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_param.h>
49 #include <lustre_log.h>
50 #include <lustre_mdc.h>
52 #include "osp_internal.h"
54 /* Slab for OSP object allocation */
55 struct kmem_cache *osp_object_kmem;
57 static struct lu_kmem_descr osp_caches[] = {
59 .ckd_cache = &osp_object_kmem,
60 .ckd_name = "osp_obj",
61 .ckd_size = sizeof(struct osp_object)
68 struct lu_object *osp_object_alloc(const struct lu_env *env,
69 const struct lu_object_header *hdr,
72 struct lu_object_header *h = NULL;
76 OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, __GFP_IO);
78 l = &o->opo_obj.do_lu;
80 /* For data object, OSP obj would always be the top
81 * object, i.e. hdr is always NULL, see lu_object_alloc.
82 * But for metadata object, we always build the object
83 * stack from MDT. i.e. mdt_object will be the top object
88 lu_object_header_init(h);
89 dt_object_init(&o->opo_obj, h, d);
90 lu_object_add_top(h, l);
92 dt_object_init(&o->opo_obj, h, d);
95 l->lo_ops = &osp_lu_obj_ops;
103 static struct dt_object
104 *osp_find_or_create(const struct lu_env *env, struct osp_device *osp,
105 struct lu_attr *attr, __u32 reg_id)
107 struct osp_thread_info *osi = osp_env_info(env);
108 struct dt_object_format dof = { 0 };
109 struct dt_object *dto;
113 lu_local_obj_fid(&osi->osi_fid, reg_id);
114 attr->la_valid = LA_MODE;
115 attr->la_mode = S_IFREG | 0644;
116 dof.dof_type = DFT_REGULAR;
117 dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
122 rc = dt_attr_get(env, dto, attr, NULL);
124 CERROR("%s: can't be initialized: rc = %d\n",
125 osp->opd_obd->obd_name, rc);
126 lu_object_put(env, &dto->do_lu);
132 static int osp_write_local_file(const struct lu_env *env,
133 struct osp_device *osp,
134 struct dt_object *dt_obj,
141 th = dt_trans_create(env, osp->opd_storage);
145 rc = dt_declare_record_write(env, dt_obj, buf->lb_len, offset, th);
148 rc = dt_trans_start_local(env, osp->opd_storage, th);
152 rc = dt_record_write(env, dt_obj, buf, &offset, th);
154 dt_trans_stop(env, osp->opd_storage, th);
158 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
160 struct osp_thread_info *osi = osp_env_info(env);
161 struct lu_fid *fid = &osp->opd_last_used_fid;
162 struct dt_object *dto;
166 dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OID);
168 RETURN(PTR_ERR(dto));
169 /* object will be released in device cleanup path */
170 if (osi->osi_attr.la_size >=
171 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
172 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
174 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
179 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
181 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
184 osp->opd_last_used_oid_file = dto;
187 /* object will be released in device cleanup path */
188 CERROR("%s: can't initialize lov_objid: rc = %d\n",
189 osp->opd_obd->obd_name, rc);
190 lu_object_put(env, &dto->do_lu);
191 osp->opd_last_used_oid_file = NULL;
195 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
197 struct osp_thread_info *osi = osp_env_info(env);
198 struct lu_fid *fid = &osp->opd_last_used_fid;
199 struct dt_object *dto;
203 dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OSEQ);
205 RETURN(PTR_ERR(dto));
207 /* object will be released in device cleanup path */
208 if (osi->osi_attr.la_size >=
209 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
210 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
212 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
217 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
219 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
222 osp->opd_last_used_seq_file = dto;
225 /* object will be released in device cleanup path */
226 CERROR("%s: can't initialize lov_seq: rc = %d\n",
227 osp->opd_obd->obd_name, rc);
228 lu_object_put(env, &dto->do_lu);
229 osp->opd_last_used_seq_file = NULL;
233 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
235 struct osp_thread_info *osi = osp_env_info(env);
239 fid_zero(&osp->opd_last_used_fid);
240 rc = osp_init_last_objid(env, osp);
242 CERROR("%s: Can not get ids %d from old objid!\n",
243 osp->opd_obd->obd_name, rc);
247 rc = osp_init_last_seq(env, osp);
249 CERROR("%s: Can not get ids %d from old objid!\n",
250 osp->opd_obd->obd_name, rc);
254 if (fid_oid(&osp->opd_last_used_fid) != 0 &&
255 fid_seq(&osp->opd_last_used_fid) == 0) {
256 /* Just upgrade from the old version,
257 * set the seq to be IDIF */
258 osp->opd_last_used_fid.f_seq =
259 fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
261 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
262 &osp->opd_last_used_fid.f_seq,
264 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
265 &osi->osi_lb, osi->osi_off);
267 CERROR("%s : Can not write seq file: rc = %d\n",
268 osp->opd_obd->obd_name, rc);
273 if (!fid_is_zero(&osp->opd_last_used_fid) &&
274 !fid_is_sane(&osp->opd_last_used_fid)) {
275 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
276 PFID(&osp->opd_last_used_fid));
277 GOTO(out, rc = -EINVAL);
280 CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
281 osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
284 if (osp->opd_last_used_oid_file != NULL) {
285 lu_object_put(env, &osp->opd_last_used_oid_file->do_lu);
286 osp->opd_last_used_oid_file = NULL;
288 if (osp->opd_last_used_seq_file != NULL) {
289 lu_object_put(env, &osp->opd_last_used_seq_file->do_lu);
290 osp->opd_last_used_seq_file = NULL;
297 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *d)
299 /* release last_used file */
300 if (d->opd_last_used_oid_file != NULL) {
301 lu_object_put(env, &d->opd_last_used_oid_file->do_lu);
302 d->opd_last_used_oid_file = NULL;
305 if (d->opd_last_used_seq_file != NULL) {
306 lu_object_put(env, &d->opd_last_used_seq_file->do_lu);
307 d->opd_last_used_seq_file = NULL;
311 static int osp_disconnect(struct osp_device *d)
313 struct obd_import *imp;
316 imp = d->opd_obd->u.cli.cl_import;
318 /* Mark import deactivated now, so we don't try to reconnect if any
319 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
320 * fully deactivate the import, or that would drop all requests. */
321 LASSERT(imp != NULL);
322 spin_lock(&imp->imp_lock);
323 imp->imp_deactive = 1;
324 spin_unlock(&imp->imp_lock);
326 ptlrpc_deactivate_import(imp);
328 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
329 * delete it regardless. (It's safe to delete an import that was
331 (void)ptlrpc_pinger_del_import(imp);
333 rc = ptlrpc_disconnect_import(imp, 0);
334 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
337 CERROR("%s: can't disconnect: rc = %d\n",
338 d->opd_obd->obd_name, rc);
340 ptlrpc_invalidate_import(imp);
345 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
352 rc = osp_disconnect(d);
354 if (!d->opd_connect_mdt) {
355 /* stop precreate thread */
356 osp_precreate_fini(d);
358 /* stop sync thread */
361 /* release last_used file */
362 osp_last_used_fini(env, d);
365 obd_fid_fini(d->opd_obd);
370 static int osp_process_config(const struct lu_env *env,
371 struct lu_device *dev, struct lustre_cfg *lcfg)
373 struct osp_device *d = lu2osp_dev(dev);
374 struct obd_device *obd = d->opd_obd;
379 switch (lcfg->lcfg_command) {
380 case LCFG_PRE_CLEANUP:
381 rc = osp_disconnect(d);
384 lu_dev_del_linkage(dev->ld_site, dev);
385 rc = osp_shutdown(env, d);
389 rc = class_process_proc_seq_param(PARAM_OSC, obd->obd_vars,
394 /* class_process_proc_param() haven't found matching
395 * parameter and returned ENOSYS so that layer(s)
396 * below could use that. But OSP is the bottom, so
398 CERROR("%s: unknown param %s\n",
399 (char *)lustre_cfg_string(lcfg, 0),
400 (char *)lustre_cfg_string(lcfg, 1));
405 CERROR("%s: unknown command %u\n",
406 (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
414 static int osp_recovery_complete(const struct lu_env *env,
415 struct lu_device *dev)
417 struct osp_device *osp = lu2osp_dev(dev);
421 osp->opd_recovery_completed = 1;
423 if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
424 wake_up(&osp->opd_pre_waitq);
428 const struct lu_device_operations osp_lu_ops = {
429 .ldo_object_alloc = osp_object_alloc,
430 .ldo_process_config = osp_process_config,
431 .ldo_recovery_complete = osp_recovery_complete,
435 * provides with statfs from corresponded OST
438 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
439 struct obd_statfs *sfs)
441 struct osp_device *d = dt2osp_dev(dev);
445 if (unlikely(d->opd_imp_active == 0))
448 if (d->opd_pre == NULL)
451 /* return recently updated data */
452 *sfs = d->opd_statfs;
455 * layer above osp (usually lod) can use ffree to estimate
456 * how many objects are available for immediate creation
458 spin_lock(&d->opd_pre_lock);
459 LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
460 fid_seq(&d->opd_pre_used_fid),
461 "last_created "DFID", next_fid "DFID"\n",
462 PFID(&d->opd_pre_last_created_fid),
463 PFID(&d->opd_pre_used_fid));
464 sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
465 fid_oid(&d->opd_pre_used_fid);
466 sfs->os_fprecreated -= d->opd_pre_reserved;
467 spin_unlock(&d->opd_pre_lock);
469 LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2);
471 CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
472 LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
473 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
474 sfs->os_files, sfs->os_ffree);
478 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
483 * XXX: wake up sync thread, command it to start flushing asap?
489 const struct dt_device_operations osp_dt_ops = {
490 .dt_statfs = osp_statfs,
492 .dt_trans_start = osp_trans_start,
493 .dt_trans_stop = osp_trans_stop,
496 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
499 struct obd_connect_data *data = NULL;
500 struct obd_device *obd;
505 LASSERT(m->opd_storage_exp == NULL);
511 obd = class_name2obd(nextdev);
513 CERROR("%s: can't locate next device: %s\n",
514 m->opd_obd->obd_name, nextdev);
515 GOTO(out, rc = -ENOTCONN);
518 rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
521 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
522 m->opd_obd->obd_name, nextdev, rc);
526 m->opd_dt_dev.dd_lu_dev.ld_site =
527 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
528 LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
529 m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
536 static int osp_init0(const struct lu_env *env, struct osp_device *m,
537 struct lu_device_type *ldt, struct lustre_cfg *cfg)
539 struct obd_device *obd;
540 struct obd_import *imp;
542 char *src, *tgt, *mdt, *osdname = NULL;
548 obd = class_name2obd(lustre_cfg_string(cfg, 0));
550 CERROR("Cannot find obd with name %s\n",
551 lustre_cfg_string(cfg, 0));
556 /* There is no record in the MDT configuration for the local disk
557 * device, so we have to extract this from elsewhere in the profile.
558 * The only information we get at setup is from the OSC records:
559 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
560 * Note that 1.8 generated configs are missing the -MDTxxxx part.
561 * We need to reconstruct the name of the underlying OSD from this:
562 * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd". We
563 * also need to determine the OST index from this - will be used
564 * to calculate the offset in shared lov_objids file later */
566 src = lustre_cfg_string(cfg, 0);
570 tgt = strrchr(src, '-');
572 CERROR("%s: invalid target name %s\n",
573 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
577 if (strncmp(tgt, "-osc", 4) == 0) {
578 /* Old OSC name fsname-OSTXXXX-osc */
579 for (tgt--; tgt > src && *tgt != '-'; tgt--)
582 CERROR("%s: invalid target name %s\n",
583 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
587 if (strncmp(tgt, "-OST", 4) != 0) {
588 CERROR("%s: invalid target name %s\n",
589 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
593 idx = simple_strtol(tgt + 4, &mdt, 16);
594 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
595 CERROR("%s: invalid OST index in '%s'\n",
596 m->opd_obd->obd_name, src);
603 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
604 if (strncmp(tgt, "-MDT", 4) != 0 &&
605 strncmp(tgt, "-OST", 4) != 0) {
606 CERROR("%s: invalid target name %s\n",
607 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
611 idx = simple_strtol(tgt + 4, &mdt, 16);
612 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
613 CERROR("%s: invalid OST index in '%s'\n",
614 m->opd_obd->obd_name, src);
618 /* Get MDT index from the name and set it to opd_group,
619 * which will be used by OSP to connect with OST */
621 if (tgt - src <= 12) {
622 CERROR("%s: invalid mdt index retrieve from %s\n",
623 m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
627 if (strncmp(tgt - 12, "-MDT", 4) == 0)
628 m->opd_connect_mdt = 1;
630 idx = simple_strtol(tgt - 8, &mdt, 16);
631 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
632 CERROR("%s: invalid OST index in '%s'\n",
633 m->opd_obd->obd_name, src);
638 idx = tgt - src - 12;
640 /* check the fsname length, and after this everything else will fit */
641 if (idx > MTI_NAME_MAXLEN) {
642 CERROR("%s: fsname too long in '%s'\n",
643 m->opd_obd->obd_name, src);
647 OBD_ALLOC(osdname, MAX_OBD_NAME);
651 memcpy(osdname, src, idx); /* copy just the fsname part */
654 mdt = strstr(mdt, "-MDT");
655 if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
656 strcat(osdname, "-MDT0000");
658 strcat(osdname, mdt);
659 strcat(osdname, "-osd");
660 CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
662 if (m->opd_connect_mdt) {
663 struct client_obd *cli = &m->opd_obd->u.cli;
665 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
666 if (!cli->cl_rpc_lock)
667 GOTO(out_fini, rc = -ENOMEM);
668 osp_init_rpc_lock(cli->cl_rpc_lock);
671 m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
672 m->opd_dt_dev.dd_ops = &osp_dt_ops;
673 obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
675 rc = osp_connect_to_osd(env, m, osdname);
679 rc = ptlrpcd_addref();
681 GOTO(out_disconnect, rc);
683 rc = client_obd_setup(obd, cfg);
685 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
691 rc = obd_fid_init(m->opd_obd, NULL, m->opd_connect_mdt ?
692 LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
694 CERROR("%s: fid init error: rc = %d\n",
695 m->opd_obd->obd_name, rc);
699 if (!m->opd_connect_mdt) {
700 /* Initialize last id from the storage - will be
701 * used in orphan cleanup. */
702 rc = osp_last_used_init(env, m);
707 /* Initialize precreation thread, it handles new
708 * connections as well. */
709 rc = osp_init_precreate(m);
711 GOTO(out_last_used, rc);
713 * Initialize synhronization mechanism taking
714 * care of propogating changes to OST in near
715 * transactional manner.
717 rc = osp_sync_init(env, m);
719 GOTO(out_precreat, rc);
723 * Initiate connect to OST
725 ll_generate_random_uuid(uuid);
726 class_uuid_unparse(uuid, &m->opd_cluuid);
728 imp = obd->u.cli.cl_import;
730 rc = ptlrpc_init_import(imp);
734 OBD_FREE(osdname, MAX_OBD_NAME);
738 if (!m->opd_connect_mdt)
739 /* stop sync thread */
742 /* stop precreate thread */
743 if (!m->opd_connect_mdt)
744 osp_precreate_fini(m);
746 if (!m->opd_connect_mdt)
747 osp_last_used_fini(env, m);
749 ptlrpc_lprocfs_unregister_obd(obd);
750 lprocfs_obd_cleanup(obd);
752 lprocfs_remove(&m->opd_symlink);
753 client_obd_cleanup(obd);
757 if (m->opd_connect_mdt) {
758 struct client_obd *cli = &m->opd_obd->u.cli;
759 if (cli->cl_rpc_lock != NULL) {
760 OBD_FREE_PTR(cli->cl_rpc_lock);
761 cli->cl_rpc_lock = NULL;
764 obd_disconnect(m->opd_storage_exp);
767 OBD_FREE(osdname, MAX_OBD_NAME);
771 static struct lu_device *osp_device_free(const struct lu_env *env,
772 struct lu_device *lu)
774 struct osp_device *m = lu2osp_dev(lu);
778 if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) {
779 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
780 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
782 dt_device_fini(&m->opd_dt_dev);
787 static struct lu_device *osp_device_alloc(const struct lu_env *env,
788 struct lu_device_type *t,
789 struct lustre_cfg *lcfg)
791 struct osp_device *m;
796 l = ERR_PTR(-ENOMEM);
801 dt_device_init(&m->opd_dt_dev, t);
802 rc = osp_init0(env, m, t, lcfg);
804 osp_device_free(env, l);
811 static struct lu_device *osp_device_fini(const struct lu_env *env,
814 struct osp_device *m = lu2osp_dev(d);
815 struct obd_import *imp;
820 if (m->opd_storage_exp)
821 obd_disconnect(m->opd_storage_exp);
823 imp = m->opd_obd->u.cli.cl_import;
825 if (imp->imp_rq_pool) {
826 ptlrpc_free_rq_pool(imp->imp_rq_pool);
827 imp->imp_rq_pool = NULL;
831 lprocfs_remove(&m->opd_symlink);
834 ptlrpc_lprocfs_unregister_obd(m->opd_obd);
835 lprocfs_obd_cleanup(m->opd_obd);
837 if (m->opd_connect_mdt) {
838 struct client_obd *cli = &m->opd_obd->u.cli;
839 if (cli->cl_rpc_lock != NULL) {
840 OBD_FREE_PTR(cli->cl_rpc_lock);
841 cli->cl_rpc_lock = NULL;
844 if (m->opd_obd->obd_proc_private != NULL)
845 lprocfs_remove((struct proc_dir_entry **)&m->opd_obd->obd_proc_private);
848 rc = client_obd_cleanup(m->opd_obd);
849 LASSERTF(rc == 0, "error %d\n", rc);
856 static int osp_reconnect(const struct lu_env *env,
857 struct obd_export *exp, struct obd_device *obd,
858 struct obd_uuid *cluuid,
859 struct obd_connect_data *data,
865 static int osp_prepare_fid_client(struct osp_device *osp)
869 LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
870 if (osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL)
873 LASSERT(osp->opd_exp != NULL);
874 osp->opd_obd->u.cli.cl_seq->lcs_exp =
875 class_export_get(osp->opd_exp);
876 if (osp->opd_pre == NULL)
879 /* Init fid for osp_precreate if necessary */
880 rc = osp_init_pre_fid(osp);
882 class_export_put(osp->opd_exp);
883 osp->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
884 CERROR("%s: init pre fid error: rc = %d\n",
885 osp->opd_obd->obd_name, rc);
893 * we use exports to track all LOD users
895 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
896 struct obd_device *obd, struct obd_uuid *cluuid,
897 struct obd_connect_data *data, void *localdata)
899 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
900 struct obd_connect_data *ocd;
901 struct obd_import *imp;
902 struct lustre_handle conn;
907 CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
909 rc = class_connect(&conn, obd, cluuid);
913 *exp = class_conn2export(&conn);
914 /* Why should there ever be more than 1 connect? */
916 LASSERT(osp->opd_connects == 1);
920 imp = osp->opd_obd->u.cli.cl_import;
921 imp->imp_dlm_handle = conn;
923 LASSERT(data != NULL);
924 LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
925 ocd = &imp->imp_connect_data;
928 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
930 ocd->ocd_version = LUSTRE_VERSION_CODE;
931 ocd->ocd_index = data->ocd_index;
932 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
934 rc = ptlrpc_connect_import(imp);
936 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
940 ptlrpc_pinger_add_import(imp);
946 * once last export (we don't count self-export) disappeared
947 * osp can be released
949 static int osp_obd_disconnect(struct obd_export *exp)
951 struct obd_device *obd = exp->exp_obd;
952 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
956 /* Only disconnect the underlying layers on the final disconnect. */
957 LASSERT(osp->opd_connects == 1);
960 rc = class_disconnect(exp);
962 CERROR("%s: class disconnect error: rc = %d\n",
967 /* destroy the device */
968 class_manual_cleanup(obd);
974 * lprocfs helpers still use OBD API, let's keep obd_statfs() support
976 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
977 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
979 struct obd_statfs *msfs;
980 struct ptlrpc_request *req;
981 struct obd_import *imp = NULL;
986 /* Since the request might also come from lprocfs, so we need
987 * sync this with client_disconnect_export Bug15684 */
988 down_read(&exp->exp_obd->u.cli.cl_sem);
989 if (exp->exp_obd->u.cli.cl_import)
990 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
991 up_read(&exp->exp_obd->u.cli.cl_sem);
995 /* We could possibly pass max_age in the request (as an absolute
996 * timestamp or a "seconds.usec ago") so the target can avoid doing
997 * extra calls into the filesystem if that isn't necessary (e.g.
998 * during mount that would help a bit). Having relative timestamps
999 * is not so great if request processing is slow, while absolute
1000 * timestamps are not ideal because they need time synchronization. */
1001 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1003 class_import_put(imp);
1008 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1010 ptlrpc_request_free(req);
1013 ptlrpc_request_set_replen(req);
1014 req->rq_request_portal = OST_CREATE_PORTAL;
1015 ptlrpc_at_set_req_timeout(req);
1017 if (flags & OBD_STATFS_NODELAY) {
1018 /* procfs requests not want stat in wait for avoid deadlock */
1019 req->rq_no_resend = 1;
1020 req->rq_no_delay = 1;
1023 rc = ptlrpc_queue_wait(req);
1027 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1029 GOTO(out, rc = -EPROTO);
1035 ptlrpc_req_finished(req);
1039 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1040 enum obd_import_event event)
1042 struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1045 case IMP_EVENT_DISCON:
1046 d->opd_got_disconnected = 1;
1047 d->opd_imp_connected = 0;
1048 if (d->opd_connect_mdt)
1051 if (d->opd_pre != NULL) {
1052 osp_pre_update_status(d, -ENODEV);
1053 wake_up(&d->opd_pre_waitq);
1056 CDEBUG(D_HA, "got disconnected\n");
1058 case IMP_EVENT_INACTIVE:
1059 d->opd_imp_active = 0;
1060 if (d->opd_connect_mdt)
1063 if (d->opd_pre != NULL) {
1064 osp_pre_update_status(d, -ENODEV);
1065 wake_up(&d->opd_pre_waitq);
1068 CDEBUG(D_HA, "got inactive\n");
1070 case IMP_EVENT_ACTIVE:
1071 d->opd_imp_active = 1;
1073 if (osp_prepare_fid_client(d) != 0)
1076 if (d->opd_got_disconnected)
1077 d->opd_new_connection = 1;
1078 d->opd_imp_connected = 1;
1079 d->opd_imp_seen_connected = 1;
1080 if (d->opd_connect_mdt)
1083 if (d->opd_pre != NULL)
1084 wake_up(&d->opd_pre_waitq);
1086 __osp_sync_check_for_work(d);
1087 CDEBUG(D_HA, "got connected\n");
1089 case IMP_EVENT_INVALIDATE:
1090 if (obd->obd_namespace == NULL)
1092 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1095 case IMP_EVENT_DEACTIVATE:
1096 case IMP_EVENT_ACTIVATE:
1099 CERROR("%s: unsupported import event: %#x\n",
1100 obd->obd_name, event);
1105 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1106 void *karg, void *uarg)
1108 struct obd_device *obd = exp->exp_obd;
1109 struct osp_device *d;
1110 struct obd_ioctl_data *data = karg;
1115 LASSERT(obd->obd_lu_dev);
1116 d = lu2osp_dev(obd->obd_lu_dev);
1117 LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1119 if (!try_module_get(THIS_MODULE)) {
1120 CERROR("%s: can't get module. Is it alive?", obd->obd_name);
1125 case OBD_IOC_CLIENT_RECOVER:
1126 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1127 data->ioc_inlbuf1, 0);
1131 case IOC_OSC_SET_ACTIVE:
1132 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1135 case OBD_IOC_PING_TARGET:
1136 rc = ptlrpc_obd_ping(obd);
1139 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1140 cmd, current_comm());
1143 module_put(THIS_MODULE);
1147 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1148 __u32 keylen, void *key, __u32 *vallen, void *val,
1149 struct lov_stripe_md *lsm)
1153 if (KEY_IS(KEY_OSP_CONNECTED)) {
1154 struct obd_device *obd = exp->exp_obd;
1155 struct osp_device *osp;
1157 if (!obd->obd_set_up || obd->obd_stopping)
1160 osp = lu2osp_dev(obd->obd_lu_dev);
1163 * 1.8/2.0 behaviour is that OST being connected once at least
1164 * is considered "healthy". and one "healthy" OST is enough to
1165 * allow lustre clients to connect to MDS
1167 RETURN(!osp->opd_imp_seen_connected);
1173 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1174 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1175 static void osp_key_exit(const struct lu_context *ctx,
1176 struct lu_context_key *key, void *data)
1178 struct osp_thread_info *info = data;
1180 info->osi_attr.la_valid = 0;
1183 struct lu_context_key osp_thread_key = {
1184 .lct_tags = LCT_MD_THREAD,
1185 .lct_init = osp_key_init,
1186 .lct_fini = osp_key_fini,
1187 .lct_exit = osp_key_exit
1190 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1191 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1193 struct lu_context_key osp_txn_key = {
1194 .lct_tags = LCT_OSP_THREAD,
1195 .lct_init = osp_txn_key_init,
1196 .lct_fini = osp_txn_key_fini
1198 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1200 static struct lu_device_type_operations osp_device_type_ops = {
1201 .ldto_init = osp_type_init,
1202 .ldto_fini = osp_type_fini,
1204 .ldto_start = osp_type_start,
1205 .ldto_stop = osp_type_stop,
1207 .ldto_device_alloc = osp_device_alloc,
1208 .ldto_device_free = osp_device_free,
1210 .ldto_device_fini = osp_device_fini
1213 static struct lu_device_type osp_device_type = {
1214 .ldt_tags = LU_DEVICE_DT,
1215 .ldt_name = LUSTRE_OSP_NAME,
1216 .ldt_ops = &osp_device_type_ops,
1217 .ldt_ctx_tags = LCT_MD_THREAD
1220 static struct obd_ops osp_obd_device_ops = {
1221 .o_owner = THIS_MODULE,
1222 .o_add_conn = client_import_add_conn,
1223 .o_del_conn = client_import_del_conn,
1224 .o_reconnect = osp_reconnect,
1225 .o_connect = osp_obd_connect,
1226 .o_disconnect = osp_obd_disconnect,
1227 .o_get_info = osp_obd_get_info,
1228 .o_import_event = osp_import_event,
1229 .o_iocontrol = osp_iocontrol,
1230 .o_statfs = osp_obd_statfs,
1231 .o_fid_init = client_fid_init,
1232 .o_fid_fini = client_fid_fini,
1235 struct llog_operations osp_mds_ost_orig_logops;
1237 static int __init osp_mod_init(void)
1241 rc = lu_kmem_init(osp_caches);
1245 rc = class_register_type(&osp_obd_device_ops, NULL, NULL,
1246 #ifndef HAVE_ONLY_PROCFS_SEQ
1249 LUSTRE_OSP_NAME, &osp_device_type);
1251 lu_kmem_fini(osp_caches);
1255 rc = class_register_type(&lwp_obd_device_ops, NULL, NULL,
1256 #ifndef HAVE_ONLY_PROCFS_SEQ
1259 LUSTRE_LWP_NAME, &lwp_device_type);
1261 class_unregister_type(LUSTRE_OSP_NAME);
1262 lu_kmem_fini(osp_caches);
1266 /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1267 osp_mds_ost_orig_logops = llog_osd_ops;
1268 osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1269 osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1273 static void __exit osp_mod_exit(void)
1275 class_unregister_type(LUSTRE_LWP_NAME);
1276 class_unregister_type(LUSTRE_OSP_NAME);
1277 lu_kmem_fini(osp_caches);
1280 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1281 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1282 MODULE_LICENSE("GPL");
1284 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);