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"
54 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
55 int stripeno, obd_off *obd_off);
57 struct lov_lock_handles {
58 struct portals_handle llh_handle;
59 atomic_t llh_refcount;
61 struct lustre_handle llh_handles[0];
64 static void lov_llh_addref(void *llhp)
66 struct lov_lock_handles *llh = llhp;
68 atomic_inc(&llh->llh_refcount);
69 CDEBUG(D_INFO, "GETting llh %p : new refcount %d\n", llh,
70 atomic_read(&llh->llh_refcount));
73 static struct lov_lock_handles *lov_llh_new(struct lov_stripe_md *lsm)
75 struct lov_lock_handles *llh;
77 OBD_ALLOC(llh, sizeof *llh +
78 sizeof(*llh->llh_handles) * lsm->lsm_stripe_count);
80 CERROR("out of memory\n");
83 atomic_set(&llh->llh_refcount, 2);
84 llh->llh_stripe_count = lsm->lsm_stripe_count;
85 INIT_LIST_HEAD(&llh->llh_handle.h_link);
86 class_handle_hash(&llh->llh_handle, lov_llh_addref);
90 static struct lov_lock_handles *lov_handle2llh(struct lustre_handle *handle)
93 LASSERT(handle != NULL);
94 RETURN(class_handle2object(handle->cookie));
97 static void lov_llh_put(struct lov_lock_handles *llh)
99 CDEBUG(D_INFO, "PUTting llh %p : new refcount %d\n", llh,
100 atomic_read(&llh->llh_refcount) - 1);
101 LASSERT(atomic_read(&llh->llh_refcount) > 0 &&
102 atomic_read(&llh->llh_refcount) < 0x5a5a);
103 if (atomic_dec_and_test(&llh->llh_refcount)) {
104 LASSERT(list_empty(&llh->llh_handle.h_link));
105 OBD_FREE(llh, sizeof *llh +
106 sizeof(*llh->llh_handles) * llh->llh_stripe_count);
110 static void lov_llh_destroy(struct lov_lock_handles *llh)
112 class_handle_unhash(&llh->llh_handle);
117 #define MAX_STRING_SIZE 128
118 static int lov_connect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt,
119 int activate, unsigned long connect_flags)
121 struct lov_obd *lov = &obd->u.lov;
122 struct obd_uuid *tgt_uuid = &tgt->uuid;
123 struct obd_device *tgt_obd;
124 struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
125 struct lustre_handle conn = {0, };
126 struct proc_dir_entry *lov_proc_dir;
130 tgt_obd = class_find_client_obd(tgt_uuid, LUSTRE_OSC_NAME,
134 CERROR("Target %s not attached\n", tgt_uuid->uuid);
138 if (!tgt_obd->obd_set_up) {
139 CERROR("Target %s not set up\n", tgt_uuid->uuid);
144 tgt_obd->obd_no_recov = 0;
145 ptlrpc_activate_import(tgt_obd->u.cli.cl_import);
148 if (tgt_obd->u.cli.cl_import->imp_invalid) {
149 CERROR("not connecting OSC %s; administratively "
150 "disabled\n", tgt_uuid->uuid);
151 rc = obd_register_observer(tgt_obd, obd);
153 CERROR("Target %s register_observer error %d; "
154 "will not be able to reactivate\n",
160 rc = obd_connect(&conn, tgt_obd, &lov_osc_uuid, connect_flags);
162 CERROR("Target %s connect error %d\n", tgt_uuid->uuid, rc);
165 tgt->ltd_exp = class_conn2export(&conn);
167 rc = obd_register_observer(tgt_obd, obd);
169 CERROR("Target %s register_observer error %d\n",
171 obd_disconnect(tgt->ltd_exp, 0);
177 lov->desc.ld_active_tgt_count++;
179 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
181 struct obd_device *osc_obd = class_conn2obd(&conn);
182 struct proc_dir_entry *osc_symlink;
183 char name[MAX_STRING_SIZE + 1];
185 LASSERT(osc_obd != NULL);
186 LASSERT(osc_obd->obd_type != NULL);
187 LASSERT(osc_obd->obd_type->typ_name != NULL);
188 name[MAX_STRING_SIZE] = '\0';
189 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
190 osc_obd->obd_type->typ_name,
192 osc_symlink = proc_symlink(osc_obd->obd_name, lov_proc_dir,
194 if (osc_symlink == NULL) {
195 CERROR("could not register LOV target "
196 "/proc/fs/lustre/%s/%s/target_obds/%s\n",
197 obd->obd_type->typ_name, obd->obd_name,
199 lprocfs_remove(lov_proc_dir);
207 static int lov_connect(struct lustre_handle *conn, struct obd_device *obd,
208 struct obd_uuid *cluuid, unsigned long connect_flags)
210 struct lov_obd *lov = &obd->u.lov;
211 struct lov_tgt_desc *tgt;
212 struct obd_export *exp;
213 struct proc_dir_entry *lov_proc_dir;
217 rc = class_connect(conn, obd, cluuid);
221 exp = class_conn2export(conn);
223 /* We don't want to actually do the underlying connections more than
224 * once, so keep track. */
226 if (lov->refcount > 1) {
227 class_export_put(exp);
231 lov_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
233 if (IS_ERR(lov_proc_dir)) {
234 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
235 obd->obd_type->typ_name, obd->obd_name);
239 /* connect_flags is the MDS number, save for use in lov_add_obd */
240 lov->lov_connect_flags = connect_flags;
241 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
242 if (obd_uuid_empty(&tgt->uuid))
244 rc = lov_connect_obd(obd, tgt, 0, connect_flags);
249 class_export_put(exp);
254 lprocfs_remove(lov_proc_dir);
257 struct obd_uuid uuid;
259 --lov->desc.ld_active_tgt_count;
261 /* save for CERROR below; (we know it's terminated) */
263 rc2 = obd_disconnect(tgt->ltd_exp, 0);
265 CERROR("error: LOV target %s disconnect on OST idx %d: "
266 "rc = %d\n", uuid.uuid, i, rc2);
268 class_disconnect(exp, 0);
272 static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt,
275 struct proc_dir_entry *lov_proc_dir;
276 struct obd_device *osc_obd = class_exp2obd(tgt->ltd_exp);
277 struct lov_obd *lov = &obd->u.lov;
281 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
283 struct proc_dir_entry *osc_symlink;
285 osc_symlink = lprocfs_srch(lov_proc_dir, osc_obd->obd_name);
287 lprocfs_remove(osc_symlink);
289 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
290 obd->obd_type->typ_name, obd->obd_name,
295 if (obd->obd_no_recov) {
296 /* Pass it on to our clients.
297 * XXX This should be an argument to disconnect,
298 * XXX not a back-door flag on the OBD. Ah well.
301 osc_obd->obd_no_recov = 1;
304 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
306 rc = obd_disconnect(tgt->ltd_exp, flags);
309 CERROR("Target %s disconnect error %d\n",
317 lov->desc.ld_active_tgt_count--;
324 static int lov_disconnect(struct obd_export *exp, int flags)
326 struct obd_device *obd = class_exp2obd(exp);
327 struct lov_obd *lov = &obd->u.lov;
328 struct lov_tgt_desc *tgt;
329 struct proc_dir_entry *lov_proc_dir;
336 /* Only disconnect the underlying layers on the final disconnect. */
338 if (lov->refcount != 0)
341 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
343 lov_disconnect_obd(obd, tgt, flags);
345 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
347 lprocfs_remove(lov_proc_dir);
349 CERROR("/proc/fs/lustre/%s/%s/target_obds missing.",
350 obd->obd_type->typ_name, obd->obd_name);
355 rc = class_disconnect(exp, 0);
361 * -EINVAL : UUID can't be found in the LOV's target list
362 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
363 * -EBADF : The UUID is found, but the OBD is the wrong type (!)
365 static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid,
368 struct lov_tgt_desc *tgt;
372 CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
373 lov, uuid->uuid, activate);
375 spin_lock(&lov->lov_lock);
376 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
377 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
378 i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
379 if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0)
383 if (i == lov->desc.ld_tgt_count)
384 GOTO(out, rc = -EINVAL);
387 if (tgt->active == activate) {
388 CDEBUG(D_INFO, "OSC %s already %sactive!\n", uuid->uuid,
389 activate ? "" : "in");
393 CDEBUG(D_INFO, "Marking OSC %s %sactive\n", uuid->uuid,
394 activate ? "" : "in");
396 tgt->active = activate;
398 lov->desc.ld_active_tgt_count++;
400 lov->desc.ld_active_tgt_count--;
404 spin_unlock(&lov->lov_lock);
408 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
409 int active, void *data)
412 struct obd_uuid *uuid;
415 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
416 CERROR("unexpected notification of %s %s!\n",
417 watched->obd_type->typ_name,
421 uuid = &watched->u.cli.cl_import->imp_target_uuid;
423 /* Set OSC as active before notifying the observer, so the
424 * observer can use the OSC normally.
426 rc = lov_set_osc_active(&obd->u.lov, uuid, active);
428 CERROR("%sactivation of %s failed: %d\n",
429 active ? "" : "de", uuid->uuid, rc);
433 if (obd->obd_observer)
434 /* Pass the notification up the chain. */
435 rc = obd_notify(obd->obd_observer, watched, active, data);
440 int lov_attach(struct obd_device *dev, obd_count len, void *data)
442 struct lprocfs_static_vars lvars;
445 lprocfs_init_vars(lov, &lvars);
446 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
449 struct proc_dir_entry *entry;
451 entry = create_proc_entry("target_obd_status", 0444,
452 dev->obd_proc_entry);
456 entry->proc_fops = &lov_proc_target_fops;
464 int lov_detach(struct obd_device *dev)
466 return lprocfs_obd_detach(dev);
469 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
471 struct lustre_cfg *lcfg = buf;
472 struct lov_desc *desc;
473 struct lov_obd *lov = &obd->u.lov;
477 if (lcfg->lcfg_inllen1 < 1) {
478 CERROR("LOV setup requires a descriptor\n");
482 desc = (struct lov_desc *)lcfg->lcfg_inlbuf1;
483 if (sizeof(*desc) > lcfg->lcfg_inllen1) {
484 CERROR("descriptor size wrong: %d > %d\n",
485 (int)sizeof(*desc), lcfg->lcfg_inllen1);
489 /* Because of 64-bit divide/mod operations only work with a 32-bit
490 * divisor in a 32-bit kernel, we cannot support a stripe width
491 * of 4GB or larger on 32-bit CPUs.
494 count = desc->ld_default_stripe_count;
495 if (count && (count * desc->ld_default_stripe_size) > ~0UL) {
496 CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n",
497 desc->ld_default_stripe_size, count, ~0UL);
500 if (desc->ld_tgt_count > 0) {
501 lov->bufsize= sizeof(struct lov_tgt_desc) * desc->ld_tgt_count;
503 lov->bufsize = sizeof(struct lov_tgt_desc) * LOV_MAX_TGT_COUNT;
505 OBD_ALLOC(lov->tgts, lov->bufsize);
506 if (lov->tgts == NULL) {
508 CERROR("couldn't allocate %d bytes for target table.\n",
513 desc->ld_tgt_count = 0;
514 desc->ld_active_tgt_count = 0;
516 spin_lock_init(&lov->lov_lock);
517 sema_init(&lov->lov_llog_sem, 1);
522 static int lov_cleanup(struct obd_device *obd, int flags)
524 struct lov_obd *lov = &obd->u.lov;
526 OBD_FREE(lov->tgts, lov->bufsize);
531 lov_add_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
533 struct lov_obd *lov = &obd->u.lov;
534 struct lov_tgt_desc *tgt;
538 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
539 uuidp->uuid, index, gen);
541 if ((index < 0) || (index >= LOV_MAX_TGT_COUNT)) {
542 CERROR("request to add OBD %s at invalid index: %d\n",
548 CERROR("request to add OBD %s with invalid generation: %d\n",
553 tgt = lov->tgts + index;
554 if (!obd_uuid_empty(&tgt->uuid)) {
555 CERROR("OBD already assigned at LOV target index %d\n",
561 /* XXX - add a sanity check on the generation number. */
564 if (index >= lov->desc.ld_tgt_count)
565 lov->desc.ld_tgt_count = index + 1;
567 CDEBUG(D_CONFIG, "idx: %d ltd_gen: %d ld_tgt_count: %d\n",
568 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
570 if (lov->refcount == 0)
574 struct obd_device *osc_obd;
576 osc_obd = class_exp2obd(tgt->ltd_exp);
578 osc_obd->obd_no_recov = 0;
581 rc = lov_connect_obd(obd, tgt, 1, lov->lov_connect_flags);
585 if (obd->obd_observer) {
586 /* tell the mds_lov about the new target */
587 rc = obd_notify(obd->obd_observer, tgt->ltd_exp->exp_obd, 1,
593 if (rc && tgt->ltd_exp != NULL)
594 lov_disconnect_obd(obd, tgt, 0);
599 lov_del_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
601 struct lov_obd *lov = &obd->u.lov;
602 struct lov_tgt_desc *tgt;
603 int count = lov->desc.ld_tgt_count;
607 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
608 uuidp->uuid, index, gen);
610 if (index >= count) {
611 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
616 tgt = lov->tgts + index;
618 if (obd_uuid_empty(&tgt->uuid)) {
619 CERROR("LOV target at index %d is not setup.\n", index);
623 if (strncmp(uuidp->uuid, tgt->uuid.uuid, sizeof uuidp->uuid) != 0) {
624 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
625 tgt->uuid.uuid, index, uuidp->uuid);
630 struct obd_device *osc_obd;
632 osc_obd = class_exp2obd(tgt->ltd_exp);
634 osc_obd->obd_no_recov = 1;
635 rc = obd_llog_finish(osc_obd, &osc_obd->obd_llogs, 1);
637 CERROR("osc_llog_finish error: %d\n", rc);
639 lov_disconnect_obd(obd, tgt, 0);
642 /* XXX - right now there is a dependency on ld_tgt_count being the
643 * maximum tgt index for computing the mds_max_easize. So we can't
646 /* lt_gen = 0 will mean it will not match the gen of any valid loi */
647 memset(tgt, 0, sizeof(*tgt));
649 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
650 tgt->uuid.uuid, index, tgt->ltd_gen, tgt->ltd_exp, tgt->active);
655 static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
657 struct lustre_cfg *lcfg = buf;
658 struct obd_uuid obd_uuid;
665 switch(cmd = lcfg->lcfg_command) {
666 case LCFG_LOV_ADD_OBD:
667 case LCFG_LOV_DEL_OBD: {
668 if (lcfg->lcfg_inllen1 > sizeof(obd_uuid.uuid))
669 GOTO(out, rc = -EINVAL);
671 obd_str2uuid(&obd_uuid, lcfg->lcfg_inlbuf1);
673 if (sscanf(lcfg->lcfg_inlbuf2, "%d", &index) != 1)
674 GOTO(out, rc = -EINVAL);
675 if (sscanf(lcfg->lcfg_inlbuf3, "%d", &gen) != 1)
676 GOTO(out, rc = -EINVAL);
677 if (cmd == LCFG_LOV_ADD_OBD)
678 rc = lov_add_obd(obd, &obd_uuid, index, gen);
680 rc = lov_del_obd(obd, &obd_uuid, index, gen);
684 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
685 GOTO(out, rc = -EINVAL);
693 /* compute object size given "stripeno" and the ost size */
694 static obd_size lov_stripe_size(struct lov_stripe_md *lsm, obd_size ost_size,
697 unsigned long ssize = lsm->lsm_stripe_size;
698 unsigned long swidth = ssize * lsm->lsm_stripe_count;
699 unsigned long stripe_size;
705 /* do_div(a, b) returns a % b, and a = a / b */
706 stripe_size = do_div(ost_size, ssize);
708 lov_size = ost_size * swidth + stripeno * ssize + stripe_size;
710 lov_size = (ost_size - 1) * swidth + (stripeno + 1) * ssize;
715 static void lov_merge_attrs(struct obdo *tgt, struct obdo *src, obd_flag valid,
716 struct lov_stripe_md *lsm, int stripeno, int *set)
718 valid &= src->o_valid;
721 if (valid & OBD_MD_FLSIZE) {
722 /* this handles sparse files properly */
725 lov_size = lov_stripe_size(lsm, src->o_size, stripeno);
726 if (lov_size > tgt->o_size)
727 tgt->o_size = lov_size;
729 if (valid & OBD_MD_FLBLOCKS)
730 tgt->o_blocks += src->o_blocks;
731 if (valid & OBD_MD_FLBLKSZ)
732 tgt->o_blksize += src->o_blksize;
733 if (valid & OBD_MD_FLCTIME && tgt->o_ctime < src->o_ctime)
734 tgt->o_ctime = src->o_ctime;
735 if (valid & OBD_MD_FLMTIME && tgt->o_mtime < src->o_mtime)
736 tgt->o_mtime = src->o_mtime;
738 memcpy(tgt, src, sizeof(*tgt));
739 tgt->o_id = lsm->lsm_object_id;
740 if (valid & OBD_MD_FLSIZE)
741 tgt->o_size = lov_stripe_size(lsm,src->o_size,stripeno);
747 #define log2(n) ffz(~(n))
750 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
751 struct lov_stripe_md **ea,
752 struct obd_trans_info *oti)
756 struct obd_uuid *ost_uuid = NULL;
760 LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
761 src_oa->o_flags == OBD_FL_DELORPHAN);
763 lov = &export->exp_obd->u.lov;
765 tmp_oa = obdo_alloc();
769 if (src_oa->o_valid & OBD_MD_FLINLINE) {
770 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
771 CDEBUG(D_HA, "clearing orphans only for %s\n",
775 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
776 struct lov_stripe_md obj_md;
777 struct lov_stripe_md *obj_mdp = &obj_md;
780 /* if called for a specific target, we don't
781 care if it is not active. */
782 if (lov->tgts[i].active == 0 && ost_uuid == NULL) {
783 CDEBUG(D_HA, "lov idx %d inactive\n", i);
787 if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid))
790 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
792 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
793 err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, &obj_mdp, oti);
795 /* This export will be disabled until it is recovered,
796 and then orphan recovery will be completed. */
797 CERROR("error in orphan recovery on OST idx %d/%d: "
798 "rc = %d\n", i, lov->desc.ld_tgt_count, err);
807 #define LOV_CREATE_RESEED_INTERVAL 1000
809 /* the LOV expects oa->o_id to be set to the LOV object id */
810 static int lov_create(struct obd_export *exp, struct obdo *src_oa,
811 struct lov_stripe_md **ea, struct obd_trans_info *oti)
813 static int ost_start_idx, ost_start_count;
815 struct lov_stripe_md *lsm;
816 struct lov_oinfo *loi = NULL;
817 struct obdo *tmp_oa, *ret_oa;
818 struct llog_cookie *cookies = NULL;
819 unsigned ost_count, ost_idx;
820 int set = 0, obj_alloc = 0, cookie_sent = 0, rc = 0, i;
825 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
826 src_oa->o_flags == OBD_FL_DELORPHAN) {
827 rc = lov_clear_orphans(exp, src_oa, ea, oti);
834 lov = &exp->exp_obd->u.lov;
836 if (!lov->desc.ld_active_tgt_count)
839 /* Recreate a specific object id at the given OST index */
840 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
841 (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) {
842 struct lov_stripe_md obj_md;
843 struct lov_stripe_md *obj_mdp = &obj_md;
845 ost_idx = src_oa->o_nlink;
849 if (ost_idx >= lov->desc.ld_tgt_count)
851 for (i = 0; i < lsm->lsm_stripe_count; i++) {
852 if (lsm->lsm_oinfo[i].loi_ost_idx == ost_idx) {
853 if (lsm->lsm_oinfo[i].loi_id != src_oa->o_id ||
854 lsm->lsm_oinfo[i].loi_gr != src_oa->o_gr) {
860 if (i == lsm->lsm_stripe_count)
863 rc = obd_create(lov->tgts[ost_idx].ltd_exp, src_oa,
868 ret_oa = obdo_alloc();
872 tmp_oa = obdo_alloc();
874 GOTO(out_oa, rc = -ENOMEM);
879 ost_count = lov_get_stripecnt(lov, 0);
881 /* If the MDS file was truncated up to some size, stripe over
882 * enough OSTs to allow the file to be created at that size. */
883 if (src_oa->o_valid & OBD_MD_FLSIZE) {
884 stripes=((src_oa->o_size+LUSTRE_STRIPE_MAXBYTES)>>12)-1;
885 do_div(stripes, (__u32)(LUSTRE_STRIPE_MAXBYTES >> 12));
887 if (stripes > lov->desc.ld_active_tgt_count)
889 if (stripes < ost_count)
895 rc = lov_alloc_memmd(&lsm, stripes, lov->desc.ld_pattern ?
896 lov->desc.ld_pattern : LOV_PATTERN_RAID0);
903 ost_count = lov->desc.ld_tgt_count;
905 LASSERT(src_oa->o_gr > 0);
906 LASSERT(src_oa->o_valid & OBD_MD_FLID);
907 lsm->lsm_object_id = src_oa->o_id;
908 lsm->lsm_object_gr = src_oa->o_gr;
909 if (!lsm->lsm_stripe_size)
910 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
911 if (!lsm->lsm_pattern) {
912 lsm->lsm_pattern = lov->desc.ld_pattern ?
913 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
916 if (*ea == NULL || lsm->lsm_oinfo[0].loi_ost_idx >= ost_count) {
917 if (--ost_start_count <= 0) {
918 ost_start_idx = ll_insecure_random_int();
919 ost_start_count = LOV_CREATE_RESEED_INTERVAL;
920 } else if (lsm->lsm_stripe_count >=
921 lov->desc.ld_active_tgt_count) {
922 /* If we allocate from all of the stripes, make the
923 * next file start on the next OST. */
926 ost_idx = ost_start_idx % ost_count;
928 ost_idx = lsm->lsm_oinfo[0].loi_ost_idx;
931 CDEBUG(D_INODE, "allocating %d subobjs for objid "LPX64" at idx %d\n",
932 lsm->lsm_stripe_count, lsm->lsm_object_id, ost_idx);
934 /* XXX LOV STACKING: need to figure out how many real OSCs */
935 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
936 oti_alloc_cookies(oti, lsm->lsm_stripe_count);
937 if (!oti->oti_logcookies)
938 GOTO(out_cleanup, rc = -ENOMEM);
939 cookies = oti->oti_logcookies;
942 loi = lsm->lsm_oinfo;
943 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
944 struct lov_stripe_md obj_md;
945 struct lov_stripe_md *obj_mdp = &obj_md;
949 if (lov->tgts[ost_idx].active == 0) {
950 if (!obd_uuid_empty(&lov->tgts[ost_idx].uuid))
951 CDEBUG(D_HA, "lov idx %d inactive\n", ost_idx);
955 /* create data objects with "parent" OA */
956 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
958 /* XXX When we start creating objects on demand, we need to
959 * make sure that we always create the object on the
960 * stripe which holds the existing file size.
962 if (src_oa->o_valid & OBD_MD_FLSIZE) {
963 if (lov_stripe_offset(lsm, src_oa->o_size, i,
964 &tmp_oa->o_size) < 0 &&
968 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
969 i, tmp_oa->o_size, src_oa->o_size);
973 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
974 err = obd_create(lov->tgts[ost_idx].ltd_exp, tmp_oa, &obj_mdp,
977 if (lov->tgts[ost_idx].active) {
978 CERROR("error creating objid "LPX64" sub-object"
979 " on OST idx %d/%d: rc = %d\n",
980 src_oa->o_id, ost_idx,
981 lsm->lsm_stripe_count, err);
983 CERROR("obd_create returned invalid "
993 oti->oti_objid[ost_idx] = tmp_oa->o_id;
994 loi->loi_id = tmp_oa->o_id;
995 loi->loi_gr = tmp_oa->o_gr;
996 loi->loi_ost_idx = ost_idx;
997 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
998 CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64" at "
999 "idx %d gen %d\n", lsm->lsm_object_id, loi->loi_id,
1000 ost_idx, loi->loi_ost_gen);
1002 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1007 ++oti->oti_logcookies;
1008 if (tmp_oa->o_valid & OBD_MD_FLCOOKIE)
1013 /* If we have allocated enough objects, we are OK */
1014 if (obj_alloc == lsm->lsm_stripe_count)
1015 GOTO(out_done, rc = 0);
1018 if (obj_alloc == 0) {
1021 GOTO(out_cleanup, rc);
1024 /* If we were passed specific striping params, then a failure to
1025 * meet those requirements is an error, since we can't reallocate
1026 * that memory (it might be part of a larger array or something).
1028 * We can only get here if lsm_stripe_count was originally > 1.
1031 CERROR("can't lstripe objid "LPX64": have %u want %u, rc %d\n",
1032 lsm->lsm_object_id, obj_alloc, lsm->lsm_stripe_count,rc);
1035 GOTO(out_cleanup, rc);
1037 struct lov_stripe_md *lsm_new;
1038 /* XXX LOV STACKING call into osc for sizes */
1039 unsigned oldsize, newsize;
1041 if (oti && cookies && cookie_sent) {
1042 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
1043 newsize = obj_alloc * sizeof(*cookies);
1045 oti_alloc_cookies(oti, obj_alloc);
1046 if (oti->oti_logcookies) {
1047 memcpy(oti->oti_logcookies, cookies, newsize);
1048 OBD_FREE(cookies, oldsize);
1049 cookies = oti->oti_logcookies;
1051 CWARN("'leaking' %d bytes\n", oldsize-newsize);
1055 CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
1056 lsm->lsm_object_id, lsm->lsm_stripe_count, obj_alloc);
1057 oldsize = lov_stripe_md_size(lsm->lsm_stripe_count);
1058 newsize = lov_stripe_md_size(obj_alloc);
1059 OBD_ALLOC(lsm_new, newsize);
1060 if (lsm_new != NULL) {
1061 memcpy(lsm_new, lsm, newsize);
1062 lsm_new->lsm_stripe_count = obj_alloc;
1063 OBD_FREE(lsm, oldsize);
1066 CWARN("'leaking' %d bytes\n", oldsize - newsize);
1073 if (src_oa->o_valid & OBD_MD_FLSIZE &&
1074 ret_oa->o_size != src_oa->o_size) {
1075 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
1076 src_oa->o_size, ret_oa->o_size);
1079 ret_oa->o_id = src_oa->o_id;
1080 ret_oa->o_gr = src_oa->o_gr;
1081 ret_oa->o_valid |= OBD_MD_FLGROUP;
1082 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1088 if (oti && cookies) {
1089 oti->oti_logcookies = cookies;
1091 oti_free_cookies(oti);
1092 src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
1094 src_oa->o_valid |= OBD_MD_FLCOOKIE;
1100 while (obj_alloc-- > 0) {
1101 struct obd_export *sub_exp;
1105 sub_exp = lov->tgts[loi->loi_ost_idx].ltd_exp;
1106 /* destroy already created objects here */
1107 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1108 tmp_oa->o_id = loi->loi_id;
1110 err = obd_destroy(sub_exp, tmp_oa, NULL, oti);
1112 CERROR("Failed to uncreate objid "LPX64" subobj "LPX64
1113 " on OST idx %d: rc = %d\n", src_oa->o_id,
1114 loi->loi_id, loi->loi_ost_idx, err);
1117 obd_free_memmd(exp, &lsm);
1121 static int lov_revalidate_policy(struct lov_obd *lov, struct lov_stripe_md *lsm)
1123 static int next_idx = 0;
1124 struct lov_tgt_desc *tgt;
1127 /* XXX - we should do something clever and take lsm
1128 * into account but just do round robin for now. */
1130 /* last_idx must always be less that count because
1131 * ld_tgt_count currently cannot shrink. */
1132 count = lov->desc.ld_tgt_count;
1134 for (i = next_idx, tgt = lov->tgts + i; i < count; i++, tgt++) {
1136 next_idx = (i + 1) % count;
1141 for (i = 0, tgt = lov->tgts; i < next_idx; i++, tgt++) {
1143 next_idx = (i + 1) % count;
1151 #define lsm_bad_magic(LSMP) \
1153 struct lov_stripe_md *_lsm__ = (LSMP); \
1156 CERROR("LOV requires striping ea\n"); \
1158 } else if (_lsm__->lsm_magic != LOV_MAGIC) { \
1159 CERROR("LOV striping magic bad %#x != %#x\n", \
1160 _lsm__->lsm_magic, LOV_MAGIC); \
1166 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
1167 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1170 struct lov_obd *lov;
1171 struct lov_oinfo *loi;
1175 if (lsm_bad_magic(lsm))
1178 if (!exp || !exp->exp_obd)
1181 lov = &exp->exp_obd->u.lov;
1182 loi = lsm->lsm_oinfo;
1183 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1186 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1187 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1188 /* Orphan clean up will (someday) fix this up. */
1189 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
1190 oti->oti_logcookies++;
1194 memcpy(&tmp, oa, sizeof(tmp));
1195 tmp.o_id = loi->loi_id;
1196 err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
1198 if (err && lov->tgts[loi->loi_ost_idx].active) {
1199 CDEBUG(D_INODE, "error: destroying objid "LPX64" subobj "
1200 LPX64" on OST idx %d: rc = %d\n",
1201 oa->o_id, loi->loi_id, loi->loi_ost_idx, err);
1209 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
1210 struct lov_stripe_md *lsm)
1213 struct lov_obd *lov;
1214 struct lov_oinfo *loi;
1215 int i, rc = 0, set = 0;
1218 if (lsm_bad_magic(lsm))
1221 if (!exp || !exp->exp_obd)
1224 lov = &exp->exp_obd->u.lov;
1226 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1227 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1228 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1231 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1232 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1236 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1237 "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
1238 /* create data objects with "parent" OA */
1239 memcpy(&tmp, oa, sizeof(tmp));
1240 tmp.o_id = loi->loi_id;
1242 err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
1245 if (lov->tgts[loi->loi_ost_idx].active) {
1246 CERROR("error: getattr objid "LPX64" subobj "
1247 LPX64" on OST idx %d: rc = %d\n",
1248 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1253 lov_merge_attrs(oa, &tmp, tmp.o_valid, lsm, i, &set);
1261 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
1264 struct lov_getattr_async_args *aa = data;
1265 struct lov_stripe_md *lsm = aa->aa_lsm;
1266 struct obdo *oa = aa->aa_oa;
1267 struct obdo *obdos = aa->aa_obdos;
1268 struct lov_oinfo *loi;
1274 /* NB all stripe requests succeeded to get here */
1276 loi = lsm->lsm_oinfo;
1277 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1278 if (obdos[i].o_valid == 0) /* inactive stripe */
1281 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1286 CERROR ("No stripes had valid attrs\n");
1291 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1295 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
1296 struct lov_stripe_md *lsm,
1297 struct ptlrpc_request_set *rqset)
1300 struct lov_obd *lov;
1301 struct lov_oinfo *loi;
1302 struct lov_getattr_async_args *aa;
1303 int i, rc = 0, set = 0;
1306 if (lsm_bad_magic(lsm))
1309 if (!exp || !exp->exp_obd)
1312 lov = &exp->exp_obd->u.lov;
1314 OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1318 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1319 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1320 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1323 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1324 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1325 /* leaves obdos[i].obd_valid unset */
1329 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at "
1330 "idx %u gen %d\n", oa->o_id, i, loi->loi_id,
1331 loi->loi_ost_idx, loi->loi_ost_gen);
1333 /* create data objects with "parent" OA */
1334 memcpy(&obdos[i], oa, sizeof(obdos[i]));
1335 obdos[i].o_id = loi->loi_id;
1337 err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp,
1338 &obdos[i], NULL, rqset);
1340 CERROR("error: getattr objid "LPX64" subobj "
1341 LPX64" on OST idx %d: rc = %d\n",
1342 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1344 GOTO(out_obdos, rc = err);
1349 GOTO (out_obdos, rc = -EIO);
1351 LASSERT (rqset->set_interpret == NULL);
1352 rqset->set_interpret = lov_getattr_interpret;
1353 LASSERT (sizeof (rqset->set_args) >= sizeof (*aa));
1354 aa = (struct lov_getattr_async_args *)&rqset->set_args;
1357 aa->aa_obdos = obdos;
1362 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1368 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
1369 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1371 struct obdo *tmp_oa, *ret_oa;
1372 struct lov_obd *lov;
1373 struct lov_oinfo *loi;
1374 int rc = 0, i, set = 0;
1377 if (lsm_bad_magic(lsm))
1380 if (!exp || !exp->exp_obd)
1383 /* for now, we only expect time updates here */
1384 LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
1385 OBD_MD_FLATIME | OBD_MD_FLMTIME |
1386 OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
1387 OBD_MD_FLSIZE | OBD_MD_FLGROUP)));
1389 LASSERT(!(src_oa->o_valid & OBD_MD_FLGROUP) || src_oa->o_gr > 0);
1391 ret_oa = obdo_alloc();
1395 tmp_oa = obdo_alloc();
1397 GOTO(out_oa, rc = -ENOMEM);
1399 lov = &exp->exp_obd->u.lov;
1400 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1403 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1404 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1408 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1409 tmp_oa->o_id = loi->loi_id;
1410 LASSERT(!(tmp_oa->o_valid & OBD_MD_FLGROUP) || tmp_oa->o_gr>0);
1412 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1413 if (lov_stripe_offset(lsm, src_oa->o_size, i,
1414 &tmp_oa->o_size) < 0 &&
1418 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1419 i, tmp_oa->o_size, src_oa->o_size);
1422 err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa,
1425 if (lov->tgts[loi->loi_ost_idx].active) {
1426 CERROR("error: setattr objid "LPX64" subobj "
1427 LPX64" on OST idx %d: rc = %d\n",
1428 src_oa->o_id, loi->loi_id,
1429 loi->loi_ost_idx, err);
1435 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set);
1440 ret_oa->o_id = src_oa->o_id;
1441 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1450 /* we have an offset in file backed by an lov and want to find out where
1451 * that offset lands in our given stripe of the file. for the easy
1452 * case where the offset is within the stripe, we just have to scale the
1453 * offset down to make it relative to the stripe instead of the lov.
1455 * the harder case is what to do when the offset doesn't intersect the
1456 * stripe. callers will want start offsets clamped ahead to the start
1457 * of the nearest stripe in the file. end offsets similarly clamped to the
1458 * nearest ending byte of a stripe in the file:
1460 * all this function does is move offsets to the nearest region of the
1461 * stripe, and it does its work "mod" the full length of all the stripes.
1462 * consider a file with 3 stripes:
1465 * ---------------------------------------------------------------------
1466 * | 0 | 1 | 2 | 0 | 1 | 2 |
1467 * ---------------------------------------------------------------------
1469 * to find stripe 1's offsets for S and E, it divides by the full stripe
1470 * width and does its math in the context of a single set of stripes:
1473 * -----------------------------------
1475 * -----------------------------------
1477 * it'll notice that E is outside stripe 1 and clamp it to the end of the
1478 * stripe, then multiply it back out by lov_off to give the real offsets in
1482 * ---------------------------------------------------------------------
1483 * | 1 | 1 | 1 | 1 | 1 | 1 |
1484 * ---------------------------------------------------------------------
1486 * it would have done similarly and pulled S forward to the start of a 1
1487 * stripe if, say, S had landed in a 0 stripe.
1489 * this rounding isn't always correct. consider an E lov offset that lands
1490 * on a 0 stripe, the "mod stripe width" math will pull it forward to the
1491 * start of a 1 stripe, when in fact it wanted to be rounded back to the end
1492 * of a previous 1 stripe. this logic is handled by callers and this is why:
1494 * this function returns < 0 when the offset was "before" the stripe and
1495 * was moved forward to the start of the stripe in question; 0 when it
1496 * falls in the stripe and no shifting was done; > 0 when the offset
1497 * was outside the stripe and was pulled back to its final byte. */
1498 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
1499 int stripeno, obd_off *obd_off)
1501 unsigned long ssize = lsm->lsm_stripe_size;
1502 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1503 unsigned long stripe_off, this_stripe;
1506 if (lov_off == OBD_OBJECT_EOF) {
1507 *obd_off = OBD_OBJECT_EOF;
1511 /* do_div(a, b) returns a % b, and a = a / b */
1512 stripe_off = do_div(lov_off, swidth);
1514 this_stripe = stripeno * ssize;
1515 if (stripe_off < this_stripe) {
1519 stripe_off -= this_stripe;
1521 if (stripe_off >= ssize) {
1527 *obd_off = lov_off * ssize + stripe_off;
1531 /* Given a whole-file size and a stripe number, give the file size which
1532 * corresponds to the individual object of that stripe.
1534 * This behaves basically in the same was as lov_stripe_offset, except that
1535 * file sizes falling before the beginning of a stripe are clamped to the end
1536 * of the previous stripe, not the beginning of the next:
1539 * ---------------------------------------------------------------------
1540 * | 0 | 1 | 2 | 0 | 1 | 2 |
1541 * ---------------------------------------------------------------------
1543 * if clamped to stripe 2 becomes:
1546 * ---------------------------------------------------------------------
1547 * | 0 | 1 | 2 | 0 | 1 | 2 |
1548 * ---------------------------------------------------------------------
1550 static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size,
1553 unsigned long ssize = lsm->lsm_stripe_size;
1554 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1555 unsigned long stripe_off, this_stripe;
1557 if (file_size == OBD_OBJECT_EOF)
1558 return OBD_OBJECT_EOF;
1560 /* do_div(a, b) returns a % b, and a = a / b */
1561 stripe_off = do_div(file_size, swidth);
1563 this_stripe = stripeno * ssize;
1564 if (stripe_off < this_stripe) {
1565 /* Move to end of previous stripe, or zero */
1566 if (file_size > 0) {
1573 stripe_off -= this_stripe;
1575 if (stripe_off >= ssize) {
1576 /* Clamp to end of this stripe */
1581 return (file_size * ssize + stripe_off);
1584 /* given an extent in an lov and a stripe, calculate the extent of the stripe
1585 * that is contained within the lov extent. this returns true if the given
1586 * stripe does intersect with the lov extent. */
1587 static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno,
1588 obd_off start, obd_off end,
1589 obd_off *obd_start, obd_off *obd_end)
1591 int start_side = 0, end_side = 0;
1593 switch (lsm->lsm_pattern) {
1594 case LOV_PATTERN_RAID0:
1595 start_side = lov_stripe_offset(lsm, start, stripeno, obd_start);
1596 end_side = lov_stripe_offset(lsm, end, stripeno, obd_end);
1598 case LOV_PATTERN_CMOBD:
1601 start_side = end_side = 0;
1607 CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n",
1608 start, end, start_side, *obd_start, *obd_end, end_side);
1610 /* this stripe doesn't intersect the file extent when neither
1611 * start or the end intersected the stripe and obd_start and
1612 * obd_end got rounded up to the save value. */
1613 if (start_side != 0 && end_side != 0 && *obd_start == *obd_end)
1616 /* as mentioned in the lov_stripe_offset commentary, end
1617 * might have been shifted in the wrong direction. This
1618 * happens when an end offset is before the stripe when viewed
1619 * through the "mod stripe size" math. we detect it being shifted
1620 * in the wrong direction and touch it up.
1621 * interestingly, this can't underflow since end must be > start
1622 * if we passed through the previous check.
1623 * (should we assert for that somewhere?) */
1630 /* compute which stripe number "lov_off" will be written into */
1631 static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off)
1633 unsigned long ssize = lsm->lsm_stripe_size;
1634 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1635 unsigned long stripe_off;
1637 if (lsm->lsm_pattern == LOV_PATTERN_CMOBD)
1640 stripe_off = do_div(lov_off, swidth);
1642 return stripe_off / ssize;
1645 static int lov_revalidate_md(struct obd_export *exp, struct obdo *src_oa,
1646 struct lov_stripe_md *ea,
1647 struct obd_trans_info *oti)
1649 struct obd_export *osc_exp;
1650 struct lov_obd *lov = &exp->exp_obd->u.lov;
1651 struct lov_stripe_md *lsm = ea;
1652 struct lov_stripe_md obj_md;
1653 struct lov_stripe_md *obj_mdp = &obj_md;
1654 struct lov_oinfo *loi;
1655 struct obdo *tmp_oa;
1656 int ost_idx, updates = 0, i;
1659 tmp_oa = obdo_alloc();
1663 loi = lsm->lsm_oinfo;
1664 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1666 if (!obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
1669 ost_idx = lov_revalidate_policy(lov, lsm);
1671 /* FIXME: punt for now. */
1672 CERROR("lov_revalidate_policy failed; no active "
1677 /* create a new object */
1678 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1679 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
1680 osc_exp = lov->tgts[ost_idx].ltd_exp;
1681 rc = obd_create(osc_exp, tmp_oa, &obj_mdp, oti);
1683 CERROR("error creating new subobj at idx %d; "
1684 "rc = %d\n", ost_idx, rc);
1688 oti->oti_objid[ost_idx] = tmp_oa->o_id;
1689 loi->loi_id = tmp_oa->o_id;
1690 loi->loi_ost_idx = ost_idx;
1691 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
1692 CDEBUG(D_INODE, "replacing objid "LPX64" subobj "LPX64
1693 " with idx %d gen %d.\n", lsm->lsm_object_id,
1694 loi->loi_id, ost_idx, loi->loi_ost_gen);
1698 /* If we got an error revalidating an entry there's no need to
1699 * cleanup up objects we allocated here because the bad entry
1700 * still points to a deleted OST. */
1706 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1707 * we can send this 'punch' to just the authoritative node and the nodes
1708 * that the punch will affect. */
1709 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1710 struct lov_stripe_md *lsm,
1711 obd_off start, obd_off end, struct obd_trans_info *oti)
1714 struct lov_obd *lov;
1715 struct lov_oinfo *loi;
1719 if (lsm_bad_magic(lsm))
1722 if (!exp || !exp->exp_obd)
1725 lov = &exp->exp_obd->u.lov;
1726 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1727 obd_off starti, endi;
1730 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1731 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1735 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1738 /* create data objects with "parent" OA */
1739 memcpy(&tmp, oa, sizeof(tmp));
1740 tmp.o_id = loi->loi_id;
1742 err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, NULL,
1743 starti, endi, NULL);
1745 if (lov->tgts[loi->loi_ost_idx].active) {
1746 CERROR("error: punch objid "LPX64" subobj "LPX64
1747 " on OST idx %d: rc = %d\n", oa->o_id,
1748 loi->loi_id, loi->loi_ost_idx, err);
1753 loi->loi_kms = loi->loi_rss = starti;
1759 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1760 struct lov_stripe_md *lsm, obd_off start, obd_off end)
1763 struct lov_obd *lov;
1764 struct lov_oinfo *loi;
1768 if (lsm_bad_magic(lsm))
1778 lov = &exp->exp_obd->u.lov;
1779 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1780 obd_off starti, endi;
1783 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1784 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1788 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1791 memcpy(tmp, oa, sizeof(*tmp));
1792 tmp->o_id = loi->loi_id;
1794 err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL,
1797 if (lov->tgts[loi->loi_ost_idx].active) {
1798 CERROR("error: fsync objid "LPX64" subobj "LPX64
1799 " on OST idx %d: rc = %d\n", oa->o_id,
1800 loi->loi_id, loi->loi_ost_idx, err);
1811 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1812 struct lov_stripe_md *lsm,
1813 obd_count oa_bufs, struct brw_page *pga)
1818 /* The caller just wants to know if there's a chance that this
1819 * I/O can succeed */
1820 for (i = 0; i < oa_bufs; i++) {
1821 int stripe = lov_stripe_number(lsm, pga[i].disk_offset);
1822 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1825 if (!lov_stripe_intersects(lsm, i, pga[i].disk_offset,
1826 pga[i].disk_offset + pga[i].count,
1830 if (lov->tgts[ost].active == 0) {
1831 CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1834 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[ost].ltd_exp, oa,
1835 NULL, 1, &pga[i], NULL);
1842 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1843 struct lov_stripe_md *lsm, obd_count oa_bufs,
1844 struct brw_page *pga, struct obd_trans_info *oti)
1850 struct lov_stripe_md lsm;
1852 } *stripeinfo, *si, *si_last;
1853 struct obdo *ret_oa = NULL, *tmp_oa = NULL;
1854 struct lov_obd *lov;
1855 struct brw_page *ioarr;
1856 struct lov_oinfo *loi;
1857 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0;
1860 if (lsm_bad_magic(lsm))
1863 lov = &exp->exp_obd->u.lov;
1865 if (cmd == OBD_BRW_CHECK) {
1866 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1870 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1874 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1876 GOTO(out_sinfo, rc = -ENOMEM);
1878 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1880 GOTO(out_where, rc = -ENOMEM);
1883 ret_oa = obdo_alloc();
1885 GOTO(out_ioarr, rc = -ENOMEM);
1887 tmp_oa = obdo_alloc();
1889 GOTO(out_oa, rc = -ENOMEM);
1892 for (i = 0; i < oa_bufs; i++) {
1893 where[i] = lov_stripe_number(lsm, pga[i].disk_offset);
1894 stripeinfo[where[i]].bufct++;
1897 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1898 i < stripe_count; i++, loi++, si_last = si, si++) {
1900 si->index = si_last->index + si_last->bufct;
1901 si->lsm.lsm_object_id = loi->loi_id;
1902 si->lsm.lsm_object_gr = lsm->lsm_object_gr;
1903 si->ost_idx = loi->loi_ost_idx;
1906 for (i = 0; i < oa_bufs; i++) {
1907 int which = where[i];
1910 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1911 LASSERT(shift < oa_bufs);
1912 ioarr[shift] = pga[i];
1913 lov_stripe_offset(lsm, pga[i].disk_offset, which,
1914 &ioarr[shift].disk_offset);
1915 stripeinfo[which].subcount++;
1918 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1919 int shift = si->index;
1921 if (lov->tgts[si->ost_idx].active == 0) {
1922 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1923 GOTO(out_oa, rc = -EIO);
1927 LASSERT(shift < oa_bufs);
1929 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1931 tmp_oa->o_id = si->lsm.lsm_object_id;
1932 rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp,
1933 tmp_oa, &si->lsm, si->bufct,
1934 &ioarr[shift], oti);
1938 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1943 ret_oa->o_id = src_oa->o_id;
1944 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1953 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1955 OBD_FREE(where, sizeof(*where) * oa_bufs);
1957 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1961 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1964 struct lov_brw_async_args *aa = data;
1965 struct lov_stripe_md *lsm = aa->aa_lsm;
1966 obd_count oa_bufs = aa->aa_oa_bufs;
1967 struct obdo *oa = aa->aa_oa;
1968 struct obdo *obdos = aa->aa_obdos;
1969 struct brw_page *ioarr = aa->aa_ioarr;
1970 struct lov_oinfo *loi;
1975 /* NB all stripe requests succeeded to get here */
1977 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1979 if (obdos[i].o_valid == 0) /* inactive stripe */
1982 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1987 CERROR("No stripes had valid attrs\n");
1991 oa->o_id = lsm->lsm_object_id;
1993 OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos));
1994 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1998 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1999 struct lov_stripe_md *lsm, obd_count oa_bufs,
2000 struct brw_page *pga, struct ptlrpc_request_set *set,
2001 struct obd_trans_info *oti)
2007 struct lov_stripe_md lsm;
2009 } *stripeinfo, *si, *si_last;
2010 struct lov_obd *lov;
2011 struct brw_page *ioarr;
2012 struct obdo *obdos = NULL;
2013 struct lov_oinfo *loi;
2014 struct lov_brw_async_args *aa;
2015 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count;
2018 if (lsm_bad_magic(lsm))
2021 lov = &exp->exp_obd->u.lov;
2023 if (cmd == OBD_BRW_CHECK) {
2024 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
2028 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
2032 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
2034 GOTO(out_sinfo, rc = -ENOMEM);
2037 OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count);
2039 GOTO(out_where, rc = -ENOMEM);
2042 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
2044 GOTO(out_obdos, rc = -ENOMEM);
2046 for (i = 0; i < oa_bufs; i++) {
2047 where[i] = lov_stripe_number(lsm, pga[i].disk_offset);
2048 stripeinfo[where[i]].bufct++;
2051 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
2052 i < stripe_count; i++, loi++, si_last = si, si++) {
2054 si->index = si_last->index + si_last->bufct;
2055 si->lsm.lsm_object_id = loi->loi_id;
2056 si->ost_idx = loi->loi_ost_idx;
2059 memcpy(&obdos[i], oa, sizeof(*obdos));
2060 obdos[i].o_id = si->lsm.lsm_object_id;
2064 for (i = 0; i < oa_bufs; i++) {
2065 int which = where[i];
2068 shift = stripeinfo[which].index + stripeinfo[which].subcount;
2069 LASSERT(shift < oa_bufs);
2070 ioarr[shift] = pga[i];
2071 lov_stripe_offset(lsm, pga[i].disk_offset, which,
2072 &ioarr[shift].disk_offset);
2073 stripeinfo[which].subcount++;
2076 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
2077 int shift = si->index;
2082 if (lov->tgts[si->ost_idx].active == 0) {
2083 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
2084 GOTO(out_ioarr, rc = -EIO);
2087 LASSERT(shift < oa_bufs);
2089 rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp,
2090 &obdos[i], &si->lsm, si->bufct,
2091 &ioarr[shift], set, oti);
2093 GOTO(out_ioarr, rc);
2096 LASSERT(set->set_interpret == NULL);
2097 set->set_interpret = (set_interpreter_func)lov_brw_interpret;
2098 LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args));
2099 aa = (struct lov_brw_async_args *)&set->set_args;
2101 aa->aa_obdos = obdos;
2103 aa->aa_ioarr = ioarr;
2104 aa->aa_oa_bufs = oa_bufs;
2106 /* Don't free ioarr or obdos - that's done in lov_brw_interpret */
2107 GOTO(out_where, rc);
2110 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
2112 OBD_FREE(obdos, stripe_count * sizeof(*obdos));
2114 OBD_FREE(where, sizeof(*where) * oa_bufs);
2116 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
2120 struct lov_async_page *lap_from_cookie(void *cookie)
2122 struct lov_async_page *lap = cookie;
2123 if (lap->lap_magic != LAP_MAGIC)
2124 return ERR_PTR(-EINVAL);
2128 static int lov_ap_make_ready(void *data, int cmd)
2130 struct lov_async_page *lap = lap_from_cookie(data);
2131 /* XXX should these assert? */
2135 return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
2137 static int lov_ap_refresh_count(void *data, int cmd)
2139 struct lov_async_page *lap = lap_from_cookie(data);
2143 return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
2146 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
2148 struct lov_async_page *lap = lap_from_cookie(data);
2149 /* XXX should these assert? */
2153 lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
2154 /* XXX woah, shouldn't we be altering more here? size? */
2155 oa->o_id = lap->lap_loi_id;
2158 static void lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
2160 struct lov_async_page *lap = lap_from_cookie(data);
2164 /* in a raid1 regime this would down a count of many ios
2165 * in flight, onl calling the caller_ops completion when all
2166 * the raid1 ios are complete */
2167 lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, oa, rc);
2170 static struct obd_async_page_ops lov_async_page_ops = {
2171 .ap_make_ready = lov_ap_make_ready,
2172 .ap_refresh_count = lov_ap_refresh_count,
2173 .ap_fill_obdo = lov_ap_fill_obdo,
2174 .ap_completion = lov_ap_completion,
2177 static int lov_prep_async_page(struct obd_export *exp,
2178 struct lov_stripe_md *lsm,
2179 struct lov_oinfo *loi, struct page *page,
2180 obd_off offset, struct obd_async_page_ops *ops,
2181 void *data, void **res)
2183 struct lov_obd *lov = &exp->exp_obd->u.lov;
2184 struct lov_async_page *lap;
2188 if (lsm_bad_magic(lsm))
2190 LASSERT(loi == NULL);
2192 stripe = lov_stripe_number(lsm, offset);
2193 loi = &lsm->lsm_oinfo[stripe];
2195 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
2197 if (lov->tgts[loi->loi_ost_idx].active == 0)
2199 if (lov->tgts[loi->loi_ost_idx].ltd_exp == NULL) {
2200 CERROR("ltd_exp == NULL, but OST idx %d doesn't appear to be "
2201 "deleted or inactive.\n", loi->loi_ost_idx);
2205 OBD_ALLOC(lap, sizeof(*lap));
2209 lap->lap_magic = LAP_MAGIC;
2210 lap->lap_caller_ops = ops;
2211 lap->lap_caller_data = data;
2213 /* FIXME handle multiple oscs after landing b_raid1 */
2214 lap->lap_stripe = stripe;
2215 switch (lsm->lsm_pattern) {
2216 case LOV_PATTERN_RAID0:
2217 lov_stripe_offset(lsm, offset, lap->lap_stripe,
2218 &lap->lap_sub_offset);
2220 case LOV_PATTERN_CMOBD:
2221 lap->lap_sub_offset = offset;
2227 /* so the callback doesn't need the lsm */
2228 lap->lap_loi_id = loi->loi_id;
2230 rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2231 lsm, loi, page, lap->lap_sub_offset,
2232 &lov_async_page_ops, lap,
2233 &lap->lap_sub_cookie);
2235 OBD_FREE(lap, sizeof(*lap));
2238 CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
2239 lap->lap_sub_cookie, offset);
2244 static int lov_queue_async_io(struct obd_export *exp,
2245 struct lov_stripe_md *lsm,
2246 struct lov_oinfo *loi, void *cookie,
2247 int cmd, obd_off off, int count,
2248 obd_flag brw_flags, obd_flag async_flags)
2250 struct lov_obd *lov = &exp->exp_obd->u.lov;
2251 struct lov_async_page *lap;
2254 LASSERT(loi == NULL);
2256 if (lsm_bad_magic(lsm))
2259 lap = lap_from_cookie(cookie);
2261 RETURN(PTR_ERR(lap));
2263 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2265 rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
2266 loi, lap->lap_sub_cookie, cmd, off, count,
2267 brw_flags, async_flags);
2271 static int lov_set_async_flags(struct obd_export *exp,
2272 struct lov_stripe_md *lsm,
2273 struct lov_oinfo *loi, void *cookie,
2274 obd_flag async_flags)
2276 struct lov_obd *lov = &exp->exp_obd->u.lov;
2277 struct lov_async_page *lap;
2280 LASSERT(loi == NULL);
2282 if (lsm_bad_magic(lsm))
2285 lap = lap_from_cookie(cookie);
2287 RETURN(PTR_ERR(lap));
2289 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2291 rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
2292 lsm, loi, lap->lap_sub_cookie, async_flags);
2296 static int lov_queue_group_io(struct obd_export *exp,
2297 struct lov_stripe_md *lsm,
2298 struct lov_oinfo *loi,
2299 struct obd_io_group *oig, void *cookie,
2300 int cmd, obd_off off, int count,
2301 obd_flag brw_flags, obd_flag async_flags)
2303 struct lov_obd *lov = &exp->exp_obd->u.lov;
2304 struct lov_async_page *lap;
2307 LASSERT(loi == NULL);
2309 if (lsm_bad_magic(lsm))
2312 lap = lap_from_cookie(cookie);
2314 RETURN(PTR_ERR(lap));
2316 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2318 rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
2319 oig, lap->lap_sub_cookie, cmd, off, count,
2320 brw_flags, async_flags);
2324 /* this isn't exactly optimal. we may have queued sync io in oscs on
2325 * all stripes, but we don't record that fact at queue time. so we
2326 * trigger sync io on all stripes. */
2327 static int lov_trigger_group_io(struct obd_export *exp,
2328 struct lov_stripe_md *lsm,
2329 struct lov_oinfo *loi,
2330 struct obd_io_group *oig)
2332 struct lov_obd *lov = &exp->exp_obd->u.lov;
2335 LASSERT(loi == NULL);
2337 if (lsm_bad_magic(lsm))
2340 loi = lsm->lsm_oinfo;
2341 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
2342 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2343 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2347 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
2349 if (rc == 0 && err != 0)
2355 static int lov_teardown_async_page(struct obd_export *exp,
2356 struct lov_stripe_md *lsm,
2357 struct lov_oinfo *loi, void *cookie)
2359 struct lov_obd *lov = &exp->exp_obd->u.lov;
2360 struct lov_async_page *lap;
2363 LASSERT(loi == NULL);
2365 if (lsm_bad_magic(lsm))
2368 lap = lap_from_cookie(cookie);
2370 RETURN(PTR_ERR(lap));
2372 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2374 rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2375 lsm, loi, lap->lap_sub_cookie);
2377 CERROR("unable to teardown sub cookie %p: %d\n",
2378 lap->lap_sub_cookie, rc);
2381 OBD_FREE(lap, sizeof(*lap));
2385 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2386 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2387 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2388 void *data,__u32 lvb_len, void *lvb_swabber,
2389 struct lustre_handle *lockh)
2391 struct lov_lock_handles *lov_lockh = NULL;
2392 struct lustre_handle *lov_lockhp;
2393 struct lov_obd *lov;
2394 struct lov_oinfo *loi;
2395 char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)];
2396 struct lov_stripe_md *submd = (void *)submd_buf;
2398 int i, save_flags = *flags;
2401 if (lsm_bad_magic(lsm))
2404 /* we should never be asked to replay a lock this way. */
2405 LASSERT((*flags & LDLM_FL_REPLAY) == 0);
2407 if (!exp || !exp->exp_obd)
2410 if (lsm->lsm_stripe_count > 1) {
2411 lov_lockh = lov_llh_new(lsm);
2412 if (lov_lockh == NULL)
2415 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2416 lov_lockhp = lov_lockh->llh_handles;
2421 lov = &exp->exp_obd->u.lov;
2422 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2423 i++, loi++, lov_lockhp++) {
2424 ldlm_policy_data_t sub_ext;
2427 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2428 policy->l_extent.end, &start,
2432 sub_ext.l_extent.start = start;
2433 sub_ext.l_extent.end = end;
2434 sub_ext.l_extent.gid = policy->l_extent.gid;
2436 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2437 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2441 /* XXX LOV STACKING: submd should be from the subobj */
2442 submd->lsm_object_id = loi->loi_id;
2443 submd->lsm_object_gr = lsm->lsm_object_gr;
2444 submd->lsm_stripe_count = 0;
2445 submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid;
2446 submd->lsm_oinfo->loi_rss = loi->loi_rss;
2447 submd->lsm_oinfo->loi_kms = loi->loi_kms;
2448 submd->lsm_oinfo->loi_blocks = loi->loi_blocks;
2449 loi->loi_mtime = submd->lsm_oinfo->loi_mtime;
2450 /* XXX submd is not fully initialized here */
2451 *flags = save_flags;
2452 rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd,
2453 type, &sub_ext, mode, flags, bl_cb, cp_cb,
2454 gl_cb, data, lvb_len, lvb_swabber, lov_lockhp);
2456 /* XXX FIXME: This unpleasantness doesn't belong here at *all*.
2457 * It belongs in the OSC, except that the OSC doesn't have
2458 * access to the real LOI -- it gets a copy, that we created
2459 * above, and that copy can be arbitrarily out of date.
2461 * The LOV API is due for a serious rewriting anyways, and this
2462 * can be addressed then. */
2463 if (rc == ELDLM_OK) {
2464 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2465 __u64 tmp = submd->lsm_oinfo->loi_rss;
2467 LASSERT(lock != NULL);
2469 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2470 /* Extend KMS up to the end of this lock and no further
2471 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2472 if (tmp > lock->l_policy_data.l_extent.end)
2473 tmp = lock->l_policy_data.l_extent.end + 1;
2474 if (tmp >= loi->loi_kms) {
2475 CDEBUG(D_INODE, "lock acquired, setting rss="
2476 LPU64", kms="LPU64"\n", loi->loi_rss,
2479 loi->loi_kms_valid = 1;
2481 CDEBUG(D_INODE, "lock acquired, setting rss="
2482 LPU64"; leaving kms="LPU64", end="LPU64
2483 "\n", loi->loi_rss, loi->loi_kms,
2484 lock->l_policy_data.l_extent.end);
2486 ldlm_lock_allow_match(lock);
2487 LDLM_LOCK_PUT(lock);
2488 } else if (rc == ELDLM_LOCK_ABORTED &&
2489 save_flags & LDLM_FL_HAS_INTENT) {
2490 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2491 loi->loi_rss = submd->lsm_oinfo->loi_rss;
2492 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2493 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2494 " kms="LPU64"\n", loi->loi_rss, loi->loi_kms);
2496 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2497 if (lov->tgts[loi->loi_ost_idx].active) {
2498 CERROR("error: enqueue objid "LPX64" subobj "
2499 LPX64" on OST idx %d: rc = %d\n",
2500 lsm->lsm_object_id, loi->loi_id,
2501 loi->loi_ost_idx, rc);
2502 GOTO(out_locks, rc);
2506 if (lsm->lsm_stripe_count > 1)
2507 lov_llh_put(lov_lockh);
2511 while (loi--, lov_lockhp--, i-- > 0) {
2512 struct lov_stripe_md submd;
2515 if (lov_lockhp->cookie == 0)
2518 /* XXX LOV STACKING: submd should be from the subobj */
2519 submd.lsm_object_id = loi->loi_id;
2520 submd.lsm_object_gr = lsm->lsm_object_gr;
2521 submd.lsm_stripe_count = 0;
2522 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2524 if (err && lov->tgts[loi->loi_ost_idx].active) {
2525 CERROR("error: cancelling objid "LPX64" on OST "
2526 "idx %d after enqueue error: rc = %d\n",
2527 loi->loi_id, loi->loi_ost_idx, err);
2531 if (lsm->lsm_stripe_count > 1) {
2532 lov_llh_destroy(lov_lockh);
2533 lov_llh_put(lov_lockh);
2538 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2539 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2540 int *flags, void *data, struct lustre_handle *lockh)
2542 struct lov_lock_handles *lov_lockh = NULL;
2543 struct lustre_handle *lov_lockhp;
2544 struct lov_obd *lov;
2545 struct lov_oinfo *loi;
2546 struct lov_stripe_md submd;
2547 ldlm_error_t rc = 0;
2551 if (lsm_bad_magic(lsm))
2554 if (!exp || !exp->exp_obd)
2557 if (lsm->lsm_stripe_count > 1) {
2558 lov_lockh = lov_llh_new(lsm);
2559 if (lov_lockh == NULL)
2562 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2563 lov_lockhp = lov_lockh->llh_handles;
2568 lov = &exp->exp_obd->u.lov;
2569 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2570 i++, loi++, lov_lockhp++) {
2571 ldlm_policy_data_t sub_ext;
2575 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2576 policy->l_extent.end, &start, &end))
2579 sub_ext.l_extent.start = start;
2580 sub_ext.l_extent.end = end;
2582 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid)) {
2583 CDEBUG(D_HA, "lov idx %d deleted\n", loi->loi_ost_idx);
2586 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2587 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2592 /* XXX LOV STACKING: submd should be from the subobj */
2593 submd.lsm_object_id = loi->loi_id;
2594 submd.lsm_object_gr = lsm->lsm_object_gr;
2595 submd.lsm_stripe_count = 0;
2597 /* XXX submd is not fully initialized here */
2598 rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2599 type, &sub_ext, mode, &lov_flags, data,
2605 if (lsm->lsm_stripe_count > 1) {
2606 if (*flags & LDLM_FL_TEST_LOCK)
2607 lov_llh_destroy(lov_lockh);
2608 lov_llh_put(lov_lockh);
2613 while (loi--, lov_lockhp--, i-- > 0) {
2614 struct lov_stripe_md submd;
2617 if (lov_lockhp->cookie == 0)
2620 /* XXX LOV STACKING: submd should be from the subobj */
2621 submd.lsm_object_id = loi->loi_id;
2622 submd.lsm_object_gr = lsm->lsm_object_gr;
2623 submd.lsm_stripe_count = 0;
2624 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2626 if (err && lov->tgts[loi->loi_ost_idx].active) {
2627 CERROR("error: cancelling objid "LPX64" on OST "
2628 "idx %d after match failure: rc = %d\n",
2629 loi->loi_id, loi->loi_ost_idx, err);
2633 if (lsm->lsm_stripe_count > 1) {
2634 lov_llh_destroy(lov_lockh);
2635 lov_llh_put(lov_lockh);
2640 static int lov_change_cbdata(struct obd_export *exp,
2641 struct lov_stripe_md *lsm, ldlm_iterator_t it,
2644 struct lov_obd *lov;
2645 struct lov_oinfo *loi;
2649 if (lsm_bad_magic(lsm))
2652 if (!exp || !exp->exp_obd)
2655 LASSERT(lsm->lsm_object_gr > 0);
2657 lov = &exp->exp_obd->u.lov;
2658 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2659 struct lov_stripe_md submd;
2660 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2661 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2665 submd.lsm_object_id = loi->loi_id;
2666 submd.lsm_object_gr = lsm->lsm_object_gr;
2667 submd.lsm_stripe_count = 0;
2668 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
2674 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
2675 __u32 mode, struct lustre_handle *lockh)
2677 struct lov_lock_handles *lov_lockh = NULL;
2678 struct lustre_handle *lov_lockhp;
2679 struct lov_obd *lov;
2680 struct lov_oinfo *loi;
2684 if (lsm_bad_magic(lsm))
2687 if (!exp || !exp->exp_obd)
2690 LASSERT(lsm->lsm_object_gr > 0);
2693 if (lsm->lsm_stripe_count > 1) {
2694 lov_lockh = lov_handle2llh(lockh);
2696 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2700 lov_lockhp = lov_lockh->llh_handles;
2705 lov = &exp->exp_obd->u.lov;
2706 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2707 i++, loi++, lov_lockhp++) {
2708 struct lov_stripe_md submd;
2711 if (lov_lockhp->cookie == 0) {
2712 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2713 loi->loi_ost_idx, loi->loi_id);
2717 /* XXX LOV STACKING: submd should be from the subobj */
2718 submd.lsm_object_id = loi->loi_id;
2719 submd.lsm_object_gr = lsm->lsm_object_gr;
2720 submd.lsm_stripe_count = 0;
2721 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2724 if (lov->tgts[loi->loi_ost_idx].active) {
2725 CERROR("error: cancel objid "LPX64" subobj "
2726 LPX64" on OST idx %d: rc = %d\n",
2728 loi->loi_id, loi->loi_ost_idx, err);
2735 if (lsm->lsm_stripe_count > 1)
2736 lov_llh_destroy(lov_lockh);
2737 if (lov_lockh != NULL)
2738 lov_llh_put(lov_lockh);
2742 static int lov_cancel_unused(struct obd_export *exp,
2743 struct lov_stripe_md *lsm, int flags, void *opaque)
2745 struct lov_obd *lov;
2746 struct lov_oinfo *loi;
2750 lov = &exp->exp_obd->u.lov;
2752 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2753 int err = obd_cancel_unused(lov->tgts[i].ltd_exp, NULL,
2761 if (lsm_bad_magic(lsm))
2764 if (!exp || !exp->exp_obd)
2767 LASSERT(lsm->lsm_object_gr > 0);
2769 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2770 struct lov_stripe_md submd;
2773 if (lov->tgts[loi->loi_ost_idx].active == 0)
2774 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2776 submd.lsm_object_id = loi->loi_id;
2777 submd.lsm_object_gr = lsm->lsm_object_gr;
2778 submd.lsm_stripe_count = 0;
2779 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
2780 &submd, flags, opaque);
2781 if (err && lov->tgts[loi->loi_ost_idx].active) {
2782 CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2783 " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2784 loi->loi_id, loi->loi_ost_idx, err);
2792 #define LOV_U64_MAX ((__u64)~0ULL)
2793 #define LOV_SUM_MAX(tot, add) \
2795 if ((tot) + (add) < (tot)) \
2796 (tot) = LOV_U64_MAX; \
2801 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2802 unsigned long max_age)
2804 struct lov_obd *lov = &obd->u.lov;
2805 struct obd_statfs lov_sfs;
2812 /* We only get block data from the OBD */
2813 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2815 if (!lov->tgts[i].active) {
2816 CDEBUG(D_HA, "lov idx %d inactive\n", i);
2820 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
2823 if (lov->tgts[i].active && !rc)
2829 memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
2832 osfs->os_bfree += lov_sfs.os_bfree;
2833 osfs->os_bavail += lov_sfs.os_bavail;
2834 osfs->os_blocks += lov_sfs.os_blocks;
2835 /* XXX not sure about this one - depends on policy.
2836 * - could be minimum if we always stripe on all OBDs
2837 * (but that would be wrong for any other policy,
2838 * if one of the OBDs has no more objects left)
2839 * - could be sum if we stripe whole objects
2840 * - could be average, just to give a nice number
2842 * To give a "reasonable" (if not wholly accurate)
2843 * number, we divide the total number of free objects
2844 * by expected stripe count (watch out for overflow).
2846 LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
2847 LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
2852 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
2853 lov->desc.ld_default_stripe_count :
2854 lov->desc.ld_active_tgt_count;
2856 if (osfs->os_files != LOV_U64_MAX)
2857 do_div(osfs->os_files, expected_stripes);
2858 if (osfs->os_ffree != LOV_U64_MAX)
2859 do_div(osfs->os_ffree, expected_stripes);
2866 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2867 void *karg, void *uarg)
2869 struct obd_device *obddev = class_exp2obd(exp);
2870 struct lov_obd *lov = &obddev->u.lov;
2871 int i, rc, count = lov->desc.ld_tgt_count;
2872 struct obd_uuid *uuidp;
2876 case OBD_IOC_LOV_GET_CONFIG: {
2877 struct obd_ioctl_data *data = karg;
2878 struct lov_tgt_desc *tgtdesc;
2879 struct lov_desc *desc;
2885 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2888 data = (struct obd_ioctl_data *)buf;
2890 if (sizeof(*desc) > data->ioc_inllen1) {
2891 obd_ioctl_freedata(buf, len);
2895 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2896 obd_ioctl_freedata(buf, len);
2900 if (sizeof(__u32) * count > data->ioc_inllen3) {
2901 obd_ioctl_freedata(buf, len);
2905 desc = (struct lov_desc *)data->ioc_inlbuf1;
2906 memcpy(desc, &(lov->desc), sizeof(*desc));
2908 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2909 genp = (__u32 *)data->ioc_inlbuf3;
2910 tgtdesc = lov->tgts;
2911 /* the uuid will be empty for deleted OSTs */
2912 for (i = 0; i < count; i++, uuidp++, genp++, tgtdesc++) {
2913 obd_str2uuid(uuidp, tgtdesc->uuid.uuid);
2914 *genp = tgtdesc->ltd_gen;
2917 rc = copy_to_user((void *)uarg, buf, len);
2920 obd_ioctl_freedata(buf, len);
2923 case LL_IOC_LOV_SETSTRIPE:
2924 rc = lov_setstripe(exp, karg, uarg);
2926 case LL_IOC_LOV_GETSTRIPE:
2927 rc = lov_getstripe(exp, karg, uarg);
2929 case LL_IOC_LOV_SETEA:
2930 rc = lov_setea(exp, karg, uarg);
2937 for (i = 0; i < count; i++) {
2940 /* OST was deleted */
2941 if (obd_uuid_empty(&lov->tgts[i].uuid))
2944 err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
2947 if (lov->tgts[i].active) {
2948 CERROR("error: iocontrol OSC %s on OST"
2949 "idx %d: cmd %x err = %d\n",
2950 lov->tgts[i].uuid.uuid, i,
2966 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2967 void *key, __u32 *vallen, void *val)
2969 struct obd_device *obddev = class_exp2obd(exp);
2970 struct lov_obd *lov = &obddev->u.lov;
2974 if (!vallen || !val)
2977 if (keylen > strlen("lock_to_stripe") &&
2978 strcmp(key, "lock_to_stripe") == 0) {
2981 struct ldlm_lock *lock;
2982 struct lov_stripe_md *lsm;
2984 struct lov_oinfo *loi;
2985 struct ldlm_res_id *res_id = &data->lock->l_resource->lr_name;
2986 __u32 *stripe = val;
2988 if (*vallen < sizeof(*stripe))
2990 *vallen = sizeof(*stripe);
2992 /* XXX This is another one of those bits that will need to
2993 * change if we ever actually support nested LOVs. It uses
2994 * the lock's export to find out which stripe it is. */
2995 /* XXX - it's assumed all the locks for deleted OSTs have
2996 * been cancelled. Also, the export for deleted OSTs will
2997 * be NULL and won't match the lock's export. */
2998 for (i = 0, loi = data->lsm->lsm_oinfo;
2999 i < data->lsm->lsm_stripe_count;
3001 if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
3002 data->lock->l_conn_export &&
3003 loi->loi_id == res_id->name[0] &&
3004 loi->loi_gr == res_id->name[2]) {
3009 LDLM_ERROR(data->lock, "lock on inode without such object\n");
3010 dump_lsm(D_ERROR, data->lsm);
3012 } else if (keylen >= strlen("size_to_stripe") &&
3013 strcmp(key, "size_to_stripe") == 0) {
3017 struct lov_stripe_md *lsm;
3020 if (*vallen < sizeof(*data))
3023 data->size = lov_size_to_stripe(data->lsm, data->size,
3024 data->stripe_number);
3026 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
3028 int rc, size = sizeof(obd_id);
3029 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3030 if (!lov->tgts[i].active)
3032 rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key,
3038 } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
3039 struct lov_desc *desc_ret = val;
3040 *desc_ret = lov->desc;
3048 static int lov_set_info(struct obd_export *exp, obd_count keylen,
3049 void *key, obd_count vallen, void *val)
3051 struct obd_device *obddev = class_exp2obd(exp);
3052 struct lov_obd *lov = &obddev->u.lov;
3056 #define KEY_IS(str) \
3057 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
3059 if (KEY_IS("next_id")) {
3060 if (vallen != lov->desc.ld_tgt_count)
3062 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3065 /* OST was deleted */
3066 if (obd_uuid_empty(&lov->tgts[i].uuid))
3069 /* initialize all OSCs, even inactive ones */
3071 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key,
3072 sizeof(obd_id), ((obd_id*)val) + i);
3078 if (KEY_IS("growth_count")) {
3079 if (vallen != sizeof(int))
3081 } else if (KEY_IS("mds_conn")) {
3082 if (vallen != sizeof(__u32))
3084 } else if (KEY_IS("unlinked") || KEY_IS("unrecovery")) {
3091 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3094 /* OST was deleted */
3095 if (obd_uuid_empty(&lov->tgts[i].uuid))
3098 if (!val && !lov->tgts[i].active)
3101 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen,
3111 /* Merge rss if @kms_only == 0
3113 * Even when merging RSS, we will take the KMS value if it's larger.
3114 * This prevents getattr from stomping on dirty cached pages which
3115 * extend the file size. */
3116 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms_only)
3118 struct lov_oinfo *loi;
3122 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3124 obd_size lov_size, tmpsize;
3126 tmpsize = loi->loi_kms;
3127 if (kms_only == 0 && loi->loi_rss > tmpsize)
3128 tmpsize = loi->loi_rss;
3130 lov_size = lov_stripe_size(lsm, tmpsize, i);
3131 if (lov_size > size)
3137 EXPORT_SYMBOL(lov_merge_size);
3140 __u64 lov_merge_blocks(struct lov_stripe_md *lsm)
3142 struct lov_oinfo *loi;
3146 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3148 blocks += loi->loi_blocks;
3152 EXPORT_SYMBOL(lov_merge_blocks);
3154 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time)
3156 struct lov_oinfo *loi;
3159 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3161 if (loi->loi_mtime > current_time)
3162 current_time = loi->loi_mtime;
3164 return current_time;
3166 EXPORT_SYMBOL(lov_merge_mtime);
3169 struct lov_multi_wait {
3170 struct ldlm_lock *lock;
3176 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
3177 struct lustre_handle *lockh)
3179 struct lov_lock_handles *lov_lockh = NULL;
3180 struct lustre_handle *lov_lockhp;
3181 struct lov_obd *lov;
3182 struct lov_oinfo *loi;
3183 struct lov_multi_wait *queues;
3187 if (lsm_bad_magic(lsm))
3190 if (!exp || !exp->exp_obd)
3193 LASSERT(lockh != NULL);
3194 if (lsm->lsm_stripe_count > 1) {
3195 lov_lockh = lov_handle2llh(lockh);
3196 if (lov_lockh == NULL) {
3197 CERROR("LOV: invalid lov lock handle %p\n", lockh);
3201 lov_lockhp = lov_lockh->llh_handles;
3206 OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
3208 GOTO(out, rc = -ENOMEM);
3210 lov = &exp->exp_obd->u.lov;
3211 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3212 i++, loi++, lov_lockhp++) {
3213 struct ldlm_lock *lock;
3214 struct obd_device *obd;
3215 unsigned long irqflags;
3217 lock = ldlm_handle2lock(lov_lockhp);
3219 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
3220 loi->loi_ost_idx, loi->loi_id);
3221 queues[i].completed = 1;
3225 queues[i].lock = lock;
3226 init_waitqueue_entry(&(queues[i].wait), current);
3227 add_wait_queue(lock->l_waitq, &(queues[i].wait));
3229 obd = class_exp2obd(lock->l_conn_export);
3231 imp = obd->u.cli.cl_import;
3233 spin_lock_irqsave(&imp->imp_lock, irqflags);
3234 queues[i].generation = imp->imp_generation;
3235 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
3239 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
3240 interrupted_completion_wait, &lwd);
3241 rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
3243 for (i = 0; i < lsm->lsm_stripe_count; i++)
3244 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
3246 if (rc == -EINTR || rc == -ETIMEDOUT) {
3252 if (lov_lockh != NULL)
3253 lov_llh_put(lov_lockh);
3258 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
3261 struct lov_oinfo *loi;
3267 stripe = lov_stripe_number(lsm, size - 1);
3268 kms = lov_size_to_stripe(lsm, size, stripe);
3269 loi = &(lsm->lsm_oinfo[stripe]);
3271 CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n",
3272 stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms);
3273 if (kms > loi->loi_kms)
3277 EXPORT_SYMBOL(lov_increase_kms);
3279 struct obd_ops lov_obd_ops = {
3280 .o_owner = THIS_MODULE,
3281 .o_attach = lov_attach,
3282 .o_detach = lov_detach,
3283 .o_setup = lov_setup,
3284 .o_cleanup = lov_cleanup,
3285 .o_process_config = lov_process_config,
3286 .o_connect = lov_connect,
3287 .o_disconnect = lov_disconnect,
3288 .o_statfs = lov_statfs,
3289 .o_packmd = lov_packmd,
3290 .o_unpackmd = lov_unpackmd,
3291 .o_revalidate_md = lov_revalidate_md,
3292 .o_create = lov_create,
3293 .o_destroy = lov_destroy,
3294 .o_getattr = lov_getattr,
3295 .o_getattr_async = lov_getattr_async,
3296 .o_setattr = lov_setattr,
3298 .o_brw_async = lov_brw_async,
3299 .o_prep_async_page = lov_prep_async_page,
3300 .o_queue_async_io = lov_queue_async_io,
3301 .o_set_async_flags = lov_set_async_flags,
3302 .o_queue_group_io = lov_queue_group_io,
3303 .o_trigger_group_io = lov_trigger_group_io,
3304 .o_teardown_async_page = lov_teardown_async_page,
3305 .o_punch = lov_punch,
3307 .o_enqueue = lov_enqueue,
3308 .o_match = lov_match,
3309 .o_change_cbdata = lov_change_cbdata,
3310 .o_cancel = lov_cancel,
3311 .o_cancel_unused = lov_cancel_unused,
3312 .o_iocontrol = lov_iocontrol,
3313 .o_get_info = lov_get_info,
3314 .o_set_info = lov_set_info,
3315 .o_llog_init = lov_llog_init,
3316 .o_llog_finish = lov_llog_finish,
3317 .o_notify = lov_notify,
3320 int __init lov_init(void)
3322 struct lprocfs_static_vars lvars;
3326 lprocfs_init_vars(lov, &lvars);
3327 rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars,
3328 OBD_LOV_DEVICENAME);
3333 static void /*__exit*/ lov_exit(void)
3335 class_unregister_type(OBD_LOV_DEVICENAME);
3338 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3339 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
3340 MODULE_LICENSE("GPL");
3342 module_init(lov_init);
3343 module_exit(lov_exit);