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/
31 * lustre/osp/osp_dev.c
33 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
34 * Author: Mikhail Pershin <mike.pershin@intel.com>
35 * Author: Di Wang <di.wang@intel.com>
38 * The Object Storage Proxy (OSP) module provides an implementation of
39 * the DT API for remote MDTs and OSTs. Every local OSP device (or
40 * object) is a proxy for a remote OSD device (or object). Thus OSP
41 * converts DT operations into RPCs, which are sent to the OUT service
42 * on a remote target, converted back to DT operations, and
43 * executed. Of course there are many ways in which this description
44 * is inaccurate but it's a good enough mental model. OSP is used by
45 * the MDT stack in several ways:
47 * - OSP devices allocate FIDs for the stripe sub-objects of a striped
50 * - OSP objects represent the remote MDT and OST objects that are
51 * the stripes of a striped object.
53 * - OSP devices log, send, and track synchronous operations (setattr
54 * and unlink) to remote targets.
56 * - OSP objects are the bottom slice of the compound LU object
57 * representing a remote MDT object: MDT/MDD/LOD/OSP.
59 * - OSP objects are used by LFSCK to represent remote OST objects
60 * during the verification of MDT-OST consistency.
62 * - OSP devices batch idempotent requests (declare_attr_get() and
63 * declare_xattr_get()) to the remote target and cache their results.
65 * In addition the OSP layer implements a subset of the OBD device API
66 * to support being a client of a remote target, connecting to other
67 * layers, and FID allocation.
70 #define DEBUG_SUBSYSTEM S_MDS
72 #include <linux/kthread.h>
74 #include <uapi/linux/lustre/lustre_ioctl.h>
75 #include <lustre_log.h>
76 #include <lustre_obdo.h>
77 #include <uapi/linux/lustre/lustre_param.h>
78 #include <obd_class.h>
80 #include "osp_internal.h"
82 /* Slab for OSP object allocation */
83 struct kmem_cache *osp_object_kmem;
85 static struct lu_kmem_descr osp_caches[] = {
87 .ckd_cache = &osp_object_kmem,
88 .ckd_name = "osp_obj",
89 .ckd_size = sizeof(struct osp_object)
97 * Implementation of lu_device_operations::ldo_object_alloc
99 * Allocates an OSP object in memory, whose FID is on the remote target.
101 * \param[in] env execution environment
102 * \param[in] hdr The header of the object stack. If it is NULL, it
103 * means the object is not built from top device, i.e.
104 * it is a sub-stripe object of striped directory or
106 * \param[in] d OSP device
108 * \retval object object being created if the creation succeed.
109 * \retval NULL NULL if the creation failed.
111 static struct lu_object *osp_object_alloc(const struct lu_env *env,
112 const struct lu_object_header *hdr,
115 struct lu_object_header *h = NULL;
116 struct osp_object *o;
119 OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, GFP_NOFS);
121 l = &o->opo_obj.do_lu;
123 /* If hdr is NULL, it means the object is not built
124 * from the top dev(MDT/OST), usually it happens when
125 * building striped object, like data object on MDT or
126 * striped object for directory */
129 lu_object_header_init(h);
130 dt_object_init(&o->opo_obj, h, d);
131 lu_object_add_top(h, l);
133 dt_object_init(&o->opo_obj, h, d);
136 l->lo_ops = &osp_lu_obj_ops;
145 * Find or create the local object
147 * Finds or creates the local file referenced by \a reg_id and return the
148 * attributes of the local file.
150 * \param[in] env execution environment
151 * \param[in] osp OSP device
152 * \param[out] attr attributes of the object
153 * \param[in] reg_id the local object ID of the file. It will be used
154 * to compose a local FID{FID_SEQ_LOCAL_FILE, reg_id, 0}
155 * to identify the object.
157 * \retval object object(dt_object) found or created
158 * \retval ERR_PTR(errno) ERR_PTR(errno) if not get the object.
160 static struct dt_object
161 *osp_find_or_create_local_file(const struct lu_env *env, struct osp_device *osp,
162 struct lu_attr *attr, __u32 reg_id)
164 struct osp_thread_info *osi = osp_env_info(env);
165 struct dt_object_format dof = { 0 };
166 struct dt_object *dto;
170 lu_local_obj_fid(&osi->osi_fid, reg_id);
171 attr->la_valid = LA_MODE;
172 attr->la_mode = S_IFREG | 0644;
173 dof.dof_type = DFT_REGULAR;
174 /* Find or create the local object by osi_fid. */
175 dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
180 /* Get attributes of the local object. */
181 rc = dt_attr_get(env, dto, attr);
183 CERROR("%s: can't be initialized: rc = %d\n",
184 osp->opd_obd->obd_name, rc);
185 dt_object_put(env, dto);
192 * Write data buffer to a local file object.
194 * \param[in] env execution environment
195 * \param[in] osp OSP device
196 * \param[in] dt_obj object written to
197 * \param[in] buf buffer containing byte array and length
198 * \param[in] offset write offset in the object in bytes
200 * \retval 0 0 if write succeed
201 * \retval -EFAULT -EFAULT if only part of buffer is written.
202 * \retval negative other negative errno if write failed.
204 static int osp_write_local_file(const struct lu_env *env,
205 struct osp_device *osp,
206 struct dt_object *dt_obj,
213 if (osp->opd_storage->dd_rdonly)
216 th = dt_trans_create(env, osp->opd_storage);
220 rc = dt_declare_record_write(env, dt_obj, buf, offset, th);
223 rc = dt_trans_start_local(env, osp->opd_storage, th);
227 rc = dt_record_write(env, dt_obj, buf, &offset, th);
229 dt_trans_stop(env, osp->opd_storage, th);
234 * Initialize last ID object.
236 * This function initializes the LAST_ID file, which stores the current last
237 * used id of data objects. The MDT will use the last used id and the last_seq
238 * (\see osp_init_last_seq()) to synchronize the precreate object cache with
241 * \param[in] env execution environment
242 * \param[in] osp OSP device
244 * \retval 0 0 if initialization succeed
245 * \retval negative negative errno if initialization failed
247 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
249 struct osp_thread_info *osi = osp_env_info(env);
250 struct lu_fid *fid = &osp->opd_last_used_fid;
251 struct dt_object *dto;
255 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
258 RETURN(PTR_ERR(dto));
260 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &osp->opd_last_id,
263 /* object will be released in device cleanup path */
264 if (osi->osi_attr.la_size >= (osi->osi_off + osi->osi_lb.lb_len)) {
265 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
266 if (rc != 0 && rc != -EFAULT)
268 /* In case of idif bits 32-48 go to f_seq
269 * (see osp_init_last_seq). So don't care
270 * about u64->u32 convertion. */
271 fid->f_oid = osp->opd_last_id;
274 if (rc == -EFAULT) { /* fresh LAST_ID */
275 osp->opd_last_id = 0;
277 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
282 osp->opd_last_used_oid_file = dto;
285 /* object will be released in device cleanup path */
286 CERROR("%s: can't initialize lov_objid: rc = %d\n",
287 osp->opd_obd->obd_name, rc);
288 dt_object_put(env, dto);
289 osp->opd_last_used_oid_file = NULL;
294 * Initialize last sequence object.
296 * This function initializes the LAST_SEQ file in the local OSD, which stores
297 * the current last used sequence of data objects. The MDT will use the last
298 * sequence and last id (\see osp_init_last_objid()) to synchronize the
299 * precreate object cache with OSTs.
301 * \param[in] env execution environment
302 * \param[in] osp OSP device
304 * \retval 0 0 if initialization succeed
305 * \retval negative negative errno if initialization failed
307 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
309 struct osp_thread_info *osi = osp_env_info(env);
310 struct lu_fid *fid = &osp->opd_last_used_fid;
311 struct dt_object *dto;
315 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
318 RETURN(PTR_ERR(dto));
320 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
323 /* object will be released in device cleanup path */
324 if (osi->osi_attr.la_size >= (osi->osi_off + osi->osi_lb.lb_len)) {
325 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
326 if (rc != 0 && rc != -EFAULT)
328 if (fid_is_idif(fid))
329 fid->f_seq = fid_idif_seq(osp->opd_last_id,
333 if (rc == -EFAULT) { /* fresh OSP */
335 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
340 osp->opd_last_used_seq_file = dto;
343 /* object will be released in device cleanup path */
344 CERROR("%s: can't initialize lov_seq: rc = %d\n",
345 osp->opd_obd->obd_name, rc);
346 dt_object_put(env, dto);
347 osp->opd_last_used_seq_file = NULL;
352 * Initialize last OID and sequence object.
354 * If the MDT is just upgraded to 2.4 from the lower version, where the
355 * LAST_SEQ file does not exist, the file will be created and IDIF sequence
356 * will be written into the file.
358 * \param[in] env execution environment
359 * \param[in] osp OSP device
361 * \retval 0 0 if initialization succeed
362 * \retval negative negative error if initialization failed
364 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
366 struct osp_thread_info *osi = osp_env_info(env);
370 fid_zero(&osp->opd_last_used_fid);
371 rc = osp_init_last_objid(env, osp);
373 CERROR("%s: Can not get ids %d from old objid!\n",
374 osp->opd_obd->obd_name, rc);
378 rc = osp_init_last_seq(env, osp);
380 CERROR("%s: Can not get sequence %d from old objseq!\n",
381 osp->opd_obd->obd_name, rc);
385 if (fid_oid(&osp->opd_last_used_fid) != 0 &&
386 fid_seq(&osp->opd_last_used_fid) == 0) {
387 /* Just upgrade from the old version,
388 * set the seq to be IDIF */
389 osp->opd_last_used_fid.f_seq =
390 fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
392 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
393 &osp->opd_last_used_fid.f_seq,
395 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
396 &osi->osi_lb, osi->osi_off);
398 CERROR("%s : Can not write seq file: rc = %d\n",
399 osp->opd_obd->obd_name, rc);
404 if (!fid_is_zero(&osp->opd_last_used_fid) &&
405 !fid_is_sane(&osp->opd_last_used_fid)) {
406 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
407 PFID(&osp->opd_last_used_fid));
408 GOTO(out, rc = -EINVAL);
411 osp_fid_to_obdid(&osp->opd_last_used_fid, &osp->opd_last_id);
412 CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
413 osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
416 if (osp->opd_last_used_oid_file != NULL) {
417 dt_object_put(env, osp->opd_last_used_oid_file);
418 osp->opd_last_used_oid_file = NULL;
420 if (osp->opd_last_used_seq_file != NULL) {
421 dt_object_put(env, osp->opd_last_used_seq_file);
422 osp->opd_last_used_seq_file = NULL;
430 * Release the last sequence and OID file objects in OSP device.
432 * \param[in] env execution environment
433 * \param[in] osp OSP device
435 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *osp)
437 /* release last_used file */
438 if (osp->opd_last_used_oid_file != NULL) {
439 dt_object_put(env, osp->opd_last_used_oid_file);
440 osp->opd_last_used_oid_file = NULL;
443 if (osp->opd_last_used_seq_file != NULL) {
444 dt_object_put(env, osp->opd_last_used_seq_file);
445 osp->opd_last_used_seq_file = NULL;
450 * Disconnects the connection between OSP and its correspondent MDT or OST, and
451 * the import will be marked as inactive. It will only be called during OSP
454 * \param[in] d OSP device being disconnected
456 * \retval 0 0 if disconnection succeed
457 * \retval negative negative errno if disconnection failed
459 static int osp_disconnect(struct osp_device *d)
461 struct obd_device *obd = d->opd_obd;
462 struct obd_import *imp;
465 imp = obd->u.cli.cl_import;
467 /* Mark import deactivated now, so we don't try to reconnect if any
468 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
469 * fully deactivate the import, or that would drop all requests. */
470 LASSERT(imp != NULL);
471 spin_lock(&imp->imp_lock);
472 imp->imp_deactive = 1;
473 spin_unlock(&imp->imp_lock);
475 ptlrpc_deactivate_import(imp);
477 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
478 * delete it regardless. (It's safe to delete an import that was
480 (void)ptlrpc_pinger_del_import(imp);
482 /* Send disconnect on healthy import, do force disconnect otherwise */
483 spin_lock(&imp->imp_lock);
484 imp->imp_obd->obd_force = imp->imp_state != LUSTRE_IMP_FULL;
485 spin_unlock(&imp->imp_lock);
487 rc = ptlrpc_disconnect_import(imp, 0);
489 CERROR("%s: can't disconnect: rc = %d\n", obd->obd_name, rc);
491 ptlrpc_invalidate_import(imp);
497 * Initialize the osp_update structure in OSP device
499 * Allocate osp update structure and start update thread.
501 * \param[in] osp OSP device
503 * \retval 0 if initialization succeeds.
504 * \retval negative errno if initialization fails.
506 static int osp_update_init(struct osp_device *osp)
508 struct task_struct *task;
513 LASSERT(osp->opd_connect_mdt);
515 if (osp->opd_storage->dd_rdonly)
518 OBD_ALLOC_PTR(osp->opd_update);
519 if (osp->opd_update == NULL)
522 init_waitqueue_head(&osp->opd_update->ou_waitq);
523 spin_lock_init(&osp->opd_update->ou_lock);
524 INIT_LIST_HEAD(&osp->opd_update->ou_list);
525 osp->opd_update->ou_rpc_version = 1;
526 osp->opd_update->ou_version = 1;
527 osp->opd_update->ou_generation = 0;
529 rc = lu_env_init(&osp->opd_update->ou_env,
530 osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
532 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
534 OBD_FREE_PTR(osp->opd_update);
535 osp->opd_update = NULL;
538 /* start thread handling sending updates to the remote MDT */
539 task = kthread_create(osp_send_update_thread, osp,
540 "osp_up%u-%u", osp->opd_index, osp->opd_group);
542 int rc = PTR_ERR(task);
544 lu_env_fini(&osp->opd_update->ou_env);
545 OBD_FREE_PTR(osp->opd_update);
546 osp->opd_update = NULL;
547 CERROR("%s: can't start precreate thread: rc = %d\n",
548 osp->opd_obd->obd_name, rc);
552 osp->opd_update->ou_update_task = task;
553 wake_up_process(task);
559 * Finialize osp_update structure in OSP device
561 * Stop the OSP update sending thread, then delete the left
562 * osp thandle in the sending list.
564 * \param [in] osp OSP device.
566 static void osp_update_fini(const struct lu_env *env, struct osp_device *osp)
568 struct osp_update_request *our;
569 struct osp_update_request *tmp;
570 struct osp_updates *ou = osp->opd_update;
575 kthread_stop(ou->ou_update_task);
576 lu_env_fini(&ou->ou_env);
578 /* Remove the left osp thandle from the list */
579 spin_lock(&ou->ou_lock);
580 list_for_each_entry_safe(our, tmp, &ou->ou_list,
582 list_del_init(&our->our_list);
583 LASSERT(our->our_th != NULL);
584 osp_trans_callback(env, our->our_th, -EIO);
585 /* our will be destroyed in osp_thandle_put() */
586 osp_thandle_put(env, our->our_th);
588 spin_unlock(&ou->ou_lock);
591 osp->opd_update = NULL;
595 * Cleanup OSP, which includes disconnect import, cleanup unlink log, stop
596 * precreate threads etc.
598 * \param[in] env execution environment.
599 * \param[in] d OSP device being disconnected.
601 * \retval 0 0 if cleanup succeed
602 * \retval negative negative errno if cleanup failed
604 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
611 rc = osp_disconnect(d);
615 if (!d->opd_connect_mdt) {
616 /* stop sync thread */
619 /* stop precreate thread */
620 osp_precreate_fini(d);
622 /* release last_used file */
623 osp_last_used_fini(env, d);
626 obd_fid_fini(d->opd_obd);
632 * Implementation of osp_lu_ops::ldo_process_config
634 * This function processes config log records in OSP layer. It is usually
635 * called from the top layer of MDT stack, and goes through the stack by calling
636 * ldo_process_config of next layer.
638 * \param[in] env execution environment
639 * \param[in] dev lu_device of OSP
640 * \param[in] lcfg config log
642 * \retval 0 0 if the config log record is executed correctly.
643 * \retval negative negative errno if the record execution fails.
645 static int osp_process_config(const struct lu_env *env,
646 struct lu_device *dev, struct lustre_cfg *lcfg)
648 struct osp_device *d = lu2osp_dev(dev);
649 struct dt_device *dt = lu2dt_dev(dev);
650 struct obd_device *obd = d->opd_obd;
656 switch (lcfg->lcfg_command) {
657 case LCFG_PRE_CLEANUP:
658 rc = osp_disconnect(d);
659 osp_update_fini(env, d);
660 if (obd->obd_namespace != NULL)
661 ldlm_namespace_free_prior(obd->obd_namespace, NULL, 1);
664 lu_dev_del_linkage(dev->ld_site, dev);
665 rc = osp_shutdown(env, d);
668 count = class_modify_config(lcfg, d->opd_connect_mdt ?
669 PARAM_OSP : PARAM_OSC,
672 /* class_modify_config() haven't found matching
673 * parameter and returned an error so that layer(s)
674 * below could use that. But OSP is the bottom, so
677 CERROR("%s: unknown param %s\n",
678 (char *)lustre_cfg_string(lcfg, 0),
679 (char *)lustre_cfg_string(lcfg, 1));
684 CERROR("%s: unknown command %u\n",
685 (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
694 * Implementation of osp_lu_ops::ldo_recovery_complete
696 * This function is called after recovery is finished, and OSP layer
697 * will wake up precreate thread here.
699 * \param[in] env execution environment
700 * \param[in] dev lu_device of OSP
702 * \retval 0 0 unconditionally
704 static int osp_recovery_complete(const struct lu_env *env,
705 struct lu_device *dev)
707 struct osp_device *osp = lu2osp_dev(dev);
710 osp->opd_recovery_completed = 1;
712 if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
713 wake_up(&osp->opd_pre_waitq);
719 * Implementation of lu_device_operations::ldo_fid_alloc() for OSP
721 * Allocate FID from remote MDT.
723 * see include/lu_object.h for the details.
725 static int osp_fid_alloc(const struct lu_env *env, struct lu_device *d,
726 struct lu_fid *fid, struct lu_object *parent,
727 const struct lu_name *name)
729 struct osp_device *osp = lu2osp_dev(d);
730 struct client_obd *cli = &osp->opd_obd->u.cli;
731 struct lu_client_seq *seq = cli->cl_seq;
736 /* Sigh, fid client is not ready yet */
737 if (!osp->opd_obd->u.cli.cl_seq)
740 if (!osp->opd_obd->u.cli.cl_seq->lcs_exp)
743 rc = seq_client_alloc_fid(env, seq, fid);
748 const struct lu_device_operations osp_lu_ops = {
749 .ldo_object_alloc = osp_object_alloc,
750 .ldo_process_config = osp_process_config,
751 .ldo_recovery_complete = osp_recovery_complete,
752 .ldo_fid_alloc = osp_fid_alloc,
756 * Implementation of dt_device_operations::dt_statfs
758 * This function provides statfs status (for precreation) from
759 * corresponding OST. Note: this function only retrieves the status
760 * from the OSP device, and the real statfs RPC happens inside
761 * precreate thread (\see osp_statfs_update). Note: OSP for MDT does
762 * not need to retrieve statfs data for now.
764 * \param[in] env execution environment.
765 * \param[in] dev dt_device of OSP.
766 * \param[out] sfs holds the retrieved statfs data.
768 * \retval 0 0 statfs data was retrieved successfully or
769 * retrieval was not needed
770 * \retval negative negative errno if get statfs failed.
772 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
773 struct obd_statfs *sfs, struct obd_statfs_info *info)
775 struct osp_device *d = dt2osp_dev(dev);
776 struct obd_import *imp = d->opd_obd->u.cli.cl_import;
780 if (imp->imp_state == LUSTRE_IMP_CLOSED)
783 if (unlikely(d->opd_imp_active == 0))
786 /* return recently updated data */
787 *sfs = d->opd_statfs;
789 info->os_reserved_mb_low = d->opd_reserved_mb_low;
790 info->os_reserved_mb_high = d->opd_reserved_mb_high;
793 if (d->opd_pre == NULL)
796 CDEBUG(D_OTHER, "%s: %llu blocks, %llu free, %llu avail, "
797 "%u bsize, %u reserved mb low, %u reserved mb high, "
798 "%llu files, %llu free files\n", d->opd_obd->obd_name,
799 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail, sfs->os_bsize,
800 d->opd_reserved_mb_low, d->opd_reserved_mb_high,
801 sfs->os_files, sfs->os_ffree);
804 if (info && !info->os_enable_pre)
808 * The layer above osp (usually lod) can use f_precreated to
809 * estimate how many objects are available for immediate usage.
811 spin_lock(&d->opd_pre_lock);
812 sfs->os_fprecreated = osp_fid_diff(&d->opd_pre_last_created_fid,
813 &d->opd_pre_used_fid);
814 sfs->os_fprecreated -= d->opd_pre_reserved;
815 LASSERTF(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2,
816 "last_created "DFID", next_fid "DFID", reserved %llu\n",
817 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_pre_used_fid),
818 d->opd_pre_reserved);
819 spin_unlock(&d->opd_pre_lock);
824 * Implementation of dt_device_operations::dt_sync
826 * This function synchronizes the OSP cache to the remote target. It wakes
827 * up unlink log threads and sends out unlink records to the remote OST.
829 * \param[in] env execution environment
830 * \param[in] dev dt_device of OSP
832 * \retval 0 0 if synchronization succeeds
833 * \retval negative negative errno if synchronization fails
835 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
837 struct osp_device *d = dt2osp_dev(dev);
838 time64_t start = ktime_get_seconds();
844 /* No Sync between MDTs yet. */
845 if (d->opd_connect_mdt)
848 recs = atomic_read(&d->opd_sync_changes);
849 old = atomic64_read(&d->opd_sync_processed_recs);
851 osp_sync_force(env, dt2osp_dev(dev));
853 if (unlikely(d->opd_imp_active == 0))
856 down_write(&d->opd_async_updates_rwsem);
858 CDEBUG(D_OTHER, "%s: async updates %d\n", d->opd_obd->obd_name,
859 atomic_read(&d->opd_async_updates_count));
861 /* make sure the connection is fine */
862 rc = wait_event_idle_timeout(
863 d->opd_sync_barrier_waitq,
864 atomic_read(&d->opd_async_updates_count) == 0,
865 cfs_time_seconds(obd_timeout));
871 up_write(&d->opd_async_updates_rwsem);
875 CDEBUG(D_CACHE, "%s: processed %llu\n", d->opd_obd->obd_name,
876 (unsigned long long)atomic64_read(&d->opd_sync_processed_recs));
878 while (atomic64_read(&d->opd_sync_processed_recs) < old + recs) {
879 __u64 last = atomic64_read(&d->opd_sync_processed_recs);
880 /* make sure the connection is fine */
881 wait_event_idle_timeout(
882 d->opd_sync_barrier_waitq,
883 atomic64_read(&d->opd_sync_processed_recs)
885 cfs_time_seconds(obd_timeout));
887 if (atomic64_read(&d->opd_sync_processed_recs) >= old + recs)
890 if (atomic64_read(&d->opd_sync_processed_recs) != last) {
891 /* some progress have been made,
896 /* no changes and expired, something is wrong */
897 GOTO(out, rc = -ETIMEDOUT);
900 /* block new processing (barrier>0 - few callers are possible */
901 atomic_inc(&d->opd_sync_barrier);
903 CDEBUG(D_CACHE, "%s: %u in flight\n", d->opd_obd->obd_name,
904 atomic_read(&d->opd_sync_rpcs_in_flight));
906 /* wait till all-in-flight are replied, so executed by the target */
907 /* XXX: this is used by LFSCK at the moment, which doesn't require
908 * all the changes to be committed, but in general it'd be
909 * better to wait till commit */
910 while (atomic_read(&d->opd_sync_rpcs_in_flight) > 0) {
911 old = atomic_read(&d->opd_sync_rpcs_in_flight);
913 wait_event_idle_timeout(
914 d->opd_sync_barrier_waitq,
915 atomic_read(&d->opd_sync_rpcs_in_flight) == 0,
916 cfs_time_seconds(obd_timeout));
918 if (atomic_read(&d->opd_sync_rpcs_in_flight) == 0)
921 if (atomic_read(&d->opd_sync_rpcs_in_flight) != old) {
922 /* some progress have been made */
926 /* no changes and expired, something is wrong */
927 GOTO(out, rc = -ETIMEDOUT);
931 /* resume normal processing (barrier=0) */
932 atomic_dec(&d->opd_sync_barrier);
933 osp_sync_check_for_work(d);
935 CDEBUG(D_CACHE, "%s: done in %lld: rc = %d\n", d->opd_obd->obd_name,
936 ktime_get_seconds() - start, rc);
941 static const struct dt_device_operations osp_dt_ops = {
942 .dt_statfs = osp_statfs,
944 .dt_trans_create = osp_trans_create,
945 .dt_trans_start = osp_trans_start,
946 .dt_trans_stop = osp_trans_stop,
947 .dt_trans_cb_add = osp_trans_cb_add,
951 * Connect OSP to local OSD.
953 * Locate the local OSD referenced by \a nextdev and connect to it. Sometimes,
954 * OSP needs to access the local OSD to store some information. For example,
955 * during precreate, it needs to update last used OID and sequence file
956 * (LAST_SEQ) in local OSD.
958 * \param[in] env execution environment
959 * \param[in] osp OSP device
960 * \param[in] nextdev the name of local OSD
962 * \retval 0 0 connection succeeded
963 * \retval negative negative errno connection failed
965 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *osp,
968 struct obd_connect_data *data = NULL;
969 struct obd_device *obd;
974 LASSERT(osp->opd_storage_exp == NULL);
980 obd = class_name2obd(nextdev);
982 CERROR("%s: can't locate next device: %s\n",
983 osp->opd_obd->obd_name, nextdev);
984 GOTO(out, rc = -ENOTCONN);
987 rc = obd_connect(env, &osp->opd_storage_exp, obd, &obd->obd_uuid, data,
990 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
991 osp->opd_obd->obd_name, nextdev, rc);
995 osp->opd_dt_dev.dd_lu_dev.ld_site =
996 osp->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
997 LASSERT(osp->opd_dt_dev.dd_lu_dev.ld_site);
998 osp->opd_storage = lu2dt_dev(osp->opd_storage_exp->exp_obd->obd_lu_dev);
1006 * Determine if the lock needs to be cancelled
1008 * Determine if the unused lock should be cancelled before replay, see
1009 * (ldlm_cancel_no_wait_policy()). Currently, only inode bits lock exists
1012 * \param[in] lock lock to be checked.
1014 * \retval 1 if the lock needs to be cancelled before replay.
1015 * \retval 0 if the lock does not need to be cancelled before
1018 static int osp_cancel_weight(struct ldlm_lock *lock)
1020 if (lock->l_resource->lr_type != LDLM_IBITS)
1027 * Initialize OSP device according to the parameters in the configuration
1030 * Reconstruct the local device name from the configuration profile, and
1031 * initialize necessary threads and structures according to the OSP type
1034 * Since there is no record in the MDT configuration for the local disk
1035 * device, we have to extract this from elsewhere in the profile.
1036 * The only information we get at setup is from the OSC records:
1037 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
1039 * Note: configs generated by Lustre 1.8 are missing the -MDTxxxx part,
1040 * so, we need to reconstruct the name of the underlying OSD from this:
1041 * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".
1043 * \param[in] env execution environment
1044 * \param[in] osp OSP device
1045 * \param[in] ldt lu device type of OSP
1046 * \param[in] cfg configuration log
1048 * \retval 0 0 if OSP initialization succeeded.
1049 * \retval negative negative errno if OSP initialization failed.
1051 static int osp_init0(const struct lu_env *env, struct osp_device *osp,
1052 struct lu_device_type *ldt, struct lustre_cfg *cfg)
1054 struct obd_device *obd;
1055 struct obd_import *imp;
1056 char *src, *tgt, *osdname = NULL;
1063 mutex_init(&osp->opd_async_requests_mutex);
1064 INIT_LIST_HEAD(&osp->opd_async_updates);
1065 init_rwsem(&osp->opd_async_updates_rwsem);
1066 atomic_set(&osp->opd_async_updates_count, 0);
1068 obd = class_name2obd(lustre_cfg_string(cfg, 0));
1070 CERROR("Cannot find obd with name %s\n",
1071 lustre_cfg_string(cfg, 0));
1076 src = lustre_cfg_string(cfg, 0);
1080 tgt = strrchr(src, '-');
1082 CERROR("%s: invalid target name %s: rc = %d\n",
1083 osp->opd_obd->obd_name, lustre_cfg_string(cfg, 0),
1088 if (strncmp(tgt, "-osc", 4) == 0) {
1089 /* Old OSC name fsname-OSTXXXX-osc */
1090 for (tgt--; tgt > src && *tgt != '-'; tgt--)
1093 CERROR("%s: invalid target name %s: rc = %d\n",
1094 osp->opd_obd->obd_name,
1095 lustre_cfg_string(cfg, 0), -EINVAL);
1099 if (strncmp(tgt, "-OST", 4) != 0) {
1100 CERROR("%s: invalid target name %s: rc = %d\n",
1101 osp->opd_obd->obd_name,
1102 lustre_cfg_string(cfg, 0), -EINVAL);
1106 rc = target_name2index(tgt + 1, &idx, &mdt);
1107 if (rc < 0 || rc & LDD_F_SV_ALL || mdt[0] != '-') {
1108 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1109 osp->opd_obd->obd_name, src, -EINVAL);
1112 osp->opd_index = idx;
1116 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
1117 if (strncmp(tgt, "-MDT", 4) != 0 &&
1118 strncmp(tgt, "-OST", 4) != 0) {
1119 CERROR("%s: invalid target name %s: rc = %d\n",
1120 osp->opd_obd->obd_name,
1121 lustre_cfg_string(cfg, 0), -EINVAL);
1125 rc = target_name2index(tgt + 1, &idx, &mdt);
1126 if (rc < 0 || rc & LDD_F_SV_ALL || *mdt != '\0') {
1127 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1128 osp->opd_obd->obd_name, src, -EINVAL);
1132 /* Get MDT index from the name and set it to opd_group,
1133 * which will be used by OSP to connect with OST */
1134 osp->opd_group = idx;
1135 if (tgt - src <= 12) {
1136 CERROR("%s: invalid mdt index from %s: rc =%d\n",
1137 osp->opd_obd->obd_name,
1138 lustre_cfg_string(cfg, 0), -EINVAL);
1142 if (strncmp(tgt - 12, "-MDT", 4) == 0)
1143 osp->opd_connect_mdt = 1;
1145 rc = target_name2index(tgt - 11, &idx, &mdt);
1146 if (rc < 0 || rc & LDD_F_SV_ALL || mdt[0] != '-') {
1147 CERROR("%s: invalid OST index in '%s': rc =%d\n",
1148 osp->opd_obd->obd_name, src, -EINVAL);
1152 osp->opd_index = idx;
1153 idx = tgt - src - 12;
1155 /* check the fsname length, and after this everything else will fit */
1156 if (idx > MTI_NAME_MAXLEN) {
1157 CERROR("%s: fsname too long in '%s': rc = %d\n",
1158 osp->opd_obd->obd_name, src, -EINVAL);
1162 OBD_ALLOC(osdname, MAX_OBD_NAME);
1163 if (osdname == NULL)
1166 memcpy(osdname, src, idx); /* copy just the fsname part */
1167 osdname[idx] = '\0';
1169 mdt = strstr(mdt, "-MDT");
1170 if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
1171 strcat(osdname, "-MDT0000");
1173 strcat(osdname, mdt);
1174 strcat(osdname, "-osd");
1175 CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
1177 osp_init_rpc_lock(osp);
1179 osp->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
1180 osp->opd_dt_dev.dd_ops = &osp_dt_ops;
1182 obd->obd_lu_dev = &osp->opd_dt_dev.dd_lu_dev;
1184 rc = osp_connect_to_osd(env, osp, osdname);
1188 rc = ptlrpcd_addref();
1190 GOTO(out_disconnect, rc);
1192 rc = client_obd_setup(obd, cfg);
1194 CERROR("%s: can't setup obd: rc = %d\n", osp->opd_obd->obd_name,
1199 osp_tunables_init(osp);
1201 rc = obd_fid_init(osp->opd_obd, NULL, osp->opd_connect_mdt ?
1202 LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
1204 CERROR("%s: fid init error: rc = %d\n",
1205 osp->opd_obd->obd_name, rc);
1209 if (!osp->opd_connect_mdt) {
1210 /* Initialize last id from the storage - will be
1211 * used in orphan cleanup. */
1212 if (!osp->opd_storage->dd_rdonly) {
1213 rc = osp_last_used_init(env, osp);
1218 /* Initialize precreation thread, it handles new
1219 * connections as well. */
1220 rc = osp_init_precreate(osp);
1222 GOTO(out_last_used, rc);
1225 * Initialize synhronization mechanism taking
1226 * care of propogating changes to OST in near
1227 * transactional manner.
1229 rc = osp_sync_init(env, osp);
1231 GOTO(out_precreat, rc);
1233 osp->opd_got_disconnected = 1;
1234 rc = osp_update_init(osp);
1239 rc = osp_init_statfs(osp);
1241 GOTO(out_precreat, rc);
1243 ns_register_cancel(obd->obd_namespace, osp_cancel_weight);
1246 * Initiate connect to OST
1248 imp = obd->u.cli.cl_import;
1250 rc = ptlrpc_init_import(imp);
1254 OBD_FREE(osdname, MAX_OBD_NAME);
1258 if (!osp->opd_connect_mdt)
1259 /* stop sync thread */
1262 /* stop precreate thread */
1263 if (!osp->opd_connect_mdt)
1264 osp_precreate_fini(osp);
1266 osp_update_fini(env, osp);
1268 if (!osp->opd_connect_mdt)
1269 osp_last_used_fini(env, osp);
1271 obd_fid_fini(osp->opd_obd);
1273 osp_tunables_fini(osp);
1274 client_obd_cleanup(obd);
1278 obd_disconnect(osp->opd_storage_exp);
1281 OBD_FREE(osdname, MAX_OBD_NAME);
1286 * Implementation of lu_device_type_operations::ldto_device_free
1288 * Free the OSP device in memory. No return value is needed for now,
1289 * so always return NULL to comply with the interface.
1291 * \param[in] env execution environment
1292 * \param[in] lu lu_device of OSP
1294 * \retval NULL NULL unconditionally
1296 static struct lu_device *osp_device_free(const struct lu_env *env,
1297 struct lu_device *lu)
1299 struct osp_device *osp = lu2osp_dev(lu);
1301 lu_site_print(env, lu->ld_site, &lu->ld_ref, D_ERROR,
1303 dt_device_fini(&osp->opd_dt_dev);
1310 * Implementation of lu_device_type_operations::ldto_device_alloc
1312 * This function allocates and initializes OSP device in memory according to
1315 * \param[in] env execution environment
1316 * \param[in] type device type of OSP
1317 * \param[in] lcfg config log
1319 * \retval pointer the pointer of allocated OSP if succeed.
1320 * \retval ERR_PTR(errno) ERR_PTR(errno) if failed.
1322 static struct lu_device *osp_device_alloc(const struct lu_env *env,
1323 struct lu_device_type *type,
1324 struct lustre_cfg *lcfg)
1326 struct osp_device *osp;
1327 struct lu_device *ld;
1331 ld = ERR_PTR(-ENOMEM);
1335 ld = osp2lu_dev(osp);
1336 dt_device_init(&osp->opd_dt_dev, type);
1337 rc = osp_init0(env, osp, type, lcfg);
1339 osp_device_free(env, ld);
1347 * Implementation of lu_device_type_operations::ldto_device_fini
1349 * This function cleans up the OSP device, i.e. release and free those
1350 * attached items in osp_device.
1352 * \param[in] env execution environment
1353 * \param[in] ld lu_device of OSP
1355 * \retval NULL NULL if cleanup succeeded.
1356 * \retval ERR_PTR(errno) ERR_PTR(errno) if cleanup failed.
1358 static struct lu_device *osp_device_fini(const struct lu_env *env,
1359 struct lu_device *ld)
1361 struct osp_device *osp = lu2osp_dev(ld);
1366 if (osp->opd_async_requests != NULL) {
1367 osp_update_request_destroy(env, osp->opd_async_requests);
1368 osp->opd_async_requests = NULL;
1371 if (osp->opd_storage_exp) {
1372 /* wait for the commit callbacks to complete */
1373 wait_event(osp->opd_sync_waitq,
1374 atomic_read(&osp->opd_commits_registered) == 0);
1375 obd_disconnect(osp->opd_storage_exp);
1378 LASSERT(osp->opd_obd);
1380 rc = client_obd_cleanup(osp->opd_obd);
1383 RETURN(ERR_PTR(rc));
1386 osp_tunables_fini(osp);
1394 * Implementation of obd_ops::o_reconnect
1396 * This function is empty and does not need to do anything for now.
1398 static int osp_reconnect(const struct lu_env *env,
1399 struct obd_export *exp, struct obd_device *obd,
1400 struct obd_uuid *cluuid,
1401 struct obd_connect_data *data,
1408 * Implementation of obd_ops::o_connect
1410 * Connect OSP to the remote target (MDT or OST). Allocate the
1411 * export and return it to the LOD, which calls this function
1412 * for each OSP to connect it to the remote target. This function
1413 * is currently only called once per OSP.
1415 * \param[in] env execution environment
1416 * \param[out] exp export connected to OSP
1417 * \param[in] obd OSP device
1418 * \param[in] cluuid OSP device client uuid
1419 * \param[in] data connect_data to be used to connect to the remote
1421 * \param[in] localdata necessary for the API interface, but not used in
1424 * \retval 0 0 if the connection succeeded.
1425 * \retval negative negative errno if the connection failed.
1427 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
1428 struct obd_device *obd, struct obd_uuid *cluuid,
1429 struct obd_connect_data *data, void *localdata)
1431 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1436 LASSERT(data != NULL);
1437 LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
1439 rc = client_connect_import(env, &osp->opd_exp, obd, cluuid, data,
1444 osp->opd_obd->u.cli.cl_seq->lcs_exp = class_export_get(osp->opd_exp);
1445 *exp = osp->opd_exp;
1451 * Implementation of obd_ops::o_disconnect
1453 * Disconnect the export for the OSP. This is called by LOD to release the
1454 * OSP during cleanup (\see lod_del_device()). The OSP will be released after
1455 * the export is released.
1457 * \param[in] exp export to be disconnected.
1459 * \retval 0 0 if disconnection succeed
1460 * \retval negative negative errno if disconnection failed
1462 static int osp_obd_disconnect(struct obd_export *exp)
1464 struct obd_device *obd = exp->exp_obd;
1468 rc = class_disconnect(exp);
1470 CERROR("%s: class disconnect error: rc = %d\n",
1475 /* destroy the device */
1476 class_manual_cleanup(obd);
1482 * Implementation of obd_ops::o_statfs
1484 * Send a RPC to the remote target to get statfs status. This is only used
1485 * in lprocfs helpers by obd_statfs.
1487 * \param[in] env execution environment
1488 * \param[in] exp connection state from this OSP to the parent (LOD)
1490 * \param[out] osfs hold the statfs result
1491 * \param[in] unused Not used in this function for now
1492 * \param[in] flags flags to indicate how OSP will issue the RPC
1494 * \retval 0 0 if statfs succeeded.
1495 * \retval negative negative errno if statfs failed.
1497 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
1498 struct obd_statfs *osfs, time64_t unused, __u32 flags)
1500 struct obd_statfs *msfs;
1501 struct ptlrpc_request *req;
1502 struct obd_import *imp = NULL, *imp0;
1507 /* Since the request might also come from lprocfs, so we need
1508 * sync this with client_disconnect_export Bug15684
1510 with_imp_locked(exp->exp_obd, imp0, rc)
1511 imp = class_import_get(imp0);
1515 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1517 class_import_put(imp);
1522 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1524 ptlrpc_request_free(req);
1527 ptlrpc_request_set_replen(req);
1528 req->rq_request_portal = OST_CREATE_PORTAL;
1529 ptlrpc_at_set_req_timeout(req);
1531 if (flags & OBD_STATFS_NODELAY) {
1532 /* procfs requests not want stat in wait for avoid deadlock */
1533 req->rq_no_resend = 1;
1534 req->rq_no_delay = 1;
1537 rc = ptlrpc_queue_wait(req);
1541 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1543 GOTO(out, rc = -EPROTO);
1549 ptlrpc_req_finished(req);
1554 * Implementation of obd_ops::o_import_event
1556 * This function is called when some related import event happens. It will
1557 * mark the necessary flags according to the event and notify the necessary
1558 * threads (mainly precreate thread).
1560 * \param[in] obd OSP OBD device
1561 * \param[in] imp import attached from OSP to remote (OST/MDT) service
1562 * \param[in] event event related to remote service (IMP_EVENT_*)
1564 * \retval 0 0 if the event handling succeeded.
1565 * \retval negative negative errno if the event handling failed.
1567 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1568 enum obd_import_event event)
1570 struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1574 case IMP_EVENT_DISCON:
1575 d->opd_got_disconnected = 1;
1576 d->opd_imp_connected = 0;
1577 if (d->opd_connect_mdt)
1580 if (d->opd_pre != NULL) {
1581 osp_pre_update_status(d, -ENODEV);
1582 wake_up(&d->opd_pre_waitq);
1585 CDEBUG(D_HA, "got disconnected\n");
1587 case IMP_EVENT_INACTIVE:
1588 d->opd_imp_active = 0;
1589 d->opd_imp_connected = 0;
1590 d->opd_obd->obd_inactive = 1;
1591 if (d->opd_connect_mdt)
1593 if (d->opd_pre != NULL) {
1594 /* Import is invalid, we can`t get stripes so
1596 rc = imp->imp_deactive ? -ESHUTDOWN : -ENODEV;
1597 osp_pre_update_status(d, rc);
1598 wake_up(&d->opd_pre_waitq);
1601 CDEBUG(D_HA, "got inactive\n");
1603 case IMP_EVENT_ACTIVE:
1604 d->opd_imp_active = 1;
1606 d->opd_new_connection = 1;
1607 d->opd_imp_connected = 1;
1608 d->opd_imp_seen_connected = 1;
1609 d->opd_obd->obd_inactive = 0;
1610 wake_up(&d->opd_pre_waitq);
1611 if (d->opd_connect_mdt)
1614 osp_sync_check_for_work(d);
1615 CDEBUG(D_HA, "got connected\n");
1617 case IMP_EVENT_INVALIDATE:
1618 if (d->opd_connect_mdt)
1619 osp_invalidate_request(d);
1621 if (obd->obd_namespace == NULL)
1623 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1626 case IMP_EVENT_DEACTIVATE:
1627 case IMP_EVENT_ACTIVATE:
1630 CERROR("%s: unsupported import event: %#x\n",
1631 obd->obd_name, event);
1637 * Implementation of obd_ops: o_iocontrol
1639 * This function is the ioctl handler for OSP. Note: lctl will access the OSP
1640 * directly by ioctl, instead of through the MDS stack.
1642 * param[in] cmd ioctl command.
1643 * param[in] exp export of this OSP.
1644 * param[in] len data length of \a karg.
1645 * param[in] karg input argument which is packed as
1647 * param[out] uarg pointer to userspace buffer (must access by
1650 * \retval 0 0 if the ioctl handling succeeded.
1651 * \retval negative negative errno if the ioctl handling failed.
1653 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1654 void *karg, void __user *uarg)
1656 struct obd_device *obd = exp->exp_obd;
1657 struct osp_device *d;
1658 struct obd_ioctl_data *data = karg;
1663 LASSERT(obd->obd_lu_dev);
1664 d = lu2osp_dev(obd->obd_lu_dev);
1665 LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1667 if (!try_module_get(THIS_MODULE)) {
1668 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1669 module_name(THIS_MODULE));
1674 case OBD_IOC_CLIENT_RECOVER:
1675 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1676 data->ioc_inlbuf1, 0);
1680 case IOC_OSC_SET_ACTIVE:
1681 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1685 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1686 cmd, current->comm);
1689 module_put(THIS_MODULE);
1695 * Implementation of obd_ops::o_get_info
1697 * Retrieve information by key. Retrieval starts from the top layer
1698 * (MDT) of the MDS stack and traverses the stack by calling the
1699 * obd_get_info() method of the next sub-layer.
1701 * \param[in] env execution environment
1702 * \param[in] exp export of this OSP
1703 * \param[in] keylen length of \a key
1704 * \param[in] key the key
1705 * \param[out] vallen length of \a val
1706 * \param[out] val holds the value returned by the key
1708 * \retval 0 0 if getting information succeeded.
1709 * \retval negative negative errno if getting information failed.
1711 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1712 __u32 keylen, void *key, __u32 *vallen, void *val)
1716 if (KEY_IS(KEY_OSP_CONNECTED)) {
1717 struct obd_device *obd = exp->exp_obd;
1718 struct osp_device *osp;
1720 if (!obd->obd_set_up || obd->obd_stopping)
1723 osp = lu2osp_dev(obd->obd_lu_dev);
1726 * 1.8/2.0 behaviour is that OST being connected once at least
1727 * is considered "healthy". and one "healthy" OST is enough to
1728 * allow lustre clients to connect to MDS
1730 RETURN(!osp->opd_imp_seen_connected);
1736 static int osp_obd_set_info_async(const struct lu_env *env,
1737 struct obd_export *exp,
1738 u32 keylen, void *key,
1739 u32 vallen, void *val,
1740 struct ptlrpc_request_set *set)
1742 struct obd_device *obd = exp->exp_obd;
1743 struct obd_import *imp = obd->u.cli.cl_import;
1744 struct osp_device *osp;
1745 struct ptlrpc_request *req;
1749 if (KEY_IS(KEY_SPTLRPC_CONF)) {
1750 sptlrpc_conf_client_adapt(exp->exp_obd);
1754 LASSERT(set != NULL);
1755 if (!obd->obd_set_up || obd->obd_stopping)
1757 osp = lu2osp_dev(obd->obd_lu_dev);
1759 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
1763 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
1764 RCL_CLIENT, keylen);
1765 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
1766 RCL_CLIENT, vallen);
1767 if (osp->opd_connect_mdt)
1768 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SET_INFO);
1770 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
1772 ptlrpc_request_free(req);
1776 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1777 memcpy(tmp, key, keylen);
1778 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1779 memcpy(tmp, val, vallen);
1781 ptlrpc_request_set_replen(req);
1782 ptlrpc_set_add_req(set, req);
1783 ptlrpc_check_set(NULL, set);
1788 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1789 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1790 static void osp_key_exit(const struct lu_context *ctx,
1791 struct lu_context_key *key, void *data)
1793 struct osp_thread_info *info = data;
1795 info->osi_attr.la_valid = 0;
1798 struct lu_context_key osp_thread_key = {
1799 .lct_tags = LCT_MD_THREAD,
1800 .lct_init = osp_key_init,
1801 .lct_fini = osp_key_fini,
1802 .lct_exit = osp_key_exit
1805 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1806 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1808 struct lu_context_key osp_txn_key = {
1809 .lct_tags = LCT_OSP_THREAD,
1810 .lct_init = osp_txn_key_init,
1811 .lct_fini = osp_txn_key_fini
1813 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1815 static const struct lu_device_type_operations osp_device_type_ops = {
1816 .ldto_init = osp_type_init,
1817 .ldto_fini = osp_type_fini,
1819 .ldto_start = osp_type_start,
1820 .ldto_stop = osp_type_stop,
1822 .ldto_device_alloc = osp_device_alloc,
1823 .ldto_device_free = osp_device_free,
1825 .ldto_device_fini = osp_device_fini
1828 static struct lu_device_type osp_device_type = {
1829 .ldt_tags = LU_DEVICE_DT,
1830 .ldt_name = LUSTRE_OSP_NAME,
1831 .ldt_ops = &osp_device_type_ops,
1832 .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1835 static const struct obd_ops osp_obd_device_ops = {
1836 .o_owner = THIS_MODULE,
1837 .o_add_conn = client_import_add_conn,
1838 .o_del_conn = client_import_del_conn,
1839 .o_reconnect = osp_reconnect,
1840 .o_connect = osp_obd_connect,
1841 .o_disconnect = osp_obd_disconnect,
1842 .o_get_info = osp_obd_get_info,
1843 .o_set_info_async = osp_obd_set_info_async,
1844 .o_import_event = osp_import_event,
1845 .o_iocontrol = osp_iocontrol,
1846 .o_statfs = osp_obd_statfs,
1847 .o_fid_init = client_fid_init,
1848 .o_fid_fini = client_fid_fini,
1852 * Initialize OSP module.
1854 * Register device types OSP and Light Weight Proxy (LWP) (\see lwp_dev.c)
1855 * in obd_types (\see class_obd.c). Initialize procfs for the
1856 * the OSP device. Note: OSP was called OSC before Lustre 2.4,
1857 * so for compatibility it still uses the name "osc" in procfs.
1858 * This is called at module load time.
1860 * \retval 0 0 if initialization succeeds.
1861 * \retval negative negative errno if initialization failed.
1863 static int __init osp_init(void)
1865 struct obd_type *sym;
1868 rc = lu_kmem_init(osp_caches);
1872 rc = class_register_type(&osp_obd_device_ops, NULL, false,
1873 LUSTRE_OSP_NAME, &osp_device_type);
1875 lu_kmem_fini(osp_caches);
1879 rc = class_register_type(&lwp_obd_device_ops, NULL, false,
1880 LUSTRE_LWP_NAME, &lwp_device_type);
1882 class_unregister_type(LUSTRE_OSP_NAME);
1883 lu_kmem_fini(osp_caches);
1887 /* create "osc" entry for compatibility purposes */
1888 sym = class_add_symlinks(LUSTRE_OSC_NAME, true);
1891 /* does real "osc" already exist ? */
1900 * Finalize OSP module.
1902 * This callback is called when kernel unloads OSP module from memory, and
1903 * it will deregister OSP and LWP device type from obd_types (\see class_obd.c).
1905 static void __exit osp_exit(void)
1907 struct obd_type *sym = class_search_type(LUSTRE_OSC_NAME);
1909 /* if this was never fully initialized by the osc layer
1910 * then we are responsible for freeing this obd_type
1913 /* final put if we manage this obd type */
1914 if (sym->typ_sym_filter)
1915 kobject_put(&sym->typ_kobj);
1916 /* put reference taken by class_search_type */
1917 kobject_put(&sym->typ_kobj);
1920 class_unregister_type(LUSTRE_LWP_NAME);
1921 class_unregister_type(LUSTRE_OSP_NAME);
1922 lu_kmem_fini(osp_caches);
1925 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1926 MODULE_DESCRIPTION("Lustre OSD Storage Proxy ("LUSTRE_OSP_NAME")");
1927 MODULE_VERSION(LUSTRE_VERSION_STRING);
1928 MODULE_LICENSE("GPL");
1930 module_init(osp_init);
1931 module_exit(osp_exit);