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, 2017, 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);
798 struct l_wait_info lwi = { 0 };
799 time64_t start = ktime_get_seconds();
805 /* No Sync between MDTs yet. */
806 if (d->opd_connect_mdt)
809 recs = atomic_read(&d->opd_sync_changes);
810 old = atomic64_read(&d->opd_sync_processed_recs);
812 osp_sync_force(env, dt2osp_dev(dev));
814 if (unlikely(d->opd_imp_active == 0))
817 down_write(&d->opd_async_updates_rwsem);
819 CDEBUG(D_OTHER, "%s: async updates %d\n", d->opd_obd->obd_name,
820 atomic_read(&d->opd_async_updates_count));
822 /* make sure the connection is fine */
823 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout), osp_sync_timeout, d);
824 rc = l_wait_event(d->opd_sync_barrier_waitq,
825 atomic_read(&d->opd_async_updates_count) == 0,
827 up_write(&d->opd_async_updates_rwsem);
831 CDEBUG(D_CACHE, "%s: processed %lu\n", d->opd_obd->obd_name,
832 atomic64_read(&d->opd_sync_processed_recs));
834 while (atomic64_read(&d->opd_sync_processed_recs) < old + recs) {
835 __u64 last = atomic64_read(&d->opd_sync_processed_recs);
836 /* make sure the connection is fine */
837 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
838 osp_sync_timeout, d);
839 l_wait_event(d->opd_sync_barrier_waitq,
840 atomic64_read(&d->opd_sync_processed_recs)
841 >= old + recs, &lwi);
843 if (atomic64_read(&d->opd_sync_processed_recs) >= old + recs)
846 if (atomic64_read(&d->opd_sync_processed_recs) != last) {
847 /* some progress have been made,
852 /* no changes and expired, something is wrong */
853 GOTO(out, rc = -ETIMEDOUT);
856 /* block new processing (barrier>0 - few callers are possible */
857 atomic_inc(&d->opd_sync_barrier);
859 CDEBUG(D_CACHE, "%s: %u in flight\n", d->opd_obd->obd_name,
860 atomic_read(&d->opd_sync_rpcs_in_flight));
862 /* wait till all-in-flight are replied, so executed by the target */
863 /* XXX: this is used by LFSCK at the moment, which doesn't require
864 * all the changes to be committed, but in general it'd be
865 * better to wait till commit */
866 while (atomic_read(&d->opd_sync_rpcs_in_flight) > 0) {
867 old = atomic_read(&d->opd_sync_rpcs_in_flight);
869 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
870 osp_sync_timeout, d);
871 l_wait_event(d->opd_sync_barrier_waitq,
872 atomic_read(&d->opd_sync_rpcs_in_flight) == 0,
875 if (atomic_read(&d->opd_sync_rpcs_in_flight) == 0)
878 if (atomic_read(&d->opd_sync_rpcs_in_flight) != old) {
879 /* some progress have been made */
883 /* no changes and expired, something is wrong */
884 GOTO(out, rc = -ETIMEDOUT);
888 /* resume normal processing (barrier=0) */
889 atomic_dec(&d->opd_sync_barrier);
890 osp_sync_check_for_work(d);
892 CDEBUG(D_CACHE, "%s: done in %lld: rc = %d\n", d->opd_obd->obd_name,
893 ktime_get_seconds() - start, rc);
898 const struct dt_device_operations osp_dt_ops = {
899 .dt_statfs = osp_statfs,
901 .dt_trans_create = osp_trans_create,
902 .dt_trans_start = osp_trans_start,
903 .dt_trans_stop = osp_trans_stop,
904 .dt_trans_cb_add = osp_trans_cb_add,
908 * Connect OSP to local OSD.
910 * Locate the local OSD referenced by \a nextdev and connect to it. Sometimes,
911 * OSP needs to access the local OSD to store some information. For example,
912 * during precreate, it needs to update last used OID and sequence file
913 * (LAST_SEQ) in local OSD.
915 * \param[in] env execution environment
916 * \param[in] osp OSP device
917 * \param[in] nextdev the name of local OSD
919 * \retval 0 0 connection succeeded
920 * \retval negative negative errno connection failed
922 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *osp,
925 struct obd_connect_data *data = NULL;
926 struct obd_device *obd;
931 LASSERT(osp->opd_storage_exp == NULL);
937 obd = class_name2obd(nextdev);
939 CERROR("%s: can't locate next device: %s\n",
940 osp->opd_obd->obd_name, nextdev);
941 GOTO(out, rc = -ENOTCONN);
944 rc = obd_connect(env, &osp->opd_storage_exp, obd, &obd->obd_uuid, data,
947 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
948 osp->opd_obd->obd_name, nextdev, rc);
952 osp->opd_dt_dev.dd_lu_dev.ld_site =
953 osp->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
954 LASSERT(osp->opd_dt_dev.dd_lu_dev.ld_site);
955 osp->opd_storage = lu2dt_dev(osp->opd_storage_exp->exp_obd->obd_lu_dev);
963 * Determine if the lock needs to be cancelled
965 * Determine if the unused lock should be cancelled before replay, see
966 * (ldlm_cancel_no_wait_policy()). Currently, only inode bits lock exists
969 * \param[in] lock lock to be checked.
971 * \retval 1 if the lock needs to be cancelled before replay.
972 * \retval 0 if the lock does not need to be cancelled before
975 static int osp_cancel_weight(struct ldlm_lock *lock)
977 if (lock->l_resource->lr_type != LDLM_IBITS)
984 * Initialize OSP device according to the parameters in the configuration
987 * Reconstruct the local device name from the configuration profile, and
988 * initialize necessary threads and structures according to the OSP type
991 * Since there is no record in the MDT configuration for the local disk
992 * device, we have to extract this from elsewhere in the profile.
993 * The only information we get at setup is from the OSC records:
994 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
996 * Note: configs generated by Lustre 1.8 are missing the -MDTxxxx part,
997 * so, we need to reconstruct the name of the underlying OSD from this:
998 * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".
1000 * \param[in] env execution environment
1001 * \param[in] osp OSP device
1002 * \param[in] ldt lu device type of OSP
1003 * \param[in] cfg configuration log
1005 * \retval 0 0 if OSP initialization succeeded.
1006 * \retval negative negative errno if OSP initialization failed.
1008 static int osp_init0(const struct lu_env *env, struct osp_device *osp,
1009 struct lu_device_type *ldt, struct lustre_cfg *cfg)
1011 struct obd_device *obd;
1012 struct obd_import *imp;
1014 char *src, *tgt, *mdt, *osdname = NULL;
1020 mutex_init(&osp->opd_async_requests_mutex);
1021 INIT_LIST_HEAD(&osp->opd_async_updates);
1022 init_rwsem(&osp->opd_async_updates_rwsem);
1023 atomic_set(&osp->opd_async_updates_count, 0);
1025 obd = class_name2obd(lustre_cfg_string(cfg, 0));
1027 CERROR("Cannot find obd with name %s\n",
1028 lustre_cfg_string(cfg, 0));
1033 src = lustre_cfg_string(cfg, 0);
1037 tgt = strrchr(src, '-');
1039 CERROR("%s: invalid target name %s: rc = %d\n",
1040 osp->opd_obd->obd_name, lustre_cfg_string(cfg, 0),
1045 if (strncmp(tgt, "-osc", 4) == 0) {
1046 /* Old OSC name fsname-OSTXXXX-osc */
1047 for (tgt--; tgt > src && *tgt != '-'; tgt--)
1050 CERROR("%s: invalid target name %s: rc = %d\n",
1051 osp->opd_obd->obd_name,
1052 lustre_cfg_string(cfg, 0), -EINVAL);
1056 if (strncmp(tgt, "-OST", 4) != 0) {
1057 CERROR("%s: invalid target name %s: rc = %d\n",
1058 osp->opd_obd->obd_name,
1059 lustre_cfg_string(cfg, 0), -EINVAL);
1063 idx = simple_strtol(tgt + 4, &mdt, 16);
1064 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1065 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1066 osp->opd_obd->obd_name, src, -EINVAL);
1069 osp->opd_index = idx;
1073 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
1074 if (strncmp(tgt, "-MDT", 4) != 0 &&
1075 strncmp(tgt, "-OST", 4) != 0) {
1076 CERROR("%s: invalid target name %s: rc = %d\n",
1077 osp->opd_obd->obd_name,
1078 lustre_cfg_string(cfg, 0), -EINVAL);
1082 idx = simple_strtol(tgt + 4, &mdt, 16);
1083 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
1084 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1085 osp->opd_obd->obd_name, src, -EINVAL);
1089 /* Get MDT index from the name and set it to opd_group,
1090 * which will be used by OSP to connect with OST */
1091 osp->opd_group = idx;
1092 if (tgt - src <= 12) {
1093 CERROR("%s: invalid mdt index from %s: rc =%d\n",
1094 osp->opd_obd->obd_name,
1095 lustre_cfg_string(cfg, 0), -EINVAL);
1099 if (strncmp(tgt - 12, "-MDT", 4) == 0)
1100 osp->opd_connect_mdt = 1;
1102 idx = simple_strtol(tgt - 8, &mdt, 16);
1103 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1104 CERROR("%s: invalid OST index in '%s': rc =%d\n",
1105 osp->opd_obd->obd_name, src, -EINVAL);
1109 osp->opd_index = idx;
1110 idx = tgt - src - 12;
1112 /* check the fsname length, and after this everything else will fit */
1113 if (idx > MTI_NAME_MAXLEN) {
1114 CERROR("%s: fsname too long in '%s': rc = %d\n",
1115 osp->opd_obd->obd_name, src, -EINVAL);
1119 OBD_ALLOC(osdname, MAX_OBD_NAME);
1120 if (osdname == NULL)
1123 memcpy(osdname, src, idx); /* copy just the fsname part */
1124 osdname[idx] = '\0';
1126 mdt = strstr(mdt, "-MDT");
1127 if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
1128 strcat(osdname, "-MDT0000");
1130 strcat(osdname, mdt);
1131 strcat(osdname, "-osd");
1132 CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
1134 if (osp->opd_connect_mdt) {
1135 struct client_obd *cli = &osp->opd_obd->u.cli;
1137 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
1138 if (!cli->cl_rpc_lock)
1139 GOTO(out_fini, rc = -ENOMEM);
1140 osp_init_rpc_lock(cli->cl_rpc_lock);
1143 osp->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
1144 osp->opd_dt_dev.dd_ops = &osp_dt_ops;
1146 obd->obd_lu_dev = &osp->opd_dt_dev.dd_lu_dev;
1148 rc = osp_connect_to_osd(env, osp, osdname);
1152 rc = ptlrpcd_addref();
1154 GOTO(out_disconnect, rc);
1156 rc = client_obd_setup(obd, cfg);
1158 CERROR("%s: can't setup obd: rc = %d\n", osp->opd_obd->obd_name,
1163 osp_lprocfs_init(osp);
1165 rc = obd_fid_init(osp->opd_obd, NULL, osp->opd_connect_mdt ?
1166 LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
1168 CERROR("%s: fid init error: rc = %d\n",
1169 osp->opd_obd->obd_name, rc);
1173 if (!osp->opd_connect_mdt) {
1174 /* Initialize last id from the storage - will be
1175 * used in orphan cleanup. */
1176 if (!osp->opd_storage->dd_rdonly) {
1177 rc = osp_last_used_init(env, osp);
1182 /* Initialize precreation thread, it handles new
1183 * connections as well. */
1184 rc = osp_init_precreate(osp);
1186 GOTO(out_last_used, rc);
1189 * Initialize synhronization mechanism taking
1190 * care of propogating changes to OST in near
1191 * transactional manner.
1193 rc = osp_sync_init(env, osp);
1195 GOTO(out_precreat, rc);
1197 rc = osp_update_init(osp);
1202 ns_register_cancel(obd->obd_namespace, osp_cancel_weight);
1205 * Initiate connect to OST
1207 ll_generate_random_uuid(uuid);
1208 class_uuid_unparse(uuid, &osp->opd_cluuid);
1210 imp = obd->u.cli.cl_import;
1212 rc = ptlrpc_init_import(imp);
1216 OBD_FREE(osdname, MAX_OBD_NAME);
1220 if (!osp->opd_connect_mdt)
1221 /* stop sync thread */
1224 /* stop precreate thread */
1225 if (!osp->opd_connect_mdt)
1226 osp_precreate_fini(osp);
1228 osp_update_fini(env, osp);
1230 if (!osp->opd_connect_mdt)
1231 osp_last_used_fini(env, osp);
1233 obd_fid_fini(osp->opd_obd);
1235 osp_lprocfs_fini(osp);
1236 client_obd_cleanup(obd);
1240 if (osp->opd_connect_mdt) {
1241 struct client_obd *cli = &osp->opd_obd->u.cli;
1242 if (cli->cl_rpc_lock != NULL) {
1243 OBD_FREE_PTR(cli->cl_rpc_lock);
1244 cli->cl_rpc_lock = NULL;
1247 obd_disconnect(osp->opd_storage_exp);
1250 OBD_FREE(osdname, MAX_OBD_NAME);
1255 * Implementation of lu_device_type_operations::ldto_device_free
1257 * Free the OSP device in memory. No return value is needed for now,
1258 * so always return NULL to comply with the interface.
1260 * \param[in] env execution environment
1261 * \param[in] lu lu_device of OSP
1263 * \retval NULL NULL unconditionally
1265 static struct lu_device *osp_device_free(const struct lu_env *env,
1266 struct lu_device *lu)
1268 struct osp_device *osp = lu2osp_dev(lu);
1270 if (atomic_read(&lu->ld_ref) && lu->ld_site) {
1271 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1272 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
1274 dt_device_fini(&osp->opd_dt_dev);
1281 * Implementation of lu_device_type_operations::ldto_device_alloc
1283 * This function allocates and initializes OSP device in memory according to
1286 * \param[in] env execution environment
1287 * \param[in] type device type of OSP
1288 * \param[in] lcfg config log
1290 * \retval pointer the pointer of allocated OSP if succeed.
1291 * \retval ERR_PTR(errno) ERR_PTR(errno) if failed.
1293 static struct lu_device *osp_device_alloc(const struct lu_env *env,
1294 struct lu_device_type *type,
1295 struct lustre_cfg *lcfg)
1297 struct osp_device *osp;
1298 struct lu_device *ld;
1302 ld = ERR_PTR(-ENOMEM);
1306 ld = osp2lu_dev(osp);
1307 dt_device_init(&osp->opd_dt_dev, type);
1308 rc = osp_init0(env, osp, type, lcfg);
1310 osp_device_free(env, ld);
1318 * Implementation of lu_device_type_operations::ldto_device_fini
1320 * This function cleans up the OSP device, i.e. release and free those
1321 * attached items in osp_device.
1323 * \param[in] env execution environment
1324 * \param[in] ld lu_device of OSP
1326 * \retval NULL NULL if cleanup succeeded.
1327 * \retval ERR_PTR(errno) ERR_PTR(errno) if cleanup failed.
1329 static struct lu_device *osp_device_fini(const struct lu_env *env,
1330 struct lu_device *ld)
1332 struct osp_device *osp = lu2osp_dev(ld);
1337 if (osp->opd_async_requests != NULL) {
1338 osp_update_request_destroy(env, osp->opd_async_requests);
1339 osp->opd_async_requests = NULL;
1342 if (osp->opd_storage_exp) {
1343 /* wait for the commit callbacks to complete */
1344 wait_event(osp->opd_sync_waitq,
1345 atomic_read(&osp->opd_commits_registered) == 0);
1346 obd_disconnect(osp->opd_storage_exp);
1349 LASSERT(osp->opd_obd);
1350 osp_lprocfs_fini(osp);
1352 if (osp->opd_connect_mdt) {
1353 struct client_obd *cli = &osp->opd_obd->u.cli;
1354 if (cli->cl_rpc_lock != NULL) {
1355 OBD_FREE_PTR(cli->cl_rpc_lock);
1356 cli->cl_rpc_lock = NULL;
1360 rc = client_obd_cleanup(osp->opd_obd);
1363 RETURN(ERR_PTR(rc));
1372 * Implementation of obd_ops::o_reconnect
1374 * This function is empty and does not need to do anything for now.
1376 static int osp_reconnect(const struct lu_env *env,
1377 struct obd_export *exp, struct obd_device *obd,
1378 struct obd_uuid *cluuid,
1379 struct obd_connect_data *data,
1386 * Implementation of obd_ops::o_connect
1388 * Connect OSP to the remote target (MDT or OST). Allocate the
1389 * export and return it to the LOD, which calls this function
1390 * for each OSP to connect it to the remote target. This function
1391 * is currently only called once per OSP.
1393 * \param[in] env execution environment
1394 * \param[out] exp export connected to OSP
1395 * \param[in] obd OSP device
1396 * \param[in] cluuid OSP device client uuid
1397 * \param[in] data connect_data to be used to connect to the remote
1399 * \param[in] localdata necessary for the API interface, but not used in
1402 * \retval 0 0 if the connection succeeded.
1403 * \retval negative negative errno if the connection failed.
1405 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
1406 struct obd_device *obd, struct obd_uuid *cluuid,
1407 struct obd_connect_data *data, void *localdata)
1409 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1410 struct obd_connect_data *ocd;
1411 struct obd_import *imp;
1412 struct lustre_handle conn;
1417 CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
1419 rc = class_connect(&conn, obd, cluuid);
1423 *exp = class_conn2export(&conn);
1424 /* Why should there ever be more than 1 connect? */
1425 osp->opd_connects++;
1426 LASSERT(osp->opd_connects == 1);
1428 osp->opd_exp = *exp;
1430 imp = osp->opd_obd->u.cli.cl_import;
1431 imp->imp_dlm_handle = conn;
1433 LASSERT(data != NULL);
1434 LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
1435 ocd = &imp->imp_connect_data;
1438 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
1439 imp->imp_connect_flags2_orig = ocd->ocd_connect_flags2;
1441 ocd->ocd_version = LUSTRE_VERSION_CODE;
1442 ocd->ocd_index = data->ocd_index;
1444 rc = ptlrpc_connect_import(imp);
1446 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
1449 osp->opd_obd->u.cli.cl_seq->lcs_exp =
1450 class_export_get(osp->opd_exp);
1453 ptlrpc_pinger_add_import(imp);
1459 * Implementation of obd_ops::o_disconnect
1461 * Disconnect the export for the OSP. This is called by LOD to release the
1462 * OSP during cleanup (\see lod_del_device()). The OSP will be released after
1463 * the export is released.
1465 * \param[in] exp export to be disconnected.
1467 * \retval 0 0 if disconnection succeed
1468 * \retval negative negative errno if disconnection failed
1470 static int osp_obd_disconnect(struct obd_export *exp)
1472 struct obd_device *obd = exp->exp_obd;
1473 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1477 /* Only disconnect the underlying layers on the final disconnect. */
1478 LASSERT(osp->opd_connects == 1);
1479 osp->opd_connects--;
1481 rc = class_disconnect(exp);
1483 CERROR("%s: class disconnect error: rc = %d\n",
1488 /* destroy the device */
1489 class_manual_cleanup(obd);
1495 * Implementation of obd_ops::o_statfs
1497 * Send a RPC to the remote target to get statfs status. This is only used
1498 * in lprocfs helpers by obd_statfs.
1500 * \param[in] env execution environment
1501 * \param[in] exp connection state from this OSP to the parent (LOD)
1503 * \param[out] osfs hold the statfs result
1504 * \param[in] unused Not used in this function for now
1505 * \param[in] flags flags to indicate how OSP will issue the RPC
1507 * \retval 0 0 if statfs succeeded.
1508 * \retval negative negative errno if statfs failed.
1510 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
1511 struct obd_statfs *osfs, time64_t unused, __u32 flags)
1513 struct obd_statfs *msfs;
1514 struct ptlrpc_request *req;
1515 struct obd_import *imp = NULL;
1520 /* Since the request might also come from lprocfs, so we need
1521 * sync this with client_disconnect_export Bug15684 */
1522 down_read(&exp->exp_obd->u.cli.cl_sem);
1523 if (exp->exp_obd->u.cli.cl_import)
1524 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
1525 up_read(&exp->exp_obd->u.cli.cl_sem);
1529 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1531 class_import_put(imp);
1536 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1538 ptlrpc_request_free(req);
1541 ptlrpc_request_set_replen(req);
1542 req->rq_request_portal = OST_CREATE_PORTAL;
1543 ptlrpc_at_set_req_timeout(req);
1545 if (flags & OBD_STATFS_NODELAY) {
1546 /* procfs requests not want stat in wait for avoid deadlock */
1547 req->rq_no_resend = 1;
1548 req->rq_no_delay = 1;
1551 rc = ptlrpc_queue_wait(req);
1555 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1557 GOTO(out, rc = -EPROTO);
1563 ptlrpc_req_finished(req);
1568 * Implementation of obd_ops::o_import_event
1570 * This function is called when some related import event happens. It will
1571 * mark the necessary flags according to the event and notify the necessary
1572 * threads (mainly precreate thread).
1574 * \param[in] obd OSP OBD device
1575 * \param[in] imp import attached from OSP to remote (OST/MDT) service
1576 * \param[in] event event related to remote service (IMP_EVENT_*)
1578 * \retval 0 0 if the event handling succeeded.
1579 * \retval negative negative errno if the event handling failed.
1581 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1582 enum obd_import_event event)
1584 struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1588 case IMP_EVENT_DISCON:
1589 d->opd_got_disconnected = 1;
1590 d->opd_imp_connected = 0;
1591 if (d->opd_connect_mdt)
1594 if (d->opd_pre != NULL) {
1595 osp_pre_update_status(d, -ENODEV);
1596 wake_up(&d->opd_pre_waitq);
1599 CDEBUG(D_HA, "got disconnected\n");
1601 case IMP_EVENT_INACTIVE:
1602 d->opd_imp_active = 0;
1603 d->opd_imp_connected = 0;
1604 d->opd_obd->obd_inactive = 1;
1605 if (d->opd_connect_mdt)
1607 if (d->opd_pre != NULL) {
1608 /* Import is invalid, we can`t get stripes so
1610 rc = imp->imp_deactive ? -ESHUTDOWN : -ENODEV;
1611 osp_pre_update_status(d, rc);
1612 wake_up(&d->opd_pre_waitq);
1615 CDEBUG(D_HA, "got inactive\n");
1617 case IMP_EVENT_ACTIVE:
1618 d->opd_imp_active = 1;
1620 if (d->opd_got_disconnected)
1621 d->opd_new_connection = 1;
1622 d->opd_imp_connected = 1;
1623 d->opd_imp_seen_connected = 1;
1624 d->opd_obd->obd_inactive = 0;
1625 if (d->opd_connect_mdt)
1628 if (d->opd_pre != NULL)
1629 wake_up(&d->opd_pre_waitq);
1631 osp_sync_check_for_work(d);
1632 CDEBUG(D_HA, "got connected\n");
1634 case IMP_EVENT_INVALIDATE:
1635 if (d->opd_connect_mdt)
1636 osp_invalidate_request(d);
1638 if (obd->obd_namespace == NULL)
1640 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1643 case IMP_EVENT_DEACTIVATE:
1644 case IMP_EVENT_ACTIVATE:
1647 CERROR("%s: unsupported import event: %#x\n",
1648 obd->obd_name, event);
1654 * Implementation of obd_ops: o_iocontrol
1656 * This function is the ioctl handler for OSP. Note: lctl will access the OSP
1657 * directly by ioctl, instead of through the MDS stack.
1659 * param[in] cmd ioctl command.
1660 * param[in] exp export of this OSP.
1661 * param[in] len data length of \a karg.
1662 * param[in] karg input argument which is packed as
1664 * param[out] uarg pointer to userspace buffer (must access by
1667 * \retval 0 0 if the ioctl handling succeeded.
1668 * \retval negative negative errno if the ioctl handling failed.
1670 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1671 void *karg, void __user *uarg)
1673 struct obd_device *obd = exp->exp_obd;
1674 struct osp_device *d;
1675 struct obd_ioctl_data *data = karg;
1680 LASSERT(obd->obd_lu_dev);
1681 d = lu2osp_dev(obd->obd_lu_dev);
1682 LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1684 if (!try_module_get(THIS_MODULE)) {
1685 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1686 module_name(THIS_MODULE));
1691 case OBD_IOC_CLIENT_RECOVER:
1692 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1693 data->ioc_inlbuf1, 0);
1697 case IOC_OSC_SET_ACTIVE:
1698 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1701 case OBD_IOC_PING_TARGET:
1702 rc = ptlrpc_obd_ping(obd);
1705 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1706 cmd, current_comm());
1709 module_put(THIS_MODULE);
1715 * Implementation of obd_ops::o_get_info
1717 * Retrieve information by key. Retrieval starts from the top layer
1718 * (MDT) of the MDS stack and traverses the stack by calling the
1719 * obd_get_info() method of the next sub-layer.
1721 * \param[in] env execution environment
1722 * \param[in] exp export of this OSP
1723 * \param[in] keylen length of \a key
1724 * \param[in] key the key
1725 * \param[out] vallen length of \a val
1726 * \param[out] val holds the value returned by the key
1728 * \retval 0 0 if getting information succeeded.
1729 * \retval negative negative errno if getting information failed.
1731 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1732 __u32 keylen, void *key, __u32 *vallen, void *val)
1736 if (KEY_IS(KEY_OSP_CONNECTED)) {
1737 struct obd_device *obd = exp->exp_obd;
1738 struct osp_device *osp;
1740 if (!obd->obd_set_up || obd->obd_stopping)
1743 osp = lu2osp_dev(obd->obd_lu_dev);
1746 * 1.8/2.0 behaviour is that OST being connected once at least
1747 * is considered "healthy". and one "healthy" OST is enough to
1748 * allow lustre clients to connect to MDS
1750 RETURN(!osp->opd_imp_seen_connected);
1756 static int osp_obd_set_info_async(const struct lu_env *env,
1757 struct obd_export *exp,
1758 u32 keylen, void *key,
1759 u32 vallen, void *val,
1760 struct ptlrpc_request_set *set)
1762 struct obd_device *obd = exp->exp_obd;
1763 struct obd_import *imp = obd->u.cli.cl_import;
1764 struct osp_device *osp;
1765 struct ptlrpc_request *req;
1769 if (KEY_IS(KEY_SPTLRPC_CONF)) {
1770 sptlrpc_conf_client_adapt(exp->exp_obd);
1774 LASSERT(set != NULL);
1775 if (!obd->obd_set_up || obd->obd_stopping)
1777 osp = lu2osp_dev(obd->obd_lu_dev);
1779 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
1783 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
1784 RCL_CLIENT, keylen);
1785 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
1786 RCL_CLIENT, vallen);
1787 if (osp->opd_connect_mdt)
1788 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SET_INFO);
1790 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
1792 ptlrpc_request_free(req);
1796 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1797 memcpy(tmp, key, keylen);
1798 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1799 memcpy(tmp, val, vallen);
1801 ptlrpc_request_set_replen(req);
1802 ptlrpc_set_add_req(set, req);
1803 ptlrpc_check_set(NULL, set);
1809 * Implementation of obd_ops: o_fid_alloc
1811 * Allocate a FID. There are two cases in which OSP performs
1814 * 1. FID precreation for data objects, which is done in
1815 * osp_precreate_fids() instead of this function.
1816 * 2. FID allocation for each sub-stripe of a striped directory.
1817 * Similar to other FID clients, OSP requests the sequence
1818 * from its corresponding remote MDT, which in turn requests
1819 * sequences from the sequence controller (MDT0).
1821 * \param[in] env execution environment
1822 * \param[in] exp export of the OSP
1823 * \param[out] fid FID being allocated
1824 * \param[in] unused necessary for the interface but unused.
1826 * \retval 0 0 FID allocated successfully.
1827 * \retval 1 1 FID allocated successfully and new sequence
1828 * requested from seq meta server
1829 * \retval negative negative errno if FID allocation failed.
1831 static int osp_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1832 struct lu_fid *fid, struct md_op_data *unused)
1834 struct client_obd *cli = &exp->exp_obd->u.cli;
1835 struct osp_device *osp = lu2osp_dev(exp->exp_obd->obd_lu_dev);
1836 struct lu_client_seq *seq = cli->cl_seq;
1839 LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
1840 /* Sigh, fid client is not ready yet */
1841 LASSERT(osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1843 RETURN(seq_client_alloc_fid(env, seq, fid));
1846 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1847 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1848 static void osp_key_exit(const struct lu_context *ctx,
1849 struct lu_context_key *key, void *data)
1851 struct osp_thread_info *info = data;
1853 info->osi_attr.la_valid = 0;
1856 struct lu_context_key osp_thread_key = {
1857 .lct_tags = LCT_MD_THREAD,
1858 .lct_init = osp_key_init,
1859 .lct_fini = osp_key_fini,
1860 .lct_exit = osp_key_exit
1863 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1864 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1866 struct lu_context_key osp_txn_key = {
1867 .lct_tags = LCT_OSP_THREAD,
1868 .lct_init = osp_txn_key_init,
1869 .lct_fini = osp_txn_key_fini
1871 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1873 static struct lu_device_type_operations osp_device_type_ops = {
1874 .ldto_init = osp_type_init,
1875 .ldto_fini = osp_type_fini,
1877 .ldto_start = osp_type_start,
1878 .ldto_stop = osp_type_stop,
1880 .ldto_device_alloc = osp_device_alloc,
1881 .ldto_device_free = osp_device_free,
1883 .ldto_device_fini = osp_device_fini
1886 static struct lu_device_type osp_device_type = {
1887 .ldt_tags = LU_DEVICE_DT,
1888 .ldt_name = LUSTRE_OSP_NAME,
1889 .ldt_ops = &osp_device_type_ops,
1890 .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1893 static struct obd_ops osp_obd_device_ops = {
1894 .o_owner = THIS_MODULE,
1895 .o_add_conn = client_import_add_conn,
1896 .o_del_conn = client_import_del_conn,
1897 .o_reconnect = osp_reconnect,
1898 .o_connect = osp_obd_connect,
1899 .o_disconnect = osp_obd_disconnect,
1900 .o_get_info = osp_obd_get_info,
1901 .o_set_info_async = osp_obd_set_info_async,
1902 .o_import_event = osp_import_event,
1903 .o_iocontrol = osp_iocontrol,
1904 .o_statfs = osp_obd_statfs,
1905 .o_fid_init = client_fid_init,
1906 .o_fid_fini = client_fid_fini,
1907 .o_fid_alloc = osp_fid_alloc,
1910 struct llog_operations osp_mds_ost_orig_logops;
1912 static struct obd_type sym;
1915 * Initialize OSP module.
1917 * Register device types OSP and Light Weight Proxy (LWP) (\see lwp_dev.c)
1918 * in obd_types (\see class_obd.c). Initialize procfs for the
1919 * the OSP device. Note: OSP was called OSC before Lustre 2.4,
1920 * so for compatibility it still uses the name "osc" in procfs.
1921 * This is called at module load time.
1923 * \retval 0 0 if initialization succeeds.
1924 * \retval negative negative errno if initialization failed.
1926 static int __init osp_init(void)
1928 struct dentry *symlink;
1929 struct obd_type *type;
1930 struct kobject *kobj;
1934 rc = lu_kmem_init(osp_caches);
1938 rc = class_register_type(&osp_obd_device_ops, NULL, true, NULL,
1939 LUSTRE_OSP_NAME, &osp_device_type);
1941 lu_kmem_fini(osp_caches);
1945 rc = class_register_type(&lwp_obd_device_ops, NULL, false, NULL,
1946 LUSTRE_LWP_NAME, &lwp_device_type);
1948 class_unregister_type(LUSTRE_OSP_NAME);
1949 lu_kmem_fini(osp_caches);
1953 /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1954 osp_mds_ost_orig_logops = llog_osd_ops;
1955 osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1956 osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1958 /* create "osc" entry for compatibility purposes */
1960 dname.len = strlen(dname.name);
1961 dname.hash = ll_full_name_hash(debugfs_lustre_root, dname.name,
1963 symlink = d_lookup(debugfs_lustre_root, &dname);
1965 symlink = debugfs_create_dir(dname.name, debugfs_lustre_root);
1966 if (IS_ERR_OR_NULL(symlink)) {
1967 rc = symlink ? PTR_ERR(symlink) : -ENOMEM;
1970 sym.typ_debugfs_entry = symlink;
1975 kobj = kset_find_obj(lustre_kset, dname.name);
1981 kobj = class_setup_tunables(dname.name);
1984 if (sym.typ_debugfs_entry)
1985 ldebugfs_remove(&sym.typ_debugfs_entry);
1988 sym.typ_kobj = kobj;
1991 type = class_search_type(LUSTRE_OSC_NAME);
1992 if (type != NULL && type->typ_procroot != NULL)
1995 type = class_search_type(LUSTRE_OSP_NAME);
1996 type->typ_procsym = lprocfs_register("osc", proc_lustre_root,
1998 if (IS_ERR(type->typ_procsym)) {
1999 CERROR("osp: can't create compat entry \"osc\": %d\n",
2000 (int) PTR_ERR(type->typ_procsym));
2001 type->typ_procsym = NULL;
2008 * Finalize OSP module.
2010 * This callback is called when kernel unloads OSP module from memory, and
2011 * it will deregister OSP and LWP device type from obd_types (\see class_obd.c).
2013 static void __exit osp_exit(void)
2015 ldebugfs_remove(&sym.typ_debugfs_entry);
2016 kobject_put(sym.typ_kobj);
2017 class_unregister_type(LUSTRE_LWP_NAME);
2018 class_unregister_type(LUSTRE_OSP_NAME);
2019 lu_kmem_fini(osp_caches);
2022 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2023 MODULE_DESCRIPTION("Lustre OSD Storage Proxy ("LUSTRE_OSP_NAME")");
2024 MODULE_VERSION(LUSTRE_VERSION_STRING);
2025 MODULE_LICENSE("GPL");
2027 module_init(osp_init);
2028 module_exit(osp_exit);