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 obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
122 struct obd_uuid *tgt_uuid = &tgt->uuid;
125 struct proc_dir_entry *lov_proc_dir;
127 struct lov_obd *lov = &obd->u.lov;
128 struct lustre_handle conn = {0, };
129 struct obd_device *tgt_obd;
133 tgt_obd = class_find_client_obd(tgt_uuid, LUSTRE_OSC_NAME,
137 CERROR("Target %s not attached\n", tgt_uuid->uuid);
141 if (!tgt_obd->obd_set_up) {
142 CERROR("Target %s not set up\n", tgt_uuid->uuid);
147 tgt_obd->obd_no_recov = 0;
148 ptlrpc_activate_import(tgt_obd->u.cli.cl_import);
151 if (tgt_obd->u.cli.cl_import->imp_invalid) {
152 CERROR("not connecting OSC %s; administratively "
153 "disabled\n", tgt_uuid->uuid);
154 rc = obd_register_observer(tgt_obd, obd);
156 CERROR("Target %s register_observer error %d; "
157 "will not be able to reactivate\n",
163 rc = obd_connect(&conn, tgt_obd, &lov_osc_uuid, connect_flags);
165 CERROR("Target %s connect error %d\n", tgt_uuid->uuid, rc);
168 tgt->ltd_exp = class_conn2export(&conn);
170 rc = obd_register_observer(tgt_obd, obd);
172 CERROR("Target %s register_observer error %d\n",
174 obd_disconnect(tgt->ltd_exp, 0);
180 lov->desc.ld_active_tgt_count++;
183 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
185 struct obd_device *osc_obd = class_conn2obd(&conn);
186 struct proc_dir_entry *osc_symlink;
187 char name[MAX_STRING_SIZE + 1];
189 LASSERT(osc_obd != NULL);
190 LASSERT(osc_obd->obd_type != NULL);
191 LASSERT(osc_obd->obd_type->typ_name != NULL);
192 name[MAX_STRING_SIZE] = '\0';
193 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
194 osc_obd->obd_type->typ_name,
196 osc_symlink = proc_symlink(osc_obd->obd_name, lov_proc_dir,
198 if (osc_symlink == NULL) {
199 CERROR("could not register LOV target "
200 "/proc/fs/lustre/%s/%s/target_obds/%s\n",
201 obd->obd_type->typ_name, obd->obd_name,
203 lprocfs_remove(lov_proc_dir);
212 static int lov_connect(struct lustre_handle *conn, struct obd_device *obd,
213 struct obd_uuid *cluuid, unsigned long flags)
216 struct proc_dir_entry *lov_proc_dir;
218 struct lov_obd *lov = &obd->u.lov;
219 struct lov_tgt_desc *tgt;
220 struct obd_export *exp;
224 rc = class_connect(conn, obd, cluuid);
228 exp = class_conn2export(conn);
230 /* We don't want to actually do the underlying connections more than
231 * once, so keep track. */
233 if (lov->refcount > 1) {
234 class_export_put(exp);
239 lov_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
241 if (IS_ERR(lov_proc_dir)) {
242 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
243 obd->obd_type->typ_name, obd->obd_name);
248 /* connect_flags is the MDS number, save for use in lov_add_obd */
249 lov->lov_connect_flags = flags;
250 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
251 if (obd_uuid_empty(&tgt->uuid))
253 rc = lov_connect_obd(obd, tgt, 0, flags);
258 class_export_put(exp);
264 lprocfs_remove(lov_proc_dir);
268 struct obd_uuid uuid;
270 --lov->desc.ld_active_tgt_count;
272 /* save for CERROR below; (we know it's terminated) */
274 rc2 = obd_disconnect(tgt->ltd_exp, 0);
276 CERROR("error: LOV target %s disconnect on OST idx %d: "
277 "rc = %d\n", uuid.uuid, i, rc2);
279 class_disconnect(exp, 0);
283 static int lov_disconnect_obd(struct obd_device *obd,
284 struct lov_tgt_desc *tgt,
288 struct proc_dir_entry *lov_proc_dir;
290 struct obd_device *osc_obd = class_exp2obd(tgt->ltd_exp);
291 struct lov_obd *lov = &obd->u.lov;
296 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
298 struct proc_dir_entry *osc_symlink;
300 osc_symlink = lprocfs_srch(lov_proc_dir, osc_obd->obd_name);
302 lprocfs_remove(osc_symlink);
304 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
305 obd->obd_type->typ_name, obd->obd_name,
311 if (obd->obd_no_recov) {
312 /* Pass it on to our clients.
313 * XXX This should be an argument to disconnect,
314 * XXX not a back-door flag on the OBD. Ah well.
317 osc_obd->obd_no_recov = 1;
320 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
321 rc = obd_disconnect(tgt->ltd_exp, flags);
324 CERROR("Target %s disconnect error %d\n",
332 lov->desc.ld_active_tgt_count--;
339 static int lov_disconnect(struct obd_export *exp, unsigned long flags)
341 struct obd_device *obd = class_exp2obd(exp);
343 struct proc_dir_entry *lov_proc_dir;
345 struct lov_obd *lov = &obd->u.lov;
346 struct lov_tgt_desc *tgt;
353 /* Only disconnect the underlying layers on the final disconnect. */
355 if (lov->refcount != 0)
358 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
360 lov_disconnect_obd(obd, tgt, flags);
364 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
366 lprocfs_remove(lov_proc_dir);
368 CERROR("/proc/fs/lustre/%s/%s/target_obds missing.",
369 obd->obd_type->typ_name, obd->obd_name);
374 rc = class_disconnect(exp, 0);
380 * -EINVAL : UUID can't be found in the LOV's target list
381 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
382 * -EBADF : The UUID is found, but the OBD is the wrong type (!)
384 static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid,
387 struct lov_tgt_desc *tgt;
391 CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
392 lov, uuid->uuid, activate);
394 spin_lock(&lov->lov_lock);
395 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
396 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
397 i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
398 if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0)
402 if (i == lov->desc.ld_tgt_count)
403 GOTO(out, rc = -EINVAL);
406 if (tgt->active == activate) {
407 CDEBUG(D_INFO, "OSC %s already %sactive!\n", uuid->uuid,
408 activate ? "" : "in");
412 CDEBUG(D_INFO, "Marking OSC %s %sactive\n", uuid->uuid,
413 activate ? "" : "in");
415 tgt->active = activate;
417 lov->desc.ld_active_tgt_count++;
419 lov->desc.ld_active_tgt_count--;
423 spin_unlock(&lov->lov_lock);
427 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
428 int active, void *data)
430 struct obd_uuid *uuid;
434 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
435 CERROR("unexpected notification of %s %s!\n",
436 watched->obd_type->typ_name,
440 uuid = &watched->u.cli.cl_import->imp_target_uuid;
442 /* Set OSC as active before notifying the observer, so the
443 * observer can use the OSC normally.
445 rc = lov_set_osc_active(&obd->u.lov, uuid, active);
447 CERROR("%sactivation of %s failed: %d\n",
448 active ? "" : "de", uuid->uuid, rc);
452 if (obd->obd_observer)
453 /* Pass the notification up the chain. */
454 rc = obd_notify(obd->obd_observer, watched, active, data);
459 int lov_attach(struct obd_device *dev, obd_count len, void *data)
461 struct lprocfs_static_vars lvars;
464 lprocfs_init_vars(lov, &lvars);
465 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
468 struct proc_dir_entry *entry;
470 entry = create_proc_entry("target_obd_status", 0444,
471 dev->obd_proc_entry);
475 entry->proc_fops = &lov_proc_target_fops;
483 int lov_detach(struct obd_device *dev)
485 return lprocfs_obd_detach(dev);
488 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
490 struct lov_obd *lov = &obd->u.lov;
491 struct lustre_cfg *lcfg = buf;
492 struct lov_desc *desc;
496 if (lcfg->lcfg_inllen1 < 1) {
497 CERROR("LOV setup requires a descriptor\n");
501 desc = (struct lov_desc *)lcfg->lcfg_inlbuf1;
502 if (sizeof(*desc) > lcfg->lcfg_inllen1) {
503 CERROR("descriptor size wrong: %d > %d\n",
504 (int)sizeof(*desc), lcfg->lcfg_inllen1);
508 /* Because of 64-bit divide/mod operations only work with a 32-bit
509 * divisor in a 32-bit kernel, we cannot support a stripe width
510 * of 4GB or larger on 32-bit CPUs.
513 count = desc->ld_default_stripe_count;
514 if (count && (count * desc->ld_default_stripe_size) > ~0UL) {
515 CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n",
516 desc->ld_default_stripe_size, count, ~0UL);
519 if (desc->ld_tgt_count > 0) {
520 lov->bufsize= sizeof(struct lov_tgt_desc) * desc->ld_tgt_count;
522 lov->bufsize = sizeof(struct lov_tgt_desc) * LOV_MAX_TGT_COUNT;
524 OBD_ALLOC(lov->tgts, lov->bufsize);
525 if (lov->tgts == NULL) {
527 CERROR("couldn't allocate %d bytes for target table.\n",
532 desc->ld_tgt_count = 0;
533 desc->ld_active_tgt_count = 0;
535 spin_lock_init(&lov->lov_lock);
536 sema_init(&lov->lov_llog_sem, 1);
541 static int lov_cleanup(struct obd_device *obd, int flags)
543 struct lov_obd *lov = &obd->u.lov;
545 OBD_FREE(lov->tgts, lov->bufsize);
550 lov_add_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
552 struct lov_obd *lov = &obd->u.lov;
553 struct lov_tgt_desc *tgt;
557 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
558 uuidp->uuid, index, gen);
560 if ((index < 0) || (index >= LOV_MAX_TGT_COUNT)) {
561 CERROR("request to add OBD %s at invalid index: %d\n",
567 CERROR("request to add OBD %s with invalid generation: %d\n",
572 tgt = lov->tgts + index;
573 if (!obd_uuid_empty(&tgt->uuid)) {
574 CERROR("OBD already assigned at LOV target index %d\n",
580 /* XXX - add a sanity check on the generation number. */
583 if (index >= lov->desc.ld_tgt_count)
584 lov->desc.ld_tgt_count = index + 1;
586 CDEBUG(D_CONFIG, "idx: %d ltd_gen: %d ld_tgt_count: %d\n",
587 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
589 if (lov->refcount == 0)
593 struct obd_device *osc_obd;
595 osc_obd = class_exp2obd(tgt->ltd_exp);
597 osc_obd->obd_no_recov = 0;
600 rc = lov_connect_obd(obd, tgt, 1, lov->lov_connect_flags);
604 if (obd->obd_observer) {
605 /* tell the mds_lov about the new target */
606 rc = obd_notify(obd->obd_observer, tgt->ltd_exp->exp_obd, 1,
612 if (rc && tgt->ltd_exp != NULL)
613 lov_disconnect_obd(obd, tgt, 0);
618 lov_del_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
620 struct lov_obd *lov = &obd->u.lov;
621 struct lov_tgt_desc *tgt;
622 int count = lov->desc.ld_tgt_count;
626 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
627 uuidp->uuid, index, gen);
629 if (index >= count) {
630 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
635 tgt = lov->tgts + index;
637 if (obd_uuid_empty(&tgt->uuid)) {
638 CERROR("LOV target at index %d is not setup.\n", index);
642 if (strncmp(uuidp->uuid, tgt->uuid.uuid, sizeof uuidp->uuid) != 0) {
643 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
644 tgt->uuid.uuid, index, uuidp->uuid);
649 struct obd_device *osc_obd;
651 osc_obd = class_exp2obd(tgt->ltd_exp);
653 osc_obd->obd_no_recov = 1;
654 rc = obd_llog_finish(osc_obd, &osc_obd->obd_llogs, 1);
656 CERROR("osc_llog_finish error: %d\n", rc);
658 lov_disconnect_obd(obd, tgt, 0);
661 /* XXX - right now there is a dependency on ld_tgt_count being the
662 * maximum tgt index for computing the mds_max_easize. So we can't
665 /* lt_gen = 0 will mean it will not match the gen of any valid loi */
666 memset(tgt, 0, sizeof(*tgt));
668 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
669 tgt->uuid.uuid, index, tgt->ltd_gen, tgt->ltd_exp, tgt->active);
674 static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
676 struct lustre_cfg *lcfg = buf;
677 struct obd_uuid obd_uuid;
684 switch(cmd = lcfg->lcfg_command) {
685 case LCFG_LOV_ADD_OBD:
686 case LCFG_LOV_DEL_OBD: {
687 if (lcfg->lcfg_inllen1 > sizeof(obd_uuid.uuid))
688 GOTO(out, rc = -EINVAL);
690 obd_str2uuid(&obd_uuid, lcfg->lcfg_inlbuf1);
692 if (sscanf(lcfg->lcfg_inlbuf2, "%d", &index) != 1)
693 GOTO(out, rc = -EINVAL);
694 if (sscanf(lcfg->lcfg_inlbuf3, "%d", &gen) != 1)
695 GOTO(out, rc = -EINVAL);
696 if (cmd == LCFG_LOV_ADD_OBD)
697 rc = lov_add_obd(obd, &obd_uuid, index, gen);
699 rc = lov_del_obd(obd, &obd_uuid, index, gen);
703 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
704 GOTO(out, rc = -EINVAL);
712 /* compute object size given "stripeno" and the ost size */
713 static obd_size lov_stripe_size(struct lov_stripe_md *lsm, obd_size ost_size,
716 unsigned long ssize = lsm->lsm_stripe_size;
717 unsigned long swidth = ssize * lsm->lsm_stripe_count;
718 unsigned long stripe_size;
724 /* do_div(a, b) returns a % b, and a = a / b */
725 stripe_size = do_div(ost_size, ssize);
727 lov_size = ost_size * swidth + stripeno * ssize + stripe_size;
729 lov_size = (ost_size - 1) * swidth + (stripeno + 1) * ssize;
734 static void lov_merge_attrs(struct obdo *tgt, struct obdo *src, obd_valid valid,
735 struct lov_stripe_md *lsm, int stripeno, int *set)
737 valid &= src->o_valid;
740 if (valid & OBD_MD_FLSIZE) {
741 /* this handles sparse files properly */
744 lov_size = lov_stripe_size(lsm, src->o_size, stripeno);
745 if (lov_size > tgt->o_size)
746 tgt->o_size = lov_size;
748 if (valid & OBD_MD_FLBLOCKS)
749 tgt->o_blocks += src->o_blocks;
750 if (valid & OBD_MD_FLBLKSZ)
751 tgt->o_blksize += src->o_blksize;
752 if (valid & OBD_MD_FLCTIME && tgt->o_ctime < src->o_ctime)
753 tgt->o_ctime = src->o_ctime;
754 if (valid & OBD_MD_FLMTIME && tgt->o_mtime < src->o_mtime)
755 tgt->o_mtime = src->o_mtime;
757 memcpy(tgt, src, sizeof(*tgt));
758 tgt->o_id = lsm->lsm_object_id;
759 if (valid & OBD_MD_FLSIZE)
760 tgt->o_size = lov_stripe_size(lsm,src->o_size,stripeno);
766 #define log2(n) ffz(~(n))
769 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
770 struct lov_stripe_md **ea,
771 struct obd_trans_info *oti)
775 struct obd_uuid *ost_uuid = NULL;
779 LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
780 src_oa->o_flags == OBD_FL_DELORPHAN);
782 lov = &export->exp_obd->u.lov;
784 tmp_oa = obdo_alloc();
788 if (src_oa->o_valid & OBD_MD_FLINLINE) {
789 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
790 CDEBUG(D_HA, "clearing orphans only for %s\n",
794 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
795 struct lov_stripe_md obj_md;
796 struct lov_stripe_md *obj_mdp = &obj_md;
799 /* if called for a specific target, we don't
800 care if it is not active. */
801 if (lov->tgts[i].active == 0 && ost_uuid == NULL) {
802 CDEBUG(D_HA, "lov idx %d inactive\n", i);
806 if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid))
809 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
811 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
812 err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, &obj_mdp, oti);
814 /* This export will be disabled until it is recovered,
815 and then orphan recovery will be completed. */
816 CERROR("error in orphan recovery on OST idx %d/%d: "
817 "rc = %d\n", i, lov->desc.ld_tgt_count, err);
826 #define LOV_CREATE_RESEED_INTERVAL 1000
828 /* the LOV expects oa->o_id to be set to the LOV object id */
829 static int lov_create(struct obd_export *exp, struct obdo *src_oa,
830 struct lov_stripe_md **ea, struct obd_trans_info *oti)
832 static int ost_start_idx, ost_start_count;
834 struct lov_stripe_md *lsm;
835 struct lov_oinfo *loi = NULL;
836 struct obdo *tmp_oa, *ret_oa;
837 struct llog_cookie *cookies = NULL;
838 unsigned ost_count, ost_idx;
839 int set = 0, obj_alloc = 0, cookie_sent = 0, rc = 0, i;
844 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
845 src_oa->o_flags == OBD_FL_DELORPHAN) {
846 rc = lov_clear_orphans(exp, src_oa, ea, oti);
853 lov = &exp->exp_obd->u.lov;
855 if (!lov->desc.ld_active_tgt_count)
858 /* Recreate a specific object id at the given OST index */
859 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
860 (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) {
861 struct lov_stripe_md obj_md;
862 struct lov_stripe_md *obj_mdp = &obj_md;
864 ost_idx = src_oa->o_nlink;
868 if (ost_idx >= lov->desc.ld_tgt_count)
870 for (i = 0; i < lsm->lsm_stripe_count; i++) {
871 if (lsm->lsm_oinfo[i].loi_ost_idx == ost_idx) {
872 if (lsm->lsm_oinfo[i].loi_id != src_oa->o_id ||
873 lsm->lsm_oinfo[i].loi_gr != src_oa->o_gr) {
879 if (i == lsm->lsm_stripe_count)
882 rc = obd_create(lov->tgts[ost_idx].ltd_exp, src_oa,
887 ret_oa = obdo_alloc();
891 tmp_oa = obdo_alloc();
893 GOTO(out_oa, rc = -ENOMEM);
898 ost_count = lov_get_stripecnt(lov, 0);
900 /* If the MDS file was truncated up to some size, stripe over
901 * enough OSTs to allow the file to be created at that size. */
902 if (src_oa->o_valid & OBD_MD_FLSIZE) {
903 stripes=((src_oa->o_size+LUSTRE_STRIPE_MAXBYTES)>>12)-1;
904 do_div(stripes, (__u32)(LUSTRE_STRIPE_MAXBYTES >> 12));
906 if (stripes > lov->desc.ld_active_tgt_count)
908 if (stripes < ost_count)
914 rc = lov_alloc_memmd(&lsm, stripes, lov->desc.ld_pattern ?
915 lov->desc.ld_pattern : LOV_PATTERN_RAID0);
922 ost_count = lov->desc.ld_tgt_count;
924 LASSERT(src_oa->o_gr > 0);
925 LASSERT(src_oa->o_valid & OBD_MD_FLID);
926 lsm->lsm_object_id = src_oa->o_id;
927 lsm->lsm_object_gr = src_oa->o_gr;
928 if (!lsm->lsm_stripe_size)
929 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
930 if (!lsm->lsm_pattern) {
931 lsm->lsm_pattern = lov->desc.ld_pattern ?
932 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
935 if (*ea == NULL || lsm->lsm_oinfo[0].loi_ost_idx >= ost_count) {
936 if (--ost_start_count <= 0) {
937 ost_start_idx = ll_insecure_random_int();
938 ost_start_count = LOV_CREATE_RESEED_INTERVAL;
939 } else if (lsm->lsm_stripe_count >=
940 lov->desc.ld_active_tgt_count) {
941 /* If we allocate from all of the stripes, make the
942 * next file start on the next OST. */
945 ost_idx = ost_start_idx % ost_count;
947 ost_idx = lsm->lsm_oinfo[0].loi_ost_idx;
950 CDEBUG(D_INODE, "allocating %d subobjs for objid "LPX64" at idx %d\n",
951 lsm->lsm_stripe_count, lsm->lsm_object_id, ost_idx);
953 /* XXX LOV STACKING: need to figure out how many real OSCs */
954 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
955 oti_alloc_cookies(oti, lsm->lsm_stripe_count);
956 if (!oti->oti_logcookies)
957 GOTO(out_cleanup, rc = -ENOMEM);
958 cookies = oti->oti_logcookies;
961 loi = lsm->lsm_oinfo;
962 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
963 struct lov_stripe_md obj_md;
964 struct lov_stripe_md *obj_mdp = &obj_md;
968 if (lov->tgts[ost_idx].active == 0) {
969 if (!obd_uuid_empty(&lov->tgts[ost_idx].uuid))
970 CDEBUG(D_HA, "lov idx %d inactive\n", ost_idx);
974 /* create data objects with "parent" OA */
975 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
977 /* XXX When we start creating objects on demand, we need to
978 * make sure that we always create the object on the
979 * stripe which holds the existing file size.
981 if (src_oa->o_valid & OBD_MD_FLSIZE) {
982 if (lov_stripe_offset(lsm, src_oa->o_size, i,
983 &tmp_oa->o_size) < 0 &&
987 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
988 i, tmp_oa->o_size, src_oa->o_size);
992 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
993 err = obd_create(lov->tgts[ost_idx].ltd_exp, tmp_oa, &obj_mdp,
996 if (lov->tgts[ost_idx].active) {
997 CERROR("error creating objid "LPX64" sub-object"
998 " on OST idx %d/%d: rc = %d\n",
999 src_oa->o_id, ost_idx,
1000 lsm->lsm_stripe_count, err);
1002 CERROR("obd_create returned invalid "
1012 oti->oti_objid[ost_idx] = tmp_oa->o_id;
1013 loi->loi_id = tmp_oa->o_id;
1014 loi->loi_gr = tmp_oa->o_gr;
1015 loi->loi_ost_idx = ost_idx;
1016 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
1017 CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64" at "
1018 "idx %d gen %d\n", lsm->lsm_object_id, loi->loi_id,
1019 ost_idx, loi->loi_ost_gen);
1021 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1026 ++oti->oti_logcookies;
1027 if (tmp_oa->o_valid & OBD_MD_FLCOOKIE)
1032 /* If we have allocated enough objects, we are OK */
1033 if (obj_alloc == lsm->lsm_stripe_count)
1034 GOTO(out_done, rc = 0);
1037 if (obj_alloc == 0) {
1040 GOTO(out_cleanup, rc);
1043 /* If we were passed specific striping params, then a failure to
1044 * meet those requirements is an error, since we can't reallocate
1045 * that memory (it might be part of a larger array or something).
1047 * We can only get here if lsm_stripe_count was originally > 1.
1050 CERROR("can't lstripe objid "LPX64": have %u want %u, rc %d\n",
1051 lsm->lsm_object_id, obj_alloc, lsm->lsm_stripe_count,rc);
1054 GOTO(out_cleanup, rc);
1056 struct lov_stripe_md *lsm_new;
1057 /* XXX LOV STACKING call into osc for sizes */
1058 unsigned oldsize, newsize;
1060 if (oti && cookies && cookie_sent) {
1061 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
1062 newsize = obj_alloc * sizeof(*cookies);
1064 oti_alloc_cookies(oti, obj_alloc);
1065 if (oti->oti_logcookies) {
1066 memcpy(oti->oti_logcookies, cookies, newsize);
1067 OBD_FREE(cookies, oldsize);
1068 cookies = oti->oti_logcookies;
1070 CWARN("'leaking' %d bytes\n", oldsize-newsize);
1074 CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
1075 lsm->lsm_object_id, lsm->lsm_stripe_count, obj_alloc);
1076 oldsize = lov_stripe_md_size(lsm->lsm_stripe_count);
1077 newsize = lov_stripe_md_size(obj_alloc);
1078 OBD_ALLOC(lsm_new, newsize);
1079 if (lsm_new != NULL) {
1080 memcpy(lsm_new, lsm, newsize);
1081 lsm_new->lsm_stripe_count = obj_alloc;
1082 OBD_FREE(lsm, oldsize);
1085 CWARN("'leaking' %d bytes\n", oldsize - newsize);
1092 if (src_oa->o_valid & OBD_MD_FLSIZE &&
1093 ret_oa->o_size != src_oa->o_size) {
1094 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
1095 src_oa->o_size, ret_oa->o_size);
1098 ret_oa->o_id = src_oa->o_id;
1099 ret_oa->o_gr = src_oa->o_gr;
1100 ret_oa->o_valid |= OBD_MD_FLGROUP;
1101 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1107 if (oti && cookies) {
1108 oti->oti_logcookies = cookies;
1110 oti_free_cookies(oti);
1111 src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
1113 src_oa->o_valid |= OBD_MD_FLCOOKIE;
1119 while (obj_alloc-- > 0) {
1120 struct obd_export *sub_exp;
1124 sub_exp = lov->tgts[loi->loi_ost_idx].ltd_exp;
1125 /* destroy already created objects here */
1126 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1127 tmp_oa->o_id = loi->loi_id;
1129 err = obd_destroy(sub_exp, tmp_oa, NULL, oti);
1131 CERROR("Failed to uncreate objid "LPX64" subobj "LPX64
1132 " on OST idx %d: rc = %d\n", src_oa->o_id,
1133 loi->loi_id, loi->loi_ost_idx, err);
1136 obd_free_memmd(exp, &lsm);
1140 static int lov_revalidate_policy(struct lov_obd *lov, struct lov_stripe_md *lsm)
1142 static int next_idx = 0;
1143 struct lov_tgt_desc *tgt;
1146 /* XXX - we should do something clever and take lsm
1147 * into account but just do round robin for now. */
1149 /* last_idx must always be less that count because
1150 * ld_tgt_count currently cannot shrink. */
1151 count = lov->desc.ld_tgt_count;
1153 for (i = next_idx, tgt = lov->tgts + i; i < count; i++, tgt++) {
1155 next_idx = (i + 1) % count;
1160 for (i = 0, tgt = lov->tgts; i < next_idx; i++, tgt++) {
1162 next_idx = (i + 1) % count;
1170 #define lsm_bad_magic(LSMP) \
1172 struct lov_stripe_md *_lsm__ = (LSMP); \
1175 CERROR("LOV requires striping ea\n"); \
1177 } else if (_lsm__->lsm_magic != LOV_MAGIC) { \
1178 CERROR("LOV striping magic bad %#x != %#x\n", \
1179 _lsm__->lsm_magic, LOV_MAGIC); \
1185 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
1186 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1188 struct obdo *tmp = NULL;
1189 struct lov_oinfo *loi;
1190 struct lov_obd *lov;
1194 if (lsm_bad_magic(lsm))
1197 if (!exp || !exp->exp_obd)
1200 lov = &exp->exp_obd->u.lov;
1201 loi = lsm->lsm_oinfo;
1202 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1205 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1206 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1207 /* Orphan clean up will (someday) fix this up. */
1208 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
1209 oti->oti_logcookies++;
1216 memcpy(tmp, oa, sizeof(*tmp));
1217 tmp->o_id = loi->loi_id;
1218 err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp,
1221 if (err && lov->tgts[loi->loi_ost_idx].active) {
1222 CDEBUG(D_INODE, "error: destroying objid "LPX64" subobj "
1223 LPX64" on OST idx %d: rc = %d\n",
1224 oa->o_id, loi->loi_id, loi->loi_ost_idx, err);
1232 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
1233 struct lov_stripe_md *lsm)
1235 struct obdo *tmp = NULL;
1236 int i, rc = 0, set = 0;
1237 struct lov_oinfo *loi;
1238 struct lov_obd *lov;
1241 if (lsm_bad_magic(lsm))
1244 if (!exp || !exp->exp_obd)
1247 lov = &exp->exp_obd->u.lov;
1249 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1250 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1251 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1254 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1255 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1259 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1260 "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
1261 /* create data objects with "parent" OA */
1265 memcpy(tmp, oa, sizeof(*tmp));
1266 tmp->o_id = loi->loi_id;
1268 err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp,
1271 if (lov->tgts[loi->loi_ost_idx].active) {
1272 CERROR("error: getattr objid "LPX64" subobj "
1273 LPX64" on OST idx %d: rc = %d\n",
1274 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1280 lov_merge_attrs(oa, tmp, tmp->o_valid, lsm, i, &set);
1289 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
1292 struct lov_getattr_async_args *aa = data;
1293 struct lov_stripe_md *lsm = aa->aa_lsm;
1294 struct obdo *oa = aa->aa_oa;
1295 struct obdo *obdos = aa->aa_obdos;
1296 struct lov_oinfo *loi;
1302 /* NB all stripe requests succeeded to get here */
1304 loi = lsm->lsm_oinfo;
1305 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1306 if (obdos[i].o_valid == 0) /* inactive stripe */
1309 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1314 CERROR ("No stripes had valid attrs\n");
1319 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1323 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
1324 struct lov_stripe_md *lsm,
1325 struct ptlrpc_request_set *rqset)
1328 struct lov_obd *lov;
1329 struct lov_oinfo *loi;
1330 struct lov_getattr_async_args *aa;
1331 int i, rc = 0, set = 0;
1334 if (lsm_bad_magic(lsm))
1337 if (!exp || !exp->exp_obd)
1340 lov = &exp->exp_obd->u.lov;
1342 OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1346 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1347 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1348 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1351 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1352 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1353 /* leaves obdos[i].obd_valid unset */
1357 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at "
1358 "idx %u gen %d\n", oa->o_id, i, loi->loi_id,
1359 loi->loi_ost_idx, loi->loi_ost_gen);
1361 /* create data objects with "parent" OA */
1362 memcpy(&obdos[i], oa, sizeof(obdos[i]));
1363 obdos[i].o_id = loi->loi_id;
1365 err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp,
1366 &obdos[i], NULL, rqset);
1368 CERROR("error: getattr objid "LPX64" subobj "
1369 LPX64" on OST idx %d: rc = %d\n",
1370 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1372 GOTO(out_obdos, rc = err);
1377 GOTO (out_obdos, rc = -EIO);
1379 LASSERT (rqset->set_interpret == NULL);
1380 rqset->set_interpret = lov_getattr_interpret;
1381 LASSERT (sizeof (rqset->set_args) >= sizeof (*aa));
1382 aa = (struct lov_getattr_async_args *)&rqset->set_args;
1385 aa->aa_obdos = obdos;
1390 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1396 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
1397 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1399 struct obdo *tmp_oa, *ret_oa;
1400 struct lov_obd *lov;
1401 struct lov_oinfo *loi;
1402 int rc = 0, i, set = 0;
1405 if (lsm_bad_magic(lsm))
1408 if (!exp || !exp->exp_obd)
1411 /* for now, we only expect time updates here */
1412 LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
1413 OBD_MD_FLATIME | OBD_MD_FLMTIME |
1414 OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
1415 OBD_MD_FLSIZE | OBD_MD_FLGROUP)));
1417 LASSERT(!(src_oa->o_valid & OBD_MD_FLGROUP) || src_oa->o_gr > 0);
1419 ret_oa = obdo_alloc();
1423 tmp_oa = obdo_alloc();
1425 GOTO(out_oa, rc = -ENOMEM);
1427 lov = &exp->exp_obd->u.lov;
1428 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1431 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1432 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1436 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1437 tmp_oa->o_id = loi->loi_id;
1438 LASSERT(!(tmp_oa->o_valid & OBD_MD_FLGROUP) || tmp_oa->o_gr>0);
1440 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1441 if (lov_stripe_offset(lsm, src_oa->o_size, i,
1442 &tmp_oa->o_size) < 0 &&
1446 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1447 i, tmp_oa->o_size, src_oa->o_size);
1450 err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa,
1453 if (lov->tgts[loi->loi_ost_idx].active) {
1454 CERROR("error: setattr objid "LPX64" subobj "
1455 LPX64" on OST idx %d: rc = %d\n",
1456 src_oa->o_id, loi->loi_id,
1457 loi->loi_ost_idx, err);
1463 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set);
1468 ret_oa->o_id = src_oa->o_id;
1469 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1478 /* we have an offset in file backed by an lov and want to find out where
1479 * that offset lands in our given stripe of the file. for the easy
1480 * case where the offset is within the stripe, we just have to scale the
1481 * offset down to make it relative to the stripe instead of the lov.
1483 * the harder case is what to do when the offset doesn't intersect the
1484 * stripe. callers will want start offsets clamped ahead to the start
1485 * of the nearest stripe in the file. end offsets similarly clamped to the
1486 * nearest ending byte of a stripe in the file:
1488 * all this function does is move offsets to the nearest region of the
1489 * stripe, and it does its work "mod" the full length of all the stripes.
1490 * consider a file with 3 stripes:
1493 * ---------------------------------------------------------------------
1494 * | 0 | 1 | 2 | 0 | 1 | 2 |
1495 * ---------------------------------------------------------------------
1497 * to find stripe 1's offsets for S and E, it divides by the full stripe
1498 * width and does its math in the context of a single set of stripes:
1501 * -----------------------------------
1503 * -----------------------------------
1505 * it'll notice that E is outside stripe 1 and clamp it to the end of the
1506 * stripe, then multiply it back out by lov_off to give the real offsets in
1510 * ---------------------------------------------------------------------
1511 * | 1 | 1 | 1 | 1 | 1 | 1 |
1512 * ---------------------------------------------------------------------
1514 * it would have done similarly and pulled S forward to the start of a 1
1515 * stripe if, say, S had landed in a 0 stripe.
1517 * this rounding isn't always correct. consider an E lov offset that lands
1518 * on a 0 stripe, the "mod stripe width" math will pull it forward to the
1519 * start of a 1 stripe, when in fact it wanted to be rounded back to the end
1520 * of a previous 1 stripe. this logic is handled by callers and this is why:
1522 * this function returns < 0 when the offset was "before" the stripe and
1523 * was moved forward to the start of the stripe in question; 0 when it
1524 * falls in the stripe and no shifting was done; > 0 when the offset
1525 * was outside the stripe and was pulled back to its final byte. */
1526 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
1527 int stripeno, obd_off *obd_off)
1529 unsigned long ssize = lsm->lsm_stripe_size;
1530 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1531 unsigned long stripe_off, this_stripe;
1534 if (lov_off == OBD_OBJECT_EOF) {
1535 *obd_off = OBD_OBJECT_EOF;
1539 /* do_div(a, b) returns a % b, and a = a / b */
1540 stripe_off = do_div(lov_off, swidth);
1542 this_stripe = stripeno * ssize;
1543 if (stripe_off < this_stripe) {
1547 stripe_off -= this_stripe;
1549 if (stripe_off >= ssize) {
1555 *obd_off = lov_off * ssize + stripe_off;
1559 /* Given a whole-file size and a stripe number, give the file size which
1560 * corresponds to the individual object of that stripe.
1562 * This behaves basically in the same was as lov_stripe_offset, except that
1563 * file sizes falling before the beginning of a stripe are clamped to the end
1564 * of the previous stripe, not the beginning of the next:
1567 * ---------------------------------------------------------------------
1568 * | 0 | 1 | 2 | 0 | 1 | 2 |
1569 * ---------------------------------------------------------------------
1571 * if clamped to stripe 2 becomes:
1574 * ---------------------------------------------------------------------
1575 * | 0 | 1 | 2 | 0 | 1 | 2 |
1576 * ---------------------------------------------------------------------
1578 static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size,
1581 unsigned long ssize = lsm->lsm_stripe_size;
1582 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1583 unsigned long stripe_off, this_stripe;
1585 if (file_size == OBD_OBJECT_EOF)
1586 return OBD_OBJECT_EOF;
1588 /* do_div(a, b) returns a % b, and a = a / b */
1589 stripe_off = do_div(file_size, swidth);
1591 this_stripe = stripeno * ssize;
1592 if (stripe_off < this_stripe) {
1593 /* Move to end of previous stripe, or zero */
1594 if (file_size > 0) {
1601 stripe_off -= this_stripe;
1603 if (stripe_off >= ssize) {
1604 /* Clamp to end of this stripe */
1609 return (file_size * ssize + stripe_off);
1612 /* given an extent in an lov and a stripe, calculate the extent of the stripe
1613 * that is contained within the lov extent. this returns true if the given
1614 * stripe does intersect with the lov extent. */
1615 static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno,
1616 obd_off start, obd_off end,
1617 obd_off *obd_start, obd_off *obd_end)
1619 int start_side = 0, end_side = 0;
1621 switch (lsm->lsm_pattern) {
1622 case LOV_PATTERN_RAID0:
1623 start_side = lov_stripe_offset(lsm, start, stripeno, obd_start);
1624 end_side = lov_stripe_offset(lsm, end, stripeno, obd_end);
1626 case LOV_PATTERN_CMOBD:
1629 start_side = end_side = 0;
1635 CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n",
1636 start, end, start_side, *obd_start, *obd_end, end_side);
1638 /* this stripe doesn't intersect the file extent when neither
1639 * start or the end intersected the stripe and obd_start and
1640 * obd_end got rounded up to the save value. */
1641 if (start_side != 0 && end_side != 0 && *obd_start == *obd_end)
1644 /* as mentioned in the lov_stripe_offset commentary, end
1645 * might have been shifted in the wrong direction. This
1646 * happens when an end offset is before the stripe when viewed
1647 * through the "mod stripe size" math. we detect it being shifted
1648 * in the wrong direction and touch it up.
1649 * interestingly, this can't underflow since end must be > start
1650 * if we passed through the previous check.
1651 * (should we assert for that somewhere?) */
1658 /* compute which stripe number "lov_off" will be written into */
1659 static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off)
1661 unsigned long ssize = lsm->lsm_stripe_size;
1662 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1663 unsigned long stripe_off;
1665 if (lsm->lsm_pattern == LOV_PATTERN_CMOBD)
1668 stripe_off = do_div(lov_off, swidth);
1670 return stripe_off / ssize;
1673 static int lov_revalidate_md(struct obd_export *exp, struct obdo *src_oa,
1674 struct lov_stripe_md *ea,
1675 struct obd_trans_info *oti)
1677 struct obd_export *osc_exp;
1678 struct lov_obd *lov = &exp->exp_obd->u.lov;
1679 struct lov_stripe_md *lsm = ea;
1680 struct lov_stripe_md obj_md;
1681 struct lov_stripe_md *obj_mdp = &obj_md;
1682 struct lov_oinfo *loi;
1683 struct obdo *tmp_oa;
1684 int ost_idx, updates = 0, i;
1687 tmp_oa = obdo_alloc();
1691 loi = lsm->lsm_oinfo;
1692 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1694 if (!obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
1697 ost_idx = lov_revalidate_policy(lov, lsm);
1699 /* FIXME: punt for now. */
1700 CERROR("lov_revalidate_policy failed; no active "
1705 /* create a new object */
1706 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1707 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
1708 osc_exp = lov->tgts[ost_idx].ltd_exp;
1709 rc = obd_create(osc_exp, tmp_oa, &obj_mdp, oti);
1711 CERROR("error creating new subobj at idx %d; "
1712 "rc = %d\n", ost_idx, rc);
1716 oti->oti_objid[ost_idx] = tmp_oa->o_id;
1717 loi->loi_id = tmp_oa->o_id;
1718 loi->loi_ost_idx = ost_idx;
1719 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
1720 CDEBUG(D_INODE, "replacing objid "LPX64" subobj "LPX64
1721 " with idx %d gen %d.\n", lsm->lsm_object_id,
1722 loi->loi_id, ost_idx, loi->loi_ost_gen);
1726 /* If we got an error revalidating an entry there's no need to
1727 * cleanup up objects we allocated here because the bad entry
1728 * still points to a deleted OST. */
1734 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1735 * we can send this 'punch' to just the authoritative node and the nodes
1736 * that the punch will affect. */
1737 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1738 struct lov_stripe_md *lsm,
1739 obd_off start, obd_off end, struct obd_trans_info *oti)
1741 struct obdo *tmp = NULL;
1742 struct lov_oinfo *loi;
1743 struct lov_obd *lov;
1747 if (lsm_bad_magic(lsm))
1750 if (!exp || !exp->exp_obd)
1753 lov = &exp->exp_obd->u.lov;
1754 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1755 obd_off starti, endi;
1758 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1759 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1763 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1766 /* create data objects with "parent" OA */
1770 memcpy(tmp, oa, sizeof(*tmp));
1771 tmp->o_id = loi->loi_id;
1773 err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp,
1774 tmp, NULL, starti, endi, NULL);
1777 if (lov->tgts[loi->loi_ost_idx].active) {
1778 CERROR("error: punch objid "LPX64" subobj "LPX64
1779 " on OST idx %d: rc = %d\n", oa->o_id,
1780 loi->loi_id, loi->loi_ost_idx, err);
1785 loi->loi_kms = loi->loi_rss = starti;
1791 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1792 struct lov_stripe_md *lsm, obd_off start, obd_off end)
1795 struct lov_obd *lov;
1796 struct lov_oinfo *loi;
1800 if (lsm_bad_magic(lsm))
1810 lov = &exp->exp_obd->u.lov;
1811 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1812 obd_off starti, endi;
1815 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1816 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1820 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1823 memcpy(tmp, oa, sizeof(*tmp));
1824 tmp->o_id = loi->loi_id;
1826 err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL,
1829 if (lov->tgts[loi->loi_ost_idx].active) {
1830 CERROR("error: fsync objid "LPX64" subobj "LPX64
1831 " on OST idx %d: rc = %d\n", oa->o_id,
1832 loi->loi_id, loi->loi_ost_idx, err);
1843 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1844 struct lov_stripe_md *lsm,
1845 obd_count oa_bufs, struct brw_page *pga)
1850 /* The caller just wants to know if there's a chance that this
1851 * I/O can succeed */
1852 for (i = 0; i < oa_bufs; i++) {
1853 int stripe = lov_stripe_number(lsm, pga[i].disk_offset);
1854 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1857 if (!lov_stripe_intersects(lsm, i, pga[i].disk_offset,
1858 pga[i].disk_offset + pga[i].count,
1862 if (lov->tgts[ost].active == 0) {
1863 CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1866 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[ost].ltd_exp, oa,
1867 NULL, 1, &pga[i], NULL);
1874 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1875 struct lov_stripe_md *lsm, obd_count oa_bufs,
1876 struct brw_page *pga, struct obd_trans_info *oti)
1882 struct lov_stripe_md lsm;
1884 } *stripeinfo, *si, *si_last;
1885 struct obdo *ret_oa = NULL, *tmp_oa = NULL;
1886 struct lov_obd *lov;
1887 struct brw_page *ioarr;
1888 struct lov_oinfo *loi;
1889 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0;
1892 if (lsm_bad_magic(lsm))
1895 lov = &exp->exp_obd->u.lov;
1897 if (cmd == OBD_BRW_CHECK) {
1898 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1902 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1906 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1908 GOTO(out_sinfo, rc = -ENOMEM);
1910 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1912 GOTO(out_where, rc = -ENOMEM);
1915 ret_oa = obdo_alloc();
1917 GOTO(out_ioarr, rc = -ENOMEM);
1919 tmp_oa = obdo_alloc();
1921 GOTO(out_oa, rc = -ENOMEM);
1924 for (i = 0; i < oa_bufs; i++) {
1925 where[i] = lov_stripe_number(lsm, pga[i].disk_offset);
1926 stripeinfo[where[i]].bufct++;
1929 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1930 i < stripe_count; i++, loi++, si_last = si, si++) {
1932 si->index = si_last->index + si_last->bufct;
1933 si->lsm.lsm_object_id = loi->loi_id;
1934 si->lsm.lsm_object_gr = lsm->lsm_object_gr;
1935 si->ost_idx = loi->loi_ost_idx;
1938 for (i = 0; i < oa_bufs; i++) {
1939 int which = where[i];
1942 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1943 LASSERT(shift < oa_bufs);
1944 ioarr[shift] = pga[i];
1945 lov_stripe_offset(lsm, pga[i].disk_offset, which,
1946 &ioarr[shift].disk_offset);
1947 stripeinfo[which].subcount++;
1950 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1951 int shift = si->index;
1953 if (lov->tgts[si->ost_idx].active == 0) {
1954 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1955 GOTO(out_oa, rc = -EIO);
1959 LASSERT(shift < oa_bufs);
1961 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1963 tmp_oa->o_id = si->lsm.lsm_object_id;
1964 rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp,
1965 tmp_oa, &si->lsm, si->bufct,
1966 &ioarr[shift], oti);
1970 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1975 ret_oa->o_id = src_oa->o_id;
1976 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1985 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1987 OBD_FREE(where, sizeof(*where) * oa_bufs);
1989 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1993 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1996 struct lov_brw_async_args *aa = data;
1997 struct lov_stripe_md *lsm = aa->aa_lsm;
1998 obd_count oa_bufs = aa->aa_oa_bufs;
1999 struct obdo *oa = aa->aa_oa;
2000 struct obdo *obdos = aa->aa_obdos;
2001 struct brw_page *ioarr = aa->aa_ioarr;
2002 struct lov_oinfo *loi;
2007 /* NB all stripe requests succeeded to get here */
2009 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2011 if (obdos[i].o_valid == 0) /* inactive stripe */
2014 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
2019 CERROR("No stripes had valid attrs\n");
2023 oa->o_id = lsm->lsm_object_id;
2025 OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos));
2026 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
2030 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
2031 struct lov_stripe_md *lsm, obd_count oa_bufs,
2032 struct brw_page *pga, struct ptlrpc_request_set *set,
2033 struct obd_trans_info *oti)
2039 struct lov_stripe_md lsm;
2041 } *stripeinfo, *si, *si_last;
2042 struct lov_obd *lov;
2043 struct brw_page *ioarr;
2044 struct obdo *obdos = NULL;
2045 struct lov_oinfo *loi;
2046 struct lov_brw_async_args *aa;
2047 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count;
2050 if (lsm_bad_magic(lsm))
2053 lov = &exp->exp_obd->u.lov;
2055 if (cmd == OBD_BRW_CHECK) {
2056 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
2060 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
2064 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
2066 GOTO(out_sinfo, rc = -ENOMEM);
2069 OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count);
2071 GOTO(out_where, rc = -ENOMEM);
2074 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
2076 GOTO(out_obdos, rc = -ENOMEM);
2078 for (i = 0; i < oa_bufs; i++) {
2079 where[i] = lov_stripe_number(lsm, pga[i].disk_offset);
2080 stripeinfo[where[i]].bufct++;
2083 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
2084 i < stripe_count; i++, loi++, si_last = si, si++) {
2086 si->index = si_last->index + si_last->bufct;
2087 si->lsm.lsm_object_id = loi->loi_id;
2088 si->ost_idx = loi->loi_ost_idx;
2091 memcpy(&obdos[i], oa, sizeof(*obdos));
2092 obdos[i].o_id = si->lsm.lsm_object_id;
2096 for (i = 0; i < oa_bufs; i++) {
2097 int which = where[i];
2100 shift = stripeinfo[which].index + stripeinfo[which].subcount;
2101 LASSERT(shift < oa_bufs);
2102 ioarr[shift] = pga[i];
2103 lov_stripe_offset(lsm, pga[i].disk_offset, which,
2104 &ioarr[shift].disk_offset);
2105 stripeinfo[which].subcount++;
2108 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
2109 int shift = si->index;
2114 if (lov->tgts[si->ost_idx].active == 0) {
2115 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
2116 GOTO(out_ioarr, rc = -EIO);
2119 LASSERT(shift < oa_bufs);
2121 rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp,
2122 &obdos[i], &si->lsm, si->bufct,
2123 &ioarr[shift], set, oti);
2125 GOTO(out_ioarr, rc);
2128 LASSERT(set->set_interpret == NULL);
2129 set->set_interpret = (set_interpreter_func)lov_brw_interpret;
2130 LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args));
2131 aa = (struct lov_brw_async_args *)&set->set_args;
2133 aa->aa_obdos = obdos;
2135 aa->aa_ioarr = ioarr;
2136 aa->aa_oa_bufs = oa_bufs;
2138 /* Don't free ioarr or obdos - that's done in lov_brw_interpret */
2139 GOTO(out_where, rc);
2142 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
2144 OBD_FREE(obdos, stripe_count * sizeof(*obdos));
2146 OBD_FREE(where, sizeof(*where) * oa_bufs);
2148 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
2152 struct lov_async_page *lap_from_cookie(void *cookie)
2154 struct lov_async_page *lap = cookie;
2155 if (lap->lap_magic != LAP_MAGIC)
2156 return ERR_PTR(-EINVAL);
2160 static int lov_ap_make_ready(void *data, int cmd)
2162 struct lov_async_page *lap = lap_from_cookie(data);
2163 /* XXX should these assert? */
2167 return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
2169 static int lov_ap_refresh_count(void *data, int cmd)
2171 struct lov_async_page *lap = lap_from_cookie(data);
2175 return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
2178 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
2180 struct lov_async_page *lap = lap_from_cookie(data);
2181 /* XXX should these assert? */
2185 lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
2186 /* XXX woah, shouldn't we be altering more here? size? */
2187 oa->o_id = lap->lap_loi_id;
2190 static void lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
2192 struct lov_async_page *lap = lap_from_cookie(data);
2196 /* in a raid1 regime this would down a count of many ios
2197 * in flight, onl calling the caller_ops completion when all
2198 * the raid1 ios are complete */
2199 lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, oa, rc);
2202 static struct obd_async_page_ops lov_async_page_ops = {
2203 .ap_make_ready = lov_ap_make_ready,
2204 .ap_refresh_count = lov_ap_refresh_count,
2205 .ap_fill_obdo = lov_ap_fill_obdo,
2206 .ap_completion = lov_ap_completion,
2209 static int lov_prep_async_page(struct obd_export *exp,
2210 struct lov_stripe_md *lsm,
2211 struct lov_oinfo *loi, struct page *page,
2212 obd_off offset, struct obd_async_page_ops *ops,
2213 void *data, void **res)
2215 struct lov_obd *lov = &exp->exp_obd->u.lov;
2216 struct lov_async_page *lap;
2220 if (lsm_bad_magic(lsm))
2222 LASSERT(loi == NULL);
2224 stripe = lov_stripe_number(lsm, offset);
2225 loi = &lsm->lsm_oinfo[stripe];
2227 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
2229 if (lov->tgts[loi->loi_ost_idx].active == 0)
2231 if (lov->tgts[loi->loi_ost_idx].ltd_exp == NULL) {
2232 CERROR("ltd_exp == NULL, but OST idx %d doesn't appear to be "
2233 "deleted or inactive.\n", loi->loi_ost_idx);
2237 OBD_ALLOC(lap, sizeof(*lap));
2241 lap->lap_magic = LAP_MAGIC;
2242 lap->lap_caller_ops = ops;
2243 lap->lap_caller_data = data;
2245 /* FIXME handle multiple oscs after landing b_raid1 */
2246 lap->lap_stripe = stripe;
2247 switch (lsm->lsm_pattern) {
2248 case LOV_PATTERN_RAID0:
2249 lov_stripe_offset(lsm, offset, lap->lap_stripe,
2250 &lap->lap_sub_offset);
2252 case LOV_PATTERN_CMOBD:
2253 lap->lap_sub_offset = offset;
2259 /* so the callback doesn't need the lsm */
2260 lap->lap_loi_id = loi->loi_id;
2262 rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2263 lsm, loi, page, lap->lap_sub_offset,
2264 &lov_async_page_ops, lap,
2265 &lap->lap_sub_cookie);
2267 OBD_FREE(lap, sizeof(*lap));
2270 CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
2271 lap->lap_sub_cookie, offset);
2276 static int lov_queue_async_io(struct obd_export *exp,
2277 struct lov_stripe_md *lsm,
2278 struct lov_oinfo *loi, void *cookie,
2279 int cmd, obd_off off, int count,
2280 obd_flags brw_flags, obd_flags async_flags)
2282 struct lov_obd *lov = &exp->exp_obd->u.lov;
2283 struct lov_async_page *lap;
2286 LASSERT(loi == NULL);
2288 if (lsm_bad_magic(lsm))
2291 lap = lap_from_cookie(cookie);
2293 RETURN(PTR_ERR(lap));
2295 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2297 rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
2298 loi, lap->lap_sub_cookie, cmd, off, count,
2299 brw_flags, async_flags);
2303 static int lov_set_async_flags(struct obd_export *exp,
2304 struct lov_stripe_md *lsm,
2305 struct lov_oinfo *loi, void *cookie,
2306 obd_flags async_flags)
2308 struct lov_obd *lov = &exp->exp_obd->u.lov;
2309 struct lov_async_page *lap;
2312 LASSERT(loi == NULL);
2314 if (lsm_bad_magic(lsm))
2317 lap = lap_from_cookie(cookie);
2319 RETURN(PTR_ERR(lap));
2321 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2323 rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
2324 lsm, loi, lap->lap_sub_cookie, async_flags);
2328 static int lov_queue_group_io(struct obd_export *exp,
2329 struct lov_stripe_md *lsm,
2330 struct lov_oinfo *loi,
2331 struct obd_io_group *oig, void *cookie,
2332 int cmd, obd_off off, int count,
2333 obd_flags brw_flags, obd_flags async_flags)
2335 struct lov_obd *lov = &exp->exp_obd->u.lov;
2336 struct lov_async_page *lap;
2339 LASSERT(loi == NULL);
2341 if (lsm_bad_magic(lsm))
2344 lap = lap_from_cookie(cookie);
2346 RETURN(PTR_ERR(lap));
2348 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2350 rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
2351 oig, lap->lap_sub_cookie, cmd, off, count,
2352 brw_flags, async_flags);
2356 /* this isn't exactly optimal. we may have queued sync io in oscs on
2357 * all stripes, but we don't record that fact at queue time. so we
2358 * trigger sync io on all stripes. */
2359 static int lov_trigger_group_io(struct obd_export *exp,
2360 struct lov_stripe_md *lsm,
2361 struct lov_oinfo *loi,
2362 struct obd_io_group *oig)
2364 struct lov_obd *lov = &exp->exp_obd->u.lov;
2367 LASSERT(loi == NULL);
2369 if (lsm_bad_magic(lsm))
2372 loi = lsm->lsm_oinfo;
2373 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
2374 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2375 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2379 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
2381 if (rc == 0 && err != 0)
2387 static int lov_teardown_async_page(struct obd_export *exp,
2388 struct lov_stripe_md *lsm,
2389 struct lov_oinfo *loi, void *cookie)
2391 struct lov_obd *lov = &exp->exp_obd->u.lov;
2392 struct lov_async_page *lap;
2395 LASSERT(loi == NULL);
2397 if (lsm_bad_magic(lsm))
2400 lap = lap_from_cookie(cookie);
2402 RETURN(PTR_ERR(lap));
2404 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2406 rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2407 lsm, loi, lap->lap_sub_cookie);
2409 CERROR("unable to teardown sub cookie %p: %d\n",
2410 lap->lap_sub_cookie, rc);
2413 OBD_FREE(lap, sizeof(*lap));
2417 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2418 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2419 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2420 void *data,__u32 lvb_len, void *lvb_swabber,
2421 struct lustre_handle *lockh)
2423 struct lov_lock_handles *lov_lockh = NULL;
2424 struct lustre_handle *lov_lockhp;
2425 struct lov_obd *lov;
2426 struct lov_oinfo *loi;
2427 char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)];
2428 struct lov_stripe_md *submd = (void *)submd_buf;
2430 int i, save_flags = *flags;
2433 if (lsm_bad_magic(lsm))
2436 /* we should never be asked to replay a lock this way. */
2437 LASSERT((*flags & LDLM_FL_REPLAY) == 0);
2439 if (!exp || !exp->exp_obd)
2442 if (lsm->lsm_stripe_count > 1) {
2443 lov_lockh = lov_llh_new(lsm);
2444 if (lov_lockh == NULL)
2447 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2448 lov_lockhp = lov_lockh->llh_handles;
2453 lov = &exp->exp_obd->u.lov;
2454 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2455 i++, loi++, lov_lockhp++) {
2456 ldlm_policy_data_t sub_ext;
2459 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2460 policy->l_extent.end, &start,
2464 sub_ext.l_extent.start = start;
2465 sub_ext.l_extent.end = end;
2466 sub_ext.l_extent.gid = policy->l_extent.gid;
2468 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2469 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2473 /* XXX LOV STACKING: submd should be from the subobj */
2474 submd->lsm_object_id = loi->loi_id;
2475 submd->lsm_object_gr = lsm->lsm_object_gr;
2476 submd->lsm_stripe_count = 0;
2477 submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid;
2478 submd->lsm_oinfo->loi_rss = loi->loi_rss;
2479 submd->lsm_oinfo->loi_kms = loi->loi_kms;
2480 submd->lsm_oinfo->loi_blocks = loi->loi_blocks;
2481 loi->loi_mtime = submd->lsm_oinfo->loi_mtime;
2482 /* XXX submd is not fully initialized here */
2483 *flags = save_flags;
2484 rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd,
2485 type, &sub_ext, mode, flags, bl_cb, cp_cb,
2486 gl_cb, data, lvb_len, lvb_swabber, lov_lockhp);
2488 /* XXX FIXME: This unpleasantness doesn't belong here at *all*.
2489 * It belongs in the OSC, except that the OSC doesn't have
2490 * access to the real LOI -- it gets a copy, that we created
2491 * above, and that copy can be arbitrarily out of date.
2493 * The LOV API is due for a serious rewriting anyways, and this
2494 * can be addressed then. */
2495 if (rc == ELDLM_OK) {
2496 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2497 __u64 tmp = submd->lsm_oinfo->loi_rss;
2499 LASSERT(lock != NULL);
2501 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2502 /* Extend KMS up to the end of this lock and no further
2503 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2504 if (tmp > lock->l_policy_data.l_extent.end)
2505 tmp = lock->l_policy_data.l_extent.end + 1;
2506 if (tmp >= loi->loi_kms) {
2507 CDEBUG(D_INODE, "lock acquired, setting rss="
2508 LPU64", kms="LPU64"\n", loi->loi_rss,
2511 loi->loi_kms_valid = 1;
2513 CDEBUG(D_INODE, "lock acquired, setting rss="
2514 LPU64"; leaving kms="LPU64", end="LPU64
2515 "\n", loi->loi_rss, loi->loi_kms,
2516 lock->l_policy_data.l_extent.end);
2518 ldlm_lock_allow_match(lock);
2519 LDLM_LOCK_PUT(lock);
2520 } else if (rc == ELDLM_LOCK_ABORTED &&
2521 save_flags & LDLM_FL_HAS_INTENT) {
2522 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2523 loi->loi_rss = submd->lsm_oinfo->loi_rss;
2524 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2525 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2526 " kms="LPU64"\n", loi->loi_rss, loi->loi_kms);
2528 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2529 if (lov->tgts[loi->loi_ost_idx].active) {
2530 CERROR("error: enqueue objid "LPX64" subobj "
2531 LPX64" on OST idx %d: rc = %d\n",
2532 lsm->lsm_object_id, loi->loi_id,
2533 loi->loi_ost_idx, rc);
2534 GOTO(out_locks, rc);
2538 if (lsm->lsm_stripe_count > 1)
2539 lov_llh_put(lov_lockh);
2543 while (loi--, lov_lockhp--, i-- > 0) {
2544 struct lov_stripe_md submd;
2547 if (lov_lockhp->cookie == 0)
2550 /* XXX LOV STACKING: submd should be from the subobj */
2551 submd.lsm_object_id = loi->loi_id;
2552 submd.lsm_object_gr = lsm->lsm_object_gr;
2553 submd.lsm_stripe_count = 0;
2554 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2556 if (err && lov->tgts[loi->loi_ost_idx].active) {
2557 CERROR("error: cancelling objid "LPX64" on OST "
2558 "idx %d after enqueue error: rc = %d\n",
2559 loi->loi_id, loi->loi_ost_idx, err);
2563 if (lsm->lsm_stripe_count > 1) {
2564 lov_llh_destroy(lov_lockh);
2565 lov_llh_put(lov_lockh);
2570 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2571 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2572 int *flags, void *data, struct lustre_handle *lockh)
2574 struct lov_lock_handles *lov_lockh = NULL;
2575 struct lustre_handle *lov_lockhp;
2576 struct lov_obd *lov;
2577 struct lov_oinfo *loi;
2578 struct lov_stripe_md submd;
2579 ldlm_error_t rc = 0;
2583 if (lsm_bad_magic(lsm))
2586 if (!exp || !exp->exp_obd)
2589 if (lsm->lsm_stripe_count > 1) {
2590 lov_lockh = lov_llh_new(lsm);
2591 if (lov_lockh == NULL)
2594 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2595 lov_lockhp = lov_lockh->llh_handles;
2600 lov = &exp->exp_obd->u.lov;
2601 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2602 i++, loi++, lov_lockhp++) {
2603 ldlm_policy_data_t sub_ext;
2607 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2608 policy->l_extent.end, &start, &end))
2611 sub_ext.l_extent.start = start;
2612 sub_ext.l_extent.end = end;
2614 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid)) {
2615 CDEBUG(D_HA, "lov idx %d deleted\n", loi->loi_ost_idx);
2618 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2619 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2624 /* XXX LOV STACKING: submd should be from the subobj */
2625 submd.lsm_object_id = loi->loi_id;
2626 submd.lsm_object_gr = lsm->lsm_object_gr;
2627 submd.lsm_stripe_count = 0;
2629 /* XXX submd is not fully initialized here */
2630 rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2631 type, &sub_ext, mode, &lov_flags, data,
2637 if (lsm->lsm_stripe_count > 1) {
2638 if (*flags & LDLM_FL_TEST_LOCK)
2639 lov_llh_destroy(lov_lockh);
2640 lov_llh_put(lov_lockh);
2645 while (loi--, lov_lockhp--, i-- > 0) {
2646 struct lov_stripe_md submd;
2649 if (lov_lockhp->cookie == 0)
2652 /* XXX LOV STACKING: submd should be from the subobj */
2653 submd.lsm_object_id = loi->loi_id;
2654 submd.lsm_object_gr = lsm->lsm_object_gr;
2655 submd.lsm_stripe_count = 0;
2656 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2658 if (err && lov->tgts[loi->loi_ost_idx].active) {
2659 CERROR("error: cancelling objid "LPX64" on OST "
2660 "idx %d after match failure: rc = %d\n",
2661 loi->loi_id, loi->loi_ost_idx, err);
2665 if (lsm->lsm_stripe_count > 1) {
2666 lov_llh_destroy(lov_lockh);
2667 lov_llh_put(lov_lockh);
2672 static int lov_change_cbdata(struct obd_export *exp,
2673 struct lov_stripe_md *lsm, ldlm_iterator_t it,
2676 struct lov_obd *lov;
2677 struct lov_oinfo *loi;
2681 if (lsm_bad_magic(lsm))
2684 if (!exp || !exp->exp_obd)
2687 LASSERT(lsm->lsm_object_gr > 0);
2689 lov = &exp->exp_obd->u.lov;
2690 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2691 struct lov_stripe_md submd;
2692 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2693 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2697 submd.lsm_object_id = loi->loi_id;
2698 submd.lsm_object_gr = lsm->lsm_object_gr;
2699 submd.lsm_stripe_count = 0;
2700 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
2706 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
2707 __u32 mode, struct lustre_handle *lockh)
2709 struct lov_lock_handles *lov_lockh = NULL;
2710 struct lustre_handle *lov_lockhp;
2711 struct lov_obd *lov;
2712 struct lov_oinfo *loi;
2716 if (lsm_bad_magic(lsm))
2719 if (!exp || !exp->exp_obd)
2722 LASSERT(lsm->lsm_object_gr > 0);
2725 if (lsm->lsm_stripe_count > 1) {
2726 lov_lockh = lov_handle2llh(lockh);
2728 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2732 lov_lockhp = lov_lockh->llh_handles;
2737 lov = &exp->exp_obd->u.lov;
2738 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2739 i++, loi++, lov_lockhp++) {
2740 struct lov_stripe_md submd;
2743 if (lov_lockhp->cookie == 0) {
2744 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2745 loi->loi_ost_idx, loi->loi_id);
2749 /* XXX LOV STACKING: submd should be from the subobj */
2750 submd.lsm_object_id = loi->loi_id;
2751 submd.lsm_object_gr = lsm->lsm_object_gr;
2752 submd.lsm_stripe_count = 0;
2753 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2756 if (lov->tgts[loi->loi_ost_idx].active) {
2757 CERROR("error: cancel objid "LPX64" subobj "
2758 LPX64" on OST idx %d: rc = %d\n",
2760 loi->loi_id, loi->loi_ost_idx, err);
2767 if (lsm->lsm_stripe_count > 1)
2768 lov_llh_destroy(lov_lockh);
2769 if (lov_lockh != NULL)
2770 lov_llh_put(lov_lockh);
2774 static int lov_cancel_unused(struct obd_export *exp,
2775 struct lov_stripe_md *lsm, int flags, void *opaque)
2777 struct lov_obd *lov;
2778 struct lov_oinfo *loi;
2782 lov = &exp->exp_obd->u.lov;
2784 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2785 int err = obd_cancel_unused(lov->tgts[i].ltd_exp, NULL,
2793 if (lsm_bad_magic(lsm))
2796 if (!exp || !exp->exp_obd)
2799 LASSERT(lsm->lsm_object_gr > 0);
2801 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2802 struct lov_stripe_md submd;
2805 if (lov->tgts[loi->loi_ost_idx].active == 0)
2806 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2808 submd.lsm_object_id = loi->loi_id;
2809 submd.lsm_object_gr = lsm->lsm_object_gr;
2810 submd.lsm_stripe_count = 0;
2811 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
2812 &submd, flags, opaque);
2813 if (err && lov->tgts[loi->loi_ost_idx].active) {
2814 CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2815 " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2816 loi->loi_id, loi->loi_ost_idx, err);
2824 #define LOV_U64_MAX ((__u64)~0ULL)
2825 #define LOV_SUM_MAX(tot, add) \
2827 if ((tot) + (add) < (tot)) \
2828 (tot) = LOV_U64_MAX; \
2833 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2834 unsigned long max_age)
2836 struct lov_obd *lov = &obd->u.lov;
2837 struct obd_statfs lov_sfs;
2844 /* We only get block data from the OBD */
2845 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2847 if (!lov->tgts[i].active) {
2848 CDEBUG(D_HA, "lov idx %d inactive\n", i);
2852 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
2855 if (lov->tgts[i].active && !rc)
2861 memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
2864 osfs->os_bfree += lov_sfs.os_bfree;
2865 osfs->os_bavail += lov_sfs.os_bavail;
2866 osfs->os_blocks += lov_sfs.os_blocks;
2867 /* XXX not sure about this one - depends on policy.
2868 * - could be minimum if we always stripe on all OBDs
2869 * (but that would be wrong for any other policy,
2870 * if one of the OBDs has no more objects left)
2871 * - could be sum if we stripe whole objects
2872 * - could be average, just to give a nice number
2874 * To give a "reasonable" (if not wholly accurate)
2875 * number, we divide the total number of free objects
2876 * by expected stripe count (watch out for overflow).
2878 LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
2879 LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
2884 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
2885 lov->desc.ld_default_stripe_count :
2886 lov->desc.ld_active_tgt_count;
2888 if (osfs->os_files != LOV_U64_MAX)
2889 do_div(osfs->os_files, expected_stripes);
2890 if (osfs->os_ffree != LOV_U64_MAX)
2891 do_div(osfs->os_ffree, expected_stripes);
2898 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2899 void *karg, void *uarg)
2901 struct obd_device *obddev = class_exp2obd(exp);
2902 struct lov_obd *lov = &obddev->u.lov;
2903 int i, rc, count = lov->desc.ld_tgt_count;
2904 struct obd_uuid *uuidp;
2908 case OBD_IOC_LOV_GET_CONFIG: {
2909 struct obd_ioctl_data *data = karg;
2910 struct lov_tgt_desc *tgtdesc;
2911 struct lov_desc *desc;
2917 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2920 data = (struct obd_ioctl_data *)buf;
2922 if (sizeof(*desc) > data->ioc_inllen1) {
2923 obd_ioctl_freedata(buf, len);
2927 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2928 obd_ioctl_freedata(buf, len);
2932 if (sizeof(__u32) * count > data->ioc_inllen3) {
2933 obd_ioctl_freedata(buf, len);
2937 desc = (struct lov_desc *)data->ioc_inlbuf1;
2938 memcpy(desc, &(lov->desc), sizeof(*desc));
2940 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2941 genp = (__u32 *)data->ioc_inlbuf3;
2942 tgtdesc = lov->tgts;
2943 /* the uuid will be empty for deleted OSTs */
2944 for (i = 0; i < count; i++, uuidp++, genp++, tgtdesc++) {
2945 obd_str2uuid(uuidp, tgtdesc->uuid.uuid);
2946 *genp = tgtdesc->ltd_gen;
2949 rc = copy_to_user((void *)uarg, buf, len);
2952 obd_ioctl_freedata(buf, len);
2955 case LL_IOC_LOV_SETSTRIPE:
2956 rc = lov_setstripe(exp, karg, uarg);
2958 case LL_IOC_LOV_GETSTRIPE:
2959 rc = lov_getstripe(exp, karg, uarg);
2961 case LL_IOC_LOV_SETEA:
2962 rc = lov_setea(exp, karg, uarg);
2969 for (i = 0; i < count; i++) {
2972 /* OST was deleted */
2973 if (obd_uuid_empty(&lov->tgts[i].uuid))
2976 err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
2979 if (lov->tgts[i].active) {
2980 CERROR("error: iocontrol OSC %s on OST"
2981 "idx %d: cmd %x err = %d\n",
2982 lov->tgts[i].uuid.uuid, i,
2998 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2999 void *key, __u32 *vallen, void *val)
3001 struct obd_device *obddev = class_exp2obd(exp);
3002 struct lov_obd *lov = &obddev->u.lov;
3006 if (!vallen || !val)
3009 if (keylen > strlen("lock_to_stripe") &&
3010 strcmp(key, "lock_to_stripe") == 0) {
3013 struct ldlm_lock *lock;
3014 struct lov_stripe_md *lsm;
3016 struct lov_oinfo *loi;
3017 struct ldlm_res_id *res_id = &data->lock->l_resource->lr_name;
3018 __u32 *stripe = val;
3020 if (*vallen < sizeof(*stripe))
3022 *vallen = sizeof(*stripe);
3024 /* XXX This is another one of those bits that will need to
3025 * change if we ever actually support nested LOVs. It uses
3026 * the lock's export to find out which stripe it is. */
3027 /* XXX - it's assumed all the locks for deleted OSTs have
3028 * been cancelled. Also, the export for deleted OSTs will
3029 * be NULL and won't match the lock's export. */
3030 for (i = 0, loi = data->lsm->lsm_oinfo;
3031 i < data->lsm->lsm_stripe_count;
3033 if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
3034 data->lock->l_conn_export &&
3035 loi->loi_id == res_id->name[0] &&
3036 loi->loi_gr == res_id->name[2]) {
3041 LDLM_ERROR(data->lock, "lock on inode without such object\n");
3042 dump_lsm(D_ERROR, data->lsm);
3044 } else if (keylen >= strlen("size_to_stripe") &&
3045 strcmp(key, "size_to_stripe") == 0) {
3049 struct lov_stripe_md *lsm;
3052 if (*vallen < sizeof(*data))
3055 data->size = lov_size_to_stripe(data->lsm, data->size,
3056 data->stripe_number);
3058 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
3060 int rc, size = sizeof(obd_id);
3061 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3062 if (!lov->tgts[i].active)
3064 rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key,
3070 } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
3071 struct lov_desc *desc_ret = val;
3072 *desc_ret = lov->desc;
3080 static int lov_set_info(struct obd_export *exp, obd_count keylen,
3081 void *key, obd_count vallen, void *val)
3083 struct obd_device *obddev = class_exp2obd(exp);
3084 struct lov_obd *lov = &obddev->u.lov;
3088 #define KEY_IS(str) \
3089 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
3091 if (KEY_IS("next_id")) {
3092 if (vallen != lov->desc.ld_tgt_count)
3094 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3097 /* OST was deleted */
3098 if (obd_uuid_empty(&lov->tgts[i].uuid))
3101 /* initialize all OSCs, even inactive ones */
3103 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key,
3104 sizeof(obd_id), ((obd_id*)val) + i);
3110 if (KEY_IS("growth_count")) {
3111 if (vallen != sizeof(int))
3113 } else if (KEY_IS("mds_conn")) {
3114 if (vallen != sizeof(__u32))
3116 } else if (KEY_IS("unlinked") || KEY_IS("unrecovery")) {
3119 } else if (KEY_IS("sec")) {
3120 struct lov_tgt_desc *tgt;
3121 struct obd_export *exp;
3124 spin_lock(&lov->lov_lock);
3125 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count;
3128 /* during setup time the connections to osc might
3129 * haven't been established.
3132 struct obd_device *tgt_obd;
3134 tgt_obd = class_find_client_obd(&tgt->uuid,
3138 CERROR("can't set security flavor, "
3139 "device %s not attached?\n",
3144 exp = tgt_obd->obd_self_export;
3147 err = obd_set_info(exp, keylen, key, vallen, val);
3151 spin_unlock(&lov->lov_lock);
3158 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3161 /* OST was deleted */
3162 if (obd_uuid_empty(&lov->tgts[i].uuid))
3165 if (!val && !lov->tgts[i].active)
3168 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen,
3178 /* Merge rss if @kms_only == 0
3180 * Even when merging RSS, we will take the KMS value if it's larger.
3181 * This prevents getattr from stomping on dirty cached pages which
3182 * extend the file size. */
3183 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms_only)
3185 struct lov_oinfo *loi;
3189 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3191 obd_size lov_size, tmpsize;
3193 tmpsize = loi->loi_kms;
3194 if (kms_only == 0 && loi->loi_rss > tmpsize)
3195 tmpsize = loi->loi_rss;
3197 lov_size = lov_stripe_size(lsm, tmpsize, i);
3198 if (lov_size > size)
3204 EXPORT_SYMBOL(lov_merge_size);
3207 __u64 lov_merge_blocks(struct lov_stripe_md *lsm)
3209 struct lov_oinfo *loi;
3213 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3215 blocks += loi->loi_blocks;
3219 EXPORT_SYMBOL(lov_merge_blocks);
3221 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time)
3223 struct lov_oinfo *loi;
3226 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3228 if (loi->loi_mtime > current_time)
3229 current_time = loi->loi_mtime;
3231 return current_time;
3233 EXPORT_SYMBOL(lov_merge_mtime);
3236 struct lov_multi_wait {
3237 struct ldlm_lock *lock;
3243 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
3244 struct lustre_handle *lockh)
3246 struct lov_lock_handles *lov_lockh = NULL;
3247 struct lustre_handle *lov_lockhp;
3248 struct lov_obd *lov;
3249 struct lov_oinfo *loi;
3250 struct lov_multi_wait *queues;
3254 if (lsm_bad_magic(lsm))
3257 if (!exp || !exp->exp_obd)
3260 LASSERT(lockh != NULL);
3261 if (lsm->lsm_stripe_count > 1) {
3262 lov_lockh = lov_handle2llh(lockh);
3263 if (lov_lockh == NULL) {
3264 CERROR("LOV: invalid lov lock handle %p\n", lockh);
3268 lov_lockhp = lov_lockh->llh_handles;
3273 OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
3275 GOTO(out, rc = -ENOMEM);
3277 lov = &exp->exp_obd->u.lov;
3278 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3279 i++, loi++, lov_lockhp++) {
3280 struct ldlm_lock *lock;
3281 struct obd_device *obd;
3282 unsigned long irqflags;
3284 lock = ldlm_handle2lock(lov_lockhp);
3286 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
3287 loi->loi_ost_idx, loi->loi_id);
3288 queues[i].completed = 1;
3292 queues[i].lock = lock;
3293 init_waitqueue_entry(&(queues[i].wait), current);
3294 add_wait_queue(lock->l_waitq, &(queues[i].wait));
3296 obd = class_exp2obd(lock->l_conn_export);
3298 imp = obd->u.cli.cl_import;
3300 spin_lock_irqsave(&imp->imp_lock, irqflags);
3301 queues[i].generation = imp->imp_generation;
3302 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
3306 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
3307 interrupted_completion_wait, &lwd);
3308 rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
3310 for (i = 0; i < lsm->lsm_stripe_count; i++)
3311 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
3313 if (rc == -EINTR || rc == -ETIMEDOUT) {
3319 if (lov_lockh != NULL)
3320 lov_llh_put(lov_lockh);
3325 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
3328 struct lov_oinfo *loi;
3334 stripe = lov_stripe_number(lsm, size - 1);
3335 kms = lov_size_to_stripe(lsm, size, stripe);
3336 loi = &(lsm->lsm_oinfo[stripe]);
3338 CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n",
3339 stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms);
3340 if (kms > loi->loi_kms)
3344 EXPORT_SYMBOL(lov_increase_kms);
3346 struct obd_ops lov_obd_ops = {
3347 .o_owner = THIS_MODULE,
3348 .o_attach = lov_attach,
3349 .o_detach = lov_detach,
3350 .o_setup = lov_setup,
3351 .o_cleanup = lov_cleanup,
3352 .o_process_config = lov_process_config,
3353 .o_connect = lov_connect,
3354 .o_disconnect = lov_disconnect,
3355 .o_statfs = lov_statfs,
3356 .o_packmd = lov_packmd,
3357 .o_unpackmd = lov_unpackmd,
3358 .o_revalidate_md = lov_revalidate_md,
3359 .o_create = lov_create,
3360 .o_destroy = lov_destroy,
3361 .o_getattr = lov_getattr,
3362 .o_getattr_async = lov_getattr_async,
3363 .o_setattr = lov_setattr,
3365 .o_brw_async = lov_brw_async,
3366 .o_prep_async_page = lov_prep_async_page,
3367 .o_queue_async_io = lov_queue_async_io,
3368 .o_set_async_flags = lov_set_async_flags,
3369 .o_queue_group_io = lov_queue_group_io,
3370 .o_trigger_group_io = lov_trigger_group_io,
3371 .o_teardown_async_page = lov_teardown_async_page,
3372 .o_punch = lov_punch,
3374 .o_enqueue = lov_enqueue,
3375 .o_match = lov_match,
3376 .o_change_cbdata = lov_change_cbdata,
3377 .o_cancel = lov_cancel,
3378 .o_cancel_unused = lov_cancel_unused,
3379 .o_iocontrol = lov_iocontrol,
3380 .o_get_info = lov_get_info,
3381 .o_set_info = lov_set_info,
3382 .o_llog_init = lov_llog_init,
3383 .o_llog_finish = lov_llog_finish,
3384 .o_notify = lov_notify,
3387 int __init lov_init(void)
3389 struct lprocfs_static_vars lvars;
3393 lprocfs_init_vars(lov, &lvars);
3394 rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars,
3395 OBD_LOV_DEVICENAME);
3400 static void /*__exit*/ lov_exit(void)
3402 class_unregister_type(OBD_LOV_DEVICENAME);
3405 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3406 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
3407 MODULE_LICENSE("GPL");
3409 module_init(lov_init);
3410 module_exit(lov_exit);