Whamcloud - gitweb
LU-11838 osp: atomic64_read() returns s64
[fs/lustre-release.git] / lustre / osp / osp_dev.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/osp/osp_dev.c
33  *
34  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
35  * Author: Mikhail Pershin <mike.pershin@intel.com>
36  * Author: Di Wang <di.wang@intel.com>
37  */
38 /*
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:
47  *
48  * - OSP devices allocate FIDs for the stripe sub-objects of a striped
49  *   file or directory.
50  *
51  * - OSP objects represent the remote MDT and OST objects that are
52  *   the stripes of a striped object.
53  *
54  * - OSP devices log, send, and track synchronous operations (setattr
55  *   and unlink) to remote targets.
56  *
57  * - OSP objects are the bottom slice of the compound LU object
58  *   representing a remote MDT object: MDT/MDD/LOD/OSP.
59  *
60  * - OSP objects are used by LFSCK to represent remote OST objects
61  *   during the verification of MDT-OST consistency.
62  *
63  * - OSP devices batch idempotent requests (declare_attr_get() and
64  *   declare_xattr_get()) to the remote target and cache their results.
65  *
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.
69  */
70
71 #define DEBUG_SUBSYSTEM S_MDS
72
73 #include <linux/kthread.h>
74
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>
80
81 #include "osp_internal.h"
82
83 /* Slab for OSP object allocation */
84 struct kmem_cache *osp_object_kmem;
85
86 static struct lu_kmem_descr osp_caches[] = {
87         {
88                 .ckd_cache = &osp_object_kmem,
89                 .ckd_name  = "osp_obj",
90                 .ckd_size  = sizeof(struct osp_object)
91         },
92         {
93                 .ckd_cache = NULL
94         }
95 };
96
97 /**
98  * Implementation of lu_device_operations::ldo_object_alloc
99  *
100  * Allocates an OSP object in memory, whose FID is on the remote target.
101  *
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
106  *                      an OST object.
107  * \param[in] d         OSP device
108  *
109  * \retval object       object being created if the creation succeed.
110  * \retval NULL         NULL if the creation failed.
111  */
112 static struct lu_object *osp_object_alloc(const struct lu_env *env,
113                                           const struct lu_object_header *hdr,
114                                           struct lu_device *d)
115 {
116         struct lu_object_header *h = NULL;
117         struct osp_object       *o;
118         struct lu_object        *l;
119
120         OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, GFP_NOFS);
121         if (o != NULL) {
122                 l = &o->opo_obj.do_lu;
123
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 */
128                 if (hdr == NULL) {
129                         h = &o->opo_header;
130                         lu_object_header_init(h);
131                         dt_object_init(&o->opo_obj, h, d);
132                         lu_object_add_top(h, l);
133                 } else {
134                         dt_object_init(&o->opo_obj, h, d);
135                 }
136
137                 l->lo_ops = &osp_lu_obj_ops;
138
139                 return l;
140         } else {
141                 return NULL;
142         }
143 }
144
145 /**
146  * Find or create the local object
147  *
148  * Finds or creates the local file referenced by \a reg_id and return the
149  * attributes of the local file.
150  *
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.
157  *
158  * \retval object               object(dt_object) found or created
159  * \retval ERR_PTR(errno)       ERR_PTR(errno) if not get the object.
160  */
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)
164 {
165         struct osp_thread_info *osi = osp_env_info(env);
166         struct dt_object_format dof = { 0 };
167         struct dt_object       *dto;
168         int                  rc;
169         ENTRY;
170
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,
177                                 &dof, attr);
178         if (IS_ERR(dto))
179                 RETURN(dto);
180
181         /* Get attributes of the local object. */
182         rc = dt_attr_get(env, dto, attr);
183         if (rc) {
184                 CERROR("%s: can't be initialized: rc = %d\n",
185                        osp->opd_obd->obd_name, rc);
186                 dt_object_put(env, dto);
187                 RETURN(ERR_PTR(rc));
188         }
189         RETURN(dto);
190 }
191
192 /**
193  * Write data buffer to a local file object.
194  *
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
200  *
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.
204  */
205 static int osp_write_local_file(const struct lu_env *env,
206                                 struct osp_device *osp,
207                                 struct dt_object *dt_obj,
208                                 struct lu_buf *buf,
209                                 loff_t offset)
210 {
211         struct thandle *th;
212         int rc;
213
214         if (osp->opd_storage->dd_rdonly)
215                 RETURN(0);
216
217         th = dt_trans_create(env, osp->opd_storage);
218         if (IS_ERR(th))
219                 RETURN(PTR_ERR(th));
220
221         rc = dt_declare_record_write(env, dt_obj, buf, offset, th);
222         if (rc)
223                 GOTO(out, rc);
224         rc = dt_trans_start_local(env, osp->opd_storage, th);
225         if (rc)
226                 GOTO(out, rc);
227
228         rc = dt_record_write(env, dt_obj, buf, &offset, th);
229 out:
230         dt_trans_stop(env, osp->opd_storage, th);
231         RETURN(rc);
232 }
233
234 /**
235  * Initialize last ID object.
236  *
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
240  * OSTs.
241  *
242  * \param[in] env       execution environment
243  * \param[in] osp       OSP device
244  *
245  * \retval 0            0 if initialization succeed
246  * \retval negative     negative errno if initialization failed
247  */
248 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
249 {
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;
253         int                     rc = -EFAULT;
254         ENTRY;
255
256         dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
257                                             MDD_LOV_OBJ_OID);
258         if (IS_ERR(dto))
259                 RETURN(PTR_ERR(dto));
260
261         osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &osp->opd_last_id,
262                            osp->opd_index);
263
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)
268                         GOTO(out, rc);
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;
273         }
274
275         if (rc == -EFAULT) { /* fresh LAST_ID */
276                 osp->opd_last_id = 0;
277                 fid->f_oid = 0;
278                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
279                                           osi->osi_off);
280                 if (rc != 0)
281                         GOTO(out, rc);
282         }
283         osp->opd_last_used_oid_file = dto;
284         RETURN(0);
285 out:
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;
291         RETURN(rc);
292 }
293
294 /**
295  * Initialize last sequence object.
296  *
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.
301  *
302  * \param[in] env       execution environment
303  * \param[in] osp       OSP device
304  *
305  * \retval 0            0 if initialization succeed
306  * \retval negative     negative errno if initialization failed
307  */
308 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
309 {
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;
313         int                     rc = -EFAULT;
314         ENTRY;
315
316         dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr,
317                                             MDD_LOV_OBJ_OSEQ);
318         if (IS_ERR(dto))
319                 RETURN(PTR_ERR(dto));
320
321         osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
322                            osp->opd_index);
323
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)
328                         GOTO(out, rc);
329                 if (fid_is_idif(fid))
330                         fid->f_seq = fid_idif_seq(osp->opd_last_id,
331                                                   osp->opd_index);
332         }
333
334         if (rc == -EFAULT) { /* fresh OSP */
335                 fid->f_seq = 0;
336                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
337                                           osi->osi_off);
338                 if (rc != 0)
339                         GOTO(out, rc);
340         }
341         osp->opd_last_used_seq_file = dto;
342         RETURN(0);
343 out:
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;
349         RETURN(rc);
350 }
351
352 /**
353  * Initialize last OID and sequence object.
354  *
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.
358  *
359  * \param[in] env       execution environment
360  * \param[in] osp       OSP device
361  *
362  * \retval 0            0 if initialization succeed
363  * \retval negative     negative error if initialization failed
364  */
365 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
366 {
367         struct osp_thread_info *osi = osp_env_info(env);
368         int                  rc;
369         ENTRY;
370
371         fid_zero(&osp->opd_last_used_fid);
372         rc = osp_init_last_objid(env, osp);
373         if (rc < 0) {
374                 CERROR("%s: Can not get ids %d from old objid!\n",
375                        osp->opd_obd->obd_name, rc);
376                 RETURN(rc);
377         }
378
379         rc = osp_init_last_seq(env, osp);
380         if (rc < 0) {
381                 CERROR("%s: Can not get sequence %d from old objseq!\n",
382                        osp->opd_obd->obd_name, rc);
383                 GOTO(out, rc);
384         }
385
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),
392                                 osp->opd_index);
393                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
394                                     &osp->opd_last_used_fid.f_seq,
395                                     osp->opd_index);
396                 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
397                                           &osi->osi_lb, osi->osi_off);
398                 if (rc) {
399                         CERROR("%s : Can not write seq file: rc = %d\n",
400                                osp->opd_obd->obd_name, rc);
401                         GOTO(out, rc);
402                 }
403         }
404
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);
410         }
411
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));
415 out:
416         if (rc != 0) {
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;
420                 }
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;
424                 }
425         }
426
427         RETURN(rc);
428 }
429
430 /**
431  * Release the last sequence and OID file objects in OSP device.
432  *
433  * \param[in] env       execution environment
434  * \param[in] osp       OSP device
435  */
436 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *osp)
437 {
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;
442         }
443
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;
447         }
448 }
449
450 /**
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
453  * cleanup process.
454  *
455  * \param[in] d         OSP device being disconnected
456  *
457  * \retval 0            0 if disconnection succeed
458  * \retval negative     negative errno if disconnection failed
459  */
460 static int osp_disconnect(struct osp_device *d)
461 {
462         struct obd_device *obd = d->opd_obd;
463         struct obd_import *imp;
464         int rc = 0;
465
466         imp = obd->u.cli.cl_import;
467
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);
475
476         ptlrpc_deactivate_import(imp);
477
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
480          * never added.) */
481         (void)ptlrpc_pinger_del_import(imp);
482
483         rc = ptlrpc_disconnect_import(imp, 0);
484         if (rc != 0)
485                 CERROR("%s: can't disconnect: rc = %d\n", obd->obd_name, rc);
486
487         ptlrpc_invalidate_import(imp);
488
489         RETURN(rc);
490 }
491
492 /**
493  * Initialize the osp_update structure in OSP device
494  *
495  * Allocate osp update structure and start update thread.
496  *
497  * \param[in] osp       OSP device
498  *
499  * \retval              0 if initialization succeeds.
500  * \retval              negative errno if initialization fails.
501  */
502 static int osp_update_init(struct osp_device *osp)
503 {
504         struct l_wait_info      lwi = { 0 };
505         struct task_struct      *task;
506
507         ENTRY;
508
509         LASSERT(osp->opd_connect_mdt);
510
511         if (osp->opd_storage->dd_rdonly)
512                 RETURN(0);
513
514         OBD_ALLOC_PTR(osp->opd_update);
515         if (osp->opd_update == NULL)
516                 RETURN(-ENOMEM);
517
518         init_waitqueue_head(&osp->opd_update_thread.t_ctl_waitq);
519         init_waitqueue_head(&osp->opd_update->ou_waitq);
520         spin_lock_init(&osp->opd_update->ou_lock);
521         INIT_LIST_HEAD(&osp->opd_update->ou_list);
522         osp->opd_update->ou_rpc_version = 1;
523         osp->opd_update->ou_version = 1;
524         osp->opd_update->ou_generation = 0;
525
526         /* start thread handling sending updates to the remote MDT */
527         task = kthread_run(osp_send_update_thread, osp,
528                            "osp_up%u-%u", osp->opd_index, osp->opd_group);
529         if (IS_ERR(task)) {
530                 int rc = PTR_ERR(task);
531
532                 OBD_FREE_PTR(osp->opd_update);
533                 osp->opd_update = NULL;
534                 CERROR("%s: can't start precreate thread: rc = %d\n",
535                        osp->opd_obd->obd_name, rc);
536                 RETURN(rc);
537         }
538
539         l_wait_event(osp->opd_update_thread.t_ctl_waitq,
540                      osp_send_update_thread_running(osp) ||
541                      osp_send_update_thread_stopped(osp), &lwi);
542
543         RETURN(0);
544 }
545
546 /**
547  * Finialize osp_update structure in OSP device
548  *
549  * Stop the OSP update sending thread, then delete the left
550  * osp thandle in the sending list.
551  *
552  * \param [in] osp      OSP device.
553  */
554 static void osp_update_fini(const struct lu_env *env, struct osp_device *osp)
555 {
556         struct osp_update_request *our;
557         struct osp_update_request *tmp;
558         struct osp_updates *ou = osp->opd_update;
559
560         if (ou == NULL)
561                 return;
562
563         osp->opd_update_thread.t_flags = SVC_STOPPING;
564         wake_up(&ou->ou_waitq);
565
566         wait_event(osp->opd_update_thread.t_ctl_waitq,
567                    osp->opd_update_thread.t_flags & SVC_STOPPED);
568
569         /* Remove the left osp thandle from the list */
570         spin_lock(&ou->ou_lock);
571         list_for_each_entry_safe(our, tmp, &ou->ou_list,
572                                  our_list) {
573                 list_del_init(&our->our_list);
574                 LASSERT(our->our_th != NULL);
575                 osp_trans_callback(env, our->our_th, -EIO);
576                 /* our will be destroyed in osp_thandle_put() */
577                 osp_thandle_put(env, our->our_th);
578         }
579         spin_unlock(&ou->ou_lock);
580
581         OBD_FREE_PTR(ou);
582         osp->opd_update = NULL;
583 }
584
585 /**
586  * Cleanup OSP, which includes disconnect import, cleanup unlink log, stop
587  * precreate threads etc.
588  *
589  * \param[in] env       execution environment.
590  * \param[in] d         OSP device being disconnected.
591  *
592  * \retval 0            0 if cleanup succeed
593  * \retval negative     negative errno if cleanup failed
594  */
595 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
596 {
597         int                      rc = 0;
598         ENTRY;
599
600         LASSERT(env);
601
602         rc = osp_disconnect(d);
603
604         osp_statfs_fini(d);
605
606         if (!d->opd_connect_mdt) {
607                 /* stop sync thread */
608                 osp_sync_fini(d);
609
610                 /* stop precreate thread */
611                 osp_precreate_fini(d);
612
613                 /* release last_used file */
614                 osp_last_used_fini(env, d);
615         }
616
617         obd_fid_fini(d->opd_obd);
618
619         RETURN(rc);
620 }
621
622 /**
623  * Implementation of osp_lu_ops::ldo_process_config
624  *
625  * This function processes config log records in OSP layer. It is usually
626  * called from the top layer of MDT stack, and goes through the stack by calling
627  * ldo_process_config of next layer.
628  *
629  * \param[in] env       execution environment
630  * \param[in] dev       lu_device of OSP
631  * \param[in] lcfg      config log
632  *
633  * \retval 0            0 if the config log record is executed correctly.
634  * \retval negative     negative errno if the record execution fails.
635  */
636 static int osp_process_config(const struct lu_env *env,
637                               struct lu_device *dev, struct lustre_cfg *lcfg)
638 {
639         struct osp_device *d = lu2osp_dev(dev);
640         struct dt_device *dt = lu2dt_dev(dev);
641         struct obd_device *obd = d->opd_obd;
642         ssize_t count;
643         int rc;
644
645         ENTRY;
646
647         switch (lcfg->lcfg_command) {
648         case LCFG_PRE_CLEANUP:
649                 rc = osp_disconnect(d);
650                 osp_update_fini(env, d);
651                 if (obd->obd_namespace != NULL)
652                         ldlm_namespace_free_prior(obd->obd_namespace, NULL, 1);
653                 break;
654         case LCFG_CLEANUP:
655                 lu_dev_del_linkage(dev->ld_site, dev);
656                 rc = osp_shutdown(env, d);
657                 break;
658         case LCFG_PARAM:
659                 count = class_modify_config(lcfg, d->opd_connect_mdt ?
660                                                   PARAM_OSP : PARAM_OSC,
661                                             &dt->dd_kobj);
662                 if (count < 0) {
663                         /* class_modify_config() haven't found matching
664                          * parameter and returned an error so that layer(s)
665                          * below could use that. But OSP is the bottom, so
666                          * just ignore it
667                          */
668                         CERROR("%s: unknown param %s\n",
669                                (char *)lustre_cfg_string(lcfg, 0),
670                                (char *)lustre_cfg_string(lcfg, 1));
671                 }
672                 rc = 0;
673                 break;
674         default:
675                 CERROR("%s: unknown command %u\n",
676                        (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
677                 rc = 0;
678                 break;
679         }
680
681         RETURN(rc);
682 }
683
684 /**
685  * Implementation of osp_lu_ops::ldo_recovery_complete
686  *
687  * This function is called after recovery is finished, and OSP layer
688  * will wake up precreate thread here.
689  *
690  * \param[in] env       execution environment
691  * \param[in] dev       lu_device of OSP
692  *
693  * \retval 0            0 unconditionally
694  */
695 static int osp_recovery_complete(const struct lu_env *env,
696                                  struct lu_device *dev)
697 {
698         struct osp_device       *osp = lu2osp_dev(dev);
699
700         ENTRY;
701         osp->opd_recovery_completed = 1;
702
703         if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
704                 wake_up(&osp->opd_pre_waitq);
705
706         RETURN(0);
707 }
708
709 const struct lu_device_operations osp_lu_ops = {
710         .ldo_object_alloc       = osp_object_alloc,
711         .ldo_process_config     = osp_process_config,
712         .ldo_recovery_complete  = osp_recovery_complete,
713 };
714
715 /**
716  * Implementation of dt_device_operations::dt_statfs
717  *
718  * This function provides statfs status (for precreation) from
719  * corresponding OST. Note: this function only retrieves the status
720  * from the OSP device, and the real statfs RPC happens inside
721  * precreate thread (\see osp_statfs_update). Note: OSP for MDT does
722  * not need to retrieve statfs data for now.
723  *
724  * \param[in] env       execution environment.
725  * \param[in] dev       dt_device of OSP.
726  * \param[out] sfs      holds the retrieved statfs data.
727  *
728  * \retval 0            0 statfs data was retrieved successfully or
729  *                      retrieval was not needed
730  * \retval negative     negative errno if get statfs failed.
731  */
732 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
733                       struct obd_statfs *sfs)
734 {
735         struct osp_device *d = dt2osp_dev(dev);
736         struct obd_import *imp = d->opd_obd->u.cli.cl_import;
737
738         ENTRY;
739
740         if (imp->imp_state == LUSTRE_IMP_CLOSED)
741                 RETURN(-ESHUTDOWN);
742
743         if (unlikely(d->opd_imp_active == 0))
744                 RETURN(-ENOTCONN);
745
746         /* return recently updated data */
747         *sfs = d->opd_statfs;
748
749         if (d->opd_pre == NULL)
750                 RETURN(0);
751
752         /*
753          * layer above osp (usually lod) can use ffree to estimate
754          * how many objects are available for immediate creation
755          */
756         spin_lock(&d->opd_pre_lock);
757         LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
758                  fid_seq(&d->opd_pre_used_fid),
759                  "last_created "DFID", next_fid "DFID"\n",
760                  PFID(&d->opd_pre_last_created_fid),
761                  PFID(&d->opd_pre_used_fid));
762         sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
763                               fid_oid(&d->opd_pre_used_fid);
764         sfs->os_fprecreated -= d->opd_pre_reserved;
765         LASSERTF(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2,
766                  "last_created "DFID", next_fid "DFID", reserved %llu\n",
767                  PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_pre_used_fid),
768                  d->opd_pre_reserved);
769         spin_unlock(&d->opd_pre_lock);
770
771         CDEBUG(D_OTHER, "%s: %llu blocks, %llu free, %llu avail, "
772                "%llu files, %llu free files\n", d->opd_obd->obd_name,
773                sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
774                sfs->os_files, sfs->os_ffree);
775         RETURN(0);
776 }
777
778 static int osp_sync_timeout(void *data)
779 {
780         return 1;
781 }
782
783 /**
784  * Implementation of dt_device_operations::dt_sync
785  *
786  * This function synchronizes the OSP cache to the remote target. It wakes
787  * up unlink log threads and sends out unlink records to the remote OST.
788  *
789  * \param[in] env       execution environment
790  * \param[in] dev       dt_device of OSP
791  *
792  * \retval 0            0 if synchronization succeeds
793  * \retval negative     negative errno if synchronization fails
794  */
795 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
796 {
797         struct osp_device *d = dt2osp_dev(dev);
798         struct l_wait_info lwi = { 0 };
799         time64_t start = ktime_get_seconds();
800         int recs, rc = 0;
801         u64 old;
802
803         ENTRY;
804
805         /* No Sync between MDTs yet. */
806         if (d->opd_connect_mdt)
807                 RETURN(0);
808
809         recs = atomic_read(&d->opd_sync_changes);
810         old = atomic64_read(&d->opd_sync_processed_recs);
811
812         osp_sync_force(env, dt2osp_dev(dev));
813
814         if (unlikely(d->opd_imp_active == 0))
815                 RETURN(-ENOTCONN);
816
817         down_write(&d->opd_async_updates_rwsem);
818
819         CDEBUG(D_OTHER, "%s: async updates %d\n", d->opd_obd->obd_name,
820                atomic_read(&d->opd_async_updates_count));
821
822         /* make sure the connection is fine */
823         lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout), osp_sync_timeout, d);
824         rc = l_wait_event(d->opd_sync_barrier_waitq,
825                           atomic_read(&d->opd_async_updates_count) == 0,
826                           &lwi);
827         up_write(&d->opd_async_updates_rwsem);
828         if (rc != 0)
829                 GOTO(out, rc);
830
831         CDEBUG(D_CACHE, "%s: processed %llu\n", d->opd_obd->obd_name,
832                (unsigned long long)atomic64_read(&d->opd_sync_processed_recs));
833
834         while (atomic64_read(&d->opd_sync_processed_recs) < old + recs) {
835                 __u64 last = atomic64_read(&d->opd_sync_processed_recs);
836                 /* make sure the connection is fine */
837                 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
838                                   osp_sync_timeout, d);
839                 l_wait_event(d->opd_sync_barrier_waitq,
840                              atomic64_read(&d->opd_sync_processed_recs)
841                              >= old + recs, &lwi);
842
843                 if (atomic64_read(&d->opd_sync_processed_recs) >= old + recs)
844                         break;
845
846                 if (atomic64_read(&d->opd_sync_processed_recs) != last) {
847                         /* some progress have been made,
848                          * keep trying... */
849                         continue;
850                 }
851
852                 /* no changes and expired, something is wrong */
853                 GOTO(out, rc = -ETIMEDOUT);
854         }
855
856         /* block new processing (barrier>0 - few callers are possible */
857         atomic_inc(&d->opd_sync_barrier);
858
859         CDEBUG(D_CACHE, "%s: %u in flight\n", d->opd_obd->obd_name,
860                atomic_read(&d->opd_sync_rpcs_in_flight));
861
862         /* wait till all-in-flight are replied, so executed by the target */
863         /* XXX: this is used by LFSCK at the moment, which doesn't require
864          *      all the changes to be committed, but in general it'd be
865          *      better to wait till commit */
866         while (atomic_read(&d->opd_sync_rpcs_in_flight) > 0) {
867                 old = atomic_read(&d->opd_sync_rpcs_in_flight);
868
869                 lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
870                                   osp_sync_timeout, d);
871                 l_wait_event(d->opd_sync_barrier_waitq,
872                              atomic_read(&d->opd_sync_rpcs_in_flight) == 0,
873                              &lwi);
874
875                 if (atomic_read(&d->opd_sync_rpcs_in_flight) == 0)
876                         break;
877
878                 if (atomic_read(&d->opd_sync_rpcs_in_flight) != old) {
879                         /* some progress have been made */
880                         continue;
881                 }
882
883                 /* no changes and expired, something is wrong */
884                 GOTO(out, rc = -ETIMEDOUT);
885         }
886
887 out:
888         /* resume normal processing (barrier=0) */
889         atomic_dec(&d->opd_sync_barrier);
890         osp_sync_check_for_work(d);
891
892         CDEBUG(D_CACHE, "%s: done in %lld: rc = %d\n", d->opd_obd->obd_name,
893                ktime_get_seconds() - start, rc);
894
895         RETURN(rc);
896 }
897
898 const struct dt_device_operations osp_dt_ops = {
899         .dt_statfs       = osp_statfs,
900         .dt_sync         = osp_sync,
901         .dt_trans_create = osp_trans_create,
902         .dt_trans_start  = osp_trans_start,
903         .dt_trans_stop   = osp_trans_stop,
904         .dt_trans_cb_add   = osp_trans_cb_add,
905 };
906
907 /**
908  * Connect OSP to local OSD.
909  *
910  * Locate the local OSD referenced by \a nextdev and connect to it. Sometimes,
911  * OSP needs to access the local OSD to store some information. For example,
912  * during precreate, it needs to update last used OID and sequence file
913  * (LAST_SEQ) in local OSD.
914  *
915  * \param[in] env       execution environment
916  * \param[in] osp       OSP device
917  * \param[in] nextdev   the name of local OSD
918  *
919  * \retval 0            0 connection succeeded
920  * \retval negative     negative errno connection failed
921  */
922 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *osp,
923                               const char *nextdev)
924 {
925         struct obd_connect_data *data = NULL;
926         struct obd_device       *obd;
927         int                      rc;
928
929         ENTRY;
930
931         LASSERT(osp->opd_storage_exp == NULL);
932
933         OBD_ALLOC_PTR(data);
934         if (data == NULL)
935                 RETURN(-ENOMEM);
936
937         obd = class_name2obd(nextdev);
938         if (obd == NULL) {
939                 CERROR("%s: can't locate next device: %s\n",
940                        osp->opd_obd->obd_name, nextdev);
941                 GOTO(out, rc = -ENOTCONN);
942         }
943
944         rc = obd_connect(env, &osp->opd_storage_exp, obd, &obd->obd_uuid, data,
945                          NULL);
946         if (rc) {
947                 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
948                        osp->opd_obd->obd_name, nextdev, rc);
949                 GOTO(out, rc);
950         }
951
952         osp->opd_dt_dev.dd_lu_dev.ld_site =
953                 osp->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
954         LASSERT(osp->opd_dt_dev.dd_lu_dev.ld_site);
955         osp->opd_storage = lu2dt_dev(osp->opd_storage_exp->exp_obd->obd_lu_dev);
956
957 out:
958         OBD_FREE_PTR(data);
959         RETURN(rc);
960 }
961
962 /**
963  * Determine if the lock needs to be cancelled
964  *
965  * Determine if the unused lock should be cancelled before replay, see
966  * (ldlm_cancel_no_wait_policy()). Currently, only inode bits lock exists
967  * between MDTs.
968  *
969  * \param[in] lock      lock to be checked.
970  *
971  * \retval              1 if the lock needs to be cancelled before replay.
972  * \retval              0 if the lock does not need to be cancelled before
973  *                      replay.
974  */
975 static int osp_cancel_weight(struct ldlm_lock *lock)
976 {
977         if (lock->l_resource->lr_type != LDLM_IBITS)
978                 RETURN(0);
979
980         RETURN(1);
981 }
982
983 /**
984  * Initialize OSP device according to the parameters in the configuration
985  * log \a cfg.
986  *
987  * Reconstruct the local device name from the configuration profile, and
988  * initialize necessary threads and structures according to the OSP type
989  * (MDT or OST).
990  *
991  * Since there is no record in the MDT configuration for the local disk
992  * device, we have to extract this from elsewhere in the profile.
993  * The only information we get at setup is from the OSC records:
994  * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
995  *
996  * Note: configs generated by Lustre 1.8 are missing the -MDTxxxx part,
997  * so, we need to reconstruct the name of the underlying OSD from this:
998  * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".
999  *
1000  * \param[in] env       execution environment
1001  * \param[in] osp       OSP device
1002  * \param[in] ldt       lu device type of OSP
1003  * \param[in] cfg       configuration log
1004  *
1005  * \retval 0            0 if OSP initialization succeeded.
1006  * \retval negative     negative errno if OSP initialization failed.
1007  */
1008 static int osp_init0(const struct lu_env *env, struct osp_device *osp,
1009                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
1010 {
1011         struct obd_device       *obd;
1012         struct obd_import       *imp;
1013         char                    *src, *tgt, *mdt, *osdname = NULL;
1014         int                     rc;
1015         long                    idx;
1016
1017         ENTRY;
1018
1019         mutex_init(&osp->opd_async_requests_mutex);
1020         INIT_LIST_HEAD(&osp->opd_async_updates);
1021         init_rwsem(&osp->opd_async_updates_rwsem);
1022         atomic_set(&osp->opd_async_updates_count, 0);
1023
1024         obd = class_name2obd(lustre_cfg_string(cfg, 0));
1025         if (obd == NULL) {
1026                 CERROR("Cannot find obd with name %s\n",
1027                        lustre_cfg_string(cfg, 0));
1028                 RETURN(-ENODEV);
1029         }
1030         osp->opd_obd = obd;
1031
1032         src = lustre_cfg_string(cfg, 0);
1033         if (src == NULL)
1034                 RETURN(-EINVAL);
1035
1036         tgt = strrchr(src, '-');
1037         if (tgt == NULL) {
1038                 CERROR("%s: invalid target name %s: rc = %d\n",
1039                        osp->opd_obd->obd_name, lustre_cfg_string(cfg, 0),
1040                        -EINVAL);
1041                 RETURN(-EINVAL);
1042         }
1043
1044         if (strncmp(tgt, "-osc", 4) == 0) {
1045                 /* Old OSC name fsname-OSTXXXX-osc */
1046                 for (tgt--; tgt > src && *tgt != '-'; tgt--)
1047                         ;
1048                 if (tgt == src) {
1049                         CERROR("%s: invalid target name %s: rc = %d\n",
1050                                osp->opd_obd->obd_name,
1051                                lustre_cfg_string(cfg, 0), -EINVAL);
1052                         RETURN(-EINVAL);
1053                 }
1054
1055                 if (strncmp(tgt, "-OST", 4) != 0) {
1056                         CERROR("%s: invalid target name %s: rc = %d\n",
1057                                osp->opd_obd->obd_name,
1058                                lustre_cfg_string(cfg, 0), -EINVAL);
1059                         RETURN(-EINVAL);
1060                 }
1061
1062                 idx = simple_strtol(tgt + 4, &mdt, 16);
1063                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1064                         CERROR("%s: invalid OST index in '%s': rc = %d\n",
1065                                osp->opd_obd->obd_name, src, -EINVAL);
1066                         RETURN(-EINVAL);
1067                 }
1068                 osp->opd_index = idx;
1069                 osp->opd_group = 0;
1070                 idx = tgt - src;
1071         } else {
1072                 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
1073                 if (strncmp(tgt, "-MDT", 4) != 0 &&
1074                     strncmp(tgt, "-OST", 4) != 0) {
1075                         CERROR("%s: invalid target name %s: rc = %d\n",
1076                                osp->opd_obd->obd_name,
1077                                lustre_cfg_string(cfg, 0), -EINVAL);
1078                         RETURN(-EINVAL);
1079                 }
1080
1081                 idx = simple_strtol(tgt + 4, &mdt, 16);
1082                 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
1083                         CERROR("%s: invalid OST index in '%s': rc = %d\n",
1084                                osp->opd_obd->obd_name, src, -EINVAL);
1085                         RETURN(-EINVAL);
1086                 }
1087
1088                 /* Get MDT index from the name and set it to opd_group,
1089                  * which will be used by OSP to connect with OST */
1090                 osp->opd_group = idx;
1091                 if (tgt - src <= 12) {
1092                         CERROR("%s: invalid mdt index from %s: rc =%d\n",
1093                                osp->opd_obd->obd_name,
1094                                lustre_cfg_string(cfg, 0), -EINVAL);
1095                         RETURN(-EINVAL);
1096                 }
1097
1098                 if (strncmp(tgt - 12, "-MDT", 4) == 0)
1099                         osp->opd_connect_mdt = 1;
1100
1101                 idx = simple_strtol(tgt - 8, &mdt, 16);
1102                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
1103                         CERROR("%s: invalid OST index in '%s': rc =%d\n",
1104                                osp->opd_obd->obd_name, src, -EINVAL);
1105                         RETURN(-EINVAL);
1106                 }
1107
1108                 osp->opd_index = idx;
1109                 idx = tgt - src - 12;
1110         }
1111         /* check the fsname length, and after this everything else will fit */
1112         if (idx > MTI_NAME_MAXLEN) {
1113                 CERROR("%s: fsname too long in '%s': rc = %d\n",
1114                        osp->opd_obd->obd_name, src, -EINVAL);
1115                 RETURN(-EINVAL);
1116         }
1117
1118         OBD_ALLOC(osdname, MAX_OBD_NAME);
1119         if (osdname == NULL)
1120                 RETURN(-ENOMEM);
1121
1122         memcpy(osdname, src, idx); /* copy just the fsname part */
1123         osdname[idx] = '\0';
1124
1125         mdt = strstr(mdt, "-MDT");
1126         if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
1127                 strcat(osdname, "-MDT0000");
1128         else
1129                 strcat(osdname, mdt);
1130         strcat(osdname, "-osd");
1131         CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
1132
1133         osp_init_rpc_lock(osp);
1134
1135         osp->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
1136         osp->opd_dt_dev.dd_ops = &osp_dt_ops;
1137
1138         obd->obd_lu_dev = &osp->opd_dt_dev.dd_lu_dev;
1139
1140         rc = osp_connect_to_osd(env, osp, osdname);
1141         if (rc)
1142                 GOTO(out_fini, rc);
1143
1144         rc = ptlrpcd_addref();
1145         if (rc)
1146                 GOTO(out_disconnect, rc);
1147
1148         rc = client_obd_setup(obd, cfg);
1149         if (rc) {
1150                 CERROR("%s: can't setup obd: rc = %d\n", osp->opd_obd->obd_name,
1151                        rc);
1152                 GOTO(out_ref, rc);
1153         }
1154
1155         osp_tunables_init(osp);
1156
1157         rc = obd_fid_init(osp->opd_obd, NULL, osp->opd_connect_mdt ?
1158                           LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
1159         if (rc) {
1160                 CERROR("%s: fid init error: rc = %d\n",
1161                        osp->opd_obd->obd_name, rc);
1162                 GOTO(out_proc, rc);
1163         }
1164
1165         if (!osp->opd_connect_mdt) {
1166                 /* Initialize last id from the storage - will be
1167                  * used in orphan cleanup. */
1168                 if (!osp->opd_storage->dd_rdonly) {
1169                         rc = osp_last_used_init(env, osp);
1170                         if (rc)
1171                                 GOTO(out_fid, rc);
1172                 }
1173
1174                 /* Initialize precreation thread, it handles new
1175                  * connections as well. */
1176                 rc = osp_init_precreate(osp);
1177                 if (rc)
1178                         GOTO(out_last_used, rc);
1179
1180                 /*
1181                  * Initialize synhronization mechanism taking
1182                  * care of propogating changes to OST in near
1183                  * transactional manner.
1184                  */
1185                 rc = osp_sync_init(env, osp);
1186                 if (rc < 0)
1187                         GOTO(out_precreat, rc);
1188         } else {
1189                 osp->opd_got_disconnected = 1;
1190                 rc = osp_update_init(osp);
1191                 if (rc != 0)
1192                         GOTO(out_fid, rc);
1193         }
1194
1195         rc = osp_init_statfs(osp);
1196         if (rc)
1197                 GOTO(out_precreat, rc);
1198
1199         ns_register_cancel(obd->obd_namespace, osp_cancel_weight);
1200
1201         /*
1202          * Initiate connect to OST
1203          */
1204         imp = obd->u.cli.cl_import;
1205
1206         rc = ptlrpc_init_import(imp);
1207         if (rc)
1208                 GOTO(out, rc);
1209         if (osdname)
1210                 OBD_FREE(osdname, MAX_OBD_NAME);
1211         RETURN(0);
1212
1213 out:
1214         if (!osp->opd_connect_mdt)
1215                 /* stop sync thread */
1216                 osp_sync_fini(osp);
1217 out_precreat:
1218         /* stop precreate thread */
1219         if (!osp->opd_connect_mdt)
1220                 osp_precreate_fini(osp);
1221         else
1222                 osp_update_fini(env, osp);
1223 out_last_used:
1224         if (!osp->opd_connect_mdt)
1225                 osp_last_used_fini(env, osp);
1226 out_fid:
1227         obd_fid_fini(osp->opd_obd);
1228 out_proc:
1229         osp_tunables_fini(osp);
1230         client_obd_cleanup(obd);
1231 out_ref:
1232         ptlrpcd_decref();
1233 out_disconnect:
1234         obd_disconnect(osp->opd_storage_exp);
1235 out_fini:
1236         if (osdname)
1237                 OBD_FREE(osdname, MAX_OBD_NAME);
1238         RETURN(rc);
1239 }
1240
1241 /**
1242  * Implementation of lu_device_type_operations::ldto_device_free
1243  *
1244  * Free the OSP device in memory.  No return value is needed for now,
1245  * so always return NULL to comply with the interface.
1246  *
1247  * \param[in] env       execution environment
1248  * \param[in] lu        lu_device of OSP
1249  *
1250  * \retval NULL         NULL unconditionally
1251  */
1252 static struct lu_device *osp_device_free(const struct lu_env *env,
1253                                          struct lu_device *lu)
1254 {
1255         struct osp_device *osp = lu2osp_dev(lu);
1256
1257         if (atomic_read(&lu->ld_ref) && lu->ld_site) {
1258                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1259                 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
1260         }
1261         dt_device_fini(&osp->opd_dt_dev);
1262         OBD_FREE_PTR(osp);
1263
1264         return NULL;
1265 }
1266
1267 /**
1268  * Implementation of lu_device_type_operations::ldto_device_alloc
1269  *
1270  * This function allocates and initializes OSP device in memory according to
1271  * the config log.
1272  *
1273  * \param[in] env       execution environment
1274  * \param[in] type      device type of OSP
1275  * \param[in] lcfg      config log
1276  *
1277  * \retval pointer              the pointer of allocated OSP if succeed.
1278  * \retval ERR_PTR(errno)       ERR_PTR(errno) if failed.
1279  */
1280 static struct lu_device *osp_device_alloc(const struct lu_env *env,
1281                                           struct lu_device_type *type,
1282                                           struct lustre_cfg *lcfg)
1283 {
1284         struct osp_device *osp;
1285         struct lu_device  *ld;
1286
1287         OBD_ALLOC_PTR(osp);
1288         if (osp == NULL) {
1289                 ld = ERR_PTR(-ENOMEM);
1290         } else {
1291                 int rc;
1292
1293                 ld = osp2lu_dev(osp);
1294                 dt_device_init(&osp->opd_dt_dev, type);
1295                 rc = osp_init0(env, osp, type, lcfg);
1296                 if (rc != 0) {
1297                         osp_device_free(env, ld);
1298                         ld = ERR_PTR(rc);
1299                 }
1300         }
1301         return ld;
1302 }
1303
1304 /**
1305  * Implementation of lu_device_type_operations::ldto_device_fini
1306  *
1307  * This function cleans up the OSP device, i.e. release and free those
1308  * attached items in osp_device.
1309  *
1310  * \param[in] env       execution environment
1311  * \param[in] ld        lu_device of OSP
1312  *
1313  * \retval NULL                 NULL if cleanup succeeded.
1314  * \retval ERR_PTR(errno)       ERR_PTR(errno) if cleanup failed.
1315  */
1316 static struct lu_device *osp_device_fini(const struct lu_env *env,
1317                                          struct lu_device *ld)
1318 {
1319         struct osp_device *osp = lu2osp_dev(ld);
1320         int                rc;
1321
1322         ENTRY;
1323
1324         if (osp->opd_async_requests != NULL) {
1325                 osp_update_request_destroy(env, osp->opd_async_requests);
1326                 osp->opd_async_requests = NULL;
1327         }
1328
1329         if (osp->opd_storage_exp) {
1330                 /* wait for the commit callbacks to complete */
1331                 wait_event(osp->opd_sync_waitq,
1332                           atomic_read(&osp->opd_commits_registered) == 0);
1333                 obd_disconnect(osp->opd_storage_exp);
1334         }
1335
1336         LASSERT(osp->opd_obd);
1337         osp_tunables_fini(osp);
1338
1339         rc = client_obd_cleanup(osp->opd_obd);
1340         if (rc != 0) {
1341                 ptlrpcd_decref();
1342                 RETURN(ERR_PTR(rc));
1343         }
1344
1345         ptlrpcd_decref();
1346
1347         RETURN(NULL);
1348 }
1349
1350 /**
1351  * Implementation of obd_ops::o_reconnect
1352  *
1353  * This function is empty and does not need to do anything for now.
1354  */
1355 static int osp_reconnect(const struct lu_env *env,
1356                          struct obd_export *exp, struct obd_device *obd,
1357                          struct obd_uuid *cluuid,
1358                          struct obd_connect_data *data,
1359                          void *localdata)
1360 {
1361         return 0;
1362 }
1363
1364 /*
1365  * Implementation of obd_ops::o_connect
1366  *
1367  * Connect OSP to the remote target (MDT or OST). Allocate the
1368  * export and return it to the LOD, which calls this function
1369  * for each OSP to connect it to the remote target. This function
1370  * is currently only called once per OSP.
1371  *
1372  * \param[in] env       execution environment
1373  * \param[out] exp      export connected to OSP
1374  * \param[in] obd       OSP device
1375  * \param[in] cluuid    OSP device client uuid
1376  * \param[in] data      connect_data to be used to connect to the remote
1377  *                      target
1378  * \param[in] localdata necessary for the API interface, but not used in
1379  *                      this function
1380  *
1381  * \retval 0            0 if the connection succeeded.
1382  * \retval negative     negative errno if the connection failed.
1383  */
1384 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
1385                            struct obd_device *obd, struct obd_uuid *cluuid,
1386                            struct obd_connect_data *data, void *localdata)
1387 {
1388         struct osp_device       *osp = lu2osp_dev(obd->obd_lu_dev);
1389         struct obd_connect_data *ocd;
1390         struct obd_import       *imp;
1391         struct lustre_handle     conn;
1392         int                      rc;
1393
1394         ENTRY;
1395
1396         CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
1397
1398         rc = class_connect(&conn, obd, cluuid);
1399         if (rc)
1400                 RETURN(rc);
1401
1402         *exp = class_conn2export(&conn);
1403         /* Why should there ever be more than 1 connect? */
1404         osp->opd_connects++;
1405         LASSERT(osp->opd_connects == 1);
1406
1407         osp->opd_exp = *exp;
1408
1409         imp = osp->opd_obd->u.cli.cl_import;
1410         imp->imp_dlm_handle = conn;
1411
1412         LASSERT(data != NULL);
1413         LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
1414         ocd = &imp->imp_connect_data;
1415         *ocd = *data;
1416
1417         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
1418         imp->imp_connect_flags2_orig = ocd->ocd_connect_flags2;
1419
1420         ocd->ocd_version = LUSTRE_VERSION_CODE;
1421         ocd->ocd_index = data->ocd_index;
1422
1423         rc = ptlrpc_connect_import(imp);
1424         if (rc) {
1425                 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
1426                 GOTO(out, rc);
1427         } else {
1428                 osp->opd_obd->u.cli.cl_seq->lcs_exp =
1429                                 class_export_get(osp->opd_exp);
1430         }
1431
1432         ptlrpc_pinger_add_import(imp);
1433 out:
1434         RETURN(rc);
1435 }
1436
1437 /**
1438  * Implementation of obd_ops::o_disconnect
1439  *
1440  * Disconnect the export for the OSP.  This is called by LOD to release the
1441  * OSP during cleanup (\see lod_del_device()). The OSP will be released after
1442  * the export is released.
1443  *
1444  * \param[in] exp       export to be disconnected.
1445  *
1446  * \retval 0            0 if disconnection succeed
1447  * \retval negative     negative errno if disconnection failed
1448  */
1449 static int osp_obd_disconnect(struct obd_export *exp)
1450 {
1451         struct obd_device *obd = exp->exp_obd;
1452         struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
1453         int                rc;
1454         ENTRY;
1455
1456         /* Only disconnect the underlying layers on the final disconnect. */
1457         LASSERT(osp->opd_connects == 1);
1458         osp->opd_connects--;
1459
1460         rc = class_disconnect(exp);
1461         if (rc) {
1462                 CERROR("%s: class disconnect error: rc = %d\n",
1463                        obd->obd_name, rc);
1464                 RETURN(rc);
1465         }
1466
1467         /* destroy the device */
1468         class_manual_cleanup(obd);
1469
1470         RETURN(rc);
1471 }
1472
1473 /**
1474  * Implementation of obd_ops::o_statfs
1475  *
1476  * Send a RPC to the remote target to get statfs status. This is only used
1477  * in lprocfs helpers by obd_statfs.
1478  *
1479  * \param[in] env       execution environment
1480  * \param[in] exp       connection state from this OSP to the parent (LOD)
1481  *                      device
1482  * \param[out] osfs     hold the statfs result
1483  * \param[in] unused    Not used in this function for now
1484  * \param[in] flags     flags to indicate how OSP will issue the RPC
1485  *
1486  * \retval 0            0 if statfs succeeded.
1487  * \retval negative     negative errno if statfs failed.
1488  */
1489 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
1490                           struct obd_statfs *osfs, time64_t unused, __u32 flags)
1491 {
1492         struct obd_statfs       *msfs;
1493         struct ptlrpc_request   *req;
1494         struct obd_import       *imp = NULL;
1495         int                      rc;
1496
1497         ENTRY;
1498
1499         /* Since the request might also come from lprocfs, so we need
1500          * sync this with client_disconnect_export Bug15684 */
1501         down_read(&exp->exp_obd->u.cli.cl_sem);
1502         if (exp->exp_obd->u.cli.cl_import)
1503                 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
1504         up_read(&exp->exp_obd->u.cli.cl_sem);
1505         if (!imp)
1506                 RETURN(-ENODEV);
1507
1508         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1509
1510         class_import_put(imp);
1511
1512         if (req == NULL)
1513                 RETURN(-ENOMEM);
1514
1515         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1516         if (rc) {
1517                 ptlrpc_request_free(req);
1518                 RETURN(rc);
1519         }
1520         ptlrpc_request_set_replen(req);
1521         req->rq_request_portal = OST_CREATE_PORTAL;
1522         ptlrpc_at_set_req_timeout(req);
1523
1524         if (flags & OBD_STATFS_NODELAY) {
1525                 /* procfs requests not want stat in wait for avoid deadlock */
1526                 req->rq_no_resend = 1;
1527                 req->rq_no_delay = 1;
1528         }
1529
1530         rc = ptlrpc_queue_wait(req);
1531         if (rc)
1532                 GOTO(out, rc);
1533
1534         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1535         if (msfs == NULL)
1536                 GOTO(out, rc = -EPROTO);
1537
1538         *osfs = *msfs;
1539
1540         EXIT;
1541 out:
1542         ptlrpc_req_finished(req);
1543         return rc;
1544 }
1545
1546 /**
1547  * Implementation of obd_ops::o_import_event
1548  *
1549  * This function is called when some related import event happens. It will
1550  * mark the necessary flags according to the event and notify the necessary
1551  * threads (mainly precreate thread).
1552  *
1553  * \param[in] obd       OSP OBD device
1554  * \param[in] imp       import attached from OSP to remote (OST/MDT) service
1555  * \param[in] event     event related to remote service (IMP_EVENT_*)
1556  *
1557  * \retval 0            0 if the event handling succeeded.
1558  * \retval negative     negative errno if the event handling failed.
1559  */
1560 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1561                             enum obd_import_event event)
1562 {
1563         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1564         int rc;
1565
1566         switch (event) {
1567         case IMP_EVENT_DISCON:
1568                 d->opd_got_disconnected = 1;
1569                 d->opd_imp_connected = 0;
1570                 if (d->opd_connect_mdt)
1571                         break;
1572
1573                 if (d->opd_pre != NULL) {
1574                         osp_pre_update_status(d, -ENODEV);
1575                         wake_up(&d->opd_pre_waitq);
1576                 }
1577
1578                 CDEBUG(D_HA, "got disconnected\n");
1579                 break;
1580         case IMP_EVENT_INACTIVE:
1581                 d->opd_imp_active = 0;
1582                 d->opd_imp_connected = 0;
1583                 d->opd_obd->obd_inactive = 1;
1584                 if (d->opd_connect_mdt)
1585                         break;
1586                 if (d->opd_pre != NULL) {
1587                         /* Import is invalid, we can`t get stripes so
1588                          * wakeup waiters */
1589                         rc = imp->imp_deactive ? -ESHUTDOWN : -ENODEV;
1590                         osp_pre_update_status(d, rc);
1591                         wake_up(&d->opd_pre_waitq);
1592                 }
1593
1594                 CDEBUG(D_HA, "got inactive\n");
1595                 break;
1596         case IMP_EVENT_ACTIVE:
1597                 d->opd_imp_active = 1;
1598
1599                 if (d->opd_got_disconnected)
1600                         d->opd_new_connection = 1;
1601                 d->opd_imp_connected = 1;
1602                 d->opd_imp_seen_connected = 1;
1603                 d->opd_obd->obd_inactive = 0;
1604                 wake_up(&d->opd_pre_waitq);
1605                 if (d->opd_connect_mdt)
1606                         break;
1607
1608                 osp_sync_check_for_work(d);
1609                 CDEBUG(D_HA, "got connected\n");
1610                 break;
1611         case IMP_EVENT_INVALIDATE:
1612                 if (d->opd_connect_mdt)
1613                         osp_invalidate_request(d);
1614
1615                 if (obd->obd_namespace == NULL)
1616                         break;
1617                 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1618                 break;
1619         case IMP_EVENT_OCD:
1620         case IMP_EVENT_DEACTIVATE:
1621         case IMP_EVENT_ACTIVATE:
1622                 break;
1623         default:
1624                 CERROR("%s: unsupported import event: %#x\n",
1625                        obd->obd_name, event);
1626         }
1627         return 0;
1628 }
1629
1630 /**
1631  * Implementation of obd_ops: o_iocontrol
1632  *
1633  * This function is the ioctl handler for OSP. Note: lctl will access the OSP
1634  * directly by ioctl, instead of through the MDS stack.
1635  *
1636  * param[in] cmd        ioctl command.
1637  * param[in] exp        export of this OSP.
1638  * param[in] len        data length of \a karg.
1639  * param[in] karg       input argument which is packed as
1640  *                      obd_ioctl_data
1641  * param[out] uarg      pointer to userspace buffer (must access by
1642  *                      copy_to_user()).
1643  *
1644  * \retval 0            0 if the ioctl handling succeeded.
1645  * \retval negative     negative errno if the ioctl handling failed.
1646  */
1647 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1648                          void *karg, void __user *uarg)
1649 {
1650         struct obd_device       *obd = exp->exp_obd;
1651         struct osp_device       *d;
1652         struct obd_ioctl_data   *data = karg;
1653         int                      rc = 0;
1654
1655         ENTRY;
1656
1657         LASSERT(obd->obd_lu_dev);
1658         d = lu2osp_dev(obd->obd_lu_dev);
1659         LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1660
1661         if (!try_module_get(THIS_MODULE)) {
1662                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1663                        module_name(THIS_MODULE));
1664                 return -EINVAL;
1665         }
1666
1667         switch (cmd) {
1668         case OBD_IOC_CLIENT_RECOVER:
1669                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1670                                            data->ioc_inlbuf1, 0);
1671                 if (rc > 0)
1672                         rc = 0;
1673                 break;
1674         case IOC_OSC_SET_ACTIVE:
1675                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1676                                               data->ioc_offset);
1677                 break;
1678         default:
1679                 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1680                        cmd, current_comm());
1681                 rc = -ENOTTY;
1682         }
1683         module_put(THIS_MODULE);
1684         return rc;
1685 }
1686
1687
1688 /**
1689  * Implementation of obd_ops::o_get_info
1690  *
1691  * Retrieve information by key. Retrieval starts from the top layer
1692  * (MDT) of the MDS stack and traverses the stack by calling the
1693  * obd_get_info() method of the next sub-layer.
1694  *
1695  * \param[in] env       execution environment
1696  * \param[in] exp       export of this OSP
1697  * \param[in] keylen    length of \a key
1698  * \param[in] key       the key
1699  * \param[out] vallen   length of \a val
1700  * \param[out] val      holds the value returned by the key
1701  *
1702  * \retval 0            0 if getting information succeeded.
1703  * \retval negative     negative errno if getting information failed.
1704  */
1705 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1706                             __u32 keylen, void *key, __u32 *vallen, void *val)
1707 {
1708         int rc = -EINVAL;
1709
1710         if (KEY_IS(KEY_OSP_CONNECTED)) {
1711                 struct obd_device       *obd = exp->exp_obd;
1712                 struct osp_device       *osp;
1713
1714                 if (!obd->obd_set_up || obd->obd_stopping)
1715                         RETURN(-EAGAIN);
1716
1717                 osp = lu2osp_dev(obd->obd_lu_dev);
1718                 LASSERT(osp);
1719                 /*
1720                  * 1.8/2.0 behaviour is that OST being connected once at least
1721                  * is considered "healthy". and one "healthy" OST is enough to
1722                  * allow lustre clients to connect to MDS
1723                  */
1724                 RETURN(!osp->opd_imp_seen_connected);
1725         }
1726
1727         RETURN(rc);
1728 }
1729
1730 static int osp_obd_set_info_async(const struct lu_env *env,
1731                                   struct obd_export *exp,
1732                                   u32 keylen, void *key,
1733                                   u32 vallen, void *val,
1734                                   struct ptlrpc_request_set *set)
1735 {
1736         struct obd_device       *obd = exp->exp_obd;
1737         struct obd_import       *imp = obd->u.cli.cl_import;
1738         struct osp_device       *osp;
1739         struct ptlrpc_request   *req;
1740         char                    *tmp;
1741         int                      rc;
1742
1743         if (KEY_IS(KEY_SPTLRPC_CONF)) {
1744                 sptlrpc_conf_client_adapt(exp->exp_obd);
1745                 RETURN(0);
1746         }
1747
1748         LASSERT(set != NULL);
1749         if (!obd->obd_set_up || obd->obd_stopping)
1750                 RETURN(-EAGAIN);
1751         osp = lu2osp_dev(obd->obd_lu_dev);
1752
1753         req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
1754         if (req == NULL)
1755                 RETURN(-ENOMEM);
1756
1757         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
1758                              RCL_CLIENT, keylen);
1759         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
1760                              RCL_CLIENT, vallen);
1761         if (osp->opd_connect_mdt)
1762                 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SET_INFO);
1763         else
1764                 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
1765         if (rc) {
1766                 ptlrpc_request_free(req);
1767                 RETURN(rc);
1768         }
1769
1770         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1771         memcpy(tmp, key, keylen);
1772         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1773         memcpy(tmp, val, vallen);
1774
1775         ptlrpc_request_set_replen(req);
1776         ptlrpc_set_add_req(set, req);
1777         ptlrpc_check_set(NULL, set);
1778
1779         RETURN(0);
1780 }
1781
1782 /**
1783  * Implementation of obd_ops: o_fid_alloc
1784  *
1785  * Allocate a FID. There are two cases in which OSP performs
1786  * FID allocation.
1787  *
1788  * 1. FID precreation for data objects, which is done in
1789  *    osp_precreate_fids() instead of this function.
1790  * 2. FID allocation for each sub-stripe of a striped directory.
1791  *    Similar to other FID clients, OSP requests the sequence
1792  *    from its corresponding remote MDT, which in turn requests
1793  *    sequences from the sequence controller (MDT0).
1794  *
1795  * \param[in] env       execution environment
1796  * \param[in] exp       export of the OSP
1797  * \param[out] fid      FID being allocated
1798  * \param[in] unused    necessary for the interface but unused.
1799  *
1800  * \retval 0            0 FID allocated successfully.
1801  * \retval 1            1 FID allocated successfully and new sequence
1802  *                      requested from seq meta server
1803  * \retval negative     negative errno if FID allocation failed.
1804  */
1805 static int osp_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1806                          struct lu_fid *fid, struct md_op_data *unused)
1807 {
1808         struct client_obd       *cli = &exp->exp_obd->u.cli;
1809         struct osp_device       *osp = lu2osp_dev(exp->exp_obd->obd_lu_dev);
1810         struct lu_client_seq    *seq = cli->cl_seq;
1811         ENTRY;
1812
1813         LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
1814         /* Sigh, fid client is not ready yet */
1815         LASSERT(osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL);
1816
1817         RETURN(seq_client_alloc_fid(env, seq, fid));
1818 }
1819
1820 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1821 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1822 static void osp_key_exit(const struct lu_context *ctx,
1823                          struct lu_context_key *key, void *data)
1824 {
1825         struct osp_thread_info *info = data;
1826
1827         info->osi_attr.la_valid = 0;
1828 }
1829
1830 struct lu_context_key osp_thread_key = {
1831         .lct_tags = LCT_MD_THREAD,
1832         .lct_init = osp_key_init,
1833         .lct_fini = osp_key_fini,
1834         .lct_exit = osp_key_exit
1835 };
1836
1837 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1838 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1839
1840 struct lu_context_key osp_txn_key = {
1841         .lct_tags = LCT_OSP_THREAD,
1842         .lct_init = osp_txn_key_init,
1843         .lct_fini = osp_txn_key_fini
1844 };
1845 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1846
1847 static struct lu_device_type_operations osp_device_type_ops = {
1848         .ldto_init           = osp_type_init,
1849         .ldto_fini           = osp_type_fini,
1850
1851         .ldto_start          = osp_type_start,
1852         .ldto_stop           = osp_type_stop,
1853
1854         .ldto_device_alloc   = osp_device_alloc,
1855         .ldto_device_free    = osp_device_free,
1856
1857         .ldto_device_fini    = osp_device_fini
1858 };
1859
1860 static struct lu_device_type osp_device_type = {
1861         .ldt_tags     = LU_DEVICE_DT,
1862         .ldt_name     = LUSTRE_OSP_NAME,
1863         .ldt_ops      = &osp_device_type_ops,
1864         .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1865 };
1866
1867 static struct obd_ops osp_obd_device_ops = {
1868         .o_owner        = THIS_MODULE,
1869         .o_add_conn     = client_import_add_conn,
1870         .o_del_conn     = client_import_del_conn,
1871         .o_reconnect    = osp_reconnect,
1872         .o_connect      = osp_obd_connect,
1873         .o_disconnect   = osp_obd_disconnect,
1874         .o_get_info     = osp_obd_get_info,
1875         .o_set_info_async = osp_obd_set_info_async,
1876         .o_import_event = osp_import_event,
1877         .o_iocontrol    = osp_iocontrol,
1878         .o_statfs       = osp_obd_statfs,
1879         .o_fid_init     = client_fid_init,
1880         .o_fid_fini     = client_fid_fini,
1881         .o_fid_alloc    = osp_fid_alloc,
1882 };
1883
1884 static struct obd_type *sym;
1885
1886 /**
1887  * Initialize OSP module.
1888  *
1889  * Register device types OSP and Light Weight Proxy (LWP) (\see lwp_dev.c)
1890  * in obd_types (\see class_obd.c).  Initialize procfs for the
1891  * the OSP device.  Note: OSP was called OSC before Lustre 2.4,
1892  * so for compatibility it still uses the name "osc" in procfs.
1893  * This is called at module load time.
1894  *
1895  * \retval 0            0 if initialization succeeds.
1896  * \retval negative     negative errno if initialization failed.
1897  */
1898 static int __init osp_init(void)
1899 {
1900         struct dentry *symlink;
1901         struct obd_type *type;
1902         struct qstr dname;
1903         int rc;
1904
1905         rc = lu_kmem_init(osp_caches);
1906         if (rc)
1907                 return rc;
1908
1909         rc = class_register_type(&osp_obd_device_ops, NULL, true, NULL,
1910                                  LUSTRE_OSP_NAME, &osp_device_type);
1911         if (rc != 0) {
1912                 lu_kmem_fini(osp_caches);
1913                 return rc;
1914         }
1915
1916         rc = class_register_type(&lwp_obd_device_ops, NULL, false, NULL,
1917                                  LUSTRE_LWP_NAME, &lwp_device_type);
1918         if (rc != 0) {
1919                 class_unregister_type(LUSTRE_OSP_NAME);
1920                 lu_kmem_fini(osp_caches);
1921                 return rc;
1922         }
1923
1924         sym = class_setup_tunables(LUSTRE_OSC_NAME);
1925         if (IS_ERR(sym)) {
1926                 rc = PTR_ERR(sym);
1927                 /* does real "osc" already exist ? */
1928                 if (rc == -EEXIST)
1929                         GOTO(try_proc, rc = 0);
1930                 GOTO(no_osc, rc);
1931         }
1932
1933         /* create "osc" entry for compatibility purposes */
1934         dname.name = "osc";
1935         dname.len = strlen(dname.name);
1936         dname.hash = ll_full_name_hash(debugfs_lustre_root, dname.name,
1937                                        dname.len);
1938         symlink = d_lookup(debugfs_lustre_root, &dname);
1939         if (!symlink) {
1940                 symlink = debugfs_create_dir(dname.name, debugfs_lustre_root);
1941                 if (IS_ERR_OR_NULL(symlink)) {
1942                         rc = symlink ? PTR_ERR(symlink) : -ENOMEM;
1943                         GOTO(no_osc, rc);
1944                 }
1945                 sym->typ_debugfs_entry = symlink;
1946         } else {
1947                 dput(symlink);
1948         }
1949
1950 try_proc:
1951         type = class_search_type(LUSTRE_OSC_NAME);
1952         if (type != NULL && type->typ_procroot != NULL)
1953                 GOTO(no_osc, rc);
1954
1955         type = class_search_type(LUSTRE_OSP_NAME);
1956         type->typ_procsym = lprocfs_register("osc", proc_lustre_root,
1957                                              NULL, NULL);
1958         if (IS_ERR(type->typ_procsym)) {
1959                 CERROR("osp: can't create compat entry \"osc\": %d\n",
1960                        (int) PTR_ERR(type->typ_procsym));
1961                 type->typ_procsym = NULL;
1962         }
1963 no_osc:
1964         return rc;
1965 }
1966
1967 /**
1968  * Finalize OSP module.
1969  *
1970  * This callback is called when kernel unloads OSP module from memory, and
1971  * it will deregister OSP and LWP device type from obd_types (\see class_obd.c).
1972  */
1973 static void __exit osp_exit(void)
1974 {
1975         if (!IS_ERR_OR_NULL(sym)) {
1976                 ldebugfs_remove(&sym->typ_debugfs_entry);
1977                 kobject_put(&sym->typ_kobj);
1978         }
1979         class_unregister_type(LUSTRE_LWP_NAME);
1980         class_unregister_type(LUSTRE_OSP_NAME);
1981         lu_kmem_fini(osp_caches);
1982 }
1983
1984 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1985 MODULE_DESCRIPTION("Lustre OSD Storage Proxy ("LUSTRE_OSP_NAME")");
1986 MODULE_VERSION(LUSTRE_VERSION_STRING);
1987 MODULE_LICENSE("GPL");
1988
1989 module_init(osp_init);
1990 module_exit(osp_exit);