4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/lov/lov_obd.c
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>
40 #define DEBUG_SUBSYSTEM S_LOV
41 #include <libcfs/libcfs.h>
43 #include <lustre/lustre_idl.h>
45 #include <cl_object.h>
46 #include <lustre_dlm.h>
47 #include <lustre_fid.h>
48 #include <lustre_ioctl.h>
49 #include <lustre_lib.h>
50 #include <lustre_mds.h>
51 #include <lustre_net.h>
52 #include <lustre_param.h>
53 #include <lustre_swab.h>
54 #include <lprocfs_status.h>
55 #include <obd_class.h>
56 #include <obd_support.h>
58 #include "lov_internal.h"
60 /* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion.
61 Any function that expects lov_tgts to remain stationary must take a ref. */
62 static void lov_getref(struct obd_device *obd)
64 struct lov_obd *lov = &obd->u.lov;
66 /* nobody gets through here until lov_putref is done */
67 mutex_lock(&lov->lov_lock);
68 atomic_inc(&lov->lov_refcount);
69 mutex_unlock(&lov->lov_lock);
73 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt);
75 static void lov_putref(struct obd_device *obd)
77 struct lov_obd *lov = &obd->u.lov;
79 mutex_lock(&lov->lov_lock);
80 /* ok to dec to 0 more than once -- ltd_exp's will be null */
81 if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
82 struct list_head kill = LIST_HEAD_INIT(kill);
83 struct lov_tgt_desc *tgt, *n;
86 CDEBUG(D_CONFIG, "destroying %d lov targets\n",
88 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
89 tgt = lov->lov_tgts[i];
91 if (!tgt || !tgt->ltd_reap)
93 list_add(&tgt->ltd_kill, &kill);
94 /* XXX - right now there is a dependency on ld_tgt_count
95 * being the maximum tgt index for computing the
96 * mds_max_easize. So we can't shrink it. */
97 lov_ost_pool_remove(&lov->lov_packed, i);
98 lov->lov_tgts[i] = NULL;
101 mutex_unlock(&lov->lov_lock);
103 list_for_each_entry_safe(tgt, n, &kill, ltd_kill) {
104 list_del(&tgt->ltd_kill);
106 __lov_del_obd(obd, tgt);
109 mutex_unlock(&lov->lov_lock);
113 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
114 enum obd_notify_event ev);
115 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
116 enum obd_notify_event ev, void *data);
118 int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
119 struct obd_connect_data *data)
121 struct lov_obd *lov = &obd->u.lov;
122 struct obd_uuid *tgt_uuid;
123 struct obd_device *tgt_obd;
124 static struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
125 struct obd_import *imp;
129 if (lov->lov_tgts[index] == NULL)
132 tgt_uuid = &lov->lov_tgts[index]->ltd_uuid;
133 tgt_obd = lov->lov_tgts[index]->ltd_obd;
135 if (!tgt_obd->obd_set_up) {
136 CERROR("Target %s not set up\n", obd_uuid2str(tgt_uuid));
140 /* override the sp_me from lov */
141 tgt_obd->u.cli.cl_sp_me = lov->lov_sp_me;
143 if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX))
144 data->ocd_index = index;
147 * Divine LOV knows that OBDs under it are OSCs.
149 imp = tgt_obd->u.cli.cl_import;
152 tgt_obd->obd_no_recov = 0;
153 /* FIXME this is probably supposed to be
154 ptlrpc_set_import_active. Horrible naming. */
155 ptlrpc_activate_import(imp);
158 rc = obd_register_observer(tgt_obd, obd);
160 CERROR("Target %s register_observer error %d\n",
161 obd_uuid2str(tgt_uuid), rc);
166 if (imp->imp_invalid) {
167 CDEBUG(D_CONFIG, "not connecting OSC %s; administratively "
168 "disabled\n", obd_uuid2str(tgt_uuid));
172 rc = obd_connect(NULL, &lov->lov_tgts[index]->ltd_exp, tgt_obd,
173 &lov_osc_uuid, data, NULL);
174 if (rc || !lov->lov_tgts[index]->ltd_exp) {
175 CERROR("Target %s connect error %d\n",
176 obd_uuid2str(tgt_uuid), rc);
180 lov->lov_tgts[index]->ltd_reap = 0;
182 CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index,
183 obd_uuid2str(tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
185 if (lov->targets_proc_entry != NULL) {
186 struct proc_dir_entry *osc_symlink;
187 struct obd_device *osc_obd;
189 osc_obd = lov->lov_tgts[index]->ltd_exp->exp_obd;
191 LASSERT(osc_obd != NULL);
192 LASSERT(osc_obd->obd_magic == OBD_DEVICE_MAGIC);
193 LASSERT(osc_obd->obd_type->typ_name != NULL);
195 osc_symlink = lprocfs_add_symlink(osc_obd->obd_name,
196 lov->targets_proc_entry,
198 osc_obd->obd_type->typ_name,
200 if (osc_symlink == NULL) {
201 CERROR("cannot register LOV target "
202 "/proc/fs/lustre/%s/%s/target_obds/%s\n",
203 obd->obd_type->typ_name, obd->obd_name,
210 static int lov_connect(const struct lu_env *env,
211 struct obd_export **exp, struct obd_device *obd,
212 struct obd_uuid *cluuid, struct obd_connect_data *data,
215 struct lov_obd *lov = &obd->u.lov;
216 struct lov_tgt_desc *tgt;
217 struct lustre_handle conn;
221 CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects);
223 rc = class_connect(&conn, obd, cluuid);
227 *exp = class_conn2export(&conn);
229 /* Why should there ever be more than 1 connect? */
231 LASSERT(lov->lov_connects == 1);
233 memset(&lov->lov_ocd, 0, sizeof(lov->lov_ocd));
235 lov->lov_ocd = *data;
237 lov->targets_proc_entry = lprocfs_register("target_obds",
240 if (IS_ERR(lov->targets_proc_entry)) {
241 CERROR("%s: cannot register "
242 "/proc/fs/lustre/%s/%s/target_obds\n",
243 obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
244 lov->targets_proc_entry = NULL;
248 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
249 tgt = lov->lov_tgts[i];
250 if (!tgt || obd_uuid_empty(&tgt->ltd_uuid))
252 /* Flags will be lowest common denominator */
253 rc = lov_connect_obd(obd, i, tgt->ltd_activate, &lov->lov_ocd);
255 CERROR("%s: lov connect tgt %d failed: %d\n",
256 obd->obd_name, i, rc);
259 /* connect to administrative disabled ost */
260 if (!lov->lov_tgts[i]->ltd_exp)
263 rc = lov_notify(obd, lov->lov_tgts[i]->ltd_exp->exp_obd,
264 OBD_NOTIFY_CONNECT, (void *)&i);
266 CERROR("%s error sending notify %d\n",
275 static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
277 struct lov_obd *lov = &obd->u.lov;
278 struct obd_device *osc_obd;
282 osc_obd = class_exp2obd(tgt->ltd_exp);
283 CDEBUG(D_CONFIG, "%s: disconnecting target %s\n",
284 obd->obd_name, osc_obd->obd_name);
286 if (tgt->ltd_active) {
288 lov->desc.ld_active_tgt_count--;
289 tgt->ltd_exp->exp_obd->obd_inactive = 1;
293 /* Pass it on to our clients.
294 * XXX This should be an argument to disconnect,
295 * XXX not a back-door flag on the OBD. Ah well.
297 osc_obd->obd_force = obd->obd_force;
298 osc_obd->obd_fail = obd->obd_fail;
299 osc_obd->obd_no_recov = obd->obd_no_recov;
301 if (lov->targets_proc_entry != NULL)
302 lprocfs_remove_proc_entry(osc_obd->obd_name,
303 lov->targets_proc_entry);
306 obd_register_observer(osc_obd, NULL);
308 rc = obd_disconnect(tgt->ltd_exp);
310 CERROR("Target %s disconnect error %d\n",
311 tgt->ltd_uuid.uuid, rc);
319 static int lov_disconnect(struct obd_export *exp)
321 struct obd_device *obd = class_exp2obd(exp);
322 struct lov_obd *lov = &obd->u.lov;
329 /* Only disconnect the underlying layers on the final disconnect. */
331 if (lov->lov_connects != 0) {
332 /* why should there be more than 1 connect? */
333 CERROR("disconnect #%d\n", lov->lov_connects);
337 /* Let's hold another reference so lov_del_obd doesn't spin through
341 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
342 if (lov->lov_tgts[i] && lov->lov_tgts[i]->ltd_exp) {
343 /* Disconnection is the last we know about an obd */
344 lov_del_target(obd, i, NULL, lov->lov_tgts[i]->ltd_gen);
349 if (lov->targets_proc_entry != NULL)
350 lprocfs_remove(&lov->targets_proc_entry);
353 rc = class_disconnect(exp); /* bz 9811 */
359 * -EINVAL : UUID can't be found in the LOV's target list
360 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
361 * -EBADF : The UUID is found, but the OBD is the wrong type (!)
362 * any >= 0 : is log target index
364 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
365 enum obd_notify_event ev)
367 struct lov_obd *lov = &obd->u.lov;
368 struct lov_tgt_desc *tgt;
369 int index, activate, active;
372 CDEBUG(D_INFO, "Searching in lov %p for uuid %s event(%d)\n",
373 lov, uuid->uuid, ev);
376 for (index = 0; index < lov->desc.ld_tgt_count; index++) {
377 tgt = lov->lov_tgts[index];
381 * LU-642, initially inactive OSC could miss the obd_connect,
382 * we make up for it here.
384 if (ev == OBD_NOTIFY_ACTIVATE && tgt->ltd_exp == NULL &&
385 obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
386 struct obd_uuid lov_osc_uuid = {"LOV_OSC_UUID"};
388 obd_connect(NULL, &tgt->ltd_exp, tgt->ltd_obd,
389 &lov_osc_uuid, &lov->lov_ocd, NULL);
394 CDEBUG(D_INFO, "lov idx %d is %s conn %#llx\n",
395 index, obd_uuid2str(&tgt->ltd_uuid),
396 tgt->ltd_exp->exp_handle.h_cookie);
397 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
401 if (index == lov->desc.ld_tgt_count)
402 GOTO(out, index = -EINVAL);
404 if (ev == OBD_NOTIFY_DEACTIVATE || ev == OBD_NOTIFY_ACTIVATE) {
405 activate = (ev == OBD_NOTIFY_ACTIVATE) ? 1 : 0;
407 if (lov->lov_tgts[index]->ltd_activate == activate) {
408 CDEBUG(D_INFO, "OSC %s already %sactivate!\n",
409 uuid->uuid, activate ? "" : "de");
411 lov->lov_tgts[index]->ltd_activate = activate;
412 CDEBUG(D_CONFIG, "%sactivate OSC %s\n",
413 activate ? "" : "de", obd_uuid2str(uuid));
416 } else if (ev == OBD_NOTIFY_INACTIVE || ev == OBD_NOTIFY_ACTIVE) {
417 active = (ev == OBD_NOTIFY_ACTIVE) ? 1 : 0;
419 if (lov->lov_tgts[index]->ltd_active == active) {
420 CDEBUG(D_INFO, "OSC %s already %sactive!\n",
421 uuid->uuid, active ? "" : "in");
424 CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n",
425 obd_uuid2str(uuid), active ? "" : "in");
428 lov->lov_tgts[index]->ltd_active = active;
430 lov->desc.ld_active_tgt_count++;
431 lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
433 lov->desc.ld_active_tgt_count--;
434 lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
437 CERROR("%s: unknown event %d for uuid %s\n", obd->obd_name,
446 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
447 enum obd_notify_event ev, void *data)
450 struct lov_obd *lov = &obd->u.lov;
453 down_read(&lov->lov_notify_lock);
454 if (!lov->lov_connects) {
455 up_read(&lov->lov_notify_lock);
459 if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE ||
460 ev == OBD_NOTIFY_ACTIVATE || ev == OBD_NOTIFY_DEACTIVATE) {
461 struct obd_uuid *uuid;
465 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
466 up_read(&lov->lov_notify_lock);
467 CERROR("unexpected notification of %s %s!\n",
468 watched->obd_type->typ_name,
472 uuid = &watched->u.cli.cl_target_uuid;
474 /* Set OSC as active before notifying the observer, so the
475 * observer can use the OSC normally.
477 rc = lov_set_osc_active(obd, uuid, ev);
479 up_read(&lov->lov_notify_lock);
480 CERROR("event(%d) of %s failed: %d\n", ev,
481 obd_uuid2str(uuid), rc);
484 /* active event should be pass lov target index as data */
488 /* Pass the notification up the chain. */
490 rc = obd_notify_observer(obd, watched, ev, data);
492 /* NULL watched means all osc's in the lov (only for syncs) */
493 /* sync event should be send lov idx as data */
494 struct lov_obd *lov = &obd->u.lov;
498 is_sync = (ev == OBD_NOTIFY_SYNC) ||
499 (ev == OBD_NOTIFY_SYNC_NONBLOCK);
502 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
503 if (!lov->lov_tgts[i])
506 /* don't send sync event if target not
507 * connected/activated */
508 if (is_sync && !lov->lov_tgts[i]->ltd_active)
511 rc = obd_notify_observer(obd, lov->lov_tgts[i]->ltd_obd,
514 CERROR("%s: notify %s of %s failed %d\n",
516 obd->obd_observer->obd_name,
517 lov->lov_tgts[i]->ltd_obd->obd_name,
524 up_read(&lov->lov_notify_lock);
528 static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
529 __u32 index, int gen, int active)
531 struct lov_obd *lov = &obd->u.lov;
532 struct lov_tgt_desc *tgt;
533 struct obd_device *tgt_obd;
537 CDEBUG(D_CONFIG, "uuid:%s idx:%d gen:%d active:%d\n",
538 uuidp->uuid, index, gen, active);
541 CERROR("request to add OBD %s with invalid generation: %d\n",
546 tgt_obd = class_find_client_obd(uuidp, LUSTRE_OSC_NAME,
551 mutex_lock(&lov->lov_lock);
553 if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) {
554 tgt = lov->lov_tgts[index];
555 CERROR("UUID %s already assigned at LOV target index %d\n",
556 obd_uuid2str(&tgt->ltd_uuid), index);
557 mutex_unlock(&lov->lov_lock);
561 if (index >= lov->lov_tgt_size) {
562 /* We need to reallocate the lov target array. */
563 struct lov_tgt_desc **newtgts, **old = NULL;
564 __u32 newsize, oldsize = 0;
566 newsize = max(lov->lov_tgt_size, (__u32)2);
567 while (newsize < index + 1)
568 newsize = newsize << 1;
569 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
570 if (newtgts == NULL) {
571 mutex_unlock(&lov->lov_lock);
575 if (lov->lov_tgt_size) {
576 memcpy(newtgts, lov->lov_tgts, sizeof(*newtgts) *
579 oldsize = lov->lov_tgt_size;
582 lov->lov_tgts = newtgts;
583 lov->lov_tgt_size = newsize;
586 OBD_FREE(old, sizeof(*old) * oldsize);
588 CDEBUG(D_CONFIG, "tgts: %p size: %d\n",
589 lov->lov_tgts, lov->lov_tgt_size);
594 mutex_unlock(&lov->lov_lock);
598 rc = lov_ost_pool_add(&lov->lov_packed, index, lov->lov_tgt_size);
600 mutex_unlock(&lov->lov_lock);
605 tgt->ltd_uuid = *uuidp;
606 tgt->ltd_obd = tgt_obd;
607 /* XXX - add a sanity check on the generation number. */
609 tgt->ltd_index = index;
610 tgt->ltd_activate = active;
611 lov->lov_tgts[index] = tgt;
612 if (index >= lov->desc.ld_tgt_count)
613 lov->desc.ld_tgt_count = index + 1;
615 mutex_unlock(&lov->lov_lock);
617 CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n",
618 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
620 if (lov->lov_connects == 0) {
621 /* lov_connect hasn't been called yet. We'll do the
622 lov_connect_obd on this target when that fn first runs,
623 because we don't know the connect flags yet. */
629 rc = lov_connect_obd(obd, index, active, &lov->lov_ocd);
633 /* connect to administrative disabled ost */
637 if (lov->lov_cache != NULL) {
638 rc = obd_set_info_async(NULL, tgt->ltd_exp,
639 sizeof(KEY_CACHE_SET), KEY_CACHE_SET,
640 sizeof(struct cl_client_cache), lov->lov_cache,
646 rc = lov_notify(obd, tgt->ltd_exp->exp_obd,
647 active ? OBD_NOTIFY_CONNECT : OBD_NOTIFY_INACTIVE,
652 CERROR("add failed (%d), deleting %s\n", rc,
653 obd_uuid2str(&tgt->ltd_uuid));
654 lov_del_target(obd, index, NULL, 0);
660 /* Schedule a target for deletion */
661 int lov_del_target(struct obd_device *obd, __u32 index,
662 struct obd_uuid *uuidp, int gen)
664 struct lov_obd *lov = &obd->u.lov;
665 int count = lov->desc.ld_tgt_count;
669 if (index >= count) {
670 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
675 /* to make sure there's no ongoing lov_notify() now */
676 down_write(&lov->lov_notify_lock);
679 if (!lov->lov_tgts[index]) {
680 CERROR("LOV target at index %d is not setup.\n", index);
681 GOTO(out, rc = -EINVAL);
684 if (uuidp && !obd_uuid_equals(uuidp, &lov->lov_tgts[index]->ltd_uuid)) {
685 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
686 lov_uuid2str(lov, index), index,
687 obd_uuid2str(uuidp));
688 GOTO(out, rc = -EINVAL);
691 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
692 lov_uuid2str(lov, index), index,
693 lov->lov_tgts[index]->ltd_gen, lov->lov_tgts[index]->ltd_exp,
694 lov->lov_tgts[index]->ltd_active);
696 lov->lov_tgts[index]->ltd_reap = 1;
697 lov->lov_death_row++;
698 /* we really delete it from obd_putref */
701 up_write(&lov->lov_notify_lock);
706 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
708 struct obd_device *osc_obd;
711 LASSERT(tgt->ltd_reap);
713 osc_obd = class_exp2obd(tgt->ltd_exp);
715 CDEBUG(D_CONFIG, "Removing tgt %s : %s\n",
717 osc_obd ? osc_obd->obd_name : "<no obd>");
720 lov_disconnect_obd(obd, tgt);
724 /* Manual cleanup - no cleanup logs to clean up the osc's. We must
725 do it ourselves. And we can't do it from lov_cleanup,
726 because we just lost our only reference to it. */
728 class_manual_cleanup(osc_obd);
731 void lov_fix_desc_stripe_size(__u64 *val)
733 if (*val < LOV_MIN_STRIPE_SIZE) {
735 LCONSOLE_INFO("Increasing default stripe size to "
737 LOV_DESC_STRIPE_SIZE_DEFAULT);
738 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
739 } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
740 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
741 LCONSOLE_WARN("Changing default stripe size to %llu (a "
743 *val, LOV_MIN_STRIPE_SIZE);
747 void lov_fix_desc_stripe_count(__u32 *val)
753 void lov_fix_desc_pattern(__u32 *val)
755 /* from lov_setstripe */
756 if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
757 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
762 void lov_fix_desc_qos_maxage(__u32 *val)
765 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
768 void lov_fix_desc(struct lov_desc *desc)
770 lov_fix_desc_stripe_size(&desc->ld_default_stripe_size);
771 lov_fix_desc_stripe_count(&desc->ld_default_stripe_count);
772 lov_fix_desc_pattern(&desc->ld_pattern);
773 lov_fix_desc_qos_maxage(&desc->ld_qos_maxage);
776 int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
778 struct lov_desc *desc;
779 struct lov_obd *lov = &obd->u.lov;
780 #ifdef CONFIG_PROC_FS
781 struct obd_type *type;
786 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
787 CERROR("LOV setup requires a descriptor\n");
791 desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
793 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
794 CERROR("descriptor size wrong: %d > %d\n",
795 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
799 if (desc->ld_magic != LOV_DESC_MAGIC) {
800 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
801 CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
802 obd->obd_name, desc);
803 lustre_swab_lov_desc(desc);
805 CERROR("%s: Bad lov desc magic: %#x\n",
806 obd->obd_name, desc->ld_magic);
813 desc->ld_active_tgt_count = 0;
815 lov->lov_tgt_size = 0;
817 mutex_init(&lov->lov_lock);
818 atomic_set(&lov->lov_refcount, 0);
819 lov->lov_sp_me = LUSTRE_SP_CLI;
821 init_rwsem(&lov->lov_notify_lock);
823 lov->lov_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
825 HASH_POOLS_BKT_BITS, 0,
828 &pool_hash_operations,
830 INIT_LIST_HEAD(&lov->lov_pool_list);
831 lov->lov_pool_count = 0;
832 rc = lov_ost_pool_init(&lov->lov_packed, 0);
836 #ifdef CONFIG_PROC_FS
837 obd->obd_vars = lprocfs_lov_obd_vars;
838 /* If this is true then both client (lov) and server
839 * (lod) are on the same node. The lod layer if loaded
840 * first will register the lov proc directory. In that
841 * case obd->obd_type->typ_procroot will be not set.
842 * Instead we use type->typ_procsym as the parent. */
843 type = class_search_type(LUSTRE_LOD_NAME);
844 if (type != NULL && type->typ_procsym != NULL) {
845 obd->obd_proc_entry = lprocfs_register(obd->obd_name,
848 if (IS_ERR(obd->obd_proc_entry)) {
849 rc = PTR_ERR(obd->obd_proc_entry);
850 CERROR("error %d setting up lprocfs for %s\n", rc,
852 obd->obd_proc_entry = NULL;
855 rc = lprocfs_obd_setup(obd);
859 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
860 0444, &lov_proc_target_fops, obd);
862 CWARN("Error adding the target_obd file\n");
864 lov->lov_pool_proc_entry = lprocfs_register("pools",
867 if (IS_ERR(lov->lov_pool_proc_entry)) {
868 rc = PTR_ERR(lov->lov_pool_proc_entry);
869 CERROR("error %d setting up lprocfs for pools\n", rc);
870 lov->lov_pool_proc_entry = NULL;
880 static int lov_cleanup(struct obd_device *obd)
882 struct lov_obd *lov = &obd->u.lov;
883 struct list_head *pos, *tmp;
884 struct pool_desc *pool;
887 list_for_each_safe(pos, tmp, &lov->lov_pool_list) {
888 pool = list_entry(pos, struct pool_desc, pool_list);
889 /* free pool structs */
890 CDEBUG(D_INFO, "delete pool %p\n", pool);
891 /* In the function below, .hs_keycmp resolves to
892 * pool_hashkey_keycmp() */
893 /* coverity[overrun-buffer-val] */
894 lov_pool_del(obd, pool->pool_name);
896 cfs_hash_putref(lov->lov_pools_hash_body);
897 lov_ost_pool_free(&lov->lov_packed);
899 lprocfs_obd_cleanup(obd);
903 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
904 if (!lov->lov_tgts[i])
907 /* Inactive targets may never have connected */
908 if (lov->lov_tgts[i]->ltd_active ||
909 atomic_read(&lov->lov_refcount))
910 /* We should never get here - these
911 * should have been removed in the
913 CERROR("%s: lov tgt %d not cleaned! "
914 "deathrow=%d, lovrc=%d\n",
915 obd->obd_name, i, lov->lov_death_row,
916 atomic_read(&lov->lov_refcount));
917 lov_del_target(obd, i, NULL, 0);
920 OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) *
922 lov->lov_tgt_size = 0;
925 if (lov->lov_cache != NULL) {
926 cl_cache_decref(lov->lov_cache);
927 lov->lov_cache = NULL;
933 int lov_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg,
934 __u32 *indexp, int *genp)
936 struct obd_uuid obd_uuid;
941 switch(cmd = lcfg->lcfg_command) {
942 case LCFG_LOV_ADD_OBD:
943 case LCFG_LOV_ADD_INA:
944 case LCFG_LOV_DEL_OBD: {
947 /* lov_modify_tgts add 0:lov_mdsA 1:ost1_UUID 2:0 3:1 */
948 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
949 GOTO(out, rc = -EINVAL);
951 obd_str2uuid(&obd_uuid, lustre_cfg_buf(lcfg, 1));
953 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", indexp) != 1)
954 GOTO(out, rc = -EINVAL);
955 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", genp) != 1)
956 GOTO(out, rc = -EINVAL);
959 if (cmd == LCFG_LOV_ADD_OBD)
960 rc = lov_add_target(obd, &obd_uuid, index, gen, 1);
961 else if (cmd == LCFG_LOV_ADD_INA)
962 rc = lov_add_target(obd, &obd_uuid, index, gen, 0);
964 rc = lov_del_target(obd, index, &obd_uuid, gen);
968 struct lov_desc *desc = &(obd->u.lov.desc);
971 GOTO(out, rc = -EINVAL);
973 rc = class_process_proc_param(PARAM_LOV, obd->obd_vars,
986 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
987 GOTO(out, rc = -EINVAL);
996 lov_statfs_interpret(struct ptlrpc_request_set *rqset, void *data, int rc)
998 struct lov_request_set *lovset = (struct lov_request_set *)data;
1003 atomic_set(&lovset->set_completes, 0);
1005 err = lov_fini_statfs_set(lovset);
1006 RETURN(rc ? rc : err);
1009 static int lov_statfs_async(struct obd_export *exp, struct obd_info *oinfo,
1010 __u64 max_age, struct ptlrpc_request_set *rqset)
1012 struct obd_device *obd = class_exp2obd(exp);
1013 struct lov_request_set *set;
1014 struct lov_request *req;
1015 struct list_head *pos;
1016 struct lov_obd *lov;
1020 LASSERT(oinfo != NULL);
1021 LASSERT(oinfo->oi_osfs != NULL);
1024 rc = lov_prep_statfs_set(obd, oinfo, &set);
1028 list_for_each(pos, &set->set_list) {
1029 req = list_entry(pos, struct lov_request, rq_link);
1030 rc = obd_statfs_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1031 &req->rq_oi, max_age, rqset);
1036 if (rc || list_empty(&rqset->set_requests)) {
1039 atomic_set(&set->set_completes, 0);
1040 err = lov_fini_statfs_set(set);
1041 RETURN(rc ? rc : err);
1044 LASSERT(rqset->set_interpret == NULL);
1045 rqset->set_interpret = lov_statfs_interpret;
1046 rqset->set_arg = (void *)set;
1050 static int lov_statfs(const struct lu_env *env, struct obd_export *exp,
1051 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1053 struct ptlrpc_request_set *set = NULL;
1054 struct obd_info oinfo = {
1062 /* for obdclass we forbid using obd_statfs_rqset, but prefer using async
1063 * statfs requests */
1064 set = ptlrpc_prep_set();
1068 rc = lov_statfs_async(exp, &oinfo, max_age, set);
1070 rc = ptlrpc_set_wait(set);
1072 ptlrpc_set_destroy(set);
1077 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1078 void *karg, void __user *uarg)
1080 struct obd_device *obddev = class_exp2obd(exp);
1081 struct lov_obd *lov = &obddev->u.lov;
1082 int i = 0, rc = 0, count = lov->desc.ld_tgt_count;
1083 struct obd_uuid *uuidp;
1087 case IOC_OBD_STATFS: {
1088 struct obd_ioctl_data *data = karg;
1089 struct obd_device *osc_obd;
1090 struct obd_statfs stat_buf = {0};
1094 memcpy(&index, data->ioc_inlbuf2, sizeof(index));
1095 if ((index >= count))
1098 if (!lov->lov_tgts[index])
1099 /* Try again with the next index */
1101 if (!lov->lov_tgts[index]->ltd_active)
1104 osc_obd = class_exp2obd(lov->lov_tgts[index]->ltd_exp);
1109 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(osc_obd),
1110 min_t(unsigned long, data->ioc_plen2,
1111 sizeof(struct obd_uuid))))
1114 memcpy(&flags, data->ioc_inlbuf1, sizeof(flags));
1115 flags = flags & LL_STATFS_NODELAY ? OBD_STATFS_NODELAY : 0;
1117 /* got statfs data */
1118 rc = obd_statfs(NULL, lov->lov_tgts[index]->ltd_exp, &stat_buf,
1119 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1123 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1124 min_t(unsigned long, data->ioc_plen1,
1125 sizeof(struct obd_statfs))))
1129 case OBD_IOC_LOV_GET_CONFIG: {
1130 struct obd_ioctl_data *data;
1131 struct lov_desc *desc;
1136 if (obd_ioctl_getdata(&buf, &len, uarg))
1139 data = (struct obd_ioctl_data *)buf;
1141 if (sizeof(*desc) > data->ioc_inllen1) {
1142 obd_ioctl_freedata(buf, len);
1146 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
1147 obd_ioctl_freedata(buf, len);
1151 if (sizeof(__u32) * count > data->ioc_inllen3) {
1152 obd_ioctl_freedata(buf, len);
1156 desc = (struct lov_desc *)data->ioc_inlbuf1;
1157 memcpy(desc, &(lov->desc), sizeof(*desc));
1159 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
1160 genp = (__u32 *)data->ioc_inlbuf3;
1161 /* the uuid will be empty for deleted OSTs */
1162 for (i = 0; i < count; i++, uuidp++, genp++) {
1163 if (!lov->lov_tgts[i])
1165 *uuidp = lov->lov_tgts[i]->ltd_uuid;
1166 *genp = lov->lov_tgts[i]->ltd_gen;
1169 if (copy_to_user(uarg, buf, len))
1171 obd_ioctl_freedata(buf, len);
1174 case OBD_IOC_QUOTACTL: {
1175 struct if_quotactl *qctl = karg;
1176 struct lov_tgt_desc *tgt = NULL;
1177 struct obd_quotactl *oqctl;
1179 if (qctl->qc_valid == QC_OSTIDX) {
1180 if (count <= qctl->qc_idx)
1183 tgt = lov->lov_tgts[qctl->qc_idx];
1184 if (!tgt || !tgt->ltd_exp)
1186 } else if (qctl->qc_valid == QC_UUID) {
1187 for (i = 0; i < count; i++) {
1188 tgt = lov->lov_tgts[i];
1190 !obd_uuid_equals(&tgt->ltd_uuid,
1194 if (tgt->ltd_exp == NULL)
1206 LASSERT(tgt && tgt->ltd_exp);
1207 OBD_ALLOC_PTR(oqctl);
1211 QCTL_COPY(oqctl, qctl);
1212 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1214 QCTL_COPY(qctl, oqctl);
1215 qctl->qc_valid = QC_OSTIDX;
1216 qctl->obd_uuid = tgt->ltd_uuid;
1218 OBD_FREE_PTR(oqctl);
1227 for (i = 0; i < count; i++) {
1229 struct obd_device *osc_obd;
1231 /* OST was disconnected */
1232 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
1235 /* ll_umount_begin() sets force flag but for lov, not
1236 * osc. Let's pass it through */
1237 osc_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
1238 osc_obd->obd_force = obddev->obd_force;
1239 err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp,
1242 if (lov->lov_tgts[i]->ltd_active) {
1243 CDEBUG(err == -ENOTTY ?
1244 D_IOCTL : D_WARNING,
1245 "iocontrol OSC %s on OST "
1246 "idx %d cmd %x: err = %d\n",
1247 lov_uuid2str(lov, i),
1264 static int lov_get_info(const struct lu_env *env, struct obd_export *exp,
1265 __u32 keylen, void *key, __u32 *vallen, void *val)
1267 struct obd_device *obddev = class_exp2obd(exp);
1268 struct lov_obd *lov = &obddev->u.lov;
1269 struct lov_desc *ld = &lov->desc;
1273 if (vallen == NULL || val == NULL)
1278 if (KEY_IS(KEY_MAX_EASIZE)) {
1279 u32 max_stripe_count = min_t(u32, ld->ld_active_tgt_count,
1280 LOV_MAX_STRIPE_COUNT);
1282 *((u32 *)val) = lov_mds_md_size(max_stripe_count, LOV_MAGIC_V3);
1283 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
1284 u32 def_stripe_count = min_t(u32, ld->ld_default_stripe_count,
1285 LOV_MAX_STRIPE_COUNT);
1287 *((u32 *)val) = lov_mds_md_size(def_stripe_count, LOV_MAGIC_V3);
1288 } else if (KEY_IS(KEY_TGT_COUNT)) {
1289 *((int *)val) = lov->desc.ld_tgt_count;
1299 static int lov_set_info_async(const struct lu_env *env, struct obd_export *exp,
1300 __u32 keylen, void *key,
1301 __u32 vallen, void *val,
1302 struct ptlrpc_request_set *set)
1304 struct obd_device *obddev = class_exp2obd(exp);
1305 struct lov_obd *lov = &obddev->u.lov;
1306 struct lov_tgt_desc *tgt;
1307 int do_inactive = 0;
1317 set = ptlrpc_prep_set();
1323 count = lov->desc.ld_tgt_count;
1325 if (KEY_IS(KEY_CHECKSUM)) {
1327 } else if (KEY_IS(KEY_CACHE_SET)) {
1328 LASSERT(lov->lov_cache == NULL);
1329 lov->lov_cache = val;
1331 cl_cache_incref(lov->lov_cache);
1334 for (i = 0; i < count; i++) {
1335 tgt = lov->lov_tgts[i];
1337 /* OST was disconnected */
1338 if (!tgt || !tgt->ltd_exp)
1341 /* OST is inactive and we don't want inactive OSCs */
1342 if (!tgt->ltd_active && !do_inactive)
1345 err = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1353 err = ptlrpc_set_wait(set);
1356 ptlrpc_set_destroy(set);
1361 void lov_stripe_lock(struct lov_stripe_md *md)
1362 __acquires(&md->lsm_lock)
1364 LASSERT(md->lsm_lock_owner != current_pid());
1365 spin_lock(&md->lsm_lock);
1366 LASSERT(md->lsm_lock_owner == 0);
1367 md->lsm_lock_owner = current_pid();
1370 void lov_stripe_unlock(struct lov_stripe_md *md)
1371 __releases(&md->lsm_lock)
1373 LASSERT(md->lsm_lock_owner == current_pid());
1374 md->lsm_lock_owner = 0;
1375 spin_unlock(&md->lsm_lock);
1378 static int lov_quotactl(struct obd_device *obd, struct obd_export *exp,
1379 struct obd_quotactl *oqctl)
1381 struct lov_obd *lov = &obd->u.lov;
1382 struct lov_tgt_desc *tgt;
1384 __u64 bhardlimit = 0;
1388 if (oqctl->qc_cmd != Q_GETOQUOTA &&
1389 oqctl->qc_cmd != LUSTRE_Q_SETQUOTA) {
1390 CERROR("%s: bad quota opc %x for lov obd\n",
1391 obd->obd_name, oqctl->qc_cmd);
1397 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1400 tgt = lov->lov_tgts[i];
1405 if (!tgt->ltd_active || tgt->ltd_reap) {
1406 if (oqctl->qc_cmd == Q_GETOQUOTA &&
1407 lov->lov_tgts[i]->ltd_activate) {
1409 CERROR("ost %d is inactive\n", i);
1411 CDEBUG(D_HA, "ost %d is inactive\n", i);
1416 err = obd_quotactl(tgt->ltd_exp, oqctl);
1418 if (tgt->ltd_active && !rc)
1423 if (oqctl->qc_cmd == Q_GETOQUOTA) {
1424 curspace += oqctl->qc_dqblk.dqb_curspace;
1425 bhardlimit += oqctl->qc_dqblk.dqb_bhardlimit;
1430 if (oqctl->qc_cmd == Q_GETOQUOTA) {
1431 oqctl->qc_dqblk.dqb_curspace = curspace;
1432 oqctl->qc_dqblk.dqb_bhardlimit = bhardlimit;
1437 static struct obd_ops lov_obd_ops = {
1438 .o_owner = THIS_MODULE,
1439 .o_setup = lov_setup,
1440 .o_cleanup = lov_cleanup,
1441 .o_connect = lov_connect,
1442 .o_disconnect = lov_disconnect,
1443 .o_statfs = lov_statfs,
1444 .o_statfs_async = lov_statfs_async,
1445 .o_iocontrol = lov_iocontrol,
1446 .o_get_info = lov_get_info,
1447 .o_set_info_async = lov_set_info_async,
1448 .o_notify = lov_notify,
1449 .o_pool_new = lov_pool_new,
1450 .o_pool_rem = lov_pool_remove,
1451 .o_pool_add = lov_pool_add,
1452 .o_pool_del = lov_pool_del,
1453 .o_getref = lov_getref,
1454 .o_putref = lov_putref,
1455 .o_quotactl = lov_quotactl,
1458 struct kmem_cache *lov_oinfo_slab;
1460 static int __init lov_init(void)
1462 bool enable_proc = true;
1463 struct obd_type *type;
1467 /* print an address of _any_ initialized kernel symbol from this
1468 * module, to allow debugging with gdb that doesn't support data
1469 * symbols from modules.*/
1470 CDEBUG(D_INFO, "Lustre LOV module (%p).\n", &lov_caches);
1472 rc = lu_kmem_init(lov_caches);
1476 lov_oinfo_slab = kmem_cache_create("lov_oinfo",
1477 sizeof(struct lov_oinfo), 0,
1478 SLAB_HWCACHE_ALIGN, NULL);
1479 if (lov_oinfo_slab == NULL) {
1480 lu_kmem_fini(lov_caches);
1484 type = class_search_type(LUSTRE_LOD_NAME);
1485 if (type != NULL && type->typ_procsym != NULL)
1486 enable_proc = false;
1488 rc = class_register_type(&lov_obd_ops, NULL, enable_proc, NULL,
1489 LUSTRE_LOV_NAME, &lov_device_type);
1492 kmem_cache_destroy(lov_oinfo_slab);
1493 lu_kmem_fini(lov_caches);
1499 static void __exit lov_exit(void)
1501 class_unregister_type(LUSTRE_LOV_NAME);
1502 kmem_cache_destroy(lov_oinfo_slab);
1503 lu_kmem_fini(lov_caches);
1506 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1507 MODULE_DESCRIPTION("Lustre Logical Object Volume");
1508 MODULE_VERSION(LUSTRE_VERSION_STRING);
1509 MODULE_LICENSE("GPL");
1511 module_init(lov_init);
1512 module_exit(lov_exit);