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);
347 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
351 CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
352 d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
353 (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
354 (unsigned long) osi->osi_cookie.lgc_index, rc);
359 if (likely(rc == 0)) {
360 spin_lock(&d->opd_syn_lock);
361 d->opd_syn_changes++;
362 spin_unlock(&d->opd_syn_lock);
368 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
369 llog_op_type type, struct thandle *th,
370 const struct lu_attr *attr)
372 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
373 lu_object_fid(&o->opo_obj.do_lu), type, 1,
377 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
378 struct lu_fid *fid, int lost, struct thandle *th)
380 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
384 * it's quite obvious we can't maintain all the structures in the memory:
385 * while OST is down, MDS can be processing thousands and thousands of unlinks
386 * filling persistent llogs and in-core respresentation
388 * this doesn't scale at all. so we need basically the following:
389 * a) destroy/setattr append llog records
390 * b) once llog has grown to X records, we process first Y committed records
392 * once record R is found via llog_process(), it becomes committed after any
393 * subsequent commit callback (at the most)
397 * ptlrpc commit callback.
399 * The callback is called by PTLRPC when a RPC is reported committed by the
400 * target (OST). We register the callback for the every RPC applying a change
401 * from the llog. This way we know then the llog records can be cancelled.
402 * Notice the callback can be called when OSP is finishing. We can detect this
403 * checking that actual transno in the request is less or equal of known
404 * committed transno (see osp_sync_process_committed() for the details).
405 * XXX: this is pretty expensive and can be improved later using batching.
407 * \param[in] req request
409 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
411 struct osp_device *d = req->rq_cb_data;
412 struct osp_job_req_args *jra;
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 jra = ptlrpc_req_async_args(req);
424 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
425 LASSERT(list_empty(&jra->jra_link));
427 ptlrpc_request_addref(req);
429 spin_lock(&d->opd_syn_lock);
430 list_add(&jra->jra_link, &d->opd_syn_committed_there);
431 spin_unlock(&d->opd_syn_lock);
433 /* XXX: some batching wouldn't hurt */
434 wake_up(&d->opd_syn_waitq);
438 * RPC interpretation callback.
440 * The callback is called by ptlrpc when RPC is replied. Now we have to decide
442 * - put request on a special list to wait until it's committed by the target,
443 * if the request is succesful
444 * - schedule llog record cancel if no target object is found
445 * - try later (essentially after reboot) in case of unexpected error
447 * \param[in] env LU environment provided by the caller
448 * \param[in] req request replied
449 * \param[in] aa callback data
450 * \param[in] rc result of RPC
454 static int osp_sync_interpret(const struct lu_env *env,
455 struct ptlrpc_request *req, void *aa, int rc)
457 struct osp_device *d = req->rq_cb_data;
458 struct osp_job_req_args *jra = aa;
460 if (jra->jra_magic != OSP_JOB_MAGIC) {
461 DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
466 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
467 atomic_read(&req->rq_refcount),
468 rc, (unsigned) req->rq_transno);
469 LASSERT(rc || req->rq_transno);
473 * we tried to destroy object or update attributes,
474 * but object doesn't exist anymore - cancell llog record
476 LASSERT(req->rq_transno == 0);
477 LASSERT(list_empty(&jra->jra_link));
479 ptlrpc_request_addref(req);
481 spin_lock(&d->opd_syn_lock);
482 list_add(&jra->jra_link, &d->opd_syn_committed_there);
483 spin_unlock(&d->opd_syn_lock);
485 wake_up(&d->opd_syn_waitq);
487 struct obd_import *imp = req->rq_import;
489 * error happened, we'll try to repeat on next boot ?
491 LASSERTF(req->rq_transno == 0 ||
492 req->rq_import_generation < imp->imp_generation,
493 "transno "LPU64", rc %d, gen: req %d, imp %d\n",
494 req->rq_transno, rc, req->rq_import_generation,
495 imp->imp_generation);
496 if (req->rq_transno == 0) {
497 /* this is the last time we see the request
498 * if transno is not zero, then commit cb
499 * will be called at some point */
500 LASSERT(d->opd_syn_rpc_in_progress > 0);
501 spin_lock(&d->opd_syn_lock);
502 d->opd_syn_rpc_in_progress--;
503 spin_unlock(&d->opd_syn_lock);
506 wake_up(&d->opd_syn_waitq);
507 } else if (d->opd_pre != NULL &&
508 unlikely(d->opd_pre_status == -ENOSPC)) {
510 * if current status is -ENOSPC (lack of free space on OST)
511 * then we should poll OST immediately once object destroy
514 osp_statfs_need_now(d);
517 LASSERT(d->opd_syn_rpc_in_flight > 0);
518 spin_lock(&d->opd_syn_lock);
519 d->opd_syn_rpc_in_flight--;
520 spin_unlock(&d->opd_syn_lock);
521 if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
522 wake_up(&d->opd_syn_barrier_waitq);
523 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
524 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
525 d->opd_syn_rpc_in_progress);
527 osp_sync_check_for_work(d);
533 ** Add request to ptlrpc queue.
535 * This is just a tiny helper function to put the request on the sending list
537 * \param[in] d OSP device
538 * \param[in] req request
540 static void osp_sync_send_new_rpc(struct osp_device *d,
541 struct ptlrpc_request *req)
543 struct osp_job_req_args *jra;
545 LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
547 jra = ptlrpc_req_async_args(req);
548 jra->jra_magic = OSP_JOB_MAGIC;
549 INIT_LIST_HEAD(&jra->jra_link);
551 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
556 * Allocate and prepare RPC for a new change.
558 * The function allocates and initializes an RPC which will be sent soon to
559 * apply the change to the target OST. The request is initialized from the
560 * llog record passed. Notice only the fields common to all type of changes
563 * \param[in] d OSP device
564 * \param[in] llh llog handle where the record is stored
565 * \param[in] h llog record
566 * \param[in] op type of the change
567 * \param[in] format request format to be used
569 * \retval pointer new request on success
570 * \retval ERR_PTR(errno) on error
572 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
573 struct llog_handle *llh,
574 struct llog_rec_hdr *h,
576 const struct req_format *format)
578 struct ptlrpc_request *req;
579 struct ost_body *body;
580 struct obd_import *imp;
583 /* Prepare the request */
584 imp = d->opd_obd->u.cli.cl_import;
586 req = ptlrpc_request_alloc(imp, format);
588 RETURN(ERR_PTR(-ENOMEM));
590 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
592 ptlrpc_req_finished(req);
597 * this is a trick: to save on memory allocations we put cookie
598 * into the request, but don't set corresponded flag in o_valid
599 * so that OST doesn't interpret this cookie. once the request
600 * is committed on OST we take cookie from the request and cancel
602 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
604 body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
605 body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
606 body->oa.o_lcookie.lgc_index = h->lrh_index;
608 req->rq_interpret_reply = osp_sync_interpret;
609 req->rq_commit_cb = osp_sync_request_commit_cb;
612 ptlrpc_request_set_replen(req);
618 * Generate a request for setattr change.
620 * The function prepares a new RPC, initializes it with setattr specific
621 * bits and send the RPC.
623 * \param[in] d OSP device
624 * \param[in] llh llog handle where the record is stored
625 * \param[in] h llog record
627 * \retval 0 on success
628 * \retval negative negated errno on error
630 static int osp_sync_new_setattr_job(struct osp_device *d,
631 struct llog_handle *llh,
632 struct llog_rec_hdr *h)
634 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
635 struct ptlrpc_request *req;
636 struct ost_body *body;
639 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
641 /* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
642 * so no bits other than these should be set. */
643 if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
644 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
645 d->opd_obd->obd_name, rec->lsr_valid);
646 /* return 0 so that sync thread can continue processing
651 req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
653 RETURN(PTR_ERR(req));
655 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
657 body->oa.o_oi = rec->lsr_oi;
658 body->oa.o_uid = rec->lsr_uid;
659 body->oa.o_gid = rec->lsr_gid;
660 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
661 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
662 * we assume that both UID and GID are valid in that case. */
663 if (rec->lsr_valid == 0)
664 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
666 body->oa.o_valid |= rec->lsr_valid;
668 osp_sync_send_new_rpc(d, req);
673 * Generate a request for unlink change.
675 * The function prepares a new RPC, initializes it with unlink(destroy)
676 * specific bits and sends the RPC. The function is used to handle
677 * llog_unlink_rec which were used in the older versions of Lustre.
678 * Current version uses llog_unlink_rec64.
680 * \param[in] d OSP device
681 * \param[in] llh llog handle where the record is stored
682 * \param[in] h llog record
684 * \retval 0 on success
685 * \retval negative negated errno on error
687 static int osp_sync_new_unlink_job(struct osp_device *d,
688 struct llog_handle *llh,
689 struct llog_rec_hdr *h)
691 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
692 struct ptlrpc_request *req;
693 struct ost_body *body;
696 LASSERT(h->lrh_type == MDS_UNLINK_REC);
698 req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
700 RETURN(PTR_ERR(req));
702 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
704 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
705 ostid_set_id(&body->oa.o_oi, rec->lur_oid);
706 body->oa.o_misc = rec->lur_count;
707 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
709 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
711 osp_sync_send_new_rpc(d, req);
716 * Prepare OUT-based object destroy RPC.
718 * The function allocates a new RPC with OUT format. Then initializes the RPC
719 * to contain OUT_DESTROY update against the object specified in the llog
720 * record provided by the caller.
722 * \param[in] env LU environment provided by the caller
723 * \param[in] osp OSP device
724 * \param[in] llh llog handle where the record is stored
725 * \param[in] h llog record
726 * \param[out] reqp request prepared
728 * \retval 0 on success
729 * \retval negative negated errno on error
731 static int osp_prep_unlink_update_req(const struct lu_env *env,
732 struct osp_device *osp,
733 struct llog_handle *llh,
734 struct llog_rec_hdr *h,
735 struct ptlrpc_request **reqp)
737 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
738 struct dt_update_request *update = NULL;
739 struct ptlrpc_request *req;
740 struct llog_cookie lcookie;
746 update = dt_update_request_create(&osp->opd_dt_dev);
748 RETURN(PTR_ERR(update));
750 /* This can only happens for unlink slave directory, so decrease
751 * ref for ".." and "." */
752 rc = out_update_pack(env, &update->dur_buf, OUT_REF_DEL, &rec->lur_fid,
757 rc = out_update_pack(env, &update->dur_buf, OUT_REF_DEL, &rec->lur_fid,
762 lcookie.lgc_lgl = llh->lgh_id;
763 lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
764 lcookie.lgc_index = h->lrh_index;
765 size = sizeof(lcookie);
768 rc = out_update_pack(env, &update->dur_buf, OUT_DESTROY, &rec->lur_fid,
773 rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
774 update->dur_buf.ub_req, &req);
778 req->rq_interpret_reply = osp_sync_interpret;
779 req->rq_commit_cb = osp_sync_request_commit_cb;
780 req->rq_cb_data = osp;
782 ptlrpc_request_set_replen(req);
786 dt_update_request_destroy(update);
792 * Generate a request for unlink change.
794 * The function prepares a new RPC, initializes it with unlink(destroy)
795 * specific bits and sends the RPC. Depending on the target (MDT or OST)
796 * two different protocols are used. For MDT we use OUT (basically OSD API
797 * updates transferred via a network). For OST we still use the old
798 * protocol (OBD?), originally for compatibility. Later we can start to
799 * use OUT for OST as well, this will allow batching and better code
802 * \param[in] env LU environment provided by the caller
803 * \param[in] d OSP device
804 * \param[in] llh llog handle where the record is stored
805 * \param[in] h llog record
807 * \retval 0 on success
808 * \retval negative negated errno on error
810 static int osp_sync_new_unlink64_job(const struct lu_env *env,
811 struct osp_device *d,
812 struct llog_handle *llh,
813 struct llog_rec_hdr *h)
815 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
816 struct ptlrpc_request *req = NULL;
817 struct ost_body *body;
821 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
823 if (d->opd_connect_mdt) {
824 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
828 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
831 RETURN(PTR_ERR(req));
833 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
836 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
839 body->oa.o_misc = rec->lur_count;
840 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
843 osp_sync_send_new_rpc(d, req);
848 * Process llog records.
850 * This function is called to process the llog records committed locally.
851 * In the recovery model used by OSP we can apply a change to a remote
852 * target once corresponding transaction (like posix unlink) is committed
853 * locally so can't revert.
854 * Depending on the llog record type, a given handler is called that is
855 * responsible for preparing and sending the RPC to apply the change.
856 * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
858 * \param[in] env LU environment provided by the caller
859 * \param[in] d OSP device
860 * \param[in] llh llog handle where the record is stored
861 * \param[in] rec llog record
863 * \retval 0 on success
864 * \retval negative negated errno on error
866 static int osp_sync_process_record(const struct lu_env *env,
867 struct osp_device *d,
868 struct llog_handle *llh,
869 struct llog_rec_hdr *rec)
871 struct llog_cookie cookie;
874 cookie.lgc_lgl = llh->lgh_id;
875 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
876 cookie.lgc_index = rec->lrh_index;
878 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
879 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
881 /* we're waiting for the record generated by this instance */
882 LASSERT(d->opd_syn_prev_done == 0);
883 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
884 sizeof(gen->lgr_gen))) {
885 CDEBUG(D_HA, "processed all old entries\n");
886 d->opd_syn_prev_done = 1;
889 /* cancel any generation record */
890 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
897 * now we prepare and fill requests to OST, put them on the queue
898 * and fire after next commit callback
901 /* notice we increment counters before sending RPC, to be consistent
902 * in RPC interpret callback which may happen very quickly */
903 spin_lock(&d->opd_syn_lock);
904 d->opd_syn_rpc_in_flight++;
905 d->opd_syn_rpc_in_progress++;
906 spin_unlock(&d->opd_syn_lock);
908 switch (rec->lrh_type) {
909 /* case MDS_UNLINK_REC is kept for compatibility */
911 rc = osp_sync_new_unlink_job(d, llh, rec);
913 case MDS_UNLINK64_REC:
914 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
916 case MDS_SETATTR64_REC:
917 rc = osp_sync_new_setattr_job(d, llh, rec);
920 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
922 /* we should continue processing */
925 /* rc > 0 means sync RPC being added to the queue */
926 if (likely(rc > 0)) {
927 spin_lock(&d->opd_syn_lock);
928 if (d->opd_syn_prev_done) {
929 LASSERT(d->opd_syn_changes > 0);
930 LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
932 * NOTE: it's possible to meet same id if
933 * OST stores few stripes of same file
935 if (rec->lrh_id > d->opd_syn_last_processed_id) {
936 d->opd_syn_last_processed_id = rec->lrh_id;
937 wake_up(&d->opd_syn_barrier_waitq);
940 d->opd_syn_changes--;
942 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
943 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
944 d->opd_syn_rpc_in_progress);
945 spin_unlock(&d->opd_syn_lock);
948 spin_lock(&d->opd_syn_lock);
949 d->opd_syn_rpc_in_flight--;
950 d->opd_syn_rpc_in_progress--;
951 spin_unlock(&d->opd_syn_lock);
954 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
955 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
960 * Cancel llog records for the committed changes.
962 * The function walks through the list of the committed RPCs and cancels
963 * corresponding llog records. see osp_sync_request_commit_cb() for the
966 * \param[in] env LU environment provided by the caller
967 * \param[in] d OSP device
969 static void osp_sync_process_committed(const struct lu_env *env,
970 struct osp_device *d)
972 struct obd_device *obd = d->opd_obd;
973 struct obd_import *imp = obd->u.cli.cl_import;
974 struct ost_body *body;
975 struct ptlrpc_request *req;
976 struct llog_ctxt *ctxt;
977 struct llog_handle *llh;
978 struct list_head list;
983 if (list_empty(&d->opd_syn_committed_there))
987 * if current status is -ENOSPC (lack of free space on OST)
988 * then we should poll OST immediately once object destroy
990 * notice: we do this upon commit as well because some backends
991 * (like DMU) do not release space right away.
993 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
994 osp_statfs_need_now(d);
997 * now cancel them all
998 * XXX: can we improve this using some batching?
999 * with batch RPC that'll happen automatically?
1000 * XXX: can we store ctxt in lod_device and save few cycles ?
1002 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1005 llh = ctxt->loc_handle;
1008 INIT_LIST_HEAD(&list);
1009 spin_lock(&d->opd_syn_lock);
1010 list_splice(&d->opd_syn_committed_there, &list);
1011 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1012 spin_unlock(&d->opd_syn_lock);
1014 while (!list_empty(&list)) {
1015 struct llog_cookie *lcookie = NULL;
1016 struct osp_job_req_args *jra;
1018 jra = list_entry(list.next, struct osp_job_req_args, jra_link);
1019 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1020 list_del_init(&jra->jra_link);
1022 req = container_of((void *)jra, struct ptlrpc_request,
1024 if (d->opd_connect_mdt) {
1025 struct object_update_request *ureq;
1026 struct object_update *update;
1027 ureq = req_capsule_client_get(&req->rq_pill,
1029 LASSERT(ureq != NULL &&
1030 ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
1032 /* 1st/2nd is for decref . and .., 3rd one is for
1033 * destroy, where the log cookie is stored.
1034 * See osp_prep_unlink_update_req */
1035 update = object_update_request_get(ureq, 2, NULL);
1036 LASSERT(update != NULL);
1037 lcookie = object_update_param_get(update, 0, NULL);
1038 LASSERT(lcookie != NULL);
1040 body = req_capsule_client_get(&req->rq_pill,
1043 lcookie = &body->oa.o_lcookie;
1045 /* import can be closing, thus all commit cb's are
1046 * called we can check committness directly */
1047 if (req->rq_transno <= imp->imp_peer_committed_transno) {
1048 rc = llog_cat_cancel_records(env, llh, 1, lcookie);
1050 CERROR("%s: can't cancel record: %d\n",
1053 DEBUG_REQ(D_HA, req, "not committed");
1056 ptlrpc_req_finished(req);
1060 llog_ctxt_put(ctxt);
1062 LASSERT(d->opd_syn_rpc_in_progress >= done);
1063 spin_lock(&d->opd_syn_lock);
1064 d->opd_syn_rpc_in_progress -= done;
1065 spin_unlock(&d->opd_syn_lock);
1066 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
1067 d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
1068 d->opd_syn_rpc_in_progress);
1070 osp_sync_check_for_work(d);
1072 /* wake up the thread if requested to stop:
1073 * it might be waiting for in-progress to complete */
1074 if (unlikely(osp_sync_running(d) == 0))
1075 wake_up(&d->opd_syn_waitq);
1081 * The core of the syncing mechanism.
1083 * This is a callback called by the llog processing function. Essentially it
1084 * suspends llog processing until there is a record to process (it's supposed
1085 * to be committed locally). The function handles RPCs committed by the target
1086 * and cancels corresponding llog records.
1088 * \param[in] env LU environment provided by the caller
1089 * \param[in] llh llog handle we're processing
1090 * \param[in] rec current llog record
1091 * \param[in] data callback data containing a pointer to the device
1093 * \retval 0 to ask the caller (llog_process()) to continue
1094 * \retval LLOG_PROC_BREAK to ask the caller to break
1096 static int osp_sync_process_queues(const struct lu_env *env,
1097 struct llog_handle *llh,
1098 struct llog_rec_hdr *rec,
1101 struct osp_device *d = data;
1105 struct l_wait_info lwi = { 0 };
1107 if (!osp_sync_running(d)) {
1108 CDEBUG(D_HA, "stop llog processing\n");
1109 return LLOG_PROC_BREAK;
1112 /* process requests committed by OST */
1113 osp_sync_process_committed(env, d);
1115 /* if we there are changes to be processed and we have
1116 * resources for this ... do now */
1117 if (osp_sync_can_process_new(d, rec)) {
1119 /* ask llog for another record */
1120 CDEBUG(D_HA, "%lu changes, %u in progress,"
1123 d->opd_syn_rpc_in_progress,
1124 d->opd_syn_rpc_in_flight);
1129 * try to send, in case of disconnection, suspend
1130 * processing till we can send this request
1133 rc = osp_sync_process_record(env, d, llh, rec);
1135 * XXX: probably different handling is needed
1136 * for some bugs, like immediate exit or if
1140 CERROR("can't send: %d\n", rc);
1141 l_wait_event(d->opd_syn_waitq,
1142 !osp_sync_running(d) ||
1143 osp_sync_has_work(d),
1146 } while (rc != 0 && osp_sync_running(d));
1152 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
1153 osp_sync_remove_from_tracker(d);
1155 l_wait_event(d->opd_syn_waitq,
1156 !osp_sync_running(d) ||
1157 osp_sync_can_process_new(d, rec) ||
1158 !list_empty(&d->opd_syn_committed_there),
1166 * This thread runs llog_cat_process() scanner calling our callback
1167 * to process llog records. in the callback we implement tricky
1168 * state machine as we don't want to start scanning of the llog again
1169 * and again, also we don't want to process too many records and send
1170 * too many RPCs a time. so, depending on current load (num of changes
1171 * being synced to OST) the callback can suspend awaiting for some
1172 * new conditions, like syncs completed.
1174 * In order to process llog records left by previous boots and to allow
1175 * llog_process_thread() to find something (otherwise it'd just exit
1176 * immediately) we add a special GENERATATION record on each boot.
1178 * \param[in] _arg a pointer to thread's arguments
1180 * \retval 0 on success
1181 * \retval negative negated errno on error
1183 static int osp_sync_thread(void *_arg)
1185 struct osp_device *d = _arg;
1186 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1187 struct l_wait_info lwi = { 0 };
1188 struct llog_ctxt *ctxt;
1189 struct obd_device *obd = d->opd_obd;
1190 struct llog_handle *llh;
1196 rc = lu_env_init(&env, LCT_LOCAL);
1198 CERROR("%s: can't initialize env: rc = %d\n",
1203 spin_lock(&d->opd_syn_lock);
1204 thread->t_flags = SVC_RUNNING;
1205 spin_unlock(&d->opd_syn_lock);
1206 wake_up(&thread->t_ctl_waitq);
1208 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1210 CERROR("can't get appropriate context\n");
1211 GOTO(out, rc = -EINVAL);
1214 llh = ctxt->loc_handle;
1216 CERROR("can't get llh\n");
1217 llog_ctxt_put(ctxt);
1218 GOTO(out, rc = -EINVAL);
1221 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
1222 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1223 "%lu changes, %u in progress, %u in flight: %d\n",
1224 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1225 d->opd_syn_rpc_in_flight, rc);
1227 /* we don't expect llog_process_thread() to exit till umount */
1228 LASSERTF(thread->t_flags != SVC_RUNNING,
1229 "%lu changes, %u in progress, %u in flight\n",
1230 d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1231 d->opd_syn_rpc_in_flight);
1233 /* wait till all the requests are completed */
1235 while (d->opd_syn_rpc_in_progress > 0) {
1236 osp_sync_process_committed(&env, d);
1238 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1239 rc = l_wait_event(d->opd_syn_waitq,
1240 d->opd_syn_rpc_in_progress == 0,
1242 if (rc == -ETIMEDOUT)
1244 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1245 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1246 d->opd_syn_rpc_in_flight,
1247 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1251 llog_cat_close(&env, llh);
1252 rc = llog_cleanup(&env, ctxt);
1254 CERROR("can't cleanup llog: %d\n", rc);
1256 LASSERTF(d->opd_syn_rpc_in_progress == 0,
1257 "%s: %d %d %sempty\n",
1258 d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1259 d->opd_syn_rpc_in_flight,
1260 list_empty(&d->opd_syn_committed_there) ? "" : "!");
1262 thread->t_flags = SVC_STOPPED;
1264 wake_up(&thread->t_ctl_waitq);
1274 * Initializes the llog. Specific llog to be used depends on the type of the
1275 * target OSP represents (OST or MDT). The function adds appends a new llog
1276 * record to mark the place where the records associated with this boot
1279 * \param[in] env LU environment provided by the caller
1280 * \param[in] d OSP device
1282 * \retval 0 on success
1283 * \retval negative negated errno on error
1285 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1287 struct osp_thread_info *osi = osp_env_info(env);
1288 struct lu_fid *fid = &osi->osi_fid;
1289 struct llog_handle *lgh = NULL;
1290 struct obd_device *obd = d->opd_obd;
1291 struct llog_ctxt *ctxt;
1299 * open llog corresponding to our OST
1301 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1302 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1304 if (d->opd_connect_mdt)
1305 lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
1307 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1309 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1310 &osi->osi_cid, fid);
1312 if (rc != -EFAULT) {
1313 CERROR("%s: can't get id from catalogs: rc = %d\n",
1318 /* After sparse OST indices is supported, the CATALOG file
1319 * may become a sparse file that results in failure on
1320 * reading. Skip this error as the llog will be created
1322 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1326 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1327 obd->obd_name, d->opd_index,
1328 POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1329 osi->osi_cid.lci_logid.lgl_ogen);
1331 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1332 &osp_mds_ost_orig_logops);
1336 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1339 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1340 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1342 /* re-create llog if it is missing */
1344 logid_set_id(&osi->osi_cid.lci_logid, 0);
1346 GOTO(out_cleanup, rc);
1349 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1350 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1352 GOTO(out_cleanup, rc);
1353 osi->osi_cid.lci_logid = lgh->lgh_id;
1356 LASSERT(lgh != NULL);
1357 ctxt->loc_handle = lgh;
1359 rc = llog_cat_init_and_process(env, lgh);
1361 GOTO(out_close, rc);
1363 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1364 &osi->osi_cid, fid);
1366 GOTO(out_close, rc);
1369 * put a mark in the llog till which we'll be processing
1370 * old records restless
1372 d->opd_syn_generation.mnt_cnt = cfs_time_current();
1373 d->opd_syn_generation.conn_cnt = cfs_time_current();
1375 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1376 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1378 memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1379 sizeof(osi->osi_gen.lgr_gen));
1381 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1383 GOTO(out_close, rc);
1384 llog_ctxt_put(ctxt);
1387 llog_cat_close(env, lgh);
1389 llog_cleanup(env, ctxt);
1394 * Cleanup llog used for syncing.
1396 * Closes and cleanups the llog. The function is called when the device is
1399 * \param[in] env LU environment provided by the caller
1400 * \param[in] d OSP device
1402 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1404 struct llog_ctxt *ctxt;
1406 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1408 llog_cat_close(env, ctxt->loc_handle);
1409 llog_cleanup(env, ctxt);
1413 * Initialization of the sync component of OSP.
1415 * Initializes the llog and starts a new thread to handle the changes to
1416 * the remote target (OST or MDT).
1418 * \param[in] env LU environment provided by the caller
1419 * \param[in] d OSP device
1421 * \retval 0 on success
1422 * \retval negative negated errno on error
1424 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1426 struct l_wait_info lwi = { 0 };
1427 struct task_struct *task;
1432 rc = osp_sync_id_traction_init(d);
1437 * initialize llog storing changes
1439 rc = osp_sync_llog_init(env, d);
1441 CERROR("%s: can't initialize llog: rc = %d\n",
1442 d->opd_obd->obd_name, rc);
1447 * Start synchronization thread
1449 d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1450 d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1451 spin_lock_init(&d->opd_syn_lock);
1452 init_waitqueue_head(&d->opd_syn_waitq);
1453 init_waitqueue_head(&d->opd_syn_barrier_waitq);
1454 init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1455 INIT_LIST_HEAD(&d->opd_syn_committed_there);
1457 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1458 d->opd_index, d->opd_group);
1461 CERROR("%s: cannot start sync thread: rc = %d\n",
1462 d->opd_obd->obd_name, rc);
1466 l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1467 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1471 osp_sync_llog_fini(env, d);
1473 osp_sync_id_traction_fini(d);
1478 * Stop the syncing thread.
1480 * Asks the syncing thread to stop and wait until it's stopped.
1482 * \param[in] d OSP device
1486 int osp_sync_fini(struct osp_device *d)
1488 struct ptlrpc_thread *thread = &d->opd_syn_thread;
1492 thread->t_flags = SVC_STOPPING;
1493 wake_up(&d->opd_syn_waitq);
1494 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1497 * unregister transaction callbacks only when sync thread
1498 * has finished operations with llog
1500 osp_sync_id_traction_fini(d);
1505 static DEFINE_MUTEX(osp_id_tracker_sem);
1506 static struct list_head osp_id_tracker_list =
1507 LIST_HEAD_INIT(osp_id_tracker_list);
1510 * OSD commit callback.
1512 * The function is used as a local OSD commit callback to track the highest
1513 * committed llog record id. see osp_sync_id_traction_init() for the details.
1515 * \param[in] th local transaction handle committed
1516 * \param[in] cookie commit callback data (our private structure)
1518 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1520 struct osp_id_tracker *tr = cookie;
1521 struct osp_device *d;
1522 struct osp_txn_info *txn;
1526 txn = osp_txn_info(&th->th_ctx);
1527 if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1530 spin_lock(&tr->otr_lock);
1531 if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1532 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1533 tr->otr_committed_id, txn->oti_current_id);
1534 tr->otr_committed_id = txn->oti_current_id;
1536 list_for_each_entry(d, &tr->otr_wakeup_list,
1538 d->opd_syn_last_committed_id = tr->otr_committed_id;
1539 wake_up(&d->opd_syn_waitq);
1542 spin_unlock(&tr->otr_lock);
1546 * Initialize commit tracking mechanism.
1548 * Some setups may have thousands of OSTs and each will be represented by OSP.
1549 * Meaning order of magnitute many more changes to apply every second. In order
1550 * to keep the number of commit callbacks low this mechanism was introduced.
1551 * The mechanism is very similar to transno used by MDT service: it's an single
1552 * ID stream which can be assigned by any OSP to its llog records. The tricky
1553 * part is that ID is stored in per-transaction data and re-used by all the OSPs
1554 * involved in that transaction. Then all these OSPs are woken up utilizing a single OSD commit callback.
1556 * The function initializes the data used by the tracker described above.
1557 * A singler tracker per OSD device is created.
1559 * \param[in] d OSP device
1561 * \retval 0 on success
1562 * \retval negative negated errno on error
1564 static int osp_sync_id_traction_init(struct osp_device *d)
1566 struct osp_id_tracker *tr, *found = NULL;
1570 LASSERT(d->opd_storage);
1571 LASSERT(d->opd_syn_tracker == NULL);
1572 INIT_LIST_HEAD(&d->opd_syn_ontrack);
1574 mutex_lock(&osp_id_tracker_sem);
1575 list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1576 if (tr->otr_dev == d->opd_storage) {
1577 LASSERT(atomic_read(&tr->otr_refcount));
1578 atomic_inc(&tr->otr_refcount);
1579 d->opd_syn_tracker = tr;
1585 if (found == NULL) {
1589 d->opd_syn_tracker = tr;
1590 spin_lock_init(&tr->otr_lock);
1591 tr->otr_dev = d->opd_storage;
1592 tr->otr_next_id = 1;
1593 tr->otr_committed_id = 0;
1594 atomic_set(&tr->otr_refcount, 1);
1595 INIT_LIST_HEAD(&tr->otr_wakeup_list);
1596 list_add(&tr->otr_list, &osp_id_tracker_list);
1597 tr->otr_tx_cb.dtc_txn_commit =
1598 osp_sync_tracker_commit_cb;
1599 tr->otr_tx_cb.dtc_cookie = tr;
1600 tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1601 dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1605 mutex_unlock(&osp_id_tracker_sem);
1611 * Release commit tracker.
1613 * Decrease a refcounter on the tracker used by the given OSP device \a d.
1614 * If no more users left, then the tracker is released.
1616 * \param[in] d OSP device
1618 static void osp_sync_id_traction_fini(struct osp_device *d)
1620 struct osp_id_tracker *tr;
1625 tr = d->opd_syn_tracker;
1631 osp_sync_remove_from_tracker(d);
1633 mutex_lock(&osp_id_tracker_sem);
1634 if (atomic_dec_and_test(&tr->otr_refcount)) {
1635 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1636 LASSERT(list_empty(&tr->otr_wakeup_list));
1637 list_del(&tr->otr_list);
1639 d->opd_syn_tracker = NULL;
1641 mutex_unlock(&osp_id_tracker_sem);
1647 * Generate a new ID on a tracker.
1649 * Generates a new ID using the tracker associated with the given OSP device
1650 * \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
1651 * the wakeup list, so OSP won't miss when a transaction using the ID is
1652 * committed. Notice ID is 32bit, but llog doesn't support >2^32 records anyway.
1654 * \param[in] d OSP device
1655 * \param[in] id 0 or ID generated previously
1657 * \retval ID the caller should use
1659 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1661 struct osp_id_tracker *tr;
1663 tr = d->opd_syn_tracker;
1666 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1667 spin_lock(&tr->otr_lock);
1668 if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1669 spin_unlock(&tr->otr_lock);
1670 CERROR("%s: next %u, last synced %lu\n",
1671 d->opd_obd->obd_name, tr->otr_next_id,
1672 d->opd_syn_last_used_id);
1677 id = tr->otr_next_id++;
1678 if (id > d->opd_syn_last_used_id)
1679 d->opd_syn_last_used_id = id;
1680 if (list_empty(&d->opd_syn_ontrack))
1681 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1682 spin_unlock(&tr->otr_lock);
1683 CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1689 * Stop to propagate commit status to OSP.
1691 * If the OSP does not have any llog records she's waiting to commit, then
1692 * it is possible to unsubscribe from wakeups from the tracking using this
1695 * \param[in] d OSP device not willing to wakeup
1697 static void osp_sync_remove_from_tracker(struct osp_device *d)
1699 struct osp_id_tracker *tr;
1701 tr = d->opd_syn_tracker;
1704 if (list_empty(&d->opd_syn_ontrack))
1707 spin_lock(&tr->otr_lock);
1708 list_del_init(&d->opd_syn_ontrack);
1709 spin_unlock(&tr->otr_lock);