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 <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
98 * Return status: whether OSP thread should keep running
100 * \param[in] d OSP device
102 * \retval 1 should keep running
103 * \retval 0 should stop
105 static inline int osp_sync_running(struct osp_device *d)
107 return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
111 * Check status: whether OSP thread has stopped
113 * \param[in] d OSP device
115 * \retval 0 still running
118 static inline int osp_sync_stopped(struct osp_device *d)
120 return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
124 ** Check for new changes to sync
126 * \param[in] d OSP device
128 * \retval 1 there are changes
129 * \retval 0 there are no changes
131 static inline int osp_sync_has_new_job(struct osp_device *d)
133 return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
134 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
135 || (d->opd_syn_prev_done == 0);
138 static inline int osp_sync_low_in_progress(struct osp_device *d)
140 return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
144 * Check for room in the network pipe to OST
146 * \param[in] d OSP device
148 * \retval 1 there is room
149 * \retval 0 no room, the pipe is full
151 static inline int osp_sync_low_in_flight(struct osp_device *d)
153 return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
157 * Wake up check for the main sync thread
159 * \param[in] d OSP device
161 * \retval 1 time to wake up
162 * \retval 0 no need to wake up
164 static inline int osp_sync_has_work(struct osp_device *d)
166 /* has new/old changes and low in-progress? */
167 if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
168 osp_sync_low_in_flight(d) && d->opd_imp_connected)
171 /* has remotely committed? */
172 if (!list_empty(&d->opd_syn_committed_there))
178 #define osp_sync_check_for_work(d) \
180 if (osp_sync_has_work(d)) { \
181 wake_up(&d->opd_syn_waitq); \
185 void __osp_sync_check_for_work(struct osp_device *d)
187 osp_sync_check_for_work(d);
191 * Check and return ready-for-new status.
193 * The thread processing llog record uses this function to check whether
194 * it's time to take another record and process it. The number of conditions
195 * must be met: the connection should be ready, RPCs in flight not exceeding
196 * the limit, the record is committed locally, etc (see the lines below).
198 * \param[in] d OSP device
199 * \param[in] rec next llog record to process
201 * \retval 0 not ready
204 static inline int osp_sync_can_process_new(struct osp_device *d,
205 struct llog_rec_hdr *rec)
209 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
211 if (!osp_sync_low_in_progress(d))
213 if (!osp_sync_low_in_flight(d))
215 if (!d->opd_imp_connected)
217 if (d->opd_syn_prev_done == 0)
219 if (d->opd_syn_changes == 0)
221 if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
227 * Declare intention to add a new change.
229 * With regard to OSD API, we have to declare any changes ahead. In this
230 * case we declare an intention to add a llog record representing the
231 * change on the local storage.
233 * \param[in] env LU environment provided by the caller
234 * \param[in] o OSP object
235 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
236 * \param[in] th transaction handle (local)
238 * \retval 0 on success
239 * \retval negative negated errno on error
241 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
242 llog_op_type type, struct thandle *th)
244 struct osp_thread_info *osi = osp_env_info(env);
245 struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
246 struct llog_ctxt *ctxt;
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_dev == d->opd_storage);
256 case MDS_UNLINK64_REC:
257 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
259 case MDS_SETATTR64_REC:
260 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
266 /* we want ->dt_trans_start() to allocate per-thandle structure */
267 th->th_tags |= LCT_OSP_THREAD;
269 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
272 rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
279 * Generate a llog record for a given change.
281 * Generates a llog record for the change passed. The change can be of two
282 * types: unlink and setattr. The record gets an ID which later will be
283 * used to track commit status of the change. For unlink changes, the caller
284 * can supply a starting FID and the count of the objects to destroy. For
285 * setattr the caller should apply attributes to apply.
288 * \param[in] env LU environment provided by the caller
289 * \param[in] d OSP device
290 * \param[in] fid fid of the object the change should be applied to
291 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
292 * \param[in] count count of objects to destroy
293 * \param[in] th transaction handle (local)
294 * \param[in] attr attributes for setattr
296 * \retval 0 on success
297 * \retval negative negated errno on error
299 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
300 const struct lu_fid *fid, llog_op_type type,
301 int count, struct thandle *th,
302 const struct lu_attr *attr)
304 struct osp_thread_info *osi = osp_env_info(env);
305 struct llog_ctxt *ctxt;
306 struct osp_txn_info *txn;
311 /* it's a layering violation, to access internals of th,
312 * but we can do this as a sanity check, for a while */
313 LASSERT(th->th_dev == d->opd_storage);
316 case MDS_UNLINK64_REC:
317 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
318 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
319 osi->osi_unlink.lur_fid = *fid;
320 osi->osi_unlink.lur_count = count;
322 case MDS_SETATTR64_REC:
323 rc = fid_to_ostid(fid, &osi->osi_oi);
325 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
326 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
327 osi->osi_setattr.lsr_oi = osi->osi_oi;
329 osi->osi_setattr.lsr_uid = attr->la_uid;
330 osi->osi_setattr.lsr_gid = attr->la_gid;
331 osi->osi_setattr.lsr_valid =
332 ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
333 ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0);
339 txn = osp_txn_info(&th->th_ctx);
342 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
343 osi->osi_hdr.lrh_id = txn->oti_current_id;
345 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
348 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
352 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
353 d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
354 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
355 (unsigned long) osi->osi_cookie.lgc_index, rc);
360 if (likely(rc == 0)) {
361 spin_lock(&d->opd_syn_lock);
362 d->opd_syn_changes++;
363 spin_unlock(&d->opd_syn_lock);
369 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
370 llog_op_type type, struct thandle *th,
371 const struct lu_attr *attr)
373 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
374 lu_object_fid(&o->opo_obj.do_lu), type, 1,
378 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
379 struct lu_fid *fid, int lost, struct thandle *th)
381 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
385 * it's quite obvious we can't maintain all the structures in the memory:
386 * while OST is down, MDS can be processing thousands and thousands of unlinks
387 * filling persistent llogs and in-core respresentation
389 * this doesn't scale at all. so we need basically the following:
390 * a) destroy/setattr append llog records
391 * b) once llog has grown to X records, we process first Y committed records
393 * once record R is found via llog_process(), it becomes committed after any
394 * subsequent commit callback (at the most)
398 * ptlrpc commit callback.
400 * The callback is called by PTLRPC when a RPC is reported committed by the
401 * target (OST). We register the callback for the every RPC applying a change
402 * from the llog. This way we know then the llog records can be cancelled.
403 * Notice the callback can be called when OSP is finishing. We can detect this
404 * checking that actual transno in the request is less or equal of known
405 * committed transno (see osp_sync_process_committed() for the details).
406 * XXX: this is pretty expensive and can be improved later using batching.
408 * \param[in] req request
410 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
412 struct osp_device *d = req->rq_cb_data;
414 CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
416 if (unlikely(req->rq_transno == 0))
419 /* do not do any opd_dyn_rpc_* accounting here
420 * it's done in osp_sync_interpret sooner or later */
423 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
424 LASSERT(list_empty(&req->rq_exp_list));
426 ptlrpc_request_addref(req);
428 spin_lock(&d->opd_syn_lock);
429 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
430 spin_unlock(&d->opd_syn_lock);
432 /* XXX: some batching wouldn't hurt */
433 wake_up(&d->opd_syn_waitq);
437 * RPC interpretation callback.
439 * The callback is called by ptlrpc when RPC is replied. Now we have to decide
441 * - put request on a special list to wait until it's committed by the target,
442 * if the request is succesful
443 * - schedule llog record cancel if no target object is found
444 * - try later (essentially after reboot) in case of unexpected error
446 * \param[in] env LU environment provided by the caller
447 * \param[in] req request replied
448 * \param[in] aa callback data
449 * \param[in] rc result of RPC
453 static int osp_sync_interpret(const struct lu_env *env,
454 struct ptlrpc_request *req, void *aa, int rc)
456 struct osp_device *d = req->rq_cb_data;
458 if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
459 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
460 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
463 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
464 atomic_read(&req->rq_refcount),
465 rc, (unsigned) req->rq_transno);
466 LASSERT(rc || req->rq_transno);
470 * we tried to destroy object or update attributes,
471 * but object doesn't exist anymore - cancell llog record
473 LASSERT(req->rq_transno == 0);
474 LASSERT(list_empty(&req->rq_exp_list));
476 ptlrpc_request_addref(req);
478 spin_lock(&d->opd_syn_lock);
479 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
480 spin_unlock(&d->opd_syn_lock);
482 wake_up(&d->opd_syn_waitq);
484 struct obd_import *imp = req->rq_import;
486 * error happened, we'll try to repeat on next boot ?
488 LASSERTF(req->rq_transno == 0 ||
489 req->rq_import_generation < imp->imp_generation,
490 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
491 req->rq_transno, rc, req->rq_import_generation,
492 imp->imp_generation);
493 if (req->rq_transno == 0) {
494 /* this is the last time we see the request
495 * if transno is not zero, then commit cb
496 * will be called at some point */
497 LASSERT(d->opd_syn_rpc_in_progress > 0);
498 spin_lock(&d->opd_syn_lock);
499 d->opd_syn_rpc_in_progress--;
500 spin_unlock(&d->opd_syn_lock);
503 wake_up(&d->opd_syn_waitq);
504 } else if (d->opd_pre != NULL &&
505 unlikely(d->opd_pre_status == -ENOSPC)) {
507 * if current status is -ENOSPC (lack of free space on OST)
508 * then we should poll OST immediately once object destroy
511 osp_statfs_need_now(d);
514 LASSERT(d->opd_syn_rpc_in_flight > 0);
515 spin_lock(&d->opd_syn_lock);
516 d->opd_syn_rpc_in_flight--;
517 spin_unlock(&d->opd_syn_lock);
518 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
519 wake_up(&d->opd_syn_barrier_waitq);
520 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
521 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
522 d->opd_syn_rpc_in_progress);
524 osp_sync_check_for_work(d);
530 ** Add request to ptlrpc queue.
532 * This is just a tiny helper function to put the request on the sending list
534 * \param[in] d OSP device
535 * \param[in] req request
537 static void osp_sync_send_new_rpc(struct osp_device *d,
538 struct ptlrpc_request *req)
540 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
541 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
543 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
548 * Allocate and prepare RPC for a new change.
550 * The function allocates and initializes an RPC which will be sent soon to
551 * apply the change to the target OST. The request is initialized from the
552 * llog record passed. Notice only the fields common to all type of changes
555 * \param[in] d OSP device
556 * \param[in] llh llog handle where the record is stored
557 * \param[in] h llog record
558 * \param[in] op type of the change
559 * \param[in] format request format to be used
561 * \retval pointer new request on success
562 * \retval ERR_PTR(errno) on error
564 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
565 struct llog_handle *llh,
566 struct llog_rec_hdr *h,
568 const struct req_format *format)
570 struct ptlrpc_request *req;
571 struct ost_body *body;
572 struct obd_import *imp;
575 /* Prepare the request */
576 imp = d->opd_obd->u.cli.cl_import;
578 req = ptlrpc_request_alloc(imp, format);
580 RETURN(ERR_PTR(-ENOMEM));
582 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
584 ptlrpc_req_finished(req);
589 * this is a trick: to save on memory allocations we put cookie
590 * into the request, but don't set corresponded flag in o_valid
591 * so that OST doesn't interpret this cookie. once the request
592 * is committed on OST we take cookie from the request and cancel
594 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
596 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
597 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
598 body->oa.o_lcookie.lgc_index = h->lrh_index;
599 INIT_LIST_HEAD(&req->rq_exp_list);
600 req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
602 req->rq_interpret_reply = osp_sync_interpret;
603 req->rq_commit_cb = osp_sync_request_commit_cb;
606 ptlrpc_request_set_replen(req);
612 * Generate a request for setattr change.
614 * The function prepares a new RPC, initializes it with setattr specific
615 * bits and send the RPC.
617 * \param[in] d OSP device
618 * \param[in] llh llog handle where the record is stored
619 * \param[in] h llog record
621 * \retval 0 on success
622 * \retval negative negated errno on error
624 static int osp_sync_new_setattr_job(struct osp_device *d,
625 struct llog_handle *llh,
626 struct llog_rec_hdr *h)
628 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
629 struct ptlrpc_request *req;
630 struct ost_body *body;
633 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
635 /* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
636 * so no bits other than these should be set. */
637 if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
638 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
639 d->opd_obd->obd_name, rec->lsr_valid);
640 /* return 0 so that sync thread can continue processing
645 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
647 RETURN(PTR_ERR(req));
649 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
651 body->oa.o_oi = rec->lsr_oi;
652 body->oa.o_uid = rec->lsr_uid;
653 body->oa.o_gid = rec->lsr_gid;
654 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
655 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
656 * we assume that both UID and GID are valid in that case. */
657 if (rec->lsr_valid == 0)
658 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
660 body->oa.o_valid |= rec->lsr_valid;
662 osp_sync_send_new_rpc(d, req);
667 * Generate a request for unlink change.
669 * The function prepares a new RPC, initializes it with unlink(destroy)
670 * specific bits and sends the RPC. The function is used to handle
671 * llog_unlink_rec which were used in the older versions of Lustre.
672 * Current version uses llog_unlink_rec64.
674 * \param[in] d OSP device
675 * \param[in] llh llog handle where the record is stored
676 * \param[in] h llog record
678 * \retval 0 on success
679 * \retval negative negated errno on error
681 static int osp_sync_new_unlink_job(struct osp_device *d,
682 struct llog_handle *llh,
683 struct llog_rec_hdr *h)
685 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
686 struct ptlrpc_request *req;
687 struct ost_body *body;
690 LASSERT(h->lrh_type == MDS_UNLINK_REC);
692 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
694 RETURN(PTR_ERR(req));
696 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
698 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
699 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
700 body->oa.o_misc = rec->lur_count;
701 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
703 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
705 osp_sync_send_new_rpc(d, req);
710 * Prepare OUT-based object destroy RPC.
712 * The function allocates a new RPC with OUT format. Then initializes the RPC
713 * to contain OUT_DESTROY update against the object specified in the llog
714 * record provided by the caller.
716 * \param[in] env LU environment provided by the caller
717 * \param[in] osp OSP device
718 * \param[in] llh llog handle where the record is stored
719 * \param[in] h llog record
720 * \param[out] reqp request prepared
722 * \retval 0 on success
723 * \retval negative negated errno on error
725 static int osp_prep_unlink_update_req(const struct lu_env *env,
726 struct osp_device *osp,
727 struct llog_handle *llh,
728 struct llog_rec_hdr *h,
729 struct ptlrpc_request **reqp)
731 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
732 struct dt_update_request *update = NULL;
733 struct ptlrpc_request *req;
734 struct llog_cookie lcookie;
740 update = dt_update_request_create(&osp->opd_dt_dev);
742 RETURN(PTR_ERR(update));
744 /* This can only happens for unlink slave directory, so decrease
745 * ref for ".." and "." */
746 rc = out_update_pack(env, &update->dur_buf, OUT_REF_DEL, &rec->lur_fid,
751 rc = out_update_pack(env, &update->dur_buf, OUT_REF_DEL, &rec->lur_fid,
756 lcookie.lgc_lgl = llh->lgh_id;
757 lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
758 lcookie.lgc_index = h->lrh_index;
759 size = sizeof(lcookie);
762 rc = out_update_pack(env, &update->dur_buf, OUT_DESTROY, &rec->lur_fid,
767 rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
768 update->dur_buf.ub_req, &req);
772 INIT_LIST_HEAD(&req->rq_exp_list);
773 req->rq_svc_thread = (void *)OSP_JOB_MAGIC;
775 req->rq_interpret_reply = osp_sync_interpret;
776 req->rq_commit_cb = osp_sync_request_commit_cb;
777 req->rq_cb_data = osp;
779 ptlrpc_request_set_replen(req);
783 dt_update_request_destroy(update);
789 * Generate a request for unlink change.
791 * The function prepares a new RPC, initializes it with unlink(destroy)
792 * specific bits and sends the RPC. Depending on the target (MDT or OST)
793 * two different protocols are used. For MDT we use OUT (basically OSD API
794 * updates transferred via a network). For OST we still use the old
795 * protocol (OBD?), originally for compatibility. Later we can start to
796 * use OUT for OST as well, this will allow batching and better code
799 * \param[in] env LU environment provided by the caller
800 * \param[in] d OSP device
801 * \param[in] llh llog handle where the record is stored
802 * \param[in] h llog record
804 * \retval 0 on success
805 * \retval negative negated errno on error
807 static int osp_sync_new_unlink64_job(const struct lu_env *env,
808 struct osp_device *d,
809 struct llog_handle *llh,
810 struct llog_rec_hdr *h)
812 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
813 struct ptlrpc_request *req = NULL;
814 struct ost_body *body;
818 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
820 if (d->opd_connect_mdt) {
821 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
825 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
828 RETURN(PTR_ERR(req));
830 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
833 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
836 body->oa.o_misc = rec->lur_count;
837 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
840 osp_sync_send_new_rpc(d, req);
845 * Process llog records.
847 * This function is called to process the llog records committed locally.
848 * In the recovery model used by OSP we can apply a change to a remote
849 * target once corresponding transaction (like posix unlink) is committed
850 * locally so can't revert.
851 * Depending on the llog record type, a given handler is called that is
852 * responsible for preparing and sending the RPC to apply the change.
853 * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
855 * \param[in] env LU environment provided by the caller
856 * \param[in] d OSP device
857 * \param[in] llh llog handle where the record is stored
858 * \param[in] rec llog record
860 * \retval 0 on success
861 * \retval negative negated errno on error
863 static int osp_sync_process_record(const struct lu_env *env,
864 struct osp_device *d,
865 struct llog_handle *llh,
866 struct llog_rec_hdr *rec)
868 struct llog_cookie cookie;
871 cookie.lgc_lgl = llh->lgh_id;
872 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
873 cookie.lgc_index = rec->lrh_index;
875 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
876 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
878 /* we're waiting for the record generated by this instance */
879 LASSERT(d->opd_syn_prev_done == 0);
880 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
881 sizeof(gen->lgr_gen))) {
882 CDEBUG(D_HA, "processed all old entries\n");
883 d->opd_syn_prev_done = 1;
886 /* cancel any generation record */
887 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
894 * now we prepare and fill requests to OST, put them on the queue
895 * and fire after next commit callback
898 /* notice we increment counters before sending RPC, to be consistent
899 * in RPC interpret callback which may happen very quickly */
900 spin_lock(&d->opd_syn_lock);
901 d->opd_syn_rpc_in_flight++;
902 d->opd_syn_rpc_in_progress++;
903 spin_unlock(&d->opd_syn_lock);
905 switch (rec->lrh_type) {
906 /* case MDS_UNLINK_REC is kept for compatibility */
908 rc = osp_sync_new_unlink_job(d, llh, rec);
910 case MDS_UNLINK64_REC:
911 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
913 case MDS_SETATTR64_REC:
914 rc = osp_sync_new_setattr_job(d, llh, rec);
917 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
919 /* we should continue processing */
922 /* rc > 0 means sync RPC being added to the queue */
923 if (likely(rc > 0)) {
924 spin_lock(&d->opd_syn_lock);
925 if (d->opd_syn_prev_done) {
926 LASSERT(d->opd_syn_changes > 0);
927 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
929 * NOTE: it's possible to meet same id if
930 * OST stores few stripes of same file
932 if (rec->lrh_id > d->opd_syn_last_processed_id) {
933 d->opd_syn_last_processed_id = rec->lrh_id;
934 wake_up(&d->opd_syn_barrier_waitq);
937 d->opd_syn_changes--;
939 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
940 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
941 d->opd_syn_rpc_in_progress);
942 spin_unlock(&d->opd_syn_lock);
945 spin_lock(&d->opd_syn_lock);
946 d->opd_syn_rpc_in_flight--;
947 d->opd_syn_rpc_in_progress--;
948 spin_unlock(&d->opd_syn_lock);
951 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
952 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
957 * Cancel llog records for the committed changes.
959 * The function walks through the list of the committed RPCs and cancels
960 * corresponding llog records. see osp_sync_request_commit_cb() for the
963 * \param[in] env LU environment provided by the caller
964 * \param[in] d OSP device
966 static void osp_sync_process_committed(const struct lu_env *env,
967 struct osp_device *d)
969 struct obd_device *obd = d->opd_obd;
970 struct obd_import *imp = obd->u.cli.cl_import;
971 struct ost_body *body;
972 struct ptlrpc_request *req, *tmp;
973 struct llog_ctxt *ctxt;
974 struct llog_handle *llh;
975 struct list_head list;
980 if (list_empty(&d->opd_syn_committed_there))
984 * if current status is -ENOSPC (lack of free space on OST)
985 * then we should poll OST immediately once object destroy
987 * notice: we do this upon commit as well because some backends
988 * (like DMU) do not release space right away.
990 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
991 osp_statfs_need_now(d);
994 * now cancel them all
995 * XXX: can we improve this using some batching?
996 * with batch RPC that'll happen automatically?
997 * XXX: can we store ctxt in lod_device and save few cycles ?
999 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1002 llh = ctxt->loc_handle;
1005 INIT_LIST_HEAD(&list);
1006 spin_lock(&d->opd_syn_lock);
1007 list_splice(&d->opd_syn_committed_there, &list);
1008 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1009 spin_unlock(&d->opd_syn_lock);
1011 list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
1012 struct llog_cookie *lcookie = NULL;
1014 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
1015 list_del_init(&req->rq_exp_list);
1017 if (d->opd_connect_mdt) {
1018 struct object_update_request *ureq;
1019 struct object_update *update;
1020 ureq = req_capsule_client_get(&req->rq_pill,
1022 LASSERT(ureq != NULL &&
1023 ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
1025 /* 1st/2nd is for decref . and .., 3rd one is for
1026 * destroy, where the log cookie is stored.
1027 * See osp_prep_unlink_update_req */
1028 update = object_update_request_get(ureq, 2, NULL);
1029 LASSERT(update != NULL);
1030 lcookie = object_update_param_get(update, 0, NULL);
1031 LASSERT(lcookie != NULL);
1033 body = req_capsule_client_get(&req->rq_pill,
1036 lcookie = &body->oa.o_lcookie;
1038 /* import can be closing, thus all commit cb's are
1039 * called we can check committness directly */
1040 if (req->rq_transno <= imp->imp_peer_committed_transno) {
1041 rc = llog_cat_cancel_records(env, llh, 1, lcookie);
1043 CERROR("%s: can't cancel record: %d\n",
1046 DEBUG_REQ(D_HA, req, "not committed");
1049 ptlrpc_req_finished(req);
1053 llog_ctxt_put(ctxt);
1055 LASSERT(d->opd_syn_rpc_in_progress >= done);
1056 spin_lock(&d->opd_syn_lock);
1057 d->opd_syn_rpc_in_progress -= done;
1058 spin_unlock(&d->opd_syn_lock);
1059 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
1060 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
1061 d->opd_syn_rpc_in_progress);
1063 osp_sync_check_for_work(d);
1065 /* wake up the thread if requested to stop:
1066 * it might be waiting for in-progress to complete */
1067 if (unlikely(osp_sync_running(d) == 0))
1068 wake_up(&d->opd_syn_waitq);
1074 * The core of the syncing mechanism.
1076 * This is a callback called by the llog processing function. Essentially it
1077 * suspends llog processing until there is a record to process (it's supposed
1078 * to be committed locally). The function handles RPCs committed by the target
1079 * and cancels corresponding llog records.
1081 * \param[in] env LU environment provided by the caller
1082 * \param[in] llh llog handle we're processing
1083 * \param[in] rec current llog record
1084 * \param[in] data callback data containing a pointer to the device
1086 * \retval 0 to ask the caller (llog_process()) to continue
1087 * \retval LLOG_PROC_BREAK to ask the caller to break
1089 static int osp_sync_process_queues(const struct lu_env *env,
1090 struct llog_handle *llh,
1091 struct llog_rec_hdr *rec,
1094 struct osp_device *d = data;
1098 struct l_wait_info lwi = { 0 };
1100 if (!osp_sync_running(d)) {
1101 CDEBUG(D_HA, "stop llog processing\n");
1102 return LLOG_PROC_BREAK;
1105 /* process requests committed by OST */
1106 osp_sync_process_committed(env, d);
1108 /* if we there are changes to be processed and we have
1109 * resources for this ... do now */
1110 if (osp_sync_can_process_new(d, rec)) {
1112 /* ask llog for another record */
1113 CDEBUG(D_HA, "%lu changes, %u in progress,"
1116 d->opd_syn_rpc_in_progress,
1117 d->opd_syn_rpc_in_flight);
1122 * try to send, in case of disconnection, suspend
1123 * processing till we can send this request
1126 rc = osp_sync_process_record(env, d, llh, rec);
1128 * XXX: probably different handling is needed
1129 * for some bugs, like immediate exit or if
1133 CERROR("can't send: %d\n", rc);
1134 l_wait_event(d->opd_syn_waitq,
1135 !osp_sync_running(d) ||
1136 osp_sync_has_work(d),
1139 } while (rc != 0 && osp_sync_running(d));
1145 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
1146 osp_sync_remove_from_tracker(d);
1148 l_wait_event(d->opd_syn_waitq,
1149 !osp_sync_running(d) ||
1150 osp_sync_can_process_new(d, rec) ||
1151 !list_empty(&d->opd_syn_committed_there),
1159 * This thread runs llog_cat_process() scanner calling our callback
1160 * to process llog records. in the callback we implement tricky
1161 * state machine as we don't want to start scanning of the llog again
1162 * and again, also we don't want to process too many records and send
1163 * too many RPCs a time. so, depending on current load (num of changes
1164 * being synced to OST) the callback can suspend awaiting for some
1165 * new conditions, like syncs completed.
1167 * In order to process llog records left by previous boots and to allow
1168 * llog_process_thread() to find something (otherwise it'd just exit
1169 * immediately) we add a special GENERATATION record on each boot.
1171 * \param[in] _arg a pointer to thread's arguments
1173 * \retval 0 on success
1174 * \retval negative negated errno on error
1176 static int osp_sync_thread(void *_arg)
1178 struct osp_device *d = _arg;
1179 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1180 struct l_wait_info lwi = { 0 };
1181 struct llog_ctxt *ctxt;
1182 struct obd_device *obd = d->opd_obd;
1183 struct llog_handle *llh;
1189 rc = lu_env_init(&env, LCT_LOCAL);
1191 CERROR("%s: can't initialize env: rc = %d\n",
1196 spin_lock(&d->opd_syn_lock);
1197 thread->t_flags = SVC_RUNNING;
1198 spin_unlock(&d->opd_syn_lock);
1199 wake_up(&thread->t_ctl_waitq);
1201 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1203 CERROR("can't get appropriate context\n");
1204 GOTO(out, rc = -EINVAL);
1207 llh = ctxt->loc_handle;
1209 CERROR("can't get llh\n");
1210 llog_ctxt_put(ctxt);
1211 GOTO(out, rc = -EINVAL);
1214 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
1215 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1216 "%lu changes, %u in progress, %u in flight: %d\n",
1217 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1218 d->opd_syn_rpc_in_flight, rc);
1220 /* we don't expect llog_process_thread() to exit till umount */
1221 LASSERTF(thread->t_flags != SVC_RUNNING,
1222 "%lu changes, %u in progress, %u in flight\n",
1223 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1224 d->opd_syn_rpc_in_flight);
1226 /* wait till all the requests are completed */
1228 while (d->opd_syn_rpc_in_progress > 0) {
1229 osp_sync_process_committed(&env, d);
1231 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1232 rc = l_wait_event(d->opd_syn_waitq,
1233 d->opd_syn_rpc_in_progress == 0,
1235 if (rc == -ETIMEDOUT)
1237 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1238 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1239 d->opd_syn_rpc_in_flight,
1240 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1244 llog_cat_close(&env, llh);
1245 rc = llog_cleanup(&env, ctxt);
1247 CERROR("can't cleanup llog: %d\n", rc);
1249 LASSERTF(d->opd_syn_rpc_in_progress == 0,
1250 "%s: %d %d %sempty\n",
1251 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1252 d->opd_syn_rpc_in_flight,
1253 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1255 thread->t_flags = SVC_STOPPED;
1257 wake_up(&thread->t_ctl_waitq);
1267 * Initializes the llog. Specific llog to be used depends on the type of the
1268 * target OSP represents (OST or MDT). The function adds appends a new llog
1269 * record to mark the place where the records associated with this boot
1272 * \param[in] env LU environment provided by the caller
1273 * \param[in] d OSP device
1275 * \retval 0 on success
1276 * \retval negative negated errno on error
1278 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1280 struct osp_thread_info *osi = osp_env_info(env);
1281 struct lu_fid *fid = &osi->osi_fid;
1282 struct llog_handle *lgh = NULL;
1283 struct obd_device *obd = d->opd_obd;
1284 struct llog_ctxt *ctxt;
1292 * open llog corresponding to our OST
1294 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1295 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1297 if (d->opd_connect_mdt)
1298 lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
1300 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1302 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1303 &osi->osi_cid, fid);
1305 if (rc != -EFAULT) {
1306 CERROR("%s: can't get id from catalogs: rc = %d\n",
1311 /* After sparse OST indices is supported, the CATALOG file
1312 * may become a sparse file that results in failure on
1313 * reading. Skip this error as the llog will be created
1315 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1319 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1320 obd->obd_name, d->opd_index,
1321 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1322 osi->osi_cid.lci_logid.lgl_ogen);
1324 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1325 &osp_mds_ost_orig_logops);
1329 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1332 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1333 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1335 /* re-create llog if it is missing */
1337 logid_set_id(&osi->osi_cid.lci_logid, 0);
1339 GOTO(out_cleanup, rc);
1342 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1343 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1345 GOTO(out_cleanup, rc);
1346 osi->osi_cid.lci_logid = lgh->lgh_id;
1349 LASSERT(lgh != NULL);
1350 ctxt->loc_handle = lgh;
1352 rc = llog_cat_init_and_process(env, lgh);
1354 GOTO(out_close, rc);
1356 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1357 &osi->osi_cid, fid);
1359 GOTO(out_close, rc);
1362 * put a mark in the llog till which we'll be processing
1363 * old records restless
1365 d->opd_syn_generation.mnt_cnt = cfs_time_current();
1366 d->opd_syn_generation.conn_cnt = cfs_time_current();
1368 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1369 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1371 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1372 sizeof(osi->osi_gen.lgr_gen));
1374 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1376 GOTO(out_close, rc);
1377 llog_ctxt_put(ctxt);
1380 llog_cat_close(env, lgh);
1382 llog_cleanup(env, ctxt);
1387 * Cleanup llog used for syncing.
1389 * Closes and cleanups the llog. The function is called when the device is
1392 * \param[in] env LU environment provided by the caller
1393 * \param[in] d OSP device
1395 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1397 struct llog_ctxt *ctxt;
1399 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1401 llog_cat_close(env, ctxt->loc_handle);
1402 llog_cleanup(env, ctxt);
1406 * Initialization of the sync component of OSP.
1408 * Initializes the llog and starts a new thread to handle the changes to
1409 * the remote target (OST or MDT).
1411 * \param[in] env LU environment provided by the caller
1412 * \param[in] d OSP device
1414 * \retval 0 on success
1415 * \retval negative negated errno on error
1417 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1419 struct l_wait_info lwi = { 0 };
1420 struct task_struct *task;
1425 rc = osp_sync_id_traction_init(d);
1430 * initialize llog storing changes
1432 rc = osp_sync_llog_init(env, d);
1434 CERROR("%s: can't initialize llog: rc = %d\n",
1435 d->opd_obd->obd_name, rc);
1440 * Start synchronization thread
1442 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1443 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1444 spin_lock_init(&d->opd_syn_lock);
1445 init_waitqueue_head(&d->opd_syn_waitq);
1446 init_waitqueue_head(&d->opd_syn_barrier_waitq);
1447 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1448 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1450 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1451 d->opd_index, d->opd_group);
1454 CERROR("%s: cannot start sync thread: rc = %d\n",
1455 d->opd_obd->obd_name, rc);
1459 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1460 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1464 osp_sync_llog_fini(env, d);
1466 osp_sync_id_traction_fini(d);
1471 * Stop the syncing thread.
1473 * Asks the syncing thread to stop and wait until it's stopped.
1475 * \param[in] d OSP device
1479 int osp_sync_fini(struct osp_device *d)
1481 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1485 thread->t_flags = SVC_STOPPING;
1486 wake_up(&d->opd_syn_waitq);
1487 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1490 * unregister transaction callbacks only when sync thread
1491 * has finished operations with llog
1493 osp_sync_id_traction_fini(d);
1498 static DEFINE_MUTEX(osp_id_tracker_sem);
1499 static struct list_head osp_id_tracker_list =
1500 LIST_HEAD_INIT(osp_id_tracker_list);
1503 * OSD commit callback.
1505 * The function is used as a local OSD commit callback to track the highest
1506 * committed llog record id. see osp_sync_id_traction_init() for the details.
1508 * \param[in] th local transaction handle committed
1509 * \param[in] cookie commit callback data (our private structure)
1511 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1513 struct osp_id_tracker *tr = cookie;
1514 struct osp_device *d;
1515 struct osp_txn_info *txn;
1519 txn = osp_txn_info(&th->th_ctx);
1520 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1523 spin_lock(&tr->otr_lock);
1524 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1525 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1526 tr->otr_committed_id, txn->oti_current_id);
1527 tr->otr_committed_id = txn->oti_current_id;
1529 list_for_each_entry(d, &tr->otr_wakeup_list,
1531 d->opd_syn_last_committed_id = tr->otr_committed_id;
1532 wake_up(&d->opd_syn_waitq);
1535 spin_unlock(&tr->otr_lock);
1539 * Initialize commit tracking mechanism.
1541 * Some setups may have thousands of OSTs and each will be represented by OSP.
1542 * Meaning order of magnitute many more changes to apply every second. In order
1543 * to keep the number of commit callbacks low this mechanism was introduced.
1544 * The mechanism is very similar to transno used by MDT service: it's an single
1545 * ID stream which can be assigned by any OSP to its llog records. The tricky
1546 * part is that ID is stored in per-transaction data and re-used by all the OSPs
1547 * involved in that transaction. Then all these OSPs are woken up utilizing a single OSD commit callback.
1549 * The function initializes the data used by the tracker described above.
1550 * A singler tracker per OSD device is created.
1552 * \param[in] d OSP device
1554 * \retval 0 on success
1555 * \retval negative negated errno on error
1557 static int osp_sync_id_traction_init(struct osp_device *d)
1559 struct osp_id_tracker *tr, *found = NULL;
1563 LASSERT(d->opd_storage);
1564 LASSERT(d->opd_syn_tracker == NULL);
1565 INIT_LIST_HEAD(&d->opd_syn_ontrack);
1567 mutex_lock(&osp_id_tracker_sem);
1568 list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1569 if (tr->otr_dev == d->opd_storage) {
1570 LASSERT(atomic_read(&tr->otr_refcount));
1571 atomic_inc(&tr->otr_refcount);
1572 d->opd_syn_tracker = tr;
1578 if (found == NULL) {
1582 d->opd_syn_tracker = tr;
1583 spin_lock_init(&tr->otr_lock);
1584 tr->otr_dev = d->opd_storage;
1585 tr->otr_next_id = 1;
1586 tr->otr_committed_id = 0;
1587 atomic_set(&tr->otr_refcount, 1);
1588 INIT_LIST_HEAD(&tr->otr_wakeup_list);
1589 list_add(&tr->otr_list, &osp_id_tracker_list);
1590 tr->otr_tx_cb.dtc_txn_commit =
1591 osp_sync_tracker_commit_cb;
1592 tr->otr_tx_cb.dtc_cookie = tr;
1593 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1594 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1598 mutex_unlock(&osp_id_tracker_sem);
1604 * Release commit tracker.
1606 * Decrease a refcounter on the tracker used by the given OSP device \a d.
1607 * If no more users left, then the tracker is released.
1609 * \param[in] d OSP device
1611 static void osp_sync_id_traction_fini(struct osp_device *d)
1613 struct osp_id_tracker *tr;
1618 tr = d->opd_syn_tracker;
1624 osp_sync_remove_from_tracker(d);
1626 mutex_lock(&osp_id_tracker_sem);
1627 if (atomic_dec_and_test(&tr->otr_refcount)) {
1628 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1629 LASSERT(list_empty(&tr->otr_wakeup_list));
1630 list_del(&tr->otr_list);
1632 d->opd_syn_tracker = NULL;
1634 mutex_unlock(&osp_id_tracker_sem);
1640 * Generate a new ID on a tracker.
1642 * Generates a new ID using the tracker associated with the given OSP device
1643 * \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
1644 * the wakeup list, so OSP won't miss when a transaction using the ID is
1645 * committed. Notice ID is 32bit, but llog doesn't support >2^32 records anyway.
1647 * \param[in] d OSP device
1648 * \param[in] id 0 or ID generated previously
1650 * \retval ID the caller should use
1652 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1654 struct osp_id_tracker *tr;
1656 tr = d->opd_syn_tracker;
1659 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1660 spin_lock(&tr->otr_lock);
1661 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1662 spin_unlock(&tr->otr_lock);
1663 CERROR("%s: next %u, last synced %lu\n",
1664 d->opd_obd->obd_name, tr->otr_next_id,
1665 d->opd_syn_last_used_id);
1670 id = tr->otr_next_id++;
1671 if (id > d->opd_syn_last_used_id)
1672 d->opd_syn_last_used_id = id;
1673 if (list_empty(&d->opd_syn_ontrack))
1674 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1675 spin_unlock(&tr->otr_lock);
1676 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1682 * Stop to propagate commit status to OSP.
1684 * If the OSP does not have any llog records she's waiting to commit, then
1685 * it is possible to unsubscribe from wakeups from the tracking using this
1688 * \param[in] d OSP device not willing to wakeup
1690 static void osp_sync_remove_from_tracker(struct osp_device *d)
1692 struct osp_id_tracker *tr;
1694 tr = d->opd_syn_tracker;
1697 if (list_empty(&d->opd_syn_ontrack))
1700 spin_lock(&tr->otr_lock);
1701 list_del_init(&d->opd_syn_ontrack);
1702 spin_unlock(&tr->otr_lock);