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_dev.c
34 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
35 * Author: Mikhail Pershin <mike.pershin@intel.com>
36 * Author: Di Wang <di.wang@intel.com>
39 * The Object Storage Proxy (OSP) module provides an implementation of
40 * the DT API for remote MDTs and OSTs. Every local OSP device (or
41 * object) is a proxy for a remote OSD device (or object). Thus OSP
42 * converts DT operations into RPCs, which are sent to the OUT service
43 * on a remote target, converted back to DT operations, and
44 * executed. Of course there are many ways in which this description
45 * is inaccurate but it's a good enough mental model. OSP is used by
46 * the MDT stack in several ways:
48 * - OSP devices allocate FIDs for the stripe sub-objects of a striped
51 * - OSP objects represent the remote MDT and OST objects that are
52 * the stripes of a striped object.
54 * - OSP devices log, send, and track synchronous operations (setattr
55 * and unlink) to remote targets.
57 * - OSP objects are the bottom slice of the compound LU object
58 * representing a remote MDT object: MDT/MDD/LOD/OSP.
60 * - OSP objects are used by LFSCK to represent remote OST objects
61 * during the verification of MDT-OST consistency.
63 * - OSP devices batch idempotent requests (declare_attr_get() and
64 * declare_xattr_get()) to the remote target and cache their results.
66 * In addition the OSP layer implements a subset of the OBD device API
67 * to support being a client of a remote target, connecting to other
68 * layers, and FID allocation.
71 #define DEBUG_SUBSYSTEM S_MDS
73 #include <linux/kthread.h>
75 #include <uapi/linux/lustre/lustre_ioctl.h>
76 #include <lustre_log.h>
77 #include <lustre_obdo.h>
78 #include <uapi/linux/lustre/lustre_param.h>
79 #include <obd_class.h>
81 #include "osp_internal.h"
83 /* Slab for OSP object allocation */
84 struct kmem_cache *osp_object_kmem;
86 static struct lu_kmem_descr osp_caches[] = {
88 .ckd_cache = &osp_object_kmem,
89 .ckd_name = "osp_obj",
90 .ckd_size = sizeof(struct osp_object)
98 * Implementation of lu_device_operations::ldo_object_alloc
100 * Allocates an OSP object in memory, whose FID is on the remote target.
102 * \param[in] env execution environment
103 * \param[in] hdr The header of the object stack. If it is NULL, it
104 * means the object is not built from top device, i.e.
105 * it is a sub-stripe object of striped directory or
107 * \param[in] d OSP device
109 * \retval object object being created if the creation succeed.
110 * \retval NULL NULL if the creation failed.
112 static struct lu_object *osp_object_alloc(const struct lu_env *env,
113 const struct lu_object_header *hdr,
116 struct lu_object_header *h = NULL;
117 struct osp_object *o;
120 OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, GFP_NOFS);
122 l = &o->opo_obj.do_lu;
124 /* If hdr is NULL, it means the object is not built
125 * from the top dev(MDT/OST), usually it happens when
126 * building striped object, like data object on MDT or
127 * striped object for directory */
130 lu_object_header_init(h);
131 dt_object_init(&o->opo_obj, h, d);
132 lu_object_add_top(h, l);
134 dt_object_init(&o->opo_obj, h, d);
137 l->lo_ops = &osp_lu_obj_ops;
146 * Find or create the local object
148 * Finds or creates the local file referenced by \a reg_id and return the
149 * attributes of the local file.
151 * \param[in] env execution environment
152 * \param[in] osp OSP device
153 * \param[out] attr attributes of the object
154 * \param[in] reg_id the local object ID of the file. It will be used
155 * to compose a local FID{FID_SEQ_LOCAL_FILE, reg_id, 0}
156 * to identify the object.
158 * \retval object object(dt_object) found or created
159 * \retval ERR_PTR(errno) ERR_PTR(errno) if not get the object.
161 static struct dt_object
162 *osp_find_or_create_local_file(const struct lu_env *env, struct osp_device *osp,
163 struct lu_attr *attr, __u32 reg_id)
165 struct osp_thread_info *osi = osp_env_info(env);
166 struct dt_object_format dof = { 0 };
167 struct dt_object *dto;
171 lu_local_obj_fid(&osi->osi_fid, reg_id);
172 attr->la_valid = LA_MODE;
173 attr->la_mode = S_IFREG | 0644;
174 dof.dof_type = DFT_REGULAR;
175 /* Find or create the local object by osi_fid. */
176 dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
181 /* Get attributes of the local object. */
182 rc = dt_attr_get(env, dto, attr);
184 CERROR("%s: can't be initialized: rc = %d\n",
185 osp->opd_obd->obd_name, rc);
186 dt_object_put(env, dto);
193 * Write data buffer to a local file object.
195 * \param[in] env execution environment
196 * \param[in] osp OSP device
197 * \param[in] dt_obj object written to
198 * \param[in] buf buffer containing byte array and length
199 * \param[in] offset write offset in the object in bytes
201 * \retval 0 0 if write succeed
202 * \retval -EFAULT -EFAULT if only part of buffer is written.
203 * \retval negative other negative errno if write failed.
205 static int osp_write_local_file(const struct lu_env *env,
206 struct osp_device *osp,
207 struct dt_object *dt_obj,
214 if (osp->opd_storage->dd_rdonly)
217 th = dt_trans_create(env, osp->opd_storage);
221 rc = dt_declare_record_write(env, dt_obj, buf, offset, th);
224 rc = dt_trans_start_local(env, osp->opd_storage, th);
228 rc = dt_record_write(env, dt_obj, buf, &offset, th);
230 dt_trans_stop(env, osp->opd_storage, th);
235 * Initialize last ID object.
237 * This function initializes the LAST_ID file, which stores the current last
238 * used id of data objects. The MDT will use the last used id and the last_seq
239 * (\see osp_init_last_seq()) to synchronize the precreate object cache with
242 * \param[in] env execution environment
243 * \param[in] osp OSP device
245 * \retval 0 0 if initialization succeed
246 * \retval negative negative errno if initialization failed
248 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
250 struct osp_thread_info *osi = osp_env_info(env);
251 struct lu_fid *fid = &osp->opd_last_used_fid;
252 struct dt_object *dto;
256 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
259 RETURN(PTR_ERR(dto));
261 /* object will be released in device cleanup path */
262 if (osi->osi_attr.la_size >=
263 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
264 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &osi->osi_id,
266 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
267 if (rc != 0 && rc != -EFAULT)
269 /* In case of idif bits 32-48 go to f_seq
270 * (see osp_init_last_seq). So don't care
271 * about u64->u32 convertion. */
272 fid->f_oid = osi->osi_id;
275 if (rc == -EFAULT) { /* fresh LAST_ID */
278 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &osi->osi_id,
280 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
285 osp->opd_last_used_oid_file = dto;
288 /* object will be released in device cleanup path */
289 CERROR("%s: can't initialize lov_objid: rc = %d\n",
290 osp->opd_obd->obd_name, rc);
291 dt_object_put(env, dto);
292 osp->opd_last_used_oid_file = NULL;
297 * Initialize last sequence object.
299 * This function initializes the LAST_SEQ file in the local OSD, which stores
300 * the current last used sequence of data objects. The MDT will use the last
301 * sequence and last id (\see osp_init_last_objid()) to synchronize the
302 * precreate object cache with OSTs.
304 * \param[in] env execution environment
305 * \param[in] osp OSP device
307 * \retval 0 0 if initialization succeed
308 * \retval negative negative errno if initialization failed
310 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
312 struct osp_thread_info *osi = osp_env_info(env);
313 struct lu_fid *fid = &osp->opd_last_used_fid;
314 struct dt_object *dto;
318 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
321 RETURN(PTR_ERR(dto));
323 /* object will be released in device cleanup path */
324 if (osi->osi_attr.la_size >=
325 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
326 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
328 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
329 if (rc != 0 && rc != -EFAULT)
331 if (fid_is_idif(fid))
332 fid->f_seq = fid_idif_seq(osi->osi_id, osp->opd_index);
335 if (rc == -EFAULT) { /* fresh OSP */
337 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
339 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
344 osp->opd_last_used_seq_file = dto;
347 /* object will be released in device cleanup path */
348 CERROR("%s: can't initialize lov_seq: rc = %d\n",
349 osp->opd_obd->obd_name, rc);
350 dt_object_put(env, dto);
351 osp->opd_last_used_seq_file = NULL;
356 * Initialize last OID and sequence object.
358 * If the MDT is just upgraded to 2.4 from the lower version, where the
359 * LAST_SEQ file does not exist, the file will be created and IDIF sequence
360 * will be written into the file.
362 * \param[in] env execution environment
363 * \param[in] osp OSP device
365 * \retval 0 0 if initialization succeed
366 * \retval negative negative error if initialization failed
368 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
370 struct osp_thread_info *osi = osp_env_info(env);
374 fid_zero(&osp->opd_last_used_fid);
375 rc = osp_init_last_objid(env, osp);
377 CERROR("%s: Can not get ids %d from old objid!\n",
378 osp->opd_obd->obd_name, rc);
382 rc = osp_init_last_seq(env, osp);
384 CERROR("%s: Can not get sequence %d from old objseq!\n",
385 osp->opd_obd->obd_name, rc);
389 if (fid_oid(&osp->opd_last_used_fid) != 0 &&
390 fid_seq(&osp->opd_last_used_fid) == 0) {
391 /* Just upgrade from the old version,
392 * set the seq to be IDIF */
393 osp->opd_last_used_fid.f_seq =
394 fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
396 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
397 &osp->opd_last_used_fid.f_seq,
399 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
400 &osi->osi_lb, osi->osi_off);
402 CERROR("%s : Can not write seq file: rc = %d\n",
403 osp->opd_obd->obd_name, rc);
408 if (!fid_is_zero(&osp->opd_last_used_fid) &&
409 !fid_is_sane(&osp->opd_last_used_fid)) {
410 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
411 PFID(&osp->opd_last_used_fid));
412 GOTO(out, rc = -EINVAL);
415 CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
416 osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
419 if (osp->opd_last_used_oid_file != NULL) {
420 dt_object_put(env, osp->opd_last_used_oid_file);
421 osp->opd_last_used_oid_file = NULL;
423 if (osp->opd_last_used_seq_file != NULL) {
424 dt_object_put(env, osp->opd_last_used_seq_file);
425 osp->opd_last_used_seq_file = NULL;
433 * Release the last sequence and OID file objects in OSP device.
435 * \param[in] env execution environment
436 * \param[in] osp OSP device
438 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *osp)
440 /* release last_used file */
441 if (osp->opd_last_used_oid_file != NULL) {
442 dt_object_put(env, osp->opd_last_used_oid_file);
443 osp->opd_last_used_oid_file = NULL;
446 if (osp->opd_last_used_seq_file != NULL) {
447 dt_object_put(env, osp->opd_last_used_seq_file);
448 osp->opd_last_used_seq_file = NULL;
453 * Disconnects the connection between OSP and its correspondent MDT or OST, and
454 * the import will be marked as inactive. It will only be called during OSP
457 * \param[in] d OSP device being disconnected
459 * \retval 0 0 if disconnection succeed
460 * \retval negative negative errno if disconnection failed
462 static int osp_disconnect(struct osp_device *d)
464 struct obd_device *obd = d->opd_obd;
465 struct obd_import *imp;
468 imp = obd->u.cli.cl_import;
470 /* Mark import deactivated now, so we don't try to reconnect if any
471 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
472 * fully deactivate the import, or that would drop all requests. */
473 LASSERT(imp != NULL);
474 spin_lock(&imp->imp_lock);
475 imp->imp_deactive = 1;
476 spin_unlock(&imp->imp_lock);
478 ptlrpc_deactivate_import(imp);
480 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
481 * delete it regardless. (It's safe to delete an import that was
483 (void)ptlrpc_pinger_del_import(imp);
485 rc = ptlrpc_disconnect_import(imp, 0);
487 CERROR("%s: can't disconnect: rc = %d\n", obd->obd_name, rc);
489 ptlrpc_invalidate_import(imp);
495 * Initialize the osp_update structure in OSP device
497 * Allocate osp update structure and start update thread.
499 * \param[in] osp OSP device
501 * \retval 0 if initialization succeeds.
502 * \retval negative errno if initialization fails.
504 static int osp_update_init(struct osp_device *osp)
506 struct l_wait_info lwi = { 0 };
507 struct task_struct *task;
511 LASSERT(osp->opd_connect_mdt);
513 if (osp->opd_storage->dd_rdonly)
516 OBD_ALLOC_PTR(osp->opd_update);
517 if (osp->opd_update == NULL)
520 init_waitqueue_head(&osp->opd_update_thread.t_ctl_waitq);
521 init_waitqueue_head(&osp->opd_update->ou_waitq);
522 spin_lock_init(&osp->opd_update->ou_lock);
523 INIT_LIST_HEAD(&osp->opd_update->ou_list);
524 osp->opd_update->ou_rpc_version = 1;
525 osp->opd_update->ou_version = 1;
526 osp->opd_update->ou_generation = 0;
528 /* start thread handling sending updates to the remote MDT */
529 task = kthread_run(osp_send_update_thread, osp,
530 "osp_up%u-%u", osp->opd_index, osp->opd_group);
532 int rc = PTR_ERR(task);
534 OBD_FREE_PTR(osp->opd_update);
535 osp->opd_update = NULL;
536 CERROR("%s: can't start precreate thread: rc = %d\n",
537 osp->opd_obd->obd_name, rc);
541 l_wait_event(osp->opd_update_thread.t_ctl_waitq,
542 osp_send_update_thread_running(osp) ||
543 osp_send_update_thread_stopped(osp), &lwi);
549 * Finialize osp_update structure in OSP device
551 * Stop the OSP update sending thread, then delete the left
552 * osp thandle in the sending list.
554 * \param [in] osp OSP device.
556 static void osp_update_fini(const struct lu_env *env, struct osp_device *osp)
558 struct osp_update_request *our;
559 struct osp_update_request *tmp;
560 struct osp_updates *ou = osp->opd_update;
565 osp->opd_update_thread.t_flags = SVC_STOPPING;
566 wake_up(&ou->ou_waitq);
568 wait_event(osp->opd_update_thread.t_ctl_waitq,
569 osp->opd_update_thread.t_flags & SVC_STOPPED);
571 /* Remove the left osp thandle from the list */
572 spin_lock(&ou->ou_lock);
573 list_for_each_entry_safe(our, tmp, &ou->ou_list,
575 list_del_init(&our->our_list);
576 LASSERT(our->our_th != NULL);
577 osp_trans_callback(env, our->our_th, -EIO);
578 /* our will be destroyed in osp_thandle_put() */
579 osp_thandle_put(env, our->our_th);
581 spin_unlock(&ou->ou_lock);
584 osp->opd_update = NULL;
588 * Cleanup OSP, which includes disconnect import, cleanup unlink log, stop
589 * precreate threads etc.
591 * \param[in] env execution environment.
592 * \param[in] d OSP device being disconnected.
594 * \retval 0 0 if cleanup succeed
595 * \retval negative negative errno if cleanup failed
597 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
604 rc = osp_disconnect(d);
606 if (!d->opd_connect_mdt) {
607 /* stop sync thread */
610 /* stop precreate thread */
611 osp_precreate_fini(d);
613 /* release last_used file */
614 osp_last_used_fini(env, d);
617 obd_fid_fini(d->opd_obd);
623 * Implementation of osp_lu_ops::ldo_process_config
625 * This function processes config log records in OSP layer. It is usually
626 * called from the top layer of MDT stack, and goes through the stack by calling
627 * ldo_process_config of next layer.
629 * \param[in] env execution environment
630 * \param[in] dev lu_device of OSP
631 * \param[in] lcfg config log
633 * \retval 0 0 if the config log record is executed correctly.
634 * \retval negative negative errno if the record execution fails.
636 static int osp_process_config(const struct lu_env *env,
637 struct lu_device *dev, struct lustre_cfg *lcfg)
639 struct osp_device *d = lu2osp_dev(dev);
640 struct obd_device *obd = d->opd_obd;
645 switch (lcfg->lcfg_command) {
646 case LCFG_PRE_CLEANUP:
647 rc = osp_disconnect(d);
648 osp_update_fini(env, d);
649 if (obd->obd_namespace != NULL)
650 ldlm_namespace_free_prior(obd->obd_namespace, NULL, 1);
653 lu_dev_del_linkage(dev->ld_site, dev);
654 rc = osp_shutdown(env, d);
658 rc = class_process_proc_param(d->opd_connect_mdt ?
659 PARAM_OSP : PARAM_OSC,
660 obd->obd_vars, lcfg, obd);
664 /* class_process_proc_param() haven't found matching
665 * parameter and returned ENOSYS so that layer(s)
666 * below could use that. But OSP is the bottom, so
668 CERROR("%s: unknown param %s\n",
669 (char *)lustre_cfg_string(lcfg, 0),
670 (char *)lustre_cfg_string(lcfg, 1));
675 CERROR("%s: unknown command %u\n",
676 (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
685 * Implementation of osp_lu_ops::ldo_recovery_complete
687 * This function is called after recovery is finished, and OSP layer
688 * will wake up precreate thread here.
690 * \param[in] env execution environment
691 * \param[in] dev lu_device of OSP
693 * \retval 0 0 unconditionally
695 static int osp_recovery_complete(const struct lu_env *env,
696 struct lu_device *dev)
698 struct osp_device *osp = lu2osp_dev(dev);
701 osp->opd_recovery_completed = 1;
703 if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
704 wake_up(&osp->opd_pre_waitq);
709 const struct lu_device_operations osp_lu_ops = {
710 .ldo_object_alloc = osp_object_alloc,
711 .ldo_process_config = osp_process_config,
712 .ldo_recovery_complete = osp_recovery_complete,
716 * Implementation of dt_device_operations::dt_statfs
718 * This function provides statfs status (for precreation) from
719 * corresponding OST. Note: this function only retrieves the status
720 * from the OSP device, and the real statfs RPC happens inside
721 * precreate thread (\see osp_statfs_update). Note: OSP for MDT does
722 * not need to retrieve statfs data for now.
724 * \param[in] env execution environment.
725 * \param[in] dev dt_device of OSP.
726 * \param[out] sfs holds the retrieved statfs data.
728 * \retval 0 0 statfs data was retrieved successfully or
729 * retrieval was not needed
730 * \retval negative negative errno if get statfs failed.
732 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
733 struct obd_statfs *sfs)
735 struct osp_device *d = dt2osp_dev(dev);
736 struct obd_import *imp = d->opd_obd->u.cli.cl_import;
740 if (imp->imp_state == LUSTRE_IMP_CLOSED)
743 if (unlikely(d->opd_imp_active == 0))
746 if (d->opd_pre == NULL)
749 /* return recently updated data */
750 *sfs = d->opd_statfs;
753 * layer above osp (usually lod) can use ffree to estimate
754 * how many objects are available for immediate creation
756 spin_lock(&d->opd_pre_lock);
757 LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
758 fid_seq(&d->opd_pre_used_fid),
759 "last_created "DFID", next_fid "DFID"\n",
760 PFID(&d->opd_pre_last_created_fid),
761 PFID(&d->opd_pre_used_fid));
762 sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
763 fid_oid(&d->opd_pre_used_fid);
764 sfs->os_fprecreated -= d->opd_pre_reserved;
765 LASSERTF(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2,
766 "last_created "DFID", next_fid "DFID", reserved %llu\n",
767 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_pre_used_fid),
768 d->opd_pre_reserved);
769 spin_unlock(&d->opd_pre_lock);
771 CDEBUG(D_OTHER, "%s: %llu blocks, %llu free, %llu avail, "
772 "%llu files, %llu free files\n", d->opd_obd->obd_name,
773 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
774 sfs->os_files, sfs->os_ffree);
778 static int osp_sync_timeout(void *data)
784 * Implementation of dt_device_operations::dt_sync
786 * This function synchronizes the OSP cache to the remote target. It wakes
787 * up unlink log threads and sends out unlink records to the remote OST.
789 * \param[in] env execution environment
790 * \param[in] dev dt_device of OSP
792 * \retval 0 0 if synchronization succeeds
793 * \retval negative negative errno if synchronization fails
795 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
797 struct osp_device *d = dt2osp_dev(dev);
799 struct l_wait_info lwi = { 0 };
800 unsigned long id, old;
802 unsigned long start = cfs_time_current();
805 /* No Sync between MDTs yet. */
806 if (d->opd_connect_mdt)
809 if (unlikely(d->opd_imp_active == 0))
812 id = d->opd_sync_last_used_id;
813 down_write(&d->opd_async_updates_rwsem);
815 CDEBUG(D_OTHER, "%s: async updates %d\n", d->opd_obd->obd_name,
816 atomic_read(&d->opd_async_updates_count));
818 /* make sure the connection is fine */
819 expire = cfs_time_shift(obd_timeout);
820 lwi = LWI_TIMEOUT(expire - cfs_time_current(), osp_sync_timeout, d);
821 rc = l_wait_event(d->opd_sync_barrier_waitq,
822 atomic_read(&d->opd_async_updates_count) == 0,
824 up_write(&d->opd_async_updates_rwsem);
828 CDEBUG(D_CACHE, "%s: id: used %lu, processed %llu\n",
829 d->opd_obd->obd_name, id, d->opd_sync_last_processed_id);
831 /* wait till all-in-line are processed */
832 while (d->opd_sync_last_processed_id < id) {
834 old = d->opd_sync_last_processed_id;
836 /* make sure the connection is fine */
837 expire = cfs_time_shift(obd_timeout);
838 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
839 osp_sync_timeout, d);
840 l_wait_event(d->opd_sync_barrier_waitq,
841 d->opd_sync_last_processed_id >= id,
844 if (d->opd_sync_last_processed_id >= id)
847 if (d->opd_sync_last_processed_id != old) {
848 /* some progress have been made,
853 /* no changes and expired, something is wrong */
854 GOTO(out, rc = -ETIMEDOUT);
857 /* block new processing (barrier>0 - few callers are possible */
858 atomic_inc(&d->opd_sync_barrier);
860 CDEBUG(D_CACHE, "%s: %u in flight\n", d->opd_obd->obd_name,
861 atomic_read(&d->opd_sync_rpcs_in_flight));
863 /* wait till all-in-flight are replied, so executed by the target */
864 /* XXX: this is used by LFSCK at the moment, which doesn't require
865 * all the changes to be committed, but in general it'd be
866 * better to wait till commit */
867 while (atomic_read(&d->opd_sync_rpcs_in_flight) > 0) {
868 old = atomic_read(&d->opd_sync_rpcs_in_flight);
870 expire = cfs_time_shift(obd_timeout);
871 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
872 osp_sync_timeout, d);
873 l_wait_event(d->opd_sync_barrier_waitq,
874 atomic_read(&d->opd_sync_rpcs_in_flight) == 0,
877 if (atomic_read(&d->opd_sync_rpcs_in_flight) == 0)
880 if (atomic_read(&d->opd_sync_rpcs_in_flight) != old) {
881 /* some progress have been made */
885 /* no changes and expired, something is wrong */
886 GOTO(out, rc = -ETIMEDOUT);
890 /* resume normal processing (barrier=0) */
891 atomic_dec(&d->opd_sync_barrier);
892 osp_sync_check_for_work(d);
894 CDEBUG(D_CACHE, "%s: done in %lu: rc = %d\n", d->opd_obd->obd_name,
895 cfs_time_current() - start, rc);
900 const struct dt_device_operations osp_dt_ops = {
901 .dt_statfs = osp_statfs,
903 .dt_trans_create = osp_trans_create,
904 .dt_trans_start = osp_trans_start,
905 .dt_trans_stop = osp_trans_stop,
906 .dt_trans_cb_add = osp_trans_cb_add,
910 * Connect OSP to local OSD.
912 * Locate the local OSD referenced by \a nextdev and connect to it. Sometimes,
913 * OSP needs to access the local OSD to store some information. For example,
914 * during precreate, it needs to update last used OID and sequence file
915 * (LAST_SEQ) in local OSD.
917 * \param[in] env execution environment
918 * \param[in] osp OSP device
919 * \param[in] nextdev the name of local OSD
921 * \retval 0 0 connection succeeded
922 * \retval negative negative errno connection failed
924 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *osp,
927 struct obd_connect_data *data = NULL;
928 struct obd_device *obd;
933 LASSERT(osp->opd_storage_exp == NULL);
939 obd = class_name2obd(nextdev);
941 CERROR("%s: can't locate next device: %s\n",
942 osp->opd_obd->obd_name, nextdev);
943 GOTO(out, rc = -ENOTCONN);
946 rc = obd_connect(env, &osp->opd_storage_exp, obd, &obd->obd_uuid, data,
949 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
950 osp->opd_obd->obd_name, nextdev, rc);
954 osp->opd_dt_dev.dd_lu_dev.ld_site =
955 osp->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
956 LASSERT(osp->opd_dt_dev.dd_lu_dev.ld_site);
957 osp->opd_storage = lu2dt_dev(osp->opd_storage_exp->exp_obd->obd_lu_dev);
965 * Determine if the lock needs to be cancelled
967 * Determine if the unused lock should be cancelled before replay, see
968 * (ldlm_cancel_no_wait_policy()). Currently, only inode bits lock exists
971 * \param[in] lock lock to be checked.
973 * \retval 1 if the lock needs to be cancelled before replay.
974 * \retval 0 if the lock does not need to be cancelled before
977 static int osp_cancel_weight(struct ldlm_lock *lock)
979 if (lock->l_resource->lr_type != LDLM_IBITS)
986 * Initialize OSP device according to the parameters in the configuration
989 * Reconstruct the local device name from the configuration profile, and
990 * initialize necessary threads and structures according to the OSP type
993 * Since there is no record in the MDT configuration for the local disk
994 * device, we have to extract this from elsewhere in the profile.
995 * The only information we get at setup is from the OSC records:
996 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
998 * Note: configs generated by Lustre 1.8 are missing the -MDTxxxx part,
999 * so, we need to reconstruct the name of the underlying OSD from this:
1000 * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".
1002 * \param[in] env execution environment
1003 * \param[in] osp OSP device
1004 * \param[in] ldt lu device type of OSP
1005 * \param[in] cfg configuration log
1007 * \retval 0 0 if OSP initialization succeeded.
1008 * \retval negative negative errno if OSP initialization failed.
1010 static int osp_init0(const struct lu_env *env, struct osp_device *osp,
1011 struct lu_device_type *ldt, struct lustre_cfg *cfg)
1013 struct obd_device *obd;
1014 struct obd_import *imp;
1016 char *src, *tgt, *mdt, *osdname = NULL;
1022 mutex_init(&osp->opd_async_requests_mutex);
1023 INIT_LIST_HEAD(&osp->opd_async_updates);
1024 init_rwsem(&osp->opd_async_updates_rwsem);
1025 atomic_set(&osp->opd_async_updates_count, 0);
1027 obd = class_name2obd(lustre_cfg_string(cfg, 0));
1029 CERROR("Cannot find obd with name %s\n",
1030 lustre_cfg_string(cfg, 0));
1035 src = lustre_cfg_string(cfg, 0);
1039 tgt = strrchr(src, '-');
1041 CERROR("%s: invalid target name %s: rc = %d\n",
1042 osp->opd_obd->obd_name, lustre_cfg_string(cfg, 0),
1047 if (strncmp(tgt, "-osc", 4) == 0) {
1048 /* Old OSC name fsname-OSTXXXX-osc */
1049 for (tgt--; tgt > src && *tgt != '-'; tgt--)
1052 CERROR("%s: invalid target name %s: rc = %d\n",
1053 osp->opd_obd->obd_name,
1054 lustre_cfg_string(cfg, 0), -EINVAL);
1058 if (strncmp(tgt, "-OST", 4) != 0) {
1059 CERROR("%s: invalid target name %s: rc = %d\n",
1060 osp->opd_obd->obd_name,
1061 lustre_cfg_string(cfg, 0), -EINVAL);
1065 idx = simple_strtol(tgt + 4, &mdt, 16);
1066 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1067 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1068 osp->opd_obd->obd_name, src, -EINVAL);
1071 osp->opd_index = idx;
1075 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
1076 if (strncmp(tgt, "-MDT", 4) != 0 &&
1077 strncmp(tgt, "-OST", 4) != 0) {
1078 CERROR("%s: invalid target name %s: rc = %d\n",
1079 osp->opd_obd->obd_name,
1080 lustre_cfg_string(cfg, 0), -EINVAL);
1084 idx = simple_strtol(tgt + 4, &mdt, 16);
1085 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
1086 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1087 osp->opd_obd->obd_name, src, -EINVAL);
1091 /* Get MDT index from the name and set it to opd_group,
1092 * which will be used by OSP to connect with OST */
1093 osp->opd_group = idx;
1094 if (tgt - src <= 12) {
1095 CERROR("%s: invalid mdt index from %s: rc =%d\n",
1096 osp->opd_obd->obd_name,
1097 lustre_cfg_string(cfg, 0), -EINVAL);
1101 if (strncmp(tgt - 12, "-MDT", 4) == 0)
1102 osp->opd_connect_mdt = 1;
1104 idx = simple_strtol(tgt - 8, &mdt, 16);
1105 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1106 CERROR("%s: invalid OST index in '%s': rc =%d\n",
1107 osp->opd_obd->obd_name, src, -EINVAL);
1111 osp->opd_index = idx;
1112 idx = tgt - src - 12;
1114 /* check the fsname length, and after this everything else will fit */
1115 if (idx > MTI_NAME_MAXLEN) {
1116 CERROR("%s: fsname too long in '%s': rc = %d\n",
1117 osp->opd_obd->obd_name, src, -EINVAL);
1121 OBD_ALLOC(osdname, MAX_OBD_NAME);
1122 if (osdname == NULL)
1125 memcpy(osdname, src, idx); /* copy just the fsname part */
1126 osdname[idx] = '\0';
1128 mdt = strstr(mdt, "-MDT");
1129 if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
1130 strcat(osdname, "-MDT0000");
1132 strcat(osdname, mdt);
1133 strcat(osdname, "-osd");
1134 CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
1136 if (osp->opd_connect_mdt) {
1137 struct client_obd *cli = &osp->opd_obd->u.cli;
1139 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
1140 if (!cli->cl_rpc_lock)
1141 GOTO(out_fini, rc = -ENOMEM);
1142 osp_init_rpc_lock(cli->cl_rpc_lock);
1145 osp->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
1146 osp->opd_dt_dev.dd_ops = &osp_dt_ops;
1148 obd->obd_lu_dev = &osp->opd_dt_dev.dd_lu_dev;
1150 rc = osp_connect_to_osd(env, osp, osdname);
1154 rc = ptlrpcd_addref();
1156 GOTO(out_disconnect, rc);
1158 rc = client_obd_setup(obd, cfg);
1160 CERROR("%s: can't setup obd: rc = %d\n", osp->opd_obd->obd_name,
1165 osp_lprocfs_init(osp);
1167 rc = obd_fid_init(osp->opd_obd, NULL, osp->opd_connect_mdt ?
1168 LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
1170 CERROR("%s: fid init error: rc = %d\n",
1171 osp->opd_obd->obd_name, rc);
1175 if (!osp->opd_connect_mdt) {
1176 /* Initialize last id from the storage - will be
1177 * used in orphan cleanup. */
1178 if (!osp->opd_storage->dd_rdonly) {
1179 rc = osp_last_used_init(env, osp);
1184 /* Initialize precreation thread, it handles new
1185 * connections as well. */
1186 rc = osp_init_precreate(osp);
1188 GOTO(out_last_used, rc);
1191 * Initialize synhronization mechanism taking
1192 * care of propogating changes to OST in near
1193 * transactional manner.
1195 rc = osp_sync_init(env, osp);
1197 GOTO(out_precreat, rc);
1199 rc = osp_update_init(osp);
1204 ns_register_cancel(obd->obd_namespace, osp_cancel_weight);
1207 * Initiate connect to OST
1209 ll_generate_random_uuid(uuid);
1210 class_uuid_unparse(uuid, &osp->opd_cluuid);
1212 imp = obd->u.cli.cl_import;
1214 rc = ptlrpc_init_import(imp);
1218 OBD_FREE(osdname, MAX_OBD_NAME);
1222 if (!osp->opd_connect_mdt)
1223 /* stop sync thread */
1226 /* stop precreate thread */
1227 if (!osp->opd_connect_mdt)
1228 osp_precreate_fini(osp);
1230 osp_update_fini(env, osp);
1232 if (!osp->opd_connect_mdt)
1233 osp_last_used_fini(env, osp);
1235 obd_fid_fini(osp->opd_obd);
1237 ptlrpc_lprocfs_unregister_obd(obd);
1238 lprocfs_obd_cleanup(obd);
1239 if (osp->opd_symlink)
1240 lprocfs_remove(&osp->opd_symlink);
1241 client_obd_cleanup(obd);
1245 if (osp->opd_connect_mdt) {
1246 struct client_obd *cli = &osp->opd_obd->u.cli;
1247 if (cli->cl_rpc_lock != NULL) {
1248 OBD_FREE_PTR(cli->cl_rpc_lock);
1249 cli->cl_rpc_lock = NULL;
1252 obd_disconnect(osp->opd_storage_exp);
1255 OBD_FREE(osdname, MAX_OBD_NAME);
1260 * Implementation of lu_device_type_operations::ldto_device_free
1262 * Free the OSP device in memory. No return value is needed for now,
1263 * so always return NULL to comply with the interface.
1265 * \param[in] env execution environment
1266 * \param[in] lu lu_device of OSP
1268 * \retval NULL NULL unconditionally
1270 static struct lu_device *osp_device_free(const struct lu_env *env,
1271 struct lu_device *lu)
1273 struct osp_device *osp = lu2osp_dev(lu);
1275 if (atomic_read(&lu->ld_ref) && lu->ld_site) {
1276 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1277 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
1279 dt_device_fini(&osp->opd_dt_dev);
1286 * Implementation of lu_device_type_operations::ldto_device_alloc
1288 * This function allocates and initializes OSP device in memory according to
1291 * \param[in] env execution environment
1292 * \param[in] type device type of OSP
1293 * \param[in] lcfg config log
1295 * \retval pointer the pointer of allocated OSP if succeed.
1296 * \retval ERR_PTR(errno) ERR_PTR(errno) if failed.
1298 static struct lu_device *osp_device_alloc(const struct lu_env *env,
1299 struct lu_device_type *type,
1300 struct lustre_cfg *lcfg)
1302 struct osp_device *osp;
1303 struct lu_device *ld;
1307 ld = ERR_PTR(-ENOMEM);
1311 ld = osp2lu_dev(osp);
1312 dt_device_init(&osp->opd_dt_dev, type);
1313 rc = osp_init0(env, osp, type, lcfg);
1315 osp_device_free(env, ld);
1323 * Implementation of lu_device_type_operations::ldto_device_fini
1325 * This function cleans up the OSP device, i.e. release and free those
1326 * attached items in osp_device.
1328 * \param[in] env execution environment
1329 * \param[in] ld lu_device of OSP
1331 * \retval NULL NULL if cleanup succeeded.
1332 * \retval ERR_PTR(errno) ERR_PTR(errno) if cleanup failed.
1334 static struct lu_device *osp_device_fini(const struct lu_env *env,
1335 struct lu_device *ld)
1337 struct osp_device *osp = lu2osp_dev(ld);
1342 if (osp->opd_async_requests != NULL) {
1343 osp_update_request_destroy(env, osp->opd_async_requests);
1344 osp->opd_async_requests = NULL;
1347 if (osp->opd_storage_exp)
1348 obd_disconnect(osp->opd_storage_exp);
1350 if (osp->opd_symlink)
1351 lprocfs_remove(&osp->opd_symlink);
1353 LASSERT(osp->opd_obd);
1354 ptlrpc_lprocfs_unregister_obd(osp->opd_obd);
1355 lprocfs_obd_cleanup(osp->opd_obd);
1357 if (osp->opd_connect_mdt) {
1358 struct client_obd *cli = &osp->opd_obd->u.cli;
1359 if (cli->cl_rpc_lock != NULL) {
1360 OBD_FREE_PTR(cli->cl_rpc_lock);
1361 cli->cl_rpc_lock = NULL;
1365 rc = client_obd_cleanup(osp->opd_obd);
1368 RETURN(ERR_PTR(rc));
1377 * Implementation of obd_ops::o_reconnect
1379 * This function is empty and does not need to do anything for now.
1381 static int osp_reconnect(const struct lu_env *env,
1382 struct obd_export *exp, struct obd_device *obd,
1383 struct obd_uuid *cluuid,
1384 struct obd_connect_data *data,
1391 * Implementation of obd_ops::o_connect
1393 * Connect OSP to the remote target (MDT or OST). Allocate the
1394 * export and return it to the LOD, which calls this function
1395 * for each OSP to connect it to the remote target. This function
1396 * is currently only called once per OSP.
1398 * \param[in] env execution environment
1399 * \param[out] exp export connected to OSP
1400 * \param[in] obd OSP device
1401 * \param[in] cluuid OSP device client uuid
1402 * \param[in] data connect_data to be used to connect to the remote
1404 * \param[in] localdata necessary for the API interface, but not used in
1407 * \retval 0 0 if the connection succeeded.
1408 * \retval negative negative errno if the connection failed.
1410 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
1411 struct obd_device *obd, struct obd_uuid *cluuid,
1412 struct obd_connect_data *data, void *localdata)
1414 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1415 struct obd_connect_data *ocd;
1416 struct obd_import *imp;
1417 struct lustre_handle conn;
1422 CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
1424 rc = class_connect(&conn, obd, cluuid);
1428 *exp = class_conn2export(&conn);
1429 /* Why should there ever be more than 1 connect? */
1430 osp->opd_connects++;
1431 LASSERT(osp->opd_connects == 1);
1433 osp->opd_exp = *exp;
1435 imp = osp->opd_obd->u.cli.cl_import;
1436 imp->imp_dlm_handle = conn;
1438 LASSERT(data != NULL);
1439 LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
1440 ocd = &imp->imp_connect_data;
1443 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
1444 imp->imp_connect_flags2_orig = ocd->ocd_connect_flags2;
1446 ocd->ocd_version = LUSTRE_VERSION_CODE;
1447 ocd->ocd_index = data->ocd_index;
1449 rc = ptlrpc_connect_import(imp);
1451 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
1454 osp->opd_obd->u.cli.cl_seq->lcs_exp =
1455 class_export_get(osp->opd_exp);
1458 ptlrpc_pinger_add_import(imp);
1464 * Implementation of obd_ops::o_disconnect
1466 * Disconnect the export for the OSP. This is called by LOD to release the
1467 * OSP during cleanup (\see lod_del_device()). The OSP will be released after
1468 * the export is released.
1470 * \param[in] exp export to be disconnected.
1472 * \retval 0 0 if disconnection succeed
1473 * \retval negative negative errno if disconnection failed
1475 static int osp_obd_disconnect(struct obd_export *exp)
1477 struct obd_device *obd = exp->exp_obd;
1478 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1482 /* Only disconnect the underlying layers on the final disconnect. */
1483 LASSERT(osp->opd_connects == 1);
1484 osp->opd_connects--;
1486 rc = class_disconnect(exp);
1488 CERROR("%s: class disconnect error: rc = %d\n",
1493 /* destroy the device */
1494 class_manual_cleanup(obd);
1500 * Implementation of obd_ops::o_statfs
1502 * Send a RPC to the remote target to get statfs status. This is only used
1503 * in lprocfs helpers by obd_statfs.
1505 * \param[in] env execution environment
1506 * \param[in] exp connection state from this OSP to the parent (LOD)
1508 * \param[out] osfs hold the statfs result
1509 * \param[in] unused Not used in this function for now
1510 * \param[in] flags flags to indicate how OSP will issue the RPC
1512 * \retval 0 0 if statfs succeeded.
1513 * \retval negative negative errno if statfs failed.
1515 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
1516 struct obd_statfs *osfs, __u64 unused, __u32 flags)
1518 struct obd_statfs *msfs;
1519 struct ptlrpc_request *req;
1520 struct obd_import *imp = NULL;
1525 /* Since the request might also come from lprocfs, so we need
1526 * sync this with client_disconnect_export Bug15684 */
1527 down_read(&exp->exp_obd->u.cli.cl_sem);
1528 if (exp->exp_obd->u.cli.cl_import)
1529 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
1530 up_read(&exp->exp_obd->u.cli.cl_sem);
1534 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1536 class_import_put(imp);
1541 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1543 ptlrpc_request_free(req);
1546 ptlrpc_request_set_replen(req);
1547 req->rq_request_portal = OST_CREATE_PORTAL;
1548 ptlrpc_at_set_req_timeout(req);
1550 if (flags & OBD_STATFS_NODELAY) {
1551 /* procfs requests not want stat in wait for avoid deadlock */
1552 req->rq_no_resend = 1;
1553 req->rq_no_delay = 1;
1556 rc = ptlrpc_queue_wait(req);
1560 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1562 GOTO(out, rc = -EPROTO);
1568 ptlrpc_req_finished(req);
1573 * Implementation of obd_ops::o_import_event
1575 * This function is called when some related import event happens. It will
1576 * mark the necessary flags according to the event and notify the necessary
1577 * threads (mainly precreate thread).
1579 * \param[in] obd OSP OBD device
1580 * \param[in] imp import attached from OSP to remote (OST/MDT) service
1581 * \param[in] event event related to remote service (IMP_EVENT_*)
1583 * \retval 0 0 if the event handling succeeded.
1584 * \retval negative negative errno if the event handling failed.
1586 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1587 enum obd_import_event event)
1589 struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1593 case IMP_EVENT_DISCON:
1594 d->opd_got_disconnected = 1;
1595 d->opd_imp_connected = 0;
1596 if (d->opd_connect_mdt)
1599 if (d->opd_pre != NULL) {
1600 osp_pre_update_status(d, -ENODEV);
1601 wake_up(&d->opd_pre_waitq);
1604 CDEBUG(D_HA, "got disconnected\n");
1606 case IMP_EVENT_INACTIVE:
1607 d->opd_imp_active = 0;
1608 d->opd_imp_connected = 0;
1609 d->opd_obd->obd_inactive = 1;
1610 if (d->opd_connect_mdt)
1612 if (d->opd_pre != NULL) {
1613 /* Import is invalid, we can`t get stripes so
1615 rc = imp->imp_deactive ? -ESHUTDOWN : -ENODEV;
1616 osp_pre_update_status(d, rc);
1617 wake_up(&d->opd_pre_waitq);
1620 CDEBUG(D_HA, "got inactive\n");
1622 case IMP_EVENT_ACTIVE:
1623 d->opd_imp_active = 1;
1625 if (d->opd_got_disconnected)
1626 d->opd_new_connection = 1;
1627 d->opd_imp_connected = 1;
1628 d->opd_imp_seen_connected = 1;
1629 d->opd_obd->obd_inactive = 0;
1630 if (d->opd_connect_mdt)
1633 if (d->opd_pre != NULL)
1634 wake_up(&d->opd_pre_waitq);
1636 osp_sync_check_for_work(d);
1637 CDEBUG(D_HA, "got connected\n");
1639 case IMP_EVENT_INVALIDATE:
1640 if (d->opd_connect_mdt)
1641 osp_invalidate_request(d);
1643 if (obd->obd_namespace == NULL)
1645 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1648 case IMP_EVENT_DEACTIVATE:
1649 case IMP_EVENT_ACTIVATE:
1652 CERROR("%s: unsupported import event: %#x\n",
1653 obd->obd_name, event);
1659 * Implementation of obd_ops: o_iocontrol
1661 * This function is the ioctl handler for OSP. Note: lctl will access the OSP
1662 * directly by ioctl, instead of through the MDS stack.
1664 * param[in] cmd ioctl command.
1665 * param[in] exp export of this OSP.
1666 * param[in] len data length of \a karg.
1667 * param[in] karg input argument which is packed as
1669 * param[out] uarg pointer to userspace buffer (must access by
1672 * \retval 0 0 if the ioctl handling succeeded.
1673 * \retval negative negative errno if the ioctl handling failed.
1675 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1676 void *karg, void __user *uarg)
1678 struct obd_device *obd = exp->exp_obd;
1679 struct osp_device *d;
1680 struct obd_ioctl_data *data = karg;
1685 LASSERT(obd->obd_lu_dev);
1686 d = lu2osp_dev(obd->obd_lu_dev);
1687 LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1689 if (!try_module_get(THIS_MODULE)) {
1690 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1691 module_name(THIS_MODULE));
1696 case OBD_IOC_CLIENT_RECOVER:
1697 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1698 data->ioc_inlbuf1, 0);
1702 case IOC_OSC_SET_ACTIVE:
1703 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1706 case OBD_IOC_PING_TARGET:
1707 rc = ptlrpc_obd_ping(obd);
1710 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1711 cmd, current_comm());
1714 module_put(THIS_MODULE);
1720 * Implementation of obd_ops::o_get_info
1722 * Retrieve information by key. Retrieval starts from the top layer
1723 * (MDT) of the MDS stack and traverses the stack by calling the
1724 * obd_get_info() method of the next sub-layer.
1726 * \param[in] env execution environment
1727 * \param[in] exp export of this OSP
1728 * \param[in] keylen length of \a key
1729 * \param[in] key the key
1730 * \param[out] vallen length of \a val
1731 * \param[out] val holds the value returned by the key
1733 * \retval 0 0 if getting information succeeded.
1734 * \retval negative negative errno if getting information failed.
1736 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1737 __u32 keylen, void *key, __u32 *vallen, void *val)
1741 if (KEY_IS(KEY_OSP_CONNECTED)) {
1742 struct obd_device *obd = exp->exp_obd;
1743 struct osp_device *osp;
1745 if (!obd->obd_set_up || obd->obd_stopping)
1748 osp = lu2osp_dev(obd->obd_lu_dev);
1751 * 1.8/2.0 behaviour is that OST being connected once at least
1752 * is considered "healthy". and one "healthy" OST is enough to
1753 * allow lustre clients to connect to MDS
1755 RETURN(!osp->opd_imp_seen_connected);
1761 static int osp_obd_set_info_async(const struct lu_env *env,
1762 struct obd_export *exp,
1763 u32 keylen, void *key,
1764 u32 vallen, void *val,
1765 struct ptlrpc_request_set *set)
1767 struct obd_device *obd = exp->exp_obd;
1768 struct obd_import *imp = obd->u.cli.cl_import;
1769 struct osp_device *osp;
1770 struct ptlrpc_request *req;
1774 if (KEY_IS(KEY_SPTLRPC_CONF)) {
1775 sptlrpc_conf_client_adapt(exp->exp_obd);
1779 LASSERT(set != NULL);
1780 if (!obd->obd_set_up || obd->obd_stopping)
1782 osp = lu2osp_dev(obd->obd_lu_dev);
1784 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
1788 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
1789 RCL_CLIENT, keylen);
1790 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
1791 RCL_CLIENT, vallen);
1792 if (osp->opd_connect_mdt)
1793 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SET_INFO);
1795 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
1797 ptlrpc_request_free(req);
1801 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1802 memcpy(tmp, key, keylen);
1803 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1804 memcpy(tmp, val, vallen);
1806 ptlrpc_request_set_replen(req);
1807 ptlrpc_set_add_req(set, req);
1808 ptlrpc_check_set(NULL, set);
1814 * Implementation of obd_ops: o_fid_alloc
1816 * Allocate a FID. There are two cases in which OSP performs
1819 * 1. FID precreation for data objects, which is done in
1820 * osp_precreate_fids() instead of this function.
1821 * 2. FID allocation for each sub-stripe of a striped directory.
1822 * Similar to other FID clients, OSP requests the sequence
1823 * from its corresponding remote MDT, which in turn requests
1824 * sequences from the sequence controller (MDT0).
1826 * \param[in] env execution environment
1827 * \param[in] exp export of the OSP
1828 * \param[out] fid FID being allocated
1829 * \param[in] unused necessary for the interface but unused.
1831 * \retval 0 0 FID allocated successfully.
1832 * \retval 1 1 FID allocated successfully and new sequence
1833 * requested from seq meta server
1834 * \retval negative negative errno if FID allocation failed.
1836 static int osp_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1837 struct lu_fid *fid, struct md_op_data *unused)
1839 struct client_obd *cli = &exp->exp_obd->u.cli;
1840 struct osp_device *osp = lu2osp_dev(exp->exp_obd->obd_lu_dev);
1841 struct lu_client_seq *seq = cli->cl_seq;
1844 LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
1845 /* Sigh, fid client is not ready yet */
1846 LASSERT(osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1848 RETURN(seq_client_alloc_fid(env, seq, fid));
1851 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1852 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1853 static void osp_key_exit(const struct lu_context *ctx,
1854 struct lu_context_key *key, void *data)
1856 struct osp_thread_info *info = data;
1858 info->osi_attr.la_valid = 0;
1861 struct lu_context_key osp_thread_key = {
1862 .lct_tags = LCT_MD_THREAD,
1863 .lct_init = osp_key_init,
1864 .lct_fini = osp_key_fini,
1865 .lct_exit = osp_key_exit
1868 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1869 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1871 struct lu_context_key osp_txn_key = {
1872 .lct_tags = LCT_OSP_THREAD | LCT_TX_HANDLE,
1873 .lct_init = osp_txn_key_init,
1874 .lct_fini = osp_txn_key_fini
1876 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1878 static struct lu_device_type_operations osp_device_type_ops = {
1879 .ldto_init = osp_type_init,
1880 .ldto_fini = osp_type_fini,
1882 .ldto_start = osp_type_start,
1883 .ldto_stop = osp_type_stop,
1885 .ldto_device_alloc = osp_device_alloc,
1886 .ldto_device_free = osp_device_free,
1888 .ldto_device_fini = osp_device_fini
1891 static struct lu_device_type osp_device_type = {
1892 .ldt_tags = LU_DEVICE_DT,
1893 .ldt_name = LUSTRE_OSP_NAME,
1894 .ldt_ops = &osp_device_type_ops,
1895 .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1898 static struct obd_ops osp_obd_device_ops = {
1899 .o_owner = THIS_MODULE,
1900 .o_add_conn = client_import_add_conn,
1901 .o_del_conn = client_import_del_conn,
1902 .o_reconnect = osp_reconnect,
1903 .o_connect = osp_obd_connect,
1904 .o_disconnect = osp_obd_disconnect,
1905 .o_get_info = osp_obd_get_info,
1906 .o_set_info_async = osp_obd_set_info_async,
1907 .o_import_event = osp_import_event,
1908 .o_iocontrol = osp_iocontrol,
1909 .o_statfs = osp_obd_statfs,
1910 .o_fid_init = client_fid_init,
1911 .o_fid_fini = client_fid_fini,
1912 .o_fid_alloc = osp_fid_alloc,
1915 struct llog_operations osp_mds_ost_orig_logops;
1918 * Initialize OSP module.
1920 * Register device types OSP and Light Weight Proxy (LWP) (\see lwp_dev.c)
1921 * in obd_types (\see class_obd.c). Initialize procfs for the
1922 * the OSP device. Note: OSP was called OSC before Lustre 2.4,
1923 * so for compatibility it still uses the name "osc" in procfs.
1924 * This is called at module load time.
1926 * \retval 0 0 if initialization succeeds.
1927 * \retval negative negative errno if initialization failed.
1929 static int __init osp_init(void)
1931 struct obd_type *type;
1934 rc = lu_kmem_init(osp_caches);
1939 rc = class_register_type(&osp_obd_device_ops, NULL, true, NULL,
1940 LUSTRE_OSP_NAME, &osp_device_type);
1942 lu_kmem_fini(osp_caches);
1946 rc = class_register_type(&lwp_obd_device_ops, NULL, true, NULL,
1947 LUSTRE_LWP_NAME, &lwp_device_type);
1949 class_unregister_type(LUSTRE_OSP_NAME);
1950 lu_kmem_fini(osp_caches);
1954 /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1955 osp_mds_ost_orig_logops = llog_osd_ops;
1956 osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1957 osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1959 /* create "osc" entry in procfs for compatibility purposes */
1960 type = class_search_type(LUSTRE_OSC_NAME);
1961 if (type != NULL && type->typ_procroot != NULL)
1964 type = class_search_type(LUSTRE_OSP_NAME);
1965 type->typ_procsym = lprocfs_register("osc", proc_lustre_root,
1967 if (IS_ERR(type->typ_procsym)) {
1968 CERROR("osp: can't create compat entry \"osc\": %d\n",
1969 (int) PTR_ERR(type->typ_procsym));
1970 type->typ_procsym = NULL;
1976 * Finalize OSP module.
1978 * This callback is called when kernel unloads OSP module from memory, and
1979 * it will deregister OSP and LWP device type from obd_types (\see class_obd.c).
1981 static void __exit osp_exit(void)
1983 class_unregister_type(LUSTRE_LWP_NAME);
1984 class_unregister_type(LUSTRE_OSP_NAME);
1985 lu_kmem_fini(osp_caches);
1988 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1989 MODULE_DESCRIPTION("Lustre OSD Storage Proxy ("LUSTRE_OSP_NAME")");
1990 MODULE_VERSION(LUSTRE_VERSION_STRING);
1991 MODULE_LICENSE("GPL");
1993 module_init(osp_init);
1994 module_exit(osp_exit);