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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osp/osp_sync.c
34 * Lustre OST Proxy Device
36 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
37 * Author: Mikhail Pershin <mike.pershin@intel.com>
40 #define DEBUG_SUBSYSTEM S_MDS
42 #include <linux/kthread.h>
43 #include <lustre_log.h>
44 #include <lustre_update.h>
45 #include "osp_internal.h"
47 static int osp_sync_id_traction_init(struct osp_device *d);
48 static void osp_sync_id_traction_fini(struct osp_device *d);
49 static __u64 osp_sync_id_get(struct osp_device *d, __u64 id);
50 static void osp_sync_remove_from_tracker(struct osp_device *d);
53 * this is a components of OSP implementing synchronization between MDS and OST
54 * it llogs all interesting changes (currently it's uig/gid change and object
55 * destroy) atomically, then makes sure changes hit OST storage
57 * we have 4 queues of work:
59 * the first queue is llog itself, once read a change is stored in 2nd queue
60 * in form of RPC (but RPC isn't fired yet).
62 * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
63 * commit. once change is committed locally it migrates onto 3rd queue.
65 * the third queue (opd_syn_committed_here) holds changes committed locally,
66 * but not sent to OST (as the pipe can be full). once pipe becomes non-full
67 * we take a change from the queue and fire corresponded RPC.
69 * once RPC is reported committed by OST (using regular last_committed mech.)
70 * the change jumps into 4th queue (opd_syn_committed_there), now we can
71 * cancel corresponded llog record and release RPC
73 * opd_syn_changes is a number of unread llog records (to be processed).
74 * notice this number doesn't include llog records from previous boots.
75 * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
77 * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
78 * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
79 * too much memory -- how to deal with 1000th OSTs ? batching could help?
81 * opd_syn_rpc_in_flight is a number of RPC in flight.
82 * we control this with OSP_MAX_IN_FLIGHT
85 /* XXX: do math to learn reasonable threshold
86 * should it be ~ number of changes fitting bulk? */
88 #define OSP_SYN_THRESHOLD 10
89 #define OSP_MAX_IN_FLIGHT 8
90 #define OSP_MAX_IN_PROGRESS 4096
92 #define OSP_JOB_MAGIC 0x26112005
94 struct osp_job_req_args {
95 /** bytes reserved for ptlrpc_replay_req() */
96 struct ptlrpc_replay_async_args jra_raa;
97 struct list_head jra_committed_link;
98 struct list_head jra_inflight_link;
99 struct llog_cookie jra_lcookie;
103 static inline int osp_sync_running(struct osp_device *d)
105 return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
109 * Check status: whether OSP thread has stopped
111 * \param[in] d OSP device
113 * \retval 0 still running
116 static inline int osp_sync_stopped(struct osp_device *d)
118 return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
122 ** Check for new changes to sync
124 * \param[in] d OSP device
126 * \retval 1 there are changes
127 * \retval 0 there are no changes
129 static inline int osp_sync_has_new_job(struct osp_device *d)
131 return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
132 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
133 || (d->opd_syn_prev_done == 0);
136 static inline int osp_sync_inflight_conflict(struct osp_device *d,
137 struct llog_rec_hdr *h)
139 struct osp_job_req_args *jra;
143 if (h == NULL || h->lrh_type == LLOG_GEN_REC ||
144 list_empty(&d->opd_syn_inflight_list))
147 memset(&ostid, 0, sizeof(ostid));
148 switch (h->lrh_type) {
150 ostid_set_seq(&ostid, ((struct llog_unlink_rec *)h)->lur_oseq);
151 ostid_set_id(&ostid, ((struct llog_unlink_rec *)h)->lur_oid);
153 case MDS_UNLINK64_REC:
154 fid_to_ostid(&((struct llog_unlink64_rec *)h)->lur_fid, &ostid);
156 case MDS_SETATTR64_REC:
157 ostid = ((struct llog_setattr64_rec *)h)->lsr_oi;
163 spin_lock(&d->opd_syn_lock);
164 list_for_each_entry(jra, &d->opd_syn_inflight_list, jra_inflight_link) {
165 struct ptlrpc_request *req;
166 struct ost_body *body;
168 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
170 req = container_of((void *)jra, struct ptlrpc_request,
172 body = req_capsule_client_get(&req->rq_pill,
176 if (memcmp(&ostid, &body->oa.o_oi, sizeof(ostid)) == 0) {
181 spin_unlock(&d->opd_syn_lock);
186 static inline int osp_sync_low_in_progress(struct osp_device *d)
188 return atomic_read(&d->opd_syn_rpc_in_progress) <
189 d->opd_syn_max_rpc_in_progress;
193 * Check for room in the network pipe to OST
195 * \param[in] d OSP device
197 * \retval 1 there is room
198 * \retval 0 no room, the pipe is full
200 static inline int osp_sync_low_in_flight(struct osp_device *d)
202 return atomic_read(&d->opd_syn_rpc_in_flight) <
203 d->opd_syn_max_rpc_in_flight;
207 * Wake up check for the main sync thread
209 * \param[in] d OSP device
211 * \retval 1 time to wake up
212 * \retval 0 no need to wake up
214 static inline int osp_sync_has_work(struct osp_device *d)
216 /* has new/old changes and low in-progress? */
217 if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
218 osp_sync_low_in_flight(d) && d->opd_imp_connected)
221 /* has remotely committed? */
222 if (!list_empty(&d->opd_syn_committed_there))
228 #define osp_sync_check_for_work(d) \
230 if (osp_sync_has_work(d)) { \
231 wake_up(&d->opd_syn_waitq); \
235 void __osp_sync_check_for_work(struct osp_device *d)
237 osp_sync_check_for_work(d);
240 static inline __u64 osp_sync_correct_id(struct osp_device *d,
241 struct llog_rec_hdr *rec)
244 * llog use cyclic store with 32 bit lrh_id
245 * so overflow lrh_id is possible. Range between
246 * last_processed and last_committed is less than
247 * 64745 ^ 2 and less than 2^32 - 1
249 __u64 correct_id = d->opd_syn_last_committed_id;
251 if ((correct_id & 0xffffffffULL) < rec->lrh_id)
252 correct_id -= 0x100000000ULL;
254 correct_id &= ~0xffffffffULL;
255 correct_id |= rec->lrh_id;
260 * Check and return ready-for-new status.
262 * The thread processing llog record uses this function to check whether
263 * it's time to take another record and process it. The number of conditions
264 * must be met: the connection should be ready, RPCs in flight not exceeding
265 * the limit, the record is committed locally, etc (see the lines below).
267 * \param[in] d OSP device
268 * \param[in] rec next llog record to process
270 * \retval 0 not ready
273 static inline int osp_sync_can_process_new(struct osp_device *d,
274 struct llog_rec_hdr *rec)
278 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
280 if (unlikely(osp_sync_inflight_conflict(d, rec)))
282 if (!osp_sync_low_in_progress(d))
284 if (!osp_sync_low_in_flight(d))
286 if (!d->opd_imp_connected)
288 if (d->opd_syn_prev_done == 0)
290 if (atomic_read(&d->opd_syn_changes) == 0)
293 osp_sync_correct_id(d, rec) <= d->opd_syn_last_committed_id)
299 * Declare intention to add a new change.
301 * With regard to OSD API, we have to declare any changes ahead. In this
302 * case we declare an intention to add a llog record representing the
303 * change on the local storage.
305 * \param[in] env LU environment provided by the caller
306 * \param[in] o OSP object
307 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
308 * \param[in] th transaction handle (local)
310 * \retval 0 on success
311 * \retval negative negated errno on error
313 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
314 llog_op_type type, struct thandle *th)
316 struct osp_thread_info *osi = osp_env_info(env);
317 struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
318 struct llog_ctxt *ctxt;
319 struct thandle *storage_th;
324 /* it's a layering violation, to access internals of th,
325 * but we can do this as a sanity check, for a while */
326 LASSERT(th->th_top != NULL);
327 storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
328 if (IS_ERR(storage_th))
329 RETURN(PTR_ERR(storage_th));
332 case MDS_UNLINK64_REC:
333 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
335 case MDS_SETATTR64_REC:
336 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
342 /* we want ->dt_trans_start() to allocate per-thandle structure */
343 storage_th->th_tags |= LCT_OSP_THREAD;
345 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
348 rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
356 * Generate a llog record for a given change.
358 * Generates a llog record for the change passed. The change can be of two
359 * types: unlink and setattr. The record gets an ID which later will be
360 * used to track commit status of the change. For unlink changes, the caller
361 * can supply a starting FID and the count of the objects to destroy. For
362 * setattr the caller should apply attributes to apply.
365 * \param[in] env LU environment provided by the caller
366 * \param[in] d OSP device
367 * \param[in] fid fid of the object the change should be applied to
368 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
369 * \param[in] count count of objects to destroy
370 * \param[in] th transaction handle (local)
371 * \param[in] attr attributes for setattr
373 * \retval 0 on success
374 * \retval negative negated errno on error
376 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
377 const struct lu_fid *fid, llog_op_type type,
378 int count, struct thandle *th,
379 const struct lu_attr *attr)
381 struct osp_thread_info *osi = osp_env_info(env);
382 struct llog_ctxt *ctxt;
383 struct osp_txn_info *txn;
384 struct thandle *storage_th;
389 /* it's a layering violation, to access internals of th,
390 * but we can do this as a sanity check, for a while */
391 LASSERT(th->th_top != NULL);
392 storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
393 if (IS_ERR(storage_th))
394 RETURN(PTR_ERR(storage_th));
397 case MDS_UNLINK64_REC:
398 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
399 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
400 osi->osi_unlink.lur_fid = *fid;
401 osi->osi_unlink.lur_count = count;
403 case MDS_SETATTR64_REC:
404 rc = fid_to_ostid(fid, &osi->osi_oi);
406 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
407 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
408 osi->osi_setattr.lsr_oi = osi->osi_oi;
410 osi->osi_setattr.lsr_uid = attr->la_uid;
411 osi->osi_setattr.lsr_gid = attr->la_gid;
412 osi->osi_setattr.lsr_valid =
413 ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
414 ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0);
420 txn = osp_txn_info(&storage_th->th_ctx);
423 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
424 osi->osi_hdr.lrh_id = (txn->oti_current_id & 0xffffffffULL);
425 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
429 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
433 if (likely(rc >= 0)) {
434 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
435 d->opd_obd->obd_name,
436 POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
437 (unsigned long)osi->osi_cookie.lgc_lgl.lgl_ogen,
438 (unsigned long)osi->osi_cookie.lgc_index, rc);
439 atomic_inc(&d->opd_syn_changes);
441 /* return 0 always here, error case just cause no llog record */
445 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
446 llog_op_type type, struct thandle *th,
447 const struct lu_attr *attr)
449 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
450 lu_object_fid(&o->opo_obj.do_lu), type, 1,
454 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
455 struct lu_fid *fid, int lost, struct thandle *th)
457 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
461 * it's quite obvious we can't maintain all the structures in the memory:
462 * while OST is down, MDS can be processing thousands and thousands of unlinks
463 * filling persistent llogs and in-core respresentation
465 * this doesn't scale at all. so we need basically the following:
466 * a) destroy/setattr append llog records
467 * b) once llog has grown to X records, we process first Y committed records
469 * once record R is found via llog_process(), it becomes committed after any
470 * subsequent commit callback (at the most)
474 * ptlrpc commit callback.
476 * The callback is called by PTLRPC when a RPC is reported committed by the
477 * target (OST). We register the callback for the every RPC applying a change
478 * from the llog. This way we know then the llog records can be cancelled.
479 * Notice the callback can be called when OSP is finishing. We can detect this
480 * checking that actual transno in the request is less or equal of known
481 * committed transno (see osp_sync_process_committed() for the details).
482 * XXX: this is pretty expensive and can be improved later using batching.
484 * \param[in] req request
486 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
488 struct osp_device *d = req->rq_cb_data;
489 struct osp_job_req_args *jra;
491 CDEBUG(D_HA, "commit req %p, transno %llu\n", req, req->rq_transno);
493 if (unlikely(req->rq_transno == 0))
496 /* do not do any opd_dyn_rpc_* accounting here
497 * it's done in osp_sync_interpret sooner or later */
500 jra = ptlrpc_req_async_args(req);
501 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
502 LASSERT(list_empty(&jra->jra_committed_link));
504 ptlrpc_request_addref(req);
506 spin_lock(&d->opd_syn_lock);
507 list_add(&jra->jra_committed_link, &d->opd_syn_committed_there);
508 spin_unlock(&d->opd_syn_lock);
510 /* XXX: some batching wouldn't hurt */
511 wake_up(&d->opd_syn_waitq);
515 * RPC interpretation callback.
517 * The callback is called by ptlrpc when RPC is replied. Now we have to decide
519 * - put request on a special list to wait until it's committed by the target,
520 * if the request is successful
521 * - schedule llog record cancel if no target object is found
522 * - try later (essentially after reboot) in case of unexpected error
524 * \param[in] env LU environment provided by the caller
525 * \param[in] req request replied
526 * \param[in] aa callback data
527 * \param[in] rc result of RPC
531 static int osp_sync_interpret(const struct lu_env *env,
532 struct ptlrpc_request *req, void *aa, int rc)
534 struct osp_device *d = req->rq_cb_data;
535 struct osp_job_req_args *jra = aa;
537 if (jra->jra_magic != OSP_JOB_MAGIC) {
538 DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
543 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
544 atomic_read(&req->rq_refcount),
545 rc, (unsigned) req->rq_transno);
546 LASSERT(rc || req->rq_transno);
550 * we tried to destroy object or update attributes,
551 * but object doesn't exist anymore - cancell llog record
553 LASSERT(req->rq_transno == 0);
554 LASSERT(list_empty(&jra->jra_committed_link));
556 ptlrpc_request_addref(req);
558 spin_lock(&d->opd_syn_lock);
559 list_add(&jra->jra_committed_link, &d->opd_syn_committed_there);
560 spin_unlock(&d->opd_syn_lock);
562 wake_up(&d->opd_syn_waitq);
564 struct obd_import *imp = req->rq_import;
566 * error happened, we'll try to repeat on next boot ?
568 LASSERTF(req->rq_transno == 0 ||
569 req->rq_import_generation < imp->imp_generation,
570 "transno %llu, rc %d, gen: req %d, imp %d\n",
571 req->rq_transno, rc, req->rq_import_generation,
572 imp->imp_generation);
573 if (req->rq_transno == 0) {
574 /* this is the last time we see the request
575 * if transno is not zero, then commit cb
576 * will be called at some point */
577 LASSERT(atomic_read(&d->opd_syn_rpc_in_progress) > 0);
578 atomic_dec(&d->opd_syn_rpc_in_progress);
581 wake_up(&d->opd_syn_waitq);
582 } else if (d->opd_pre != NULL &&
583 unlikely(d->opd_pre_status == -ENOSPC)) {
585 * if current status is -ENOSPC (lack of free space on OST)
586 * then we should poll OST immediately once object destroy
589 osp_statfs_need_now(d);
592 spin_lock(&d->opd_syn_lock);
593 list_del_init(&jra->jra_inflight_link);
594 spin_unlock(&d->opd_syn_lock);
595 LASSERT(atomic_read(&d->opd_syn_rpc_in_flight) > 0);
596 atomic_dec(&d->opd_syn_rpc_in_flight);
597 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
598 wake_up(&d->opd_syn_barrier_waitq);
599 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
600 d->opd_obd->obd_name, atomic_read(&d->opd_syn_rpc_in_flight),
601 atomic_read(&d->opd_syn_rpc_in_progress));
603 osp_sync_check_for_work(d);
609 ** Add request to ptlrpc queue.
611 * This is just a tiny helper function to put the request on the sending list
613 * \param[in] d OSP device
614 * \param[in] llh llog handle where the record is stored
615 * \param[in] h llog record
616 * \param[in] req request
618 static void osp_sync_send_new_rpc(struct osp_device *d,
619 struct llog_handle *llh,
620 struct llog_rec_hdr *h,
621 struct ptlrpc_request *req)
623 struct osp_job_req_args *jra;
625 LASSERT(atomic_read(&d->opd_syn_rpc_in_flight) <=
626 d->opd_syn_max_rpc_in_flight);
628 jra = ptlrpc_req_async_args(req);
629 jra->jra_magic = OSP_JOB_MAGIC;
630 jra->jra_lcookie.lgc_lgl = llh->lgh_id;
631 jra->jra_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
632 jra->jra_lcookie.lgc_index = h->lrh_index;
633 INIT_LIST_HEAD(&jra->jra_committed_link);
634 spin_lock(&d->opd_syn_lock);
635 list_add_tail(&jra->jra_inflight_link, &d->opd_syn_inflight_list);
636 spin_unlock(&d->opd_syn_lock);
638 ptlrpcd_add_req(req);
643 * Allocate and prepare RPC for a new change.
645 * The function allocates and initializes an RPC which will be sent soon to
646 * apply the change to the target OST. The request is initialized from the
647 * llog record passed. Notice only the fields common to all type of changes
650 * \param[in] d OSP device
651 * \param[in] op type of the change
652 * \param[in] format request format to be used
654 * \retval pointer new request on success
655 * \retval ERR_PTR(errno) on error
657 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
659 const struct req_format *format)
661 struct ptlrpc_request *req;
662 struct obd_import *imp;
665 /* Prepare the request */
666 imp = d->opd_obd->u.cli.cl_import;
669 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
670 RETURN(ERR_PTR(-ENOMEM));
672 req = ptlrpc_request_alloc(imp, format);
674 RETURN(ERR_PTR(-ENOMEM));
676 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
678 ptlrpc_req_finished(req);
682 req->rq_interpret_reply = osp_sync_interpret;
683 req->rq_commit_cb = osp_sync_request_commit_cb;
686 ptlrpc_request_set_replen(req);
692 * Generate a request for setattr change.
694 * The function prepares a new RPC, initializes it with setattr specific
695 * bits and send the RPC.
697 * \param[in] d OSP device
698 * \param[in] llh llog handle where the record is stored
699 * \param[in] h llog record
701 * \retval 0 on success
702 * \retval 1 on invalid record
703 * \retval negative negated errno on error
705 static int osp_sync_new_setattr_job(struct osp_device *d,
706 struct llog_handle *llh,
707 struct llog_rec_hdr *h)
709 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
710 struct ptlrpc_request *req;
711 struct ost_body *body;
714 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
716 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
718 /* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
719 * so no bits other than these should be set. */
720 if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
721 CERROR("%s: invalid setattr record, lsr_valid:%llu\n",
722 d->opd_obd->obd_name, rec->lsr_valid);
723 /* return 1 on invalid record */
727 req = osp_sync_new_job(d, OST_SETATTR, &RQF_OST_SETATTR);
729 RETURN(PTR_ERR(req));
731 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
733 body->oa.o_oi = rec->lsr_oi;
734 body->oa.o_uid = rec->lsr_uid;
735 body->oa.o_gid = rec->lsr_gid;
736 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
737 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
738 * we assume that both UID and GID are valid in that case. */
739 if (rec->lsr_valid == 0)
740 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
742 body->oa.o_valid |= rec->lsr_valid;
744 osp_sync_send_new_rpc(d, llh, h, req);
749 * Generate a request for unlink change.
751 * The function prepares a new RPC, initializes it with unlink(destroy)
752 * specific bits and sends the RPC. The function is used to handle
753 * llog_unlink_rec which were used in the older versions of Lustre.
754 * Current version uses llog_unlink_rec64.
756 * \param[in] d OSP device
757 * \param[in] llh llog handle where the record is stored
758 * \param[in] h llog record
760 * \retval 0 on success
761 * \retval negative negated errno on error
763 static int osp_sync_new_unlink_job(struct osp_device *d,
764 struct llog_handle *llh,
765 struct llog_rec_hdr *h)
767 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
768 struct ptlrpc_request *req;
769 struct ost_body *body;
772 LASSERT(h->lrh_type == MDS_UNLINK_REC);
774 req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
776 RETURN(PTR_ERR(req));
778 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
780 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
781 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
782 body->oa.o_misc = rec->lur_count;
783 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
785 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
787 osp_sync_send_new_rpc(d, llh, h, req);
792 * Generate a request for unlink change.
794 * The function prepares a new RPC, initializes it with unlink(destroy)
795 * specific bits and sends the RPC. Depending on the target (MDT or OST)
796 * two different protocols are used. For MDT we use OUT (basically OSD API
797 * updates transferred via a network). For OST we still use the old
798 * protocol (OBD?), originally for compatibility. Later we can start to
799 * use OUT for OST as well, this will allow batching and better code
802 * \param[in] d OSP device
803 * \param[in] llh llog handle where the record is stored
804 * \param[in] h llog record
806 * \retval 0 on success
807 * \retval negative negated errno on error
809 static int osp_sync_new_unlink64_job(struct osp_device *d,
810 struct llog_handle *llh,
811 struct llog_rec_hdr *h)
813 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
814 struct ptlrpc_request *req = NULL;
815 struct ost_body *body;
819 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
820 req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
822 RETURN(PTR_ERR(req));
824 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
827 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
830 body->oa.o_misc = rec->lur_count;
831 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
833 osp_sync_send_new_rpc(d, llh, h, req);
838 * Process llog records.
840 * This function is called to process the llog records committed locally.
841 * In the recovery model used by OSP we can apply a change to a remote
842 * target once corresponding transaction (like posix unlink) is committed
843 * locally so can't revert.
844 * Depending on the llog record type, a given handler is called that is
845 * responsible for preparing and sending the RPC to apply the change.
846 * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
848 * \param[in] env LU environment provided by the caller
849 * \param[in] d OSP device
850 * \param[in] llh llog handle where the record is stored
851 * \param[in] rec llog record
853 static void osp_sync_process_record(const struct lu_env *env,
854 struct osp_device *d,
855 struct llog_handle *llh,
856 struct llog_rec_hdr *rec)
858 struct llog_handle *cathandle = llh->u.phd.phd_cat_handle;
859 struct llog_cookie cookie;
864 cookie.lgc_lgl = llh->lgh_id;
865 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
866 cookie.lgc_index = rec->lrh_index;
868 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
869 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
871 /* we're waiting for the record generated by this instance */
872 LASSERT(d->opd_syn_prev_done == 0);
873 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
874 sizeof(gen->lgr_gen))) {
875 CDEBUG(D_HA, "processed all old entries\n");
876 d->opd_syn_prev_done = 1;
879 /* cancel any generation record */
880 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
886 * now we prepare and fill requests to OST, put them on the queue
887 * and fire after next commit callback
890 /* notice we increment counters before sending RPC, to be consistent
891 * in RPC interpret callback which may happen very quickly */
892 atomic_inc(&d->opd_syn_rpc_in_flight);
893 atomic_inc(&d->opd_syn_rpc_in_progress);
895 switch (rec->lrh_type) {
896 /* case MDS_UNLINK_REC is kept for compatibility */
898 rc = osp_sync_new_unlink_job(d, llh, rec);
900 case MDS_UNLINK64_REC:
901 rc = osp_sync_new_unlink64_job(d, llh, rec);
903 case MDS_SETATTR64_REC:
904 rc = osp_sync_new_setattr_job(d, llh, rec);
907 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
909 /* treat "unknown record type" as "invalid" */
914 /* For all kinds of records, not matter successful or not,
915 * we should decrease changes and bump last_processed_id.
917 if (d->opd_syn_prev_done) {
918 __u64 correct_id = osp_sync_correct_id(d, rec);
919 LASSERT(atomic_read(&d->opd_syn_changes) > 0);
920 LASSERT(correct_id <= d->opd_syn_last_committed_id);
921 /* NOTE: it's possible to meet same id if
922 * OST stores few stripes of same file
925 /* another thread may be trying to set new value */
927 if (correct_id > d->opd_syn_last_processed_id) {
928 d->opd_syn_last_processed_id = correct_id;
929 wake_up(&d->opd_syn_barrier_waitq);
933 atomic_dec(&d->opd_syn_changes);
936 atomic_dec(&d->opd_syn_rpc_in_flight);
937 atomic_dec(&d->opd_syn_rpc_in_progress);
940 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
941 d->opd_obd->obd_name, atomic_read(&d->opd_syn_rpc_in_flight),
942 atomic_read(&d->opd_syn_rpc_in_progress));
944 /* Delete the invalid record */
946 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
948 CERROR("%s: can't delete invalid record: "
949 "fid = "DFID", rec_id = %u, rc = %d\n",
950 d->opd_obd->obd_name,
951 PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
955 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
956 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);
962 * Cancel llog records for the committed changes.
964 * The function walks through the list of the committed RPCs and cancels
965 * corresponding llog records. see osp_sync_request_commit_cb() for the
968 * \param[in] env LU environment provided by the caller
969 * \param[in] d OSP device
971 static void osp_sync_process_committed(const struct lu_env *env,
972 struct osp_device *d)
974 struct obd_device *obd = d->opd_obd;
975 struct obd_import *imp = obd->u.cli.cl_import;
976 struct ost_body *body;
977 struct ptlrpc_request *req;
978 struct llog_ctxt *ctxt;
979 struct llog_handle *llh;
980 struct list_head list;
985 if (list_empty(&d->opd_syn_committed_there))
989 * if current status is -ENOSPC (lack of free space on OST)
990 * then we should poll OST immediately once object destroy
992 * notice: we do this upon commit as well because some backends
993 * (like DMU) do not release space right away.
995 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
996 osp_statfs_need_now(d);
999 * now cancel them all
1000 * XXX: can we improve this using some batching?
1001 * with batch RPC that'll happen automatically?
1002 * XXX: can we store ctxt in lod_device and save few cycles ?
1004 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1007 llh = ctxt->loc_handle;
1010 INIT_LIST_HEAD(&list);
1011 spin_lock(&d->opd_syn_lock);
1012 list_splice(&d->opd_syn_committed_there, &list);
1013 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1014 spin_unlock(&d->opd_syn_lock);
1016 while (!list_empty(&list)) {
1017 struct osp_job_req_args *jra;
1019 jra = list_entry(list.next, struct osp_job_req_args,
1020 jra_committed_link);
1021 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1022 list_del_init(&jra->jra_committed_link);
1024 req = container_of((void *)jra, struct ptlrpc_request,
1026 body = req_capsule_client_get(&req->rq_pill,
1029 /* import can be closing, thus all commit cb's are
1030 * called we can check committness directly */
1031 if (req->rq_import_generation == imp->imp_generation) {
1032 rc = llog_cat_cancel_records(env, llh, 1,
1035 CERROR("%s: can't cancel record: %d\n",
1038 DEBUG_REQ(D_OTHER, req, "imp_committed = %llu",
1039 imp->imp_peer_committed_transno);
1041 ptlrpc_req_finished(req);
1045 llog_ctxt_put(ctxt);
1047 LASSERT(atomic_read(&d->opd_syn_rpc_in_progress) >= done);
1048 atomic_sub(done, &d->opd_syn_rpc_in_progress);
1049 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
1050 d->opd_obd->obd_name, atomic_read(&d->opd_syn_rpc_in_flight),
1051 atomic_read(&d->opd_syn_rpc_in_progress));
1053 osp_sync_check_for_work(d);
1055 /* wake up the thread if requested to stop:
1056 * it might be waiting for in-progress to complete */
1057 if (unlikely(osp_sync_running(d) == 0))
1058 wake_up(&d->opd_syn_waitq);
1064 * The core of the syncing mechanism.
1066 * This is a callback called by the llog processing function. Essentially it
1067 * suspends llog processing until there is a record to process (it's supposed
1068 * to be committed locally). The function handles RPCs committed by the target
1069 * and cancels corresponding llog records.
1071 * \param[in] env LU environment provided by the caller
1072 * \param[in] llh llog handle we're processing
1073 * \param[in] rec current llog record
1074 * \param[in] data callback data containing a pointer to the device
1076 * \retval 0 to ask the caller (llog_process()) to continue
1077 * \retval LLOG_PROC_BREAK to ask the caller to break
1079 static int osp_sync_process_queues(const struct lu_env *env,
1080 struct llog_handle *llh,
1081 struct llog_rec_hdr *rec,
1084 struct osp_device *d = data;
1087 struct l_wait_info lwi = { 0 };
1089 if (!osp_sync_running(d)) {
1090 CDEBUG(D_HA, "stop llog processing\n");
1091 return LLOG_PROC_BREAK;
1094 /* process requests committed by OST */
1095 osp_sync_process_committed(env, d);
1097 /* if we there are changes to be processed and we have
1098 * resources for this ... do now */
1099 if (osp_sync_can_process_new(d, rec)) {
1101 /* ask llog for another record */
1102 CDEBUG(D_HA, "%u changes, %u in progress,"
1104 atomic_read(&d->opd_syn_changes),
1105 atomic_read(&d->opd_syn_rpc_in_progress),
1106 atomic_read(&d->opd_syn_rpc_in_flight));
1109 osp_sync_process_record(env, d, llh, rec);
1114 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
1115 osp_sync_remove_from_tracker(d);
1117 l_wait_event(d->opd_syn_waitq,
1118 !osp_sync_running(d) ||
1119 osp_sync_can_process_new(d, rec) ||
1120 !list_empty(&d->opd_syn_committed_there),
1128 * This thread runs llog_cat_process() scanner calling our callback
1129 * to process llog records. in the callback we implement tricky
1130 * state machine as we don't want to start scanning of the llog again
1131 * and again, also we don't want to process too many records and send
1132 * too many RPCs a time. so, depending on current load (num of changes
1133 * being synced to OST) the callback can suspend awaiting for some
1134 * new conditions, like syncs completed.
1136 * In order to process llog records left by previous boots and to allow
1137 * llog_process_thread() to find something (otherwise it'd just exit
1138 * immediately) we add a special GENERATATION record on each boot.
1140 * \param[in] _arg a pointer to thread's arguments
1142 * \retval 0 on success
1143 * \retval negative negated errno on error
1145 static int osp_sync_thread(void *_arg)
1147 struct osp_device *d = _arg;
1148 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1149 struct l_wait_info lwi = { 0 };
1150 struct llog_ctxt *ctxt;
1151 struct obd_device *obd = d->opd_obd;
1152 struct llog_handle *llh;
1158 rc = lu_env_init(&env, LCT_LOCAL);
1160 CERROR("%s: can't initialize env: rc = %d\n",
1163 spin_lock(&d->opd_syn_lock);
1164 thread->t_flags = SVC_STOPPED;
1165 spin_unlock(&d->opd_syn_lock);
1166 wake_up(&thread->t_ctl_waitq);
1171 spin_lock(&d->opd_syn_lock);
1172 thread->t_flags = SVC_RUNNING;
1173 spin_unlock(&d->opd_syn_lock);
1174 wake_up(&thread->t_ctl_waitq);
1176 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1178 CERROR("can't get appropriate context\n");
1179 GOTO(out, rc = -EINVAL);
1182 llh = ctxt->loc_handle;
1184 CERROR("can't get llh\n");
1185 llog_ctxt_put(ctxt);
1186 GOTO(out, rc = -EINVAL);
1189 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
1191 CERROR("%s: llog process with osp_sync_process_queues "
1192 "failed: %d\n", d->opd_obd->obd_name, rc);
1195 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1196 "%u changes, %u in progress, %u in flight: %d\n",
1197 atomic_read(&d->opd_syn_changes),
1198 atomic_read(&d->opd_syn_rpc_in_progress),
1199 atomic_read(&d->opd_syn_rpc_in_flight), rc);
1201 /* we don't expect llog_process_thread() to exit till umount */
1202 LASSERTF(thread->t_flags != SVC_RUNNING,
1203 "%u changes, %u in progress, %u in flight\n",
1204 atomic_read(&d->opd_syn_changes),
1205 atomic_read(&d->opd_syn_rpc_in_progress),
1206 atomic_read(&d->opd_syn_rpc_in_flight));
1208 /* wait till all the requests are completed */
1210 while (atomic_read(&d->opd_syn_rpc_in_progress) > 0) {
1211 osp_sync_process_committed(&env, d);
1213 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1214 rc = l_wait_event(d->opd_syn_waitq,
1215 atomic_read(&d->opd_syn_rpc_in_progress) == 0,
1217 if (rc == -ETIMEDOUT)
1219 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1220 d->opd_obd->obd_name,
1221 atomic_read(&d->opd_syn_rpc_in_progress),
1222 atomic_read(&d->opd_syn_rpc_in_flight),
1223 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1228 llog_cat_close(&env, llh);
1229 rc = llog_cleanup(&env, ctxt);
1231 CERROR("can't cleanup llog: %d\n", rc);
1233 LASSERTF(atomic_read(&d->opd_syn_rpc_in_progress) == 0,
1234 "%s: %d %d %sempty\n",
1235 d->opd_obd->obd_name, atomic_read(&d->opd_syn_rpc_in_progress),
1236 atomic_read(&d->opd_syn_rpc_in_flight),
1237 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1239 thread->t_flags = SVC_STOPPED;
1241 wake_up(&thread->t_ctl_waitq);
1251 * Initializes the llog. Specific llog to be used depends on the type of the
1252 * target OSP represents (OST or MDT). The function adds appends a new llog
1253 * record to mark the place where the records associated with this boot
1256 * \param[in] env LU environment provided by the caller
1257 * \param[in] d OSP device
1259 * \retval 0 on success
1260 * \retval negative negated errno on error
1262 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1264 struct osp_thread_info *osi = osp_env_info(env);
1265 struct lu_fid *fid = &osi->osi_fid;
1266 struct llog_handle *lgh = NULL;
1267 struct obd_device *obd = d->opd_obd;
1268 struct llog_ctxt *ctxt;
1276 * open llog corresponding to our OST
1278 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1279 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1281 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1283 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1284 &osi->osi_cid, fid);
1286 if (rc != -EFAULT) {
1287 CERROR("%s: can't get id from catalogs: rc = %d\n",
1292 /* After sparse OST indices is supported, the CATALOG file
1293 * may become a sparse file that results in failure on
1294 * reading. Skip this error as the llog will be created
1296 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1300 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1301 obd->obd_name, d->opd_index,
1302 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1303 osi->osi_cid.lci_logid.lgl_ogen);
1305 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT,
1306 d->opd_storage->dd_lu_dev.ld_obd,
1307 &osp_mds_ost_orig_logops);
1311 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1314 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1315 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1317 /* re-create llog if it is missing */
1319 logid_set_id(&osi->osi_cid.lci_logid, 0);
1321 GOTO(out_cleanup, rc);
1324 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1325 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1327 GOTO(out_cleanup, rc);
1328 osi->osi_cid.lci_logid = lgh->lgh_id;
1331 LASSERT(lgh != NULL);
1332 ctxt->loc_handle = lgh;
1334 rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
1336 GOTO(out_close, rc);
1338 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1339 &osi->osi_cid, fid);
1341 GOTO(out_close, rc);
1344 * put a mark in the llog till which we'll be processing
1345 * old records restless
1347 d->opd_syn_generation.mnt_cnt = cfs_time_current();
1348 d->opd_syn_generation.conn_cnt = cfs_time_current();
1350 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1351 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1353 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1354 sizeof(osi->osi_gen.lgr_gen));
1356 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1358 GOTO(out_close, rc);
1359 llog_ctxt_put(ctxt);
1362 llog_cat_close(env, lgh);
1364 llog_cleanup(env, ctxt);
1369 * Cleanup llog used for syncing.
1371 * Closes and cleanups the llog. The function is called when the device is
1374 * \param[in] env LU environment provided by the caller
1375 * \param[in] d OSP device
1377 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1379 struct llog_ctxt *ctxt;
1381 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1383 llog_cat_close(env, ctxt->loc_handle);
1384 llog_cleanup(env, ctxt);
1389 * Initialization of the sync component of OSP.
1391 * Initializes the llog and starts a new thread to handle the changes to
1392 * the remote target (OST or MDT).
1394 * \param[in] env LU environment provided by the caller
1395 * \param[in] d OSP device
1397 * \retval 0 on success
1398 * \retval negative negated errno on error
1400 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1402 struct l_wait_info lwi = { 0 };
1403 struct task_struct *task;
1408 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1409 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1410 spin_lock_init(&d->opd_syn_lock);
1411 init_waitqueue_head(&d->opd_syn_waitq);
1412 init_waitqueue_head(&d->opd_syn_barrier_waitq);
1413 thread_set_flags(&d->opd_syn_thread, SVC_INIT);
1414 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1415 INIT_LIST_HEAD(&d->opd_syn_inflight_list);
1416 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1418 if (d->opd_storage->dd_rdonly)
1421 rc = osp_sync_id_traction_init(d);
1426 * initialize llog storing changes
1428 rc = osp_sync_llog_init(env, d);
1430 CERROR("%s: can't initialize llog: rc = %d\n",
1431 d->opd_obd->obd_name, rc);
1436 * Start synchronization thread
1438 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1439 d->opd_index, d->opd_group);
1442 CERROR("%s: cannot start sync thread: rc = %d\n",
1443 d->opd_obd->obd_name, rc);
1447 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1448 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1452 osp_sync_llog_fini(env, d);
1454 osp_sync_id_traction_fini(d);
1459 * Stop the syncing thread.
1461 * Asks the syncing thread to stop and wait until it's stopped.
1463 * \param[in] d OSP device
1467 int osp_sync_fini(struct osp_device *d)
1469 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1473 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1474 thread->t_flags = SVC_STOPPING;
1475 wake_up(&d->opd_syn_waitq);
1476 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
1480 * unregister transaction callbacks only when sync thread
1481 * has finished operations with llog
1483 osp_sync_id_traction_fini(d);
1488 static DEFINE_MUTEX(osp_id_tracker_sem);
1489 static struct list_head osp_id_tracker_list =
1490 LIST_HEAD_INIT(osp_id_tracker_list);
1493 * OSD commit callback.
1495 * The function is used as a local OSD commit callback to track the highest
1496 * committed llog record id. see osp_sync_id_traction_init() for the details.
1498 * \param[in] th local transaction handle committed
1499 * \param[in] cookie commit callback data (our private structure)
1501 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1503 struct osp_id_tracker *tr = cookie;
1504 struct osp_device *d;
1505 struct osp_txn_info *txn;
1509 txn = osp_txn_info(&th->th_ctx);
1510 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1513 spin_lock(&tr->otr_lock);
1514 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1515 CDEBUG(D_OTHER, "committed: %llu -> %llu\n",
1516 tr->otr_committed_id, txn->oti_current_id);
1517 tr->otr_committed_id = txn->oti_current_id;
1519 list_for_each_entry(d, &tr->otr_wakeup_list,
1521 d->opd_syn_last_committed_id = tr->otr_committed_id;
1522 wake_up(&d->opd_syn_waitq);
1525 spin_unlock(&tr->otr_lock);
1529 * Initialize commit tracking mechanism.
1531 * Some setups may have thousands of OSTs and each will be represented by OSP.
1532 * Meaning order of magnitute many more changes to apply every second. In order
1533 * to keep the number of commit callbacks low this mechanism was introduced.
1534 * The mechanism is very similar to transno used by MDT service: it's an single
1535 * ID stream which can be assigned by any OSP to its llog records. The tricky
1536 * part is that ID is stored in per-transaction data and re-used by all the OSPs
1537 * involved in that transaction. Then all these OSPs are woken up utilizing a single OSD commit callback.
1539 * The function initializes the data used by the tracker described above.
1540 * A singler tracker per OSD device is created.
1542 * \param[in] d OSP device
1544 * \retval 0 on success
1545 * \retval negative negated errno on error
1547 static int osp_sync_id_traction_init(struct osp_device *d)
1549 struct osp_id_tracker *tr, *found = NULL;
1553 LASSERT(d->opd_storage);
1554 LASSERT(d->opd_syn_tracker == NULL);
1555 INIT_LIST_HEAD(&d->opd_syn_ontrack);
1557 mutex_lock(&osp_id_tracker_sem);
1558 list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1559 if (tr->otr_dev == d->opd_storage) {
1560 LASSERT(atomic_read(&tr->otr_refcount));
1561 atomic_inc(&tr->otr_refcount);
1562 d->opd_syn_tracker = tr;
1568 if (found == NULL) {
1572 d->opd_syn_tracker = tr;
1573 spin_lock_init(&tr->otr_lock);
1574 tr->otr_dev = d->opd_storage;
1575 tr->otr_next_id = 1;
1576 tr->otr_committed_id = 0;
1577 atomic_set(&tr->otr_refcount, 1);
1578 INIT_LIST_HEAD(&tr->otr_wakeup_list);
1579 list_add(&tr->otr_list, &osp_id_tracker_list);
1580 tr->otr_tx_cb.dtc_txn_commit =
1581 osp_sync_tracker_commit_cb;
1582 tr->otr_tx_cb.dtc_cookie = tr;
1583 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1584 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1588 mutex_unlock(&osp_id_tracker_sem);
1594 * Release commit tracker.
1596 * Decrease a refcounter on the tracker used by the given OSP device \a d.
1597 * If no more users left, then the tracker is released.
1599 * \param[in] d OSP device
1601 static void osp_sync_id_traction_fini(struct osp_device *d)
1603 struct osp_id_tracker *tr;
1608 tr = d->opd_syn_tracker;
1614 osp_sync_remove_from_tracker(d);
1616 mutex_lock(&osp_id_tracker_sem);
1617 if (atomic_dec_and_test(&tr->otr_refcount)) {
1618 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1619 LASSERT(list_empty(&tr->otr_wakeup_list));
1620 list_del(&tr->otr_list);
1622 d->opd_syn_tracker = NULL;
1624 mutex_unlock(&osp_id_tracker_sem);
1630 * Generate a new ID on a tracker.
1632 * Generates a new ID using the tracker associated with the given OSP device
1633 * \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
1634 * the wakeup list, so OSP won't miss when a transaction using the ID is
1637 * \param[in] d OSP device
1638 * \param[in] id 0 or ID generated previously
1640 * \retval ID the caller should use
1642 static __u64 osp_sync_id_get(struct osp_device *d, __u64 id)
1644 struct osp_id_tracker *tr;
1646 tr = d->opd_syn_tracker;
1649 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1650 spin_lock(&tr->otr_lock);
1651 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_TRACK_OVERFLOW))
1652 tr->otr_next_id = 0xfffffff0;
1654 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1655 spin_unlock(&tr->otr_lock);
1656 CERROR("%s: next %llu, last synced %llu\n",
1657 d->opd_obd->obd_name, tr->otr_next_id,
1658 d->opd_syn_last_used_id);
1663 id = tr->otr_next_id++;
1664 if (id > d->opd_syn_last_used_id)
1665 d->opd_syn_last_used_id = id;
1666 if (list_empty(&d->opd_syn_ontrack))
1667 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1668 spin_unlock(&tr->otr_lock);
1669 CDEBUG(D_OTHER, "new id %llu\n", id);
1675 * Stop to propagate commit status to OSP.
1677 * If the OSP does not have any llog records she's waiting to commit, then
1678 * it is possible to unsubscribe from wakeups from the tracking using this
1681 * \param[in] d OSP device not willing to wakeup
1683 static void osp_sync_remove_from_tracker(struct osp_device *d)
1685 struct osp_id_tracker *tr;
1687 tr = d->opd_syn_tracker;
1690 if (list_empty(&d->opd_syn_ontrack))
1693 spin_lock(&tr->otr_lock);
1694 list_del_init(&d->opd_syn_ontrack);
1695 spin_unlock(&tr->otr_lock);