4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osp/osp_dev.c
38 * Lustre OST Proxy Device
40 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
41 * Author: Mikhail Pershin <mike.pershin@intel.com>
42 * Author: Di Wang <di.wang@intel.com>
45 #define DEBUG_SUBSYSTEM S_MDS
47 #include <obd_class.h>
48 #include <lustre_ioctl.h>
49 #include <lustre_param.h>
50 #include <lustre_log.h>
51 #include <lustre_mdc.h>
53 #include "osp_internal.h"
55 /* Slab for OSP object allocation */
56 struct kmem_cache *osp_object_kmem;
58 static struct lu_kmem_descr osp_caches[] = {
60 .ckd_cache = &osp_object_kmem,
61 .ckd_name = "osp_obj",
62 .ckd_size = sizeof(struct osp_object)
69 struct lu_object *osp_object_alloc(const struct lu_env *env,
70 const struct lu_object_header *hdr,
73 struct lu_object_header *h = NULL;
77 OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, GFP_NOFS);
79 l = &o->opo_obj.do_lu;
81 /* If hdr is NULL, it means the object is not built
82 * from the top dev(MDT/OST), usually it happens when
83 * building striped object, like data object on MDT or
84 * striped object for directory */
87 lu_object_header_init(h);
88 dt_object_init(&o->opo_obj, h, d);
89 lu_object_add_top(h, l);
91 dt_object_init(&o->opo_obj, h, d);
94 l->lo_ops = &osp_lu_obj_ops;
102 static struct dt_object
103 *osp_find_or_create(const struct lu_env *env, struct osp_device *osp,
104 struct lu_attr *attr, __u32 reg_id)
106 struct osp_thread_info *osi = osp_env_info(env);
107 struct dt_object_format dof = { 0 };
108 struct dt_object *dto;
112 lu_local_obj_fid(&osi->osi_fid, reg_id);
113 attr->la_valid = LA_MODE;
114 attr->la_mode = S_IFREG | 0644;
115 dof.dof_type = DFT_REGULAR;
116 dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
121 rc = dt_attr_get(env, dto, attr, NULL);
123 CERROR("%s: can't be initialized: rc = %d\n",
124 osp->opd_obd->obd_name, rc);
125 lu_object_put(env, &dto->do_lu);
131 static int osp_write_local_file(const struct lu_env *env,
132 struct osp_device *osp,
133 struct dt_object *dt_obj,
140 th = dt_trans_create(env, osp->opd_storage);
144 rc = dt_declare_record_write(env, dt_obj, buf, offset, th);
147 rc = dt_trans_start_local(env, osp->opd_storage, th);
151 rc = dt_record_write(env, dt_obj, buf, &offset, th);
153 dt_trans_stop(env, osp->opd_storage, th);
157 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
159 struct osp_thread_info *osi = osp_env_info(env);
160 struct lu_fid *fid = &osp->opd_last_used_fid;
161 struct dt_object *dto;
165 dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OID);
167 RETURN(PTR_ERR(dto));
168 /* object will be released in device cleanup path */
169 if (osi->osi_attr.la_size >=
170 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
171 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
173 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
178 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
180 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
183 osp->opd_last_used_oid_file = dto;
186 /* object will be released in device cleanup path */
187 CERROR("%s: can't initialize lov_objid: rc = %d\n",
188 osp->opd_obd->obd_name, rc);
189 lu_object_put(env, &dto->do_lu);
190 osp->opd_last_used_oid_file = NULL;
194 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
196 struct osp_thread_info *osi = osp_env_info(env);
197 struct lu_fid *fid = &osp->opd_last_used_fid;
198 struct dt_object *dto;
202 dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OSEQ);
204 RETURN(PTR_ERR(dto));
206 /* object will be released in device cleanup path */
207 if (osi->osi_attr.la_size >=
208 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
209 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
211 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
216 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
218 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
221 osp->opd_last_used_seq_file = dto;
224 /* object will be released in device cleanup path */
225 CERROR("%s: can't initialize lov_seq: rc = %d\n",
226 osp->opd_obd->obd_name, rc);
227 lu_object_put(env, &dto->do_lu);
228 osp->opd_last_used_seq_file = NULL;
232 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
234 struct osp_thread_info *osi = osp_env_info(env);
238 fid_zero(&osp->opd_last_used_fid);
239 rc = osp_init_last_objid(env, osp);
241 CERROR("%s: Can not get ids %d from old objid!\n",
242 osp->opd_obd->obd_name, rc);
246 rc = osp_init_last_seq(env, osp);
248 CERROR("%s: Can not get ids %d from old objid!\n",
249 osp->opd_obd->obd_name, rc);
253 if (fid_oid(&osp->opd_last_used_fid) != 0 &&
254 fid_seq(&osp->opd_last_used_fid) == 0) {
255 /* Just upgrade from the old version,
256 * set the seq to be IDIF */
257 osp->opd_last_used_fid.f_seq =
258 fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
260 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
261 &osp->opd_last_used_fid.f_seq,
263 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
264 &osi->osi_lb, osi->osi_off);
266 CERROR("%s : Can not write seq file: rc = %d\n",
267 osp->opd_obd->obd_name, rc);
272 if (!fid_is_zero(&osp->opd_last_used_fid) &&
273 !fid_is_sane(&osp->opd_last_used_fid)) {
274 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
275 PFID(&osp->opd_last_used_fid));
276 GOTO(out, rc = -EINVAL);
279 CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
280 osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
283 if (osp->opd_last_used_oid_file != NULL) {
284 lu_object_put(env, &osp->opd_last_used_oid_file->do_lu);
285 osp->opd_last_used_oid_file = NULL;
287 if (osp->opd_last_used_seq_file != NULL) {
288 lu_object_put(env, &osp->opd_last_used_seq_file->do_lu);
289 osp->opd_last_used_seq_file = NULL;
296 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *d)
298 /* release last_used file */
299 if (d->opd_last_used_oid_file != NULL) {
300 lu_object_put(env, &d->opd_last_used_oid_file->do_lu);
301 d->opd_last_used_oid_file = NULL;
304 if (d->opd_last_used_seq_file != NULL) {
305 lu_object_put(env, &d->opd_last_used_seq_file->do_lu);
306 d->opd_last_used_seq_file = NULL;
310 static int osp_disconnect(struct osp_device *d)
312 struct obd_import *imp;
315 imp = d->opd_obd->u.cli.cl_import;
317 /* Mark import deactivated now, so we don't try to reconnect if any
318 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
319 * fully deactivate the import, or that would drop all requests. */
320 LASSERT(imp != NULL);
321 spin_lock(&imp->imp_lock);
322 imp->imp_deactive = 1;
323 spin_unlock(&imp->imp_lock);
325 ptlrpc_deactivate_import(imp);
327 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
328 * delete it regardless. (It's safe to delete an import that was
330 (void)ptlrpc_pinger_del_import(imp);
332 rc = ptlrpc_disconnect_import(imp, 0);
334 CERROR("%s: can't disconnect: rc = %d\n",
335 d->opd_obd->obd_name, rc);
337 ptlrpc_invalidate_import(imp);
342 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
349 rc = osp_disconnect(d);
353 if (!d->opd_connect_mdt) {
354 /* stop precreate thread */
355 osp_precreate_fini(d);
357 /* release last_used file */
358 osp_last_used_fini(env, d);
361 obd_fid_fini(d->opd_obd);
366 static int osp_process_config(const struct lu_env *env,
367 struct lu_device *dev, struct lustre_cfg *lcfg)
369 struct osp_device *d = lu2osp_dev(dev);
370 struct obd_device *obd = d->opd_obd;
375 switch (lcfg->lcfg_command) {
376 case LCFG_PRE_CLEANUP:
377 rc = osp_disconnect(d);
380 lu_dev_del_linkage(dev->ld_site, dev);
381 rc = osp_shutdown(env, d);
385 rc = class_process_proc_seq_param(PARAM_OSC, obd->obd_vars,
390 /* class_process_proc_param() haven't found matching
391 * parameter and returned ENOSYS so that layer(s)
392 * below could use that. But OSP is the bottom, so
394 CERROR("%s: unknown param %s\n",
395 (char *)lustre_cfg_string(lcfg, 0),
396 (char *)lustre_cfg_string(lcfg, 1));
401 CERROR("%s: unknown command %u\n",
402 (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
410 static int osp_recovery_complete(const struct lu_env *env,
411 struct lu_device *dev)
413 struct osp_device *osp = lu2osp_dev(dev);
417 osp->opd_recovery_completed = 1;
419 if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
420 wake_up(&osp->opd_pre_waitq);
424 const struct lu_device_operations osp_lu_ops = {
425 .ldo_object_alloc = osp_object_alloc,
426 .ldo_process_config = osp_process_config,
427 .ldo_recovery_complete = osp_recovery_complete,
431 * provides with statfs from corresponded OST
434 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
435 struct obd_statfs *sfs)
437 struct osp_device *d = dt2osp_dev(dev);
441 if (unlikely(d->opd_imp_active == 0))
444 if (d->opd_pre == NULL)
447 /* return recently updated data */
448 *sfs = d->opd_statfs;
451 * layer above osp (usually lod) can use ffree to estimate
452 * how many objects are available for immediate creation
454 spin_lock(&d->opd_pre_lock);
455 LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
456 fid_seq(&d->opd_pre_used_fid),
457 "last_created "DFID", next_fid "DFID"\n",
458 PFID(&d->opd_pre_last_created_fid),
459 PFID(&d->opd_pre_used_fid));
460 sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
461 fid_oid(&d->opd_pre_used_fid);
462 sfs->os_fprecreated -= d->opd_pre_reserved;
463 spin_unlock(&d->opd_pre_lock);
465 LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2);
467 CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
468 LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
469 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
470 sfs->os_files, sfs->os_ffree);
474 static int osp_sync_timeout(void *data)
479 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
481 struct osp_device *d = dt2osp_dev(dev);
483 struct l_wait_info lwi = { 0 };
484 unsigned long id, old;
486 unsigned long start = cfs_time_current();
489 if (unlikely(d->opd_imp_active == 0))
492 id = d->opd_syn_last_used_id;
494 CDEBUG(D_OTHER, "%s: id: used %lu, processed %lu\n",
495 d->opd_obd->obd_name, id, d->opd_syn_last_processed_id);
497 /* wait till all-in-line are processed */
498 while (d->opd_syn_last_processed_id < id) {
500 old = d->opd_syn_last_processed_id;
502 /* make sure the connection is fine */
503 expire = cfs_time_shift(obd_timeout);
504 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
505 osp_sync_timeout, d);
506 l_wait_event(d->opd_syn_barrier_waitq,
507 d->opd_syn_last_processed_id >= id,
510 if (d->opd_syn_last_processed_id >= id)
513 if (d->opd_syn_last_processed_id != old) {
514 /* some progress have been made,
519 /* no changes and expired, something is wrong */
520 GOTO(out, rc = -ETIMEDOUT);
523 /* block new processing (barrier>0 - few callers are possible */
524 atomic_inc(&d->opd_syn_barrier);
526 CDEBUG(D_OTHER, "%s: %u in flight\n", d->opd_obd->obd_name,
527 d->opd_syn_rpc_in_flight);
529 /* wait till all-in-flight are replied, so executed by the target */
530 /* XXX: this is used by LFSCK at the moment, which doesn't require
531 * all the changes to be committed, but in general it'd be
532 * better to wait till commit */
533 while (d->opd_syn_rpc_in_flight > 0) {
535 old = d->opd_syn_rpc_in_flight;
537 expire = cfs_time_shift(obd_timeout);
538 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
539 osp_sync_timeout, d);
540 l_wait_event(d->opd_syn_barrier_waitq,
541 d->opd_syn_rpc_in_flight == 0, &lwi);
543 if (d->opd_syn_rpc_in_flight == 0)
546 if (d->opd_syn_rpc_in_flight != old) {
547 /* some progress have been made */
551 /* no changes and expired, something is wrong */
552 GOTO(out, rc = -ETIMEDOUT);
555 CDEBUG(D_OTHER, "%s: done in %lu\n", d->opd_obd->obd_name,
556 cfs_time_current() - start);
558 /* resume normal processing (barrier=0) */
559 atomic_dec(&d->opd_syn_barrier);
560 __osp_sync_check_for_work(d);
565 const struct dt_device_operations osp_dt_ops = {
566 .dt_statfs = osp_statfs,
568 .dt_trans_create = osp_trans_create,
569 .dt_trans_start = osp_trans_start,
570 .dt_trans_stop = osp_trans_stop,
573 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
576 struct obd_connect_data *data = NULL;
577 struct obd_device *obd;
582 LASSERT(m->opd_storage_exp == NULL);
588 obd = class_name2obd(nextdev);
590 CERROR("%s: can't locate next device: %s\n",
591 m->opd_obd->obd_name, nextdev);
592 GOTO(out, rc = -ENOTCONN);
595 rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
598 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
599 m->opd_obd->obd_name, nextdev, rc);
603 m->opd_dt_dev.dd_lu_dev.ld_site =
604 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
605 LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
606 m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
613 static int osp_init0(const struct lu_env *env, struct osp_device *m,
614 struct lu_device_type *ldt, struct lustre_cfg *cfg)
616 struct obd_device *obd;
617 struct obd_import *imp;
619 char *src, *tgt, *mdt, *osdname = NULL;
625 mutex_init(&m->opd_async_requests_mutex);
627 obd = class_name2obd(lustre_cfg_string(cfg, 0));
629 CERROR("Cannot find obd with name %s\n",
630 lustre_cfg_string(cfg, 0));
635 /* There is no record in the MDT configuration for the local disk
636 * device, so we have to extract this from elsewhere in the profile.
637 * The only information we get at setup is from the OSC records:
638 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
639 * Note that 1.8 generated configs are missing the -MDTxxxx part.
640 * We need to reconstruct the name of the underlying OSD from this:
641 * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd". We
642 * also need to determine the OST index from this - will be used
643 * to calculate the offset in shared lov_objids file later */
645 src = lustre_cfg_string(cfg, 0);
649 tgt = strrchr(src, '-');
651 CERROR("%s: invalid target name %s\n",
652 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
656 if (strncmp(tgt, "-osc", 4) == 0) {
657 /* Old OSC name fsname-OSTXXXX-osc */
658 for (tgt--; tgt > src && *tgt != '-'; tgt--)
661 CERROR("%s: invalid target name %s\n",
662 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
666 if (strncmp(tgt, "-OST", 4) != 0) {
667 CERROR("%s: invalid target name %s\n",
668 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
672 idx = simple_strtol(tgt + 4, &mdt, 16);
673 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
674 CERROR("%s: invalid OST index in '%s'\n",
675 m->opd_obd->obd_name, src);
682 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
683 if (strncmp(tgt, "-MDT", 4) != 0 &&
684 strncmp(tgt, "-OST", 4) != 0) {
685 CERROR("%s: invalid target name %s\n",
686 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
690 idx = simple_strtol(tgt + 4, &mdt, 16);
691 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
692 CERROR("%s: invalid OST index in '%s'\n",
693 m->opd_obd->obd_name, src);
697 /* Get MDT index from the name and set it to opd_group,
698 * which will be used by OSP to connect with OST */
700 if (tgt - src <= 12) {
701 CERROR("%s: invalid mdt index retrieve from %s\n",
702 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
706 if (strncmp(tgt - 12, "-MDT", 4) == 0)
707 m->opd_connect_mdt = 1;
709 idx = simple_strtol(tgt - 8, &mdt, 16);
710 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
711 CERROR("%s: invalid OST index in '%s'\n",
712 m->opd_obd->obd_name, src);
717 idx = tgt - src - 12;
719 /* check the fsname length, and after this everything else will fit */
720 if (idx > MTI_NAME_MAXLEN) {
721 CERROR("%s: fsname too long in '%s'\n",
722 m->opd_obd->obd_name, src);
726 OBD_ALLOC(osdname, MAX_OBD_NAME);
730 memcpy(osdname, src, idx); /* copy just the fsname part */
733 mdt = strstr(mdt, "-MDT");
734 if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
735 strcat(osdname, "-MDT0000");
737 strcat(osdname, mdt);
738 strcat(osdname, "-osd");
739 CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
741 if (m->opd_connect_mdt) {
742 struct client_obd *cli = &m->opd_obd->u.cli;
744 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
745 if (!cli->cl_rpc_lock)
746 GOTO(out_fini, rc = -ENOMEM);
747 osp_init_rpc_lock(cli->cl_rpc_lock);
750 m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
751 m->opd_dt_dev.dd_ops = &osp_dt_ops;
752 obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
754 rc = osp_connect_to_osd(env, m, osdname);
758 rc = ptlrpcd_addref();
760 GOTO(out_disconnect, rc);
762 rc = client_obd_setup(obd, cfg);
764 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
770 rc = obd_fid_init(m->opd_obd, NULL, m->opd_connect_mdt ?
771 LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
773 CERROR("%s: fid init error: rc = %d\n",
774 m->opd_obd->obd_name, rc);
778 if (!m->opd_connect_mdt) {
779 /* Initialize last id from the storage - will be
780 * used in orphan cleanup. */
781 rc = osp_last_used_init(env, m);
786 /* Initialize precreation thread, it handles new
787 * connections as well. */
788 rc = osp_init_precreate(m);
790 GOTO(out_last_used, rc);
794 * Initialize synhronization mechanism taking
795 * care of propogating changes to OST in near
796 * transactional manner.
798 rc = osp_sync_init(env, m);
800 GOTO(out_precreat, rc);
803 * Initiate connect to OST
805 ll_generate_random_uuid(uuid);
806 class_uuid_unparse(uuid, &m->opd_cluuid);
808 imp = obd->u.cli.cl_import;
810 rc = ptlrpc_init_import(imp);
814 OBD_FREE(osdname, MAX_OBD_NAME);
818 /* stop sync thread */
821 /* stop precreate thread */
822 if (!m->opd_connect_mdt)
823 osp_precreate_fini(m);
825 if (!m->opd_connect_mdt)
826 osp_last_used_fini(env, m);
828 ptlrpc_lprocfs_unregister_obd(obd);
829 lprocfs_obd_cleanup(obd);
831 lprocfs_remove(&m->opd_symlink);
832 client_obd_cleanup(obd);
836 if (m->opd_connect_mdt) {
837 struct client_obd *cli = &m->opd_obd->u.cli;
838 if (cli->cl_rpc_lock != NULL) {
839 OBD_FREE_PTR(cli->cl_rpc_lock);
840 cli->cl_rpc_lock = NULL;
843 obd_disconnect(m->opd_storage_exp);
846 OBD_FREE(osdname, MAX_OBD_NAME);
850 static struct lu_device *osp_device_free(const struct lu_env *env,
851 struct lu_device *lu)
853 struct osp_device *m = lu2osp_dev(lu);
857 if (atomic_read(&lu->ld_ref) && lu->ld_site) {
858 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
859 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
861 dt_device_fini(&m->opd_dt_dev);
866 static struct lu_device *osp_device_alloc(const struct lu_env *env,
867 struct lu_device_type *t,
868 struct lustre_cfg *lcfg)
870 struct osp_device *m;
875 l = ERR_PTR(-ENOMEM);
880 dt_device_init(&m->opd_dt_dev, t);
881 rc = osp_init0(env, m, t, lcfg);
883 osp_device_free(env, l);
890 static struct lu_device *osp_device_fini(const struct lu_env *env,
893 struct osp_device *m = lu2osp_dev(d);
894 struct obd_import *imp;
899 if (m->opd_async_requests != NULL) {
900 out_destroy_update_req(m->opd_async_requests);
901 m->opd_async_requests = NULL;
904 if (m->opd_storage_exp)
905 obd_disconnect(m->opd_storage_exp);
907 imp = m->opd_obd->u.cli.cl_import;
909 if (imp->imp_rq_pool) {
910 ptlrpc_free_rq_pool(imp->imp_rq_pool);
911 imp->imp_rq_pool = NULL;
915 lprocfs_remove(&m->opd_symlink);
918 ptlrpc_lprocfs_unregister_obd(m->opd_obd);
919 lprocfs_obd_cleanup(m->opd_obd);
921 if (m->opd_connect_mdt) {
922 struct client_obd *cli = &m->opd_obd->u.cli;
923 if (cli->cl_rpc_lock != NULL) {
924 OBD_FREE_PTR(cli->cl_rpc_lock);
925 cli->cl_rpc_lock = NULL;
929 rc = client_obd_cleanup(m->opd_obd);
930 LASSERTF(rc == 0, "error %d\n", rc);
937 static int osp_reconnect(const struct lu_env *env,
938 struct obd_export *exp, struct obd_device *obd,
939 struct obd_uuid *cluuid,
940 struct obd_connect_data *data,
946 static int osp_prepare_fid_client(struct osp_device *osp)
948 LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
949 if (osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL)
952 LASSERT(osp->opd_exp != NULL);
953 osp->opd_obd->u.cli.cl_seq->lcs_exp =
954 class_export_get(osp->opd_exp);
959 * we use exports to track all LOD users
961 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
962 struct obd_device *obd, struct obd_uuid *cluuid,
963 struct obd_connect_data *data, void *localdata)
965 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
966 struct obd_connect_data *ocd;
967 struct obd_import *imp;
968 struct lustre_handle conn;
973 CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
975 rc = class_connect(&conn, obd, cluuid);
979 *exp = class_conn2export(&conn);
980 /* Why should there ever be more than 1 connect? */
982 LASSERT(osp->opd_connects == 1);
986 imp = osp->opd_obd->u.cli.cl_import;
987 imp->imp_dlm_handle = conn;
989 LASSERT(data != NULL);
990 LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
991 ocd = &imp->imp_connect_data;
994 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
996 ocd->ocd_version = LUSTRE_VERSION_CODE;
997 ocd->ocd_index = data->ocd_index;
998 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
1000 rc = ptlrpc_connect_import(imp);
1002 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
1006 ptlrpc_pinger_add_import(imp);
1012 * once last export (we don't count self-export) disappeared
1013 * osp can be released
1015 static int osp_obd_disconnect(struct obd_export *exp)
1017 struct obd_device *obd = exp->exp_obd;
1018 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1022 /* Only disconnect the underlying layers on the final disconnect. */
1023 LASSERT(osp->opd_connects == 1);
1024 osp->opd_connects--;
1026 rc = class_disconnect(exp);
1028 CERROR("%s: class disconnect error: rc = %d\n",
1033 /* destroy the device */
1034 class_manual_cleanup(obd);
1040 * lprocfs helpers still use OBD API, let's keep obd_statfs() support
1042 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
1043 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1045 struct obd_statfs *msfs;
1046 struct ptlrpc_request *req;
1047 struct obd_import *imp = NULL;
1052 /* Since the request might also come from lprocfs, so we need
1053 * sync this with client_disconnect_export Bug15684 */
1054 down_read(&exp->exp_obd->u.cli.cl_sem);
1055 if (exp->exp_obd->u.cli.cl_import)
1056 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
1057 up_read(&exp->exp_obd->u.cli.cl_sem);
1061 /* We could possibly pass max_age in the request (as an absolute
1062 * timestamp or a "seconds.usec ago") so the target can avoid doing
1063 * extra calls into the filesystem if that isn't necessary (e.g.
1064 * during mount that would help a bit). Having relative timestamps
1065 * is not so great if request processing is slow, while absolute
1066 * timestamps are not ideal because they need time synchronization. */
1067 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1069 class_import_put(imp);
1074 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1076 ptlrpc_request_free(req);
1079 ptlrpc_request_set_replen(req);
1080 req->rq_request_portal = OST_CREATE_PORTAL;
1081 ptlrpc_at_set_req_timeout(req);
1083 if (flags & OBD_STATFS_NODELAY) {
1084 /* procfs requests not want stat in wait for avoid deadlock */
1085 req->rq_no_resend = 1;
1086 req->rq_no_delay = 1;
1089 rc = ptlrpc_queue_wait(req);
1093 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1095 GOTO(out, rc = -EPROTO);
1101 ptlrpc_req_finished(req);
1105 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1106 enum obd_import_event event)
1108 struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1111 case IMP_EVENT_DISCON:
1112 d->opd_got_disconnected = 1;
1113 d->opd_imp_connected = 0;
1114 if (d->opd_connect_mdt)
1117 if (d->opd_pre != NULL) {
1118 osp_pre_update_status(d, -ENODEV);
1119 wake_up(&d->opd_pre_waitq);
1122 CDEBUG(D_HA, "got disconnected\n");
1124 case IMP_EVENT_INACTIVE:
1125 d->opd_imp_active = 0;
1126 if (d->opd_connect_mdt)
1129 if (d->opd_pre != NULL) {
1130 osp_pre_update_status(d, -ENODEV);
1131 wake_up(&d->opd_pre_waitq);
1134 CDEBUG(D_HA, "got inactive\n");
1136 case IMP_EVENT_ACTIVE:
1137 d->opd_imp_active = 1;
1139 if (osp_prepare_fid_client(d) != 0)
1142 if (d->opd_got_disconnected)
1143 d->opd_new_connection = 1;
1144 d->opd_imp_connected = 1;
1145 d->opd_imp_seen_connected = 1;
1146 if (d->opd_connect_mdt)
1149 if (d->opd_pre != NULL)
1150 wake_up(&d->opd_pre_waitq);
1152 __osp_sync_check_for_work(d);
1153 CDEBUG(D_HA, "got connected\n");
1155 case IMP_EVENT_INVALIDATE:
1156 if (obd->obd_namespace == NULL)
1158 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1161 case IMP_EVENT_DEACTIVATE:
1162 case IMP_EVENT_ACTIVATE:
1165 CERROR("%s: unsupported import event: %#x\n",
1166 obd->obd_name, event);
1171 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1172 void *karg, void *uarg)
1174 struct obd_device *obd = exp->exp_obd;
1175 struct osp_device *d;
1176 struct obd_ioctl_data *data = karg;
1181 LASSERT(obd->obd_lu_dev);
1182 d = lu2osp_dev(obd->obd_lu_dev);
1183 LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1185 if (!try_module_get(THIS_MODULE)) {
1186 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1187 module_name(THIS_MODULE));
1192 case OBD_IOC_CLIENT_RECOVER:
1193 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1194 data->ioc_inlbuf1, 0);
1198 case IOC_OSC_SET_ACTIVE:
1199 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1202 case OBD_IOC_PING_TARGET:
1203 rc = ptlrpc_obd_ping(obd);
1206 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1207 cmd, current_comm());
1210 module_put(THIS_MODULE);
1214 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1215 __u32 keylen, void *key, __u32 *vallen, void *val,
1216 struct lov_stripe_md *lsm)
1220 if (KEY_IS(KEY_OSP_CONNECTED)) {
1221 struct obd_device *obd = exp->exp_obd;
1222 struct osp_device *osp;
1224 if (!obd->obd_set_up || obd->obd_stopping)
1227 osp = lu2osp_dev(obd->obd_lu_dev);
1230 * 1.8/2.0 behaviour is that OST being connected once at least
1231 * is considered "healthy". and one "healthy" OST is enough to
1232 * allow lustre clients to connect to MDS
1234 RETURN(!osp->opd_imp_seen_connected);
1240 int osp_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1241 struct lu_fid *fid, struct md_op_data *op_data)
1243 struct client_obd *cli = &exp->exp_obd->u.cli;
1244 struct osp_device *osp = lu2osp_dev(exp->exp_obd->obd_lu_dev);
1245 struct lu_client_seq *seq = cli->cl_seq;
1248 LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
1249 /* Sigh, fid client is not ready yet */
1250 if (osp->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1253 RETURN(seq_client_alloc_fid(env, seq, fid));
1256 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1257 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1258 static void osp_key_exit(const struct lu_context *ctx,
1259 struct lu_context_key *key, void *data)
1261 struct osp_thread_info *info = data;
1263 info->osi_attr.la_valid = 0;
1266 struct lu_context_key osp_thread_key = {
1267 .lct_tags = LCT_MD_THREAD,
1268 .lct_init = osp_key_init,
1269 .lct_fini = osp_key_fini,
1270 .lct_exit = osp_key_exit
1273 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1274 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1276 struct lu_context_key osp_txn_key = {
1277 .lct_tags = LCT_OSP_THREAD | LCT_TX_HANDLE,
1278 .lct_init = osp_txn_key_init,
1279 .lct_fini = osp_txn_key_fini
1281 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1283 static struct lu_device_type_operations osp_device_type_ops = {
1284 .ldto_init = osp_type_init,
1285 .ldto_fini = osp_type_fini,
1287 .ldto_start = osp_type_start,
1288 .ldto_stop = osp_type_stop,
1290 .ldto_device_alloc = osp_device_alloc,
1291 .ldto_device_free = osp_device_free,
1293 .ldto_device_fini = osp_device_fini
1296 static struct lu_device_type osp_device_type = {
1297 .ldt_tags = LU_DEVICE_DT,
1298 .ldt_name = LUSTRE_OSP_NAME,
1299 .ldt_ops = &osp_device_type_ops,
1300 .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1303 static struct obd_ops osp_obd_device_ops = {
1304 .o_owner = THIS_MODULE,
1305 .o_add_conn = client_import_add_conn,
1306 .o_del_conn = client_import_del_conn,
1307 .o_reconnect = osp_reconnect,
1308 .o_connect = osp_obd_connect,
1309 .o_disconnect = osp_obd_disconnect,
1310 .o_get_info = osp_obd_get_info,
1311 .o_import_event = osp_import_event,
1312 .o_iocontrol = osp_iocontrol,
1313 .o_statfs = osp_obd_statfs,
1314 .o_fid_init = client_fid_init,
1315 .o_fid_fini = client_fid_fini,
1316 .o_fid_alloc = osp_fid_alloc,
1319 struct llog_operations osp_mds_ost_orig_logops;
1321 static int __init osp_mod_init(void)
1323 struct obd_type *type;
1326 rc = lu_kmem_init(osp_caches);
1331 rc = class_register_type(&osp_obd_device_ops, NULL, true, NULL,
1332 #ifndef HAVE_ONLY_PROCFS_SEQ
1335 LUSTRE_OSP_NAME, &osp_device_type);
1337 lu_kmem_fini(osp_caches);
1341 rc = class_register_type(&lwp_obd_device_ops, NULL, true, NULL,
1342 #ifndef HAVE_ONLY_PROCFS_SEQ
1345 LUSTRE_LWP_NAME, &lwp_device_type);
1347 class_unregister_type(LUSTRE_OSP_NAME);
1348 lu_kmem_fini(osp_caches);
1352 /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1353 osp_mds_ost_orig_logops = llog_osd_ops;
1354 osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1355 osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1357 /* create "osc" entry in procfs for compatibility purposes */
1358 type = class_search_type(LUSTRE_OSC_NAME);
1359 if (type != NULL && type->typ_procroot != NULL)
1362 type = class_search_type(LUSTRE_OSP_NAME);
1363 type->typ_procsym = lprocfs_seq_register("osc", proc_lustre_root,
1365 if (IS_ERR(type->typ_procsym)) {
1366 CERROR("osp: can't create compat entry \"osc\": %d\n",
1367 (int) PTR_ERR(type->typ_procsym));
1368 type->typ_procsym = NULL;
1373 static void __exit osp_mod_exit(void)
1375 class_unregister_type(LUSTRE_LWP_NAME);
1376 class_unregister_type(LUSTRE_OSP_NAME);
1377 lu_kmem_fini(osp_caches);
1380 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1381 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1382 MODULE_LICENSE("GPL");
1384 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);