4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osp/osp_sync.c
38 * Lustre OST Proxy Device
40 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
41 * Author: Mikhail Pershin <mike.pershin@intel.com>
44 #define DEBUG_SUBSYSTEM S_MDS
46 #include <lustre_log.h>
47 #include <lustre_update.h>
48 #include "osp_internal.h"
50 static int osp_sync_id_traction_init(struct osp_device *d);
51 static void osp_sync_id_traction_fini(struct osp_device *d);
52 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
53 static void osp_sync_remove_from_tracker(struct osp_device *d);
56 * this is a components of OSP implementing synchronization between MDS and OST
57 * it llogs all interesting changes (currently it's uig/gid change and object
58 * destroy) atomically, then makes sure changes hit OST storage
60 * we have 4 queues of work:
62 * the first queue is llog itself, once read a change is stored in 2nd queue
63 * in form of RPC (but RPC isn't fired yet).
65 * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
66 * commit. once change is committed locally it migrates onto 3rd queue.
68 * the third queue (opd_syn_committed_here) holds changes committed locally,
69 * but not sent to OST (as the pipe can be full). once pipe becomes non-full
70 * we take a change from the queue and fire corresponded RPC.
72 * once RPC is reported committed by OST (using regular last_committed mech.)
73 * the change jumps into 4th queue (opd_syn_committed_there), now we can
74 * cancel corresponded llog record and release RPC
76 * opd_syn_changes is a number of unread llog records (to be processed).
77 * notice this number doesn't include llog records from previous boots.
78 * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
80 * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
81 * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
82 * too much memory -- how to deal with 1000th OSTs ? batching could help?
84 * opd_syn_rpc_in_flight is a number of RPC in flight.
85 * we control this with OSP_MAX_IN_FLIGHT
88 /* XXX: do math to learn reasonable threshold
89 * should it be ~ number of changes fitting bulk? */
91 #define OSP_SYN_THRESHOLD 10
92 #define OSP_MAX_IN_FLIGHT 8
93 #define OSP_MAX_IN_PROGRESS 4096
95 #define OSP_JOB_MAGIC 0x26112005
97 struct osp_job_req_args {
98 /** bytes reserved for ptlrpc_replay_req() */
99 struct ptlrpc_replay_async_args jra_raa;
100 struct list_head jra_link;
104 static inline int osp_sync_running(struct osp_device *d)
106 return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
110 * Check status: whether OSP thread has stopped
112 * \param[in] d OSP device
114 * \retval 0 still running
117 static inline int osp_sync_stopped(struct osp_device *d)
119 return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
123 ** Check for new changes to sync
125 * \param[in] d OSP device
127 * \retval 1 there are changes
128 * \retval 0 there are no changes
130 static inline int osp_sync_has_new_job(struct osp_device *d)
132 return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
133 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
134 || (d->opd_syn_prev_done == 0);
137 static inline int osp_sync_low_in_progress(struct osp_device *d)
139 return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
143 * Check for room in the network pipe to OST
145 * \param[in] d OSP device
147 * \retval 1 there is room
148 * \retval 0 no room, the pipe is full
150 static inline int osp_sync_low_in_flight(struct osp_device *d)
152 return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
156 * Wake up check for the main sync thread
158 * \param[in] d OSP device
160 * \retval 1 time to wake up
161 * \retval 0 no need to wake up
163 static inline int osp_sync_has_work(struct osp_device *d)
165 /* has new/old changes and low in-progress? */
166 if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
167 osp_sync_low_in_flight(d) && d->opd_imp_connected)
170 /* has remotely committed? */
171 if (!list_empty(&d->opd_syn_committed_there))
177 #define osp_sync_check_for_work(d) \
179 if (osp_sync_has_work(d)) { \
180 wake_up(&d->opd_syn_waitq); \
184 void __osp_sync_check_for_work(struct osp_device *d)
186 osp_sync_check_for_work(d);
190 * Check and return ready-for-new status.
192 * The thread processing llog record uses this function to check whether
193 * it's time to take another record and process it. The number of conditions
194 * must be met: the connection should be ready, RPCs in flight not exceeding
195 * the limit, the record is committed locally, etc (see the lines below).
197 * \param[in] d OSP device
198 * \param[in] rec next llog record to process
200 * \retval 0 not ready
203 static inline int osp_sync_can_process_new(struct osp_device *d,
204 struct llog_rec_hdr *rec)
208 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
210 if (!osp_sync_low_in_progress(d))
212 if (!osp_sync_low_in_flight(d))
214 if (!d->opd_imp_connected)
216 if (d->opd_syn_prev_done == 0)
218 if (d->opd_syn_changes == 0)
220 if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
226 * Declare intention to add a new change.
228 * With regard to OSD API, we have to declare any changes ahead. In this
229 * case we declare an intention to add a llog record representing the
230 * change on the local storage.
232 * \param[in] env LU environment provided by the caller
233 * \param[in] o OSP object
234 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
235 * \param[in] th transaction handle (local)
237 * \retval 0 on success
238 * \retval negative negated errno on error
240 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
241 llog_op_type type, struct thandle *th)
243 struct osp_thread_info *osi = osp_env_info(env);
244 struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
245 struct llog_ctxt *ctxt;
250 /* it's a layering violation, to access internals of th,
251 * but we can do this as a sanity check, for a while */
252 LASSERT(th->th_dev == d->opd_storage);
255 case MDS_UNLINK64_REC:
256 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
258 case MDS_SETATTR64_REC:
259 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
265 /* we want ->dt_trans_start() to allocate per-thandle structure */
266 th->th_tags |= LCT_OSP_THREAD;
268 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
271 rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
278 * Generate a llog record for a given change.
280 * Generates a llog record for the change passed. The change can be of two
281 * types: unlink and setattr. The record gets an ID which later will be
282 * used to track commit status of the change. For unlink changes, the caller
283 * can supply a starting FID and the count of the objects to destroy. For
284 * setattr the caller should apply attributes to apply.
287 * \param[in] env LU environment provided by the caller
288 * \param[in] d OSP device
289 * \param[in] fid fid of the object the change should be applied to
290 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
291 * \param[in] count count of objects to destroy
292 * \param[in] th transaction handle (local)
293 * \param[in] attr attributes for setattr
295 * \retval 0 on success
296 * \retval negative negated errno on error
298 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
299 const struct lu_fid *fid, llog_op_type type,
300 int count, struct thandle *th,
301 const struct lu_attr *attr)
303 struct osp_thread_info *osi = osp_env_info(env);
304 struct llog_ctxt *ctxt;
305 struct osp_txn_info *txn;
310 /* it's a layering violation, to access internals of th,
311 * but we can do this as a sanity check, for a while */
312 LASSERT(th->th_dev == d->opd_storage);
315 case MDS_UNLINK64_REC:
316 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
317 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
318 osi->osi_unlink.lur_fid = *fid;
319 osi->osi_unlink.lur_count = count;
321 case MDS_SETATTR64_REC:
322 rc = fid_to_ostid(fid, &osi->osi_oi);
324 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
325 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
326 osi->osi_setattr.lsr_oi = osi->osi_oi;
328 osi->osi_setattr.lsr_uid = attr->la_uid;
329 osi->osi_setattr.lsr_gid = attr->la_gid;
330 osi->osi_setattr.lsr_valid =
331 ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
332 ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0);
338 txn = osp_txn_info(&th->th_ctx);
341 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
342 osi->osi_hdr.lrh_id = txn->oti_current_id;
344 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 if (likely(rc >= 0)) {
353 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
354 d->opd_obd->obd_name,
355 POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
356 (unsigned long)osi->osi_cookie.lgc_lgl.lgl_ogen,
357 (unsigned long)osi->osi_cookie.lgc_index, rc);
358 spin_lock(&d->opd_syn_lock);
359 d->opd_syn_changes++;
360 spin_unlock(&d->opd_syn_lock);
362 /* return 0 always here, error case just cause no llog record */
366 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
367 llog_op_type type, struct thandle *th,
368 const struct lu_attr *attr)
370 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
371 lu_object_fid(&o->opo_obj.do_lu), type, 1,
375 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
376 struct lu_fid *fid, int lost, struct thandle *th)
378 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
382 * it's quite obvious we can't maintain all the structures in the memory:
383 * while OST is down, MDS can be processing thousands and thousands of unlinks
384 * filling persistent llogs and in-core respresentation
386 * this doesn't scale at all. so we need basically the following:
387 * a) destroy/setattr append llog records
388 * b) once llog has grown to X records, we process first Y committed records
390 * once record R is found via llog_process(), it becomes committed after any
391 * subsequent commit callback (at the most)
395 * ptlrpc commit callback.
397 * The callback is called by PTLRPC when a RPC is reported committed by the
398 * target (OST). We register the callback for the every RPC applying a change
399 * from the llog. This way we know then the llog records can be cancelled.
400 * Notice the callback can be called when OSP is finishing. We can detect this
401 * checking that actual transno in the request is less or equal of known
402 * committed transno (see osp_sync_process_committed() for the details).
403 * XXX: this is pretty expensive and can be improved later using batching.
405 * \param[in] req request
407 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
409 struct osp_device *d = req->rq_cb_data;
410 struct osp_job_req_args *jra;
412 CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
414 if (unlikely(req->rq_transno == 0))
417 /* do not do any opd_dyn_rpc_* accounting here
418 * it's done in osp_sync_interpret sooner or later */
421 jra = ptlrpc_req_async_args(req);
422 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
423 LASSERT(list_empty(&jra->jra_link));
425 ptlrpc_request_addref(req);
427 spin_lock(&d->opd_syn_lock);
428 list_add(&jra->jra_link, &d->opd_syn_committed_there);
429 spin_unlock(&d->opd_syn_lock);
431 /* XXX: some batching wouldn't hurt */
432 wake_up(&d->opd_syn_waitq);
436 * RPC interpretation callback.
438 * The callback is called by ptlrpc when RPC is replied. Now we have to decide
440 * - put request on a special list to wait until it's committed by the target,
441 * if the request is succesful
442 * - schedule llog record cancel if no target object is found
443 * - try later (essentially after reboot) in case of unexpected error
445 * \param[in] env LU environment provided by the caller
446 * \param[in] req request replied
447 * \param[in] aa callback data
448 * \param[in] rc result of RPC
452 static int osp_sync_interpret(const struct lu_env *env,
453 struct ptlrpc_request *req, void *aa, int rc)
455 struct osp_device *d = req->rq_cb_data;
456 struct osp_job_req_args *jra = aa;
458 if (jra->jra_magic != OSP_JOB_MAGIC) {
459 DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
464 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
465 atomic_read(&req->rq_refcount),
466 rc, (unsigned) req->rq_transno);
467 LASSERT(rc || req->rq_transno);
471 * we tried to destroy object or update attributes,
472 * but object doesn't exist anymore - cancell llog record
474 LASSERT(req->rq_transno == 0);
475 LASSERT(list_empty(&jra->jra_link));
477 ptlrpc_request_addref(req);
479 spin_lock(&d->opd_syn_lock);
480 list_add(&jra->jra_link, &d->opd_syn_committed_there);
481 spin_unlock(&d->opd_syn_lock);
483 wake_up(&d->opd_syn_waitq);
485 struct obd_import *imp = req->rq_import;
487 * error happened, we'll try to repeat on next boot ?
489 LASSERTF(req->rq_transno == 0 ||
490 req->rq_import_generation < imp->imp_generation,
491 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
492 req->rq_transno, rc, req->rq_import_generation,
493 imp->imp_generation);
494 if (req->rq_transno == 0) {
495 /* this is the last time we see the request
496 * if transno is not zero, then commit cb
497 * will be called at some point */
498 LASSERT(d->opd_syn_rpc_in_progress > 0);
499 spin_lock(&d->opd_syn_lock);
500 d->opd_syn_rpc_in_progress--;
501 spin_unlock(&d->opd_syn_lock);
504 wake_up(&d->opd_syn_waitq);
505 } else if (d->opd_pre != NULL &&
506 unlikely(d->opd_pre_status == -ENOSPC)) {
508 * if current status is -ENOSPC (lack of free space on OST)
509 * then we should poll OST immediately once object destroy
512 osp_statfs_need_now(d);
515 LASSERT(d->opd_syn_rpc_in_flight > 0);
516 spin_lock(&d->opd_syn_lock);
517 d->opd_syn_rpc_in_flight--;
518 spin_unlock(&d->opd_syn_lock);
519 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
520 wake_up(&d->opd_syn_barrier_waitq);
521 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
522 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
523 d->opd_syn_rpc_in_progress);
525 osp_sync_check_for_work(d);
531 ** Add request to ptlrpc queue.
533 * This is just a tiny helper function to put the request on the sending list
535 * \param[in] d OSP device
536 * \param[in] req request
538 static void osp_sync_send_new_rpc(struct osp_device *d,
539 struct ptlrpc_request *req)
541 struct osp_job_req_args *jra;
543 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
545 jra = ptlrpc_req_async_args(req);
546 jra->jra_magic = OSP_JOB_MAGIC;
547 INIT_LIST_HEAD(&jra->jra_link);
549 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
554 * Allocate and prepare RPC for a new change.
556 * The function allocates and initializes an RPC which will be sent soon to
557 * apply the change to the target OST. The request is initialized from the
558 * llog record passed. Notice only the fields common to all type of changes
561 * \param[in] d OSP device
562 * \param[in] llh llog handle where the record is stored
563 * \param[in] h llog record
564 * \param[in] op type of the change
565 * \param[in] format request format to be used
567 * \retval pointer new request on success
568 * \retval ERR_PTR(errno) on error
570 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
571 struct llog_handle *llh,
572 struct llog_rec_hdr *h,
574 const struct req_format *format)
576 struct ptlrpc_request *req;
577 struct ost_body *body;
578 struct obd_import *imp;
581 /* Prepare the request */
582 imp = d->opd_obd->u.cli.cl_import;
584 req = ptlrpc_request_alloc(imp, format);
586 RETURN(ERR_PTR(-ENOMEM));
588 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
590 ptlrpc_req_finished(req);
595 * this is a trick: to save on memory allocations we put cookie
596 * into the request, but don't set corresponded flag in o_valid
597 * so that OST doesn't interpret this cookie. once the request
598 * is committed on OST we take cookie from the request and cancel
600 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
602 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
603 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
604 body->oa.o_lcookie.lgc_index = h->lrh_index;
606 req->rq_interpret_reply = osp_sync_interpret;
607 req->rq_commit_cb = osp_sync_request_commit_cb;
610 ptlrpc_request_set_replen(req);
616 * Generate a request for setattr change.
618 * The function prepares a new RPC, initializes it with setattr specific
619 * bits and send the RPC.
621 * \param[in] d OSP device
622 * \param[in] llh llog handle where the record is stored
623 * \param[in] h llog record
625 * \retval 0 on success
626 * \retval negative negated errno on error
628 static int osp_sync_new_setattr_job(struct osp_device *d,
629 struct llog_handle *llh,
630 struct llog_rec_hdr *h)
632 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
633 struct ptlrpc_request *req;
634 struct ost_body *body;
637 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
639 /* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
640 * so no bits other than these should be set. */
641 if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
642 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
643 d->opd_obd->obd_name, rec->lsr_valid);
644 /* return 0 so that sync thread can continue processing
649 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
651 RETURN(PTR_ERR(req));
653 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
655 body->oa.o_oi = rec->lsr_oi;
656 body->oa.o_uid = rec->lsr_uid;
657 body->oa.o_gid = rec->lsr_gid;
658 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
659 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
660 * we assume that both UID and GID are valid in that case. */
661 if (rec->lsr_valid == 0)
662 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
664 body->oa.o_valid |= rec->lsr_valid;
666 osp_sync_send_new_rpc(d, req);
671 * Generate a request for unlink change.
673 * The function prepares a new RPC, initializes it with unlink(destroy)
674 * specific bits and sends the RPC. The function is used to handle
675 * llog_unlink_rec which were used in the older versions of Lustre.
676 * Current version uses llog_unlink_rec64.
678 * \param[in] d OSP device
679 * \param[in] llh llog handle where the record is stored
680 * \param[in] h llog record
682 * \retval 0 on success
683 * \retval negative negated errno on error
685 static int osp_sync_new_unlink_job(struct osp_device *d,
686 struct llog_handle *llh,
687 struct llog_rec_hdr *h)
689 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
690 struct ptlrpc_request *req;
691 struct ost_body *body;
694 LASSERT(h->lrh_type == MDS_UNLINK_REC);
696 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
698 RETURN(PTR_ERR(req));
700 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
702 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
703 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
704 body->oa.o_misc = rec->lur_count;
705 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
707 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
709 osp_sync_send_new_rpc(d, req);
714 * Prepare OUT-based object destroy RPC.
716 * The function allocates a new RPC with OUT format. Then initializes the RPC
717 * to contain OUT_DESTROY update against the object specified in the llog
718 * record provided by the caller.
720 * \param[in] env LU environment provided by the caller
721 * \param[in] osp OSP device
722 * \param[in] llh llog handle where the record is stored
723 * \param[in] h llog record
724 * \param[out] reqp request prepared
726 * \retval 0 on success
727 * \retval negative negated errno on error
729 static int osp_prep_unlink_update_req(const struct lu_env *env,
730 struct osp_device *osp,
731 struct llog_handle *llh,
732 struct llog_rec_hdr *h,
733 struct ptlrpc_request **reqp)
735 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
736 struct dt_update_request *update = NULL;
737 struct ptlrpc_request *req;
738 struct llog_cookie lcookie;
744 update = dt_update_request_create(&osp->opd_dt_dev);
746 RETURN(PTR_ERR(update));
748 /* This can only happens for unlink slave directory, so decrease
749 * ref for ".." and "." */
750 rc = out_update_pack(env, &update->dur_buf, OUT_REF_DEL, &rec->lur_fid,
755 rc = out_update_pack(env, &update->dur_buf, OUT_REF_DEL, &rec->lur_fid,
760 lcookie.lgc_lgl = llh->lgh_id;
761 lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
762 lcookie.lgc_index = h->lrh_index;
763 size = sizeof(lcookie);
766 rc = out_update_pack(env, &update->dur_buf, OUT_DESTROY, &rec->lur_fid,
771 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
772 update->dur_buf.ub_req, &req);
776 req->rq_interpret_reply = osp_sync_interpret;
777 req->rq_commit_cb = osp_sync_request_commit_cb;
778 req->rq_cb_data = osp;
780 ptlrpc_request_set_replen(req);
784 dt_update_request_destroy(update);
790 * Generate a request for unlink change.
792 * The function prepares a new RPC, initializes it with unlink(destroy)
793 * specific bits and sends the RPC. Depending on the target (MDT or OST)
794 * two different protocols are used. For MDT we use OUT (basically OSD API
795 * updates transferred via a network). For OST we still use the old
796 * protocol (OBD?), originally for compatibility. Later we can start to
797 * use OUT for OST as well, this will allow batching and better code
800 * \param[in] env LU environment provided by the caller
801 * \param[in] d OSP device
802 * \param[in] llh llog handle where the record is stored
803 * \param[in] h llog record
805 * \retval 0 on success
806 * \retval negative negated errno on error
808 static int osp_sync_new_unlink64_job(const struct lu_env *env,
809 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);
821 if (d->opd_connect_mdt) {
822 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
826 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
829 RETURN(PTR_ERR(req));
831 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
834 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
837 body->oa.o_misc = rec->lur_count;
838 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
841 osp_sync_send_new_rpc(d, req);
846 * Process llog records.
848 * This function is called to process the llog records committed locally.
849 * In the recovery model used by OSP we can apply a change to a remote
850 * target once corresponding transaction (like posix unlink) is committed
851 * locally so can't revert.
852 * Depending on the llog record type, a given handler is called that is
853 * responsible for preparing and sending the RPC to apply the change.
854 * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
856 * \param[in] env LU environment provided by the caller
857 * \param[in] d OSP device
858 * \param[in] llh llog handle where the record is stored
859 * \param[in] rec llog record
861 * \retval 0 on success
862 * \retval negative negated errno on error
864 static int osp_sync_process_record(const struct lu_env *env,
865 struct osp_device *d,
866 struct llog_handle *llh,
867 struct llog_rec_hdr *rec)
869 struct llog_cookie cookie;
872 cookie.lgc_lgl = llh->lgh_id;
873 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
874 cookie.lgc_index = rec->lrh_index;
876 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
877 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
879 /* we're waiting for the record generated by this instance */
880 LASSERT(d->opd_syn_prev_done == 0);
881 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
882 sizeof(gen->lgr_gen))) {
883 CDEBUG(D_HA, "processed all old entries\n");
884 d->opd_syn_prev_done = 1;
887 /* cancel any generation record */
888 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
895 * now we prepare and fill requests to OST, put them on the queue
896 * and fire after next commit callback
899 /* notice we increment counters before sending RPC, to be consistent
900 * in RPC interpret callback which may happen very quickly */
901 spin_lock(&d->opd_syn_lock);
902 d->opd_syn_rpc_in_flight++;
903 d->opd_syn_rpc_in_progress++;
904 spin_unlock(&d->opd_syn_lock);
906 switch (rec->lrh_type) {
907 /* case MDS_UNLINK_REC is kept for compatibility */
909 rc = osp_sync_new_unlink_job(d, llh, rec);
911 case MDS_UNLINK64_REC:
912 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
914 case MDS_SETATTR64_REC:
915 rc = osp_sync_new_setattr_job(d, llh, rec);
918 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
920 /* we should continue processing */
923 /* rc > 0 means sync RPC being added to the queue */
924 if (likely(rc > 0)) {
925 spin_lock(&d->opd_syn_lock);
926 if (d->opd_syn_prev_done) {
927 LASSERT(d->opd_syn_changes > 0);
928 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
930 * NOTE: it's possible to meet same id if
931 * OST stores few stripes of same file
933 if (rec->lrh_id > d->opd_syn_last_processed_id) {
934 d->opd_syn_last_processed_id = rec->lrh_id;
935 wake_up(&d->opd_syn_barrier_waitq);
938 d->opd_syn_changes--;
940 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
941 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
942 d->opd_syn_rpc_in_progress);
943 spin_unlock(&d->opd_syn_lock);
946 spin_lock(&d->opd_syn_lock);
947 d->opd_syn_rpc_in_flight--;
948 d->opd_syn_rpc_in_progress--;
949 spin_unlock(&d->opd_syn_lock);
952 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
953 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
958 * Cancel llog records for the committed changes.
960 * The function walks through the list of the committed RPCs and cancels
961 * corresponding llog records. see osp_sync_request_commit_cb() for the
964 * \param[in] env LU environment provided by the caller
965 * \param[in] d OSP device
967 static void osp_sync_process_committed(const struct lu_env *env,
968 struct osp_device *d)
970 struct obd_device *obd = d->opd_obd;
971 struct obd_import *imp = obd->u.cli.cl_import;
972 struct ost_body *body;
973 struct ptlrpc_request *req;
974 struct llog_ctxt *ctxt;
975 struct llog_handle *llh;
976 struct list_head list;
981 if (list_empty(&d->opd_syn_committed_there))
985 * if current status is -ENOSPC (lack of free space on OST)
986 * then we should poll OST immediately once object destroy
988 * notice: we do this upon commit as well because some backends
989 * (like DMU) do not release space right away.
991 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
992 osp_statfs_need_now(d);
995 * now cancel them all
996 * XXX: can we improve this using some batching?
997 * with batch RPC that'll happen automatically?
998 * XXX: can we store ctxt in lod_device and save few cycles ?
1000 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1003 llh = ctxt->loc_handle;
1006 INIT_LIST_HEAD(&list);
1007 spin_lock(&d->opd_syn_lock);
1008 list_splice(&d->opd_syn_committed_there, &list);
1009 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1010 spin_unlock(&d->opd_syn_lock);
1012 while (!list_empty(&list)) {
1013 struct llog_cookie *lcookie = NULL;
1014 struct osp_job_req_args *jra;
1016 jra = list_entry(list.next, struct osp_job_req_args, jra_link);
1017 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1018 list_del_init(&jra->jra_link);
1020 req = container_of((void *)jra, struct ptlrpc_request,
1022 if (d->opd_connect_mdt) {
1023 struct object_update_request *ureq;
1024 struct object_update *update;
1025 ureq = req_capsule_client_get(&req->rq_pill,
1027 LASSERT(ureq != NULL &&
1028 ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
1030 /* 1st/2nd is for decref . and .., 3rd one is for
1031 * destroy, where the log cookie is stored.
1032 * See osp_prep_unlink_update_req */
1033 update = object_update_request_get(ureq, 2, NULL);
1034 LASSERT(update != NULL);
1035 lcookie = object_update_param_get(update, 0, NULL);
1036 LASSERT(lcookie != NULL);
1038 body = req_capsule_client_get(&req->rq_pill,
1041 lcookie = &body->oa.o_lcookie;
1043 /* import can be closing, thus all commit cb's are
1044 * called we can check committness directly */
1045 if (req->rq_transno <= imp->imp_peer_committed_transno) {
1046 rc = llog_cat_cancel_records(env, llh, 1, lcookie);
1048 CERROR("%s: can't cancel record: %d\n",
1051 DEBUG_REQ(D_HA, req, "not committed");
1054 ptlrpc_req_finished(req);
1058 llog_ctxt_put(ctxt);
1060 LASSERT(d->opd_syn_rpc_in_progress >= done);
1061 spin_lock(&d->opd_syn_lock);
1062 d->opd_syn_rpc_in_progress -= done;
1063 spin_unlock(&d->opd_syn_lock);
1064 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
1065 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
1066 d->opd_syn_rpc_in_progress);
1068 osp_sync_check_for_work(d);
1070 /* wake up the thread if requested to stop:
1071 * it might be waiting for in-progress to complete */
1072 if (unlikely(osp_sync_running(d) == 0))
1073 wake_up(&d->opd_syn_waitq);
1079 * The core of the syncing mechanism.
1081 * This is a callback called by the llog processing function. Essentially it
1082 * suspends llog processing until there is a record to process (it's supposed
1083 * to be committed locally). The function handles RPCs committed by the target
1084 * and cancels corresponding llog records.
1086 * \param[in] env LU environment provided by the caller
1087 * \param[in] llh llog handle we're processing
1088 * \param[in] rec current llog record
1089 * \param[in] data callback data containing a pointer to the device
1091 * \retval 0 to ask the caller (llog_process()) to continue
1092 * \retval LLOG_PROC_BREAK to ask the caller to break
1094 static int osp_sync_process_queues(const struct lu_env *env,
1095 struct llog_handle *llh,
1096 struct llog_rec_hdr *rec,
1099 struct osp_device *d = data;
1103 struct l_wait_info lwi = { 0 };
1105 if (!osp_sync_running(d)) {
1106 CDEBUG(D_HA, "stop llog processing\n");
1107 return LLOG_PROC_BREAK;
1110 /* process requests committed by OST */
1111 osp_sync_process_committed(env, d);
1113 /* if we there are changes to be processed and we have
1114 * resources for this ... do now */
1115 if (osp_sync_can_process_new(d, rec)) {
1117 /* ask llog for another record */
1118 CDEBUG(D_HA, "%lu changes, %u in progress,"
1121 d->opd_syn_rpc_in_progress,
1122 d->opd_syn_rpc_in_flight);
1127 * try to send, in case of disconnection, suspend
1128 * processing till we can send this request
1131 rc = osp_sync_process_record(env, d, llh, rec);
1133 * XXX: probably different handling is needed
1134 * for some bugs, like immediate exit or if
1138 CERROR("can't send: %d\n", rc);
1139 l_wait_event(d->opd_syn_waitq,
1140 !osp_sync_running(d) ||
1141 osp_sync_has_work(d),
1144 } while (rc != 0 && osp_sync_running(d));
1150 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
1151 osp_sync_remove_from_tracker(d);
1153 l_wait_event(d->opd_syn_waitq,
1154 !osp_sync_running(d) ||
1155 osp_sync_can_process_new(d, rec) ||
1156 !list_empty(&d->opd_syn_committed_there),
1164 * This thread runs llog_cat_process() scanner calling our callback
1165 * to process llog records. in the callback we implement tricky
1166 * state machine as we don't want to start scanning of the llog again
1167 * and again, also we don't want to process too many records and send
1168 * too many RPCs a time. so, depending on current load (num of changes
1169 * being synced to OST) the callback can suspend awaiting for some
1170 * new conditions, like syncs completed.
1172 * In order to process llog records left by previous boots and to allow
1173 * llog_process_thread() to find something (otherwise it'd just exit
1174 * immediately) we add a special GENERATATION record on each boot.
1176 * \param[in] _arg a pointer to thread's arguments
1178 * \retval 0 on success
1179 * \retval negative negated errno on error
1181 static int osp_sync_thread(void *_arg)
1183 struct osp_device *d = _arg;
1184 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1185 struct l_wait_info lwi = { 0 };
1186 struct llog_ctxt *ctxt;
1187 struct obd_device *obd = d->opd_obd;
1188 struct llog_handle *llh;
1194 rc = lu_env_init(&env, LCT_LOCAL);
1196 CERROR("%s: can't initialize env: rc = %d\n",
1201 spin_lock(&d->opd_syn_lock);
1202 thread->t_flags = SVC_RUNNING;
1203 spin_unlock(&d->opd_syn_lock);
1204 wake_up(&thread->t_ctl_waitq);
1206 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1208 CERROR("can't get appropriate context\n");
1209 GOTO(out, rc = -EINVAL);
1212 llh = ctxt->loc_handle;
1214 CERROR("can't get llh\n");
1215 llog_ctxt_put(ctxt);
1216 GOTO(out, rc = -EINVAL);
1219 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
1220 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1221 "%lu changes, %u in progress, %u in flight: %d\n",
1222 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1223 d->opd_syn_rpc_in_flight, rc);
1225 /* we don't expect llog_process_thread() to exit till umount */
1226 LASSERTF(thread->t_flags != SVC_RUNNING,
1227 "%lu changes, %u in progress, %u in flight\n",
1228 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1229 d->opd_syn_rpc_in_flight);
1231 /* wait till all the requests are completed */
1233 while (d->opd_syn_rpc_in_progress > 0) {
1234 osp_sync_process_committed(&env, d);
1236 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1237 rc = l_wait_event(d->opd_syn_waitq,
1238 d->opd_syn_rpc_in_progress == 0,
1240 if (rc == -ETIMEDOUT)
1242 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1243 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1244 d->opd_syn_rpc_in_flight,
1245 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1249 llog_cat_close(&env, llh);
1250 rc = llog_cleanup(&env, ctxt);
1252 CERROR("can't cleanup llog: %d\n", rc);
1254 LASSERTF(d->opd_syn_rpc_in_progress == 0,
1255 "%s: %d %d %sempty\n",
1256 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1257 d->opd_syn_rpc_in_flight,
1258 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1260 thread->t_flags = SVC_STOPPED;
1262 wake_up(&thread->t_ctl_waitq);
1272 * Initializes the llog. Specific llog to be used depends on the type of the
1273 * target OSP represents (OST or MDT). The function adds appends a new llog
1274 * record to mark the place where the records associated with this boot
1277 * \param[in] env LU environment provided by the caller
1278 * \param[in] d OSP device
1280 * \retval 0 on success
1281 * \retval negative negated errno on error
1283 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1285 struct osp_thread_info *osi = osp_env_info(env);
1286 struct lu_fid *fid = &osi->osi_fid;
1287 struct llog_handle *lgh = NULL;
1288 struct obd_device *obd = d->opd_obd;
1289 struct llog_ctxt *ctxt;
1297 * open llog corresponding to our OST
1299 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1300 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1302 if (d->opd_connect_mdt)
1303 lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
1305 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1307 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1308 &osi->osi_cid, fid);
1310 if (rc != -EFAULT) {
1311 CERROR("%s: can't get id from catalogs: rc = %d\n",
1316 /* After sparse OST indices is supported, the CATALOG file
1317 * may become a sparse file that results in failure on
1318 * reading. Skip this error as the llog will be created
1320 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1324 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1325 obd->obd_name, d->opd_index,
1326 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1327 osi->osi_cid.lci_logid.lgl_ogen);
1329 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1330 &osp_mds_ost_orig_logops);
1334 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1337 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1338 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1340 /* re-create llog if it is missing */
1342 logid_set_id(&osi->osi_cid.lci_logid, 0);
1344 GOTO(out_cleanup, rc);
1347 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1348 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1350 GOTO(out_cleanup, rc);
1351 osi->osi_cid.lci_logid = lgh->lgh_id;
1354 LASSERT(lgh != NULL);
1355 ctxt->loc_handle = lgh;
1357 rc = llog_cat_init_and_process(env, lgh);
1359 GOTO(out_close, rc);
1361 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1362 &osi->osi_cid, fid);
1364 GOTO(out_close, rc);
1367 * put a mark in the llog till which we'll be processing
1368 * old records restless
1370 d->opd_syn_generation.mnt_cnt = cfs_time_current();
1371 d->opd_syn_generation.conn_cnt = cfs_time_current();
1373 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1374 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1376 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1377 sizeof(osi->osi_gen.lgr_gen));
1379 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1381 GOTO(out_close, rc);
1382 llog_ctxt_put(ctxt);
1385 llog_cat_close(env, lgh);
1387 llog_cleanup(env, ctxt);
1392 * Cleanup llog used for syncing.
1394 * Closes and cleanups the llog. The function is called when the device is
1397 * \param[in] env LU environment provided by the caller
1398 * \param[in] d OSP device
1400 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1402 struct llog_ctxt *ctxt;
1404 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1406 llog_cat_close(env, ctxt->loc_handle);
1407 llog_cleanup(env, ctxt);
1411 * Initialization of the sync component of OSP.
1413 * Initializes the llog and starts a new thread to handle the changes to
1414 * the remote target (OST or MDT).
1416 * \param[in] env LU environment provided by the caller
1417 * \param[in] d OSP device
1419 * \retval 0 on success
1420 * \retval negative negated errno on error
1422 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1424 struct l_wait_info lwi = { 0 };
1425 struct task_struct *task;
1430 rc = osp_sync_id_traction_init(d);
1435 * initialize llog storing changes
1437 rc = osp_sync_llog_init(env, d);
1439 CERROR("%s: can't initialize llog: rc = %d\n",
1440 d->opd_obd->obd_name, rc);
1445 * Start synchronization thread
1447 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1448 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1449 spin_lock_init(&d->opd_syn_lock);
1450 init_waitqueue_head(&d->opd_syn_waitq);
1451 init_waitqueue_head(&d->opd_syn_barrier_waitq);
1452 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1453 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1455 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1456 d->opd_index, d->opd_group);
1459 CERROR("%s: cannot start sync thread: rc = %d\n",
1460 d->opd_obd->obd_name, rc);
1464 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1465 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1469 osp_sync_llog_fini(env, d);
1471 osp_sync_id_traction_fini(d);
1476 * Stop the syncing thread.
1478 * Asks the syncing thread to stop and wait until it's stopped.
1480 * \param[in] d OSP device
1484 int osp_sync_fini(struct osp_device *d)
1486 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1490 thread->t_flags = SVC_STOPPING;
1491 wake_up(&d->opd_syn_waitq);
1492 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1495 * unregister transaction callbacks only when sync thread
1496 * has finished operations with llog
1498 osp_sync_id_traction_fini(d);
1503 static DEFINE_MUTEX(osp_id_tracker_sem);
1504 static struct list_head osp_id_tracker_list =
1505 LIST_HEAD_INIT(osp_id_tracker_list);
1508 * OSD commit callback.
1510 * The function is used as a local OSD commit callback to track the highest
1511 * committed llog record id. see osp_sync_id_traction_init() for the details.
1513 * \param[in] th local transaction handle committed
1514 * \param[in] cookie commit callback data (our private structure)
1516 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1518 struct osp_id_tracker *tr = cookie;
1519 struct osp_device *d;
1520 struct osp_txn_info *txn;
1524 txn = osp_txn_info(&th->th_ctx);
1525 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1528 spin_lock(&tr->otr_lock);
1529 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1530 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1531 tr->otr_committed_id, txn->oti_current_id);
1532 tr->otr_committed_id = txn->oti_current_id;
1534 list_for_each_entry(d, &tr->otr_wakeup_list,
1536 d->opd_syn_last_committed_id = tr->otr_committed_id;
1537 wake_up(&d->opd_syn_waitq);
1540 spin_unlock(&tr->otr_lock);
1544 * Initialize commit tracking mechanism.
1546 * Some setups may have thousands of OSTs and each will be represented by OSP.
1547 * Meaning order of magnitute many more changes to apply every second. In order
1548 * to keep the number of commit callbacks low this mechanism was introduced.
1549 * The mechanism is very similar to transno used by MDT service: it's an single
1550 * ID stream which can be assigned by any OSP to its llog records. The tricky
1551 * part is that ID is stored in per-transaction data and re-used by all the OSPs
1552 * involved in that transaction. Then all these OSPs are woken up utilizing a single OSD commit callback.
1554 * The function initializes the data used by the tracker described above.
1555 * A singler tracker per OSD device is created.
1557 * \param[in] d OSP device
1559 * \retval 0 on success
1560 * \retval negative negated errno on error
1562 static int osp_sync_id_traction_init(struct osp_device *d)
1564 struct osp_id_tracker *tr, *found = NULL;
1568 LASSERT(d->opd_storage);
1569 LASSERT(d->opd_syn_tracker == NULL);
1570 INIT_LIST_HEAD(&d->opd_syn_ontrack);
1572 mutex_lock(&osp_id_tracker_sem);
1573 list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1574 if (tr->otr_dev == d->opd_storage) {
1575 LASSERT(atomic_read(&tr->otr_refcount));
1576 atomic_inc(&tr->otr_refcount);
1577 d->opd_syn_tracker = tr;
1583 if (found == NULL) {
1587 d->opd_syn_tracker = tr;
1588 spin_lock_init(&tr->otr_lock);
1589 tr->otr_dev = d->opd_storage;
1590 tr->otr_next_id = 1;
1591 tr->otr_committed_id = 0;
1592 atomic_set(&tr->otr_refcount, 1);
1593 INIT_LIST_HEAD(&tr->otr_wakeup_list);
1594 list_add(&tr->otr_list, &osp_id_tracker_list);
1595 tr->otr_tx_cb.dtc_txn_commit =
1596 osp_sync_tracker_commit_cb;
1597 tr->otr_tx_cb.dtc_cookie = tr;
1598 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1599 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1603 mutex_unlock(&osp_id_tracker_sem);
1609 * Release commit tracker.
1611 * Decrease a refcounter on the tracker used by the given OSP device \a d.
1612 * If no more users left, then the tracker is released.
1614 * \param[in] d OSP device
1616 static void osp_sync_id_traction_fini(struct osp_device *d)
1618 struct osp_id_tracker *tr;
1623 tr = d->opd_syn_tracker;
1629 osp_sync_remove_from_tracker(d);
1631 mutex_lock(&osp_id_tracker_sem);
1632 if (atomic_dec_and_test(&tr->otr_refcount)) {
1633 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1634 LASSERT(list_empty(&tr->otr_wakeup_list));
1635 list_del(&tr->otr_list);
1637 d->opd_syn_tracker = NULL;
1639 mutex_unlock(&osp_id_tracker_sem);
1645 * Generate a new ID on a tracker.
1647 * Generates a new ID using the tracker associated with the given OSP device
1648 * \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
1649 * the wakeup list, so OSP won't miss when a transaction using the ID is
1650 * committed. Notice ID is 32bit, but llog doesn't support >2^32 records anyway.
1652 * \param[in] d OSP device
1653 * \param[in] id 0 or ID generated previously
1655 * \retval ID the caller should use
1657 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1659 struct osp_id_tracker *tr;
1661 tr = d->opd_syn_tracker;
1664 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1665 spin_lock(&tr->otr_lock);
1666 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1667 spin_unlock(&tr->otr_lock);
1668 CERROR("%s: next %u, last synced %lu\n",
1669 d->opd_obd->obd_name, tr->otr_next_id,
1670 d->opd_syn_last_used_id);
1675 id = tr->otr_next_id++;
1676 if (id > d->opd_syn_last_used_id)
1677 d->opd_syn_last_used_id = id;
1678 if (list_empty(&d->opd_syn_ontrack))
1679 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1680 spin_unlock(&tr->otr_lock);
1681 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1687 * Stop to propagate commit status to OSP.
1689 * If the OSP does not have any llog records she's waiting to commit, then
1690 * it is possible to unsubscribe from wakeups from the tracking using this
1693 * \param[in] d OSP device not willing to wakeup
1695 static void osp_sync_remove_from_tracker(struct osp_device *d)
1697 struct osp_id_tracker *tr;
1699 tr = d->opd_syn_tracker;
1702 if (list_empty(&d->opd_syn_ontrack))
1705 spin_lock(&tr->otr_lock);
1706 list_del_init(&d->opd_syn_ontrack);
1707 spin_unlock(&tr->otr_lock);