent llogs and in-core respresentation
298 * this doesn't scale at all. so we need basically the following:
299 * a) destroy/setattr append llog records
300 * b) once llog has grown to X records, we process first Y committed records
302 * once record R is found via llog_process(), it becomes committed after any
303 * subsequent commit callback (at the most)
307 * called for each atomic on-disk change (not once per transaction batch)
308 * and goes over the list
309 * XXX: should be optimized?
313 * called for each RPC reported committed
315 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
317 struct osp_device *d = req->rq_cb_data;
319 CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
321 if (unlikely(req->rq_transno == 0))
324 /* do not do any opd_dyn_rpc_* accounting here
325 * it's done in osp_sync_interpret sooner or later */
328 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
329 LASSERT(list_empty(&req->rq_exp_list));
331 ptlrpc_request_addref(req);
333 spin_lock(&d->opd_syn_lock);
334 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
335 spin_unlock(&d->opd_syn_lock);
337 /* XXX: some batching wouldn't hurt */
338 wake_up(&d->opd_syn_waitq);
341 static int osp_sync_interpret(const struct lu_env *env,
342 struct ptlrpc_request *req, void *aa, int rc)
344 struct osp_device *d = req->rq_cb_data;
346 if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
347 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
348 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
351 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
352 atomic_read(&req->rq_refcount),
353 rc, (unsigned) req->rq_transno);
354 LASSERT(rc || req->rq_transno);
358 * we tried to destroy object or update attributes,
359 * but object doesn't exist anymore - cancell llog record
361 LASSERT(req->rq_transno == 0);
362 LASSERT(list_empty(&req->rq_exp_list));
364 ptlrpc_request_addref(req);
366 spin_lock(&d->opd_syn_lock);
367 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
368 spin_unlock(&d->opd_syn_lock);
370 wake_up(&d->opd_syn_waitq);
372 struct obd_import *imp = req->rq_import;
374 * error happened, we'll try to repeat on next boot ?
376 LASSERTF(req->rq_transno == 0 ||
377 req->rq_import_generation < imp->imp_generation,
378 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
379 req->rq_transno, rc, req->rq_import_generation,
380 imp->imp_generation);
381 if (req->rq_transno == 0) {
382 /* this is the last time we see the request
383 * if transno is not zero, then commit cb
384 * will be called at some point */
385 LASSERT(d->opd_syn_rpc_in_progress > 0);
386 spin_lock(&d->opd_syn_lock);
387 d->opd_syn_rpc_in_progress--;
388 spin_unlock(&d->opd_syn_lock);
391 wake_up(&d->opd_syn_waitq);
392 } else if (d->opd_pre != NULL &&
393 unlikely(d->opd_pre_status == -ENOSPC)) {
395 * if current status is -ENOSPC (lack of free space on OST)
396 * then we should poll OST immediately once object destroy
399 osp_statfs_need_now(d);
402 LASSERT(d->opd_syn_rpc_in_flight > 0);
403 spin_lock(&d->opd_syn_lock);
404 d->opd_syn_rpc_in_flight--;
405 spin_unlock(&d->opd_syn_lock);
406 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
407 wake_up(&d->opd_syn_barrier_waitq);
408 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
409 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
410 d->opd_syn_rpc_in_progress);
412 osp_sync_check_for_work(d);
418 * the function walks through list of committed locally changes
419 * and send them to RPC until the pipe is full
421 static void osp_sync_send_new_rpc(struct osp_device *d,
422 struct ptlrpc_request *req)
424 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
425 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
427 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
430 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
431 struct llog_handle *llh,
432 struct llog_rec_hdr *h,
434 const struct req_format *format)
436 struct ptlrpc_request *req;
437 struct ost_body *body;
438 struct obd_import *imp;
441 /* Prepare the request */
442 imp = d->opd_obd->u.cli.cl_import;
444 req = ptlrpc_request_alloc(imp, format);
446 RETURN(ERR_PTR(-ENOMEM));
448 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
450 ptlrpc_req_finished(req);
455 * this is a trick: to save on memory allocations we put cookie
456 * into the request, but don't set corresponded flag in o_valid
457 * so that OST doesn't interpret this cookie. once the request
458 * is committed on OST we take cookie from the request and cancel
460 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
462 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
463 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
464 body->oa.o_lcookie.lgc_index = h->lrh_index;
465 INIT_LIST_HEAD(&req->rq_exp_list);
466 req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
468 req->rq_interpret_reply = osp_sync_interpret;
469 req->rq_commit_cb = osp_sync_request_commit_cb;
472 ptlrpc_request_set_replen(req);
477 static int osp_sync_new_setattr_job(struct osp_device *d,
478 struct llog_handle *llh,
479 struct llog_rec_hdr *h)
481 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
482 struct ptlrpc_request *req;
483 struct ost_body *body;
486 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
488 /* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
489 * so no bits other than these should be set. */
490 if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
491 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
492 d->opd_obd->obd_name, rec->lsr_valid);
493 /* return 0 so that sync thread can continue processing
498 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
500 RETURN(PTR_ERR(req));
502 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
504 body->oa.o_oi = rec->lsr_oi;
505 body->oa.o_uid = rec->lsr_uid;
506 body->oa.o_gid = rec->lsr_gid;
507 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
508 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
509 * we assume that both UID and GID are valid in that case. */
510 if (rec->lsr_valid == 0)
511 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
513 body->oa.o_valid |= rec->lsr_valid;
515 osp_sync_send_new_rpc(d, req);
519 static int osp_sync_new_unlink_job(struct osp_device *d,
520 struct llog_handle *llh,
521 struct llog_rec_hdr *h)
523 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
524 struct ptlrpc_request *req;
525 struct ost_body *body;
528 LASSERT(h->lrh_type == MDS_UNLINK_REC);
530 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
532 RETURN(PTR_ERR(req));
534 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
536 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
537 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
538 body->oa.o_misc = rec->lur_count;
539 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
541 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
543 osp_sync_send_new_rpc(d, req);
547 static int osp_prep_unlink_update_req(const struct lu_env *env,
548 struct osp_device *osp,
549 struct llog_handle *llh,
550 struct llog_rec_hdr *h,
551 struct ptlrpc_request **reqp)
553 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
554 struct dt_update_request *update = NULL;
555 struct ptlrpc_request *req;
557 struct llog_cookie lcookie;
562 update = out_create_update_req(&osp->opd_dt_dev);
564 RETURN(PTR_ERR(update));
566 /* This can only happens for unlink slave directory, so decrease
567 * ref for ".." and "." */
568 rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
573 rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
578 lcookie.lgc_lgl = llh->lgh_id;
579 lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
580 lcookie.lgc_index = h->lrh_index;
581 size = sizeof(lcookie);
582 buf = (const char *)&lcookie;
584 rc = out_insert_update(env, update, OUT_DESTROY, &rec->lur_fid, 1,
589 rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
590 update->dur_req, &req);
594 INIT_LIST_HEAD(&req->rq_exp_list);
595 req->rq_svc_thread = (void *)OSP_JOB_MAGIC;
597 req->rq_interpret_reply = osp_sync_interpret;
598 req->rq_commit_cb = osp_sync_request_commit_cb;
599 req->rq_cb_data = osp;
601 ptlrpc_request_set_replen(req);
605 out_destroy_update_req(update);
610 static int osp_sync_new_unlink64_job(const struct lu_env *env,
611 struct osp_device *d,
612 struct llog_handle *llh,
613 struct llog_rec_hdr *h)
615 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
616 struct ptlrpc_request *req = NULL;
617 struct ost_body *body;
621 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
623 if (d->opd_connect_mdt) {
624 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
628 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
631 RETURN(PTR_ERR(req));
633 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
636 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
639 body->oa.o_misc = rec->lur_count;
640 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
643 osp_sync_send_new_rpc(d, req);
647 static int osp_sync_process_record(const struct lu_env *env,
648 struct osp_device *d,
649 struct llog_handle *llh,
650 struct llog_rec_hdr *rec)
652 struct llog_cookie cookie;
655 cookie.lgc_lgl = llh->lgh_id;
656 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
657 cookie.lgc_index = rec->lrh_index;
659 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
660 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
662 /* we're waiting for the record generated by this instance */
663 LASSERT(d->opd_syn_prev_done == 0);
664 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
665 sizeof(gen->lgr_gen))) {
666 CDEBUG(D_HA, "processed all old entries\n");
667 d->opd_syn_prev_done = 1;
670 /* cancel any generation record */
671 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
678 * now we prepare and fill requests to OST, put them on the queue
679 * and fire after next commit callback
682 /* notice we increment counters before sending RPC, to be consistent
683 * in RPC interpret callback which may happen very quickly */
684 spin_lock(&d->opd_syn_lock);
685 d->opd_syn_rpc_in_flight++;
686 d->opd_syn_rpc_in_progress++;
687 spin_unlock(&d->opd_syn_lock);
689 switch (rec->lrh_type) {
690 /* case MDS_UNLINK_REC is kept for compatibility */
692 rc = osp_sync_new_unlink_job(d, llh, rec);
694 case MDS_UNLINK64_REC:
695 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
697 case MDS_SETATTR64_REC:
698 rc = osp_sync_new_setattr_job(d, llh, rec);
701 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
703 /* we should continue processing */
706 /* rc > 0 means sync RPC being added to the queue */
707 if (likely(rc > 0)) {
708 spin_lock(&d->opd_syn_lock);
709 if (d->opd_syn_prev_done) {
710 LASSERT(d->opd_syn_changes > 0);
711 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
713 * NOTE: it's possible to meet same id if
714 * OST stores few stripes of same file
716 if (rec->lrh_id > d->opd_syn_last_processed_id) {
717 d->opd_syn_last_processed_id = rec->lrh_id;
718 wake_up(&d->opd_syn_barrier_waitq);
721 d->opd_syn_changes--;
723 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
724 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
725 d->opd_syn_rpc_in_progress);
726 spin_unlock(&d->opd_syn_lock);
729 spin_lock(&d->opd_syn_lock);
730 d->opd_syn_rpc_in_flight--;
731 d->opd_syn_rpc_in_progress--;
732 spin_unlock(&d->opd_syn_lock);
735 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
736 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
740 static void osp_sync_process_committed(const struct lu_env *env,
741 struct osp_device *d)
743 struct obd_device *obd = d->opd_obd;
744 struct obd_import *imp = obd->u.cli.cl_import;
745 struct ost_body *body;
746 struct ptlrpc_request *req, *tmp;
747 struct llog_ctxt *ctxt;
748 struct llog_handle *llh;
749 struct list_head list;
754 if (list_empty(&d->opd_syn_committed_there))
758 * if current status is -ENOSPC (lack of free space on OST)
759 * then we should poll OST immediately once object destroy
761 * notice: we do this upon commit as well because some backends
762 * (like DMU) do not release space right away.
764 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
765 osp_statfs_need_now(d);
768 * now cancel them all
769 * XXX: can we improve this using some batching?
770 * with batch RPC that'll happen automatically?
771 * XXX: can we store ctxt in lod_device and save few cycles ?
773 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
776 llh = ctxt->loc_handle;
779 INIT_LIST_HEAD(&list);
780 spin_lock(&d->opd_syn_lock);
781 list_splice(&d->opd_syn_committed_there, &list);
782 INIT_LIST_HEAD(&d->opd_syn_committed_there);
783 spin_unlock(&d->opd_syn_lock);
785 list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
786 struct llog_cookie *lcookie = NULL;
788 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
789 list_del_init(&req->rq_exp_list);
791 if (d->opd_connect_mdt) {
792 struct object_update_request *ureq;
793 struct object_update *update;
794 ureq = req_capsule_client_get(&req->rq_pill,
796 LASSERT(ureq != NULL &&
797 ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
799 /* 1st/2nd is for decref . and .., 3rd one is for
800 * destroy, where the log cookie is stored.
801 * See osp_prep_unlink_update_req */
802 update = object_update_request_get(ureq, 2, NULL);
803 LASSERT(update != NULL);
804 lcookie = object_update_param_get(update, 0, NULL);
805 LASSERT(lcookie != NULL);
807 body = req_capsule_client_get(&req->rq_pill,
810 lcookie = &body->oa.o_lcookie;
812 /* import can be closing, thus all commit cb's are
813 * called we can check committness directly */
814 if (req->rq_transno <= imp->imp_peer_committed_transno) {
815 rc = llog_cat_cancel_records(env, llh, 1, lcookie);
817 CERROR("%s: can't cancel record: %d\n",
820 DEBUG_REQ(D_HA, req, "not committed");
823 ptlrpc_req_finished(req);
829 LASSERT(d->opd_syn_rpc_in_progress >= done);
830 spin_lock(&d->opd_syn_lock);
831 d->opd_syn_rpc_in_progress -= done;
832 spin_unlock(&d->opd_syn_lock);
833 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
834 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
835 d->opd_syn_rpc_in_progress);
837 osp_sync_check_for_work(d);
839 /* wake up the thread if requested to stop:
840 * it might be waiting for in-progress to complete */
841 if (unlikely(osp_sync_running(d) == 0))
842 wake_up(&d->opd_syn_waitq);
848 * this is where most of queues processing happens
850 static int osp_sync_process_queues(const struct lu_env *env,
851 struct llog_handle *llh,
852 struct llog_rec_hdr *rec,
855 struct osp_device *d = data;
859 struct l_wait_info lwi = { 0 };
861 if (!osp_sync_running(d)) {
862 CDEBUG(D_HA, "stop llog processing\n");
863 return LLOG_PROC_BREAK;
866 /* process requests committed by OST */
867 osp_sync_process_committed(env, d);
869 /* if we there are changes to be processed and we have
870 * resources for this ... do now */
871 if (osp_sync_can_process_new(d, rec)) {
873 /* ask llog for another record */
874 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
876 d->opd_syn_rpc_in_progress,
877 d->opd_syn_rpc_in_flight);
882 * try to send, in case of disconnection, suspend
883 * processing till we can send this request
886 rc = osp_sync_process_record(env, d, llh, rec);
888 * XXX: probably different handling is needed
889 * for some bugs, like immediate exit or if
893 CERROR("can't send: %d\n", rc);
894 l_wait_event(d->opd_syn_waitq,
895 !osp_sync_running(d) ||
896 osp_sync_has_work(d),
899 } while (rc != 0 && osp_sync_running(d));
905 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
906 osp_sync_remove_from_tracker(d);
908 l_wait_event(d->opd_syn_waitq,
909 !osp_sync_running(d) ||
910 osp_sync_can_process_new(d, rec) ||
911 !list_empty(&d->opd_syn_committed_there),
917 * this thread runs llog_cat_process() scanner calling our callback
918 * to process llog records. in the callback we implement tricky
919 * state machine as we don't want to start scanning of the llog again
920 * and again, also we don't want to process too many records and send
921 * too many RPCs a time. so, depending on current load (num of changes
922 * being synced to OST) the callback can suspend awaiting for some
923 * new conditions, like syncs completed.
925 * in order to process llog records left by previous boots and to allow
926 * llog_process_thread() to find something (otherwise it'd just exit
927 * immediately) we add a special GENERATATION record on each boot.
929 static int osp_sync_thread(void *_arg)
931 struct osp_device *d = _arg;
932 struct ptlrpc_thread *thread = &d->opd_syn_thread;
933 struct l_wait_info lwi = { 0 };
934 struct llog_ctxt *ctxt;
935 struct obd_device *obd = d->opd_obd;
936 struct llog_handle *llh;
942 rc = lu_env_init(&env, LCT_LOCAL);
944 CERROR("%s: can't initialize env: rc = %d\n",
949 spin_lock(&d->opd_syn_lock);
950 thread->t_flags = SVC_RUNNING;
951 spin_unlock(&d->opd_syn_lock);
952 wake_up(&thread->t_ctl_waitq);
954 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
956 CERROR("can't get appropriate context\n");
957 GOTO(out, rc = -EINVAL);
960 llh = ctxt->loc_handle;
962 CERROR("can't get llh\n");
964 GOTO(out, rc = -EINVAL);
967 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
968 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
969 "%lu changes, %u in progress, %u in flight: %d\n",
970 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
971 d->opd_syn_rpc_in_flight, rc);
973 /* we don't expect llog_process_thread() to exit till umount */
974 LASSERTF(thread->t_flags != SVC_RUNNING,
975 "%lu changes, %u in progress, %u in flight\n",
976 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
977 d->opd_syn_rpc_in_flight);
979 /* wait till all the requests are completed */
981 while (d->opd_syn_rpc_in_progress > 0) {
982 osp_sync_process_committed(&env, d);
984 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
985 rc = l_wait_event(d->opd_syn_waitq,
986 d->opd_syn_rpc_in_progress == 0,
988 if (rc == -ETIMEDOUT)
990 LASSERTF(count < 10, "%s: %d %d %sempty\n",
991 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
992 d->opd_syn_rpc_in_flight,
993 list_empty(&d->opd_syn_committed_there) ? "" : "!");
997 llog_cat_close(&env, llh);
998 rc = llog_cleanup(&env, ctxt);
1000 CERROR("can't cleanup llog: %d\n", rc);
1002 LASSERTF(d->opd_syn_rpc_in_progress == 0,
1003 "%s: %d %d %sempty\n",
1004 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1005 d->opd_syn_rpc_in_flight,
1006 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1008 thread->t_flags = SVC_STOPPED;
1010 wake_up(&thread->t_ctl_waitq);
1017 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1019 struct osp_thread_info *osi = osp_env_info(env);
1020 struct lu_fid *fid = &osi->osi_fid;
1021 struct llog_handle *lgh = NULL;
1022 struct obd_device *obd = d->opd_obd;
1023 struct llog_ctxt *ctxt;
1031 * open llog corresponding to our OST
1033 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1034 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1036 if (d->opd_connect_mdt)
1037 lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
1039 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1041 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1042 &osi->osi_cid, fid);
1044 CERROR("%s: can't get id from catalogs: rc = %d\n",
1049 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1050 obd->obd_name, d->opd_index,
1051 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1052 osi->osi_cid.lci_logid.lgl_ogen);
1054 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1055 &osp_mds_ost_orig_logops);
1059 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1062 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1063 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1065 /* re-create llog if it is missing */
1067 logid_set_id(&osi->osi_cid.lci_logid, 0);
1069 GOTO(out_cleanup, rc);
1072 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1073 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1075 GOTO(out_cleanup, rc);
1076 osi->osi_cid.lci_logid = lgh->lgh_id;
1079 LASSERT(lgh != NULL);
1080 ctxt->loc_handle = lgh;
1082 rc = llog_cat_init_and_process(env, lgh);
1084 GOTO(out_close, rc);
1086 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1087 &osi->osi_cid, fid);
1089 GOTO(out_close, rc);
1092 * put a mark in the llog till which we'll be processing
1093 * old records restless
1095 d->opd_syn_generation.mnt_cnt = cfs_time_current();
1096 d->opd_syn_generation.conn_cnt = cfs_time_current();
1098 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1099 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1101 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1102 sizeof(osi->osi_gen.lgr_gen));
1104 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1106 GOTO(out_close, rc);
1107 llog_ctxt_put(ctxt);
1110 llog_cat_close(env, lgh);
1112 llog_cleanup(env, ctxt);
1116 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1118 struct llog_ctxt *ctxt;
1120 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1122 llog_cat_close(env, ctxt->loc_handle);
1123 llog_cleanup(env, ctxt);
1127 * initializes sync component of OSP
1129 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1131 struct l_wait_info lwi = { 0 };
1132 struct task_struct *task;
1137 rc = osp_sync_id_traction_init(d);
1142 * initialize llog storing changes
1144 rc = osp_sync_llog_init(env, d);
1146 CERROR("%s: can't initialize llog: rc = %d\n",
1147 d->opd_obd->obd_name, rc);
1152 * Start synchronization thread
1154 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1155 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1156 spin_lock_init(&d->opd_syn_lock);
1157 init_waitqueue_head(&d->opd_syn_waitq);
1158 init_waitqueue_head(&d->opd_syn_barrier_waitq);
1159 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1160 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1162 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1163 d->opd_index, d->opd_group);
1166 CERROR("%s: cannot start sync thread: rc = %d\n",
1167 d->opd_obd->obd_name, rc);
1171 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1172 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1176 osp_sync_llog_fini(env, d);
1178 osp_sync_id_traction_fini(d);
1182 int osp_sync_fini(struct osp_device *d)
1184 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1188 thread->t_flags = SVC_STOPPING;
1189 wake_up(&d->opd_syn_waitq);
1190 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1193 * unregister transaction callbacks only when sync thread
1194 * has finished operations with llog
1196 osp_sync_id_traction_fini(d);
1201 static DEFINE_MUTEX(osp_id_tracker_sem);
1202 static struct list_head osp_id_tracker_list =
1203 LIST_HEAD_INIT(osp_id_tracker_list);
1205 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1207 struct osp_id_tracker *tr = cookie;
1208 struct osp_device *d;
1209 struct osp_txn_info *txn;
1213 txn = osp_txn_info(&th->th_ctx);
1214 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1217 spin_lock(&tr->otr_lock);
1218 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1219 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1220 tr->otr_committed_id, txn->oti_current_id);
1221 tr->otr_committed_id = txn->oti_current_id;
1223 list_for_each_entry(d, &tr->otr_wakeup_list,
1225 d->opd_syn_last_committed_id = tr->otr_committed_id;
1226 wake_up(&d->opd_syn_waitq);
1229 spin_unlock(&tr->otr_lock);
1232 static int osp_sync_id_traction_init(struct osp_device *d)
1234 struct osp_id_tracker *tr, *found = NULL;
1238 LASSERT(d->opd_storage);
1239 LASSERT(d->opd_syn_tracker == NULL);
1240 INIT_LIST_HEAD(&d->opd_syn_ontrack);
1242 mutex_lock(&osp_id_tracker_sem);
1243 list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1244 if (tr->otr_dev == d->opd_storage) {
1245 LASSERT(atomic_read(&tr->otr_refcount));
1246 atomic_inc(&tr->otr_refcount);
1247 d->opd_syn_tracker = tr;
1253 if (found == NULL) {
1257 d->opd_syn_tracker = tr;
1258 spin_lock_init(&tr->otr_lock);
1259 tr->otr_dev = d->opd_storage;
1260 tr->otr_next_id = 1;
1261 tr->otr_committed_id = 0;
1262 atomic_set(&tr->otr_refcount, 1);
1263 INIT_LIST_HEAD(&tr->otr_wakeup_list);
1264 list_add(&tr->otr_list, &osp_id_tracker_list);
1265 tr->otr_tx_cb.dtc_txn_commit =
1266 osp_sync_tracker_commit_cb;
1267 tr->otr_tx_cb.dtc_cookie = tr;
1268 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1269 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1273 mutex_unlock(&osp_id_tracker_sem);
1278 static void osp_sync_id_traction_fini(struct osp_device *d)
1280 struct osp_id_tracker *tr;
1285 tr = d->opd_syn_tracker;
1291 osp_sync_remove_from_tracker(d);
1293 mutex_lock(&osp_id_tracker_sem);
1294 if (atomic_dec_and_test(&tr->otr_refcount)) {
1295 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1296 LASSERT(list_empty(&tr->otr_wakeup_list));
1297 list_del(&tr->otr_list);
1299 d->opd_syn_tracker = NULL;
1301 mutex_unlock(&osp_id_tracker_sem);
1307 * generates id for the tracker
1309 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1311 struct osp_id_tracker *tr;
1313 tr = d->opd_syn_tracker;
1316 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1317 spin_lock(&tr->otr_lock);
1318 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1319 spin_unlock(&tr->otr_lock);
1320 CERROR("%s: next %u, last synced %lu\n",
1321 d->opd_obd->obd_name, tr->otr_next_id,
1322 d->opd_syn_last_used_id);
1327 id = tr->otr_next_id++;
1328 if (id > d->opd_syn_last_used_id)
1329 d->opd_syn_last_used_id = id;
1330 if (list_empty(&d->opd_syn_ontrack))
1331 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1332 spin_unlock(&tr->otr_lock);
1333 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1338 static void osp_sync_remove_from_tracker(struct osp_device *d)
1340 struct osp_id_tracker *tr;
1342 tr = d->opd_syn_tracker;
1345 if (list_empty(&d->opd_syn_ontrack))
1348 spin_lock(&tr->otr_lock);
1349 list_del_init(&d->opd_syn_ontrack);
1350 spin_unlock(&tr->otr_lock);