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_ioctl.h>
76 #include <lustre_log.h>
77 #include <lustre_obdo.h>
78 #include <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, &fid->f_oid,
266 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
267 if (rc != 0 && rc != -EFAULT)
271 if (rc == -EFAULT) { /* fresh LAST_ID */
273 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
275 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
280 osp->opd_last_used_oid_file = dto;
283 /* object will be released in device cleanup path */
284 CERROR("%s: can't initialize lov_objid: rc = %d\n",
285 osp->opd_obd->obd_name, rc);
286 dt_object_put(env, dto);
287 osp->opd_last_used_oid_file = NULL;
292 * Initialize last sequence object.
294 * This function initializes the LAST_SEQ file in the local OSD, which stores
295 * the current last used sequence of data objects. The MDT will use the last
296 * sequence and last id (\see osp_init_last_objid()) to synchronize the
297 * precreate object cache with OSTs.
299 * \param[in] env execution environment
300 * \param[in] osp OSP device
302 * \retval 0 0 if initialization succeed
303 * \retval negative negative errno if initialization failed
305 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
307 struct osp_thread_info *osi = osp_env_info(env);
308 struct lu_fid *fid = &osp->opd_last_used_fid;
309 struct dt_object *dto;
313 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
316 RETURN(PTR_ERR(dto));
318 /* object will be released in device cleanup path */
319 if (osi->osi_attr.la_size >=
320 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
321 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
323 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
324 if (rc != 0 && rc != -EFAULT)
328 if (rc == -EFAULT) { /* fresh OSP */
330 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
332 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
337 osp->opd_last_used_seq_file = dto;
340 /* object will be released in device cleanup path */
341 CERROR("%s: can't initialize lov_seq: rc = %d\n",
342 osp->opd_obd->obd_name, rc);
343 dt_object_put(env, dto);
344 osp->opd_last_used_seq_file = NULL;
349 * Initialize last OID and sequence object.
351 * If the MDT is just upgraded to 2.4 from the lower version, where the
352 * LAST_SEQ file does not exist, the file will be created and IDIF sequence
353 * will be written into the file.
355 * \param[in] env execution environment
356 * \param[in] osp OSP device
358 * \retval 0 0 if initialization succeed
359 * \retval negative negative error if initialization failed
361 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
363 struct osp_thread_info *osi = osp_env_info(env);
367 fid_zero(&osp->opd_last_used_fid);
368 rc = osp_init_last_objid(env, osp);
370 CERROR("%s: Can not get ids %d from old objid!\n",
371 osp->opd_obd->obd_name, rc);
375 rc = osp_init_last_seq(env, osp);
377 CERROR("%s: Can not get ids %d from old objid!\n",
378 osp->opd_obd->obd_name, rc);
382 if (fid_oid(&osp->opd_last_used_fid) != 0 &&
383 fid_seq(&osp->opd_last_used_fid) == 0) {
384 /* Just upgrade from the old version,
385 * set the seq to be IDIF */
386 osp->opd_last_used_fid.f_seq =
387 fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
389 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
390 &osp->opd_last_used_fid.f_seq,
392 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
393 &osi->osi_lb, osi->osi_off);
395 CERROR("%s : Can not write seq file: rc = %d\n",
396 osp->opd_obd->obd_name, rc);
401 if (!fid_is_zero(&osp->opd_last_used_fid) &&
402 !fid_is_sane(&osp->opd_last_used_fid)) {
403 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
404 PFID(&osp->opd_last_used_fid));
405 GOTO(out, rc = -EINVAL);
408 CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
409 osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
412 if (osp->opd_last_used_oid_file != NULL) {
413 dt_object_put(env, osp->opd_last_used_oid_file);
414 osp->opd_last_used_oid_file = NULL;
416 if (osp->opd_last_used_seq_file != NULL) {
417 dt_object_put(env, osp->opd_last_used_seq_file);
418 osp->opd_last_used_seq_file = NULL;
426 * Release the last sequence and OID file objects in OSP device.
428 * \param[in] env execution environment
429 * \param[in] osp OSP device
431 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *osp)
433 /* release last_used file */
434 if (osp->opd_last_used_oid_file != NULL) {
435 dt_object_put(env, osp->opd_last_used_oid_file);
436 osp->opd_last_used_oid_file = NULL;
439 if (osp->opd_last_used_seq_file != NULL) {
440 dt_object_put(env, osp->opd_last_used_seq_file);
441 osp->opd_last_used_seq_file = NULL;
446 * Disconnects the connection between OSP and its correspondent MDT or OST, and
447 * the import will be marked as inactive. It will only be called during OSP
450 * \param[in] d OSP device being disconnected
452 * \retval 0 0 if disconnection succeed
453 * \retval negative negative errno if disconnection failed
455 static int osp_disconnect(struct osp_device *d)
457 struct obd_device *obd = d->opd_obd;
458 struct obd_import *imp;
461 imp = obd->u.cli.cl_import;
463 /* Mark import deactivated now, so we don't try to reconnect if any
464 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
465 * fully deactivate the import, or that would drop all requests. */
466 LASSERT(imp != NULL);
467 spin_lock(&imp->imp_lock);
468 imp->imp_deactive = 1;
469 spin_unlock(&imp->imp_lock);
471 ptlrpc_deactivate_import(imp);
473 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
474 * delete it regardless. (It's safe to delete an import that was
476 (void)ptlrpc_pinger_del_import(imp);
478 rc = ptlrpc_disconnect_import(imp, 0);
480 CERROR("%s: can't disconnect: rc = %d\n", obd->obd_name, rc);
482 ptlrpc_invalidate_import(imp);
488 * Initialize the osp_update structure in OSP device
490 * Allocate osp update structure and start update thread.
492 * \param[in] osp OSP device
494 * \retval 0 if initialization succeeds.
495 * \retval negative errno if initialization fails.
497 static int osp_update_init(struct osp_device *osp)
499 struct l_wait_info lwi = { 0 };
500 struct task_struct *task;
504 LASSERT(osp->opd_connect_mdt);
506 if (osp->opd_storage->dd_rdonly)
509 OBD_ALLOC_PTR(osp->opd_update);
510 if (osp->opd_update == NULL)
513 init_waitqueue_head(&osp->opd_update_thread.t_ctl_waitq);
514 init_waitqueue_head(&osp->opd_update->ou_waitq);
515 spin_lock_init(&osp->opd_update->ou_lock);
516 INIT_LIST_HEAD(&osp->opd_update->ou_list);
517 osp->opd_update->ou_rpc_version = 1;
518 osp->opd_update->ou_version = 1;
519 osp->opd_update->ou_generation = 0;
521 /* start thread handling sending updates to the remote MDT */
522 task = kthread_run(osp_send_update_thread, osp,
523 "osp_up%u-%u", osp->opd_index, osp->opd_group);
525 int rc = PTR_ERR(task);
527 OBD_FREE_PTR(osp->opd_update);
528 osp->opd_update = NULL;
529 CERROR("%s: can't start precreate thread: rc = %d\n",
530 osp->opd_obd->obd_name, rc);
534 l_wait_event(osp->opd_update_thread.t_ctl_waitq,
535 osp_send_update_thread_running(osp) ||
536 osp_send_update_thread_stopped(osp), &lwi);
542 * Finialize osp_update structure in OSP device
544 * Stop the OSP update sending thread, then delete the left
545 * osp thandle in the sending list.
547 * \param [in] osp OSP device.
549 static void osp_update_fini(const struct lu_env *env, struct osp_device *osp)
551 struct osp_update_request *our;
552 struct osp_update_request *tmp;
553 struct osp_updates *ou = osp->opd_update;
558 osp->opd_update_thread.t_flags = SVC_STOPPING;
559 wake_up(&ou->ou_waitq);
561 wait_event(osp->opd_update_thread.t_ctl_waitq,
562 osp->opd_update_thread.t_flags & SVC_STOPPED);
564 /* Remove the left osp thandle from the list */
565 spin_lock(&ou->ou_lock);
566 list_for_each_entry_safe(our, tmp, &ou->ou_list,
568 list_del_init(&our->our_list);
569 LASSERT(our->our_th != NULL);
570 osp_trans_callback(env, our->our_th, -EIO);
571 /* our will be destroyed in osp_thandle_put() */
572 osp_thandle_put(env, our->our_th);
574 spin_unlock(&ou->ou_lock);
577 osp->opd_update = NULL;
581 * Cleanup OSP, which includes disconnect import, cleanup unlink log, stop
582 * precreate threads etc.
584 * \param[in] env execution environment.
585 * \param[in] d OSP device being disconnected.
587 * \retval 0 0 if cleanup succeed
588 * \retval negative negative errno if cleanup failed
590 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
597 rc = osp_disconnect(d);
599 if (!d->opd_connect_mdt) {
600 /* stop sync thread */
603 /* stop precreate thread */
604 osp_precreate_fini(d);
606 /* release last_used file */
607 osp_last_used_fini(env, d);
610 obd_fid_fini(d->opd_obd);
616 * Implementation of osp_lu_ops::ldo_process_config
618 * This function processes config log records in OSP layer. It is usually
619 * called from the top layer of MDT stack, and goes through the stack by calling
620 * ldo_process_config of next layer.
622 * \param[in] env execution environment
623 * \param[in] dev lu_device of OSP
624 * \param[in] lcfg config log
626 * \retval 0 0 if the config log record is executed correctly.
627 * \retval negative negative errno if the record execution fails.
629 static int osp_process_config(const struct lu_env *env,
630 struct lu_device *dev, struct lustre_cfg *lcfg)
632 struct osp_device *d = lu2osp_dev(dev);
633 struct obd_device *obd = d->opd_obd;
638 switch (lcfg->lcfg_command) {
639 case LCFG_PRE_CLEANUP:
640 rc = osp_disconnect(d);
641 osp_update_fini(env, d);
642 if (obd->obd_namespace != NULL)
643 ldlm_namespace_free_prior(obd->obd_namespace, NULL, 1);
646 lu_dev_del_linkage(dev->ld_site, dev);
647 rc = osp_shutdown(env, d);
651 rc = class_process_proc_param(d->opd_connect_mdt ?
652 PARAM_OSP : PARAM_OSC,
653 obd->obd_vars, lcfg, obd);
657 /* class_process_proc_param() haven't found matching
658 * parameter and returned ENOSYS so that layer(s)
659 * below could use that. But OSP is the bottom, so
661 CERROR("%s: unknown param %s\n",
662 (char *)lustre_cfg_string(lcfg, 0),
663 (char *)lustre_cfg_string(lcfg, 1));
668 CERROR("%s: unknown command %u\n",
669 (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
678 * Implementation of osp_lu_ops::ldo_recovery_complete
680 * This function is called after recovery is finished, and OSP layer
681 * will wake up precreate thread here.
683 * \param[in] env execution environment
684 * \param[in] dev lu_device of OSP
686 * \retval 0 0 unconditionally
688 static int osp_recovery_complete(const struct lu_env *env,
689 struct lu_device *dev)
691 struct osp_device *osp = lu2osp_dev(dev);
694 osp->opd_recovery_completed = 1;
696 if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
697 wake_up(&osp->opd_pre_waitq);
702 const struct lu_device_operations osp_lu_ops = {
703 .ldo_object_alloc = osp_object_alloc,
704 .ldo_process_config = osp_process_config,
705 .ldo_recovery_complete = osp_recovery_complete,
709 * Implementation of dt_device_operations::dt_statfs
711 * This function provides statfs status (for precreation) from
712 * corresponding OST. Note: this function only retrieves the status
713 * from the OSP device, and the real statfs RPC happens inside
714 * precreate thread (\see osp_statfs_update). Note: OSP for MDT does
715 * not need to retrieve statfs data for now.
717 * \param[in] env execution environment.
718 * \param[in] dev dt_device of OSP.
719 * \param[out] sfs holds the retrieved statfs data.
721 * \retval 0 0 statfs data was retrieved successfully or
722 * retrieval was not needed
723 * \retval negative negative errno if get statfs failed.
725 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
726 struct obd_statfs *sfs)
728 struct osp_device *d = dt2osp_dev(dev);
729 struct obd_import *imp = d->opd_obd->u.cli.cl_import;
733 if (imp->imp_state == LUSTRE_IMP_CLOSED)
736 if (unlikely(d->opd_imp_active == 0))
739 if (d->opd_pre == NULL)
742 /* return recently updated data */
743 *sfs = d->opd_statfs;
746 * layer above osp (usually lod) can use ffree to estimate
747 * how many objects are available for immediate creation
749 spin_lock(&d->opd_pre_lock);
750 LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
751 fid_seq(&d->opd_pre_used_fid),
752 "last_created "DFID", next_fid "DFID"\n",
753 PFID(&d->opd_pre_last_created_fid),
754 PFID(&d->opd_pre_used_fid));
755 sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
756 fid_oid(&d->opd_pre_used_fid);
757 sfs->os_fprecreated -= d->opd_pre_reserved;
758 LASSERTF(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2,
759 "last_created "DFID", next_fid "DFID", reserved %llu\n",
760 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_pre_used_fid),
761 d->opd_pre_reserved);
762 spin_unlock(&d->opd_pre_lock);
764 CDEBUG(D_OTHER, "%s: %llu blocks, %llu free, %llu avail, "
765 "%llu files, %llu free files\n", d->opd_obd->obd_name,
766 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
767 sfs->os_files, sfs->os_ffree);
771 static int osp_sync_timeout(void *data)
777 * Implementation of dt_device_operations::dt_sync
779 * This function synchronizes the OSP cache to the remote target. It wakes
780 * up unlink log threads and sends out unlink records to the remote OST.
782 * \param[in] env execution environment
783 * \param[in] dev dt_device of OSP
785 * \retval 0 0 if synchronization succeeds
786 * \retval negative negative errno if synchronization fails
788 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
790 struct osp_device *d = dt2osp_dev(dev);
792 struct l_wait_info lwi = { 0 };
793 unsigned long id, old;
795 unsigned long start = cfs_time_current();
798 /* No Sync between MDTs yet. */
799 if (d->opd_connect_mdt)
802 if (unlikely(d->opd_imp_active == 0))
805 id = d->opd_syn_last_used_id;
806 down_write(&d->opd_async_updates_rwsem);
808 CDEBUG(D_OTHER, "%s: async updates %d\n", d->opd_obd->obd_name,
809 atomic_read(&d->opd_async_updates_count));
811 /* make sure the connection is fine */
812 expire = cfs_time_shift(obd_timeout);
813 lwi = LWI_TIMEOUT(expire - cfs_time_current(), osp_sync_timeout, d);
814 rc = l_wait_event(d->opd_syn_barrier_waitq,
815 atomic_read(&d->opd_async_updates_count) == 0,
817 up_write(&d->opd_async_updates_rwsem);
821 CDEBUG(D_CACHE, "%s: id: used %lu, processed %llu\n",
822 d->opd_obd->obd_name, id, d->opd_syn_last_processed_id);
824 /* wait till all-in-line are processed */
825 while (d->opd_syn_last_processed_id < id) {
827 old = d->opd_syn_last_processed_id;
829 /* make sure the connection is fine */
830 expire = cfs_time_shift(obd_timeout);
831 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
832 osp_sync_timeout, d);
833 l_wait_event(d->opd_syn_barrier_waitq,
834 d->opd_syn_last_processed_id >= id,
837 if (d->opd_syn_last_processed_id >= id)
840 if (d->opd_syn_last_processed_id != old) {
841 /* some progress have been made,
846 /* no changes and expired, something is wrong */
847 GOTO(out, rc = -ETIMEDOUT);
850 /* block new processing (barrier>0 - few callers are possible */
851 atomic_inc(&d->opd_syn_barrier);
853 CDEBUG(D_CACHE, "%s: %u in flight\n", d->opd_obd->obd_name,
854 atomic_read(&d->opd_syn_rpc_in_flight));
856 /* wait till all-in-flight are replied, so executed by the target */
857 /* XXX: this is used by LFSCK at the moment, which doesn't require
858 * all the changes to be committed, but in general it'd be
859 * better to wait till commit */
860 while (atomic_read(&d->opd_syn_rpc_in_flight) > 0) {
862 old = atomic_read(&d->opd_syn_rpc_in_flight);
864 expire = cfs_time_shift(obd_timeout);
865 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
866 osp_sync_timeout, d);
867 l_wait_event(d->opd_syn_barrier_waitq,
868 atomic_read(&d->opd_syn_rpc_in_flight) == 0,
871 if (atomic_read(&d->opd_syn_rpc_in_flight) == 0)
874 if (atomic_read(&d->opd_syn_rpc_in_flight) != old) {
875 /* some progress have been made */
879 /* no changes and expired, something is wrong */
880 GOTO(out, rc = -ETIMEDOUT);
884 /* resume normal processing (barrier=0) */
885 atomic_dec(&d->opd_syn_barrier);
886 __osp_sync_check_for_work(d);
888 CDEBUG(D_CACHE, "%s: done in %lu: rc = %d\n", d->opd_obd->obd_name,
889 cfs_time_current() - start, rc);
894 const struct dt_device_operations osp_dt_ops = {
895 .dt_statfs = osp_statfs,
897 .dt_trans_create = osp_trans_create,
898 .dt_trans_start = osp_trans_start,
899 .dt_trans_stop = osp_trans_stop,
900 .dt_trans_cb_add = osp_trans_cb_add,
904 * Connect OSP to local OSD.
906 * Locate the local OSD referenced by \a nextdev and connect to it. Sometimes,
907 * OSP needs to access the local OSD to store some information. For example,
908 * during precreate, it needs to update last used OID and sequence file
909 * (LAST_SEQ) in local OSD.
911 * \param[in] env execution environment
912 * \param[in] osp OSP device
913 * \param[in] nextdev the name of local OSD
915 * \retval 0 0 connection succeeded
916 * \retval negative negative errno connection failed
918 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *osp,
921 struct obd_connect_data *data = NULL;
922 struct obd_device *obd;
927 LASSERT(osp->opd_storage_exp == NULL);
933 obd = class_name2obd(nextdev);
935 CERROR("%s: can't locate next device: %s\n",
936 osp->opd_obd->obd_name, nextdev);
937 GOTO(out, rc = -ENOTCONN);
940 rc = obd_connect(env, &osp->opd_storage_exp, obd, &obd->obd_uuid, data,
943 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
944 osp->opd_obd->obd_name, nextdev, rc);
948 osp->opd_dt_dev.dd_lu_dev.ld_site =
949 osp->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
950 LASSERT(osp->opd_dt_dev.dd_lu_dev.ld_site);
951 osp->opd_storage = lu2dt_dev(osp->opd_storage_exp->exp_obd->obd_lu_dev);
959 * Determine if the lock needs to be cancelled
961 * Determine if the unused lock should be cancelled before replay, see
962 * (ldlm_cancel_no_wait_policy()). Currently, only inode bits lock exists
965 * \param[in] lock lock to be checked.
967 * \retval 1 if the lock needs to be cancelled before replay.
968 * \retval 0 if the lock does not need to be cancelled before
971 static int osp_cancel_weight(struct ldlm_lock *lock)
973 if (lock->l_resource->lr_type != LDLM_IBITS)
980 * Initialize OSP device according to the parameters in the configuration
983 * Reconstruct the local device name from the configuration profile, and
984 * initialize necessary threads and structures according to the OSP type
987 * Since there is no record in the MDT configuration for the local disk
988 * device, we have to extract this from elsewhere in the profile.
989 * The only information we get at setup is from the OSC records:
990 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
992 * Note: configs generated by Lustre 1.8 are missing the -MDTxxxx part,
993 * so, we need to reconstruct the name of the underlying OSD from this:
994 * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".
996 * \param[in] env execution environment
997 * \param[in] osp OSP device
998 * \param[in] ldt lu device type of OSP
999 * \param[in] cfg configuration log
1001 * \retval 0 0 if OSP initialization succeeded.
1002 * \retval negative negative errno if OSP initialization failed.
1004 static int osp_init0(const struct lu_env *env, struct osp_device *osp,
1005 struct lu_device_type *ldt, struct lustre_cfg *cfg)
1007 struct obd_device *obd;
1008 struct obd_import *imp;
1010 char *src, *tgt, *mdt, *osdname = NULL;
1016 mutex_init(&osp->opd_async_requests_mutex);
1017 INIT_LIST_HEAD(&osp->opd_async_updates);
1018 init_rwsem(&osp->opd_async_updates_rwsem);
1019 atomic_set(&osp->opd_async_updates_count, 0);
1021 obd = class_name2obd(lustre_cfg_string(cfg, 0));
1023 CERROR("Cannot find obd with name %s\n",
1024 lustre_cfg_string(cfg, 0));
1029 src = lustre_cfg_string(cfg, 0);
1033 tgt = strrchr(src, '-');
1035 CERROR("%s: invalid target name %s: rc = %d\n",
1036 osp->opd_obd->obd_name, lustre_cfg_string(cfg, 0),
1041 if (strncmp(tgt, "-osc", 4) == 0) {
1042 /* Old OSC name fsname-OSTXXXX-osc */
1043 for (tgt--; tgt > src && *tgt != '-'; tgt--)
1046 CERROR("%s: invalid target name %s: rc = %d\n",
1047 osp->opd_obd->obd_name,
1048 lustre_cfg_string(cfg, 0), -EINVAL);
1052 if (strncmp(tgt, "-OST", 4) != 0) {
1053 CERROR("%s: invalid target name %s: rc = %d\n",
1054 osp->opd_obd->obd_name,
1055 lustre_cfg_string(cfg, 0), -EINVAL);
1059 idx = simple_strtol(tgt + 4, &mdt, 16);
1060 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1061 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1062 osp->opd_obd->obd_name, src, -EINVAL);
1065 osp->opd_index = idx;
1069 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
1070 if (strncmp(tgt, "-MDT", 4) != 0 &&
1071 strncmp(tgt, "-OST", 4) != 0) {
1072 CERROR("%s: invalid target name %s: rc = %d\n",
1073 osp->opd_obd->obd_name,
1074 lustre_cfg_string(cfg, 0), -EINVAL);
1078 idx = simple_strtol(tgt + 4, &mdt, 16);
1079 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
1080 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1081 osp->opd_obd->obd_name, src, -EINVAL);
1085 /* Get MDT index from the name and set it to opd_group,
1086 * which will be used by OSP to connect with OST */
1087 osp->opd_group = idx;
1088 if (tgt - src <= 12) {
1089 CERROR("%s: invalid mdt index from %s: rc =%d\n",
1090 osp->opd_obd->obd_name,
1091 lustre_cfg_string(cfg, 0), -EINVAL);
1095 if (strncmp(tgt - 12, "-MDT", 4) == 0)
1096 osp->opd_connect_mdt = 1;
1098 idx = simple_strtol(tgt - 8, &mdt, 16);
1099 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1100 CERROR("%s: invalid OST index in '%s': rc =%d\n",
1101 osp->opd_obd->obd_name, src, -EINVAL);
1105 osp->opd_index = idx;
1106 idx = tgt - src - 12;
1108 /* check the fsname length, and after this everything else will fit */
1109 if (idx > MTI_NAME_MAXLEN) {
1110 CERROR("%s: fsname too long in '%s': rc = %d\n",
1111 osp->opd_obd->obd_name, src, -EINVAL);
1115 OBD_ALLOC(osdname, MAX_OBD_NAME);
1116 if (osdname == NULL)
1119 memcpy(osdname, src, idx); /* copy just the fsname part */
1120 osdname[idx] = '\0';
1122 mdt = strstr(mdt, "-MDT");
1123 if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
1124 strcat(osdname, "-MDT0000");
1126 strcat(osdname, mdt);
1127 strcat(osdname, "-osd");
1128 CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
1130 if (osp->opd_connect_mdt) {
1131 struct client_obd *cli = &osp->opd_obd->u.cli;
1133 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
1134 if (!cli->cl_rpc_lock)
1135 GOTO(out_fini, rc = -ENOMEM);
1136 osp_init_rpc_lock(cli->cl_rpc_lock);
1139 osp->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
1140 osp->opd_dt_dev.dd_ops = &osp_dt_ops;
1142 obd->obd_lu_dev = &osp->opd_dt_dev.dd_lu_dev;
1144 rc = osp_connect_to_osd(env, osp, osdname);
1148 rc = ptlrpcd_addref();
1150 GOTO(out_disconnect, rc);
1152 rc = client_obd_setup(obd, cfg);
1154 CERROR("%s: can't setup obd: rc = %d\n", osp->opd_obd->obd_name,
1159 osp_lprocfs_init(osp);
1161 rc = obd_fid_init(osp->opd_obd, NULL, osp->opd_connect_mdt ?
1162 LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
1164 CERROR("%s: fid init error: rc = %d\n",
1165 osp->opd_obd->obd_name, rc);
1169 if (!osp->opd_connect_mdt) {
1170 /* Initialize last id from the storage - will be
1171 * used in orphan cleanup. */
1172 if (!osp->opd_storage->dd_rdonly) {
1173 rc = osp_last_used_init(env, osp);
1178 /* Initialize precreation thread, it handles new
1179 * connections as well. */
1180 rc = osp_init_precreate(osp);
1182 GOTO(out_last_used, rc);
1185 * Initialize synhronization mechanism taking
1186 * care of propogating changes to OST in near
1187 * transactional manner.
1189 rc = osp_sync_init(env, osp);
1191 GOTO(out_precreat, rc);
1193 rc = osp_update_init(osp);
1198 ns_register_cancel(obd->obd_namespace, osp_cancel_weight);
1201 * Initiate connect to OST
1203 ll_generate_random_uuid(uuid);
1204 class_uuid_unparse(uuid, &osp->opd_cluuid);
1206 imp = obd->u.cli.cl_import;
1208 rc = ptlrpc_init_import(imp);
1212 OBD_FREE(osdname, MAX_OBD_NAME);
1216 if (!osp->opd_connect_mdt)
1217 /* stop sync thread */
1220 /* stop precreate thread */
1221 if (!osp->opd_connect_mdt)
1222 osp_precreate_fini(osp);
1224 osp_update_fini(env, osp);
1226 if (!osp->opd_connect_mdt)
1227 osp_last_used_fini(env, osp);
1229 obd_fid_fini(osp->opd_obd);
1231 ptlrpc_lprocfs_unregister_obd(obd);
1232 lprocfs_obd_cleanup(obd);
1233 if (osp->opd_symlink)
1234 lprocfs_remove(&osp->opd_symlink);
1235 client_obd_cleanup(obd);
1239 if (osp->opd_connect_mdt) {
1240 struct client_obd *cli = &osp->opd_obd->u.cli;
1241 if (cli->cl_rpc_lock != NULL) {
1242 OBD_FREE_PTR(cli->cl_rpc_lock);
1243 cli->cl_rpc_lock = NULL;
1246 obd_disconnect(osp->opd_storage_exp);
1249 OBD_FREE(osdname, MAX_OBD_NAME);
1254 * Implementation of lu_device_type_operations::ldto_device_free
1256 * Free the OSP device in memory. No return value is needed for now,
1257 * so always return NULL to comply with the interface.
1259 * \param[in] env execution environment
1260 * \param[in] lu lu_device of OSP
1262 * \retval NULL NULL unconditionally
1264 static struct lu_device *osp_device_free(const struct lu_env *env,
1265 struct lu_device *lu)
1267 struct osp_device *osp = lu2osp_dev(lu);
1269 if (atomic_read(&lu->ld_ref) && lu->ld_site) {
1270 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1271 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
1273 dt_device_fini(&osp->opd_dt_dev);
1280 * Implementation of lu_device_type_operations::ldto_device_alloc
1282 * This function allocates and initializes OSP device in memory according to
1285 * \param[in] env execution environment
1286 * \param[in] type device type of OSP
1287 * \param[in] lcfg config log
1289 * \retval pointer the pointer of allocated OSP if succeed.
1290 * \retval ERR_PTR(errno) ERR_PTR(errno) if failed.
1292 static struct lu_device *osp_device_alloc(const struct lu_env *env,
1293 struct lu_device_type *type,
1294 struct lustre_cfg *lcfg)
1296 struct osp_device *osp;
1297 struct lu_device *ld;
1301 ld = ERR_PTR(-ENOMEM);
1305 ld = osp2lu_dev(osp);
1306 dt_device_init(&osp->opd_dt_dev, type);
1307 rc = osp_init0(env, osp, type, lcfg);
1309 osp_device_free(env, ld);
1317 * Implementation of lu_device_type_operations::ldto_device_fini
1319 * This function cleans up the OSP device, i.e. release and free those
1320 * attached items in osp_device.
1322 * \param[in] env execution environment
1323 * \param[in] ld lu_device of OSP
1325 * \retval NULL NULL if cleanup succeeded.
1326 * \retval ERR_PTR(errno) ERR_PTR(errno) if cleanup failed.
1328 static struct lu_device *osp_device_fini(const struct lu_env *env,
1329 struct lu_device *ld)
1331 struct osp_device *osp = lu2osp_dev(ld);
1336 if (osp->opd_async_requests != NULL) {
1337 osp_update_request_destroy(env, osp->opd_async_requests);
1338 osp->opd_async_requests = NULL;
1341 if (osp->opd_storage_exp)
1342 obd_disconnect(osp->opd_storage_exp);
1344 if (osp->opd_symlink)
1345 lprocfs_remove(&osp->opd_symlink);
1347 LASSERT(osp->opd_obd);
1348 ptlrpc_lprocfs_unregister_obd(osp->opd_obd);
1349 lprocfs_obd_cleanup(osp->opd_obd);
1351 if (osp->opd_connect_mdt) {
1352 struct client_obd *cli = &osp->opd_obd->u.cli;
1353 if (cli->cl_rpc_lock != NULL) {
1354 OBD_FREE_PTR(cli->cl_rpc_lock);
1355 cli->cl_rpc_lock = NULL;
1359 rc = client_obd_cleanup(osp->opd_obd);
1362 RETURN(ERR_PTR(rc));
1371 * Implementation of obd_ops::o_reconnect
1373 * This function is empty and does not need to do anything for now.
1375 static int osp_reconnect(const struct lu_env *env,
1376 struct obd_export *exp, struct obd_device *obd,
1377 struct obd_uuid *cluuid,
1378 struct obd_connect_data *data,
1385 * Implementation of obd_ops::o_connect
1387 * Connect OSP to the remote target (MDT or OST). Allocate the
1388 * export and return it to the LOD, which calls this function
1389 * for each OSP to connect it to the remote target. This function
1390 * is currently only called once per OSP.
1392 * \param[in] env execution environment
1393 * \param[out] exp export connected to OSP
1394 * \param[in] obd OSP device
1395 * \param[in] cluuid OSP device client uuid
1396 * \param[in] data connect_data to be used to connect to the remote
1398 * \param[in] localdata necessary for the API interface, but not used in
1401 * \retval 0 0 if the connection succeeded.
1402 * \retval negative negative errno if the connection failed.
1404 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
1405 struct obd_device *obd, struct obd_uuid *cluuid,
1406 struct obd_connect_data *data, void *localdata)
1408 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1409 struct obd_connect_data *ocd;
1410 struct obd_import *imp;
1411 struct lustre_handle conn;
1416 CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
1418 rc = class_connect(&conn, obd, cluuid);
1422 *exp = class_conn2export(&conn);
1423 /* Why should there ever be more than 1 connect? */
1424 osp->opd_connects++;
1425 LASSERT(osp->opd_connects == 1);
1427 osp->opd_exp = *exp;
1429 imp = osp->opd_obd->u.cli.cl_import;
1430 imp->imp_dlm_handle = conn;
1432 LASSERT(data != NULL);
1433 LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
1434 ocd = &imp->imp_connect_data;
1437 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
1438 imp->imp_connect_flags2_orig = ocd->ocd_connect_flags2;
1440 ocd->ocd_version = LUSTRE_VERSION_CODE;
1441 ocd->ocd_index = data->ocd_index;
1443 rc = ptlrpc_connect_import(imp);
1445 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
1448 osp->opd_obd->u.cli.cl_seq->lcs_exp =
1449 class_export_get(osp->opd_exp);
1452 ptlrpc_pinger_add_import(imp);
1458 * Implementation of obd_ops::o_disconnect
1460 * Disconnect the export for the OSP. This is called by LOD to release the
1461 * OSP during cleanup (\see lod_del_device()). The OSP will be released after
1462 * the export is released.
1464 * \param[in] exp export to be disconnected.
1466 * \retval 0 0 if disconnection succeed
1467 * \retval negative negative errno if disconnection failed
1469 static int osp_obd_disconnect(struct obd_export *exp)
1471 struct obd_device *obd = exp->exp_obd;
1472 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1476 /* Only disconnect the underlying layers on the final disconnect. */
1477 LASSERT(osp->opd_connects == 1);
1478 osp->opd_connects--;
1480 rc = class_disconnect(exp);
1482 CERROR("%s: class disconnect error: rc = %d\n",
1487 /* destroy the device */
1488 class_manual_cleanup(obd);
1494 * Implementation of obd_ops::o_statfs
1496 * Send a RPC to the remote target to get statfs status. This is only used
1497 * in lprocfs helpers by obd_statfs.
1499 * \param[in] env execution environment
1500 * \param[in] exp connection state from this OSP to the parent (LOD)
1502 * \param[out] osfs hold the statfs result
1503 * \param[in] unused Not used in this function for now
1504 * \param[in] flags flags to indicate how OSP will issue the RPC
1506 * \retval 0 0 if statfs succeeded.
1507 * \retval negative negative errno if statfs failed.
1509 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
1510 struct obd_statfs *osfs, __u64 unused, __u32 flags)
1512 struct obd_statfs *msfs;
1513 struct ptlrpc_request *req;
1514 struct obd_import *imp = NULL;
1519 /* Since the request might also come from lprocfs, so we need
1520 * sync this with client_disconnect_export Bug15684 */
1521 down_read(&exp->exp_obd->u.cli.cl_sem);
1522 if (exp->exp_obd->u.cli.cl_import)
1523 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
1524 up_read(&exp->exp_obd->u.cli.cl_sem);
1528 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1530 class_import_put(imp);
1535 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1537 ptlrpc_request_free(req);
1540 ptlrpc_request_set_replen(req);
1541 req->rq_request_portal = OST_CREATE_PORTAL;
1542 ptlrpc_at_set_req_timeout(req);
1544 if (flags & OBD_STATFS_NODELAY) {
1545 /* procfs requests not want stat in wait for avoid deadlock */
1546 req->rq_no_resend = 1;
1547 req->rq_no_delay = 1;
1550 rc = ptlrpc_queue_wait(req);
1554 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1556 GOTO(out, rc = -EPROTO);
1562 ptlrpc_req_finished(req);
1567 * Implementation of obd_ops::o_import_event
1569 * This function is called when some related import event happens. It will
1570 * mark the necessary flags according to the event and notify the necessary
1571 * threads (mainly precreate thread).
1573 * \param[in] obd OSP OBD device
1574 * \param[in] imp import attached from OSP to remote (OST/MDT) service
1575 * \param[in] event event related to remote service (IMP_EVENT_*)
1577 * \retval 0 0 if the event handling succeeded.
1578 * \retval negative negative errno if the event handling failed.
1580 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1581 enum obd_import_event event)
1583 struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1587 case IMP_EVENT_DISCON:
1588 d->opd_got_disconnected = 1;
1589 d->opd_imp_connected = 0;
1590 if (d->opd_connect_mdt)
1593 if (d->opd_pre != NULL) {
1594 osp_pre_update_status(d, -ENODEV);
1595 wake_up(&d->opd_pre_waitq);
1598 CDEBUG(D_HA, "got disconnected\n");
1600 case IMP_EVENT_INACTIVE:
1601 d->opd_imp_active = 0;
1602 d->opd_imp_connected = 0;
1603 d->opd_obd->obd_inactive = 1;
1604 if (d->opd_connect_mdt)
1606 if (d->opd_pre != NULL) {
1607 /* Import is invalid, we can`t get stripes so
1609 rc = imp->imp_deactive ? -ESHUTDOWN : -ENODEV;
1610 osp_pre_update_status(d, rc);
1611 wake_up(&d->opd_pre_waitq);
1614 CDEBUG(D_HA, "got inactive\n");
1616 case IMP_EVENT_ACTIVE:
1617 d->opd_imp_active = 1;
1619 if (d->opd_got_disconnected)
1620 d->opd_new_connection = 1;
1621 d->opd_imp_connected = 1;
1622 d->opd_imp_seen_connected = 1;
1623 d->opd_obd->obd_inactive = 0;
1624 if (d->opd_connect_mdt)
1627 if (d->opd_pre != NULL)
1628 wake_up(&d->opd_pre_waitq);
1630 __osp_sync_check_for_work(d);
1631 CDEBUG(D_HA, "got connected\n");
1633 case IMP_EVENT_INVALIDATE:
1634 if (d->opd_connect_mdt)
1635 osp_invalidate_request(d);
1637 if (obd->obd_namespace == NULL)
1639 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1642 case IMP_EVENT_DEACTIVATE:
1643 case IMP_EVENT_ACTIVATE:
1646 CERROR("%s: unsupported import event: %#x\n",
1647 obd->obd_name, event);
1653 * Implementation of obd_ops: o_iocontrol
1655 * This function is the ioctl handler for OSP. Note: lctl will access the OSP
1656 * directly by ioctl, instead of through the MDS stack.
1658 * param[in] cmd ioctl command.
1659 * param[in] exp export of this OSP.
1660 * param[in] len data length of \a karg.
1661 * param[in] karg input argument which is packed as
1663 * param[out] uarg pointer to userspace buffer (must access by
1666 * \retval 0 0 if the ioctl handling succeeded.
1667 * \retval negative negative errno if the ioctl handling failed.
1669 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1670 void *karg, void __user *uarg)
1672 struct obd_device *obd = exp->exp_obd;
1673 struct osp_device *d;
1674 struct obd_ioctl_data *data = karg;
1679 LASSERT(obd->obd_lu_dev);
1680 d = lu2osp_dev(obd->obd_lu_dev);
1681 LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1683 if (!try_module_get(THIS_MODULE)) {
1684 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1685 module_name(THIS_MODULE));
1690 case OBD_IOC_CLIENT_RECOVER:
1691 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1692 data->ioc_inlbuf1, 0);
1696 case IOC_OSC_SET_ACTIVE:
1697 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1700 case OBD_IOC_PING_TARGET:
1701 rc = ptlrpc_obd_ping(obd);
1704 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1705 cmd, current_comm());
1708 module_put(THIS_MODULE);
1714 * Implementation of obd_ops::o_get_info
1716 * Retrieve information by key. Retrieval starts from the top layer
1717 * (MDT) of the MDS stack and traverses the stack by calling the
1718 * obd_get_info() method of the next sub-layer.
1720 * \param[in] env execution environment
1721 * \param[in] exp export of this OSP
1722 * \param[in] keylen length of \a key
1723 * \param[in] key the key
1724 * \param[out] vallen length of \a val
1725 * \param[out] val holds the value returned by the key
1727 * \retval 0 0 if getting information succeeded.
1728 * \retval negative negative errno if getting information failed.
1730 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1731 __u32 keylen, void *key, __u32 *vallen, void *val)
1735 if (KEY_IS(KEY_OSP_CONNECTED)) {
1736 struct obd_device *obd = exp->exp_obd;
1737 struct osp_device *osp;
1739 if (!obd->obd_set_up || obd->obd_stopping)
1742 osp = lu2osp_dev(obd->obd_lu_dev);
1745 * 1.8/2.0 behaviour is that OST being connected once at least
1746 * is considered "healthy". and one "healthy" OST is enough to
1747 * allow lustre clients to connect to MDS
1749 RETURN(!osp->opd_imp_seen_connected);
1755 static int osp_obd_set_info_async(const struct lu_env *env,
1756 struct obd_export *exp,
1757 u32 keylen, void *key,
1758 u32 vallen, void *val,
1759 struct ptlrpc_request_set *set)
1761 struct obd_device *obd = exp->exp_obd;
1762 struct obd_import *imp = obd->u.cli.cl_import;
1763 struct osp_device *osp;
1764 struct ptlrpc_request *req;
1768 if (KEY_IS(KEY_SPTLRPC_CONF)) {
1769 sptlrpc_conf_client_adapt(exp->exp_obd);
1773 LASSERT(set != NULL);
1774 if (!obd->obd_set_up || obd->obd_stopping)
1776 osp = lu2osp_dev(obd->obd_lu_dev);
1778 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
1782 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
1783 RCL_CLIENT, keylen);
1784 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
1785 RCL_CLIENT, vallen);
1786 if (osp->opd_connect_mdt)
1787 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SET_INFO);
1789 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
1791 ptlrpc_request_free(req);
1795 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1796 memcpy(tmp, key, keylen);
1797 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1798 memcpy(tmp, val, vallen);
1800 ptlrpc_request_set_replen(req);
1801 ptlrpc_set_add_req(set, req);
1802 ptlrpc_check_set(NULL, set);
1808 * Implementation of obd_ops: o_fid_alloc
1810 * Allocate a FID. There are two cases in which OSP performs
1813 * 1. FID precreation for data objects, which is done in
1814 * osp_precreate_fids() instead of this function.
1815 * 2. FID allocation for each sub-stripe of a striped directory.
1816 * Similar to other FID clients, OSP requests the sequence
1817 * from its corresponding remote MDT, which in turn requests
1818 * sequences from the sequence controller (MDT0).
1820 * \param[in] env execution environment
1821 * \param[in] exp export of the OSP
1822 * \param[out] fid FID being allocated
1823 * \param[in] unused necessary for the interface but unused.
1825 * \retval 0 0 FID allocated successfully.
1826 * \retval 1 1 FID allocated successfully and new sequence
1827 * requested from seq meta server
1828 * \retval negative negative errno if FID allocation failed.
1830 static int osp_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1831 struct lu_fid *fid, struct md_op_data *unused)
1833 struct client_obd *cli = &exp->exp_obd->u.cli;
1834 struct osp_device *osp = lu2osp_dev(exp->exp_obd->obd_lu_dev);
1835 struct lu_client_seq *seq = cli->cl_seq;
1838 LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
1839 /* Sigh, fid client is not ready yet */
1840 LASSERT(osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1842 RETURN(seq_client_alloc_fid(env, seq, fid));
1845 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1846 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1847 static void osp_key_exit(const struct lu_context *ctx,
1848 struct lu_context_key *key, void *data)
1850 struct osp_thread_info *info = data;
1852 info->osi_attr.la_valid = 0;
1855 struct lu_context_key osp_thread_key = {
1856 .lct_tags = LCT_MD_THREAD,
1857 .lct_init = osp_key_init,
1858 .lct_fini = osp_key_fini,
1859 .lct_exit = osp_key_exit
1862 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1863 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1865 struct lu_context_key osp_txn_key = {
1866 .lct_tags = LCT_OSP_THREAD | LCT_TX_HANDLE,
1867 .lct_init = osp_txn_key_init,
1868 .lct_fini = osp_txn_key_fini
1870 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1872 static struct lu_device_type_operations osp_device_type_ops = {
1873 .ldto_init = osp_type_init,
1874 .ldto_fini = osp_type_fini,
1876 .ldto_start = osp_type_start,
1877 .ldto_stop = osp_type_stop,
1879 .ldto_device_alloc = osp_device_alloc,
1880 .ldto_device_free = osp_device_free,
1882 .ldto_device_fini = osp_device_fini
1885 static struct lu_device_type osp_device_type = {
1886 .ldt_tags = LU_DEVICE_DT,
1887 .ldt_name = LUSTRE_OSP_NAME,
1888 .ldt_ops = &osp_device_type_ops,
1889 .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1892 static struct obd_ops osp_obd_device_ops = {
1893 .o_owner = THIS_MODULE,
1894 .o_add_conn = client_import_add_conn,
1895 .o_del_conn = client_import_del_conn,
1896 .o_reconnect = osp_reconnect,
1897 .o_connect = osp_obd_connect,
1898 .o_disconnect = osp_obd_disconnect,
1899 .o_get_info = osp_obd_get_info,
1900 .o_set_info_async = osp_obd_set_info_async,
1901 .o_import_event = osp_import_event,
1902 .o_iocontrol = osp_iocontrol,
1903 .o_statfs = osp_obd_statfs,
1904 .o_fid_init = client_fid_init,
1905 .o_fid_fini = client_fid_fini,
1906 .o_fid_alloc = osp_fid_alloc,
1909 struct llog_operations osp_mds_ost_orig_logops;
1912 * Initialize OSP module.
1914 * Register device types OSP and Light Weight Proxy (LWP) (\see lwp_dev.c)
1915 * in obd_types (\see class_obd.c). Initialize procfs for the
1916 * the OSP device. Note: OSP was called OSC before Lustre 2.4,
1917 * so for compatibility it still uses the name "osc" in procfs.
1918 * This is called at module load time.
1920 * \retval 0 0 if initialization succeeds.
1921 * \retval negative negative errno if initialization failed.
1923 static int __init osp_init(void)
1925 struct obd_type *type;
1928 rc = lu_kmem_init(osp_caches);
1933 rc = class_register_type(&osp_obd_device_ops, NULL, true, NULL,
1934 LUSTRE_OSP_NAME, &osp_device_type);
1936 lu_kmem_fini(osp_caches);
1940 rc = class_register_type(&lwp_obd_device_ops, NULL, true, NULL,
1941 LUSTRE_LWP_NAME, &lwp_device_type);
1943 class_unregister_type(LUSTRE_OSP_NAME);
1944 lu_kmem_fini(osp_caches);
1948 /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1949 osp_mds_ost_orig_logops = llog_osd_ops;
1950 osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1951 osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1953 /* create "osc" entry in procfs for compatibility purposes */
1954 type = class_search_type(LUSTRE_OSC_NAME);
1955 if (type != NULL && type->typ_procroot != NULL)
1958 type = class_search_type(LUSTRE_OSP_NAME);
1959 type->typ_procsym = lprocfs_register("osc", proc_lustre_root,
1961 if (IS_ERR(type->typ_procsym)) {
1962 CERROR("osp: can't create compat entry \"osc\": %d\n",
1963 (int) PTR_ERR(type->typ_procsym));
1964 type->typ_procsym = NULL;
1970 * Finalize OSP module.
1972 * This callback is called when kernel unloads OSP module from memory, and
1973 * it will deregister OSP and LWP device type from obd_types (\see class_obd.c).
1975 static void __exit osp_exit(void)
1977 class_unregister_type(LUSTRE_LWP_NAME);
1978 class_unregister_type(LUSTRE_OSP_NAME);
1979 lu_kmem_fini(osp_caches);
1982 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1983 MODULE_DESCRIPTION("Lustre OSD Storage Proxy ("LUSTRE_OSP_NAME")");
1984 MODULE_VERSION(LUSTRE_VERSION_STRING);
1985 MODULE_LICENSE("GPL");
1987 module_init(osp_init);
1988 module_exit(osp_exit);