4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osp/osp_dev.c
38 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
39 * Author: Mikhail Pershin <mike.pershin@intel.com>
40 * Author: Di Wang <di.wang@intel.com>
43 * The Object Storage Proxy (OSP) module provides an implementation of
44 * the DT API for remote MDTs and OSTs. Every local OSP device (or
45 * object) is a proxy for a remote OSD device (or object). Thus OSP
46 * converts DT operations into RPCs, which are sent to the OUT service
47 * on a remote target, converted back to DT operations, and
48 * executed. Of course there are many ways in which this description
49 * is inaccurate but it's a good enough mental model. OSP is used by
50 * the MDT stack in several ways:
52 * - OSP devices allocate FIDs for the stripe sub-objects of a striped
55 * - OSP objects represent the remote MDT and OST objects that are
56 * the stripes of a striped object.
58 * - OSP devices log, send, and track synchronous operations (setattr
59 * and unlink) to remote targets.
61 * - OSP objects are the bottom slice of the compound LU object
62 * representing a remote MDT object: MDT/MDD/LOD/OSP.
64 * - OSP objects are used by LFSCK to represent remote OST objects
65 * during the verification of MDT-OST consistency.
67 * - OSP devices batch idempotent requests (declare_attr_get() and
68 * declare_xattr_get()) to the remote target and cache their results.
70 * In addition the OSP layer implements a subset of the OBD device API
71 * to support being a client of a remote target, connecting to other
72 * layers, and FID allocation.
75 #define DEBUG_SUBSYSTEM S_MDS
77 #include <linux/kthread.h>
79 #include <lustre_ioctl.h>
80 #include <lustre_log.h>
81 #include <lustre_obdo.h>
82 #include <lustre_param.h>
83 #include <obd_class.h>
85 #include "osp_internal.h"
87 /* Slab for OSP object allocation */
88 struct kmem_cache *osp_object_kmem;
90 static struct lu_kmem_descr osp_caches[] = {
92 .ckd_cache = &osp_object_kmem,
93 .ckd_name = "osp_obj",
94 .ckd_size = sizeof(struct osp_object)
102 * Implementation of lu_device_operations::ldo_object_alloc
104 * Allocates an OSP object in memory, whose FID is on the remote target.
106 * \param[in] env execution environment
107 * \param[in] hdr The header of the object stack. If it is NULL, it
108 * means the object is not built from top device, i.e.
109 * it is a sub-stripe object of striped directory or
111 * \param[in] d OSP device
113 * \retval object object being created if the creation succeed.
114 * \retval NULL NULL if the creation failed.
116 static struct lu_object *osp_object_alloc(const struct lu_env *env,
117 const struct lu_object_header *hdr,
120 struct lu_object_header *h = NULL;
121 struct osp_object *o;
124 OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, GFP_NOFS);
126 l = &o->opo_obj.do_lu;
128 /* If hdr is NULL, it means the object is not built
129 * from the top dev(MDT/OST), usually it happens when
130 * building striped object, like data object on MDT or
131 * striped object for directory */
134 lu_object_header_init(h);
135 dt_object_init(&o->opo_obj, h, d);
136 lu_object_add_top(h, l);
138 dt_object_init(&o->opo_obj, h, d);
141 l->lo_ops = &osp_lu_obj_ops;
150 * Find or create the local object
152 * Finds or creates the local file referenced by \a reg_id and return the
153 * attributes of the local file.
155 * \param[in] env execution environment
156 * \param[in] osp OSP device
157 * \param[out] attr attributes of the object
158 * \param[in] reg_id the local object ID of the file. It will be used
159 * to compose a local FID{FID_SEQ_LOCAL_FILE, reg_id, 0}
160 * to identify the object.
162 * \retval object object(dt_object) found or created
163 * \retval ERR_PTR(errno) ERR_PTR(errno) if not get the object.
165 static struct dt_object
166 *osp_find_or_create_local_file(const struct lu_env *env, struct osp_device *osp,
167 struct lu_attr *attr, __u32 reg_id)
169 struct osp_thread_info *osi = osp_env_info(env);
170 struct dt_object_format dof = { 0 };
171 struct dt_object *dto;
175 lu_local_obj_fid(&osi->osi_fid, reg_id);
176 attr->la_valid = LA_MODE;
177 attr->la_mode = S_IFREG | 0644;
178 dof.dof_type = DFT_REGULAR;
179 /* Find or create the local object by osi_fid. */
180 dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
185 /* Get attributes of the local object. */
186 rc = dt_attr_get(env, dto, attr);
188 CERROR("%s: can't be initialized: rc = %d\n",
189 osp->opd_obd->obd_name, rc);
190 lu_object_put(env, &dto->do_lu);
197 * Write data buffer to a local file object.
199 * \param[in] env execution environment
200 * \param[in] osp OSP device
201 * \param[in] dt_obj object written to
202 * \param[in] buf buffer containing byte array and length
203 * \param[in] offset write offset in the object in bytes
205 * \retval 0 0 if write succeed
206 * \retval -EFAULT -EFAULT if only part of buffer is written.
207 * \retval negative other negative errno if write failed.
209 static int osp_write_local_file(const struct lu_env *env,
210 struct osp_device *osp,
211 struct dt_object *dt_obj,
218 th = dt_trans_create(env, osp->opd_storage);
222 rc = dt_declare_record_write(env, dt_obj, buf, offset, th);
225 rc = dt_trans_start_local(env, osp->opd_storage, th);
229 rc = dt_record_write(env, dt_obj, buf, &offset, th);
231 dt_trans_stop(env, osp->opd_storage, th);
236 * Initialize last ID object.
238 * This function initializes the LAST_ID file, which stores the current last
239 * used id of data objects. The MDT will use the last used id and the last_seq
240 * (\see osp_init_last_seq()) to synchronize the precreate object cache with
243 * \param[in] env execution environment
244 * \param[in] osp OSP device
246 * \retval 0 0 if initialization succeed
247 * \retval negative negative errno if initialization failed
249 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
251 struct osp_thread_info *osi = osp_env_info(env);
252 struct lu_fid *fid = &osp->opd_last_used_fid;
253 struct dt_object *dto;
257 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
260 RETURN(PTR_ERR(dto));
262 /* object will be released in device cleanup path */
263 if (osi->osi_attr.la_size >=
264 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
265 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
267 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
268 if (rc != 0 && rc != -EFAULT)
272 if (rc == -EFAULT) { /* fresh LAST_ID */
274 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
276 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
281 osp->opd_last_used_oid_file = dto;
284 /* object will be released in device cleanup path */
285 CERROR("%s: can't initialize lov_objid: rc = %d\n",
286 osp->opd_obd->obd_name, rc);
287 lu_object_put(env, &dto->do_lu);
288 osp->opd_last_used_oid_file = NULL;
293 * Initialize last sequence object.
295 * This function initializes the LAST_SEQ file in the local OSD, which stores
296 * the current last used sequence of data objects. The MDT will use the last
297 * sequence and last id (\see osp_init_last_objid()) to synchronize the
298 * precreate object cache with OSTs.
300 * \param[in] env execution environment
301 * \param[in] osp OSP device
303 * \retval 0 0 if initialization succeed
304 * \retval negative negative errno if initialization failed
306 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
308 struct osp_thread_info *osi = osp_env_info(env);
309 struct lu_fid *fid = &osp->opd_last_used_fid;
310 struct dt_object *dto;
314 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
317 RETURN(PTR_ERR(dto));
319 /* object will be released in device cleanup path */
320 if (osi->osi_attr.la_size >=
321 sizeof(osi->osi_id) * (osp->opd_index + 1)) {
322 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
324 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
325 if (rc != 0 && rc != -EFAULT)
329 if (rc == -EFAULT) { /* fresh OSP */
331 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
333 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
338 osp->opd_last_used_seq_file = dto;
341 /* object will be released in device cleanup path */
342 CERROR("%s: can't initialize lov_seq: rc = %d\n",
343 osp->opd_obd->obd_name, rc);
344 lu_object_put(env, &dto->do_lu);
345 osp->opd_last_used_seq_file = NULL;
350 * Initialize last OID and sequence object.
352 * If the MDT is just upgraded to 2.4 from the lower version, where the
353 * LAST_SEQ file does not exist, the file will be created and IDIF sequence
354 * will be written into the file.
356 * \param[in] env execution environment
357 * \param[in] osp OSP device
359 * \retval 0 0 if initialization succeed
360 * \retval negative negative error if initialization failed
362 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
364 struct osp_thread_info *osi = osp_env_info(env);
368 fid_zero(&osp->opd_last_used_fid);
369 rc = osp_init_last_objid(env, osp);
371 CERROR("%s: Can not get ids %d from old objid!\n",
372 osp->opd_obd->obd_name, rc);
376 rc = osp_init_last_seq(env, osp);
378 CERROR("%s: Can not get ids %d from old objid!\n",
379 osp->opd_obd->obd_name, rc);
383 if (fid_oid(&osp->opd_last_used_fid) != 0 &&
384 fid_seq(&osp->opd_last_used_fid) == 0) {
385 /* Just upgrade from the old version,
386 * set the seq to be IDIF */
387 osp->opd_last_used_fid.f_seq =
388 fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
390 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
391 &osp->opd_last_used_fid.f_seq,
393 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
394 &osi->osi_lb, osi->osi_off);
396 CERROR("%s : Can not write seq file: rc = %d\n",
397 osp->opd_obd->obd_name, rc);
402 if (!fid_is_zero(&osp->opd_last_used_fid) &&
403 !fid_is_sane(&osp->opd_last_used_fid)) {
404 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
405 PFID(&osp->opd_last_used_fid));
406 GOTO(out, rc = -EINVAL);
409 CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
410 osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
413 if (osp->opd_last_used_oid_file != NULL) {
414 lu_object_put(env, &osp->opd_last_used_oid_file->do_lu);
415 osp->opd_last_used_oid_file = NULL;
417 if (osp->opd_last_used_seq_file != NULL) {
418 lu_object_put(env, &osp->opd_last_used_seq_file->do_lu);
419 osp->opd_last_used_seq_file = NULL;
427 * Release the last sequence and OID file objects in OSP device.
429 * \param[in] env execution environment
430 * \param[in] osp OSP device
432 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *osp)
434 /* release last_used file */
435 if (osp->opd_last_used_oid_file != NULL) {
436 lu_object_put(env, &osp->opd_last_used_oid_file->do_lu);
437 osp->opd_last_used_oid_file = NULL;
440 if (osp->opd_last_used_seq_file != NULL) {
441 lu_object_put(env, &osp->opd_last_used_seq_file->do_lu);
442 osp->opd_last_used_seq_file = NULL;
447 * Disconnects the connection between OSP and its correspondent MDT or OST, and
448 * the import will be marked as inactive. It will only be called during OSP
451 * \param[in] d OSP device being disconnected
453 * \retval 0 0 if disconnection succeed
454 * \retval negative negative errno if disconnection failed
456 static int osp_disconnect(struct osp_device *d)
458 struct obd_device *obd = d->opd_obd;
459 struct obd_import *imp;
462 imp = obd->u.cli.cl_import;
464 /* Mark import deactivated now, so we don't try to reconnect if any
465 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
466 * fully deactivate the import, or that would drop all requests. */
467 LASSERT(imp != NULL);
468 spin_lock(&imp->imp_lock);
469 imp->imp_deactive = 1;
470 spin_unlock(&imp->imp_lock);
472 ptlrpc_deactivate_import(imp);
474 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
475 * delete it regardless. (It's safe to delete an import that was
477 (void)ptlrpc_pinger_del_import(imp);
479 rc = ptlrpc_disconnect_import(imp, 0);
481 CERROR("%s: can't disconnect: rc = %d\n", obd->obd_name, rc);
483 ptlrpc_invalidate_import(imp);
489 * Initialize the osp_update structure in OSP device
491 * Allocate osp update structure and start update thread.
493 * \param[in] osp OSP device
495 * \retval 0 if initialization succeeds.
496 * \retval negative errno if initialization fails.
498 static int osp_update_init(struct osp_device *osp)
500 struct l_wait_info lwi = { 0 };
501 struct task_struct *task;
505 LASSERT(osp->opd_connect_mdt);
507 OBD_ALLOC_PTR(osp->opd_update);
508 if (osp->opd_update == NULL)
511 init_waitqueue_head(&osp->opd_update_thread.t_ctl_waitq);
512 init_waitqueue_head(&osp->opd_update->ou_waitq);
513 spin_lock_init(&osp->opd_update->ou_lock);
514 INIT_LIST_HEAD(&osp->opd_update->ou_list);
515 osp->opd_update->ou_rpc_version = 1;
516 osp->opd_update->ou_version = 1;
518 /* start thread handling sending updates to the remote MDT */
519 task = kthread_run(osp_send_update_thread, osp,
520 "osp_up%u-%u", osp->opd_index, osp->opd_group);
522 int rc = PTR_ERR(task);
524 OBD_FREE_PTR(osp->opd_update);
525 osp->opd_update = NULL;
526 CERROR("%s: can't start precreate thread: rc = %d\n",
527 osp->opd_obd->obd_name, rc);
531 l_wait_event(osp->opd_update_thread.t_ctl_waitq,
532 osp_send_update_thread_running(osp) ||
533 osp_send_update_thread_stopped(osp), &lwi);
539 * Finialize osp_update structure in OSP device
541 * Stop the OSP update sending thread, then delete the left
542 * osp thandle in the sending list.
544 * \param [in] osp OSP device.
546 static void osp_update_fini(const struct lu_env *env, struct osp_device *osp)
548 struct osp_update_request *our;
549 struct osp_update_request *tmp;
550 struct osp_updates *ou = osp->opd_update;
555 osp->opd_update_thread.t_flags = SVC_STOPPING;
556 wake_up(&ou->ou_waitq);
558 wait_event(osp->opd_update_thread.t_ctl_waitq,
559 osp->opd_update_thread.t_flags & SVC_STOPPED);
561 /* Remove the left osp thandle from the list */
562 spin_lock(&ou->ou_lock);
563 list_for_each_entry_safe(our, tmp, &ou->ou_list,
565 list_del_init(&our->our_list);
566 LASSERT(our->our_th != NULL);
567 osp_trans_callback(env, our->our_th, -EIO);
568 /* our will be destroyed in osp_thandle_put() */
569 osp_thandle_put(env, our->our_th);
571 spin_unlock(&ou->ou_lock);
574 osp->opd_update = NULL;
578 * Cleanup OSP, which includes disconnect import, cleanup unlink log, stop
579 * precreate threads etc.
581 * \param[in] env execution environment.
582 * \param[in] d OSP device being disconnected.
584 * \retval 0 0 if cleanup succeed
585 * \retval negative negative errno if cleanup failed
587 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
594 rc = osp_disconnect(d);
596 if (!d->opd_connect_mdt) {
597 /* stop sync thread */
600 /* stop precreate thread */
601 osp_precreate_fini(d);
603 /* release last_used file */
604 osp_last_used_fini(env, d);
607 obd_fid_fini(d->opd_obd);
613 * Implementation of osp_lu_ops::ldo_process_config
615 * This function processes config log records in OSP layer. It is usually
616 * called from the top layer of MDT stack, and goes through the stack by calling
617 * ldo_process_config of next layer.
619 * \param[in] env execution environment
620 * \param[in] dev lu_device of OSP
621 * \param[in] lcfg config log
623 * \retval 0 0 if the config log record is executed correctly.
624 * \retval negative negative errno if the record execution fails.
626 static int osp_process_config(const struct lu_env *env,
627 struct lu_device *dev, struct lustre_cfg *lcfg)
629 struct osp_device *d = lu2osp_dev(dev);
630 struct obd_device *obd = d->opd_obd;
635 switch (lcfg->lcfg_command) {
636 case LCFG_PRE_CLEANUP:
637 rc = osp_disconnect(d);
638 osp_update_fini(env, d);
639 if (obd->obd_namespace != NULL)
640 ldlm_namespace_free_prior(obd->obd_namespace, NULL, 1);
643 lu_dev_del_linkage(dev->ld_site, dev);
644 rc = osp_shutdown(env, d);
648 rc = class_process_proc_param(d->opd_connect_mdt ?
649 PARAM_OSP : PARAM_OSC,
650 obd->obd_vars, lcfg, obd);
654 /* class_process_proc_param() haven't found matching
655 * parameter and returned ENOSYS so that layer(s)
656 * below could use that. But OSP is the bottom, so
658 CERROR("%s: unknown param %s\n",
659 (char *)lustre_cfg_string(lcfg, 0),
660 (char *)lustre_cfg_string(lcfg, 1));
665 CERROR("%s: unknown command %u\n",
666 (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
675 * Implementation of osp_lu_ops::ldo_recovery_complete
677 * This function is called after recovery is finished, and OSP layer
678 * will wake up precreate thread here.
680 * \param[in] env execution environment
681 * \param[in] dev lu_device of OSP
683 * \retval 0 0 unconditionally
685 static int osp_recovery_complete(const struct lu_env *env,
686 struct lu_device *dev)
688 struct osp_device *osp = lu2osp_dev(dev);
691 osp->opd_recovery_completed = 1;
693 if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
694 wake_up(&osp->opd_pre_waitq);
699 const struct lu_device_operations osp_lu_ops = {
700 .ldo_object_alloc = osp_object_alloc,
701 .ldo_process_config = osp_process_config,
702 .ldo_recovery_complete = osp_recovery_complete,
706 * Implementation of dt_device_operations::dt_statfs
708 * This function provides statfs status (for precreation) from
709 * corresponding OST. Note: this function only retrieves the status
710 * from the OSP device, and the real statfs RPC happens inside
711 * precreate thread (\see osp_statfs_update). Note: OSP for MDT does
712 * not need to retrieve statfs data for now.
714 * \param[in] env execution environment.
715 * \param[in] dev dt_device of OSP.
716 * \param[out] sfs holds the retrieved statfs data.
718 * \retval 0 0 statfs data was retrieved successfully or
719 * retrieval was not needed
720 * \retval negative negative errno if get statfs failed.
722 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
723 struct obd_statfs *sfs)
725 struct osp_device *d = dt2osp_dev(dev);
726 struct obd_import *imp = d->opd_obd->u.cli.cl_import;
730 if (imp->imp_state == LUSTRE_IMP_CLOSED)
733 if (unlikely(d->opd_imp_active == 0))
736 if (d->opd_pre == NULL)
739 /* return recently updated data */
740 *sfs = d->opd_statfs;
743 * layer above osp (usually lod) can use ffree to estimate
744 * how many objects are available for immediate creation
746 spin_lock(&d->opd_pre_lock);
747 LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
748 fid_seq(&d->opd_pre_used_fid),
749 "last_created "DFID", next_fid "DFID"\n",
750 PFID(&d->opd_pre_last_created_fid),
751 PFID(&d->opd_pre_used_fid));
752 sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
753 fid_oid(&d->opd_pre_used_fid);
754 sfs->os_fprecreated -= d->opd_pre_reserved;
755 spin_unlock(&d->opd_pre_lock);
757 LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2);
759 CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
760 LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
761 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
762 sfs->os_files, sfs->os_ffree);
766 static int osp_sync_timeout(void *data)
772 * Implementation of dt_device_operations::dt_sync
774 * This function synchronizes the OSP cache to the remote target. It wakes
775 * up unlink log threads and sends out unlink records to the remote OST.
777 * \param[in] env execution environment
778 * \param[in] dev dt_device of OSP
780 * \retval 0 0 if synchronization succeeds
781 * \retval negative negative errno if synchronization fails
783 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
785 struct osp_device *d = dt2osp_dev(dev);
787 struct l_wait_info lwi = { 0 };
788 unsigned long id, old;
790 unsigned long start = cfs_time_current();
793 /* No Sync between MDTs yet. */
794 if (d->opd_connect_mdt)
797 if (unlikely(d->opd_imp_active == 0))
800 id = d->opd_syn_last_used_id;
801 down_write(&d->opd_async_updates_rwsem);
803 CDEBUG(D_OTHER, "%s: async updates %d\n", d->opd_obd->obd_name,
804 atomic_read(&d->opd_async_updates_count));
806 /* make sure the connection is fine */
807 expire = cfs_time_shift(obd_timeout);
808 lwi = LWI_TIMEOUT(expire - cfs_time_current(), osp_sync_timeout, d);
809 rc = l_wait_event(d->opd_syn_barrier_waitq,
810 atomic_read(&d->opd_async_updates_count) == 0,
812 up_write(&d->opd_async_updates_rwsem);
816 CDEBUG(D_CACHE, "%s: id: used %lu, processed "LPU64"\n",
817 d->opd_obd->obd_name, id, d->opd_syn_last_processed_id);
819 /* wait till all-in-line are processed */
820 while (d->opd_syn_last_processed_id < id) {
822 old = d->opd_syn_last_processed_id;
824 /* make sure the connection is fine */
825 expire = cfs_time_shift(obd_timeout);
826 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
827 osp_sync_timeout, d);
828 l_wait_event(d->opd_syn_barrier_waitq,
829 d->opd_syn_last_processed_id >= id,
832 if (d->opd_syn_last_processed_id >= id)
835 if (d->opd_syn_last_processed_id != old) {
836 /* some progress have been made,
841 /* no changes and expired, something is wrong */
842 GOTO(out, rc = -ETIMEDOUT);
845 /* block new processing (barrier>0 - few callers are possible */
846 atomic_inc(&d->opd_syn_barrier);
848 CDEBUG(D_CACHE, "%s: %u in flight\n", d->opd_obd->obd_name,
849 atomic_read(&d->opd_syn_rpc_in_flight));
851 /* wait till all-in-flight are replied, so executed by the target */
852 /* XXX: this is used by LFSCK at the moment, which doesn't require
853 * all the changes to be committed, but in general it'd be
854 * better to wait till commit */
855 while (atomic_read(&d->opd_syn_rpc_in_flight) > 0) {
857 old = atomic_read(&d->opd_syn_rpc_in_flight);
859 expire = cfs_time_shift(obd_timeout);
860 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
861 osp_sync_timeout, d);
862 l_wait_event(d->opd_syn_barrier_waitq,
863 atomic_read(&d->opd_syn_rpc_in_flight) == 0,
866 if (atomic_read(&d->opd_syn_rpc_in_flight) == 0)
869 if (atomic_read(&d->opd_syn_rpc_in_flight) != old) {
870 /* some progress have been made */
874 /* no changes and expired, something is wrong */
875 GOTO(out, rc = -ETIMEDOUT);
879 /* resume normal processing (barrier=0) */
880 atomic_dec(&d->opd_syn_barrier);
881 __osp_sync_check_for_work(d);
883 CDEBUG(D_CACHE, "%s: done in %lu: rc = %d\n", d->opd_obd->obd_name,
884 cfs_time_current() - start, rc);
889 const struct dt_device_operations osp_dt_ops = {
890 .dt_statfs = osp_statfs,
892 .dt_trans_create = osp_trans_create,
893 .dt_trans_start = osp_trans_start,
894 .dt_trans_stop = osp_trans_stop,
895 .dt_trans_cb_add = osp_trans_cb_add,
899 * Connect OSP to local OSD.
901 * Locate the local OSD referenced by \a nextdev and connect to it. Sometimes,
902 * OSP needs to access the local OSD to store some information. For example,
903 * during precreate, it needs to update last used OID and sequence file
904 * (LAST_SEQ) in local OSD.
906 * \param[in] env execution environment
907 * \param[in] osp OSP device
908 * \param[in] nextdev the name of local OSD
910 * \retval 0 0 connection succeeded
911 * \retval negative negative errno connection failed
913 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *osp,
916 struct obd_connect_data *data = NULL;
917 struct obd_device *obd;
922 LASSERT(osp->opd_storage_exp == NULL);
928 obd = class_name2obd(nextdev);
930 CERROR("%s: can't locate next device: %s\n",
931 osp->opd_obd->obd_name, nextdev);
932 GOTO(out, rc = -ENOTCONN);
935 rc = obd_connect(env, &osp->opd_storage_exp, obd, &obd->obd_uuid, data,
938 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
939 osp->opd_obd->obd_name, nextdev, rc);
943 osp->opd_dt_dev.dd_lu_dev.ld_site =
944 osp->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
945 LASSERT(osp->opd_dt_dev.dd_lu_dev.ld_site);
946 osp->opd_storage = lu2dt_dev(osp->opd_storage_exp->exp_obd->obd_lu_dev);
954 * Determine if the lock needs to be cancelled
956 * Determine if the unused lock should be cancelled before replay, see
957 * (ldlm_cancel_no_wait_policy()). Currently, only inode bits lock exists
960 * \param[in] lock lock to be checked.
962 * \retval 1 if the lock needs to be cancelled before replay.
963 * \retval 0 if the lock does not need to be cancelled before
966 static int osp_cancel_weight(struct ldlm_lock *lock)
968 if (lock->l_resource->lr_type != LDLM_IBITS)
975 * Initialize OSP device according to the parameters in the configuration
978 * Reconstruct the local device name from the configuration profile, and
979 * initialize necessary threads and structures according to the OSP type
982 * Since there is no record in the MDT configuration for the local disk
983 * device, we have to extract this from elsewhere in the profile.
984 * The only information we get at setup is from the OSC records:
985 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
987 * Note: configs generated by Lustre 1.8 are missing the -MDTxxxx part,
988 * so, we need to reconstruct the name of the underlying OSD from this:
989 * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".
991 * \param[in] env execution environment
992 * \param[in] osp OSP device
993 * \param[in] ldt lu device type of OSP
994 * \param[in] cfg configuration log
996 * \retval 0 0 if OSP initialization succeeded.
997 * \retval negative negative errno if OSP initialization failed.
999 static int osp_init0(const struct lu_env *env, struct osp_device *osp,
1000 struct lu_device_type *ldt, struct lustre_cfg *cfg)
1002 struct obd_device *obd;
1003 struct obd_import *imp;
1005 char *src, *tgt, *mdt, *osdname = NULL;
1011 mutex_init(&osp->opd_async_requests_mutex);
1012 INIT_LIST_HEAD(&osp->opd_async_updates);
1013 init_rwsem(&osp->opd_async_updates_rwsem);
1014 atomic_set(&osp->opd_async_updates_count, 0);
1016 obd = class_name2obd(lustre_cfg_string(cfg, 0));
1018 CERROR("Cannot find obd with name %s\n",
1019 lustre_cfg_string(cfg, 0));
1024 src = lustre_cfg_string(cfg, 0);
1028 tgt = strrchr(src, '-');
1030 CERROR("%s: invalid target name %s: rc = %d\n",
1031 osp->opd_obd->obd_name, lustre_cfg_string(cfg, 0),
1036 if (strncmp(tgt, "-osc", 4) == 0) {
1037 /* Old OSC name fsname-OSTXXXX-osc */
1038 for (tgt--; tgt > src && *tgt != '-'; tgt--)
1041 CERROR("%s: invalid target name %s: rc = %d\n",
1042 osp->opd_obd->obd_name,
1043 lustre_cfg_string(cfg, 0), -EINVAL);
1047 if (strncmp(tgt, "-OST", 4) != 0) {
1048 CERROR("%s: invalid target name %s: rc = %d\n",
1049 osp->opd_obd->obd_name,
1050 lustre_cfg_string(cfg, 0), -EINVAL);
1054 idx = simple_strtol(tgt + 4, &mdt, 16);
1055 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1056 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1057 osp->opd_obd->obd_name, src, -EINVAL);
1060 osp->opd_index = idx;
1064 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
1065 if (strncmp(tgt, "-MDT", 4) != 0 &&
1066 strncmp(tgt, "-OST", 4) != 0) {
1067 CERROR("%s: invalid target name %s: rc = %d\n",
1068 osp->opd_obd->obd_name,
1069 lustre_cfg_string(cfg, 0), -EINVAL);
1073 idx = simple_strtol(tgt + 4, &mdt, 16);
1074 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
1075 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1076 osp->opd_obd->obd_name, src, -EINVAL);
1080 /* Get MDT index from the name and set it to opd_group,
1081 * which will be used by OSP to connect with OST */
1082 osp->opd_group = idx;
1083 if (tgt - src <= 12) {
1084 CERROR("%s: invalid mdt index from %s: rc =%d\n",
1085 osp->opd_obd->obd_name,
1086 lustre_cfg_string(cfg, 0), -EINVAL);
1090 if (strncmp(tgt - 12, "-MDT", 4) == 0)
1091 osp->opd_connect_mdt = 1;
1093 idx = simple_strtol(tgt - 8, &mdt, 16);
1094 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1095 CERROR("%s: invalid OST index in '%s': rc =%d\n",
1096 osp->opd_obd->obd_name, src, -EINVAL);
1100 osp->opd_index = idx;
1101 idx = tgt - src - 12;
1103 /* check the fsname length, and after this everything else will fit */
1104 if (idx > MTI_NAME_MAXLEN) {
1105 CERROR("%s: fsname too long in '%s': rc = %d\n",
1106 osp->opd_obd->obd_name, src, -EINVAL);
1110 OBD_ALLOC(osdname, MAX_OBD_NAME);
1111 if (osdname == NULL)
1114 memcpy(osdname, src, idx); /* copy just the fsname part */
1115 osdname[idx] = '\0';
1117 mdt = strstr(mdt, "-MDT");
1118 if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
1119 strcat(osdname, "-MDT0000");
1121 strcat(osdname, mdt);
1122 strcat(osdname, "-osd");
1123 CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
1125 if (osp->opd_connect_mdt) {
1126 struct client_obd *cli = &osp->opd_obd->u.cli;
1128 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
1129 if (!cli->cl_rpc_lock)
1130 GOTO(out_fini, rc = -ENOMEM);
1131 osp_init_rpc_lock(cli->cl_rpc_lock);
1134 osp->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
1135 osp->opd_dt_dev.dd_ops = &osp_dt_ops;
1137 obd->obd_lu_dev = &osp->opd_dt_dev.dd_lu_dev;
1139 rc = osp_connect_to_osd(env, osp, osdname);
1143 rc = ptlrpcd_addref();
1145 GOTO(out_disconnect, rc);
1147 rc = client_obd_setup(obd, cfg);
1149 CERROR("%s: can't setup obd: rc = %d\n", osp->opd_obd->obd_name,
1154 osp_lprocfs_init(osp);
1156 rc = obd_fid_init(osp->opd_obd, NULL, osp->opd_connect_mdt ?
1157 LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
1159 CERROR("%s: fid init error: rc = %d\n",
1160 osp->opd_obd->obd_name, rc);
1164 if (!osp->opd_connect_mdt) {
1165 /* Initialize last id from the storage - will be
1166 * used in orphan cleanup. */
1167 rc = osp_last_used_init(env, osp);
1172 /* Initialize precreation thread, it handles new
1173 * connections as well. */
1174 rc = osp_init_precreate(osp);
1176 GOTO(out_last_used, rc);
1179 * Initialize synhronization mechanism taking
1180 * care of propogating changes to OST in near
1181 * transactional manner.
1183 rc = osp_sync_init(env, osp);
1185 GOTO(out_precreat, rc);
1187 rc = osp_update_init(osp);
1192 ns_register_cancel(obd->obd_namespace, osp_cancel_weight);
1195 * Initiate connect to OST
1197 ll_generate_random_uuid(uuid);
1198 class_uuid_unparse(uuid, &osp->opd_cluuid);
1200 imp = obd->u.cli.cl_import;
1202 rc = ptlrpc_init_import(imp);
1206 OBD_FREE(osdname, MAX_OBD_NAME);
1210 if (!osp->opd_connect_mdt)
1211 /* stop sync thread */
1214 /* stop precreate thread */
1215 if (!osp->opd_connect_mdt)
1216 osp_precreate_fini(osp);
1218 osp_update_fini(env, osp);
1220 if (!osp->opd_connect_mdt)
1221 osp_last_used_fini(env, osp);
1223 obd_fid_fini(osp->opd_obd);
1225 ptlrpc_lprocfs_unregister_obd(obd);
1226 lprocfs_obd_cleanup(obd);
1227 if (osp->opd_symlink)
1228 lprocfs_remove(&osp->opd_symlink);
1229 client_obd_cleanup(obd);
1233 if (osp->opd_connect_mdt) {
1234 struct client_obd *cli = &osp->opd_obd->u.cli;
1235 if (cli->cl_rpc_lock != NULL) {
1236 OBD_FREE_PTR(cli->cl_rpc_lock);
1237 cli->cl_rpc_lock = NULL;
1240 obd_disconnect(osp->opd_storage_exp);
1243 OBD_FREE(osdname, MAX_OBD_NAME);
1248 * Implementation of lu_device_type_operations::ldto_device_free
1250 * Free the OSP device in memory. No return value is needed for now,
1251 * so always return NULL to comply with the interface.
1253 * \param[in] env execution environment
1254 * \param[in] lu lu_device of OSP
1256 * \retval NULL NULL unconditionally
1258 static struct lu_device *osp_device_free(const struct lu_env *env,
1259 struct lu_device *lu)
1261 struct osp_device *osp = lu2osp_dev(lu);
1263 if (atomic_read(&lu->ld_ref) && lu->ld_site) {
1264 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1265 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
1267 dt_device_fini(&osp->opd_dt_dev);
1274 * Implementation of lu_device_type_operations::ldto_device_alloc
1276 * This function allocates and initializes OSP device in memory according to
1279 * \param[in] env execution environment
1280 * \param[in] type device type of OSP
1281 * \param[in] lcfg config log
1283 * \retval pointer the pointer of allocated OSP if succeed.
1284 * \retval ERR_PTR(errno) ERR_PTR(errno) if failed.
1286 static struct lu_device *osp_device_alloc(const struct lu_env *env,
1287 struct lu_device_type *type,
1288 struct lustre_cfg *lcfg)
1290 struct osp_device *osp;
1291 struct lu_device *ld;
1295 ld = ERR_PTR(-ENOMEM);
1299 ld = osp2lu_dev(osp);
1300 dt_device_init(&osp->opd_dt_dev, type);
1301 rc = osp_init0(env, osp, type, lcfg);
1303 osp_device_free(env, ld);
1311 * Implementation of lu_device_type_operations::ldto_device_fini
1313 * This function cleans up the OSP device, i.e. release and free those
1314 * attached items in osp_device.
1316 * \param[in] env execution environment
1317 * \param[in] ld lu_device of OSP
1319 * \retval NULL NULL if cleanup succeeded.
1320 * \retval ERR_PTR(errno) ERR_PTR(errno) if cleanup failed.
1322 static struct lu_device *osp_device_fini(const struct lu_env *env,
1323 struct lu_device *ld)
1325 struct osp_device *osp = lu2osp_dev(ld);
1330 if (osp->opd_async_requests != NULL) {
1331 osp_update_request_destroy(env, osp->opd_async_requests);
1332 osp->opd_async_requests = NULL;
1335 if (osp->opd_storage_exp)
1336 obd_disconnect(osp->opd_storage_exp);
1338 if (osp->opd_symlink)
1339 lprocfs_remove(&osp->opd_symlink);
1341 LASSERT(osp->opd_obd);
1342 ptlrpc_lprocfs_unregister_obd(osp->opd_obd);
1343 lprocfs_obd_cleanup(osp->opd_obd);
1345 if (osp->opd_connect_mdt) {
1346 struct client_obd *cli = &osp->opd_obd->u.cli;
1347 if (cli->cl_rpc_lock != NULL) {
1348 OBD_FREE_PTR(cli->cl_rpc_lock);
1349 cli->cl_rpc_lock = NULL;
1353 rc = client_obd_cleanup(osp->opd_obd);
1356 RETURN(ERR_PTR(rc));
1365 * Implementation of obd_ops::o_reconnect
1367 * This function is empty and does not need to do anything for now.
1369 static int osp_reconnect(const struct lu_env *env,
1370 struct obd_export *exp, struct obd_device *obd,
1371 struct obd_uuid *cluuid,
1372 struct obd_connect_data *data,
1379 * Implementation of obd_ops::o_connect
1381 * Connect OSP to the remote target (MDT or OST). Allocate the
1382 * export and return it to the LOD, which calls this function
1383 * for each OSP to connect it to the remote target. This function
1384 * is currently only called once per OSP.
1386 * \param[in] env execution environment
1387 * \param[out] exp export connected to OSP
1388 * \param[in] obd OSP device
1389 * \param[in] cluuid OSP device client uuid
1390 * \param[in] data connect_data to be used to connect to the remote
1392 * \param[in] localdata necessary for the API interface, but not used in
1395 * \retval 0 0 if the connection succeeded.
1396 * \retval negative negative errno if the connection failed.
1398 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
1399 struct obd_device *obd, struct obd_uuid *cluuid,
1400 struct obd_connect_data *data, void *localdata)
1402 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1403 struct obd_connect_data *ocd;
1404 struct obd_import *imp;
1405 struct lustre_handle conn;
1410 CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
1412 rc = class_connect(&conn, obd, cluuid);
1416 *exp = class_conn2export(&conn);
1417 /* Why should there ever be more than 1 connect? */
1418 osp->opd_connects++;
1419 LASSERT(osp->opd_connects == 1);
1421 osp->opd_exp = *exp;
1423 imp = osp->opd_obd->u.cli.cl_import;
1424 imp->imp_dlm_handle = conn;
1426 LASSERT(data != NULL);
1427 LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
1428 ocd = &imp->imp_connect_data;
1431 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
1432 imp->imp_connect_flags2_orig = ocd->ocd_connect_flags2;
1434 ocd->ocd_version = LUSTRE_VERSION_CODE;
1435 ocd->ocd_index = data->ocd_index;
1437 rc = ptlrpc_connect_import(imp);
1439 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
1442 osp->opd_obd->u.cli.cl_seq->lcs_exp =
1443 class_export_get(osp->opd_exp);
1446 ptlrpc_pinger_add_import(imp);
1452 * Implementation of obd_ops::o_disconnect
1454 * Disconnect the export for the OSP. This is called by LOD to release the
1455 * OSP during cleanup (\see lod_del_device()). The OSP will be released after
1456 * the export is released.
1458 * \param[in] exp export to be disconnected.
1460 * \retval 0 0 if disconnection succeed
1461 * \retval negative negative errno if disconnection failed
1463 static int osp_obd_disconnect(struct obd_export *exp)
1465 struct obd_device *obd = exp->exp_obd;
1466 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1470 /* Only disconnect the underlying layers on the final disconnect. */
1471 LASSERT(osp->opd_connects == 1);
1472 osp->opd_connects--;
1474 rc = class_disconnect(exp);
1476 CERROR("%s: class disconnect error: rc = %d\n",
1481 /* destroy the device */
1482 class_manual_cleanup(obd);
1488 * Implementation of obd_ops::o_statfs
1490 * Send a RPC to the remote target to get statfs status. This is only used
1491 * in lprocfs helpers by obd_statfs.
1493 * \param[in] env execution environment
1494 * \param[in] exp connection state from this OSP to the parent (LOD)
1496 * \param[out] osfs hold the statfs result
1497 * \param[in] unused Not used in this function for now
1498 * \param[in] flags flags to indicate how OSP will issue the RPC
1500 * \retval 0 0 if statfs succeeded.
1501 * \retval negative negative errno if statfs failed.
1503 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
1504 struct obd_statfs *osfs, __u64 unused, __u32 flags)
1506 struct obd_statfs *msfs;
1507 struct ptlrpc_request *req;
1508 struct obd_import *imp = NULL;
1513 /* Since the request might also come from lprocfs, so we need
1514 * sync this with client_disconnect_export Bug15684 */
1515 down_read(&exp->exp_obd->u.cli.cl_sem);
1516 if (exp->exp_obd->u.cli.cl_import)
1517 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
1518 up_read(&exp->exp_obd->u.cli.cl_sem);
1522 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1524 class_import_put(imp);
1529 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1531 ptlrpc_request_free(req);
1534 ptlrpc_request_set_replen(req);
1535 req->rq_request_portal = OST_CREATE_PORTAL;
1536 ptlrpc_at_set_req_timeout(req);
1538 if (flags & OBD_STATFS_NODELAY) {
1539 /* procfs requests not want stat in wait for avoid deadlock */
1540 req->rq_no_resend = 1;
1541 req->rq_no_delay = 1;
1544 rc = ptlrpc_queue_wait(req);
1548 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1550 GOTO(out, rc = -EPROTO);
1556 ptlrpc_req_finished(req);
1561 * Implementation of obd_ops::o_import_event
1563 * This function is called when some related import event happens. It will
1564 * mark the necessary flags according to the event and notify the necessary
1565 * threads (mainly precreate thread).
1567 * \param[in] obd OSP OBD device
1568 * \param[in] imp import attached from OSP to remote (OST/MDT) service
1569 * \param[in] event event related to remote service (IMP_EVENT_*)
1571 * \retval 0 0 if the event handling succeeded.
1572 * \retval negative negative errno if the event handling failed.
1574 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1575 enum obd_import_event event)
1577 struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1581 case IMP_EVENT_DISCON:
1582 d->opd_got_disconnected = 1;
1583 d->opd_imp_connected = 0;
1584 if (d->opd_connect_mdt)
1587 if (d->opd_pre != NULL) {
1588 osp_pre_update_status(d, -ENODEV);
1589 wake_up(&d->opd_pre_waitq);
1592 CDEBUG(D_HA, "got disconnected\n");
1594 case IMP_EVENT_INACTIVE:
1595 d->opd_imp_active = 0;
1596 d->opd_imp_connected = 0;
1597 d->opd_obd->obd_inactive = 1;
1598 if (d->opd_connect_mdt)
1600 if (d->opd_pre != NULL) {
1601 /* Import is invalid, we can`t get stripes so
1603 rc = imp->imp_deactive ? -ESHUTDOWN : -ENODEV;
1604 osp_pre_update_status(d, rc);
1605 wake_up(&d->opd_pre_waitq);
1608 CDEBUG(D_HA, "got inactive\n");
1610 case IMP_EVENT_ACTIVE:
1611 d->opd_imp_active = 1;
1613 if (d->opd_got_disconnected)
1614 d->opd_new_connection = 1;
1615 d->opd_imp_connected = 1;
1616 d->opd_imp_seen_connected = 1;
1617 d->opd_obd->obd_inactive = 0;
1618 if (d->opd_connect_mdt)
1621 if (d->opd_pre != NULL)
1622 wake_up(&d->opd_pre_waitq);
1624 __osp_sync_check_for_work(d);
1625 CDEBUG(D_HA, "got connected\n");
1627 case IMP_EVENT_INVALIDATE:
1628 if (d->opd_connect_mdt)
1629 osp_invalidate_request(d);
1631 if (obd->obd_namespace == NULL)
1633 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1636 case IMP_EVENT_DEACTIVATE:
1637 case IMP_EVENT_ACTIVATE:
1640 CERROR("%s: unsupported import event: %#x\n",
1641 obd->obd_name, event);
1647 * Implementation of obd_ops: o_iocontrol
1649 * This function is the ioctl handler for OSP. Note: lctl will access the OSP
1650 * directly by ioctl, instead of through the MDS stack.
1652 * param[in] cmd ioctl command.
1653 * param[in] exp export of this OSP.
1654 * param[in] len data length of \a karg.
1655 * param[in] karg input argument which is packed as
1657 * param[out] uarg pointer to userspace buffer (must access by
1660 * \retval 0 0 if the ioctl handling succeeded.
1661 * \retval negative negative errno if the ioctl handling failed.
1663 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1664 void *karg, void __user *uarg)
1666 struct obd_device *obd = exp->exp_obd;
1667 struct osp_device *d;
1668 struct obd_ioctl_data *data = karg;
1673 LASSERT(obd->obd_lu_dev);
1674 d = lu2osp_dev(obd->obd_lu_dev);
1675 LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1677 if (!try_module_get(THIS_MODULE)) {
1678 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1679 module_name(THIS_MODULE));
1684 case OBD_IOC_CLIENT_RECOVER:
1685 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1686 data->ioc_inlbuf1, 0);
1690 case IOC_OSC_SET_ACTIVE:
1691 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1694 case OBD_IOC_PING_TARGET:
1695 rc = ptlrpc_obd_ping(obd);
1698 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1699 cmd, current_comm());
1702 module_put(THIS_MODULE);
1708 * Implementation of obd_ops::o_get_info
1710 * Retrieve information by key. Retrieval starts from the top layer
1711 * (MDT) of the MDS stack and traverses the stack by calling the
1712 * obd_get_info() method of the next sub-layer.
1714 * \param[in] env execution environment
1715 * \param[in] exp export of this OSP
1716 * \param[in] keylen length of \a key
1717 * \param[in] key the key
1718 * \param[out] vallen length of \a val
1719 * \param[out] val holds the value returned by the key
1721 * \retval 0 0 if getting information succeeded.
1722 * \retval negative negative errno if getting information failed.
1724 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1725 __u32 keylen, void *key, __u32 *vallen, void *val)
1729 if (KEY_IS(KEY_OSP_CONNECTED)) {
1730 struct obd_device *obd = exp->exp_obd;
1731 struct osp_device *osp;
1733 if (!obd->obd_set_up || obd->obd_stopping)
1736 osp = lu2osp_dev(obd->obd_lu_dev);
1739 * 1.8/2.0 behaviour is that OST being connected once at least
1740 * is considered "healthy". and one "healthy" OST is enough to
1741 * allow lustre clients to connect to MDS
1743 RETURN(!osp->opd_imp_seen_connected);
1749 static int osp_obd_set_info_async(const struct lu_env *env,
1750 struct obd_export *exp,
1751 u32 keylen, void *key,
1752 u32 vallen, void *val,
1753 struct ptlrpc_request_set *set)
1755 struct obd_device *obd = exp->exp_obd;
1756 struct obd_import *imp = obd->u.cli.cl_import;
1757 struct osp_device *osp;
1758 struct ptlrpc_request *req;
1762 if (KEY_IS(KEY_SPTLRPC_CONF)) {
1763 sptlrpc_conf_client_adapt(exp->exp_obd);
1767 LASSERT(set != NULL);
1768 if (!obd->obd_set_up || obd->obd_stopping)
1770 osp = lu2osp_dev(obd->obd_lu_dev);
1772 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
1776 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
1777 RCL_CLIENT, keylen);
1778 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
1779 RCL_CLIENT, vallen);
1780 if (osp->opd_connect_mdt)
1781 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SET_INFO);
1783 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
1785 ptlrpc_request_free(req);
1789 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1790 memcpy(tmp, key, keylen);
1791 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1792 memcpy(tmp, val, vallen);
1794 ptlrpc_request_set_replen(req);
1795 ptlrpc_set_add_req(set, req);
1796 ptlrpc_check_set(NULL, set);
1802 * Implementation of obd_ops: o_fid_alloc
1804 * Allocate a FID. There are two cases in which OSP performs
1807 * 1. FID precreation for data objects, which is done in
1808 * osp_precreate_fids() instead of this function.
1809 * 2. FID allocation for each sub-stripe of a striped directory.
1810 * Similar to other FID clients, OSP requests the sequence
1811 * from its corresponding remote MDT, which in turn requests
1812 * sequences from the sequence controller (MDT0).
1814 * \param[in] env execution environment
1815 * \param[in] exp export of the OSP
1816 * \param[out] fid FID being allocated
1817 * \param[in] unused necessary for the interface but unused.
1819 * \retval 0 0 FID allocated successfully.
1820 * \retval 1 1 FID allocated successfully and new sequence
1821 * requested from seq meta server
1822 * \retval negative negative errno if FID allocation failed.
1824 static int osp_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1825 struct lu_fid *fid, struct md_op_data *unused)
1827 struct client_obd *cli = &exp->exp_obd->u.cli;
1828 struct osp_device *osp = lu2osp_dev(exp->exp_obd->obd_lu_dev);
1829 struct lu_client_seq *seq = cli->cl_seq;
1832 LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
1833 /* Sigh, fid client is not ready yet */
1834 LASSERT(osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1836 RETURN(seq_client_alloc_fid(env, seq, fid));
1839 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1840 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1841 static void osp_key_exit(const struct lu_context *ctx,
1842 struct lu_context_key *key, void *data)
1844 struct osp_thread_info *info = data;
1846 info->osi_attr.la_valid = 0;
1849 struct lu_context_key osp_thread_key = {
1850 .lct_tags = LCT_MD_THREAD,
1851 .lct_init = osp_key_init,
1852 .lct_fini = osp_key_fini,
1853 .lct_exit = osp_key_exit
1856 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1857 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1859 struct lu_context_key osp_txn_key = {
1860 .lct_tags = LCT_OSP_THREAD | LCT_TX_HANDLE,
1861 .lct_init = osp_txn_key_init,
1862 .lct_fini = osp_txn_key_fini
1864 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1866 static struct lu_device_type_operations osp_device_type_ops = {
1867 .ldto_init = osp_type_init,
1868 .ldto_fini = osp_type_fini,
1870 .ldto_start = osp_type_start,
1871 .ldto_stop = osp_type_stop,
1873 .ldto_device_alloc = osp_device_alloc,
1874 .ldto_device_free = osp_device_free,
1876 .ldto_device_fini = osp_device_fini
1879 static struct lu_device_type osp_device_type = {
1880 .ldt_tags = LU_DEVICE_DT,
1881 .ldt_name = LUSTRE_OSP_NAME,
1882 .ldt_ops = &osp_device_type_ops,
1883 .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1886 static struct obd_ops osp_obd_device_ops = {
1887 .o_owner = THIS_MODULE,
1888 .o_add_conn = client_import_add_conn,
1889 .o_del_conn = client_import_del_conn,
1890 .o_reconnect = osp_reconnect,
1891 .o_connect = osp_obd_connect,
1892 .o_disconnect = osp_obd_disconnect,
1893 .o_get_info = osp_obd_get_info,
1894 .o_set_info_async = osp_obd_set_info_async,
1895 .o_import_event = osp_import_event,
1896 .o_iocontrol = osp_iocontrol,
1897 .o_statfs = osp_obd_statfs,
1898 .o_fid_init = client_fid_init,
1899 .o_fid_fini = client_fid_fini,
1900 .o_fid_alloc = osp_fid_alloc,
1903 struct llog_operations osp_mds_ost_orig_logops;
1906 * Initialize OSP module.
1908 * Register device types OSP and Light Weight Proxy (LWP) (\see lwp_dev.c)
1909 * in obd_types (\see class_obd.c). Initialize procfs for the
1910 * the OSP device. Note: OSP was called OSC before Lustre 2.4,
1911 * so for compatibility it still uses the name "osc" in procfs.
1912 * This is called at module load time.
1914 * \retval 0 0 if initialization succeeds.
1915 * \retval negative negative errno if initialization failed.
1917 static int __init osp_init(void)
1919 struct obd_type *type;
1922 rc = lu_kmem_init(osp_caches);
1927 rc = class_register_type(&osp_obd_device_ops, NULL, true, NULL,
1928 LUSTRE_OSP_NAME, &osp_device_type);
1930 lu_kmem_fini(osp_caches);
1934 rc = class_register_type(&lwp_obd_device_ops, NULL, true, NULL,
1935 LUSTRE_LWP_NAME, &lwp_device_type);
1937 class_unregister_type(LUSTRE_OSP_NAME);
1938 lu_kmem_fini(osp_caches);
1942 /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1943 osp_mds_ost_orig_logops = llog_osd_ops;
1944 osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1945 osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1947 /* create "osc" entry in procfs for compatibility purposes */
1948 type = class_search_type(LUSTRE_OSC_NAME);
1949 if (type != NULL && type->typ_procroot != NULL)
1952 type = class_search_type(LUSTRE_OSP_NAME);
1953 type->typ_procsym = lprocfs_register("osc", proc_lustre_root,
1955 if (IS_ERR(type->typ_procsym)) {
1956 CERROR("osp: can't create compat entry \"osc\": %d\n",
1957 (int) PTR_ERR(type->typ_procsym));
1958 type->typ_procsym = NULL;
1964 * Finalize OSP module.
1966 * This callback is called when kernel unloads OSP module from memory, and
1967 * it will deregister OSP and LWP device type from obd_types (\see class_obd.c).
1969 static void __exit osp_exit(void)
1971 class_unregister_type(LUSTRE_LWP_NAME);
1972 class_unregister_type(LUSTRE_OSP_NAME);
1973 lu_kmem_fini(osp_caches);
1976 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1977 MODULE_DESCRIPTION("Lustre OSD Storage Proxy ("LUSTRE_OSP_NAME")");
1978 MODULE_VERSION(LUSTRE_VERSION_STRING);
1979 MODULE_LICENSE("GPL");
1981 module_init(osp_init);
1982 module_exit(osp_exit);