4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osp/osp_sync.c
34 * Lustre OST Proxy Device
36 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
37 * Author: Mikhail Pershin <mike.pershin@intel.com>
40 #define DEBUG_SUBSYSTEM S_MDS
42 #include <linux/kthread.h>
43 #include <lustre_log.h>
44 #include <lustre_update.h>
45 #include "osp_internal.h"
47 static int osp_sync_id_traction_init(struct osp_device *d);
48 static void osp_sync_id_traction_fini(struct osp_device *d);
49 static __u64 osp_sync_id_get(struct osp_device *d, __u64 id);
50 static void osp_sync_remove_from_tracker(struct osp_device *d);
53 * this is a components of OSP implementing synchronization between MDS and OST
54 * it llogs all interesting changes (currently it's uig/gid change and object
55 * destroy) atomically, then makes sure changes hit OST storage
57 * we have 4 queues of work:
59 * the first queue is llog itself, once read a change is stored in 2nd queue
60 * in form of RPC (but RPC isn't fired yet).
62 * the second queue (opd_sync_waiting_for_commit) holds changes awaiting local
63 * commit. once change is committed locally it migrates onto 3rd queue.
65 * the third queue (opd_sync_committed_here) holds changes committed locally,
66 * but not sent to OST (as the pipe can be full). once pipe becomes non-full
67 * we take a change from the queue and fire corresponded RPC.
69 * once RPC is reported committed by OST (using regular last_committed mech.)
70 * the change jumps into 4th queue (opd_sync_committed_there), now we can
71 * cancel corresponded llog record and release RPC
73 * opd_sync_changes is a number of unread llog records (to be processed).
74 * notice this number doesn't include llog records from previous boots.
75 * with OSP_SYNC_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
77 * opd_sync_rpcs_in_progress is total number of requests in above 2-4 queues.
78 * we control this with OSP_MAX_RPCS_IN_PROGRESS so that OSP don't consume
79 * too much memory -- how to deal with 1000th OSTs ? batching could help?
81 * opd_sync_rpcs_in_flight is a number of RPC in flight.
82 * we control this with OSP_MAX_RPCS_IN_FLIGHT
85 /* XXX: do math to learn reasonable threshold
86 * should it be ~ number of changes fitting bulk? */
88 #define OSP_SYNC_THRESHOLD 10
89 #define OSP_MAX_RPCS_IN_FLIGHT 8
90 #define OSP_MAX_RPCS_IN_PROGRESS 4096
92 #define OSP_JOB_MAGIC 0x26112005
94 struct osp_job_req_args {
95 /** bytes reserved for ptlrpc_replay_req() */
96 struct ptlrpc_replay_async_args jra_raa;
97 struct list_head jra_committed_link;
98 struct list_head jra_in_flight_link;
99 struct llog_cookie jra_lcookie;
103 static inline int osp_sync_running(struct osp_device *d)
105 return !!(d->opd_sync_thread.t_flags & SVC_RUNNING);
109 * Check status: whether OSP thread has stopped
111 * \param[in] d OSP device
113 * \retval 0 still running
116 static inline int osp_sync_stopped(struct osp_device *d)
118 return !!(d->opd_sync_thread.t_flags & SVC_STOPPED);
122 ** Check for new changes to sync
124 * \param[in] d OSP device
126 * \retval 1 there are changes
127 * \retval 0 there are no changes
129 static inline int osp_sync_has_new_job(struct osp_device *d)
131 return ((d->opd_sync_last_processed_id < d->opd_sync_last_used_id) &&
132 (d->opd_sync_last_processed_id < d->opd_sync_last_committed_id))
133 || (d->opd_sync_prev_done == 0);
136 static inline int osp_sync_in_flight_conflict(struct osp_device *d,
137 struct llog_rec_hdr *h)
139 struct osp_job_req_args *jra;
143 if (h == NULL || h->lrh_type == LLOG_GEN_REC ||
144 list_empty(&d->opd_sync_in_flight_list))
147 memset(&ostid, 0, sizeof(ostid));
148 switch (h->lrh_type) {
149 case MDS_UNLINK_REC: {
150 struct llog_unlink_rec *unlink = (struct llog_unlink_rec *)h;
152 ostid_set_seq(&ostid, unlink->lur_oseq);
153 if (ostid_set_id(&ostid, unlink->lur_oid)) {
154 CERROR("Bad %llu to set " DOSTID "\n",
155 (unsigned long long)(unlink->lur_oid),
161 case MDS_UNLINK64_REC:
162 fid_to_ostid(&((struct llog_unlink64_rec *)h)->lur_fid, &ostid);
164 case MDS_SETATTR64_REC:
165 ostid = ((struct llog_setattr64_rec *)h)->lsr_oi;
171 spin_lock(&d->opd_sync_lock);
172 list_for_each_entry(jra, &d->opd_sync_in_flight_list,
173 jra_in_flight_link) {
174 struct ptlrpc_request *req;
175 struct ost_body *body;
177 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
179 req = container_of((void *)jra, struct ptlrpc_request,
181 body = req_capsule_client_get(&req->rq_pill,
185 if (memcmp(&ostid, &body->oa.o_oi, sizeof(ostid)) == 0) {
190 spin_unlock(&d->opd_sync_lock);
195 static inline int osp_sync_rpcs_in_progress_low(struct osp_device *d)
197 return atomic_read(&d->opd_sync_rpcs_in_progress) <
198 d->opd_sync_max_rpcs_in_progress;
202 * Check for room in the network pipe to OST
204 * \param[in] d OSP device
206 * \retval 1 there is room
207 * \retval 0 no room, the pipe is full
209 static inline int osp_sync_rpcs_in_flight_low(struct osp_device *d)
211 return atomic_read(&d->opd_sync_rpcs_in_flight) <
212 d->opd_sync_max_rpcs_in_flight;
216 * Wake up check for the main sync thread
218 * \param[in] d OSP device
220 * \retval 1 time to wake up
221 * \retval 0 no need to wake up
223 static inline int osp_sync_has_work(struct osp_device *osp)
225 /* has new/old changes and low in-progress? */
226 if (osp_sync_has_new_job(osp) && osp_sync_rpcs_in_progress_low(osp) &&
227 osp_sync_rpcs_in_flight_low(osp) && osp->opd_imp_connected)
230 /* has remotely committed? */
231 if (!list_empty(&osp->opd_sync_committed_there))
237 void osp_sync_check_for_work(struct osp_device *osp)
239 if (osp_sync_has_work(osp))
240 wake_up(&osp->opd_sync_waitq);
243 static inline __u64 osp_sync_correct_id(struct osp_device *d,
244 struct llog_rec_hdr *rec)
247 * llog use cyclic store with 32 bit lrh_id
248 * so overflow lrh_id is possible. Range between
249 * last_processed and last_committed is less than
250 * 64745 ^ 2 and less than 2^32 - 1
252 __u64 correct_id = d->opd_sync_last_committed_id;
254 if ((correct_id & 0xffffffffULL) < rec->lrh_id)
255 correct_id -= 0x100000000ULL;
257 correct_id &= ~0xffffffffULL;
258 correct_id |= rec->lrh_id;
263 * Check and return ready-for-new status.
265 * The thread processing llog record uses this function to check whether
266 * it's time to take another record and process it. The number of conditions
267 * must be met: the connection should be ready, RPCs in flight not exceeding
268 * the limit, the record is committed locally, etc (see the lines below).
270 * \param[in] d OSP device
271 * \param[in] rec next llog record to process
273 * \retval 0 not ready
276 static inline int osp_sync_can_process_new(struct osp_device *d,
277 struct llog_rec_hdr *rec)
281 if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
283 if (unlikely(osp_sync_in_flight_conflict(d, rec)))
285 if (!osp_sync_rpcs_in_progress_low(d))
287 if (!osp_sync_rpcs_in_flight_low(d))
289 if (!d->opd_imp_connected)
291 if (d->opd_sync_prev_done == 0)
293 if (atomic_read(&d->opd_sync_changes) == 0)
296 osp_sync_correct_id(d, rec) <= d->opd_sync_last_committed_id)
302 * Declare intention to add a new change.
304 * With regard to OSD API, we have to declare any changes ahead. In this
305 * case we declare an intention to add a llog record representing the
306 * change on the local storage.
308 * \param[in] env LU environment provided by the caller
309 * \param[in] o OSP object
310 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
311 * \param[in] th transaction handle (local)
313 * \retval 0 on success
314 * \retval negative negated errno on error
316 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
317 llog_op_type type, struct thandle *th)
319 struct osp_thread_info *osi = osp_env_info(env);
320 struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
321 struct llog_ctxt *ctxt;
322 struct thandle *storage_th;
327 /* it's a layering violation, to access internals of th,
328 * but we can do this as a sanity check, for a while */
329 LASSERT(th->th_top != NULL);
330 storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
331 if (IS_ERR(storage_th))
332 RETURN(PTR_ERR(storage_th));
335 case MDS_UNLINK64_REC:
336 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
338 case MDS_SETATTR64_REC:
339 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec_v2);
345 /* we want ->dt_trans_start() to allocate per-thandle structure */
346 storage_th->th_tags |= LCT_OSP_THREAD;
348 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
351 rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
359 * Generate a llog record for a given change.
361 * Generates a llog record for the change passed. The change can be of two
362 * types: unlink and setattr. The record gets an ID which later will be
363 * used to track commit status of the change. For unlink changes, the caller
364 * can supply a starting FID and the count of the objects to destroy. For
365 * setattr the caller should apply attributes to apply.
368 * \param[in] env LU environment provided by the caller
369 * \param[in] d OSP device
370 * \param[in] fid fid of the object the change should be applied to
371 * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
372 * \param[in] count count of objects to destroy
373 * \param[in] th transaction handle (local)
374 * \param[in] attr attributes for setattr
376 * \retval 0 on success
377 * \retval negative negated errno on error
379 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
380 const struct lu_fid *fid, llog_op_type type,
381 int count, struct thandle *th,
382 const struct lu_attr *attr)
384 struct osp_thread_info *osi = osp_env_info(env);
385 struct llog_ctxt *ctxt;
386 struct osp_txn_info *txn;
387 struct thandle *storage_th;
392 /* it's a layering violation, to access internals of th,
393 * but we can do this as a sanity check, for a while */
394 LASSERT(th->th_top != NULL);
395 storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
396 if (IS_ERR(storage_th))
397 RETURN(PTR_ERR(storage_th));
400 case MDS_UNLINK64_REC:
401 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
402 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
403 osi->osi_unlink.lur_fid = *fid;
404 osi->osi_unlink.lur_count = count;
406 case MDS_SETATTR64_REC:
407 rc = fid_to_ostid(fid, &osi->osi_oi);
409 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
410 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
411 osi->osi_setattr.lsr_oi = osi->osi_oi;
413 osi->osi_setattr.lsr_uid = attr->la_uid;
414 osi->osi_setattr.lsr_gid = attr->la_gid;
415 osi->osi_setattr.lsr_projid = attr->la_projid;
416 osi->osi_setattr.lsr_valid =
417 ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
418 ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0) |
419 ((attr->la_valid & LA_PROJID) ? OBD_MD_FLPROJID : 0);
425 txn = osp_txn_info(&storage_th->th_ctx);
428 txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
429 osi->osi_hdr.lrh_id = (txn->oti_current_id & 0xffffffffULL);
430 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
434 rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
438 if (likely(rc >= 0)) {
439 CDEBUG(D_OTHER, "%s: new record "DFID":%x.%u: rc = %d\n",
440 d->opd_obd->obd_name,
441 PFID(&osi->osi_cookie.lgc_lgl.lgl_oi.oi_fid),
442 osi->osi_cookie.lgc_lgl.lgl_ogen,
443 osi->osi_cookie.lgc_index, rc);
444 atomic_inc(&d->opd_sync_changes);
446 /* return 0 always here, error case just cause no llog record */
450 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
451 llog_op_type type, struct thandle *th,
452 const struct lu_attr *attr)
454 return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
455 lu_object_fid(&o->opo_obj.do_lu), type, 1,
459 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
460 struct lu_fid *fid, int lost, struct thandle *th)
462 return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
466 * it's quite obvious we can't maintain all the structures in the memory:
467 * while OST is down, MDS can be processing thousands and thousands of unlinks
468 * filling persistent llogs and in-core respresentation
470 * this doesn't scale at all. so we need basically the following:
471 * a) destroy/setattr append llog records
472 * b) once llog has grown to X records, we process first Y committed records
474 * once record R is found via llog_process(), it becomes committed after any
475 * subsequent commit callback (at the most)
479 * ptlrpc commit callback.
481 * The callback is called by PTLRPC when a RPC is reported committed by the
482 * target (OST). We register the callback for the every RPC applying a change
483 * from the llog. This way we know then the llog records can be cancelled.
484 * Notice the callback can be called when OSP is finishing. We can detect this
485 * checking that actual transno in the request is less or equal of known
486 * committed transno (see osp_sync_process_committed() for the details).
487 * XXX: this is pretty expensive and can be improved later using batching.
489 * \param[in] req request
491 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
493 struct osp_device *d = req->rq_cb_data;
494 struct osp_job_req_args *jra;
496 CDEBUG(D_HA, "commit req %p, transno %llu\n", req, req->rq_transno);
498 if (unlikely(req->rq_transno == 0))
501 /* do not do any opd_sync_rpcs_* accounting here
502 * it's done in osp_sync_interpret sooner or later */
505 jra = ptlrpc_req_async_args(req);
506 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
507 LASSERT(list_empty(&jra->jra_committed_link));
509 ptlrpc_request_addref(req);
511 spin_lock(&d->opd_sync_lock);
512 list_add(&jra->jra_committed_link, &d->opd_sync_committed_there);
513 spin_unlock(&d->opd_sync_lock);
515 /* XXX: some batching wouldn't hurt */
516 wake_up(&d->opd_sync_waitq);
520 * RPC interpretation callback.
522 * The callback is called by ptlrpc when RPC is replied. Now we have to decide
524 * - put request on a special list to wait until it's committed by the target,
525 * if the request is successful
526 * - schedule llog record cancel if no target object is found
527 * - try later (essentially after reboot) in case of unexpected error
529 * \param[in] env LU environment provided by the caller
530 * \param[in] req request replied
531 * \param[in] aa callback data
532 * \param[in] rc result of RPC
536 static int osp_sync_interpret(const struct lu_env *env,
537 struct ptlrpc_request *req, void *aa, int rc)
539 struct osp_device *d = req->rq_cb_data;
540 struct osp_job_req_args *jra = aa;
542 if (jra->jra_magic != OSP_JOB_MAGIC) {
543 DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
548 CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
549 atomic_read(&req->rq_refcount),
550 rc, (unsigned) req->rq_transno);
551 LASSERT(rc || req->rq_transno);
555 * we tried to destroy object or update attributes,
556 * but object doesn't exist anymore - cancell llog record
558 LASSERT(req->rq_transno == 0);
559 LASSERT(list_empty(&jra->jra_committed_link));
561 ptlrpc_request_addref(req);
563 spin_lock(&d->opd_sync_lock);
564 list_add(&jra->jra_committed_link,
565 &d->opd_sync_committed_there);
566 spin_unlock(&d->opd_sync_lock);
568 wake_up(&d->opd_sync_waitq);
570 struct obd_import *imp = req->rq_import;
572 * error happened, we'll try to repeat on next boot ?
574 LASSERTF(req->rq_transno == 0 ||
575 req->rq_import_generation < imp->imp_generation,
576 "transno %llu, rc %d, gen: req %d, imp %d\n",
577 req->rq_transno, rc, req->rq_import_generation,
578 imp->imp_generation);
579 if (req->rq_transno == 0) {
580 /* this is the last time we see the request
581 * if transno is not zero, then commit cb
582 * will be called at some point */
583 LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) > 0);
584 atomic_dec(&d->opd_sync_rpcs_in_progress);
587 wake_up(&d->opd_sync_waitq);
588 } else if (d->opd_pre != NULL &&
589 unlikely(d->opd_pre_status == -ENOSPC)) {
591 * if current status is -ENOSPC (lack of free space on OST)
592 * then we should poll OST immediately once object destroy
595 osp_statfs_need_now(d);
598 spin_lock(&d->opd_sync_lock);
599 list_del_init(&jra->jra_in_flight_link);
600 spin_unlock(&d->opd_sync_lock);
601 LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) > 0);
602 atomic_dec(&d->opd_sync_rpcs_in_flight);
603 if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
604 wake_up(&d->opd_sync_barrier_waitq);
605 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
606 d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
607 atomic_read(&d->opd_sync_rpcs_in_progress));
609 osp_sync_check_for_work(d);
615 ** Add request to ptlrpc queue.
617 * This is just a tiny helper function to put the request on the sending list
619 * \param[in] d OSP device
620 * \param[in] llh llog handle where the record is stored
621 * \param[in] h llog record
622 * \param[in] req request
624 static void osp_sync_send_new_rpc(struct osp_device *d,
625 struct llog_handle *llh,
626 struct llog_rec_hdr *h,
627 struct ptlrpc_request *req)
629 struct osp_job_req_args *jra;
631 LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) <=
632 d->opd_sync_max_rpcs_in_flight);
634 jra = ptlrpc_req_async_args(req);
635 jra->jra_magic = OSP_JOB_MAGIC;
636 jra->jra_lcookie.lgc_lgl = llh->lgh_id;
637 jra->jra_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
638 jra->jra_lcookie.lgc_index = h->lrh_index;
639 INIT_LIST_HEAD(&jra->jra_committed_link);
640 spin_lock(&d->opd_sync_lock);
641 list_add_tail(&jra->jra_in_flight_link, &d->opd_sync_in_flight_list);
642 spin_unlock(&d->opd_sync_lock);
644 ptlrpcd_add_req(req);
649 * Allocate and prepare RPC for a new change.
651 * The function allocates and initializes an RPC which will be sent soon to
652 * apply the change to the target OST. The request is initialized from the
653 * llog record passed. Notice only the fields common to all type of changes
656 * \param[in] d OSP device
657 * \param[in] op type of the change
658 * \param[in] format request format to be used
660 * \retval pointer new request on success
661 * \retval ERR_PTR(errno) on error
663 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
665 const struct req_format *format)
667 struct ptlrpc_request *req;
668 struct obd_import *imp;
671 /* Prepare the request */
672 imp = d->opd_obd->u.cli.cl_import;
675 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
676 RETURN(ERR_PTR(-ENOMEM));
678 req = ptlrpc_request_alloc(imp, format);
680 RETURN(ERR_PTR(-ENOMEM));
682 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
684 ptlrpc_req_finished(req);
688 req->rq_interpret_reply = osp_sync_interpret;
689 req->rq_commit_cb = osp_sync_request_commit_cb;
692 ptlrpc_request_set_replen(req);
698 * Generate a request for setattr change.
700 * The function prepares a new RPC, initializes it with setattr specific
701 * bits and send the RPC.
703 * \param[in] d OSP device
704 * \param[in] llh llog handle where the record is stored
705 * \param[in] h llog record
707 * \retval 0 on success
708 * \retval 1 on invalid record
709 * \retval negative negated errno on error
711 static int osp_sync_new_setattr_job(struct osp_device *d,
712 struct llog_handle *llh,
713 struct llog_rec_hdr *h)
715 struct llog_setattr64_rec *rec = (struct llog_setattr64_rec *)h;
716 struct ptlrpc_request *req;
717 struct ost_body *body;
720 LASSERT(h->lrh_type == MDS_SETATTR64_REC);
722 if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
725 /* lsr_valid can only be 0 or HAVE OBD_MD_{FLUID, FLGID, FLPROJID} set,
726 * so no bits other than these should be set. */
727 if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID |
728 OBD_MD_FLPROJID)) != 0) {
729 CERROR("%s: invalid setattr record, lsr_valid:%llu\n",
730 d->opd_obd->obd_name, rec->lsr_valid);
731 /* return 1 on invalid record */
735 req = osp_sync_new_job(d, OST_SETATTR, &RQF_OST_SETATTR);
737 RETURN(PTR_ERR(req));
739 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
741 body->oa.o_oi = rec->lsr_oi;
742 body->oa.o_uid = rec->lsr_uid;
743 body->oa.o_gid = rec->lsr_gid;
744 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
745 if (h->lrh_len > sizeof(struct llog_setattr64_rec))
746 body->oa.o_projid = ((struct llog_setattr64_rec_v2 *)
749 /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
750 * we assume that both UID and GID are valid in that case. */
751 if (rec->lsr_valid == 0)
752 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
754 body->oa.o_valid |= rec->lsr_valid;
756 osp_sync_send_new_rpc(d, llh, h, req);
761 * Generate a request for unlink change.
763 * The function prepares a new RPC, initializes it with unlink(destroy)
764 * specific bits and sends the RPC. The function is used to handle
765 * llog_unlink_rec which were used in the older versions of Lustre.
766 * Current version uses llog_unlink_rec64.
768 * \param[in] d OSP device
769 * \param[in] llh llog handle where the record is stored
770 * \param[in] h llog record
772 * \retval 0 on success
773 * \retval negative negated errno on error
775 static int osp_sync_new_unlink_job(struct osp_device *d,
776 struct llog_handle *llh,
777 struct llog_rec_hdr *h)
779 struct llog_unlink_rec *rec = (struct llog_unlink_rec *)h;
780 struct ptlrpc_request *req;
781 struct ost_body *body;
785 LASSERT(h->lrh_type == MDS_UNLINK_REC);
787 req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
789 RETURN(PTR_ERR(req));
791 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
793 ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
794 rc = ostid_set_id(&body->oa.o_oi, rec->lur_oid);
797 body->oa.o_misc = rec->lur_count;
798 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
800 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
802 osp_sync_send_new_rpc(d, llh, h, req);
807 * Generate a request for unlink change.
809 * The function prepares a new RPC, initializes it with unlink(destroy)
810 * specific bits and sends the RPC. Depending on the target (MDT or OST)
811 * two different protocols are used. For MDT we use OUT (basically OSD API
812 * updates transferred via a network). For OST we still use the old
813 * protocol (OBD?), originally for compatibility. Later we can start to
814 * use OUT for OST as well, this will allow batching and better code
817 * \param[in] d OSP device
818 * \param[in] llh llog handle where the record is stored
819 * \param[in] h llog record
821 * \retval 0 on success
822 * \retval negative negated errno on error
824 static int osp_sync_new_unlink64_job(struct osp_device *d,
825 struct llog_handle *llh,
826 struct llog_rec_hdr *h)
828 struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
829 struct ptlrpc_request *req = NULL;
830 struct ost_body *body;
834 LASSERT(h->lrh_type == MDS_UNLINK64_REC);
835 req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
837 RETURN(PTR_ERR(req));
839 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
842 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
845 body->oa.o_misc = rec->lur_count;
846 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
848 osp_sync_send_new_rpc(d, llh, h, req);
853 * Process llog records.
855 * This function is called to process the llog records committed locally.
856 * In the recovery model used by OSP we can apply a change to a remote
857 * target once corresponding transaction (like posix unlink) is committed
858 * locally so can't revert.
859 * Depending on the llog record type, a given handler is called that is
860 * responsible for preparing and sending the RPC to apply the change.
861 * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
863 * \param[in] env LU environment provided by the caller
864 * \param[in] d OSP device
865 * \param[in] llh llog handle where the record is stored
866 * \param[in] rec llog record
868 static void osp_sync_process_record(const struct lu_env *env,
869 struct osp_device *d,
870 struct llog_handle *llh,
871 struct llog_rec_hdr *rec)
873 struct llog_handle *cathandle = llh->u.phd.phd_cat_handle;
874 struct llog_cookie cookie;
879 cookie.lgc_lgl = llh->lgh_id;
880 cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
881 cookie.lgc_index = rec->lrh_index;
883 if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
884 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
886 /* we're waiting for the record generated by this instance */
887 LASSERT(d->opd_sync_prev_done == 0);
888 if (!memcmp(&d->opd_sync_generation, &gen->lgr_gen,
889 sizeof(gen->lgr_gen))) {
890 CDEBUG(D_HA, "processed all old entries\n");
891 d->opd_sync_prev_done = 1;
894 /* cancel any generation record */
895 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
901 * now we prepare and fill requests to OST, put them on the queue
902 * and fire after next commit callback
905 /* notice we increment counters before sending RPC, to be consistent
906 * in RPC interpret callback which may happen very quickly */
907 atomic_inc(&d->opd_sync_rpcs_in_flight);
908 atomic_inc(&d->opd_sync_rpcs_in_progress);
910 switch (rec->lrh_type) {
911 /* case MDS_UNLINK_REC is kept for compatibility */
913 rc = osp_sync_new_unlink_job(d, llh, rec);
915 case MDS_UNLINK64_REC:
916 rc = osp_sync_new_unlink64_job(d, llh, rec);
918 case MDS_SETATTR64_REC:
919 rc = osp_sync_new_setattr_job(d, llh, rec);
922 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
924 /* treat "unknown record type" as "invalid" */
929 /* For all kinds of records, not matter successful or not,
930 * we should decrease changes and bump last_processed_id.
932 if (d->opd_sync_prev_done) {
933 __u64 correct_id = osp_sync_correct_id(d, rec);
934 LASSERT(atomic_read(&d->opd_sync_changes) > 0);
935 LASSERT(correct_id <= d->opd_sync_last_committed_id);
936 /* NOTE: it's possible to meet same id if
937 * OST stores few stripes of same file
940 /* another thread may be trying to set new value */
942 if (correct_id > d->opd_sync_last_processed_id) {
943 d->opd_sync_last_processed_id = correct_id;
944 wake_up(&d->opd_sync_barrier_waitq);
948 atomic_dec(&d->opd_sync_changes);
951 atomic_dec(&d->opd_sync_rpcs_in_flight);
952 atomic_dec(&d->opd_sync_rpcs_in_progress);
955 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
956 d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
957 atomic_read(&d->opd_sync_rpcs_in_progress));
959 /* Delete the invalid record */
961 rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
963 CERROR("%s: can't delete invalid record: "
964 "fid = "DFID", rec_id = %u, rc = %d\n",
965 d->opd_obd->obd_name,
966 PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
970 CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
971 rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);
977 * Cancel llog records for the committed changes.
979 * The function walks through the list of the committed RPCs and cancels
980 * corresponding llog records. see osp_sync_request_commit_cb() for the
983 * \param[in] env LU environment provided by the caller
984 * \param[in] d OSP device
986 static void osp_sync_process_committed(const struct lu_env *env,
987 struct osp_device *d)
989 struct obd_device *obd = d->opd_obd;
990 struct obd_import *imp = obd->u.cli.cl_import;
991 struct ost_body *body;
992 struct ptlrpc_request *req;
993 struct llog_ctxt *ctxt;
994 struct llog_handle *llh;
995 struct list_head list;
1000 if (list_empty(&d->opd_sync_committed_there))
1004 * if current status is -ENOSPC (lack of free space on OST)
1005 * then we should poll OST immediately once object destroy
1007 * notice: we do this upon commit as well because some backends
1008 * (like DMU) do not release space right away.
1010 if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
1011 osp_statfs_need_now(d);
1014 * now cancel them all
1015 * XXX: can we improve this using some batching?
1016 * with batch RPC that'll happen automatically?
1017 * XXX: can we store ctxt in lod_device and save few cycles ?
1019 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1022 llh = ctxt->loc_handle;
1025 INIT_LIST_HEAD(&list);
1026 spin_lock(&d->opd_sync_lock);
1027 list_splice(&d->opd_sync_committed_there, &list);
1028 INIT_LIST_HEAD(&d->opd_sync_committed_there);
1029 spin_unlock(&d->opd_sync_lock);
1031 while (!list_empty(&list)) {
1032 struct osp_job_req_args *jra;
1034 jra = list_entry(list.next, struct osp_job_req_args,
1035 jra_committed_link);
1036 LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
1037 list_del_init(&jra->jra_committed_link);
1039 req = container_of((void *)jra, struct ptlrpc_request,
1041 body = req_capsule_client_get(&req->rq_pill,
1044 /* import can be closing, thus all commit cb's are
1045 * called we can check committness directly */
1046 if (req->rq_import_generation == imp->imp_generation) {
1047 rc = llog_cat_cancel_records(env, llh, 1,
1050 CERROR("%s: can't cancel record: %d\n",
1053 DEBUG_REQ(D_OTHER, req, "imp_committed = %llu",
1054 imp->imp_peer_committed_transno);
1056 ptlrpc_req_finished(req);
1060 llog_ctxt_put(ctxt);
1062 LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) >= done);
1063 atomic_sub(done, &d->opd_sync_rpcs_in_progress);
1064 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
1065 d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
1066 atomic_read(&d->opd_sync_rpcs_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_sync_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;
1102 struct l_wait_info lwi = { 0 };
1104 if (!osp_sync_running(d)) {
1105 CDEBUG(D_HA, "stop llog processing\n");
1106 return LLOG_PROC_BREAK;
1109 /* process requests committed by OST */
1110 osp_sync_process_committed(env, d);
1112 /* if we there are changes to be processed and we have
1113 * resources for this ... do now */
1114 if (osp_sync_can_process_new(d, rec)) {
1116 /* ask llog for another record */
1117 CDEBUG(D_HA, "%u changes, %u in progress,"
1119 atomic_read(&d->opd_sync_changes),
1120 atomic_read(&d->opd_sync_rpcs_in_progress),
1121 atomic_read(&d->opd_sync_rpcs_in_flight));
1124 osp_sync_process_record(env, d, llh, rec);
1129 if (d->opd_sync_last_processed_id == d->opd_sync_last_used_id)
1130 osp_sync_remove_from_tracker(d);
1132 l_wait_event(d->opd_sync_waitq,
1133 !osp_sync_running(d) ||
1134 osp_sync_can_process_new(d, rec) ||
1135 !list_empty(&d->opd_sync_committed_there),
1143 * This thread runs llog_cat_process() scanner calling our callback
1144 * to process llog records. in the callback we implement tricky
1145 * state machine as we don't want to start scanning of the llog again
1146 * and again, also we don't want to process too many records and send
1147 * too many RPCs a time. so, depending on current load (num of changes
1148 * being synced to OST) the callback can suspend awaiting for some
1149 * new conditions, like syncs completed.
1151 * In order to process llog records left by previous boots and to allow
1152 * llog_process_thread() to find something (otherwise it'd just exit
1153 * immediately) we add a special GENERATATION record on each boot.
1155 * \param[in] _arg a pointer to thread's arguments
1157 * \retval 0 on success
1158 * \retval negative negated errno on error
1160 static int osp_sync_thread(void *_arg)
1162 struct osp_device *d = _arg;
1163 struct ptlrpc_thread *thread = &d->opd_sync_thread;
1164 struct l_wait_info lwi = { 0 };
1165 struct llog_ctxt *ctxt;
1166 struct obd_device *obd = d->opd_obd;
1167 struct llog_handle *llh;
1173 rc = lu_env_init(&env, LCT_LOCAL);
1175 CERROR("%s: can't initialize env: rc = %d\n",
1178 spin_lock(&d->opd_sync_lock);
1179 thread->t_flags = SVC_STOPPED;
1180 spin_unlock(&d->opd_sync_lock);
1181 wake_up(&thread->t_ctl_waitq);
1186 spin_lock(&d->opd_sync_lock);
1187 thread->t_flags = SVC_RUNNING;
1188 spin_unlock(&d->opd_sync_lock);
1189 wake_up(&thread->t_ctl_waitq);
1191 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1193 CERROR("can't get appropriate context\n");
1194 GOTO(out, rc = -EINVAL);
1197 llh = ctxt->loc_handle;
1199 CERROR("can't get llh\n");
1200 llog_ctxt_put(ctxt);
1201 GOTO(out, rc = -EINVAL);
1204 rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
1206 CERROR("%s: llog process with osp_sync_process_queues "
1207 "failed: %d\n", d->opd_obd->obd_name, rc);
1210 LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
1211 "%u changes, %u in progress, %u in flight: %d\n",
1212 atomic_read(&d->opd_sync_changes),
1213 atomic_read(&d->opd_sync_rpcs_in_progress),
1214 atomic_read(&d->opd_sync_rpcs_in_flight), rc);
1216 /* we don't expect llog_process_thread() to exit till umount */
1217 LASSERTF(thread->t_flags != SVC_RUNNING,
1218 "%u changes, %u in progress, %u in flight\n",
1219 atomic_read(&d->opd_sync_changes),
1220 atomic_read(&d->opd_sync_rpcs_in_progress),
1221 atomic_read(&d->opd_sync_rpcs_in_flight));
1223 /* wait till all the requests are completed */
1225 while (atomic_read(&d->opd_sync_rpcs_in_progress) > 0) {
1226 osp_sync_process_committed(&env, d);
1228 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
1229 rc = l_wait_event(d->opd_sync_waitq,
1230 atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1232 if (rc == -ETIMEDOUT)
1234 LASSERTF(count < 10, "%s: %d %d %sempty\n",
1235 d->opd_obd->obd_name,
1236 atomic_read(&d->opd_sync_rpcs_in_progress),
1237 atomic_read(&d->opd_sync_rpcs_in_flight),
1238 list_empty(&d->opd_sync_committed_there) ? "" : "!");
1243 llog_cat_close(&env, llh);
1244 rc = llog_cleanup(&env, ctxt);
1246 CERROR("can't cleanup llog: %d\n", rc);
1248 LASSERTF(atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
1249 "%s: %d %d %sempty\n", d->opd_obd->obd_name,
1250 atomic_read(&d->opd_sync_rpcs_in_progress),
1251 atomic_read(&d->opd_sync_rpcs_in_flight),
1252 list_empty(&d->opd_sync_committed_there) ? "" : "!");
1254 thread->t_flags = SVC_STOPPED;
1256 wake_up(&thread->t_ctl_waitq);
1266 * Initializes the llog. Specific llog to be used depends on the type of the
1267 * target OSP represents (OST or MDT). The function adds appends a new llog
1268 * record to mark the place where the records associated with this boot
1271 * \param[in] env LU environment provided by the caller
1272 * \param[in] d OSP device
1274 * \retval 0 on success
1275 * \retval negative negated errno on error
1277 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1279 struct osp_thread_info *osi = osp_env_info(env);
1280 struct lu_fid *fid = &osi->osi_fid;
1281 struct llog_handle *lgh = NULL;
1282 struct obd_device *obd = d->opd_obd;
1283 struct llog_ctxt *ctxt;
1291 * open llog corresponding to our OST
1293 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1294 obd->obd_lvfs_ctxt.dt = d->opd_storage;
1296 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1298 rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1299 &osi->osi_cid, fid);
1301 if (rc != -EFAULT) {
1302 CERROR("%s: can't get id from catalogs: rc = %d\n",
1307 /* After sparse OST indices is supported, the CATALOG file
1308 * may become a sparse file that results in failure on
1309 * reading. Skip this error as the llog will be created
1311 memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
1315 CDEBUG(D_INFO, "%s: Init llog for %d - catid "DFID":%x\n",
1316 obd->obd_name, d->opd_index,
1317 PFID(&osi->osi_cid.lci_logid.lgl_oi.oi_fid),
1318 osi->osi_cid.lci_logid.lgl_ogen);
1320 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT,
1321 d->opd_storage->dd_lu_dev.ld_obd,
1322 &osp_mds_ost_orig_logops);
1326 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1329 if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1330 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1332 /* re-create llog if it is missing */
1334 logid_set_id(&osi->osi_cid.lci_logid, 0);
1336 GOTO(out_cleanup, rc);
1339 if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1340 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1342 GOTO(out_cleanup, rc);
1343 osi->osi_cid.lci_logid = lgh->lgh_id;
1346 LASSERT(lgh != NULL);
1347 ctxt->loc_handle = lgh;
1349 rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
1351 GOTO(out_close, rc);
1353 rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1354 &osi->osi_cid, fid);
1356 GOTO(out_close, rc);
1359 * put a mark in the llog till which we'll be processing
1360 * old records restless
1362 d->opd_sync_generation.mnt_cnt = cfs_time_current();
1363 d->opd_sync_generation.conn_cnt = cfs_time_current();
1365 osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1366 osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1368 memcpy(&osi->osi_gen.lgr_gen, &d->opd_sync_generation,
1369 sizeof(osi->osi_gen.lgr_gen));
1371 rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1373 GOTO(out_close, rc);
1374 llog_ctxt_put(ctxt);
1377 llog_cat_close(env, lgh);
1379 llog_cleanup(env, ctxt);
1384 * Cleanup llog used for syncing.
1386 * Closes and cleanups the llog. The function is called when the device is
1389 * \param[in] env LU environment provided by the caller
1390 * \param[in] d OSP device
1392 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1394 struct llog_ctxt *ctxt;
1396 ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1398 llog_cat_close(env, ctxt->loc_handle);
1399 llog_cleanup(env, ctxt);
1404 * Initialization of the sync component of OSP.
1406 * Initializes the llog and starts a new thread to handle the changes to
1407 * the remote target (OST or MDT).
1409 * \param[in] env LU environment provided by the caller
1410 * \param[in] d OSP device
1412 * \retval 0 on success
1413 * \retval negative negated errno on error
1415 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1417 struct l_wait_info lwi = { 0 };
1418 struct task_struct *task;
1423 d->opd_sync_max_rpcs_in_flight = OSP_MAX_RPCS_IN_FLIGHT;
1424 d->opd_sync_max_rpcs_in_progress = OSP_MAX_RPCS_IN_PROGRESS;
1425 spin_lock_init(&d->opd_sync_lock);
1426 init_waitqueue_head(&d->opd_sync_waitq);
1427 init_waitqueue_head(&d->opd_sync_barrier_waitq);
1428 thread_set_flags(&d->opd_sync_thread, SVC_INIT);
1429 init_waitqueue_head(&d->opd_sync_thread.t_ctl_waitq);
1430 INIT_LIST_HEAD(&d->opd_sync_in_flight_list);
1431 INIT_LIST_HEAD(&d->opd_sync_committed_there);
1433 if (d->opd_storage->dd_rdonly)
1436 rc = osp_sync_id_traction_init(d);
1441 * initialize llog storing changes
1443 rc = osp_sync_llog_init(env, d);
1445 CERROR("%s: can't initialize llog: rc = %d\n",
1446 d->opd_obd->obd_name, rc);
1451 * Start synchronization thread
1453 task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1454 d->opd_index, d->opd_group);
1457 CERROR("%s: cannot start sync thread: rc = %d\n",
1458 d->opd_obd->obd_name, rc);
1462 l_wait_event(d->opd_sync_thread.t_ctl_waitq,
1463 osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1467 osp_sync_llog_fini(env, d);
1469 osp_sync_id_traction_fini(d);
1474 * Stop the syncing thread.
1476 * Asks the syncing thread to stop and wait until it's stopped.
1478 * \param[in] d OSP device
1482 int osp_sync_fini(struct osp_device *d)
1484 struct ptlrpc_thread *thread = &d->opd_sync_thread;
1488 if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1489 thread->t_flags = SVC_STOPPING;
1490 wake_up(&d->opd_sync_waitq);
1491 wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
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: %llu -> %llu\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_sync_last_committed_id = tr->otr_committed_id;
1537 wake_up(&d->opd_sync_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_sync_tracker == NULL);
1570 INIT_LIST_HEAD(&d->opd_sync_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_sync_tracker = tr;
1583 if (found == NULL) {
1587 d->opd_sync_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_sync_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_sync_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
1652 * \param[in] d OSP device
1653 * \param[in] id 0 or ID generated previously
1655 * \retval ID the caller should use
1657 static __u64 osp_sync_id_get(struct osp_device *d, __u64 id)
1659 struct osp_id_tracker *tr;
1661 tr = d->opd_sync_tracker;
1664 /* XXX: we can improve this introducing per-cpu preallocated ids? */
1665 spin_lock(&tr->otr_lock);
1666 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_TRACK_OVERFLOW))
1667 tr->otr_next_id = 0xfffffff0;
1669 if (unlikely(tr->otr_next_id <= d->opd_sync_last_used_id)) {
1670 spin_unlock(&tr->otr_lock);
1671 CERROR("%s: next %llu, last synced %llu\n",
1672 d->opd_obd->obd_name, tr->otr_next_id,
1673 d->opd_sync_last_used_id);
1678 id = tr->otr_next_id++;
1679 if (id > d->opd_sync_last_used_id)
1680 d->opd_sync_last_used_id = id;
1681 if (list_empty(&d->opd_sync_ontrack))
1682 list_add(&d->opd_sync_ontrack, &tr->otr_wakeup_list);
1683 spin_unlock(&tr->otr_lock);
1684 CDEBUG(D_OTHER, "new id %llu\n", id);
1690 * Stop to propagate commit status to OSP.
1692 * If the OSP does not have any llog records she's waiting to commit, then
1693 * it is possible to unsubscribe from wakeups from the tracking using this
1696 * \param[in] d OSP device not willing to wakeup
1698 static void osp_sync_remove_from_tracker(struct osp_device *d)
1700 struct osp_id_tracker *tr;
1702 tr = d->opd_sync_tracker;
1705 if (list_empty(&d->opd_sync_ontrack))
1708 spin_lock(&tr->otr_lock);
1709 list_del_init(&d->opd_sync_ontrack);
1710 spin_unlock(&tr->otr_lock);