Whamcloud - gitweb
LU-7236 ptlrpc: idle connections can disconnect
[fs/lustre-release.git] / lustre / lov / lov_obd.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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 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/lov/lov_obd.c
33  *
34  * Author: Phil Schwan <phil@clusterfs.com>
35  * Author: Peter Braam <braam@clusterfs.com>
36  * Author: Mike Shaver <shaver@clusterfs.com>
37  * Author: Nathan Rutman <nathan@clusterfs.com>
38  */
39
40 #define DEBUG_SUBSYSTEM S_LOV
41 #include <libcfs/libcfs.h>
42
43 #include <cl_object.h>
44 #include <lustre_dlm.h>
45 #include <lustre_fid.h>
46 #include <uapi/linux/lustre/lustre_ioctl.h>
47 #include <lustre_lib.h>
48 #include <lustre_mds.h>
49 #include <lustre_net.h>
50 #include <uapi/linux/lustre/lustre_param.h>
51 #include <lustre_swab.h>
52 #include <lprocfs_status.h>
53 #include <obd_class.h>
54 #include <obd_support.h>
55
56 #include "lov_internal.h"
57
58 /* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion.
59    Any function that expects lov_tgts to remain stationary must take a ref. */
60 static void lov_getref(struct obd_device *obd)
61 {
62         struct lov_obd *lov = &obd->u.lov;
63
64         /* nobody gets through here until lov_putref is done */
65         mutex_lock(&lov->lov_lock);
66         atomic_inc(&lov->lov_refcount);
67         mutex_unlock(&lov->lov_lock);
68         return;
69 }
70
71 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt);
72
73 static void lov_putref(struct obd_device *obd)
74 {
75         struct lov_obd *lov = &obd->u.lov;
76
77         mutex_lock(&lov->lov_lock);
78         /* ok to dec to 0 more than once -- ltd_exp's will be null */
79         if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
80                 struct list_head kill = LIST_HEAD_INIT(kill);
81                 struct lov_tgt_desc *tgt, *n;
82                 int i;
83
84                 CDEBUG(D_CONFIG, "destroying %d lov targets\n",
85                        lov->lov_death_row);
86                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
87                         tgt = lov->lov_tgts[i];
88
89                         if (!tgt || !tgt->ltd_reap)
90                                 continue;
91                         list_add(&tgt->ltd_kill, &kill);
92                         /* XXX - right now there is a dependency on ld_tgt_count
93                          * being the maximum tgt index for computing the
94                          * mds_max_easize. So we can't shrink it. */
95                         lov_ost_pool_remove(&lov->lov_packed, i);
96                         lov->lov_tgts[i] = NULL;
97                         lov->lov_death_row--;
98                 }
99                 mutex_unlock(&lov->lov_lock);
100
101                 list_for_each_entry_safe(tgt, n, &kill, ltd_kill) {
102                         list_del(&tgt->ltd_kill);
103                         /* Disconnect */
104                         __lov_del_obd(obd, tgt);
105                 }
106         } else {
107                 mutex_unlock(&lov->lov_lock);
108         }
109 }
110
111 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
112                               enum obd_notify_event ev);
113 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
114                       enum obd_notify_event ev);
115
116 int lov_connect_obd(struct obd_device *obd, u32 index, int activate,
117                     struct obd_connect_data *data)
118 {
119         struct lov_obd *lov = &obd->u.lov;
120         struct obd_uuid *tgt_uuid;
121         struct obd_device *tgt_obd;
122         static struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
123         struct obd_import *imp;
124         int rc;
125         ENTRY;
126
127         if (lov->lov_tgts[index] == NULL)
128                 RETURN(-EINVAL);
129
130         tgt_uuid = &lov->lov_tgts[index]->ltd_uuid;
131         tgt_obd = lov->lov_tgts[index]->ltd_obd;
132
133         if (!tgt_obd->obd_set_up) {
134                 CERROR("Target %s not set up\n", obd_uuid2str(tgt_uuid));
135                 RETURN(-EINVAL);
136         }
137
138         /* override the sp_me from lov */
139         tgt_obd->u.cli.cl_sp_me = lov->lov_sp_me;
140
141         if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX))
142                 data->ocd_index = index;
143
144         /*
145          * Divine LOV knows that OBDs under it are OSCs.
146          */
147         imp = tgt_obd->u.cli.cl_import;
148
149         if (activate) {
150                 tgt_obd->obd_no_recov = 0;
151                 /* FIXME this is probably supposed to be
152                    ptlrpc_set_import_active.  Horrible naming. */
153                 ptlrpc_activate_import(imp);
154         }
155
156         rc = obd_register_observer(tgt_obd, obd);
157         if (rc) {
158                 CERROR("Target %s register_observer error %d\n",
159                        obd_uuid2str(tgt_uuid), rc);
160                 RETURN(rc);
161         }
162
163
164         if (imp->imp_invalid) {
165                 CDEBUG(D_CONFIG, "not connecting OSC %s; administratively "
166                        "disabled\n", obd_uuid2str(tgt_uuid));
167                 RETURN(0);
168         }
169
170         rc = obd_connect(NULL, &lov->lov_tgts[index]->ltd_exp, tgt_obd,
171                          &lov_osc_uuid, data, NULL);
172         if (rc || !lov->lov_tgts[index]->ltd_exp) {
173                 CERROR("Target %s connect error %d\n",
174                        obd_uuid2str(tgt_uuid), rc);
175                 RETURN(-ENODEV);
176         }
177
178         lov->lov_tgts[index]->ltd_reap = 0;
179
180         CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index,
181                obd_uuid2str(tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
182
183         if (lov->lov_tgts_kobj) {
184                 /* Even if we failed, that's ok */
185                 rc = sysfs_create_link(lov->lov_tgts_kobj,
186                                        &tgt_obd->obd_kset.kobj,
187                                        tgt_obd->obd_name);
188                 if (rc) {
189                         CERROR("%s: can't register LOV target /sys/fs/lustre/%s/%s/target_obds/%s : rc = %d\n",
190                                obd->obd_name, obd->obd_type->typ_name,
191                                obd->obd_name,
192                                lov->lov_tgts[index]->ltd_exp->exp_obd->obd_name,
193                                rc);
194                 }
195         }
196         RETURN(0);
197 }
198
199 static int lov_connect(const struct lu_env *env,
200                        struct obd_export **exp, struct obd_device *obd,
201                        struct obd_uuid *cluuid, struct obd_connect_data *data,
202                        void *localdata)
203 {
204         struct lov_obd *lov = &obd->u.lov;
205         struct lov_tgt_desc *tgt;
206         struct lustre_handle conn;
207         int i, rc;
208         ENTRY;
209
210         CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects);
211
212         rc = class_connect(&conn, obd, cluuid);
213         if (rc)
214                 RETURN(rc);
215
216         *exp = class_conn2export(&conn);
217
218         /* Why should there ever be more than 1 connect? */
219         lov->lov_connects++;
220         LASSERT(lov->lov_connects == 1);
221
222         memset(&lov->lov_ocd, 0, sizeof(lov->lov_ocd));
223         if (data)
224                 lov->lov_ocd = *data;
225
226         obd_getref(obd);
227
228         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
229                 tgt = lov->lov_tgts[i];
230                 if (!tgt || obd_uuid_empty(&tgt->ltd_uuid))
231                         continue;
232                 /* Flags will be lowest common denominator */
233                 rc = lov_connect_obd(obd, i, tgt->ltd_activate, &lov->lov_ocd);
234                 if (rc) {
235                         CERROR("%s: lov connect tgt %d failed: %d\n",
236                                obd->obd_name, i, rc);
237                         continue;
238                 }
239                 /* connect to administrative disabled ost */
240                 if (!lov->lov_tgts[i]->ltd_exp)
241                         continue;
242
243                 rc = lov_notify(obd, lov->lov_tgts[i]->ltd_exp->exp_obd,
244                                 OBD_NOTIFY_CONNECT);
245                 if (rc) {
246                         CERROR("%s error sending notify %d\n",
247                                obd->obd_name, rc);
248                 }
249         }
250         obd_putref(obd);
251
252         RETURN(0);
253 }
254
255 static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
256 {
257         struct lov_obd *lov = &obd->u.lov;
258         struct obd_device *osc_obd;
259         int rc;
260         ENTRY;
261
262         osc_obd = class_exp2obd(tgt->ltd_exp);
263         CDEBUG(D_CONFIG, "%s: disconnecting target %s\n",
264                obd->obd_name, osc_obd->obd_name);
265
266         if (tgt->ltd_active) {
267                 tgt->ltd_active = 0;
268                 lov->desc.ld_active_tgt_count--;
269                 tgt->ltd_exp->exp_obd->obd_inactive = 1;
270         }
271
272         if (osc_obd) {
273                 if (lov->lov_tgts_kobj)
274                         sysfs_remove_link(lov->lov_tgts_kobj,
275                                           osc_obd->obd_name);
276
277                 /* Pass it on to our clients.
278                  * XXX This should be an argument to disconnect,
279                  * XXX not a back-door flag on the OBD.  Ah well.
280                  */
281                 osc_obd->obd_force = obd->obd_force;
282                 osc_obd->obd_fail = obd->obd_fail;
283                 osc_obd->obd_no_recov = obd->obd_no_recov;
284
285                 if (lov->targets_proc_entry != NULL)
286                         lprocfs_remove_proc_entry(osc_obd->obd_name,
287                                                   lov->targets_proc_entry);
288         }
289
290         obd_register_observer(osc_obd, NULL);
291
292         rc = obd_disconnect(tgt->ltd_exp);
293         if (rc) {
294                 CERROR("Target %s disconnect error %d\n",
295                        tgt->ltd_uuid.uuid, rc);
296                 rc = 0;
297         }
298
299         tgt->ltd_exp = NULL;
300         RETURN(0);
301 }
302
303 static int lov_disconnect(struct obd_export *exp)
304 {
305         struct obd_device *obd = class_exp2obd(exp);
306         struct lov_obd *lov = &obd->u.lov;
307         u32 index;
308         int rc;
309
310         ENTRY;
311         if (!lov->lov_tgts)
312                 goto out;
313
314         /* Only disconnect the underlying layers on the final disconnect. */
315         lov->lov_connects--;
316         if (lov->lov_connects != 0) {
317                 /* why should there be more than 1 connect? */
318                 CWARN("%s: unexpected disconnect #%d\n",
319                       obd->obd_name, lov->lov_connects);
320                 goto out;
321         }
322
323         /* hold another ref so lov_del_obd() doesn't spin in putref each time */
324         obd_getref(obd);
325
326         for (index = 0; index < lov->desc.ld_tgt_count; index++) {
327                 if (lov->lov_tgts[index] && lov->lov_tgts[index]->ltd_exp) {
328                         /* Disconnection is the last we know about an OBD */
329                         lov_del_target(obd, index, NULL,
330                                        lov->lov_tgts[index]->ltd_gen);
331                 }
332         }
333         obd_putref(obd);
334
335 out:
336         rc = class_disconnect(exp); /* bz 9811 */
337         RETURN(rc);
338 }
339
340 /* Error codes:
341  *
342  *  -EINVAL  : UUID can't be found in the LOV's target list
343  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
344  *  -EBADF   : The UUID is found, but the OBD is the wrong type (!)
345  *  any >= 0 : is log target index
346  */
347 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
348                               enum obd_notify_event ev)
349 {
350         struct lov_obd *lov = &obd->u.lov;
351         struct lov_tgt_desc *tgt;
352         int index, activate, active;
353         ENTRY;
354
355         CDEBUG(D_INFO, "Searching in lov %p for uuid %s event(%d)\n",
356                lov, uuid->uuid, ev);
357
358         obd_getref(obd);
359         for (index = 0; index < lov->desc.ld_tgt_count; index++) {
360                 tgt = lov->lov_tgts[index];
361                 if (!tgt)
362                         continue;
363                 /*
364                  * LU-642, initially inactive OSC could miss the obd_connect,
365                  * we make up for it here.
366                  */
367                 if (ev == OBD_NOTIFY_ACTIVATE && tgt->ltd_exp == NULL &&
368                     obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
369                         struct obd_uuid lov_osc_uuid = {"LOV_OSC_UUID"};
370
371                         obd_connect(NULL, &tgt->ltd_exp, tgt->ltd_obd,
372                                     &lov_osc_uuid, &lov->lov_ocd, NULL);
373                 }
374                 if (!tgt->ltd_exp)
375                         continue;
376
377                 CDEBUG(D_INFO, "lov idx %d is %s conn %#llx\n",
378                        index, obd_uuid2str(&tgt->ltd_uuid),
379                        tgt->ltd_exp->exp_handle.h_cookie);
380                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
381                         break;
382         }
383
384         if (index == lov->desc.ld_tgt_count)
385                 GOTO(out, index = -EINVAL);
386
387         if (ev == OBD_NOTIFY_DEACTIVATE || ev == OBD_NOTIFY_ACTIVATE) {
388                 activate = (ev == OBD_NOTIFY_ACTIVATE) ? 1 : 0;
389
390                 if (lov->lov_tgts[index]->ltd_activate == activate) {
391                         CDEBUG(D_INFO, "OSC %s already %sactivate!\n",
392                                uuid->uuid, activate ? "" : "de");
393                 } else {
394                         lov->lov_tgts[index]->ltd_activate = activate;
395                         CDEBUG(D_CONFIG, "%sactivate OSC %s\n",
396                                activate ? "" : "de", obd_uuid2str(uuid));
397                 }
398
399         } else if (ev == OBD_NOTIFY_INACTIVE || ev == OBD_NOTIFY_ACTIVE) {
400                 active = (ev == OBD_NOTIFY_ACTIVE) ? 1 : 0;
401
402                 if (lov->lov_tgts[index]->ltd_active == active) {
403                         CDEBUG(D_INFO, "OSC %s already %sactive!\n",
404                                uuid->uuid, active ? "" : "in");
405                         GOTO(out, index);
406                 } else {
407                         CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n",
408                                obd_uuid2str(uuid), active ? "" : "in");
409                 }
410
411                 lov->lov_tgts[index]->ltd_active = active;
412                 if (active) {
413                         lov->desc.ld_active_tgt_count++;
414                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
415                 } else {
416                         lov->desc.ld_active_tgt_count--;
417                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
418                 }
419         } else {
420                 CERROR("%s: unknown event %d for uuid %s\n", obd->obd_name,
421                        ev, uuid->uuid);
422         }
423
424  out:
425         obd_putref(obd);
426         RETURN(index);
427 }
428
429 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
430                       enum obd_notify_event ev)
431 {
432         int rc = 0;
433         struct lov_obd *lov = &obd->u.lov;
434         ENTRY;
435
436         down_read(&lov->lov_notify_lock);
437         if (!lov->lov_connects)
438                 GOTO(out_notify_lock, rc = 0);
439
440         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE ||
441             ev == OBD_NOTIFY_ACTIVATE || ev == OBD_NOTIFY_DEACTIVATE) {
442                 struct obd_uuid *uuid;
443
444                 LASSERT(watched);
445
446                 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
447                         CERROR("unexpected notification of %s %s\n",
448                                watched->obd_type->typ_name, watched->obd_name);
449                         GOTO(out_notify_lock, rc = -EINVAL);
450                 }
451
452                 uuid = &watched->u.cli.cl_target_uuid;
453
454                 /* Set OSC as active before notifying the observer, so the
455                  * observer can use the OSC normally.
456                  */
457                 rc = lov_set_osc_active(obd, uuid, ev);
458                 if (rc < 0) {
459                         CERROR("%s: event %d failed: rc = %d\n", obd->obd_name,
460                                ev, rc);
461                         GOTO(out_notify_lock, rc);
462                 }
463         }
464
465         /* Pass the notification up the chain. */
466         rc = obd_notify_observer(obd, watched, ev);
467
468 out_notify_lock:
469         up_read(&lov->lov_notify_lock);
470
471         RETURN(rc);
472 }
473
474 static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
475                           u32 index, int gen, int active)
476 {
477         struct lov_obd *lov = &obd->u.lov;
478         struct lov_tgt_desc *tgt;
479         struct obd_device *tgt_obd;
480         int rc;
481
482         ENTRY;
483         CDEBUG(D_CONFIG, "uuid:%s idx:%u gen:%d active:%d\n",
484                uuidp->uuid, index, gen, active);
485
486         if (gen <= 0) {
487                 CERROR("%s: request to add '%s' with invalid generation: %d\n",
488                        obd->obd_name, uuidp->uuid, gen);
489                 RETURN(-EINVAL);
490         }
491
492         tgt_obd = class_find_client_obd(uuidp, LUSTRE_OSC_NAME, &obd->obd_uuid);
493         if (tgt_obd == NULL)
494                 RETURN(-EINVAL);
495
496         mutex_lock(&lov->lov_lock);
497
498         if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) {
499                 tgt = lov->lov_tgts[index];
500                 rc = -EEXIST;
501                 CERROR("%s: UUID %s already assigned at index %d: rc = %d\n",
502                        obd->obd_name, obd_uuid2str(&tgt->ltd_uuid), index, rc);
503                 mutex_unlock(&lov->lov_lock);
504                 RETURN(rc);
505         }
506
507         if (index >= lov->lov_tgt_size) {
508                 /* We need to reallocate the lov target array. */
509                 struct lov_tgt_desc **newtgts, **old = NULL;
510                 __u32 newsize, oldsize = 0;
511
512                 newsize = max(lov->lov_tgt_size, (__u32)2);
513                 while (newsize < index + 1)
514                         newsize = newsize << 1;
515                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
516                 if (newtgts == NULL) {
517                         mutex_unlock(&lov->lov_lock);
518                         RETURN(-ENOMEM);
519                 }
520
521                 if (lov->lov_tgt_size) {
522                         memcpy(newtgts, lov->lov_tgts, sizeof(*newtgts) *
523                                lov->lov_tgt_size);
524                         old = lov->lov_tgts;
525                         oldsize = lov->lov_tgt_size;
526                 }
527
528                 lov->lov_tgts = newtgts;
529                 lov->lov_tgt_size = newsize;
530                 smp_rmb();
531                 if (old)
532                         OBD_FREE(old, sizeof(*old) * oldsize);
533
534                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n",
535                        lov->lov_tgts, lov->lov_tgt_size);
536         }
537
538         OBD_ALLOC_PTR(tgt);
539         if (!tgt) {
540                 mutex_unlock(&lov->lov_lock);
541                 RETURN(-ENOMEM);
542         }
543
544         rc = lov_ost_pool_add(&lov->lov_packed, index, lov->lov_tgt_size);
545         if (rc) {
546                 mutex_unlock(&lov->lov_lock);
547                 OBD_FREE_PTR(tgt);
548                 RETURN(rc);
549         }
550
551         tgt->ltd_uuid = *uuidp;
552         tgt->ltd_obd = tgt_obd;
553         /* XXX - add a sanity check on the generation number. */
554         tgt->ltd_gen = gen;
555         tgt->ltd_index = index;
556         tgt->ltd_activate = active;
557         lov->lov_tgts[index] = tgt;
558         if (index >= lov->desc.ld_tgt_count)
559                 lov->desc.ld_tgt_count = index + 1;
560
561         mutex_unlock(&lov->lov_lock);
562
563         CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n",
564                 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
565
566         if (lov->lov_connects == 0) {
567                 /* lov_connect hasn't been called yet. We'll do the
568                    lov_connect_obd on this target when that fn first runs,
569                    because we don't know the connect flags yet. */
570                 RETURN(0);
571         }
572
573         obd_getref(obd);
574
575         rc = lov_connect_obd(obd, index, active, &lov->lov_ocd);
576         if (rc)
577                 GOTO(out, rc);
578
579         /* connect to administrative disabled ost */
580         if (!tgt->ltd_exp)
581                 GOTO(out, rc = 0);
582
583         if (lov->lov_cache != NULL) {
584                 rc = obd_set_info_async(NULL, tgt->ltd_exp,
585                                 sizeof(KEY_CACHE_SET), KEY_CACHE_SET,
586                                 sizeof(struct cl_client_cache), lov->lov_cache,
587                                 NULL);
588                 if (rc < 0)
589                         GOTO(out, rc);
590         }
591
592         rc = lov_notify(obd, tgt->ltd_exp->exp_obd,
593                         active ? OBD_NOTIFY_CONNECT : OBD_NOTIFY_INACTIVE);
594
595 out:
596         if (rc) {
597                 CERROR("%s: add failed, deleting %s: rc = %d\n",
598                        obd->obd_name, obd_uuid2str(&tgt->ltd_uuid), rc);
599                 lov_del_target(obd, index, NULL, 0);
600         }
601         obd_putref(obd);
602         RETURN(rc);
603 }
604
605 /* Schedule a target for deletion */
606 int lov_del_target(struct obd_device *obd, u32 index,
607                    struct obd_uuid *uuidp, int gen)
608 {
609         struct lov_obd *lov = &obd->u.lov;
610         int count = lov->desc.ld_tgt_count;
611         int rc = 0;
612         ENTRY;
613
614         if (index >= count) {
615                 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
616                        index, count);
617                 RETURN(-EINVAL);
618         }
619
620         /* to make sure there's no ongoing lov_notify() now */
621         down_write(&lov->lov_notify_lock);
622         obd_getref(obd);
623
624         if (!lov->lov_tgts[index]) {
625                 CERROR("LOV target at index %d is not setup.\n", index);
626                 GOTO(out, rc = -EINVAL);
627         }
628
629         if (uuidp && !obd_uuid_equals(uuidp, &lov->lov_tgts[index]->ltd_uuid)) {
630                 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
631                        lov_uuid2str(lov, index), index,
632                        obd_uuid2str(uuidp));
633                 GOTO(out, rc = -EINVAL);
634         }
635
636         CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
637                lov_uuid2str(lov, index), index,
638                lov->lov_tgts[index]->ltd_gen, lov->lov_tgts[index]->ltd_exp,
639                lov->lov_tgts[index]->ltd_active);
640
641         lov->lov_tgts[index]->ltd_reap = 1;
642         lov->lov_death_row++;
643         /* we really delete it from obd_putref */
644 out:
645         obd_putref(obd);
646         up_write(&lov->lov_notify_lock);
647
648         RETURN(rc);
649 }
650
651 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
652 {
653         struct obd_device *osc_obd;
654
655         LASSERT(tgt);
656         LASSERT(tgt->ltd_reap);
657
658         osc_obd = class_exp2obd(tgt->ltd_exp);
659
660         CDEBUG(D_CONFIG, "Removing tgt %s : %s\n",
661                tgt->ltd_uuid.uuid,
662                osc_obd ? osc_obd->obd_name : "<no obd>");
663
664         if (tgt->ltd_exp)
665                 lov_disconnect_obd(obd, tgt);
666
667         OBD_FREE_PTR(tgt);
668
669         /* Manual cleanup - no cleanup logs to clean up the osc's.  We must
670            do it ourselves. And we can't do it from lov_cleanup,
671            because we just lost our only reference to it. */
672         if (osc_obd)
673                 class_manual_cleanup(osc_obd);
674 }
675
676 void lov_fix_desc_stripe_size(__u64 *val)
677 {
678         if (*val < LOV_MIN_STRIPE_SIZE) {
679                 if (*val != 0)
680                         LCONSOLE_INFO("Increasing default stripe size to "
681                                       "minimum %u\n",
682                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
683                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
684         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
685                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
686                 LCONSOLE_WARN("Changing default stripe size to %llu (a "
687                               "multiple of %u)\n",
688                               *val, LOV_MIN_STRIPE_SIZE);
689         }
690 }
691
692 void lov_fix_desc_stripe_count(__u32 *val)
693 {
694         if (*val == 0)
695                 *val = 1;
696 }
697
698 void lov_fix_desc_pattern(__u32 *val)
699 {
700         /* from lov_setstripe */
701         if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
702                 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
703                 *val = 0;
704         }
705 }
706
707 void lov_fix_desc_qos_maxage(__u32 *val)
708 {
709         if (*val == 0)
710                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
711 }
712
713 void lov_fix_desc(struct lov_desc *desc)
714 {
715         lov_fix_desc_stripe_size(&desc->ld_default_stripe_size);
716         lov_fix_desc_stripe_count(&desc->ld_default_stripe_count);
717         lov_fix_desc_pattern(&desc->ld_pattern);
718         lov_fix_desc_qos_maxage(&desc->ld_qos_maxage);
719 }
720
721 int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
722 {
723         struct lov_desc *desc;
724         struct lov_obd *lov = &obd->u.lov;
725         int rc;
726         ENTRY;
727
728         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
729                 CERROR("LOV setup requires a descriptor\n");
730                 RETURN(-EINVAL);
731         }
732
733         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
734
735         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
736                 CERROR("descriptor size wrong: %d > %d\n",
737                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
738                 RETURN(-EINVAL);
739         }
740
741         if (desc->ld_magic != LOV_DESC_MAGIC) {
742                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
743                             CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
744                                    obd->obd_name, desc);
745                             lustre_swab_lov_desc(desc);
746                 } else {
747                         CERROR("%s: Bad lov desc magic: %#x\n",
748                                obd->obd_name, desc->ld_magic);
749                         RETURN(-EINVAL);
750                 }
751         }
752
753         lov_fix_desc(desc);
754
755         desc->ld_active_tgt_count = 0;
756         lov->desc = *desc;
757         lov->lov_tgt_size = 0;
758
759         mutex_init(&lov->lov_lock);
760         atomic_set(&lov->lov_refcount, 0);
761         lov->lov_sp_me = LUSTRE_SP_CLI;
762
763         init_rwsem(&lov->lov_notify_lock);
764
765         lov->lov_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
766                                                    HASH_POOLS_MAX_BITS,
767                                                    HASH_POOLS_BKT_BITS, 0,
768                                                    CFS_HASH_MIN_THETA,
769                                                    CFS_HASH_MAX_THETA,
770                                                    &pool_hash_operations,
771                                                    CFS_HASH_DEFAULT);
772         INIT_LIST_HEAD(&lov->lov_pool_list);
773         lov->lov_pool_count = 0;
774         rc = lov_ost_pool_init(&lov->lov_packed, 0);
775         if (rc)
776                 GOTO(out, rc);
777
778         rc = lov_tunables_init(obd);
779         if (rc)
780                 GOTO(out, rc);
781
782         lov->lov_tgts_kobj = kobject_create_and_add("target_obds",
783                                                     &obd->obd_kset.kobj);
784
785 out:
786         return rc;
787 }
788
789 static int lov_cleanup(struct obd_device *obd)
790 {
791         struct lov_obd *lov = &obd->u.lov;
792         struct list_head *pos, *tmp;
793         struct pool_desc *pool;
794         ENTRY;
795
796         if (lov->lov_tgts_kobj) {
797                 kobject_put(lov->lov_tgts_kobj);
798                 lov->lov_tgts_kobj = NULL;
799         }
800
801         list_for_each_safe(pos, tmp, &lov->lov_pool_list) {
802                 pool = list_entry(pos, struct pool_desc, pool_list);
803                 /* free pool structs */
804                 CDEBUG(D_INFO, "delete pool %p\n", pool);
805                 /* In the function below, .hs_keycmp resolves to
806                  * pool_hashkey_keycmp() */
807                 /* coverity[overrun-buffer-val] */
808                 lov_pool_del(obd, pool->pool_name);
809         }
810         cfs_hash_putref(lov->lov_pools_hash_body);
811         lov_ost_pool_free(&lov->lov_packed);
812
813         lprocfs_obd_cleanup(obd);
814         if (lov->lov_tgts) {
815                 int i;
816                 obd_getref(obd);
817                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
818                         if (!lov->lov_tgts[i])
819                                 continue;
820
821                         /* Inactive targets may never have connected */
822                         if (lov->lov_tgts[i]->ltd_active ||
823                             atomic_read(&lov->lov_refcount))
824                                 /* We should never get here - these
825                                  * should have been removed in the
826                                  * disconnect. */
827                                 CERROR("%s: lov tgt %d not cleaned! "
828                                        "deathrow=%d, lovrc=%d\n",
829                                        obd->obd_name, i, lov->lov_death_row,
830                                        atomic_read(&lov->lov_refcount));
831                         lov_del_target(obd, i, NULL, 0);
832                 }
833                 obd_putref(obd);
834                 OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) *
835                          lov->lov_tgt_size);
836                 lov->lov_tgt_size = 0;
837         }
838
839         if (lov->lov_cache != NULL) {
840                 cl_cache_decref(lov->lov_cache);
841                 lov->lov_cache = NULL;
842         }
843
844         RETURN(0);
845 }
846
847 int lov_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg,
848                             u32 *indexp, int *genp)
849 {
850         struct obd_uuid obd_uuid;
851         int cmd;
852         int rc = 0;
853
854         ENTRY;
855         switch (cmd = lcfg->lcfg_command) {
856         case LCFG_ADD_MDC:
857         case LCFG_DEL_MDC:
858                 break;
859         case LCFG_LOV_ADD_OBD:
860         case LCFG_LOV_ADD_INA:
861         case LCFG_LOV_DEL_OBD: {
862                 u32 index;
863                 int gen;
864
865                 /* lov_modify_tgts add  0:lov_mdsA  1:ost1_UUID  2:0  3:1 */
866                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
867                         GOTO(out, rc = -EINVAL);
868
869                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
870
871                 rc = kstrtou32(lustre_cfg_buf(lcfg, 2), 10, indexp);
872                 if (rc)
873                         GOTO(out, rc);
874                 rc = kstrtoint(lustre_cfg_buf(lcfg, 3), 10, genp);
875                 if (rc)
876                         GOTO(out, rc);
877                 index = *indexp;
878                 gen = *genp;
879                 if (cmd == LCFG_LOV_ADD_OBD)
880                         rc = lov_add_target(obd, &obd_uuid, index, gen, 1);
881                 else if (cmd == LCFG_LOV_ADD_INA)
882                         rc = lov_add_target(obd, &obd_uuid, index, gen, 0);
883                 else
884                         rc = lov_del_target(obd, index, &obd_uuid, gen);
885
886                 GOTO(out, rc);
887         }
888         case LCFG_PARAM: {
889                 struct lov_desc *desc = &(obd->u.lov.desc);
890                 ssize_t count;
891
892                 if (!desc)
893                         GOTO(out, rc = -EINVAL);
894
895                 count = class_modify_config(lcfg, PARAM_LOV,
896                                             &obd->obd_kset.kobj);
897                 GOTO(out, rc = count < 0 ? count : 0);
898         }
899         case LCFG_POOL_NEW:
900         case LCFG_POOL_ADD:
901         case LCFG_POOL_DEL:
902         case LCFG_POOL_REM:
903                 GOTO(out, rc);
904
905         default: {
906                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
907                 GOTO(out, rc = -EINVAL);
908
909         }
910         }
911 out:
912         RETURN(rc);
913 }
914
915 static int lov_statfs(const struct lu_env *env, struct obd_export *exp,
916                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
917 {
918         struct obd_device *obd = class_exp2obd(exp);
919         struct lov_obd *lov = &obd->u.lov;
920         struct obd_info oinfo = {
921                 .oi_osfs = osfs,
922                 .oi_flags = flags,
923         };
924         struct ptlrpc_request_set *rqset;
925         struct lov_request_set *set = NULL;
926         struct lov_request *req;
927         int rc = 0;
928         int rc2;
929
930         ENTRY;
931
932         rqset = ptlrpc_prep_set();
933         if (rqset == NULL)
934                 RETURN(-ENOMEM);
935
936         rc = lov_prep_statfs_set(obd, &oinfo, &set);
937         if (rc < 0)
938                 GOTO(out_rqset, rc);
939
940         list_for_each_entry(req, &set->set_list, rq_link) {
941                 rc = obd_statfs_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
942                                       &req->rq_oi, max_age, rqset);
943                 if (rc < 0)
944                         GOTO(out_set, rc);
945         }
946
947         rc = ptlrpc_set_wait(rqset);
948
949 out_set:
950         if (rc < 0)
951                 atomic_set(&set->set_completes, 0);
952
953         rc2 = lov_fini_statfs_set(set);
954         if (rc == 0)
955                 rc = rc2;
956
957 out_rqset:
958         ptlrpc_set_destroy(rqset);
959
960         RETURN(rc);
961 }
962
963 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
964                          void *karg, void __user *uarg)
965 {
966         struct obd_device *obddev = class_exp2obd(exp);
967         struct lov_obd *lov = &obddev->u.lov;
968         int i = 0, rc = 0, count = lov->desc.ld_tgt_count;
969         struct obd_uuid *uuidp;
970         ENTRY;
971
972         switch (cmd) {
973         case IOC_OBD_STATFS: {
974                 struct obd_ioctl_data *data = karg;
975                 struct obd_device *osc_obd;
976                 struct obd_statfs stat_buf = {0};
977                 struct obd_import *imp;
978                 __u32 index;
979                 __u32 flags;
980
981                 memcpy(&index, data->ioc_inlbuf2, sizeof(index));
982                 if ((index >= count))
983                         RETURN(-ENODEV);
984
985                 if (!lov->lov_tgts[index])
986                         /* Try again with the next index */
987                         RETURN(-EAGAIN);
988                 imp = lov->lov_tgts[index]->ltd_exp->exp_obd->u.cli.cl_import;
989                 if (!lov->lov_tgts[index]->ltd_active &&
990                     imp->imp_state != LUSTRE_IMP_IDLE)
991                         RETURN(-ENODATA);
992
993                 osc_obd = class_exp2obd(lov->lov_tgts[index]->ltd_exp);
994                 if (!osc_obd)
995                         RETURN(-EINVAL);
996
997                 /* copy UUID */
998                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(osc_obd),
999                                  min_t(unsigned long, data->ioc_plen2,
1000                                        sizeof(struct obd_uuid))))
1001                         RETURN(-EFAULT);
1002
1003                 memcpy(&flags, data->ioc_inlbuf1, sizeof(flags));
1004                 flags = flags & LL_STATFS_NODELAY ? OBD_STATFS_NODELAY : 0;
1005
1006                 /* got statfs data */
1007                 rc = obd_statfs(NULL, lov->lov_tgts[index]->ltd_exp, &stat_buf,
1008                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
1009                                 flags);
1010                 if (rc)
1011                         RETURN(rc);
1012                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1013                                  min_t(unsigned long, data->ioc_plen1,
1014                                        sizeof(struct obd_statfs))))
1015                         RETURN(-EFAULT);
1016                 break;
1017         }
1018         case OBD_IOC_LOV_GET_CONFIG: {
1019                 struct obd_ioctl_data *data;
1020                 struct lov_desc *desc;
1021                 char *buf = NULL;
1022                 __u32 *genp;
1023
1024                 len = 0;
1025                 if (obd_ioctl_getdata(&buf, &len, uarg))
1026                         RETURN(-EINVAL);
1027
1028                 data = (struct obd_ioctl_data *)buf;
1029
1030                 if (sizeof(*desc) > data->ioc_inllen1) {
1031                         OBD_FREE_LARGE(buf, len);
1032                         RETURN(-EINVAL);
1033                 }
1034
1035                 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
1036                         OBD_FREE_LARGE(buf, len);
1037                         RETURN(-EINVAL);
1038                 }
1039
1040                 if (sizeof(__u32) * count > data->ioc_inllen3) {
1041                         OBD_FREE_LARGE(buf, len);
1042                         RETURN(-EINVAL);
1043                 }
1044
1045                 desc = (struct lov_desc *)data->ioc_inlbuf1;
1046                 memcpy(desc, &(lov->desc), sizeof(*desc));
1047
1048                 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
1049                 genp = (__u32 *)data->ioc_inlbuf3;
1050                 /* the uuid will be empty for deleted OSTs */
1051                 for (i = 0; i < count; i++, uuidp++, genp++) {
1052                         if (!lov->lov_tgts[i])
1053                                 continue;
1054                         *uuidp = lov->lov_tgts[i]->ltd_uuid;
1055                         *genp = lov->lov_tgts[i]->ltd_gen;
1056                 }
1057
1058                 if (copy_to_user(uarg, buf, len))
1059                         rc = -EFAULT;
1060                 OBD_FREE_LARGE(buf, len);
1061                 break;
1062         }
1063         case OBD_IOC_QUOTACTL: {
1064                 struct if_quotactl *qctl = karg;
1065                 struct lov_tgt_desc *tgt = NULL;
1066                 struct obd_quotactl *oqctl;
1067
1068                 if (qctl->qc_valid == QC_OSTIDX) {
1069                         if (count <= qctl->qc_idx)
1070                                 RETURN(-EINVAL);
1071
1072                         tgt = lov->lov_tgts[qctl->qc_idx];
1073                         if (!tgt || !tgt->ltd_exp)
1074                                 RETURN(-EINVAL);
1075                 } else if (qctl->qc_valid == QC_UUID) {
1076                         for (i = 0; i < count; i++) {
1077                                 tgt = lov->lov_tgts[i];
1078                                 if (!tgt ||
1079                                     !obd_uuid_equals(&tgt->ltd_uuid,
1080                                                      &qctl->obd_uuid))
1081                                         continue;
1082
1083                                 if (tgt->ltd_exp == NULL)
1084                                         RETURN(-EINVAL);
1085
1086                                 break;
1087                         }
1088                 } else {
1089                         RETURN(-EINVAL);
1090                 }
1091
1092                 if (i >= count)
1093                         RETURN(-EAGAIN);
1094
1095                 LASSERT(tgt && tgt->ltd_exp);
1096                 OBD_ALLOC_PTR(oqctl);
1097                 if (!oqctl)
1098                         RETURN(-ENOMEM);
1099
1100                 QCTL_COPY(oqctl, qctl);
1101                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1102                 if (rc == 0) {
1103                         QCTL_COPY(qctl, oqctl);
1104                         qctl->qc_valid = QC_OSTIDX;
1105                         qctl->obd_uuid = tgt->ltd_uuid;
1106                 }
1107                 OBD_FREE_PTR(oqctl);
1108                 break;
1109         }
1110         default: {
1111                 int set = 0;
1112
1113                 if (count == 0)
1114                         RETURN(-ENOTTY);
1115
1116                 for (i = 0; i < count; i++) {
1117                         int err;
1118                         struct obd_device *osc_obd;
1119
1120                         /* OST was disconnected */
1121                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
1122                                 continue;
1123
1124                         /* ll_umount_begin() sets force flag but for lov, not
1125                          * osc. Let's pass it through */
1126                         osc_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
1127                         osc_obd->obd_force = obddev->obd_force;
1128                         err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp,
1129                                             len, karg, uarg);
1130                         if (err) {
1131                                 if (lov->lov_tgts[i]->ltd_active) {
1132                                         CDEBUG(err == -ENOTTY ?
1133                                                D_IOCTL : D_WARNING,
1134                                                "iocontrol OSC %s on OST "
1135                                                "idx %d cmd %x: err = %d\n",
1136                                                lov_uuid2str(lov, i),
1137                                                i, cmd, err);
1138                                         if (!rc)
1139                                                 rc = err;
1140                                 }
1141                         } else {
1142                                 set = 1;
1143                         }
1144                 }
1145                 if (!set && !rc)
1146                         rc = -EIO;
1147         }
1148         }
1149
1150         RETURN(rc);
1151 }
1152
1153 static int lov_get_info(const struct lu_env *env, struct obd_export *exp,
1154                         __u32 keylen, void *key, __u32 *vallen, void *val)
1155 {
1156         struct obd_device *obddev = class_exp2obd(exp);
1157         struct lov_obd *lov = &obddev->u.lov;
1158         struct lov_desc *ld = &lov->desc;
1159         int rc = 0;
1160         ENTRY;
1161
1162         if (vallen == NULL || val == NULL)
1163                 RETURN(-EFAULT);
1164
1165         obd_getref(obddev);
1166
1167         if (KEY_IS(KEY_MAX_EASIZE)) {
1168                 u32 max_stripe_count = min_t(u32, ld->ld_active_tgt_count,
1169                                              LOV_MAX_STRIPE_COUNT);
1170
1171                 *((u32 *)val) = lov_mds_md_size(max_stripe_count, LOV_MAGIC_V3);
1172         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
1173                 u32 def_stripe_count = min_t(u32, ld->ld_default_stripe_count,
1174                                              LOV_MAX_STRIPE_COUNT);
1175
1176                 *((u32 *)val) = lov_mds_md_size(def_stripe_count, LOV_MAGIC_V3);
1177         } else if (KEY_IS(KEY_TGT_COUNT)) {
1178                 *((int *)val) = lov->desc.ld_tgt_count;
1179         } else {
1180                 rc = -EINVAL;
1181         }
1182
1183         obd_putref(obddev);
1184
1185         RETURN(rc);
1186 }
1187
1188 static int lov_set_info_async(const struct lu_env *env, struct obd_export *exp,
1189                               __u32 keylen, void *key,
1190                               __u32 vallen, void *val,
1191                               struct ptlrpc_request_set *set)
1192 {
1193         struct obd_device *obddev = class_exp2obd(exp);
1194         struct lov_obd *lov = &obddev->u.lov;
1195         struct lov_tgt_desc *tgt;
1196         bool do_inactive = false, no_set = false;
1197         u32 i;
1198         int rc = 0;
1199         int err;
1200
1201         ENTRY;
1202
1203         if (set == NULL) {
1204                 no_set = true;
1205                 set = ptlrpc_prep_set();
1206                 if (!set)
1207                         RETURN(-ENOMEM);
1208         }
1209
1210         obd_getref(obddev);
1211
1212         if (KEY_IS(KEY_CHECKSUM)) {
1213                 do_inactive = true;
1214         } else if (KEY_IS(KEY_CACHE_SET)) {
1215                 LASSERT(lov->lov_cache == NULL);
1216                 lov->lov_cache = val;
1217                 do_inactive = true;
1218                 cl_cache_incref(lov->lov_cache);
1219         }
1220
1221         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1222                 tgt = lov->lov_tgts[i];
1223
1224                 /* OST was disconnected */
1225                 if (tgt == NULL || tgt->ltd_exp == NULL)
1226                         continue;
1227
1228                 /* OST is inactive and we don't want inactive OSCs */
1229                 if (!tgt->ltd_active && !do_inactive)
1230                         continue;
1231
1232                 err = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1233                                          vallen, val, set);
1234
1235                 if (rc == 0)
1236                         rc = err;
1237         }
1238
1239         /* cycle through MDC target for Data-on-MDT */
1240         for (i = 0; i < LOV_MDC_TGT_MAX; i++) {
1241                 struct obd_device *mdc;
1242
1243                 mdc = lov->lov_mdc_tgts[i].lmtd_mdc;
1244                 if (mdc == NULL)
1245                         continue;
1246
1247                 err = obd_set_info_async(env, mdc->obd_self_export,
1248                                          keylen, key, vallen, val, set);
1249                 if (rc == 0)
1250                         rc = err;
1251         }
1252
1253         obd_putref(obddev);
1254         if (no_set) {
1255                 err = ptlrpc_set_wait(set);
1256                 if (rc == 0)
1257                         rc = err;
1258                 ptlrpc_set_destroy(set);
1259         }
1260         RETURN(rc);
1261 }
1262
1263 void lov_stripe_lock(struct lov_stripe_md *md)
1264 __acquires(&md->lsm_lock)
1265 {
1266         LASSERT(md->lsm_lock_owner != current_pid());
1267         spin_lock(&md->lsm_lock);
1268         LASSERT(md->lsm_lock_owner == 0);
1269         md->lsm_lock_owner = current_pid();
1270 }
1271
1272 void lov_stripe_unlock(struct lov_stripe_md *md)
1273 __releases(&md->lsm_lock)
1274 {
1275         LASSERT(md->lsm_lock_owner == current_pid());
1276         md->lsm_lock_owner = 0;
1277         spin_unlock(&md->lsm_lock);
1278 }
1279
1280 static int lov_quotactl(struct obd_device *obd, struct obd_export *exp,
1281                         struct obd_quotactl *oqctl)
1282 {
1283         struct lov_obd      *lov = &obd->u.lov;
1284         struct lov_tgt_desc *tgt;
1285         __u64                curspace = 0;
1286         __u64                bhardlimit = 0;
1287         int                  i, rc = 0;
1288         ENTRY;
1289
1290         if (oqctl->qc_cmd != Q_GETOQUOTA &&
1291             oqctl->qc_cmd != LUSTRE_Q_SETQUOTA) {
1292                 CERROR("%s: bad quota opc %x for lov obd\n",
1293                        obd->obd_name, oqctl->qc_cmd);
1294                 RETURN(-EFAULT);
1295         }
1296
1297         /* for lov tgt */
1298         obd_getref(obd);
1299         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1300                 int err;
1301
1302                 tgt = lov->lov_tgts[i];
1303
1304                 if (!tgt)
1305                         continue;
1306
1307                 if (!tgt->ltd_active || tgt->ltd_reap) {
1308                         if (oqctl->qc_cmd == Q_GETOQUOTA &&
1309                             lov->lov_tgts[i]->ltd_activate) {
1310                                 rc = -ENETDOWN;
1311                                 CERROR("ost %d is inactive\n", i);
1312                         } else {
1313                                 CDEBUG(D_HA, "ost %d is inactive\n", i);
1314                         }
1315                         continue;
1316                 }
1317
1318                 err = obd_quotactl(tgt->ltd_exp, oqctl);
1319                 if (err) {
1320                         if (tgt->ltd_active && !rc)
1321                                 rc = err;
1322                         continue;
1323                 }
1324
1325                 if (oqctl->qc_cmd == Q_GETOQUOTA) {
1326                         curspace += oqctl->qc_dqblk.dqb_curspace;
1327                         bhardlimit += oqctl->qc_dqblk.dqb_bhardlimit;
1328                 }
1329         }
1330         obd_putref(obd);
1331
1332         if (oqctl->qc_cmd == Q_GETOQUOTA) {
1333                 oqctl->qc_dqblk.dqb_curspace = curspace;
1334                 oqctl->qc_dqblk.dqb_bhardlimit = bhardlimit;
1335         }
1336         RETURN(rc);
1337 }
1338
1339 static struct obd_ops lov_obd_ops = {
1340         .o_owner                = THIS_MODULE,
1341         .o_setup                = lov_setup,
1342         .o_cleanup              = lov_cleanup,
1343         .o_connect              = lov_connect,
1344         .o_disconnect           = lov_disconnect,
1345         .o_statfs               = lov_statfs,
1346         .o_iocontrol            = lov_iocontrol,
1347         .o_get_info             = lov_get_info,
1348         .o_set_info_async       = lov_set_info_async,
1349         .o_notify               = lov_notify,
1350         .o_pool_new             = lov_pool_new,
1351         .o_pool_rem             = lov_pool_remove,
1352         .o_pool_add             = lov_pool_add,
1353         .o_pool_del             = lov_pool_del,
1354         .o_getref               = lov_getref,
1355         .o_putref               = lov_putref,
1356         .o_quotactl             = lov_quotactl,
1357 };
1358
1359 struct kmem_cache *lov_oinfo_slab;
1360
1361 static int __init lov_init(void)
1362 {
1363         bool enable_proc = true;
1364         struct obd_type *type;
1365         int rc;
1366         ENTRY;
1367
1368         /* print an address of _any_ initialized kernel symbol from this
1369          * module, to allow debugging with gdb that doesn't support data
1370          * symbols from modules.*/
1371         CDEBUG(D_INFO, "Lustre LOV module (%p).\n", &lov_caches);
1372
1373         rc = lu_kmem_init(lov_caches);
1374         if (rc)
1375                 return rc;
1376
1377         lov_oinfo_slab = kmem_cache_create("lov_oinfo",
1378                                            sizeof(struct lov_oinfo), 0,
1379                                            SLAB_HWCACHE_ALIGN, NULL);
1380         if (lov_oinfo_slab == NULL) {
1381                 lu_kmem_fini(lov_caches);
1382                 return -ENOMEM;
1383         }
1384
1385         type = class_search_type(LUSTRE_LOD_NAME);
1386         if (type != NULL && type->typ_procsym != NULL)
1387                 enable_proc = false;
1388
1389         rc = class_register_type(&lov_obd_ops, NULL, enable_proc, NULL,
1390                                  LUSTRE_LOV_NAME, &lov_device_type);
1391
1392         if (rc) {
1393                 kmem_cache_destroy(lov_oinfo_slab);
1394                 lu_kmem_fini(lov_caches);
1395         }
1396
1397         RETURN(rc);
1398 }
1399
1400 static void __exit lov_exit(void)
1401 {
1402         class_unregister_type(LUSTRE_LOV_NAME);
1403         kmem_cache_destroy(lov_oinfo_slab);
1404         lu_kmem_fini(lov_caches);
1405 }
1406
1407 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1408 MODULE_DESCRIPTION("Lustre Logical Object Volume");
1409 MODULE_VERSION(LUSTRE_VERSION_STRING);
1410 MODULE_LICENSE("GPL");
1411
1412 module_init(lov_init);
1413 module_exit(lov_exit);