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, 2014, 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 <lustre_update.h>
48 #include "osp_internal.h"
50 static int osp_sync_id_traction_init(struct osp_device *d);
51 static void osp_sync_id_traction_fini(struct osp_device *d);
52 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
53 static void osp_sync_remove_from_tracker(struct osp_device *d);
56 * this is a components of OSP implementing synchronization between MDS and OST
57 * it llogs all interesting changes (currently it's uig/gid change and object
58 * destroy) atomically, then makes sure changes hit OST storage
60 * we have 4 queues of work:
62 * the first queue is llog itself, once read a change is stored in 2nd queue
63 * in form of RPC (but RPC isn't fired yet).
65 * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
66 * commit. once change is committed locally it migrates onto 3rd queue.
68 * the third queue (opd_syn_committed_here) holds changes committed locally,
69 * but not sent to OST (as the pipe can be full). once pipe becomes non-full
70 * we take a change from the queue and fire corresponded RPC.
72 * once RPC is reported committed by OST (using regular last_committed mech.)
73 * the change jumps into 4th queue (opd_syn_committed_there), now we can
74 * cancel corresponded llog record and release RPC
76 * opd_syn_changes is a number of unread llog records (to be processed).
77 * notice this number doesn't include llog records from previous boots.
78 * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
80 * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
81 * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
82 * too much memory -- how to deal with 1000th OSTs ? batching could help?
84 * opd_syn_rpc_in_flight is a number of RPC in flight.
85 * we control this with OSP_MAX_IN_FLIGHT
88 /* XXX: do math to learn reasonable threshold
89 * should it be ~ number of changes fitting bulk? */
91 #define OSP_SYN_THRESHOLD 10
92 #define OSP_MAX_IN_FLIGHT 8
93 #define OSP_MAX_IN_PROGRESS 4096
95 #define OSP_JOB_MAGIC 0x26112005
97 struct osp_job_req_args {
98 /** bytes reserved for ptlrpc_replay_req() */
99 struct ptlrpc_replay_async_args jra_raa;
100 struct list_head jra_link;
104 static inline int osp_sync_running(struct osp_device *d)
106 return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
110 * Check status: whether OSP thread has stopped
112 * \param[in] d OSP device
114 * \retval 0 still running
117 static inline int osp_sync_stopped(struct osp_device *d)
119 return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
123 ** Check for new changes to sync
125 * \param[in] d OSP device
127 * \retval 1 there are changes
128 * \retval 0 there are no changes
130 static inline int osp_sync_has_new_job(struct osp_device *d)
132 return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
133 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
134 || (d->opd_syn_prev_done == 0);
137 static inline int osp_sync_low_in_progress(struct osp_device *d)
139 return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
143 * Check for room in the network pipe to OST
145 * \param[in] d OSP device
147 * \retval 1 there is room
148 * \retval 0 no room, the pipe is full
150 static inline int osp_sync_low_in_flight(struct osp_device *d)
152 return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
156 * Wake up check for the main sync thread
158 * \param[in] d OSP device
160 * \retval 1 time to wake up
161 * \retval 0 no need to wake up
163 static inline int osp_sync_has_work(struct osp_device *d)
165 /* has new/old changes and low in-progress? */
166 if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
167 osp_sync_low_in_flight(d) && d->opd_imp_connected)
170 /* has remotely committed? */
171 if (!list_empty(&d->opd_syn_committed_there))
177 #define osp_sync_check_for_work(d) \
179 if (osp_sync_has_work(d)) { \
180 wake_up(&d->opd_syn_waitq); \
184 void __osp_sync_check_for_work(struct osp_device *d)
186 osp_sync_check_for_work(d);
190 * Check and return ready-for-new status.
192 * The thread processing llog record uses this function to check whether
193 * it's time to take another record and process it. The number of conditions
194 * must be met: the connection should be ready, RPCs in flight not exceeding
195 * the limit, the record is committed locally, etc (see the lines below).
197 * \param[in] d OSP device
198 * \param[in] rec next llog record to process
200 * \retval 0 not ready
203 static inline int osp_sync_can_process_new(struct osp_device *d,
204 struct llog_rec_hdr *rec)
208 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
210 if (!osp_sync_low_in_progress(d))
212 if (!osp_sync_low_in_flight(d))
214 if (!d->opd_imp_connected)
216 if (d->opd_syn_prev_done == 0)
218 if (d->opd_syn_changes == 0)
220 if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
226 * Declare intention to add a new change.
228 * With regard to OSD API, we have to declare any changes ahead. In this
229 * case we declare an intention to add a llog record representing the
230 * change on the local storage.
232 * \param[in] env LU environment provided by the caller
233 * \param[in] o OSP object
234 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
235 * \param[in] th transaction handle (local)
237 * \retval 0 on success
238 * \retval negative negated errno on error
240 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
241 llog_op_type type, struct thandle *th)
243 struct osp_thread_info *osi = osp_env_info(env);
244 struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
245 struct llog_ctxt *ctxt;
246 struct thandle *storage_th;
251 /* it's a layering violation, to access internals of th,
252 * but we can do this as a sanity check, for a while */
253 LASSERT(th->th_top != NULL);
254 storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
255 if (IS_ERR(storage_th))
256 RETURN(PTR_ERR(storage_th));
259 case MDS_UNLINK64_REC:
260 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
262 case MDS_SETATTR64_REC:
263 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
269 /* we want ->dt_trans_start() to allocate per-thandle structure */
270 storage_th->th_tags |= LCT_OSP_THREAD;
272 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
275 rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
283 * Generate a llog record for a given change.
285 * Generates a llog record for the change passed. The change can be of two
286 * types: unlink and setattr. The record gets an ID which later will be
287 * used to track commit status of the change. For unlink changes, the caller
288 * can supply a starting FID and the count of the objects to destroy. For
289 * setattr the caller should apply attributes to apply.
292 * \param[in] env LU environment provided by the caller
293 * \param[in] d OSP device
294 * \param[in] fid fid of the object the change should be applied to
295 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
296 * \param[in] count count of objects to destroy
297 * \param[in] th transaction handle (local)
298 * \param[in] attr attributes for setattr
300 * \retval 0 on success
301 * \retval negative negated errno on error
303 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
304 const struct lu_fid *fid, llog_op_type type,
305 int count, struct thandle *th,
306 const struct lu_attr *attr)
308 struct osp_thread_info *osi = osp_env_info(env);
309 struct llog_ctxt *ctxt;
310 struct osp_txn_info *txn;
311 struct thandle *storage_th;
316 /* it's a layering violation, to access internals of th,
317 * but we can do this as a sanity check, for a while */
318 LASSERT(th->th_top != NULL);
319 storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
320 if (IS_ERR(storage_th))
321 RETURN(PTR_ERR(storage_th));
324 case MDS_UNLINK64_REC:
325 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
326 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
327 osi->osi_unlink.lur_fid = *fid;
328 osi->osi_unlink.lur_count = count;
330 case MDS_SETATTR64_REC:
331 rc = fid_to_ostid(fid, &osi->osi_oi);
333 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
334 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
335 osi->osi_setattr.lsr_oi = osi->osi_oi;
337 osi->osi_setattr.lsr_uid = attr->la_uid;
338 osi->osi_setattr.lsr_gid = attr->la_gid;
339 osi->osi_setattr.lsr_valid =
340 ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
341 ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0);
347 txn = osp_txn_info(&storage_th->th_ctx);
350 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
351 osi->osi_hdr.lrh_id = txn->oti_current_id;
353 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
357 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
361 if (likely(rc >= 0)) {
362 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
363 d->opd_obd->obd_name,
364 POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
365 (unsigned long)osi->osi_cookie.lgc_lgl.lgl_ogen,
366 (unsigned long)osi->osi_cookie.lgc_index, rc);
367 spin_lock(&d->opd_syn_lock);
368 d->opd_syn_changes++;
369 spin_unlock(&d->opd_syn_lock);
371 /* return 0 always here, error case just cause no llog record */
375 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
376 llog_op_type type, struct thandle *th,
377 const struct lu_attr *attr)
379 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
380 lu_object_fid(&o->opo_obj.do_lu), type, 1,
384 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
385 struct lu_fid *fid, int lost, struct thandle *th)
387 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
391 * it's quite obvious we can't maintain all the structures in the memory:
392 * while OST is down, MDS can be processing thousands and thousands of unlinks
393 * filling persistent llogs and in-core respresentation
395 * this doesn't scale at all. so we need basically the following:
396 * a) destroy/setattr append llog records
397 * b) once llog has grown to X records, we process first Y committed records
399 * once record R is found via llog_process(), it becomes committed after any
400 * subsequent commit callback (at the most)
404 * ptlrpc commit callback.
406 * The callback is called by PTLRPC when a RPC is reported committed by the
407 * target (OST). We register the callback for the every RPC applying a change
408 * from the llog. This way we know then the llog records can be cancelled.
409 * Notice the callback can be called when OSP is finishing. We can detect this
410 * checking that actual transno in the request is less or equal of known
411 * committed transno (see osp_sync_process_committed() for the details).
412 * XXX: this is pretty expensive and can be improved later using batching.
414 * \param[in] req request
416 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
418 struct osp_device *d = req->rq_cb_data;
419 struct osp_job_req_args *jra;
421 CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
423 if (unlikely(req->rq_transno == 0))
426 /* do not do any opd_dyn_rpc_* accounting here
427 * it's done in osp_sync_interpret sooner or later */
430 jra = ptlrpc_req_async_args(req);
431 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
432 LASSERT(list_empty(&jra->jra_link));
434 ptlrpc_request_addref(req);
436 spin_lock(&d->opd_syn_lock);
437 list_add(&jra->jra_link, &d->opd_syn_committed_there);
438 spin_unlock(&d->opd_syn_lock);
440 /* XXX: some batching wouldn't hurt */
441 wake_up(&d->opd_syn_waitq);
445 * RPC interpretation callback.
447 * The callback is called by ptlrpc when RPC is replied. Now we have to decide
449 * - put request on a special list to wait until it's committed by the target,
450 * if the request is successful
451 * - schedule llog record cancel if no target object is found
452 * - try later (essentially after reboot) in case of unexpected error
454 * \param[in] env LU environment provided by the caller
455 * \param[in] req request replied
456 * \param[in] aa callback data
457 * \param[in] rc result of RPC
461 static int osp_sync_interpret(const struct lu_env *env,
462 struct ptlrpc_request *req, void *aa, int rc)
464 struct osp_device *d = req->rq_cb_data;
465 struct osp_job_req_args *jra = aa;
467 if (jra->jra_magic != OSP_JOB_MAGIC) {
468 DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
473 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
474 atomic_read(&req->rq_refcount),
475 rc, (unsigned) req->rq_transno);
476 LASSERT(rc || req->rq_transno);
480 * we tried to destroy object or update attributes,
481 * but object doesn't exist anymore - cancell llog record
483 LASSERT(req->rq_transno == 0);
484 LASSERT(list_empty(&jra->jra_link));
486 ptlrpc_request_addref(req);
488 spin_lock(&d->opd_syn_lock);
489 list_add(&jra->jra_link, &d->opd_syn_committed_there);
490 spin_unlock(&d->opd_syn_lock);
492 wake_up(&d->opd_syn_waitq);
494 struct obd_import *imp = req->rq_import;
496 * error happened, we'll try to repeat on next boot ?
498 LASSERTF(req->rq_transno == 0 ||
499 req->rq_import_generation < imp->imp_generation,
500 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
501 req->rq_transno, rc, req->rq_import_generation,
502 imp->imp_generation);
503 if (req->rq_transno == 0) {
504 /* this is the last time we see the request
505 * if transno is not zero, then commit cb
506 * will be called at some point */
507 LASSERT(d->opd_syn_rpc_in_progress > 0);
508 spin_lock(&d->opd_syn_lock);
509 d->opd_syn_rpc_in_progress--;
510 spin_unlock(&d->opd_syn_lock);
513 wake_up(&d->opd_syn_waitq);
514 } else if (d->opd_pre != NULL &&
515 unlikely(d->opd_pre_status == -ENOSPC)) {
517 * if current status is -ENOSPC (lack of free space on OST)
518 * then we should poll OST immediately once object destroy
521 osp_statfs_need_now(d);
524 LASSERT(d->opd_syn_rpc_in_flight > 0);
525 spin_lock(&d->opd_syn_lock);
526 d->opd_syn_rpc_in_flight--;
527 spin_unlock(&d->opd_syn_lock);
528 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
529 wake_up(&d->opd_syn_barrier_waitq);
530 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
531 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
532 d->opd_syn_rpc_in_progress);
534 osp_sync_check_for_work(d);
540 ** Add request to ptlrpc queue.
542 * This is just a tiny helper function to put the request on the sending list
544 * \param[in] d OSP device
545 * \param[in] req request
547 static void osp_sync_send_new_rpc(struct osp_device *d,
548 struct ptlrpc_request *req)
550 struct osp_job_req_args *jra;
552 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
554 jra = ptlrpc_req_async_args(req);
555 jra->jra_magic = OSP_JOB_MAGIC;
556 INIT_LIST_HEAD(&jra->jra_link);
558 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
563 * Allocate and prepare RPC for a new change.
565 * The function allocates and initializes an RPC which will be sent soon to
566 * apply the change to the target OST. The request is initialized from the
567 * llog record passed. Notice only the fields common to all type of changes
570 * \param[in] d OSP device
571 * \param[in] llh llog handle where the record is stored
572 * \param[in] h llog record
573 * \param[in] op type of the change
574 * \param[in] format request format to be used
576 * \retval pointer new request on success
577 * \retval ERR_PTR(errno) on error
579 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
580 struct llog_handle *llh,
581 struct llog_rec_hdr *h,
583 const struct req_format *format)
585 struct ptlrpc_request *req;
586 struct ost_body *body;
587 struct obd_import *imp;
590 /* Prepare the request */
591 imp = d->opd_obd->u.cli.cl_import;
593 req = ptlrpc_request_alloc(imp, format);
595 RETURN(ERR_PTR(-ENOMEM));
597 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
599 ptlrpc_req_finished(req);
604 * this is a trick: to save on memory allocations we put cookie
605 * into the request, but don't set corresponded flag in o_valid
606 * so that OST doesn't interpret this cookie. once the request
607 * is committed on OST we take cookie from the request and cancel
609 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
611 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
612 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
613 body->oa.o_lcookie.lgc_index = h->lrh_index;
615 req->rq_interpret_reply = osp_sync_interpret;
616 req->rq_commit_cb = osp_sync_request_commit_cb;
619 ptlrpc_request_set_replen(req);
625 * Generate a request for setattr change.
627 * The function prepares a new RPC, initializes it with setattr specific
628 * bits and send the RPC.
630 * \param[in] d OSP device
631 * \param[in] llh llog handle where the record is stored
632 * \param[in] h llog record
634 * \retval 0 on success
635 * \retval negative negated errno on error
637 static int osp_sync_new_setattr_job(struct osp_device *d,
638 struct llog_handle *llh,
639 struct llog_rec_hdr *h)
641 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
642 struct ptlrpc_request *req;
643 struct ost_body *body;
646 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
648 /* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
649 * so no bits other than these should be set. */
650 if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
651 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
652 d->opd_obd->obd_name, rec->lsr_valid);
653 /* return 0 so that sync thread can continue processing
658 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
660 RETURN(PTR_ERR(req));
662 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
664 body->oa.o_oi = rec->lsr_oi;
665 body->oa.o_uid = rec->lsr_uid;
666 body->oa.o_gid = rec->lsr_gid;
667 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
668 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
669 * we assume that both UID and GID are valid in that case. */
670 if (rec->lsr_valid == 0)
671 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
673 body->oa.o_valid |= rec->lsr_valid;
675 osp_sync_send_new_rpc(d, req);
680 * Generate a request for unlink change.
682 * The function prepares a new RPC, initializes it with unlink(destroy)
683 * specific bits and sends the RPC. The function is used to handle
684 * llog_unlink_rec which were used in the older versions of Lustre.
685 * Current version uses llog_unlink_rec64.
687 * \param[in] d OSP device
688 * \param[in] llh llog handle where the record is stored
689 * \param[in] h llog record
691 * \retval 0 on success
692 * \retval negative negated errno on error
694 static int osp_sync_new_unlink_job(struct osp_device *d,
695 struct llog_handle *llh,
696 struct llog_rec_hdr *h)
698 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
699 struct ptlrpc_request *req;
700 struct ost_body *body;
703 LASSERT(h->lrh_type == MDS_UNLINK_REC);
705 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
707 RETURN(PTR_ERR(req));
709 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
711 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
712 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
713 body->oa.o_misc = rec->lur_count;
714 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
716 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
718 osp_sync_send_new_rpc(d, req);
723 * Generate a request for unlink change.
725 * The function prepares a new RPC, initializes it with unlink(destroy)
726 * specific bits and sends the RPC. Depending on the target (MDT or OST)
727 * two different protocols are used. For MDT we use OUT (basically OSD API
728 * updates transferred via a network). For OST we still use the old
729 * protocol (OBD?), originally for compatibility. Later we can start to
730 * use OUT for OST as well, this will allow batching and better code
733 * \param[in] env LU environment provided by the caller
734 * \param[in] d OSP device
735 * \param[in] llh llog handle where the record is stored
736 * \param[in] h llog record
738 * \retval 0 on success
739 * \retval negative negated errno on error
741 static int osp_sync_new_unlink64_job(const struct lu_env *env,
742 struct osp_device *d,
743 struct llog_handle *llh,
744 struct llog_rec_hdr *h)
746 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
747 struct ptlrpc_request *req = NULL;
748 struct ost_body *body;
752 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
753 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
756 RETURN(PTR_ERR(req));
758 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
761 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
764 body->oa.o_misc = rec->lur_count;
765 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
767 osp_sync_send_new_rpc(d, req);
772 * Process llog records.
774 * This function is called to process the llog records committed locally.
775 * In the recovery model used by OSP we can apply a change to a remote
776 * target once corresponding transaction (like posix unlink) is committed
777 * locally so can't revert.
778 * Depending on the llog record type, a given handler is called that is
779 * responsible for preparing and sending the RPC to apply the change.
780 * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
782 * \param[in] env LU environment provided by the caller
783 * \param[in] d OSP device
784 * \param[in] llh llog handle where the record is stored
785 * \param[in] rec llog record
787 * \retval 0 on success
788 * \retval negative negated errno on error
790 static int osp_sync_process_record(const struct lu_env *env,
791 struct osp_device *d,
792 struct llog_handle *llh,
793 struct llog_rec_hdr *rec)
795 struct llog_cookie cookie;
798 cookie.lgc_lgl = llh->lgh_id;
799 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
800 cookie.lgc_index = rec->lrh_index;
802 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
803 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
805 /* we're waiting for the record generated by this instance */
806 LASSERT(d->opd_syn_prev_done == 0);
807 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
808 sizeof(gen->lgr_gen))) {
809 CDEBUG(D_HA, "processed all old entries\n");
810 d->opd_syn_prev_done = 1;
813 /* cancel any generation record */
814 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
821 * now we prepare and fill requests to OST, put them on the queue
822 * and fire after next commit callback
825 /* notice we increment counters before sending RPC, to be consistent
826 * in RPC interpret callback which may happen very quickly */
827 spin_lock(&d->opd_syn_lock);
828 d->opd_syn_rpc_in_flight++;
829 d->opd_syn_rpc_in_progress++;
830 spin_unlock(&d->opd_syn_lock);
832 switch (rec->lrh_type) {
833 /* case MDS_UNLINK_REC is kept for compatibility */
835 rc = osp_sync_new_unlink_job(d, llh, rec);
837 case MDS_UNLINK64_REC:
838 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
840 case MDS_SETATTR64_REC:
841 rc = osp_sync_new_setattr_job(d, llh, rec);
844 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
846 /* we should continue processing */
849 /* rc > 0 means sync RPC being added to the queue */
850 if (likely(rc > 0)) {
851 spin_lock(&d->opd_syn_lock);
852 if (d->opd_syn_prev_done) {
853 LASSERT(d->opd_syn_changes > 0);
854 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
856 * NOTE: it's possible to meet same id if
857 * OST stores few stripes of same file
859 if (rec->lrh_id > d->opd_syn_last_processed_id) {
860 d->opd_syn_last_processed_id = rec->lrh_id;
861 wake_up(&d->opd_syn_barrier_waitq);
864 d->opd_syn_changes--;
866 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
867 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
868 d->opd_syn_rpc_in_progress);
869 spin_unlock(&d->opd_syn_lock);
872 spin_lock(&d->opd_syn_lock);
873 d->opd_syn_rpc_in_flight--;
874 d->opd_syn_rpc_in_progress--;
875 spin_unlock(&d->opd_syn_lock);
878 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
879 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
884 * Cancel llog records for the committed changes.
886 * The function walks through the list of the committed RPCs and cancels
887 * corresponding llog records. see osp_sync_request_commit_cb() for the
890 * \param[in] env LU environment provided by the caller
891 * \param[in] d OSP device
893 static void osp_sync_process_committed(const struct lu_env *env,
894 struct osp_device *d)
896 struct obd_device *obd = d->opd_obd;
897 struct obd_import *imp = obd->u.cli.cl_import;
898 struct ost_body *body;
899 struct ptlrpc_request *req;
900 struct llog_ctxt *ctxt;
901 struct llog_handle *llh;
902 struct list_head list;
907 if (list_empty(&d->opd_syn_committed_there))
911 * if current status is -ENOSPC (lack of free space on OST)
912 * then we should poll OST immediately once object destroy
914 * notice: we do this upon commit as well because some backends
915 * (like DMU) do not release space right away.
917 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
918 osp_statfs_need_now(d);
921 * now cancel them all
922 * XXX: can we improve this using some batching?
923 * with batch RPC that'll happen automatically?
924 * XXX: can we store ctxt in lod_device and save few cycles ?
926 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
929 llh = ctxt->loc_handle;
932 INIT_LIST_HEAD(&list);
933 spin_lock(&d->opd_syn_lock);
934 list_splice(&d->opd_syn_committed_there, &list);
935 INIT_LIST_HEAD(&d->opd_syn_committed_there);
936 spin_unlock(&d->opd_syn_lock);
938 while (!list_empty(&list)) {
939 struct llog_cookie *lcookie = NULL;
940 struct osp_job_req_args *jra;
942 jra = list_entry(list.next, struct osp_job_req_args, jra_link);
943 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
944 list_del_init(&jra->jra_link);
946 req = container_of((void *)jra, struct ptlrpc_request,
948 body = req_capsule_client_get(&req->rq_pill,
951 lcookie = &body->oa.o_lcookie;
952 /* import can be closing, thus all commit cb's are
953 * called we can check committness directly */
954 if (req->rq_transno <= imp->imp_peer_committed_transno) {
955 rc = llog_cat_cancel_records(env, llh, 1, lcookie);
957 CERROR("%s: can't cancel record: %d\n",
960 DEBUG_REQ(D_HA, req, "not committed");
963 ptlrpc_req_finished(req);
969 LASSERT(d->opd_syn_rpc_in_progress >= done);
970 spin_lock(&d->opd_syn_lock);
971 d->opd_syn_rpc_in_progress -= done;
972 spin_unlock(&d->opd_syn_lock);
973 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
974 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
975 d->opd_syn_rpc_in_progress);
977 osp_sync_check_for_work(d);
979 /* wake up the thread if requested to stop:
980 * it might be waiting for in-progress to complete */
981 if (unlikely(osp_sync_running(d) == 0))
982 wake_up(&d->opd_syn_waitq);
988 * The core of the syncing mechanism.
990 * This is a callback called by the llog processing function. Essentially it
991 * suspends llog processing until there is a record to process (it's supposed
992 * to be committed locally). The function handles RPCs committed by the target
993 * and cancels corresponding llog records.
995 * \param[in] env LU environment provided by the caller
996 * \param[in] llh llog handle we're processing
997 * \param[in] rec current llog record
998 * \param[in] data callback data containing a pointer to the device
1000 * \retval 0 to ask the caller (llog_process()) to continue
1001 * \retval LLOG_PROC_BREAK to ask the caller to break
1003 static int osp_sync_process_queues(const struct lu_env *env,
1004 struct llog_handle *llh,
1005 struct llog_rec_hdr *rec,
1008 struct osp_device *d = data;
1012 struct l_wait_info lwi = { 0 };
1014 if (!osp_sync_running(d)) {
1015 CDEBUG(D_HA, "stop llog processing\n");
1016 return LLOG_PROC_BREAK;
1019 /* process requests committed by OST */
1020 osp_sync_process_committed(env, d);
1022 /* if we there are changes to be processed and we have
1023 * resources for this ... do now */
1024 if (osp_sync_can_process_new(d, rec)) {
1026 /* ask llog for another record */
1027 CDEBUG(D_HA, "%lu changes, %u in progress,"
1030 d->opd_syn_rpc_in_progress,
1031 d->opd_syn_rpc_in_flight);
1036 * try to send, in case of disconnection, suspend
1037 * processing till we can send this request
1040 rc = osp_sync_process_record(env, d, llh, rec);
1042 * XXX: probably different handling is needed
1043 * for some bugs, like immediate exit or if
1047 CERROR("can't send: %d\n", rc);
1048 l_wait_event(d->opd_syn_waitq,
1049 !osp_sync_running(d) ||
1050 osp_sync_has_work(d),
1053 } while (rc != 0 && osp_sync_running(d));
1059 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
1060 osp_sync_remove_from_tracker(d);
1062 l_wait_event(d->opd_syn_waitq,
1063 !osp_sync_running(d) ||
1064 osp_sync_can_process_new(d, rec) ||
1065 !list_empty(&d->opd_syn_committed_there),
1073 * This thread runs llog_cat_process() scanner calling our callback
1074 * to process llog records. in the callback we implement tricky
1075 * state machine as we don't want to start scanning of the llog again
1076 * and again, also we don't want to process too many records and send
1077 * too many RPCs a time. so, depending on current load (num of changes
1078 * being synced to OST) the callback can suspend awaiting for some
1079 * new conditions, like syncs completed.
1081 * In order to process llog records left by previous boots and to allow
1082 * llog_process_thread() to find something (otherwise it'd just exit
1083 * immediately) we add a special GENERATATION record on each boot.
1085 * \param[in] _arg a pointer to thread's arguments
1087 * \retval 0 on success
1088 * \retval negative negated errno on error
1090 static int osp_sync_thread(void *_arg)
1092 struct osp_device *d = _arg;
1093 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1094 struct l_wait_info lwi = { 0 };
1095 struct llog_ctxt *ctxt;
1096 struct obd_device *obd = d->opd_obd;
1097 struct llog_handle *llh;
1103 rc = lu_env_init(&env, LCT_LOCAL);
1105 CERROR("%s: can't initialize env: rc = %d\n",
1110 spin_lock(&d->opd_syn_lock);
1111 thread->t_flags = SVC_RUNNING;
1112 spin_unlock(&d->opd_syn_lock);
1113 wake_up(&thread->t_ctl_waitq);
1115 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1117 CERROR("can't get appropriate context\n");
1118 GOTO(out, rc = -EINVAL);
1121 llh = ctxt->loc_handle;
1123 CERROR("can't get llh\n");
1124 llog_ctxt_put(ctxt);
1125 GOTO(out, rc = -EINVAL);
1128 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
1129 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1130 "%lu changes, %u in progress, %u in flight: %d\n",
1131 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1132 d->opd_syn_rpc_in_flight, rc);
1134 /* we don't expect llog_process_thread() to exit till umount */
1135 LASSERTF(thread->t_flags != SVC_RUNNING,
1136 "%lu changes, %u in progress, %u in flight\n",
1137 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1138 d->opd_syn_rpc_in_flight);
1140 /* wait till all the requests are completed */
1142 while (d->opd_syn_rpc_in_progress > 0) {
1143 osp_sync_process_committed(&env, d);
1145 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1146 rc = l_wait_event(d->opd_syn_waitq,
1147 d->opd_syn_rpc_in_progress == 0,
1149 if (rc == -ETIMEDOUT)
1151 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1152 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1153 d->opd_syn_rpc_in_flight,
1154 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1158 llog_cat_close(&env, llh);
1159 rc = llog_cleanup(&env, ctxt);
1161 CERROR("can't cleanup llog: %d\n", rc);
1163 LASSERTF(d->opd_syn_rpc_in_progress == 0,
1164 "%s: %d %d %sempty\n",
1165 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1166 d->opd_syn_rpc_in_flight,
1167 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1169 thread->t_flags = SVC_STOPPED;
1171 wake_up(&thread->t_ctl_waitq);
1181 * Initializes the llog. Specific llog to be used depends on the type of the
1182 * target OSP represents (OST or MDT). The function adds appends a new llog
1183 * record to mark the place where the records associated with this boot
1186 * \param[in] env LU environment provided by the caller
1187 * \param[in] d OSP device
1189 * \retval 0 on success
1190 * \retval negative negated errno on error
1192 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1194 struct osp_thread_info *osi = osp_env_info(env);
1195 struct lu_fid *fid = &osi->osi_fid;
1196 struct llog_handle *lgh = NULL;
1197 struct obd_device *obd = d->opd_obd;
1198 struct llog_ctxt *ctxt;
1206 * open llog corresponding to our OST
1208 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1209 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1211 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1213 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1214 &osi->osi_cid, fid);
1216 if (rc != -EFAULT) {
1217 CERROR("%s: can't get id from catalogs: rc = %d\n",
1222 /* After sparse OST indices is supported, the CATALOG file
1223 * may become a sparse file that results in failure on
1224 * reading. Skip this error as the llog will be created
1226 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1230 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1231 obd->obd_name, d->opd_index,
1232 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1233 osi->osi_cid.lci_logid.lgl_ogen);
1235 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1236 &osp_mds_ost_orig_logops);
1240 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1243 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1244 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1246 /* re-create llog if it is missing */
1248 logid_set_id(&osi->osi_cid.lci_logid, 0);
1250 GOTO(out_cleanup, rc);
1253 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1254 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1256 GOTO(out_cleanup, rc);
1257 osi->osi_cid.lci_logid = lgh->lgh_id;
1260 LASSERT(lgh != NULL);
1261 ctxt->loc_handle = lgh;
1263 rc = llog_cat_init_and_process(env, lgh);
1265 GOTO(out_close, rc);
1267 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1268 &osi->osi_cid, fid);
1270 GOTO(out_close, rc);
1273 * put a mark in the llog till which we'll be processing
1274 * old records restless
1276 d->opd_syn_generation.mnt_cnt = cfs_time_current();
1277 d->opd_syn_generation.conn_cnt = cfs_time_current();
1279 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1280 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1282 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1283 sizeof(osi->osi_gen.lgr_gen));
1285 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1287 GOTO(out_close, rc);
1288 llog_ctxt_put(ctxt);
1291 llog_cat_close(env, lgh);
1293 llog_cleanup(env, ctxt);
1298 * Cleanup llog used for syncing.
1300 * Closes and cleanups the llog. The function is called when the device is
1303 * \param[in] env LU environment provided by the caller
1304 * \param[in] d OSP device
1306 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1308 struct llog_ctxt *ctxt;
1310 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1312 llog_cat_close(env, ctxt->loc_handle);
1313 llog_cleanup(env, ctxt);
1317 * Initialization of the sync component of OSP.
1319 * Initializes the llog and starts a new thread to handle the changes to
1320 * the remote target (OST or MDT).
1322 * \param[in] env LU environment provided by the caller
1323 * \param[in] d OSP device
1325 * \retval 0 on success
1326 * \retval negative negated errno on error
1328 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1330 struct l_wait_info lwi = { 0 };
1331 struct task_struct *task;
1336 rc = osp_sync_id_traction_init(d);
1341 * initialize llog storing changes
1343 rc = osp_sync_llog_init(env, d);
1345 CERROR("%s: can't initialize llog: rc = %d\n",
1346 d->opd_obd->obd_name, rc);
1351 * Start synchronization thread
1353 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1354 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1355 spin_lock_init(&d->opd_syn_lock);
1356 init_waitqueue_head(&d->opd_syn_waitq);
1357 init_waitqueue_head(&d->opd_syn_barrier_waitq);
1358 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1359 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1361 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1362 d->opd_index, d->opd_group);
1365 CERROR("%s: cannot start sync thread: rc = %d\n",
1366 d->opd_obd->obd_name, rc);
1370 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1371 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1375 osp_sync_llog_fini(env, d);
1377 osp_sync_id_traction_fini(d);
1382 * Stop the syncing thread.
1384 * Asks the syncing thread to stop and wait until it's stopped.
1386 * \param[in] d OSP device
1390 int osp_sync_fini(struct osp_device *d)
1392 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1396 thread->t_flags = SVC_STOPPING;
1397 wake_up(&d->opd_syn_waitq);
1398 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1401 * unregister transaction callbacks only when sync thread
1402 * has finished operations with llog
1404 osp_sync_id_traction_fini(d);
1409 static DEFINE_MUTEX(osp_id_tracker_sem);
1410 static struct list_head osp_id_tracker_list =
1411 LIST_HEAD_INIT(osp_id_tracker_list);
1414 * OSD commit callback.
1416 * The function is used as a local OSD commit callback to track the highest
1417 * committed llog record id. see osp_sync_id_traction_init() for the details.
1419 * \param[in] th local transaction handle committed
1420 * \param[in] cookie commit callback data (our private structure)
1422 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1424 struct osp_id_tracker *tr = cookie;
1425 struct osp_device *d;
1426 struct osp_txn_info *txn;
1430 txn = osp_txn_info(&th->th_ctx);
1431 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1434 spin_lock(&tr->otr_lock);
1435 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1436 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1437 tr->otr_committed_id, txn->oti_current_id);
1438 tr->otr_committed_id = txn->oti_current_id;
1440 list_for_each_entry(d, &tr->otr_wakeup_list,
1442 d->opd_syn_last_committed_id = tr->otr_committed_id;
1443 wake_up(&d->opd_syn_waitq);
1446 spin_unlock(&tr->otr_lock);
1450 * Initialize commit tracking mechanism.
1452 * Some setups may have thousands of OSTs and each will be represented by OSP.
1453 * Meaning order of magnitute many more changes to apply every second. In order
1454 * to keep the number of commit callbacks low this mechanism was introduced.
1455 * The mechanism is very similar to transno used by MDT service: it's an single
1456 * ID stream which can be assigned by any OSP to its llog records. The tricky
1457 * part is that ID is stored in per-transaction data and re-used by all the OSPs
1458 * involved in that transaction. Then all these OSPs are woken up utilizing a single OSD commit callback.
1460 * The function initializes the data used by the tracker described above.
1461 * A singler tracker per OSD device is created.
1463 * \param[in] d OSP device
1465 * \retval 0 on success
1466 * \retval negative negated errno on error
1468 static int osp_sync_id_traction_init(struct osp_device *d)
1470 struct osp_id_tracker *tr, *found = NULL;
1474 LASSERT(d->opd_storage);
1475 LASSERT(d->opd_syn_tracker == NULL);
1476 INIT_LIST_HEAD(&d->opd_syn_ontrack);
1478 mutex_lock(&osp_id_tracker_sem);
1479 list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1480 if (tr->otr_dev == d->opd_storage) {
1481 LASSERT(atomic_read(&tr->otr_refcount));
1482 atomic_inc(&tr->otr_refcount);
1483 d->opd_syn_tracker = tr;
1489 if (found == NULL) {
1493 d->opd_syn_tracker = tr;
1494 spin_lock_init(&tr->otr_lock);
1495 tr->otr_dev = d->opd_storage;
1496 tr->otr_next_id = 1;
1497 tr->otr_committed_id = 0;
1498 atomic_set(&tr->otr_refcount, 1);
1499 INIT_LIST_HEAD(&tr->otr_wakeup_list);
1500 list_add(&tr->otr_list, &osp_id_tracker_list);
1501 tr->otr_tx_cb.dtc_txn_commit =
1502 osp_sync_tracker_commit_cb;
1503 tr->otr_tx_cb.dtc_cookie = tr;
1504 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1505 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1509 mutex_unlock(&osp_id_tracker_sem);
1515 * Release commit tracker.
1517 * Decrease a refcounter on the tracker used by the given OSP device \a d.
1518 * If no more users left, then the tracker is released.
1520 * \param[in] d OSP device
1522 static void osp_sync_id_traction_fini(struct osp_device *d)
1524 struct osp_id_tracker *tr;
1529 tr = d->opd_syn_tracker;
1535 osp_sync_remove_from_tracker(d);
1537 mutex_lock(&osp_id_tracker_sem);
1538 if (atomic_dec_and_test(&tr->otr_refcount)) {
1539 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1540 LASSERT(list_empty(&tr->otr_wakeup_list));
1541 list_del(&tr->otr_list);
1543 d->opd_syn_tracker = NULL;
1545 mutex_unlock(&osp_id_tracker_sem);
1551 * Generate a new ID on a tracker.
1553 * Generates a new ID using the tracker associated with the given OSP device
1554 * \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
1555 * the wakeup list, so OSP won't miss when a transaction using the ID is
1556 * committed. Notice ID is 32bit, but llog doesn't support >2^32 records anyway.
1558 * \param[in] d OSP device
1559 * \param[in] id 0 or ID generated previously
1561 * \retval ID the caller should use
1563 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1565 struct osp_id_tracker *tr;
1567 tr = d->opd_syn_tracker;
1570 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1571 spin_lock(&tr->otr_lock);
1572 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1573 spin_unlock(&tr->otr_lock);
1574 CERROR("%s: next %u, last synced %lu\n",
1575 d->opd_obd->obd_name, tr->otr_next_id,
1576 d->opd_syn_last_used_id);
1581 id = tr->otr_next_id++;
1582 if (id > d->opd_syn_last_used_id)
1583 d->opd_syn_last_used_id = id;
1584 if (list_empty(&d->opd_syn_ontrack))
1585 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1586 spin_unlock(&tr->otr_lock);
1587 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1593 * Stop to propagate commit status to OSP.
1595 * If the OSP does not have any llog records she's waiting to commit, then
1596 * it is possible to unsubscribe from wakeups from the tracking using this
1599 * \param[in] d OSP device not willing to wakeup
1601 static void osp_sync_remove_from_tracker(struct osp_device *d)
1603 struct osp_id_tracker *tr;
1605 tr = d->opd_syn_tracker;
1608 if (list_empty(&d->opd_syn_ontrack))
1611 spin_lock(&tr->otr_lock);
1612 list_del_init(&d->opd_syn_ontrack);
1613 spin_unlock(&tr->otr_lock);