1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Phil Schwan <phil@clusterfs.com>
6 * Peter Braam <braam@clusterfs.com>
7 * Mike Shaver <shaver@clusterfs.com>
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
26 # define EXPORT_SYMTAB
28 #define DEBUG_SUBSYSTEM S_LOV
30 #include <linux/slab.h>
31 #include <linux/module.h>
32 #include <linux/init.h>
33 #include <linux/slab.h>
34 #include <linux/pagemap.h>
35 #include <linux/seq_file.h>
36 #include <asm/div64.h>
38 #include <liblustre.h>
41 #include <linux/obd_support.h>
42 #include <linux/lustre_lib.h>
43 #include <linux/lustre_net.h>
44 #include <linux/lustre_idl.h>
45 #include <linux/lustre_dlm.h>
46 #include <linux/lustre_mds.h>
47 #include <linux/obd_class.h>
48 #include <linux/obd_lov.h>
49 #include <linux/obd_ost.h>
50 #include <linux/lprocfs_status.h>
52 #include "lov_internal.h"
55 #define MAX_STRING_SIZE 128
56 static int lov_connect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt,
57 int activate, struct obd_connect_data *conn_data,
58 unsigned long connect_flags)
60 struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
61 struct obd_uuid *tgt_uuid = &tgt->uuid;
64 struct proc_dir_entry *lov_proc_dir;
66 struct lov_obd *lov = &obd->u.lov;
67 struct lustre_handle conn = {0, };
68 struct obd_device *tgt_obd;
72 tgt_obd = class_find_client_obd(tgt_uuid, OBD_OSC_DEVICENAME,
76 CERROR("Target %s not attached\n", tgt_uuid->uuid);
80 if (!tgt_obd->obd_set_up) {
81 CERROR("Target %s not set up\n", tgt_uuid->uuid);
86 tgt_obd->obd_no_recov = 0;
87 ptlrpc_activate_import(tgt_obd->u.cli.cl_import);
90 if (tgt_obd->u.cli.cl_import->imp_invalid) {
91 CERROR("not connecting OSC %s; administratively "
92 "disabled\n", tgt_uuid->uuid);
93 rc = obd_register_observer(tgt_obd, obd);
95 CERROR("Target %s register_observer error %d; "
96 "will not be able to reactivate\n",
102 rc = obd_connect(&conn, tgt_obd, &lov_osc_uuid, conn_data,
105 CERROR("Target %s connect error %d\n", tgt_uuid->uuid, rc);
108 tgt->ltd_exp = class_conn2export(&conn);
110 rc = obd_register_observer(tgt_obd, obd);
112 CERROR("Target %s register_observer error %d\n",
114 obd_disconnect(tgt->ltd_exp, 0);
120 lov->desc.ld_active_tgt_count++;
123 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
125 struct obd_device *osc_obd = class_conn2obd(&conn);
126 struct proc_dir_entry *osc_symlink;
127 char name[MAX_STRING_SIZE + 1];
129 LASSERT(osc_obd != NULL);
130 LASSERT(osc_obd->obd_type != NULL);
131 LASSERT(osc_obd->obd_type->typ_name != NULL);
132 name[MAX_STRING_SIZE] = '\0';
133 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
134 osc_obd->obd_type->typ_name,
136 osc_symlink = proc_symlink(osc_obd->obd_name, lov_proc_dir,
138 if (osc_symlink == NULL) {
139 CERROR("could not register LOV target "
140 "/proc/fs/lustre/%s/%s/target_obds/%s\n",
141 obd->obd_type->typ_name, obd->obd_name,
143 lprocfs_remove(lov_proc_dir);
152 static int lov_connect(struct lustre_handle *conn, struct obd_device *obd,
153 struct obd_uuid *cluuid, struct obd_connect_data *data,
157 struct proc_dir_entry *lov_proc_dir;
159 struct lov_obd *lov = &obd->u.lov;
160 struct lov_tgt_desc *tgt;
161 struct obd_export *exp;
165 rc = class_connect(conn, obd, cluuid);
169 exp = class_conn2export(conn);
171 /* We don't want to actually do the underlying connections more than
172 * once, so keep track. */
174 if (lov->refcount > 1) {
175 class_export_put(exp);
180 lov_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
182 if (IS_ERR(lov_proc_dir)) {
183 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
184 obd->obd_type->typ_name, obd->obd_name);
189 /* connect_flags is the MDS number, save for use in lov_add_obd */
190 lov->lov_connect_flags = flags;
191 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
192 if (obd_uuid_empty(&tgt->uuid))
194 rc = lov_connect_obd(obd, tgt, 0, data, flags);
199 class_export_put(exp);
205 lprocfs_remove(lov_proc_dir);
209 struct obd_uuid uuid;
211 --lov->desc.ld_active_tgt_count;
213 /* save for CERROR below; (we know it's terminated) */
215 rc2 = obd_disconnect(tgt->ltd_exp, 0);
217 CERROR("error: LOV target %s disconnect on OST idx %d: "
218 "rc = %d\n", uuid.uuid, i, rc2);
220 class_disconnect(exp, 0);
224 static int lov_disconnect_obd(struct obd_device *obd,
225 struct lov_tgt_desc *tgt,
229 struct proc_dir_entry *lov_proc_dir;
231 struct obd_device *osc_obd = class_exp2obd(tgt->ltd_exp);
232 struct lov_obd *lov = &obd->u.lov;
237 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
239 struct proc_dir_entry *osc_symlink;
241 osc_symlink = lprocfs_srch(lov_proc_dir, osc_obd->obd_name);
243 lprocfs_remove(osc_symlink);
245 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
246 obd->obd_type->typ_name, obd->obd_name,
251 if (obd->obd_no_recov) {
252 /* Pass it on to our clients.
253 * XXX This should be an argument to disconnect,
254 * XXX not a back-door flag on the OBD. Ah well.
257 osc_obd->obd_no_recov = 1;
260 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
261 rc = obd_disconnect(tgt->ltd_exp, flags);
264 CERROR("Target %s disconnect error %d\n",
272 lov->desc.ld_active_tgt_count--;
278 static int lov_disconnect(struct obd_export *exp, unsigned long flags)
280 struct obd_device *obd = class_exp2obd(exp);
282 struct proc_dir_entry *lov_proc_dir;
284 struct lov_obd *lov = &obd->u.lov;
285 struct lov_tgt_desc *tgt;
292 /* Only disconnect the underlying layers on the final disconnect. */
294 if (lov->refcount != 0)
297 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
299 lov_disconnect_obd(obd, tgt, flags);
303 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
305 lprocfs_remove(lov_proc_dir);
307 CERROR("/proc/fs/lustre/%s/%s/target_obds missing.",
308 obd->obd_type->typ_name, obd->obd_name);
313 rc = class_disconnect(exp, 0);
319 * -EINVAL : UUID can't be found in the LOV's target list
320 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
321 * -EBADF : The UUID is found, but the OBD is the wrong type (!)
323 static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid,
326 struct lov_tgt_desc *tgt;
330 CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
331 lov, uuid->uuid, activate);
333 spin_lock(&lov->lov_lock);
334 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
335 if (tgt->ltd_exp == NULL)
338 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
339 i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
341 if (obd_uuid_equals(uuid, &tgt->uuid))
345 if (i == lov->desc.ld_tgt_count)
346 GOTO(out, rc = -EINVAL);
349 if (tgt->active == activate) {
350 CDEBUG(D_INFO, "OSC %s already %sactive!\n", uuid->uuid,
351 activate ? "" : "in");
355 CDEBUG(D_INFO, "Marking OSC %s %sactive\n", uuid->uuid,
356 activate ? "" : "in");
357 CDEBUG(D_ERROR, "Marking OSC %s %sactive\n", uuid->uuid,
358 activate ? "" : "in");
360 tgt->active = activate;
362 lov->desc.ld_active_tgt_count++;
364 lov->desc.ld_active_tgt_count--;
368 spin_unlock(&lov->lov_lock);
372 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
373 int active, void *data)
375 struct obd_uuid *uuid;
379 if (strcmp(watched->obd_type->typ_name, OBD_OSC_DEVICENAME)) {
380 CERROR("unexpected notification of %s %s!\n",
381 watched->obd_type->typ_name,
385 uuid = &watched->u.cli.cl_import->imp_target_uuid;
387 /* Set OSC as active before notifying the observer, so the
388 * observer can use the OSC normally.
390 rc = lov_set_osc_active(&obd->u.lov, uuid, active);
392 CERROR("%sactivation of %s failed: %d\n",
393 active ? "" : "de", uuid->uuid, rc);
397 if (obd->obd_observer)
398 /* Pass the notification up the chain. */
399 rc = obd_notify(obd->obd_observer, watched, active, data);
404 int lov_attach(struct obd_device *dev, obd_count len, void *data)
406 struct lprocfs_static_vars lvars;
409 lprocfs_init_vars(lov, &lvars);
410 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
413 struct proc_dir_entry *entry;
415 entry = create_proc_entry("target_obd_status", 0444,
416 dev->obd_proc_entry);
420 entry->proc_fops = &lov_proc_target_fops;
428 int lov_detach(struct obd_device *dev)
430 return lprocfs_obd_detach(dev);
433 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
435 struct lov_obd *lov = &obd->u.lov;
436 struct lustre_cfg *lcfg = buf;
437 struct lov_desc *desc;
441 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
442 CERROR("LOV setup requires a descriptor\n");
446 desc = (struct lov_desc *)lustre_cfg_string(lcfg, 1);
447 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
448 CERROR("descriptor size wrong: %d > %d\n",
449 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
453 /* Because of 64-bit divide/mod operations only work with a 32-bit
454 * divisor in a 32-bit kernel, we cannot support a stripe width
455 * of 4GB or larger on 32-bit CPUs.
458 count = desc->ld_default_stripe_count;
459 if (count && (count * desc->ld_default_stripe_size) > ~0UL) {
460 CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n",
461 desc->ld_default_stripe_size, count, ~0UL);
464 if (desc->ld_tgt_count > 0) {
465 lov->bufsize= sizeof(struct lov_tgt_desc) * desc->ld_tgt_count;
467 lov->bufsize = sizeof(struct lov_tgt_desc) * LOV_MAX_TGT_COUNT;
469 OBD_ALLOC(lov->tgts, lov->bufsize);
470 if (lov->tgts == NULL) {
472 CERROR("couldn't allocate %d bytes for target table.\n",
477 desc->ld_tgt_count = 0;
478 desc->ld_active_tgt_count = 0;
480 spin_lock_init(&lov->lov_lock);
481 sema_init(&lov->lov_llog_sem, 1);
486 static int lov_cleanup(struct obd_device *obd, int flags)
488 struct lov_obd *lov = &obd->u.lov;
490 OBD_FREE(lov->tgts, lov->bufsize);
495 lov_add_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
497 struct lov_obd *lov = &obd->u.lov;
498 struct lov_tgt_desc *tgt;
502 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
503 uuidp->uuid, index, gen);
505 if ((index < 0) || (index >= LOV_MAX_TGT_COUNT)) {
506 CERROR("request to add OBD %s at invalid index: %d\n",
512 CERROR("request to add OBD %s with invalid generation: %d\n",
517 tgt = lov->tgts + index;
518 if (!obd_uuid_empty(&tgt->uuid)) {
519 CERROR("OBD already assigned at LOV target index %d\n",
525 /* XXX - add a sanity check on the generation number. */
528 if (index >= lov->desc.ld_tgt_count)
529 lov->desc.ld_tgt_count = index + 1;
531 CDEBUG(D_CONFIG, "idx: %d ltd_gen: %d ld_tgt_count: %d\n",
532 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
534 if (lov->refcount == 0)
538 struct obd_device *osc_obd;
540 osc_obd = class_exp2obd(tgt->ltd_exp);
542 osc_obd->obd_no_recov = 0;
545 rc = lov_connect_obd(obd, tgt, 1, NULL, lov->lov_connect_flags);
549 if (obd->obd_observer) {
550 /* tell the mds_lov about the new target */
551 rc = obd_notify(obd->obd_observer, tgt->ltd_exp->exp_obd, 1,
557 if (rc && tgt->ltd_exp != NULL)
558 lov_disconnect_obd(obd, tgt, 0);
563 lov_del_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
565 struct lov_obd *lov = &obd->u.lov;
566 struct lov_tgt_desc *tgt;
567 int count = lov->desc.ld_tgt_count;
571 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
572 uuidp->uuid, index, gen);
574 if (index >= count) {
575 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
580 tgt = lov->tgts + index;
582 if (obd_uuid_empty(&tgt->uuid)) {
583 CERROR("LOV target at index %d is not setup.\n", index);
587 if (!obd_uuid_equals(uuidp, &tgt->uuid)) {
588 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
589 tgt->uuid.uuid, index, uuidp->uuid);
594 struct obd_device *osc_obd;
596 osc_obd = class_exp2obd(tgt->ltd_exp);
598 osc_obd->obd_no_recov = 1;
599 rc = obd_llog_finish(osc_obd, &osc_obd->obd_llogs, 1);
601 CERROR("osc_llog_finish error: %d\n", rc);
603 lov_disconnect_obd(obd, tgt, 0);
606 /* XXX - right now there is a dependency on ld_tgt_count being the
607 * maximum tgt index for computing the mds_max_easize. So we can't
610 /* lt_gen = 0 will mean it will not match the gen of any valid loi */
611 memset(tgt, 0, sizeof(*tgt));
613 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
614 tgt->uuid.uuid, index, tgt->ltd_gen, tgt->ltd_exp, tgt->active);
619 static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
621 struct lustre_cfg *lcfg = buf;
622 struct obd_uuid obd_uuid;
629 switch(cmd = lcfg->lcfg_command) {
630 case LCFG_LOV_ADD_OBD:
631 case LCFG_LOV_DEL_OBD: {
632 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
633 GOTO(out, rc = -EINVAL);
635 obd_str2uuid(&obd_uuid, lustre_cfg_string(lcfg, 1));
637 if (sscanf(lustre_cfg_buf(lcfg, 2), "%d", &index) != 1)
638 GOTO(out, rc = -EINVAL);
639 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
640 GOTO(out, rc = -EINVAL);
641 if (cmd == LCFG_LOV_ADD_OBD)
642 rc = lov_add_obd(obd, &obd_uuid, index, gen);
644 rc = lov_del_obd(obd, &obd_uuid, index, gen);
648 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
649 GOTO(out, rc = -EINVAL);
658 #define log2(n) ffz(~(n))
661 static int lov_clear_orphans(struct obd_export *export,
663 struct lov_stripe_md **ea,
664 struct obd_trans_info *oti)
668 struct obd_uuid *ost_uuid = NULL;
672 LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
673 src_oa->o_flags == OBD_FL_DELORPHAN);
675 lov = &export->exp_obd->u.lov;
677 tmp_oa = obdo_alloc();
681 if (src_oa->o_valid & OBD_MD_FLINLINE) {
682 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
683 CDEBUG(D_HA, "clearing orphans only for %s\n",
687 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
689 struct lov_stripe_md obj_md;
690 struct lov_stripe_md *obj_mdp = &obj_md;
693 * if called for a specific target, we don't care if it is not
696 if (lov->tgts[i].active == 0 && ost_uuid == NULL) {
697 CDEBUG(D_HA, "lov idx %d inactive\n", i);
701 if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid))
705 * setting up objid OSS objects should be destroyed starting
708 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
709 tmp_oa->o_valid |= OBD_MD_FLID;
710 tmp_oa->o_id = oti->oti_objid[i];
712 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
713 err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, NULL, 0,
717 * this export will be disabled until it is recovered,
718 * and then orphan recovery will be completed.
720 CERROR("error in orphan recovery on OST idx %d/%d: "
721 "rc = %d\n", i, lov->desc.ld_tgt_count, err);
731 /* the LOV expects oa->o_id to be set to the LOV object id */
733 lov_create(struct obd_export *exp, struct obdo *src_oa,
734 void *acl, int acl_size, struct lov_stripe_md **ea,
735 struct obd_trans_info *oti)
737 struct lov_request_set *set = NULL;
738 struct list_head *pos;
747 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
748 src_oa->o_flags == OBD_FL_DELORPHAN) {
749 rc = lov_clear_orphans(exp, src_oa, ea, oti);
753 lov = &exp->exp_obd->u.lov;
754 if (!lov->desc.ld_active_tgt_count)
757 LASSERT(oti->oti_flags & OBD_MODE_CROW);
759 /* main creation loop */
760 rc = lov_prep_create_set(exp, ea, src_oa, oti, &set);
764 list_for_each (pos, &set->set_list) {
765 struct lov_request *req =
766 list_entry(pos, struct lov_request, rq_link);
768 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
769 rc = obd_create(lov->tgts[req->rq_idx].ltd_exp,
770 req->rq_oa, NULL, 0, &req->rq_md, oti);
771 lov_update_create_set(set, req, rc);
773 rc = lov_fini_create_set(set, ea);
777 #define lsm_bad_magic(LSMP) \
779 struct lov_stripe_md *_lsm__ = (LSMP); \
782 CERROR("LOV requires striping ea\n"); \
784 } else if (_lsm__->lsm_magic != LOV_MAGIC) { \
785 CERROR("LOV striping magic bad %#x != %#x\n", \
786 _lsm__->lsm_magic, LOV_MAGIC); \
792 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
793 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
795 struct lov_request_set *set;
796 struct lov_request *req;
797 struct list_head *pos;
802 if (lsm_bad_magic(lsm))
805 if (!exp || !exp->exp_obd)
808 lov = &exp->exp_obd->u.lov;
809 rc = lov_prep_destroy_set(exp, oa, lsm, oti, &set);
813 list_for_each (pos, &set->set_list) {
815 req = list_entry(pos, struct lov_request, rq_link);
817 /* XXX update the cookie position */
818 oti->oti_logcookies = set->set_cookies + req->rq_stripe;
819 rc = obd_destroy(lov->tgts[req->rq_idx].ltd_exp, req->rq_oa,
821 err = lov_update_common_set(set, req, rc);
823 CERROR("error: destroying objid "LPX64" subobj "
824 LPX64" on OST idx %d: rc = %d\n",
825 set->set_oa->o_id, req->rq_oa->o_id,
831 lov_fini_destroy_set(set);
835 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
836 struct lov_stripe_md *lsm)
838 struct lov_request_set *set;
839 struct lov_request *req;
840 struct list_head *pos;
845 if (lsm_bad_magic(lsm))
848 if (!exp || !exp->exp_obd)
851 lov = &exp->exp_obd->u.lov;
853 rc = lov_prep_getattr_set(exp, oa, lsm, &set);
857 list_for_each (pos, &set->set_list) {
858 req = list_entry(pos, struct lov_request, rq_link);
860 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
861 "%u\n", oa->o_id, req->rq_stripe, req->rq_oa->o_id,
864 rc = obd_getattr(lov->tgts[req->rq_idx].ltd_exp,
866 err = lov_update_common_set(set, req, rc);
868 CERROR("error: getattr objid "LPX64" subobj "
869 LPX64" on OST idx %d: rc = %d\n",
870 set->set_oa->o_id, req->rq_oa->o_id,
876 rc = lov_fini_getattr_set(set);
882 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
885 struct lov_request_set *lovset = (struct lov_request_set *)data;
888 /* don't do attribute merge if this aysnc op failed */
890 lovset->set_completes = 0;
891 lov_fini_getattr_set(lovset);
893 rc = lov_fini_getattr_set(lovset);
898 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
899 struct lov_stripe_md *lsm,
900 struct ptlrpc_request_set *rqset)
902 struct lov_request_set *lovset;
904 struct list_head *pos;
905 struct lov_request *req;
909 if (lsm_bad_magic(lsm))
912 if (!exp || !exp->exp_obd)
915 lov = &exp->exp_obd->u.lov;
917 rc = lov_prep_getattr_set(exp, oa, lsm, &lovset);
921 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
922 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
924 list_for_each (pos, &lovset->set_list) {
925 req = list_entry(pos, struct lov_request, rq_link);
927 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
928 "%u\n", oa->o_id, req->rq_stripe, req->rq_oa->o_id,
930 rc = obd_getattr_async(lov->tgts[req->rq_idx].ltd_exp,
931 req->rq_oa, NULL, rqset);
933 CERROR("error: getattr objid "LPX64" subobj "
934 LPX64" on OST idx %d: rc = %d\n",
935 lovset->set_oa->o_id, req->rq_oa->o_id,
939 lov_update_common_set(lovset, req, rc);
943 LASSERT (rqset->set_interpret == NULL);
944 rqset->set_interpret = lov_getattr_interpret;
945 rqset->set_arg = (void *)lovset;
949 lov_fini_getattr_set(lovset);
953 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
954 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
956 struct lov_request_set *set;
958 struct list_head *pos;
959 struct lov_request *req;
963 if (lsm_bad_magic(lsm))
966 if (!exp || !exp->exp_obd)
969 LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
970 OBD_MD_FLATIME | OBD_MD_FLMTIME |
971 OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
972 OBD_MD_FLSIZE | OBD_MD_FLGROUP |
973 OBD_MD_FLUID | OBD_MD_FLGID |
974 OBD_MD_FLINLINE | OBD_MD_FLIFID)));
976 LASSERT(!(src_oa->o_valid & OBD_MD_FLGROUP) || src_oa->o_gr > 0);
978 lov = &exp->exp_obd->u.lov;
979 rc = lov_prep_setattr_set(exp, src_oa, lsm, NULL, &set);
983 list_for_each (pos, &set->set_list) {
984 req = list_entry(pos, struct lov_request, rq_link);
986 rc = obd_setattr(lov->tgts[req->rq_idx].ltd_exp, req->rq_oa,
988 err = lov_update_common_set(set, req, rc);
990 CERROR("error: setattr objid "LPX64" subobj "
991 LPX64" on OST idx %d: rc = %d\n",
992 set->set_oa->o_id, req->rq_oa->o_id,
998 err = lov_fini_setattr_set(set);
1004 static int lov_revalidate_policy(struct lov_obd *lov, struct lov_stripe_md *lsm)
1006 static int next_idx = 0;
1007 struct lov_tgt_desc *tgt;
1010 /* XXX - we should do something clever and take lsm
1011 * into account but just do round robin for now. */
1013 /* last_idx must always be less that count because
1014 * ld_tgt_count currently cannot shrink. */
1015 count = lov->desc.ld_tgt_count;
1017 for (i = next_idx, tgt = lov->tgts + i; i < count; i++, tgt++) {
1019 next_idx = (i + 1) % count;
1024 for (i = 0, tgt = lov->tgts; i < next_idx; i++, tgt++) {
1026 next_idx = (i + 1) % count;
1034 static int lov_revalidate_md(struct obd_export *exp, struct obdo *src_oa,
1035 struct lov_stripe_md *ea,
1036 struct obd_trans_info *oti)
1038 struct obd_export *osc_exp;
1039 struct lov_obd *lov = &exp->exp_obd->u.lov;
1040 struct lov_stripe_md *lsm = ea;
1041 struct lov_stripe_md obj_md;
1042 struct lov_stripe_md *obj_mdp = &obj_md;
1043 struct lov_oinfo *loi;
1044 struct obdo *tmp_oa;
1045 int ost_idx, updates = 0, i;
1048 tmp_oa = obdo_alloc();
1052 loi = lsm->lsm_oinfo;
1053 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1055 if (!obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
1058 ost_idx = lov_revalidate_policy(lov, lsm);
1060 /* FIXME: punt for now. */
1061 CERROR("lov_revalidate_policy failed; no active "
1066 /* create a new object */
1067 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1068 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
1069 osc_exp = lov->tgts[ost_idx].ltd_exp;
1070 rc = obd_create(osc_exp, tmp_oa, NULL, 0, &obj_mdp, oti);
1072 CERROR("error creating new subobj at idx %d; "
1073 "rc = %d\n", ost_idx, rc);
1077 oti->oti_objid[ost_idx] = tmp_oa->o_id;
1078 loi->loi_id = tmp_oa->o_id;
1079 loi->loi_gr = tmp_oa->o_gr;
1080 loi->loi_ost_idx = ost_idx;
1081 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
1082 CDEBUG(D_INODE, "replacing objid "LPX64" subobj "LPX64
1083 " with idx %d gen %d.\n", lsm->lsm_object_id,
1084 loi->loi_id, ost_idx, loi->loi_ost_gen);
1088 /* If we got an error revalidating an entry there's no need to
1089 * cleanup up objects we allocated here because the bad entry
1090 * still points to a deleted OST. */
1096 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1097 * we can send this 'punch' to just the authoritative node and the nodes
1098 * that the punch will affect. */
1099 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1100 struct lov_stripe_md *lsm,
1101 obd_off start, obd_off end, struct obd_trans_info *oti)
1103 struct lov_request_set *set;
1104 struct lov_obd *lov;
1105 struct list_head *pos;
1106 struct lov_request *req;
1107 int err = 0, rc = 0;
1110 if (lsm_bad_magic(lsm))
1113 if (!exp || !exp->exp_obd)
1116 lov = &exp->exp_obd->u.lov;
1117 rc = lov_prep_punch_set(exp, oa, lsm, start, end, oti, &set);
1121 list_for_each (pos, &set->set_list) {
1122 req = list_entry(pos, struct lov_request, rq_link);
1124 rc = obd_punch(lov->tgts[req->rq_idx].ltd_exp, req->rq_oa,
1125 NULL, req->rq_extent.start,
1126 req->rq_extent.end, NULL);
1127 err = lov_update_punch_set(set, req, rc);
1129 CERROR("error: punch objid "LPX64" subobj "LPX64
1130 " on OST idx %d: rc = %d\n", set->set_oa->o_id,
1131 req->rq_oa->o_id, req->rq_idx, rc);
1136 err = lov_fini_punch_set(set);
1142 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1143 struct lov_stripe_md *lsm, obd_off start, obd_off end)
1145 struct lov_request_set *set;
1146 struct lov_obd *lov;
1147 struct list_head *pos;
1148 struct lov_request *req;
1149 int err = 0, rc = 0;
1152 if (lsm_bad_magic(lsm))
1158 lov = &exp->exp_obd->u.lov;
1159 rc = lov_prep_sync_set(exp, oa, lsm, start, end, &set);
1163 list_for_each (pos, &set->set_list) {
1164 req = list_entry(pos, struct lov_request, rq_link);
1166 rc = obd_sync(lov->tgts[req->rq_idx].ltd_exp, req->rq_oa,
1167 NULL, req->rq_extent.start, req->rq_extent.end);
1168 err = lov_update_common_set(set, req, rc);
1170 CERROR("error: fsync objid "LPX64" subobj "LPX64
1171 " on OST idx %d: rc = %d\n", set->set_oa->o_id,
1172 req->rq_oa->o_id, req->rq_idx, rc);
1177 err = lov_fini_sync_set(set);
1183 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1184 struct lov_stripe_md *lsm,
1185 obd_count oa_bufs, struct brw_page *pga)
1190 /* The caller just wants to know if there's a chance that this
1191 * I/O can succeed */
1192 for (i = 0; i < oa_bufs; i++) {
1193 int stripe = lov_stripe_number(lsm, pga[i].disk_offset);
1194 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1197 if (!lov_stripe_intersects(lsm, i, pga[i].disk_offset,
1198 pga[i].disk_offset + pga[i].count,
1202 if (lov->tgts[ost].active == 0) {
1203 CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1206 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[ost].ltd_exp, oa,
1207 NULL, 1, &pga[i], NULL);
1214 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1215 struct lov_stripe_md *lsm, obd_count oa_bufs,
1216 struct brw_page *pga, struct obd_trans_info *oti)
1218 struct lov_request_set *set;
1219 struct lov_request *req;
1220 struct list_head *pos;
1221 struct lov_obd *lov = &exp->exp_obd->u.lov;
1225 if (lsm_bad_magic(lsm))
1228 if (cmd == OBD_BRW_CHECK) {
1229 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1233 rc = lov_prep_brw_set(exp, src_oa, lsm, oa_bufs, pga, oti, &set);
1237 list_for_each (pos, &set->set_list) {
1238 struct obd_export *sub_exp;
1239 struct brw_page *sub_pga;
1240 req = list_entry(pos, struct lov_request, rq_link);
1242 sub_exp = lov->tgts[req->rq_idx].ltd_exp;
1243 sub_pga = set->set_pga + req->rq_pgaidx;
1244 rc = obd_brw(cmd, sub_exp, req->rq_oa, req->rq_md,
1245 req->rq_oabufs, sub_pga, oti);
1248 lov_update_common_set(set, req, rc);
1251 err = lov_fini_brw_set(set);
1257 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1260 struct lov_request_set *lovset = (struct lov_request_set *)data;
1264 lovset->set_completes = 0;
1265 lov_fini_brw_set(lovset);
1267 rc = lov_fini_brw_set(lovset);
1273 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1274 struct lov_stripe_md *lsm, obd_count oa_bufs,
1275 struct brw_page *pga, struct ptlrpc_request_set *set,
1276 struct obd_trans_info *oti)
1278 struct lov_request_set *lovset;
1279 struct lov_request *req;
1280 struct list_head *pos;
1281 struct lov_obd *lov = &exp->exp_obd->u.lov;
1285 if (lsm_bad_magic(lsm))
1288 if (cmd == OBD_BRW_CHECK) {
1289 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
1293 rc = lov_prep_brw_set(exp, oa, lsm, oa_bufs, pga, oti, &lovset);
1297 list_for_each (pos, &lovset->set_list) {
1298 struct obd_export *sub_exp;
1299 struct brw_page *sub_pga;
1300 req = list_entry(pos, struct lov_request, rq_link);
1302 sub_exp = lov->tgts[req->rq_idx].ltd_exp;
1303 sub_pga = lovset->set_pga + req->rq_pgaidx;
1304 rc = obd_brw_async(cmd, sub_exp, req->rq_oa, req->rq_md,
1305 req->rq_oabufs, sub_pga, set, oti);
1308 lov_update_common_set(lovset, req, rc);
1311 LASSERT(set->set_interpret == NULL);
1312 set->set_interpret = (set_interpreter_func)lov_brw_interpret;
1313 set->set_arg = (void *)lovset;
1317 lov_fini_brw_set(lovset);
1321 static int lov_ap_make_ready(void *data, int cmd)
1323 struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1325 return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
1327 static int lov_ap_refresh_count(void *data, int cmd)
1329 struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1331 return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
1334 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
1336 struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1338 lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
1339 /* XXX woah, shouldn't we be altering more here? size? */
1340 oa->o_id = lap->lap_loi_id;
1343 static void lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
1345 struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1347 /* in a raid1 regime this would down a count of many ios
1348 * in flight, onl calling the caller_ops completion when all
1349 * the raid1 ios are complete */
1350 lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, oa, rc);
1353 static struct obd_async_page_ops lov_async_page_ops = {
1354 .ap_make_ready = lov_ap_make_ready,
1355 .ap_refresh_count = lov_ap_refresh_count,
1356 .ap_fill_obdo = lov_ap_fill_obdo,
1357 .ap_completion = lov_ap_completion,
1360 static int lov_prep_async_page(struct obd_export *exp,
1361 struct lov_stripe_md *lsm,
1362 struct lov_oinfo *loi, struct page *page,
1363 obd_off offset, struct obd_async_page_ops *ops,
1364 void *data, void **res)
1366 struct lov_obd *lov = &exp->exp_obd->u.lov;
1367 struct lov_async_page *lap;
1371 if (lsm_bad_magic(lsm))
1373 LASSERT(loi == NULL);
1375 stripe = lov_stripe_number(lsm, offset);
1376 loi = &lsm->lsm_oinfo[stripe];
1378 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
1380 if (lov->tgts[loi->loi_ost_idx].active == 0)
1382 if (lov->tgts[loi->loi_ost_idx].ltd_exp == NULL) {
1383 CERROR("ltd_exp == NULL, but OST idx %d doesn't appear to be "
1384 "deleted or inactive.\n", loi->loi_ost_idx);
1388 OBD_ALLOC(lap, sizeof(*lap));
1392 lap->lap_magic = LAP_MAGIC;
1393 lap->lap_caller_ops = ops;
1394 lap->lap_caller_data = data;
1396 /* FIXME handle multiple oscs after landing b_raid1 */
1397 lap->lap_stripe = stripe;
1398 switch (lsm->lsm_pattern) {
1399 case LOV_PATTERN_RAID0:
1400 lov_stripe_offset(lsm, offset, lap->lap_stripe,
1401 &lap->lap_sub_offset);
1403 case LOV_PATTERN_CMOBD:
1404 lap->lap_sub_offset = offset;
1410 /* so the callback doesn't need the lsm */
1411 lap->lap_loi_id = loi->loi_id;
1413 rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
1414 lsm, loi, page, lap->lap_sub_offset,
1415 &lov_async_page_ops, lap,
1416 &lap->lap_sub_cookie);
1418 OBD_FREE(lap, sizeof(*lap));
1421 CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
1422 lap->lap_sub_cookie, offset);
1427 static int lov_queue_async_io(struct obd_export *exp,
1428 struct lov_stripe_md *lsm,
1429 struct lov_oinfo *loi, void *cookie,
1430 int cmd, obd_off off, int count,
1431 obd_flags brw_flags, obd_flags async_flags)
1433 struct lov_obd *lov = &exp->exp_obd->u.lov;
1434 struct lov_async_page *lap;
1437 LASSERT(loi == NULL);
1439 if (lsm_bad_magic(lsm))
1442 lap = LAP_FROM_COOKIE(cookie);
1444 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1446 rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
1447 loi, lap->lap_sub_cookie, cmd, off, count,
1448 brw_flags, async_flags);
1452 static int lov_set_async_flags(struct obd_export *exp,
1453 struct lov_stripe_md *lsm,
1454 struct lov_oinfo *loi, void *cookie,
1455 obd_flags async_flags)
1457 struct lov_obd *lov = &exp->exp_obd->u.lov;
1458 struct lov_async_page *lap;
1461 LASSERT(loi == NULL);
1463 if (lsm_bad_magic(lsm))
1466 lap = LAP_FROM_COOKIE(cookie);
1468 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1470 rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
1471 lsm, loi, lap->lap_sub_cookie, async_flags);
1475 static int lov_queue_group_io(struct obd_export *exp,
1476 struct lov_stripe_md *lsm,
1477 struct lov_oinfo *loi,
1478 struct obd_io_group *oig, void *cookie,
1479 int cmd, obd_off off, int count,
1480 obd_flags brw_flags, obd_flags async_flags)
1482 struct lov_obd *lov = &exp->exp_obd->u.lov;
1483 struct lov_async_page *lap;
1486 LASSERT(loi == NULL);
1488 if (lsm_bad_magic(lsm))
1491 lap = LAP_FROM_COOKIE(cookie);
1493 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1495 rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
1496 oig, lap->lap_sub_cookie, cmd, off, count,
1497 brw_flags, async_flags);
1501 /* this isn't exactly optimal. we may have queued sync io in oscs on
1502 * all stripes, but we don't record that fact at queue time. so we
1503 * trigger sync io on all stripes. */
1504 static int lov_trigger_group_io(struct obd_export *exp,
1505 struct lov_stripe_md *lsm,
1506 struct lov_oinfo *loi,
1507 struct obd_io_group *oig)
1509 struct lov_obd *lov = &exp->exp_obd->u.lov;
1512 LASSERT(loi == NULL);
1514 if (lsm_bad_magic(lsm))
1517 loi = lsm->lsm_oinfo;
1518 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1519 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1520 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1524 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
1526 if (rc == 0 && err != 0)
1532 static int lov_teardown_async_page(struct obd_export *exp,
1533 struct lov_stripe_md *lsm,
1534 struct lov_oinfo *loi, void *cookie)
1536 struct lov_obd *lov = &exp->exp_obd->u.lov;
1537 struct lov_async_page *lap;
1540 LASSERT(loi == NULL);
1542 if (lsm_bad_magic(lsm))
1545 lap = LAP_FROM_COOKIE(cookie);
1547 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1549 rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
1550 lsm, loi, lap->lap_sub_cookie);
1552 CERROR("unable to teardown sub cookie %p: %d\n",
1553 lap->lap_sub_cookie, rc);
1556 OBD_FREE(lap, sizeof(*lap));
1560 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
1561 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
1562 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1563 void *data,__u32 lvb_len, void *lvb_swabber,
1564 struct lustre_handle *lockh)
1566 struct lov_request_set *set;
1567 struct lov_request *req;
1568 struct list_head *pos;
1569 struct lustre_handle *lov_lockhp;
1570 struct lov_obd *lov;
1572 int save_flags = *flags;
1575 if (lsm_bad_magic(lsm))
1578 /* we should never be asked to replay a lock this way. */
1579 LASSERT((*flags & LDLM_FL_REPLAY) == 0);
1581 if (!exp || !exp->exp_obd)
1584 lov = &exp->exp_obd->u.lov;
1585 rc = lov_prep_enqueue_set(exp, lsm, policy, mode, lockh, &set);
1589 list_for_each (pos, &set->set_list) {
1590 ldlm_policy_data_t sub_policy;
1591 req = list_entry(pos, struct lov_request, rq_link);
1592 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
1593 LASSERT(lov_lockhp);
1595 *flags = save_flags;
1596 sub_policy.l_extent.start = req->rq_extent.start;
1597 sub_policy.l_extent.end = req->rq_extent.end;
1599 rc = obd_enqueue(lov->tgts[req->rq_idx].ltd_exp, req->rq_md,
1600 type, &sub_policy, mode, flags, bl_cb,
1601 cp_cb, gl_cb, data, lvb_len, lvb_swabber,
1603 rc = lov_update_enqueue_set(set, req, rc, save_flags);
1608 lov_fini_enqueue_set(set, mode);
1612 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
1613 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
1614 int *flags, void *data, struct lustre_handle *lockh)
1616 struct lov_request_set *set;
1617 struct lov_request *req;
1618 struct list_head *pos;
1619 struct lov_obd *lov = &exp->exp_obd->u.lov;
1620 struct lustre_handle *lov_lockhp;
1621 int lov_flags, rc = 0;
1624 if (lsm_bad_magic(lsm))
1627 if (!exp || !exp->exp_obd)
1630 lov = &exp->exp_obd->u.lov;
1631 rc = lov_prep_match_set(exp, lsm, policy, mode, lockh, &set);
1635 list_for_each (pos, &set->set_list) {
1636 ldlm_policy_data_t sub_policy;
1637 req = list_entry(pos, struct lov_request, rq_link);
1638 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
1639 LASSERT(lov_lockhp);
1641 sub_policy.l_extent.start = req->rq_extent.start;
1642 sub_policy.l_extent.end = req->rq_extent.end;
1645 rc = obd_match(lov->tgts[req->rq_idx].ltd_exp, req->rq_md,
1646 type, &sub_policy, mode, &lov_flags, data,
1648 rc = lov_update_match_set(set, req, rc);
1652 lov_fini_match_set(set, mode, *flags);
1656 static int dump_missed_lock(struct ldlm_lock *lock, void *data)
1658 LDLM_ERROR(lock, "forgotten lock");
1659 return LDLM_ITER_CONTINUE;
1662 static int lov_change_cbdata(struct obd_export *exp,
1663 struct lov_stripe_md *lsm, ldlm_iterator_t it,
1666 struct lov_obd *lov;
1667 struct lov_oinfo *loi;
1671 if (lsm_bad_magic(lsm))
1674 if (!exp || !exp->exp_obd)
1677 LASSERT(lsm->lsm_object_gr > 0);
1679 lov = &exp->exp_obd->u.lov;
1680 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1681 struct lov_stripe_md submd;
1682 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1683 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1684 submd.lsm_object_id = loi->loi_id;
1685 submd.lsm_object_gr = lsm->lsm_object_gr;
1686 submd.lsm_stripe_count = 0;
1687 obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
1688 &submd, dump_missed_lock, NULL);
1692 submd.lsm_object_id = loi->loi_id;
1693 submd.lsm_object_gr = lsm->lsm_object_gr;
1694 submd.lsm_stripe_count = 0;
1695 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
1701 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
1702 __u32 mode, struct lustre_handle *lockh)
1704 struct lov_request_set *set;
1705 struct lov_request *req;
1706 struct list_head *pos;
1707 struct lov_obd *lov = &exp->exp_obd->u.lov;
1708 struct lustre_handle *lov_lockhp;
1709 int err = 0, rc = 0;
1712 if (lsm_bad_magic(lsm))
1715 if (!exp || !exp->exp_obd)
1718 LASSERT(lsm->lsm_object_gr > 0);
1721 lov = &exp->exp_obd->u.lov;
1722 rc = lov_prep_cancel_set(exp, lsm, mode, lockh, &set);
1726 list_for_each (pos, &set->set_list) {
1727 req = list_entry(pos, struct lov_request, rq_link);
1728 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
1730 rc = obd_cancel(lov->tgts[req->rq_idx].ltd_exp, req->rq_md,
1732 rc = lov_update_common_set(set, req, rc);
1734 CERROR("error: cancel objid "LPX64" subobj "
1735 LPX64" on OST idx %d: rc = %d\n",
1737 req->rq_md->lsm_object_id, req->rq_idx, rc);
1742 lov_fini_cancel_set(set);
1746 static int lov_cancel_unused(struct obd_export *exp,
1747 struct lov_stripe_md *lsm,
1748 int flags, void *opaque)
1750 struct lov_obd *lov;
1751 struct lov_oinfo *loi;
1755 lov = &exp->exp_obd->u.lov;
1757 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1758 int err = obd_cancel_unused(lov->tgts[i].ltd_exp,
1759 NULL, flags, opaque);
1766 if (lsm_bad_magic(lsm))
1769 if (!exp || !exp->exp_obd)
1772 LASSERT(lsm->lsm_object_gr > 0);
1774 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1775 struct lov_stripe_md submd;
1778 if (lov->tgts[loi->loi_ost_idx].active == 0)
1779 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1781 submd.lsm_object_id = loi->loi_id;
1782 submd.lsm_object_gr = lsm->lsm_object_gr;
1783 submd.lsm_stripe_count = 0;
1784 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
1785 &submd, flags, opaque);
1786 if (err && lov->tgts[loi->loi_ost_idx].active) {
1787 CERROR("error: cancel unused objid "LPX64" subobj "LPX64
1788 " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
1789 loi->loi_id, loi->loi_ost_idx, err);
1797 #define LOV_U64_MAX ((__u64)~0ULL)
1798 #define LOV_SUM_MAX(tot, add) \
1800 if ((tot) + (add) < (tot)) \
1801 (tot) = LOV_U64_MAX; \
1806 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1807 unsigned long max_age)
1809 struct lov_obd *lov = &obd->u.lov;
1810 struct obd_statfs lov_sfs;
1817 /* We only get block data from the OBD */
1818 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1820 if (!lov->tgts[i].active) {
1821 CDEBUG(D_HA, "lov idx %d inactive\n", i);
1825 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
1828 if (lov->tgts[i].active && !rc)
1834 memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
1837 osfs->os_bfree += lov_sfs.os_bfree;
1838 osfs->os_bavail += lov_sfs.os_bavail;
1839 osfs->os_blocks += lov_sfs.os_blocks;
1840 /* XXX not sure about this one - depends on policy.
1841 * - could be minimum if we always stripe on all OBDs
1842 * (but that would be wrong for any other policy,
1843 * if one of the OBDs has no more objects left)
1844 * - could be sum if we stripe whole objects
1845 * - could be average, just to give a nice number
1847 * To give a "reasonable" (if not wholly accurate)
1848 * number, we divide the total number of free objects
1849 * by expected stripe count (watch out for overflow).
1851 LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
1852 LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
1857 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
1858 lov->desc.ld_default_stripe_count :
1859 lov->desc.ld_active_tgt_count;
1861 if (osfs->os_files != LOV_U64_MAX)
1862 do_div(osfs->os_files, expected_stripes);
1863 if (osfs->os_ffree != LOV_U64_MAX)
1864 do_div(osfs->os_ffree, expected_stripes);
1871 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1872 void *karg, void *uarg)
1874 struct obd_device *obddev = class_exp2obd(exp);
1875 struct lov_obd *lov = &obddev->u.lov;
1876 int i, rc, count = lov->desc.ld_tgt_count;
1877 struct obd_uuid *uuidp;
1881 case OBD_IOC_LOV_GET_CONFIG: {
1882 struct obd_ioctl_data *data = karg;
1883 struct lov_tgt_desc *tgtdesc;
1884 struct lov_desc *desc;
1890 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
1893 data = (struct obd_ioctl_data *)buf;
1895 if (sizeof(*desc) > data->ioc_inllen1) {
1896 obd_ioctl_freedata(buf, len);
1900 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
1901 obd_ioctl_freedata(buf, len);
1905 if (sizeof(__u32) * count > data->ioc_inllen3) {
1906 obd_ioctl_freedata(buf, len);
1910 desc = (struct lov_desc *)data->ioc_inlbuf1;
1911 memcpy(desc, &(lov->desc), sizeof(*desc));
1913 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
1914 genp = (__u32 *)data->ioc_inlbuf3;
1915 tgtdesc = lov->tgts;
1916 /* the uuid will be empty for deleted OSTs */
1917 for (i = 0; i < count; i++, uuidp++, genp++, tgtdesc++) {
1918 obd_str2uuid(uuidp, (char *)tgtdesc->uuid.uuid);
1919 *genp = tgtdesc->ltd_gen;
1922 rc = copy_to_user((void *)uarg, buf, len);
1925 obd_ioctl_freedata(buf, len);
1928 case LL_IOC_LOV_SETSTRIPE:
1929 rc = lov_setstripe(exp, karg, uarg);
1931 case LL_IOC_LOV_GETSTRIPE:
1932 rc = lov_getstripe(exp, karg, uarg);
1934 case LL_IOC_LOV_SETEA:
1935 rc = lov_setea(exp, karg, uarg);
1942 for (i = 0; i < count; i++) {
1945 /* OST was deleted */
1946 if (obd_uuid_empty(&lov->tgts[i].uuid))
1949 err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
1952 if (lov->tgts[i].active) {
1953 CERROR("error: iocontrol OSC %s on OST "
1954 "idx %d cmd %x: err = %d\n",
1955 lov->tgts[i].uuid.uuid, i,
1971 static int lov_get_info(struct obd_export *exp, __u32 keylen,
1972 void *key, __u32 *vallen, void *val)
1974 struct obd_device *obddev = class_exp2obd(exp);
1975 struct lov_obd *lov = &obddev->u.lov;
1979 if (!vallen || !val)
1982 if (keylen > strlen("lock_to_stripe") &&
1983 strcmp(key, "lock_to_stripe") == 0) {
1986 struct ldlm_lock *lock;
1987 struct lov_stripe_md *lsm;
1989 struct lov_oinfo *loi;
1990 struct ldlm_res_id *res_id = &data->lock->l_resource->lr_name;
1991 __u32 *stripe = val;
1993 if (*vallen < sizeof(*stripe))
1995 *vallen = sizeof(*stripe);
1997 /* XXX This is another one of those bits that will need to
1998 * change if we ever actually support nested LOVs. It uses
1999 * the lock's export to find out which stripe it is. */
2000 /* XXX - it's assumed all the locks for deleted OSTs have
2001 * been cancelled. Also, the export for deleted OSTs will
2002 * be NULL and won't match the lock's export. */
2003 for (i = 0, loi = data->lsm->lsm_oinfo;
2004 i < data->lsm->lsm_stripe_count;
2006 if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
2007 data->lock->l_conn_export &&
2008 loi->loi_id == res_id->name[0] &&
2009 loi->loi_gr == res_id->name[2]) {
2014 LDLM_ERROR(data->lock, "lock on inode without such object");
2015 dump_lsm(D_ERROR, data->lsm);
2017 } else if (keylen >= strlen("size_to_stripe") &&
2018 strcmp(key, "size_to_stripe") == 0) {
2022 struct lov_stripe_md *lsm;
2025 if (*vallen < sizeof(*data))
2028 data->size = lov_size_to_stripe(data->lsm, data->size,
2029 data->stripe_number);
2031 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2032 __u32 size = sizeof(obd_id);
2036 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2037 if (!lov->tgts[i].active)
2039 rc = obd_get_info(lov->tgts[i].ltd_exp,
2040 keylen, key, &size, &(ids[i]));
2045 } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
2046 struct lov_desc *desc_ret = val;
2047 *desc_ret = lov->desc;
2055 static int lov_set_info(struct obd_export *exp, obd_count keylen,
2056 void *key, obd_count vallen, void *val)
2058 struct obd_device *obddev = class_exp2obd(exp);
2059 struct lov_obd *lov = &obddev->u.lov;
2063 #define KEY_IS(str) \
2064 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
2066 if (KEY_IS("async")) {
2067 struct lov_desc *desc = &lov->desc;
2068 struct lov_tgt_desc *tgts = lov->tgts;
2070 if (vallen != sizeof(int))
2072 lov->async = *((int*) val);
2074 for (i = 0; i < desc->ld_tgt_count; i++, tgts++) {
2075 struct obd_uuid *tgt_uuid = &tgts->uuid;
2076 struct obd_device *tgt_obd;
2078 tgt_obd = class_find_client_obd(tgt_uuid,
2082 CERROR("Target %s not attached\n",
2089 err = obd_set_info(tgt_obd->obd_self_export,
2090 keylen, key, vallen, val);
2092 CERROR("Failed to set async on target %s\n",
2101 if (KEY_IS("mds_conn")) {
2102 if (vallen != sizeof(__u32))
2104 } else if (KEY_IS("unlinked") || KEY_IS("unrecovery")) {
2107 } else if (KEY_IS("sec") || KEY_IS("sec_flags")) {
2108 struct lov_tgt_desc *tgt;
2109 struct obd_export *exp;
2112 spin_lock(&lov->lov_lock);
2113 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count;
2116 /* during setup time the connections to osc might
2117 * haven't been established.
2120 struct obd_device *tgt_obd;
2122 tgt_obd = class_find_client_obd(&tgt->uuid,
2126 CERROR("can't set security flavor, "
2127 "device %s not attached?\n",
2132 exp = tgt_obd->obd_self_export;
2135 err = obd_set_info(exp, keylen, key, vallen, val);
2139 spin_unlock(&lov->lov_lock);
2142 } else if (KEY_IS("flush_cred")) {
2143 struct lov_tgt_desc *tgt;
2146 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count;
2150 rc = obd_set_info(tgt->ltd_exp,
2151 keylen, key, vallen, val);
2161 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2162 if (val && !obd_uuid_equals(val, &lov->tgts[i].uuid))
2165 if (!val && !lov->tgts[i].active)
2168 err = obd_set_info(lov->tgts[i].ltd_exp,
2169 keylen, key, vallen, val);
2178 struct lov_multi_wait {
2179 struct ldlm_lock *lock;
2185 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
2186 struct lustre_handle *lockh)
2188 struct lov_lock_handles *lov_lockh = NULL;
2189 struct lustre_handle *lov_lockhp;
2190 struct lov_obd *lov;
2191 struct lov_oinfo *loi;
2192 struct lov_multi_wait *queues;
2196 if (lsm_bad_magic(lsm))
2199 if (!exp || !exp->exp_obd)
2202 LASSERT(lockh != NULL);
2203 if (lsm->lsm_stripe_count > 1) {
2204 lov_lockh = lov_handle2llh(lockh);
2205 if (lov_lockh == NULL) {
2206 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2210 lov_lockhp = lov_lockh->llh_handles;
2215 OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
2217 GOTO(out, rc = -ENOMEM);
2219 lov = &exp->exp_obd->u.lov;
2220 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2221 i++, loi++, lov_lockhp++) {
2222 struct ldlm_lock *lock;
2223 struct obd_device *obd;
2224 unsigned long irqflags;
2226 lock = ldlm_handle2lock(lov_lockhp);
2228 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2229 loi->loi_ost_idx, loi->loi_id);
2230 queues[i].completed = 1;
2234 queues[i].lock = lock;
2235 init_waitqueue_entry(&(queues[i].wait), current);
2236 add_wait_queue(lock->l_waitq, &(queues[i].wait));
2238 obd = class_exp2obd(lock->l_conn_export);
2240 imp = obd->u.cli.cl_import;
2242 spin_lock_irqsave(&imp->imp_lock, irqflags);
2243 queues[i].generation = imp->imp_generation;
2244 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
2248 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
2249 interrupted_completion_wait, &lwd);
2250 rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
2252 for (i = 0; i < lsm->lsm_stripe_count; i++)
2253 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
2255 if (rc == -EINTR || rc == -ETIMEDOUT) {
2261 if (lov_lockh != NULL)
2262 lov_llh_put(lov_lockh);
2267 struct obd_ops lov_obd_ops = {
2268 .o_owner = THIS_MODULE,
2269 .o_attach = lov_attach,
2270 .o_detach = lov_detach,
2271 .o_setup = lov_setup,
2272 .o_cleanup = lov_cleanup,
2273 .o_process_config = lov_process_config,
2274 .o_connect = lov_connect,
2275 .o_disconnect = lov_disconnect,
2276 .o_statfs = lov_statfs,
2277 .o_packmd = lov_packmd,
2278 .o_unpackmd = lov_unpackmd,
2279 .o_revalidate_md = lov_revalidate_md,
2280 .o_create = lov_create,
2281 .o_destroy = lov_destroy,
2282 .o_getattr = lov_getattr,
2283 .o_getattr_async = lov_getattr_async,
2284 .o_setattr = lov_setattr,
2286 .o_brw_async = lov_brw_async,
2287 .o_prep_async_page = lov_prep_async_page,
2288 .o_queue_async_io = lov_queue_async_io,
2289 .o_set_async_flags = lov_set_async_flags,
2290 .o_queue_group_io = lov_queue_group_io,
2291 .o_trigger_group_io = lov_trigger_group_io,
2292 .o_teardown_async_page = lov_teardown_async_page,
2293 .o_adjust_kms = lov_adjust_kms,
2294 .o_punch = lov_punch,
2296 .o_enqueue = lov_enqueue,
2297 .o_match = lov_match,
2298 .o_change_cbdata = lov_change_cbdata,
2299 .o_cancel = lov_cancel,
2300 .o_cancel_unused = lov_cancel_unused,
2301 .o_iocontrol = lov_iocontrol,
2302 .o_get_info = lov_get_info,
2303 .o_set_info = lov_set_info,
2304 .o_llog_init = lov_llog_init,
2305 .o_llog_finish = lov_llog_finish,
2306 .o_notify = lov_notify,
2309 int __init lov_init(void)
2311 struct lprocfs_static_vars lvars;
2315 lprocfs_init_vars(lov, &lvars);
2316 rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars,
2317 OBD_LOV_DEVICENAME);
2322 static void /*__exit*/ lov_exit(void)
2324 class_unregister_type(OBD_LOV_DEVICENAME);
2327 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2328 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
2329 MODULE_LICENSE("GPL");
2331 module_init(lov_init);
2332 module_exit(lov_exit);