4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osp/osp_dev.c
34 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
35 * Author: Mikhail Pershin <mike.pershin@intel.com>
36 * Author: Di Wang <di.wang@intel.com>
39 * The Object Storage Proxy (OSP) module provides an implementation of
40 * the DT API for remote MDTs and OSTs. Every local OSP device (or
41 * object) is a proxy for a remote OSD device (or object). Thus OSP
42 * converts DT operations into RPCs, which are sent to the OUT service
43 * on a remote target, converted back to DT operations, and
44 * executed. Of course there are many ways in which this description
45 * is inaccurate but it's a good enough mental model. OSP is used by
46 * the MDT stack in several ways:
48 * - OSP devices allocate FIDs for the stripe sub-objects of a striped
51 * - OSP objects represent the remote MDT and OST objects that are
52 * the stripes of a striped object.
54 * - OSP devices log, send, and track synchronous operations (setattr
55 * and unlink) to remote targets.
57 * - OSP objects are the bottom slice of the compound LU object
58 * representing a remote MDT object: MDT/MDD/LOD/OSP.
60 * - OSP objects are used by LFSCK to represent remote OST objects
61 * during the verification of MDT-OST consistency.
63 * - OSP devices batch idempotent requests (declare_attr_get() and
64 * declare_xattr_get()) to the remote target and cache their results.
66 * In addition the OSP layer implements a subset of the OBD device API
67 * to support being a client of a remote target, connecting to other
68 * layers, and FID allocation.
71 #define DEBUG_SUBSYSTEM S_MDS
73 #include <linux/kthread.h>
75 #include <uapi/linux/lustre/lustre_ioctl.h>
76 #include <lustre_log.h>
77 #include <lustre_obdo.h>
78 #include <uapi/linux/lustre/lustre_param.h>
79 #include <obd_class.h>
81 #include "osp_internal.h"
83 /* Slab for OSP object allocation */
84 struct kmem_cache *osp_object_kmem;
86 static struct lu_kmem_descr osp_caches[] = {
88 .ckd_cache = &osp_object_kmem,
89 .ckd_name = "osp_obj",
90 .ckd_size = sizeof(struct osp_object)
98 * Implementation of lu_device_operations::ldo_object_alloc
100 * Allocates an OSP object in memory, whose FID is on the remote target.
102 * \param[in] env execution environment
103 * \param[in] hdr The header of the object stack. If it is NULL, it
104 * means the object is not built from top device, i.e.
105 * it is a sub-stripe object of striped directory or
107 * \param[in] d OSP device
109 * \retval object object being created if the creation succeed.
110 * \retval NULL NULL if the creation failed.
112 static struct lu_object *osp_object_alloc(const struct lu_env *env,
113 const struct lu_object_header *hdr,
116 struct lu_object_header *h = NULL;
117 struct osp_object *o;
120 OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, GFP_NOFS);
122 l = &o->opo_obj.do_lu;
124 /* If hdr is NULL, it means the object is not built
125 * from the top dev(MDT/OST), usually it happens when
126 * building striped object, like data object on MDT or
127 * striped object for directory */
130 lu_object_header_init(h);
131 dt_object_init(&o->opo_obj, h, d);
132 lu_object_add_top(h, l);
134 dt_object_init(&o->opo_obj, h, d);
137 l->lo_ops = &osp_lu_obj_ops;
146 * Find or create the local object
148 * Finds or creates the local file referenced by \a reg_id and return the
149 * attributes of the local file.
151 * \param[in] env execution environment
152 * \param[in] osp OSP device
153 * \param[out] attr attributes of the object
154 * \param[in] reg_id the local object ID of the file. It will be used
155 * to compose a local FID{FID_SEQ_LOCAL_FILE, reg_id, 0}
156 * to identify the object.
158 * \retval object object(dt_object) found or created
159 * \retval ERR_PTR(errno) ERR_PTR(errno) if not get the object.
161 static struct dt_object
162 *osp_find_or_create_local_file(const struct lu_env *env, struct osp_device *osp,
163 struct lu_attr *attr, __u32 reg_id)
165 struct osp_thread_info *osi = osp_env_info(env);
166 struct dt_object_format dof = { 0 };
167 struct dt_object *dto;
171 lu_local_obj_fid(&osi->osi_fid, reg_id);
172 attr->la_valid = LA_MODE;
173 attr->la_mode = S_IFREG | 0644;
174 dof.dof_type = DFT_REGULAR;
175 /* Find or create the local object by osi_fid. */
176 dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
181 /* Get attributes of the local object. */
182 rc = dt_attr_get(env, dto, attr);
184 CERROR("%s: can't be initialized: rc = %d\n",
185 osp->opd_obd->obd_name, rc);
186 dt_object_put(env, dto);
193 * Write data buffer to a local file object.
195 * \param[in] env execution environment
196 * \param[in] osp OSP device
197 * \param[in] dt_obj object written to
198 * \param[in] buf buffer containing byte array and length
199 * \param[in] offset write offset in the object in bytes
201 * \retval 0 0 if write succeed
202 * \retval -EFAULT -EFAULT if only part of buffer is written.
203 * \retval negative other negative errno if write failed.
205 static int osp_write_local_file(const struct lu_env *env,
206 struct osp_device *osp,
207 struct dt_object *dt_obj,
214 if (osp->opd_storage->dd_rdonly)
217 th = dt_trans_create(env, osp->opd_storage);
221 rc = dt_declare_record_write(env, dt_obj, buf, offset, th);
224 rc = dt_trans_start_local(env, osp->opd_storage, th);
228 rc = dt_record_write(env, dt_obj, buf, &offset, th);
230 dt_trans_stop(env, osp->opd_storage, th);
235 * Initialize last ID object.
237 * This function initializes the LAST_ID file, which stores the current last
238 * used id of data objects. The MDT will use the last used id and the last_seq
239 * (\see osp_init_last_seq()) to synchronize the precreate object cache with
242 * \param[in] env execution environment
243 * \param[in] osp OSP device
245 * \retval 0 0 if initialization succeed
246 * \retval negative negative errno if initialization failed
248 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
250 struct osp_thread_info *osi = osp_env_info(env);
251 struct lu_fid *fid = &osp->opd_last_used_fid;
252 struct dt_object *dto;
256 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
259 RETURN(PTR_ERR(dto));
261 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &osp->opd_last_id,
264 /* object will be released in device cleanup path */
265 if (osi->osi_attr.la_size >= (osi->osi_off + osi->osi_lb.lb_len)) {
266 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
267 if (rc != 0 && rc != -EFAULT)
269 /* In case of idif bits 32-48 go to f_seq
270 * (see osp_init_last_seq). So don't care
271 * about u64->u32 convertion. */
272 fid->f_oid = osp->opd_last_id;
275 if (rc == -EFAULT) { /* fresh LAST_ID */
276 osp->opd_last_id = 0;
278 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
283 osp->opd_last_used_oid_file = dto;
286 /* object will be released in device cleanup path */
287 CERROR("%s: can't initialize lov_objid: rc = %d\n",
288 osp->opd_obd->obd_name, rc);
289 dt_object_put(env, dto);
290 osp->opd_last_used_oid_file = NULL;
295 * Initialize last sequence object.
297 * This function initializes the LAST_SEQ file in the local OSD, which stores
298 * the current last used sequence of data objects. The MDT will use the last
299 * sequence and last id (\see osp_init_last_objid()) to synchronize the
300 * precreate object cache with OSTs.
302 * \param[in] env execution environment
303 * \param[in] osp OSP device
305 * \retval 0 0 if initialization succeed
306 * \retval negative negative errno if initialization failed
308 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
310 struct osp_thread_info *osi = osp_env_info(env);
311 struct lu_fid *fid = &osp->opd_last_used_fid;
312 struct dt_object *dto;
316 dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
319 RETURN(PTR_ERR(dto));
321 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
324 /* object will be released in device cleanup path */
325 if (osi->osi_attr.la_size >= (osi->osi_off + osi->osi_lb.lb_len)) {
326 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
327 if (rc != 0 && rc != -EFAULT)
329 if (fid_is_idif(fid))
330 fid->f_seq = fid_idif_seq(osp->opd_last_id,
334 if (rc == -EFAULT) { /* fresh OSP */
336 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
341 osp->opd_last_used_seq_file = dto;
344 /* object will be released in device cleanup path */
345 CERROR("%s: can't initialize lov_seq: rc = %d\n",
346 osp->opd_obd->obd_name, rc);
347 dt_object_put(env, dto);
348 osp->opd_last_used_seq_file = NULL;
353 * Initialize last OID and sequence object.
355 * If the MDT is just upgraded to 2.4 from the lower version, where the
356 * LAST_SEQ file does not exist, the file will be created and IDIF sequence
357 * will be written into the file.
359 * \param[in] env execution environment
360 * \param[in] osp OSP device
362 * \retval 0 0 if initialization succeed
363 * \retval negative negative error if initialization failed
365 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
367 struct osp_thread_info *osi = osp_env_info(env);
371 fid_zero(&osp->opd_last_used_fid);
372 rc = osp_init_last_objid(env, osp);
374 CERROR("%s: Can not get ids %d from old objid!\n",
375 osp->opd_obd->obd_name, rc);
379 rc = osp_init_last_seq(env, osp);
381 CERROR("%s: Can not get sequence %d from old objseq!\n",
382 osp->opd_obd->obd_name, rc);
386 if (fid_oid(&osp->opd_last_used_fid) != 0 &&
387 fid_seq(&osp->opd_last_used_fid) == 0) {
388 /* Just upgrade from the old version,
389 * set the seq to be IDIF */
390 osp->opd_last_used_fid.f_seq =
391 fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
393 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
394 &osp->opd_last_used_fid.f_seq,
396 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
397 &osi->osi_lb, osi->osi_off);
399 CERROR("%s : Can not write seq file: rc = %d\n",
400 osp->opd_obd->obd_name, rc);
405 if (!fid_is_zero(&osp->opd_last_used_fid) &&
406 !fid_is_sane(&osp->opd_last_used_fid)) {
407 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
408 PFID(&osp->opd_last_used_fid));
409 GOTO(out, rc = -EINVAL);
412 osp_fid_to_obdid(&osp->opd_last_used_fid, &osp->opd_last_id);
413 CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
414 osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
417 if (osp->opd_last_used_oid_file != NULL) {
418 dt_object_put(env, osp->opd_last_used_oid_file);
419 osp->opd_last_used_oid_file = NULL;
421 if (osp->opd_last_used_seq_file != NULL) {
422 dt_object_put(env, osp->opd_last_used_seq_file);
423 osp->opd_last_used_seq_file = NULL;
431 * Release the last sequence and OID file objects in OSP device.
433 * \param[in] env execution environment
434 * \param[in] osp OSP device
436 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *osp)
438 /* release last_used file */
439 if (osp->opd_last_used_oid_file != NULL) {
440 dt_object_put(env, osp->opd_last_used_oid_file);
441 osp->opd_last_used_oid_file = NULL;
444 if (osp->opd_last_used_seq_file != NULL) {
445 dt_object_put(env, osp->opd_last_used_seq_file);
446 osp->opd_last_used_seq_file = NULL;
451 * Disconnects the connection between OSP and its correspondent MDT or OST, and
452 * the import will be marked as inactive. It will only be called during OSP
455 * \param[in] d OSP device being disconnected
457 * \retval 0 0 if disconnection succeed
458 * \retval negative negative errno if disconnection failed
460 static int osp_disconnect(struct osp_device *d)
462 struct obd_device *obd = d->opd_obd;
463 struct obd_import *imp;
466 imp = obd->u.cli.cl_import;
468 /* Mark import deactivated now, so we don't try to reconnect if any
469 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
470 * fully deactivate the import, or that would drop all requests. */
471 LASSERT(imp != NULL);
472 spin_lock(&imp->imp_lock);
473 imp->imp_deactive = 1;
474 spin_unlock(&imp->imp_lock);
476 ptlrpc_deactivate_import(imp);
478 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
479 * delete it regardless. (It's safe to delete an import that was
481 (void)ptlrpc_pinger_del_import(imp);
483 rc = ptlrpc_disconnect_import(imp, 0);
485 CERROR("%s: can't disconnect: rc = %d\n", obd->obd_name, rc);
487 ptlrpc_invalidate_import(imp);
493 * Initialize the osp_update structure in OSP device
495 * Allocate osp update structure and start update thread.
497 * \param[in] osp OSP device
499 * \retval 0 if initialization succeeds.
500 * \retval negative errno if initialization fails.
502 static int osp_update_init(struct osp_device *osp)
504 struct task_struct *task;
509 LASSERT(osp->opd_connect_mdt);
511 if (osp->opd_storage->dd_rdonly)
514 OBD_ALLOC_PTR(osp->opd_update);
515 if (osp->opd_update == NULL)
518 init_waitqueue_head(&osp->opd_update->ou_waitq);
519 spin_lock_init(&osp->opd_update->ou_lock);
520 INIT_LIST_HEAD(&osp->opd_update->ou_list);
521 osp->opd_update->ou_rpc_version = 1;
522 osp->opd_update->ou_version = 1;
523 osp->opd_update->ou_generation = 0;
525 rc = lu_env_init(&osp->opd_update->ou_env,
526 osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
528 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
530 OBD_FREE_PTR(osp->opd_update);
531 osp->opd_update = NULL;
534 /* start thread handling sending updates to the remote MDT */
535 task = kthread_create(osp_send_update_thread, osp,
536 "osp_up%u-%u", osp->opd_index, osp->opd_group);
538 int rc = PTR_ERR(task);
540 lu_env_fini(&osp->opd_update->ou_env);
541 OBD_FREE_PTR(osp->opd_update);
542 osp->opd_update = NULL;
543 CERROR("%s: can't start precreate thread: rc = %d\n",
544 osp->opd_obd->obd_name, rc);
548 osp->opd_update->ou_update_task = task;
549 wake_up_process(task);
555 * Finialize osp_update structure in OSP device
557 * Stop the OSP update sending thread, then delete the left
558 * osp thandle in the sending list.
560 * \param [in] osp OSP device.
562 static void osp_update_fini(const struct lu_env *env, struct osp_device *osp)
564 struct osp_update_request *our;
565 struct osp_update_request *tmp;
566 struct osp_updates *ou = osp->opd_update;
571 kthread_stop(ou->ou_update_task);
572 lu_env_fini(&ou->ou_env);
574 /* Remove the left osp thandle from the list */
575 spin_lock(&ou->ou_lock);
576 list_for_each_entry_safe(our, tmp, &ou->ou_list,
578 list_del_init(&our->our_list);
579 LASSERT(our->our_th != NULL);
580 osp_trans_callback(env, our->our_th, -EIO);
581 /* our will be destroyed in osp_thandle_put() */
582 osp_thandle_put(env, our->our_th);
584 spin_unlock(&ou->ou_lock);
587 osp->opd_update = NULL;
591 * Cleanup OSP, which includes disconnect import, cleanup unlink log, stop
592 * precreate threads etc.
594 * \param[in] env execution environment.
595 * \param[in] d OSP device being disconnected.
597 * \retval 0 0 if cleanup succeed
598 * \retval negative negative errno if cleanup failed
600 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
607 rc = osp_disconnect(d);
611 if (!d->opd_connect_mdt) {
612 /* stop sync thread */
615 /* stop precreate thread */
616 osp_precreate_fini(d);
618 /* release last_used file */
619 osp_last_used_fini(env, d);
622 obd_fid_fini(d->opd_obd);
628 * Implementation of osp_lu_ops::ldo_process_config
630 * This function processes config log records in OSP layer. It is usually
631 * called from the top layer of MDT stack, and goes through the stack by calling
632 * ldo_process_config of next layer.
634 * \param[in] env execution environment
635 * \param[in] dev lu_device of OSP
636 * \param[in] lcfg config log
638 * \retval 0 0 if the config log record is executed correctly.
639 * \retval negative negative errno if the record execution fails.
641 static int osp_process_config(const struct lu_env *env,
642 struct lu_device *dev, struct lustre_cfg *lcfg)
644 struct osp_device *d = lu2osp_dev(dev);
645 struct dt_device *dt = lu2dt_dev(dev);
646 struct obd_device *obd = d->opd_obd;
652 switch (lcfg->lcfg_command) {
653 case LCFG_PRE_CLEANUP:
654 rc = osp_disconnect(d);
655 osp_update_fini(env, d);
656 if (obd->obd_namespace != NULL)
657 ldlm_namespace_free_prior(obd->obd_namespace, NULL, 1);
660 lu_dev_del_linkage(dev->ld_site, dev);
661 rc = osp_shutdown(env, d);
664 count = class_modify_config(lcfg, d->opd_connect_mdt ?
665 PARAM_OSP : PARAM_OSC,
668 /* class_modify_config() haven't found matching
669 * parameter and returned an error so that layer(s)
670 * below could use that. But OSP is the bottom, so
673 CERROR("%s: unknown param %s\n",
674 (char *)lustre_cfg_string(lcfg, 0),
675 (char *)lustre_cfg_string(lcfg, 1));
680 CERROR("%s: unknown command %u\n",
681 (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
690 * Implementation of osp_lu_ops::ldo_recovery_complete
692 * This function is called after recovery is finished, and OSP layer
693 * will wake up precreate thread here.
695 * \param[in] env execution environment
696 * \param[in] dev lu_device of OSP
698 * \retval 0 0 unconditionally
700 static int osp_recovery_complete(const struct lu_env *env,
701 struct lu_device *dev)
703 struct osp_device *osp = lu2osp_dev(dev);
706 osp->opd_recovery_completed = 1;
708 if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
709 wake_up(&osp->opd_pre_waitq);
714 const struct lu_device_operations osp_lu_ops = {
715 .ldo_object_alloc = osp_object_alloc,
716 .ldo_process_config = osp_process_config,
717 .ldo_recovery_complete = osp_recovery_complete,
721 * Implementation of dt_device_operations::dt_statfs
723 * This function provides statfs status (for precreation) from
724 * corresponding OST. Note: this function only retrieves the status
725 * from the OSP device, and the real statfs RPC happens inside
726 * precreate thread (\see osp_statfs_update). Note: OSP for MDT does
727 * not need to retrieve statfs data for now.
729 * \param[in] env execution environment.
730 * \param[in] dev dt_device of OSP.
731 * \param[out] sfs holds the retrieved statfs data.
733 * \retval 0 0 statfs data was retrieved successfully or
734 * retrieval was not needed
735 * \retval negative negative errno if get statfs failed.
737 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
738 struct obd_statfs *sfs, struct obd_statfs_info *info)
740 struct osp_device *d = dt2osp_dev(dev);
741 struct obd_import *imp = d->opd_obd->u.cli.cl_import;
745 if (imp->imp_state == LUSTRE_IMP_CLOSED)
748 if (unlikely(d->opd_imp_active == 0))
751 /* return recently updated data */
752 *sfs = d->opd_statfs;
754 info->os_reserved_mb_low = d->opd_reserved_mb_low;
755 info->os_reserved_mb_high = d->opd_reserved_mb_high;
758 if (d->opd_pre == NULL)
762 "%s: %llu blocks, %llu free, %llu avail, %u reserved mb low, %u reserved mb high, %llu files, %llu free files\n",
763 d->opd_obd->obd_name,
764 sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
765 d->opd_reserved_mb_low, d->opd_reserved_mb_high,
766 sfs->os_files, sfs->os_ffree);
769 if (info && !info->os_enable_pre)
773 * The layer above osp (usually lod) can use f_precreated to
774 * estimate how many objects are available for immediate usage.
776 spin_lock(&d->opd_pre_lock);
777 sfs->os_fprecreated = osp_fid_diff(&d->opd_pre_last_created_fid,
778 &d->opd_pre_used_fid);
779 sfs->os_fprecreated -= d->opd_pre_reserved;
780 LASSERTF(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2,
781 "last_created "DFID", next_fid "DFID", reserved %llu\n",
782 PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_pre_used_fid),
783 d->opd_pre_reserved);
784 spin_unlock(&d->opd_pre_lock);
789 * Implementation of dt_device_operations::dt_sync
791 * This function synchronizes the OSP cache to the remote target. It wakes
792 * up unlink log threads and sends out unlink records to the remote OST.
794 * \param[in] env execution environment
795 * \param[in] dev dt_device of OSP
797 * \retval 0 0 if synchronization succeeds
798 * \retval negative negative errno if synchronization fails
800 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
802 struct osp_device *d = dt2osp_dev(dev);
803 time64_t start = ktime_get_seconds();
809 /* No Sync between MDTs yet. */
810 if (d->opd_connect_mdt)
813 recs = atomic_read(&d->opd_sync_changes);
814 old = atomic64_read(&d->opd_sync_processed_recs);
816 osp_sync_force(env, dt2osp_dev(dev));
818 if (unlikely(d->opd_imp_active == 0))
821 down_write(&d->opd_async_updates_rwsem);
823 CDEBUG(D_OTHER, "%s: async updates %d\n", d->opd_obd->obd_name,
824 atomic_read(&d->opd_async_updates_count));
826 /* make sure the connection is fine */
827 rc = wait_event_idle_timeout(
828 d->opd_sync_barrier_waitq,
829 atomic_read(&d->opd_async_updates_count) == 0,
830 cfs_time_seconds(obd_timeout));
836 up_write(&d->opd_async_updates_rwsem);
840 CDEBUG(D_CACHE, "%s: processed %llu\n", d->opd_obd->obd_name,
841 (unsigned long long)atomic64_read(&d->opd_sync_processed_recs));
843 while (atomic64_read(&d->opd_sync_processed_recs) < old + recs) {
844 __u64 last = atomic64_read(&d->opd_sync_processed_recs);
845 /* make sure the connection is fine */
846 wait_event_idle_timeout(
847 d->opd_sync_barrier_waitq,
848 atomic64_read(&d->opd_sync_processed_recs)
850 cfs_time_seconds(obd_timeout));
852 if (atomic64_read(&d->opd_sync_processed_recs) >= old + recs)
855 if (atomic64_read(&d->opd_sync_processed_recs) != last) {
856 /* some progress have been made,
861 /* no changes and expired, something is wrong */
862 GOTO(out, rc = -ETIMEDOUT);
865 /* block new processing (barrier>0 - few callers are possible */
866 atomic_inc(&d->opd_sync_barrier);
868 CDEBUG(D_CACHE, "%s: %u in flight\n", d->opd_obd->obd_name,
869 atomic_read(&d->opd_sync_rpcs_in_flight));
871 /* wait till all-in-flight are replied, so executed by the target */
872 /* XXX: this is used by LFSCK at the moment, which doesn't require
873 * all the changes to be committed, but in general it'd be
874 * better to wait till commit */
875 while (atomic_read(&d->opd_sync_rpcs_in_flight) > 0) {
876 old = atomic_read(&d->opd_sync_rpcs_in_flight);
878 wait_event_idle_timeout(
879 d->opd_sync_barrier_waitq,
880 atomic_read(&d->opd_sync_rpcs_in_flight) == 0,
881 cfs_time_seconds(obd_timeout));
883 if (atomic_read(&d->opd_sync_rpcs_in_flight) == 0)
886 if (atomic_read(&d->opd_sync_rpcs_in_flight) != old) {
887 /* some progress have been made */
891 /* no changes and expired, something is wrong */
892 GOTO(out, rc = -ETIMEDOUT);
896 /* resume normal processing (barrier=0) */
897 atomic_dec(&d->opd_sync_barrier);
898 osp_sync_check_for_work(d);
900 CDEBUG(D_CACHE, "%s: done in %lld: rc = %d\n", d->opd_obd->obd_name,
901 ktime_get_seconds() - start, rc);
906 const struct dt_device_operations osp_dt_ops = {
907 .dt_statfs = osp_statfs,
909 .dt_trans_create = osp_trans_create,
910 .dt_trans_start = osp_trans_start,
911 .dt_trans_stop = osp_trans_stop,
912 .dt_trans_cb_add = osp_trans_cb_add,
916 * Connect OSP to local OSD.
918 * Locate the local OSD referenced by \a nextdev and connect to it. Sometimes,
919 * OSP needs to access the local OSD to store some information. For example,
920 * during precreate, it needs to update last used OID and sequence file
921 * (LAST_SEQ) in local OSD.
923 * \param[in] env execution environment
924 * \param[in] osp OSP device
925 * \param[in] nextdev the name of local OSD
927 * \retval 0 0 connection succeeded
928 * \retval negative negative errno connection failed
930 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *osp,
933 struct obd_connect_data *data = NULL;
934 struct obd_device *obd;
939 LASSERT(osp->opd_storage_exp == NULL);
945 obd = class_name2obd(nextdev);
947 CERROR("%s: can't locate next device: %s\n",
948 osp->opd_obd->obd_name, nextdev);
949 GOTO(out, rc = -ENOTCONN);
952 rc = obd_connect(env, &osp->opd_storage_exp, obd, &obd->obd_uuid, data,
955 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
956 osp->opd_obd->obd_name, nextdev, rc);
960 osp->opd_dt_dev.dd_lu_dev.ld_site =
961 osp->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
962 LASSERT(osp->opd_dt_dev.dd_lu_dev.ld_site);
963 osp->opd_storage = lu2dt_dev(osp->opd_storage_exp->exp_obd->obd_lu_dev);
971 * Determine if the lock needs to be cancelled
973 * Determine if the unused lock should be cancelled before replay, see
974 * (ldlm_cancel_no_wait_policy()). Currently, only inode bits lock exists
977 * \param[in] lock lock to be checked.
979 * \retval 1 if the lock needs to be cancelled before replay.
980 * \retval 0 if the lock does not need to be cancelled before
983 static int osp_cancel_weight(struct ldlm_lock *lock)
985 if (lock->l_resource->lr_type != LDLM_IBITS)
992 * Initialize OSP device according to the parameters in the configuration
995 * Reconstruct the local device name from the configuration profile, and
996 * initialize necessary threads and structures according to the OSP type
999 * Since there is no record in the MDT configuration for the local disk
1000 * device, we have to extract this from elsewhere in the profile.
1001 * The only information we get at setup is from the OSC records:
1002 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
1004 * Note: configs generated by Lustre 1.8 are missing the -MDTxxxx part,
1005 * so, we need to reconstruct the name of the underlying OSD from this:
1006 * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".
1008 * \param[in] env execution environment
1009 * \param[in] osp OSP device
1010 * \param[in] ldt lu device type of OSP
1011 * \param[in] cfg configuration log
1013 * \retval 0 0 if OSP initialization succeeded.
1014 * \retval negative negative errno if OSP initialization failed.
1016 static int osp_init0(const struct lu_env *env, struct osp_device *osp,
1017 struct lu_device_type *ldt, struct lustre_cfg *cfg)
1019 struct obd_device *obd;
1020 struct obd_import *imp;
1021 char *src, *tgt, *osdname = NULL;
1028 mutex_init(&osp->opd_async_requests_mutex);
1029 INIT_LIST_HEAD(&osp->opd_async_updates);
1030 init_rwsem(&osp->opd_async_updates_rwsem);
1031 atomic_set(&osp->opd_async_updates_count, 0);
1033 obd = class_name2obd(lustre_cfg_string(cfg, 0));
1035 CERROR("Cannot find obd with name %s\n",
1036 lustre_cfg_string(cfg, 0));
1041 src = lustre_cfg_string(cfg, 0);
1045 tgt = strrchr(src, '-');
1047 CERROR("%s: invalid target name %s: rc = %d\n",
1048 osp->opd_obd->obd_name, lustre_cfg_string(cfg, 0),
1053 if (strncmp(tgt, "-osc", 4) == 0) {
1054 /* Old OSC name fsname-OSTXXXX-osc */
1055 for (tgt--; tgt > src && *tgt != '-'; tgt--)
1058 CERROR("%s: invalid target name %s: rc = %d\n",
1059 osp->opd_obd->obd_name,
1060 lustre_cfg_string(cfg, 0), -EINVAL);
1064 if (strncmp(tgt, "-OST", 4) != 0) {
1065 CERROR("%s: invalid target name %s: rc = %d\n",
1066 osp->opd_obd->obd_name,
1067 lustre_cfg_string(cfg, 0), -EINVAL);
1071 rc = target_name2index(tgt + 1, &idx, &mdt);
1072 if (rc < 0 || rc & LDD_F_SV_ALL || mdt[0] != '-') {
1073 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1074 osp->opd_obd->obd_name, src, -EINVAL);
1077 osp->opd_index = idx;
1081 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
1082 if (strncmp(tgt, "-MDT", 4) != 0 &&
1083 strncmp(tgt, "-OST", 4) != 0) {
1084 CERROR("%s: invalid target name %s: rc = %d\n",
1085 osp->opd_obd->obd_name,
1086 lustre_cfg_string(cfg, 0), -EINVAL);
1090 rc = target_name2index(tgt + 1, &idx, &mdt);
1091 if (rc < 0 || rc & LDD_F_SV_ALL || *mdt != '\0') {
1092 CERROR("%s: invalid OST index in '%s': rc = %d\n",
1093 osp->opd_obd->obd_name, src, -EINVAL);
1097 /* Get MDT index from the name and set it to opd_group,
1098 * which will be used by OSP to connect with OST */
1099 osp->opd_group = idx;
1100 if (tgt - src <= 12) {
1101 CERROR("%s: invalid mdt index from %s: rc =%d\n",
1102 osp->opd_obd->obd_name,
1103 lustre_cfg_string(cfg, 0), -EINVAL);
1107 if (strncmp(tgt - 12, "-MDT", 4) == 0)
1108 osp->opd_connect_mdt = 1;
1110 rc = target_name2index(tgt - 11, &idx, &mdt);
1111 if (rc < 0 || rc & LDD_F_SV_ALL || mdt[0] != '-') {
1112 CERROR("%s: invalid OST index in '%s': rc =%d\n",
1113 osp->opd_obd->obd_name, src, -EINVAL);
1117 osp->opd_index = idx;
1118 idx = tgt - src - 12;
1120 /* check the fsname length, and after this everything else will fit */
1121 if (idx > MTI_NAME_MAXLEN) {
1122 CERROR("%s: fsname too long in '%s': rc = %d\n",
1123 osp->opd_obd->obd_name, src, -EINVAL);
1127 OBD_ALLOC(osdname, MAX_OBD_NAME);
1128 if (osdname == NULL)
1131 memcpy(osdname, src, idx); /* copy just the fsname part */
1132 osdname[idx] = '\0';
1134 mdt = strstr(mdt, "-MDT");
1135 if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
1136 strcat(osdname, "-MDT0000");
1138 strcat(osdname, mdt);
1139 strcat(osdname, "-osd");
1140 CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
1142 osp_init_rpc_lock(osp);
1144 osp->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
1145 osp->opd_dt_dev.dd_ops = &osp_dt_ops;
1147 obd->obd_lu_dev = &osp->opd_dt_dev.dd_lu_dev;
1149 rc = osp_connect_to_osd(env, osp, osdname);
1153 rc = ptlrpcd_addref();
1155 GOTO(out_disconnect, rc);
1157 rc = client_obd_setup(obd, cfg);
1159 CERROR("%s: can't setup obd: rc = %d\n", osp->opd_obd->obd_name,
1164 osp_tunables_init(osp);
1166 rc = obd_fid_init(osp->opd_obd, NULL, osp->opd_connect_mdt ?
1167 LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
1169 CERROR("%s: fid init error: rc = %d\n",
1170 osp->opd_obd->obd_name, rc);
1174 if (!osp->opd_connect_mdt) {
1175 /* Initialize last id from the storage - will be
1176 * used in orphan cleanup. */
1177 if (!osp->opd_storage->dd_rdonly) {
1178 rc = osp_last_used_init(env, osp);
1183 /* Initialize precreation thread, it handles new
1184 * connections as well. */
1185 rc = osp_init_precreate(osp);
1187 GOTO(out_last_used, rc);
1190 * Initialize synhronization mechanism taking
1191 * care of propogating changes to OST in near
1192 * transactional manner.
1194 rc = osp_sync_init(env, osp);
1196 GOTO(out_precreat, rc);
1198 osp->opd_got_disconnected = 1;
1199 rc = osp_update_init(osp);
1204 rc = osp_init_statfs(osp);
1206 GOTO(out_precreat, rc);
1208 ns_register_cancel(obd->obd_namespace, osp_cancel_weight);
1211 * Initiate connect to OST
1213 imp = obd->u.cli.cl_import;
1215 rc = ptlrpc_init_import(imp);
1219 OBD_FREE(osdname, MAX_OBD_NAME);
1223 if (!osp->opd_connect_mdt)
1224 /* stop sync thread */
1227 /* stop precreate thread */
1228 if (!osp->opd_connect_mdt)
1229 osp_precreate_fini(osp);
1231 osp_update_fini(env, osp);
1233 if (!osp->opd_connect_mdt)
1234 osp_last_used_fini(env, osp);
1236 obd_fid_fini(osp->opd_obd);
1238 osp_tunables_fini(osp);
1239 client_obd_cleanup(obd);
1243 obd_disconnect(osp->opd_storage_exp);
1246 OBD_FREE(osdname, MAX_OBD_NAME);
1251 * Implementation of lu_device_type_operations::ldto_device_free
1253 * Free the OSP device in memory. No return value is needed for now,
1254 * so always return NULL to comply with the interface.
1256 * \param[in] env execution environment
1257 * \param[in] lu lu_device of OSP
1259 * \retval NULL NULL unconditionally
1261 static struct lu_device *osp_device_free(const struct lu_env *env,
1262 struct lu_device *lu)
1264 struct osp_device *osp = lu2osp_dev(lu);
1266 if (atomic_read(&lu->ld_ref) && lu->ld_site) {
1267 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1268 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
1270 dt_device_fini(&osp->opd_dt_dev);
1277 * Implementation of lu_device_type_operations::ldto_device_alloc
1279 * This function allocates and initializes OSP device in memory according to
1282 * \param[in] env execution environment
1283 * \param[in] type device type of OSP
1284 * \param[in] lcfg config log
1286 * \retval pointer the pointer of allocated OSP if succeed.
1287 * \retval ERR_PTR(errno) ERR_PTR(errno) if failed.
1289 static struct lu_device *osp_device_alloc(const struct lu_env *env,
1290 struct lu_device_type *type,
1291 struct lustre_cfg *lcfg)
1293 struct osp_device *osp;
1294 struct lu_device *ld;
1298 ld = ERR_PTR(-ENOMEM);
1302 ld = osp2lu_dev(osp);
1303 dt_device_init(&osp->opd_dt_dev, type);
1304 rc = osp_init0(env, osp, type, lcfg);
1306 osp_device_free(env, ld);
1314 * Implementation of lu_device_type_operations::ldto_device_fini
1316 * This function cleans up the OSP device, i.e. release and free those
1317 * attached items in osp_device.
1319 * \param[in] env execution environment
1320 * \param[in] ld lu_device of OSP
1322 * \retval NULL NULL if cleanup succeeded.
1323 * \retval ERR_PTR(errno) ERR_PTR(errno) if cleanup failed.
1325 static struct lu_device *osp_device_fini(const struct lu_env *env,
1326 struct lu_device *ld)
1328 struct osp_device *osp = lu2osp_dev(ld);
1333 if (osp->opd_async_requests != NULL) {
1334 osp_update_request_destroy(env, osp->opd_async_requests);
1335 osp->opd_async_requests = NULL;
1338 if (osp->opd_storage_exp) {
1339 /* wait for the commit callbacks to complete */
1340 wait_event(osp->opd_sync_waitq,
1341 atomic_read(&osp->opd_commits_registered) == 0);
1342 obd_disconnect(osp->opd_storage_exp);
1345 LASSERT(osp->opd_obd);
1347 rc = client_obd_cleanup(osp->opd_obd);
1350 RETURN(ERR_PTR(rc));
1353 osp_tunables_fini(osp);
1361 * Implementation of obd_ops::o_reconnect
1363 * This function is empty and does not need to do anything for now.
1365 static int osp_reconnect(const struct lu_env *env,
1366 struct obd_export *exp, struct obd_device *obd,
1367 struct obd_uuid *cluuid,
1368 struct obd_connect_data *data,
1375 * Implementation of obd_ops::o_connect
1377 * Connect OSP to the remote target (MDT or OST). Allocate the
1378 * export and return it to the LOD, which calls this function
1379 * for each OSP to connect it to the remote target. This function
1380 * is currently only called once per OSP.
1382 * \param[in] env execution environment
1383 * \param[out] exp export connected to OSP
1384 * \param[in] obd OSP device
1385 * \param[in] cluuid OSP device client uuid
1386 * \param[in] data connect_data to be used to connect to the remote
1388 * \param[in] localdata necessary for the API interface, but not used in
1391 * \retval 0 0 if the connection succeeded.
1392 * \retval negative negative errno if the connection failed.
1394 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
1395 struct obd_device *obd, struct obd_uuid *cluuid,
1396 struct obd_connect_data *data, void *localdata)
1398 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1399 struct obd_connect_data *ocd;
1400 struct obd_import *imp;
1401 struct lustre_handle conn;
1406 CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
1408 rc = class_connect(&conn, obd, cluuid);
1412 *exp = class_conn2export(&conn);
1413 /* Why should there ever be more than 1 connect? */
1414 osp->opd_connects++;
1415 LASSERT(osp->opd_connects == 1);
1417 osp->opd_exp = *exp;
1419 imp = osp->opd_obd->u.cli.cl_import;
1420 imp->imp_dlm_handle = conn;
1422 LASSERT(data != NULL);
1423 LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
1424 ocd = &imp->imp_connect_data;
1427 imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
1428 imp->imp_connect_flags2_orig = ocd->ocd_connect_flags2;
1430 ocd->ocd_version = LUSTRE_VERSION_CODE;
1431 ocd->ocd_index = data->ocd_index;
1433 rc = ptlrpc_connect_import(imp);
1435 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
1438 osp->opd_obd->u.cli.cl_seq->lcs_exp =
1439 class_export_get(osp->opd_exp);
1442 ptlrpc_pinger_add_import(imp);
1448 * Implementation of obd_ops::o_disconnect
1450 * Disconnect the export for the OSP. This is called by LOD to release the
1451 * OSP during cleanup (\see lod_del_device()). The OSP will be released after
1452 * the export is released.
1454 * \param[in] exp export to be disconnected.
1456 * \retval 0 0 if disconnection succeed
1457 * \retval negative negative errno if disconnection failed
1459 static int osp_obd_disconnect(struct obd_export *exp)
1461 struct obd_device *obd = exp->exp_obd;
1462 struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1466 /* Only disconnect the underlying layers on the final disconnect. */
1467 LASSERT(osp->opd_connects == 1);
1468 osp->opd_connects--;
1470 rc = class_disconnect(exp);
1472 CERROR("%s: class disconnect error: rc = %d\n",
1477 /* destroy the device */
1478 class_manual_cleanup(obd);
1484 * Implementation of obd_ops::o_statfs
1486 * Send a RPC to the remote target to get statfs status. This is only used
1487 * in lprocfs helpers by obd_statfs.
1489 * \param[in] env execution environment
1490 * \param[in] exp connection state from this OSP to the parent (LOD)
1492 * \param[out] osfs hold the statfs result
1493 * \param[in] unused Not used in this function for now
1494 * \param[in] flags flags to indicate how OSP will issue the RPC
1496 * \retval 0 0 if statfs succeeded.
1497 * \retval negative negative errno if statfs failed.
1499 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
1500 struct obd_statfs *osfs, time64_t unused, __u32 flags)
1502 struct obd_statfs *msfs;
1503 struct ptlrpc_request *req;
1504 struct obd_import *imp = NULL;
1509 /* Since the request might also come from lprocfs, so we need
1510 * sync this with client_disconnect_export Bug15684 */
1511 down_read(&exp->exp_obd->u.cli.cl_sem);
1512 if (exp->exp_obd->u.cli.cl_import)
1513 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
1514 up_read(&exp->exp_obd->u.cli.cl_sem);
1518 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1520 class_import_put(imp);
1525 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1527 ptlrpc_request_free(req);
1530 ptlrpc_request_set_replen(req);
1531 req->rq_request_portal = OST_CREATE_PORTAL;
1532 ptlrpc_at_set_req_timeout(req);
1534 if (flags & OBD_STATFS_NODELAY) {
1535 /* procfs requests not want stat in wait for avoid deadlock */
1536 req->rq_no_resend = 1;
1537 req->rq_no_delay = 1;
1540 rc = ptlrpc_queue_wait(req);
1544 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1546 GOTO(out, rc = -EPROTO);
1552 ptlrpc_req_finished(req);
1557 * Implementation of obd_ops::o_import_event
1559 * This function is called when some related import event happens. It will
1560 * mark the necessary flags according to the event and notify the necessary
1561 * threads (mainly precreate thread).
1563 * \param[in] obd OSP OBD device
1564 * \param[in] imp import attached from OSP to remote (OST/MDT) service
1565 * \param[in] event event related to remote service (IMP_EVENT_*)
1567 * \retval 0 0 if the event handling succeeded.
1568 * \retval negative negative errno if the event handling failed.
1570 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1571 enum obd_import_event event)
1573 struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1577 case IMP_EVENT_DISCON:
1578 d->opd_got_disconnected = 1;
1579 d->opd_imp_connected = 0;
1580 if (d->opd_connect_mdt)
1583 if (d->opd_pre != NULL) {
1584 osp_pre_update_status(d, -ENODEV);
1585 wake_up(&d->opd_pre_waitq);
1588 CDEBUG(D_HA, "got disconnected\n");
1590 case IMP_EVENT_INACTIVE:
1591 d->opd_imp_active = 0;
1592 d->opd_imp_connected = 0;
1593 d->opd_obd->obd_inactive = 1;
1594 if (d->opd_connect_mdt)
1596 if (d->opd_pre != NULL) {
1597 /* Import is invalid, we can`t get stripes so
1599 rc = imp->imp_deactive ? -ESHUTDOWN : -ENODEV;
1600 osp_pre_update_status(d, rc);
1601 wake_up(&d->opd_pre_waitq);
1604 CDEBUG(D_HA, "got inactive\n");
1606 case IMP_EVENT_ACTIVE:
1607 d->opd_imp_active = 1;
1609 if (d->opd_got_disconnected)
1610 d->opd_new_connection = 1;
1611 d->opd_imp_connected = 1;
1612 d->opd_imp_seen_connected = 1;
1613 d->opd_obd->obd_inactive = 0;
1614 wake_up(&d->opd_pre_waitq);
1615 if (d->opd_connect_mdt)
1618 osp_sync_check_for_work(d);
1619 CDEBUG(D_HA, "got connected\n");
1621 case IMP_EVENT_INVALIDATE:
1622 if (d->opd_connect_mdt)
1623 osp_invalidate_request(d);
1625 if (obd->obd_namespace == NULL)
1627 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1630 case IMP_EVENT_DEACTIVATE:
1631 case IMP_EVENT_ACTIVATE:
1634 CERROR("%s: unsupported import event: %#x\n",
1635 obd->obd_name, event);
1641 * Implementation of obd_ops: o_iocontrol
1643 * This function is the ioctl handler for OSP. Note: lctl will access the OSP
1644 * directly by ioctl, instead of through the MDS stack.
1646 * param[in] cmd ioctl command.
1647 * param[in] exp export of this OSP.
1648 * param[in] len data length of \a karg.
1649 * param[in] karg input argument which is packed as
1651 * param[out] uarg pointer to userspace buffer (must access by
1654 * \retval 0 0 if the ioctl handling succeeded.
1655 * \retval negative negative errno if the ioctl handling failed.
1657 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1658 void *karg, void __user *uarg)
1660 struct obd_device *obd = exp->exp_obd;
1661 struct osp_device *d;
1662 struct obd_ioctl_data *data = karg;
1667 LASSERT(obd->obd_lu_dev);
1668 d = lu2osp_dev(obd->obd_lu_dev);
1669 LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1671 if (!try_module_get(THIS_MODULE)) {
1672 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1673 module_name(THIS_MODULE));
1678 case OBD_IOC_CLIENT_RECOVER:
1679 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1680 data->ioc_inlbuf1, 0);
1684 case IOC_OSC_SET_ACTIVE:
1685 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1689 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1690 cmd, current->comm);
1693 module_put(THIS_MODULE);
1699 * Implementation of obd_ops::o_get_info
1701 * Retrieve information by key. Retrieval starts from the top layer
1702 * (MDT) of the MDS stack and traverses the stack by calling the
1703 * obd_get_info() method of the next sub-layer.
1705 * \param[in] env execution environment
1706 * \param[in] exp export of this OSP
1707 * \param[in] keylen length of \a key
1708 * \param[in] key the key
1709 * \param[out] vallen length of \a val
1710 * \param[out] val holds the value returned by the key
1712 * \retval 0 0 if getting information succeeded.
1713 * \retval negative negative errno if getting information failed.
1715 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1716 __u32 keylen, void *key, __u32 *vallen, void *val)
1720 if (KEY_IS(KEY_OSP_CONNECTED)) {
1721 struct obd_device *obd = exp->exp_obd;
1722 struct osp_device *osp;
1724 if (!obd->obd_set_up || obd->obd_stopping)
1727 osp = lu2osp_dev(obd->obd_lu_dev);
1730 * 1.8/2.0 behaviour is that OST being connected once at least
1731 * is considered "healthy". and one "healthy" OST is enough to
1732 * allow lustre clients to connect to MDS
1734 RETURN(!osp->opd_imp_seen_connected);
1740 static int osp_obd_set_info_async(const struct lu_env *env,
1741 struct obd_export *exp,
1742 u32 keylen, void *key,
1743 u32 vallen, void *val,
1744 struct ptlrpc_request_set *set)
1746 struct obd_device *obd = exp->exp_obd;
1747 struct obd_import *imp = obd->u.cli.cl_import;
1748 struct osp_device *osp;
1749 struct ptlrpc_request *req;
1753 if (KEY_IS(KEY_SPTLRPC_CONF)) {
1754 sptlrpc_conf_client_adapt(exp->exp_obd);
1758 LASSERT(set != NULL);
1759 if (!obd->obd_set_up || obd->obd_stopping)
1761 osp = lu2osp_dev(obd->obd_lu_dev);
1763 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
1767 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
1768 RCL_CLIENT, keylen);
1769 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
1770 RCL_CLIENT, vallen);
1771 if (osp->opd_connect_mdt)
1772 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SET_INFO);
1774 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
1776 ptlrpc_request_free(req);
1780 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1781 memcpy(tmp, key, keylen);
1782 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1783 memcpy(tmp, val, vallen);
1785 ptlrpc_request_set_replen(req);
1786 ptlrpc_set_add_req(set, req);
1787 ptlrpc_check_set(NULL, set);
1793 * Implementation of obd_ops: o_fid_alloc
1795 * Allocate a FID. There are two cases in which OSP performs
1798 * 1. FID precreation for data objects, which is done in
1799 * osp_precreate_fids() instead of this function.
1800 * 2. FID allocation for each sub-stripe of a striped directory.
1801 * Similar to other FID clients, OSP requests the sequence
1802 * from its corresponding remote MDT, which in turn requests
1803 * sequences from the sequence controller (MDT0).
1805 * \param[in] env execution environment
1806 * \param[in] exp export of the OSP
1807 * \param[out] fid FID being allocated
1808 * \param[in] unused necessary for the interface but unused.
1810 * \retval 0 0 FID allocated successfully.
1811 * \retval 1 1 FID allocated successfully and new sequence
1812 * requested from seq meta server
1813 * \retval negative negative errno if FID allocation failed.
1815 static int osp_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1816 struct lu_fid *fid, struct md_op_data *unused)
1818 struct client_obd *cli = &exp->exp_obd->u.cli;
1819 struct osp_device *osp = lu2osp_dev(exp->exp_obd->obd_lu_dev);
1820 struct lu_client_seq *seq = cli->cl_seq;
1823 LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
1824 /* Sigh, fid client is not ready yet */
1825 LASSERT(osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1827 RETURN(seq_client_alloc_fid(env, seq, fid));
1830 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1831 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1832 static void osp_key_exit(const struct lu_context *ctx,
1833 struct lu_context_key *key, void *data)
1835 struct osp_thread_info *info = data;
1837 info->osi_attr.la_valid = 0;
1840 struct lu_context_key osp_thread_key = {
1841 .lct_tags = LCT_MD_THREAD,
1842 .lct_init = osp_key_init,
1843 .lct_fini = osp_key_fini,
1844 .lct_exit = osp_key_exit
1847 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1848 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1850 struct lu_context_key osp_txn_key = {
1851 .lct_tags = LCT_OSP_THREAD,
1852 .lct_init = osp_txn_key_init,
1853 .lct_fini = osp_txn_key_fini
1855 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1857 static struct lu_device_type_operations osp_device_type_ops = {
1858 .ldto_init = osp_type_init,
1859 .ldto_fini = osp_type_fini,
1861 .ldto_start = osp_type_start,
1862 .ldto_stop = osp_type_stop,
1864 .ldto_device_alloc = osp_device_alloc,
1865 .ldto_device_free = osp_device_free,
1867 .ldto_device_fini = osp_device_fini
1870 static struct lu_device_type osp_device_type = {
1871 .ldt_tags = LU_DEVICE_DT,
1872 .ldt_name = LUSTRE_OSP_NAME,
1873 .ldt_ops = &osp_device_type_ops,
1874 .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1877 static const struct obd_ops osp_obd_device_ops = {
1878 .o_owner = THIS_MODULE,
1879 .o_add_conn = client_import_add_conn,
1880 .o_del_conn = client_import_del_conn,
1881 .o_reconnect = osp_reconnect,
1882 .o_connect = osp_obd_connect,
1883 .o_disconnect = osp_obd_disconnect,
1884 .o_get_info = osp_obd_get_info,
1885 .o_set_info_async = osp_obd_set_info_async,
1886 .o_import_event = osp_import_event,
1887 .o_iocontrol = osp_iocontrol,
1888 .o_statfs = osp_obd_statfs,
1889 .o_fid_init = client_fid_init,
1890 .o_fid_fini = client_fid_fini,
1891 .o_fid_alloc = osp_fid_alloc,
1895 * Initialize OSP module.
1897 * Register device types OSP and Light Weight Proxy (LWP) (\see lwp_dev.c)
1898 * in obd_types (\see class_obd.c). Initialize procfs for the
1899 * the OSP device. Note: OSP was called OSC before Lustre 2.4,
1900 * so for compatibility it still uses the name "osc" in procfs.
1901 * This is called at module load time.
1903 * \retval 0 0 if initialization succeeds.
1904 * \retval negative negative errno if initialization failed.
1906 static int __init osp_init(void)
1908 struct obd_type *sym;
1911 rc = lu_kmem_init(osp_caches);
1915 rc = class_register_type(&osp_obd_device_ops, NULL, false, NULL,
1916 LUSTRE_OSP_NAME, &osp_device_type);
1918 lu_kmem_fini(osp_caches);
1922 rc = class_register_type(&lwp_obd_device_ops, NULL, false, NULL,
1923 LUSTRE_LWP_NAME, &lwp_device_type);
1925 class_unregister_type(LUSTRE_OSP_NAME);
1926 lu_kmem_fini(osp_caches);
1930 /* create "osc" entry for compatibility purposes */
1931 sym = class_add_symlinks(LUSTRE_OSC_NAME, true);
1934 /* does real "osc" already exist ? */
1943 * Finalize OSP module.
1945 * This callback is called when kernel unloads OSP module from memory, and
1946 * it will deregister OSP and LWP device type from obd_types (\see class_obd.c).
1948 static void __exit osp_exit(void)
1950 struct obd_type *sym = class_search_type(LUSTRE_OSC_NAME);
1952 /* if this was never fully initialized by the osc layer
1953 * then we are responsible for freeing this obd_type
1956 /* final put if we manage this obd type */
1957 if (sym->typ_sym_filter)
1958 kobject_put(&sym->typ_kobj);
1959 /* put reference taken by class_search_type */
1960 kobject_put(&sym->typ_kobj);
1963 class_unregister_type(LUSTRE_LWP_NAME);
1964 class_unregister_type(LUSTRE_OSP_NAME);
1965 lu_kmem_fini(osp_caches);
1968 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1969 MODULE_DESCRIPTION("Lustre OSD Storage Proxy ("LUSTRE_OSP_NAME")");
1970 MODULE_VERSION(LUSTRE_VERSION_STRING);
1971 MODULE_LICENSE("GPL");
1973 module_init(osp_init);
1974 module_exit(osp_exit);