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.",
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.",
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 obd_device *obd;
369 struct lov_tgt_desc *tgt;
373 CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
374 lov, uuid->uuid, activate);
376 spin_lock(&lov->lov_lock);
377 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
378 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
379 i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
380 if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0)
384 if (i == lov->desc.ld_tgt_count)
385 GOTO(out, rc = -EINVAL);
387 obd = class_exp2obd(tgt->ltd_exp);
389 /* This can happen if OST failure races with node shutdown */
390 GOTO(out, rc = -ENOTCONN);
393 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LOV idx %d\n",
394 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
395 obd->obd_type->typ_name, i);
396 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_OSC_NAME) == 0);
398 if (tgt->active == activate) {
399 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
400 activate ? "" : "in");
404 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd, activate ? "" : "in");
406 tgt->active = activate;
408 lov->desc.ld_active_tgt_count++;
410 lov->desc.ld_active_tgt_count--;
414 spin_unlock(&lov->lov_lock);
418 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
419 int active, void *data)
422 struct obd_uuid *uuid;
425 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
426 CERROR("unexpected notification of %s %s!\n",
427 watched->obd_type->typ_name,
431 uuid = &watched->u.cli.cl_import->imp_target_uuid;
433 /* Set OSC as active before notifying the observer, so the
434 * observer can use the OSC normally.
436 rc = lov_set_osc_active(&obd->u.lov, uuid, active);
438 CERROR("%sactivation of %s failed: %d\n",
439 active ? "" : "de", uuid->uuid, rc);
443 if (obd->obd_observer)
444 /* Pass the notification up the chain. */
445 rc = obd_notify(obd->obd_observer, watched, active, data);
450 int lov_attach(struct obd_device *dev, obd_count len, void *data)
452 struct lprocfs_static_vars lvars;
455 lprocfs_init_vars(lov, &lvars);
456 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
459 struct proc_dir_entry *entry;
461 entry = create_proc_entry("target_obd_status", 0444,
462 dev->obd_proc_entry);
466 entry->proc_fops = &lov_proc_target_fops;
474 int lov_detach(struct obd_device *dev)
476 return lprocfs_obd_detach(dev);
479 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
481 struct lustre_cfg *lcfg = buf;
482 struct lov_desc *desc;
483 struct lov_obd *lov = &obd->u.lov;
487 if (lcfg->lcfg_inllen1 < 1) {
488 CERROR("LOV setup requires a descriptor\n");
492 desc = (struct lov_desc *)lcfg->lcfg_inlbuf1;
493 if (sizeof(*desc) > lcfg->lcfg_inllen1) {
494 CERROR("descriptor size wrong: %d > %d\n",
495 (int)sizeof(*desc), lcfg->lcfg_inllen1);
499 /* Because of 64-bit divide/mod operations only work with a 32-bit
500 * divisor in a 32-bit kernel, we cannot support a stripe width
501 * of 4GB or larger on 32-bit CPUs.
504 count = desc->ld_default_stripe_count;
505 if (count && (count * desc->ld_default_stripe_size) > ~0UL) {
506 CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n",
507 desc->ld_default_stripe_size, count, ~0UL);
511 lov->bufsize = LOV_MAX_TGT_COUNT * sizeof(struct lov_tgt_desc);
512 OBD_ALLOC(lov->tgts, lov->bufsize);
513 if (lov->tgts == NULL) {
515 CERROR("couldn't allocate %d bytes for target table.\n",
520 desc->ld_tgt_count = 0;
521 desc->ld_active_tgt_count = 0;
523 spin_lock_init(&lov->lov_lock);
524 sema_init(&lov->lov_llog_sem, 1);
529 static int lov_cleanup(struct obd_device *obd, int flags)
531 struct lov_obd *lov = &obd->u.lov;
533 OBD_FREE(lov->tgts, lov->bufsize);
538 lov_add_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
540 struct lov_obd *lov = &obd->u.lov;
541 struct lov_tgt_desc *tgt;
545 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
546 uuidp->uuid, index, gen);
548 if ((index < 0) || (index >= LOV_MAX_TGT_COUNT)) {
549 CERROR("request to add OBD %s at invalid index: %d\n",
555 CERROR("request to add OBD %s with invalid generation: %d\n",
560 tgt = lov->tgts + index;
561 if (!obd_uuid_empty(&tgt->uuid)) {
562 CERROR("OBD already assigned at LOV target index %d\n",
568 /* XXX - add a sanity check on the generation number. */
571 if (index >= lov->desc.ld_tgt_count)
572 lov->desc.ld_tgt_count = index + 1;
574 CDEBUG(D_CONFIG, "idx: %d ltd_gen: %d ld_tgt_count: %d\n",
575 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
577 if (lov->refcount == 0)
581 struct obd_device *osc_obd;
583 osc_obd = class_exp2obd(tgt->ltd_exp);
585 osc_obd->obd_no_recov = 0;
588 rc = lov_connect_obd(obd, tgt, 1, lov->lov_connect_flags);
592 if (obd->obd_observer) {
593 /* tell the mds_lov about the new target */
594 rc = obd_notify(obd->obd_observer, tgt->ltd_exp->exp_obd, 1,
600 if (rc && tgt->ltd_exp != NULL)
601 lov_disconnect_obd(obd, tgt, 0);
606 lov_del_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
608 struct lov_obd *lov = &obd->u.lov;
609 struct lov_tgt_desc *tgt;
610 int count = lov->desc.ld_tgt_count;
614 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
615 uuidp->uuid, index, gen);
617 if (index >= count) {
618 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
623 tgt = lov->tgts + index;
625 if (obd_uuid_empty(&tgt->uuid)) {
626 CERROR("LOV target at index %d is not setup.\n", index);
630 if (strncmp(uuidp->uuid, tgt->uuid.uuid, sizeof uuidp->uuid) != 0) {
631 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
632 tgt->uuid.uuid, index, uuidp->uuid);
637 struct obd_device *osc_obd;
639 osc_obd = class_exp2obd(tgt->ltd_exp);
641 osc_obd->obd_no_recov = 1;
642 rc = obd_llog_finish(osc_obd, &osc_obd->obd_llogs, 1);
644 CERROR("osc_llog_finish error: %d\n", rc);
646 lov_disconnect_obd(obd, tgt, 0);
649 /* XXX - right now there is a dependency on ld_tgt_count being the
650 * maximum tgt index for computing the mds_max_easize. So we can't
653 /* lt_gen = 0 will mean it will not match the gen of any valid loi */
654 memset(tgt, 0, sizeof(*tgt));
656 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
657 tgt->uuid.uuid, index, tgt->ltd_gen, tgt->ltd_exp, tgt->active);
662 static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
664 struct lustre_cfg *lcfg = buf;
665 struct obd_uuid obd_uuid;
672 switch(cmd = lcfg->lcfg_command) {
673 case LCFG_LOV_ADD_OBD:
674 case LCFG_LOV_DEL_OBD: {
675 if (lcfg->lcfg_inllen1 > sizeof(obd_uuid.uuid))
676 GOTO(out, rc = -EINVAL);
678 obd_str2uuid(&obd_uuid, lcfg->lcfg_inlbuf1);
680 if (sscanf(lcfg->lcfg_inlbuf2, "%d", &index) != 1)
681 GOTO(out, rc = -EINVAL);
682 if (sscanf(lcfg->lcfg_inlbuf3, "%d", &gen) != 1)
683 GOTO(out, rc = -EINVAL);
684 if (cmd == LCFG_LOV_ADD_OBD)
685 rc = lov_add_obd(obd, &obd_uuid, index, gen);
687 rc = lov_del_obd(obd, &obd_uuid, index, gen);
691 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
692 GOTO(out, rc = -EINVAL);
700 /* compute object size given "stripeno" and the ost size */
701 static obd_size lov_stripe_size(struct lov_stripe_md *lsm, obd_size ost_size,
704 unsigned long ssize = lsm->lsm_stripe_size;
705 unsigned long swidth = ssize * lsm->lsm_stripe_count;
706 unsigned long stripe_size;
712 /* do_div(a, b) returns a % b, and a = a / b */
713 stripe_size = do_div(ost_size, ssize);
715 lov_size = ost_size * swidth + stripeno * ssize + stripe_size;
717 lov_size = (ost_size - 1) * swidth + (stripeno + 1) * ssize;
722 static void lov_merge_attrs(struct obdo *tgt, struct obdo *src, obd_flag valid,
723 struct lov_stripe_md *lsm, int stripeno, int *set)
725 valid &= src->o_valid;
728 if (valid & OBD_MD_FLSIZE) {
729 /* this handles sparse files properly */
732 lov_size = lov_stripe_size(lsm, src->o_size, stripeno);
733 if (lov_size > tgt->o_size)
734 tgt->o_size = lov_size;
736 if (valid & OBD_MD_FLBLOCKS)
737 tgt->o_blocks += src->o_blocks;
738 if (valid & OBD_MD_FLBLKSZ)
739 tgt->o_blksize += src->o_blksize;
740 if (valid & OBD_MD_FLCTIME && tgt->o_ctime < src->o_ctime)
741 tgt->o_ctime = src->o_ctime;
742 if (valid & OBD_MD_FLMTIME && tgt->o_mtime < src->o_mtime)
743 tgt->o_mtime = src->o_mtime;
745 memcpy(tgt, src, sizeof(*tgt));
746 tgt->o_id = lsm->lsm_object_id;
747 if (valid & OBD_MD_FLSIZE)
748 tgt->o_size = lov_stripe_size(lsm,src->o_size,stripeno);
754 #define log2(n) ffz(~(n))
757 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
758 struct lov_stripe_md **ea,
759 struct obd_trans_info *oti)
763 struct obd_uuid *ost_uuid = NULL;
767 LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
768 src_oa->o_flags == OBD_FL_DELORPHAN);
770 lov = &export->exp_obd->u.lov;
772 tmp_oa = obdo_alloc();
776 if (src_oa->o_valid & OBD_MD_FLINLINE) {
777 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
778 CDEBUG(D_HA, "clearing orphans only for %s\n",
782 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
783 struct lov_stripe_md obj_md;
784 struct lov_stripe_md *obj_mdp = &obj_md;
787 /* if called for a specific target, we don't
788 care if it is not active. */
789 if (lov->tgts[i].active == 0 && ost_uuid == NULL) {
790 CDEBUG(D_HA, "lov idx %d inactive\n", i);
794 if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid))
797 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
799 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
800 err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, &obj_mdp, oti);
802 /* This export will be disabled until it is recovered,
803 and then orphan recovery will be completed. */
804 CERROR("error in orphan recovery on OST idx %d/%d: "
805 "rc = %d\n", i, lov->desc.ld_tgt_count, err);
814 #define LOV_CREATE_RESEED_INTERVAL 1000
816 /* the LOV expects oa->o_id to be set to the LOV object id */
817 static int lov_create(struct obd_export *exp, struct obdo *src_oa,
818 struct lov_stripe_md **ea, struct obd_trans_info *oti)
820 static int ost_start_idx, ost_start_count;
822 struct lov_stripe_md *lsm;
823 struct lov_oinfo *loi = NULL;
824 struct obdo *tmp_oa, *ret_oa;
825 struct llog_cookie *cookies = NULL;
826 unsigned ost_count, ost_idx;
827 int set = 0, obj_alloc = 0, cookie_sent = 0, rc = 0, i;
832 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
833 src_oa->o_flags == OBD_FL_DELORPHAN) {
834 rc = lov_clear_orphans(exp, src_oa, ea, oti);
841 lov = &exp->exp_obd->u.lov;
843 if (!lov->desc.ld_active_tgt_count)
846 /* Recreate a specific object id at the given OST index */
847 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
848 (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) {
849 struct lov_stripe_md obj_md;
850 struct lov_stripe_md *obj_mdp = &obj_md;
852 ost_idx = src_oa->o_nlink;
856 if (ost_idx >= lov->desc.ld_tgt_count)
858 for (i = 0; i < lsm->lsm_stripe_count; i++) {
859 if (lsm->lsm_oinfo[i].loi_ost_idx == ost_idx) {
860 if (lsm->lsm_oinfo[i].loi_id != src_oa->o_id ||
861 lsm->lsm_oinfo[i].loi_gr != src_oa->o_gr) {
867 if (i == lsm->lsm_stripe_count)
870 rc = obd_create(lov->tgts[ost_idx].ltd_exp, src_oa,
875 ret_oa = obdo_alloc();
879 tmp_oa = obdo_alloc();
881 GOTO(out_oa, rc = -ENOMEM);
886 ost_count = lov_get_stripecnt(lov, 0);
888 /* If the MDS file was truncated up to some size, stripe over
889 * enough OSTs to allow the file to be created at that size. */
890 if (src_oa->o_valid & OBD_MD_FLSIZE) {
891 stripes=((src_oa->o_size+LUSTRE_STRIPE_MAXBYTES)>>12)-1;
892 do_div(stripes, (__u32)(LUSTRE_STRIPE_MAXBYTES >> 12));
894 if (stripes > lov->desc.ld_active_tgt_count)
896 if (stripes < ost_count)
902 rc = lov_alloc_memmd(&lsm, stripes, lov->desc.ld_pattern ?
903 lov->desc.ld_pattern : LOV_PATTERN_RAID0);
910 ost_count = lov->desc.ld_tgt_count;
912 LASSERT(src_oa->o_gr > 0);
913 LASSERT(src_oa->o_valid & OBD_MD_FLID);
914 lsm->lsm_object_id = src_oa->o_id;
915 lsm->lsm_object_gr = src_oa->o_gr;
916 if (!lsm->lsm_stripe_size)
917 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
918 if (!lsm->lsm_pattern) {
919 lsm->lsm_pattern = lov->desc.ld_pattern ?
920 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
923 if (*ea == NULL || lsm->lsm_oinfo[0].loi_ost_idx >= ost_count) {
924 if (--ost_start_count <= 0) {
925 ost_start_idx = ll_insecure_random_int();
926 ost_start_count = LOV_CREATE_RESEED_INTERVAL;
927 } else if (lsm->lsm_stripe_count >=
928 lov->desc.ld_active_tgt_count) {
929 /* If we allocate from all of the stripes, make the
930 * next file start on the next OST. */
933 ost_idx = ost_start_idx % ost_count;
935 ost_idx = lsm->lsm_oinfo[0].loi_ost_idx;
938 CDEBUG(D_INODE, "allocating %d subobjs for objid "LPX64" at idx %d\n",
939 lsm->lsm_stripe_count, lsm->lsm_object_id, ost_idx);
941 /* XXX LOV STACKING: need to figure out how many real OSCs */
942 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
943 oti_alloc_cookies(oti, lsm->lsm_stripe_count);
944 if (!oti->oti_logcookies)
945 GOTO(out_cleanup, rc = -ENOMEM);
946 cookies = oti->oti_logcookies;
949 loi = lsm->lsm_oinfo;
950 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
951 struct lov_stripe_md obj_md;
952 struct lov_stripe_md *obj_mdp = &obj_md;
956 if (lov->tgts[ost_idx].active == 0) {
957 if (!obd_uuid_empty(&lov->tgts[ost_idx].uuid))
958 CDEBUG(D_HA, "lov idx %d inactive\n", ost_idx);
962 /* create data objects with "parent" OA */
963 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
965 /* XXX When we start creating objects on demand, we need to
966 * make sure that we always create the object on the
967 * stripe which holds the existing file size.
969 if (src_oa->o_valid & OBD_MD_FLSIZE) {
970 if (lov_stripe_offset(lsm, src_oa->o_size, i,
971 &tmp_oa->o_size) < 0 &&
975 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
976 i, tmp_oa->o_size, src_oa->o_size);
980 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
981 err = obd_create(lov->tgts[ost_idx].ltd_exp, tmp_oa, &obj_mdp,
984 if (lov->tgts[ost_idx].active) {
985 CERROR("error creating objid "LPX64" sub-object"
986 " on OST idx %d/%d: rc = %d\n",
987 src_oa->o_id, ost_idx,
988 lsm->lsm_stripe_count, err);
990 CERROR("obd_create returned invalid "
1000 oti->oti_objid[ost_idx] = tmp_oa->o_id;
1001 loi->loi_id = tmp_oa->o_id;
1002 loi->loi_gr = tmp_oa->o_gr;
1003 loi->loi_ost_idx = ost_idx;
1004 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
1005 CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64" at "
1006 "idx %d gen %d\n", lsm->lsm_object_id, loi->loi_id,
1007 ost_idx, loi->loi_ost_gen);
1009 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1014 ++oti->oti_logcookies;
1015 if (tmp_oa->o_valid & OBD_MD_FLCOOKIE)
1020 /* If we have allocated enough objects, we are OK */
1021 if (obj_alloc == lsm->lsm_stripe_count)
1022 GOTO(out_done, rc = 0);
1025 if (obj_alloc == 0) {
1028 GOTO(out_cleanup, rc);
1031 /* If we were passed specific striping params, then a failure to
1032 * meet those requirements is an error, since we can't reallocate
1033 * that memory (it might be part of a larger array or something).
1035 * We can only get here if lsm_stripe_count was originally > 1.
1038 CERROR("can't lstripe objid "LPX64": have %u want %u, rc %d\n",
1039 lsm->lsm_object_id, obj_alloc, lsm->lsm_stripe_count,rc);
1042 GOTO(out_cleanup, rc);
1044 struct lov_stripe_md *lsm_new;
1045 /* XXX LOV STACKING call into osc for sizes */
1046 unsigned oldsize, newsize;
1048 if (oti && cookies && cookie_sent) {
1049 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
1050 newsize = obj_alloc * sizeof(*cookies);
1052 oti_alloc_cookies(oti, obj_alloc);
1053 if (oti->oti_logcookies) {
1054 memcpy(oti->oti_logcookies, cookies, newsize);
1055 OBD_FREE(cookies, oldsize);
1056 cookies = oti->oti_logcookies;
1058 CWARN("'leaking' %d bytes\n", oldsize-newsize);
1062 CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
1063 lsm->lsm_object_id, lsm->lsm_stripe_count, obj_alloc);
1064 oldsize = lov_stripe_md_size(lsm->lsm_stripe_count);
1065 newsize = lov_stripe_md_size(obj_alloc);
1066 OBD_ALLOC(lsm_new, newsize);
1067 if (lsm_new != NULL) {
1068 memcpy(lsm_new, lsm, newsize);
1069 lsm_new->lsm_stripe_count = obj_alloc;
1070 OBD_FREE(lsm, oldsize);
1073 CWARN("'leaking' %d bytes\n", oldsize - newsize);
1080 if (src_oa->o_valid & OBD_MD_FLSIZE &&
1081 ret_oa->o_size != src_oa->o_size) {
1082 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
1083 src_oa->o_size, ret_oa->o_size);
1086 ret_oa->o_id = src_oa->o_id;
1087 ret_oa->o_gr = src_oa->o_gr;
1088 ret_oa->o_valid |= OBD_MD_FLGROUP;
1089 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1095 if (oti && cookies) {
1096 oti->oti_logcookies = cookies;
1098 oti_free_cookies(oti);
1099 src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
1101 src_oa->o_valid |= OBD_MD_FLCOOKIE;
1107 while (obj_alloc-- > 0) {
1108 struct obd_export *sub_exp;
1112 sub_exp = lov->tgts[loi->loi_ost_idx].ltd_exp;
1113 /* destroy already created objects here */
1114 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1115 tmp_oa->o_id = loi->loi_id;
1117 err = obd_destroy(sub_exp, tmp_oa, NULL, oti);
1119 CERROR("Failed to uncreate objid "LPX64" subobj "LPX64
1120 " on OST idx %d: rc = %d\n", src_oa->o_id,
1121 loi->loi_id, loi->loi_ost_idx, err);
1124 obd_free_memmd(exp, &lsm);
1128 static int lov_revalidate_policy(struct lov_obd *lov, struct lov_stripe_md *lsm)
1130 static int next_idx = 0;
1131 struct lov_tgt_desc *tgt;
1134 /* XXX - we should do something clever and take lsm
1135 * into account but just do round robin for now. */
1137 /* last_idx must always be less that count because
1138 * ld_tgt_count currently cannot shrink. */
1139 count = lov->desc.ld_tgt_count;
1141 for (i = next_idx, tgt = lov->tgts + i; i < count; i++, tgt++) {
1143 next_idx = (i + 1) % count;
1148 for (i = 0, tgt = lov->tgts; i < next_idx; i++, tgt++) {
1150 next_idx = (i + 1) % count;
1158 #define lsm_bad_magic(LSMP) \
1160 struct lov_stripe_md *_lsm__ = (LSMP); \
1163 CERROR("LOV requires striping ea\n"); \
1165 } else if (_lsm__->lsm_magic != LOV_MAGIC) { \
1166 CERROR("LOV striping magic bad %#x != %#x\n", \
1167 _lsm__->lsm_magic, LOV_MAGIC); \
1173 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
1174 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1177 struct lov_obd *lov;
1178 struct lov_oinfo *loi;
1182 if (lsm_bad_magic(lsm))
1185 if (!exp || !exp->exp_obd)
1188 lov = &exp->exp_obd->u.lov;
1189 loi = lsm->lsm_oinfo;
1190 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1193 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1194 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1195 /* Orphan clean up will (someday) fix this up. */
1196 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
1197 oti->oti_logcookies++;
1201 memcpy(&tmp, oa, sizeof(tmp));
1202 tmp.o_id = loi->loi_id;
1203 err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
1205 if (err && lov->tgts[loi->loi_ost_idx].active) {
1206 CDEBUG(D_INODE, "error: destroying objid "LPX64" subobj "
1207 LPX64" on OST idx %d: rc = %d\n",
1208 oa->o_id, loi->loi_id, loi->loi_ost_idx, err);
1216 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
1217 struct lov_stripe_md *lsm)
1220 struct lov_obd *lov;
1221 struct lov_oinfo *loi;
1222 int i, rc = 0, set = 0;
1225 if (lsm_bad_magic(lsm))
1228 if (!exp || !exp->exp_obd)
1231 lov = &exp->exp_obd->u.lov;
1233 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1234 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1235 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1238 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1239 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1243 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1244 "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
1245 /* create data objects with "parent" OA */
1246 memcpy(&tmp, oa, sizeof(tmp));
1247 tmp.o_id = loi->loi_id;
1249 err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
1252 if (lov->tgts[loi->loi_ost_idx].active) {
1253 CERROR("error: getattr objid "LPX64" subobj "
1254 LPX64" on OST idx %d: rc = %d\n",
1255 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1260 lov_merge_attrs(oa, &tmp, tmp.o_valid, lsm, i, &set);
1268 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
1271 struct lov_getattr_async_args *aa = data;
1272 struct lov_stripe_md *lsm = aa->aa_lsm;
1273 struct obdo *oa = aa->aa_oa;
1274 struct obdo *obdos = aa->aa_obdos;
1275 struct lov_oinfo *loi;
1281 /* NB all stripe requests succeeded to get here */
1283 loi = lsm->lsm_oinfo;
1284 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1285 if (obdos[i].o_valid == 0) /* inactive stripe */
1288 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1293 CERROR ("No stripes had valid attrs\n");
1298 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1302 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
1303 struct lov_stripe_md *lsm,
1304 struct ptlrpc_request_set *rqset)
1307 struct lov_obd *lov;
1308 struct lov_oinfo *loi;
1309 struct lov_getattr_async_args *aa;
1310 int i, rc = 0, set = 0;
1313 if (lsm_bad_magic(lsm))
1316 if (!exp || !exp->exp_obd)
1319 lov = &exp->exp_obd->u.lov;
1321 OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1325 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1326 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1327 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1330 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1331 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1332 /* leaves obdos[i].obd_valid unset */
1336 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at "
1337 "idx %u gen %d\n", oa->o_id, i, loi->loi_id,
1338 loi->loi_ost_idx, loi->loi_ost_gen);
1340 /* create data objects with "parent" OA */
1341 memcpy(&obdos[i], oa, sizeof(obdos[i]));
1342 obdos[i].o_id = loi->loi_id;
1344 err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp,
1345 &obdos[i], NULL, rqset);
1347 CERROR("error: getattr objid "LPX64" subobj "
1348 LPX64" on OST idx %d: rc = %d\n",
1349 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1351 GOTO(out_obdos, rc = err);
1356 GOTO (out_obdos, rc = -EIO);
1358 LASSERT (rqset->set_interpret == NULL);
1359 rqset->set_interpret = lov_getattr_interpret;
1360 LASSERT (sizeof (rqset->set_args) >= sizeof (*aa));
1361 aa = (struct lov_getattr_async_args *)&rqset->set_args;
1364 aa->aa_obdos = obdos;
1369 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1375 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
1376 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1378 struct obdo *tmp_oa, *ret_oa;
1379 struct lov_obd *lov;
1380 struct lov_oinfo *loi;
1381 int rc = 0, i, set = 0;
1384 if (lsm_bad_magic(lsm))
1387 if (!exp || !exp->exp_obd)
1390 /* for now, we only expect time updates here */
1391 LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
1392 OBD_MD_FLATIME | OBD_MD_FLMTIME |
1393 OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
1394 OBD_MD_FLSIZE | OBD_MD_FLGROUP)));
1396 LASSERT(!(src_oa->o_valid & OBD_MD_FLGROUP) || src_oa->o_gr > 0);
1398 ret_oa = obdo_alloc();
1402 tmp_oa = obdo_alloc();
1404 GOTO(out_oa, rc = -ENOMEM);
1406 lov = &exp->exp_obd->u.lov;
1407 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1410 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1411 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1415 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1416 tmp_oa->o_id = loi->loi_id;
1417 LASSERT(!(tmp_oa->o_valid & OBD_MD_FLGROUP) || tmp_oa->o_gr>0);
1419 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1420 if (lov_stripe_offset(lsm, src_oa->o_size, i,
1421 &tmp_oa->o_size) < 0 &&
1425 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1426 i, tmp_oa->o_size, src_oa->o_size);
1429 err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa,
1432 if (lov->tgts[loi->loi_ost_idx].active) {
1433 CERROR("error: setattr objid "LPX64" subobj "
1434 LPX64" on OST idx %d: rc = %d\n",
1435 src_oa->o_id, loi->loi_id,
1436 loi->loi_ost_idx, err);
1442 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set);
1447 ret_oa->o_id = src_oa->o_id;
1448 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1457 /* we have an offset in file backed by an lov and want to find out where
1458 * that offset lands in our given stripe of the file. for the easy
1459 * case where the offset is within the stripe, we just have to scale the
1460 * offset down to make it relative to the stripe instead of the lov.
1462 * the harder case is what to do when the offset doesn't intersect the
1463 * stripe. callers will want start offsets clamped ahead to the start
1464 * of the nearest stripe in the file. end offsets similarly clamped to the
1465 * nearest ending byte of a stripe in the file:
1467 * all this function does is move offsets to the nearest region of the
1468 * stripe, and it does its work "mod" the full length of all the stripes.
1469 * consider a file with 3 stripes:
1472 * ---------------------------------------------------------------------
1473 * | 0 | 1 | 2 | 0 | 1 | 2 |
1474 * ---------------------------------------------------------------------
1476 * to find stripe 1's offsets for S and E, it divides by the full stripe
1477 * width and does its math in the context of a single set of stripes:
1480 * -----------------------------------
1482 * -----------------------------------
1484 * it'll notice that E is outside stripe 1 and clamp it to the end of the
1485 * stripe, then multiply it back out by lov_off to give the real offsets in
1489 * ---------------------------------------------------------------------
1490 * | 1 | 1 | 1 | 1 | 1 | 1 |
1491 * ---------------------------------------------------------------------
1493 * it would have done similarly and pulled S forward to the start of a 1
1494 * stripe if, say, S had landed in a 0 stripe.
1496 * this rounding isn't always correct. consider an E lov offset that lands
1497 * on a 0 stripe, the "mod stripe width" math will pull it forward to the
1498 * start of a 1 stripe, when in fact it wanted to be rounded back to the end
1499 * of a previous 1 stripe. this logic is handled by callers and this is why:
1501 * this function returns < 0 when the offset was "before" the stripe and
1502 * was moved forward to the start of the stripe in question; 0 when it
1503 * falls in the stripe and no shifting was done; > 0 when the offset
1504 * was outside the stripe and was pulled back to its final byte. */
1505 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
1506 int stripeno, obd_off *obd_off)
1508 unsigned long ssize = lsm->lsm_stripe_size;
1509 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1510 unsigned long stripe_off, this_stripe;
1513 if (lov_off == OBD_OBJECT_EOF) {
1514 *obd_off = OBD_OBJECT_EOF;
1518 /* do_div(a, b) returns a % b, and a = a / b */
1519 stripe_off = do_div(lov_off, swidth);
1521 this_stripe = stripeno * ssize;
1522 if (stripe_off < this_stripe) {
1526 stripe_off -= this_stripe;
1528 if (stripe_off >= ssize) {
1534 *obd_off = lov_off * ssize + stripe_off;
1538 /* Given a whole-file size and a stripe number, give the file size which
1539 * corresponds to the individual object of that stripe.
1541 * This behaves basically in the same was as lov_stripe_offset, except that
1542 * file sizes falling before the beginning of a stripe are clamped to the end
1543 * of the previous stripe, not the beginning of the next:
1546 * ---------------------------------------------------------------------
1547 * | 0 | 1 | 2 | 0 | 1 | 2 |
1548 * ---------------------------------------------------------------------
1550 * if clamped to stripe 2 becomes:
1553 * ---------------------------------------------------------------------
1554 * | 0 | 1 | 2 | 0 | 1 | 2 |
1555 * ---------------------------------------------------------------------
1557 static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size,
1560 unsigned long ssize = lsm->lsm_stripe_size;
1561 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1562 unsigned long stripe_off, this_stripe;
1564 if (file_size == OBD_OBJECT_EOF)
1565 return OBD_OBJECT_EOF;
1567 /* do_div(a, b) returns a % b, and a = a / b */
1568 stripe_off = do_div(file_size, swidth);
1570 this_stripe = stripeno * ssize;
1571 if (stripe_off < this_stripe) {
1572 /* Move to end of previous stripe, or zero */
1573 if (file_size > 0) {
1580 stripe_off -= this_stripe;
1582 if (stripe_off >= ssize) {
1583 /* Clamp to end of this stripe */
1588 return (file_size * ssize + stripe_off);
1591 /* given an extent in an lov and a stripe, calculate the extent of the stripe
1592 * that is contained within the lov extent. this returns true if the given
1593 * stripe does intersect with the lov extent. */
1594 static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno,
1595 obd_off start, obd_off end,
1596 obd_off *obd_start, obd_off *obd_end)
1598 int start_side = 0, end_side = 0;
1600 switch (lsm->lsm_pattern) {
1601 case LOV_PATTERN_RAID0:
1602 start_side = lov_stripe_offset(lsm, start, stripeno, obd_start);
1603 end_side = lov_stripe_offset(lsm, end, stripeno, obd_end);
1605 case LOV_PATTERN_CMOBD:
1608 start_side = end_side = 0;
1614 CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n",
1615 start, end, start_side, *obd_start, *obd_end, end_side);
1617 /* this stripe doesn't intersect the file extent when neither
1618 * start or the end intersected the stripe and obd_start and
1619 * obd_end got rounded up to the save value. */
1620 if (start_side != 0 && end_side != 0 && *obd_start == *obd_end)
1623 /* as mentioned in the lov_stripe_offset commentary, end
1624 * might have been shifted in the wrong direction. This
1625 * happens when an end offset is before the stripe when viewed
1626 * through the "mod stripe size" math. we detect it being shifted
1627 * in the wrong direction and touch it up.
1628 * interestingly, this can't underflow since end must be > start
1629 * if we passed through the previous check.
1630 * (should we assert for that somewhere?) */
1637 /* compute which stripe number "lov_off" will be written into */
1638 static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off)
1640 unsigned long ssize = lsm->lsm_stripe_size;
1641 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1642 unsigned long stripe_off;
1644 if (lsm->lsm_pattern == LOV_PATTERN_CMOBD)
1647 stripe_off = do_div(lov_off, swidth);
1649 return stripe_off / ssize;
1652 static int lov_revalidate_md(struct obd_export *exp, struct obdo *src_oa,
1653 struct lov_stripe_md *ea,
1654 struct obd_trans_info *oti)
1656 struct obd_export *osc_exp;
1657 struct lov_obd *lov = &exp->exp_obd->u.lov;
1658 struct lov_stripe_md *lsm = ea;
1659 struct lov_stripe_md obj_md;
1660 struct lov_stripe_md *obj_mdp = &obj_md;
1661 struct lov_oinfo *loi;
1662 struct obdo *tmp_oa;
1663 int ost_idx, updates = 0, i;
1666 tmp_oa = obdo_alloc();
1670 loi = lsm->lsm_oinfo;
1671 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1673 if (!obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
1676 ost_idx = lov_revalidate_policy(lov, lsm);
1678 /* FIXME: punt for now. */
1679 CERROR("lov_revalidate_policy failed; no active "
1684 /* create a new object */
1685 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1686 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
1687 osc_exp = lov->tgts[ost_idx].ltd_exp;
1688 rc = obd_create(osc_exp, tmp_oa, &obj_mdp, oti);
1690 CERROR("error creating new subobj at idx %d; "
1691 "rc = %d\n", ost_idx, rc);
1695 oti->oti_objid[ost_idx] = tmp_oa->o_id;
1696 loi->loi_id = tmp_oa->o_id;
1697 loi->loi_ost_idx = ost_idx;
1698 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
1699 CDEBUG(D_INODE, "replacing objid "LPX64" subobj "LPX64
1700 " with idx %d gen %d.\n", lsm->lsm_object_id,
1701 loi->loi_id, ost_idx, loi->loi_ost_gen);
1705 /* If we got an error revalidating an entry there's no need to
1706 * cleanup up objects we allocated here because the bad entry
1707 * still points to a deleted OST. */
1713 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1714 * we can send this 'punch' to just the authoritative node and the nodes
1715 * that the punch will affect. */
1716 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1717 struct lov_stripe_md *lsm,
1718 obd_off start, obd_off end, struct obd_trans_info *oti)
1721 struct lov_obd *lov;
1722 struct lov_oinfo *loi;
1726 if (lsm_bad_magic(lsm))
1729 if (!exp || !exp->exp_obd)
1732 lov = &exp->exp_obd->u.lov;
1733 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1734 obd_off starti, endi;
1737 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1738 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1742 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1745 /* create data objects with "parent" OA */
1746 memcpy(&tmp, oa, sizeof(tmp));
1747 tmp.o_id = loi->loi_id;
1749 err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, NULL,
1750 starti, endi, NULL);
1752 if (lov->tgts[loi->loi_ost_idx].active) {
1753 CERROR("error: punch objid "LPX64" subobj "LPX64
1754 " on OST idx %d: rc = %d\n", oa->o_id,
1755 loi->loi_id, loi->loi_ost_idx, err);
1760 loi->loi_kms = loi->loi_rss = starti;
1766 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1767 struct lov_stripe_md *lsm, obd_off start, obd_off end)
1770 struct lov_obd *lov;
1771 struct lov_oinfo *loi;
1775 if (lsm_bad_magic(lsm))
1785 lov = &exp->exp_obd->u.lov;
1786 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1787 obd_off starti, endi;
1790 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1791 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1795 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1798 memcpy(tmp, oa, sizeof(*tmp));
1799 tmp->o_id = loi->loi_id;
1801 err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL,
1804 if (lov->tgts[loi->loi_ost_idx].active) {
1805 CERROR("error: fsync objid "LPX64" subobj "LPX64
1806 " on OST idx %d: rc = %d\n", oa->o_id,
1807 loi->loi_id, loi->loi_ost_idx, err);
1818 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1819 struct lov_stripe_md *lsm,
1820 obd_count oa_bufs, struct brw_page *pga)
1825 /* The caller just wants to know if there's a chance that this
1826 * I/O can succeed */
1827 for (i = 0; i < oa_bufs; i++) {
1828 int stripe = lov_stripe_number(lsm, pga[i].off);
1829 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1832 if (!lov_stripe_intersects(lsm, i, pga[i].off,
1833 pga[i].off + pga[i].count,
1837 if (lov->tgts[ost].active == 0) {
1838 CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1841 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[ost].ltd_exp, oa,
1842 NULL, 1, &pga[i], NULL);
1849 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1850 struct lov_stripe_md *lsm, obd_count oa_bufs,
1851 struct brw_page *pga, struct obd_trans_info *oti)
1857 struct lov_stripe_md lsm;
1859 } *stripeinfo, *si, *si_last;
1860 struct obdo *ret_oa = NULL, *tmp_oa = NULL;
1861 struct lov_obd *lov;
1862 struct brw_page *ioarr;
1863 struct lov_oinfo *loi;
1864 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0;
1867 if (lsm_bad_magic(lsm))
1870 lov = &exp->exp_obd->u.lov;
1872 if (cmd == OBD_BRW_CHECK) {
1873 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1877 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1881 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1883 GOTO(out_sinfo, rc = -ENOMEM);
1885 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1887 GOTO(out_where, rc = -ENOMEM);
1890 ret_oa = obdo_alloc();
1892 GOTO(out_ioarr, rc = -ENOMEM);
1894 tmp_oa = obdo_alloc();
1896 GOTO(out_oa, rc = -ENOMEM);
1899 for (i = 0; i < oa_bufs; i++) {
1900 where[i] = lov_stripe_number(lsm, pga[i].off);
1901 stripeinfo[where[i]].bufct++;
1904 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1905 i < stripe_count; i++, loi++, si_last = si, si++) {
1907 si->index = si_last->index + si_last->bufct;
1908 si->lsm.lsm_object_id = loi->loi_id;
1909 si->lsm.lsm_object_gr = lsm->lsm_object_gr;
1910 si->ost_idx = loi->loi_ost_idx;
1913 for (i = 0; i < oa_bufs; i++) {
1914 int which = where[i];
1917 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1918 LASSERT(shift < oa_bufs);
1919 ioarr[shift] = pga[i];
1920 lov_stripe_offset(lsm, pga[i].off, which,
1922 stripeinfo[which].subcount++;
1925 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1926 int shift = si->index;
1928 if (lov->tgts[si->ost_idx].active == 0) {
1929 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1930 GOTO(out_oa, rc = -EIO);
1934 LASSERT(shift < oa_bufs);
1936 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1938 tmp_oa->o_id = si->lsm.lsm_object_id;
1939 rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp,
1940 tmp_oa, &si->lsm, si->bufct,
1941 &ioarr[shift], oti);
1943 GOTO(out_ioarr, rc);
1945 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1950 ret_oa->o_id = src_oa->o_id;
1951 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1960 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1962 OBD_FREE(where, sizeof(*where) * oa_bufs);
1964 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1968 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1971 struct lov_brw_async_args *aa = data;
1972 struct lov_stripe_md *lsm = aa->aa_lsm;
1973 obd_count oa_bufs = aa->aa_oa_bufs;
1974 struct obdo *oa = aa->aa_oa;
1975 struct obdo *obdos = aa->aa_obdos;
1976 struct brw_page *ioarr = aa->aa_ioarr;
1977 struct lov_oinfo *loi;
1982 /* NB all stripe requests succeeded to get here */
1984 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1986 if (obdos[i].o_valid == 0) /* inactive stripe */
1989 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1994 CERROR("No stripes had valid attrs\n");
1998 oa->o_id = lsm->lsm_object_id;
2000 OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos));
2001 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
2005 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
2006 struct lov_stripe_md *lsm, obd_count oa_bufs,
2007 struct brw_page *pga, struct ptlrpc_request_set *set,
2008 struct obd_trans_info *oti)
2014 struct lov_stripe_md lsm;
2016 } *stripeinfo, *si, *si_last;
2017 struct lov_obd *lov;
2018 struct brw_page *ioarr;
2019 struct obdo *obdos = NULL;
2020 struct lov_oinfo *loi;
2021 struct lov_brw_async_args *aa;
2022 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count;
2025 if (lsm_bad_magic(lsm))
2028 lov = &exp->exp_obd->u.lov;
2030 if (cmd == OBD_BRW_CHECK) {
2031 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
2035 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
2039 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
2041 GOTO(out_sinfo, rc = -ENOMEM);
2044 OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count);
2046 GOTO(out_where, rc = -ENOMEM);
2049 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
2051 GOTO(out_obdos, rc = -ENOMEM);
2053 for (i = 0; i < oa_bufs; i++) {
2054 where[i] = lov_stripe_number(lsm, pga[i].off);
2055 stripeinfo[where[i]].bufct++;
2058 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
2059 i < stripe_count; i++, loi++, si_last = si, si++) {
2061 si->index = si_last->index + si_last->bufct;
2062 si->lsm.lsm_object_id = loi->loi_id;
2063 si->ost_idx = loi->loi_ost_idx;
2066 memcpy(&obdos[i], oa, sizeof(*obdos));
2067 obdos[i].o_id = si->lsm.lsm_object_id;
2071 for (i = 0; i < oa_bufs; i++) {
2072 int which = where[i];
2075 shift = stripeinfo[which].index + stripeinfo[which].subcount;
2076 LASSERT(shift < oa_bufs);
2077 ioarr[shift] = pga[i];
2078 lov_stripe_offset(lsm, pga[i].off, which,
2080 stripeinfo[which].subcount++;
2083 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
2084 int shift = si->index;
2089 if (lov->tgts[si->ost_idx].active == 0) {
2090 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
2091 GOTO(out_ioarr, rc = -EIO);
2094 LASSERT(shift < oa_bufs);
2096 rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp,
2097 &obdos[i], &si->lsm, si->bufct,
2098 &ioarr[shift], set, oti);
2100 GOTO(out_ioarr, rc);
2103 LASSERT(set->set_interpret == NULL);
2104 set->set_interpret = (set_interpreter_func)lov_brw_interpret;
2105 LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args));
2106 aa = (struct lov_brw_async_args *)&set->set_args;
2108 aa->aa_obdos = obdos;
2110 aa->aa_ioarr = ioarr;
2111 aa->aa_oa_bufs = oa_bufs;
2113 /* Don't free ioarr or obdos - that's done in lov_brw_interpret */
2114 GOTO(out_where, rc);
2117 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
2119 OBD_FREE(obdos, stripe_count * sizeof(*obdos));
2121 OBD_FREE(where, sizeof(*where) * oa_bufs);
2123 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
2127 struct lov_async_page *lap_from_cookie(void *cookie)
2129 struct lov_async_page *lap = cookie;
2130 if (lap->lap_magic != LAP_MAGIC)
2131 return ERR_PTR(-EINVAL);
2135 static int lov_ap_make_ready(void *data, int cmd)
2137 struct lov_async_page *lap = lap_from_cookie(data);
2138 /* XXX should these assert? */
2142 return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
2144 static int lov_ap_refresh_count(void *data, int cmd)
2146 struct lov_async_page *lap = lap_from_cookie(data);
2150 return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
2153 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
2155 struct lov_async_page *lap = lap_from_cookie(data);
2156 /* XXX should these assert? */
2160 lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
2161 /* XXX woah, shouldn't we be altering more here? size? */
2162 oa->o_id = lap->lap_loi_id;
2165 static void lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
2167 struct lov_async_page *lap = lap_from_cookie(data);
2171 /* in a raid1 regime this would down a count of many ios
2172 * in flight, onl calling the caller_ops completion when all
2173 * the raid1 ios are complete */
2174 lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, oa, rc);
2177 static struct obd_async_page_ops lov_async_page_ops = {
2178 .ap_make_ready = lov_ap_make_ready,
2179 .ap_refresh_count = lov_ap_refresh_count,
2180 .ap_fill_obdo = lov_ap_fill_obdo,
2181 .ap_completion = lov_ap_completion,
2184 static int lov_prep_async_page(struct obd_export *exp,
2185 struct lov_stripe_md *lsm,
2186 struct lov_oinfo *loi, struct page *page,
2187 obd_off offset, struct obd_async_page_ops *ops,
2188 void *data, void **res)
2190 struct lov_obd *lov = &exp->exp_obd->u.lov;
2191 struct lov_async_page *lap;
2195 if (lsm_bad_magic(lsm))
2197 LASSERT(loi == NULL);
2199 stripe = lov_stripe_number(lsm, offset);
2200 loi = &lsm->lsm_oinfo[stripe];
2202 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
2204 if (lov->tgts[loi->loi_ost_idx].active == 0)
2206 if (lov->tgts[loi->loi_ost_idx].ltd_exp == NULL) {
2207 CERROR("ltd_exp == NULL, but OST idx %d doesn't appear to be "
2208 "deleted or inactive.\n", loi->loi_ost_idx);
2212 OBD_ALLOC(lap, sizeof(*lap));
2216 lap->lap_magic = LAP_MAGIC;
2217 lap->lap_caller_ops = ops;
2218 lap->lap_caller_data = data;
2220 /* FIXME handle multiple oscs after landing b_raid1 */
2221 lap->lap_stripe = stripe;
2222 switch (lsm->lsm_pattern) {
2223 case LOV_PATTERN_RAID0:
2224 lov_stripe_offset(lsm, offset, lap->lap_stripe,
2225 &lap->lap_sub_offset);
2227 case LOV_PATTERN_CMOBD:
2228 lap->lap_sub_offset = offset;
2234 /* so the callback doesn't need the lsm */
2235 lap->lap_loi_id = loi->loi_id;
2237 rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2238 lsm, loi, page, lap->lap_sub_offset,
2239 &lov_async_page_ops, lap,
2240 &lap->lap_sub_cookie);
2242 OBD_FREE(lap, sizeof(*lap));
2245 CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
2246 lap->lap_sub_cookie, offset);
2251 static int lov_queue_async_io(struct obd_export *exp,
2252 struct lov_stripe_md *lsm,
2253 struct lov_oinfo *loi, void *cookie,
2254 int cmd, obd_off off, int count,
2255 obd_flag brw_flags, obd_flag async_flags)
2257 struct lov_obd *lov = &exp->exp_obd->u.lov;
2258 struct lov_async_page *lap;
2261 LASSERT(loi == NULL);
2263 if (lsm_bad_magic(lsm))
2266 lap = lap_from_cookie(cookie);
2268 RETURN(PTR_ERR(lap));
2270 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2272 rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
2273 loi, lap->lap_sub_cookie, cmd, off, count,
2274 brw_flags, async_flags);
2278 static int lov_set_async_flags(struct obd_export *exp,
2279 struct lov_stripe_md *lsm,
2280 struct lov_oinfo *loi, void *cookie,
2281 obd_flag async_flags)
2283 struct lov_obd *lov = &exp->exp_obd->u.lov;
2284 struct lov_async_page *lap;
2287 LASSERT(loi == NULL);
2289 if (lsm_bad_magic(lsm))
2292 lap = lap_from_cookie(cookie);
2294 RETURN(PTR_ERR(lap));
2296 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2298 rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
2299 lsm, loi, lap->lap_sub_cookie, async_flags);
2303 static int lov_queue_group_io(struct obd_export *exp,
2304 struct lov_stripe_md *lsm,
2305 struct lov_oinfo *loi,
2306 struct obd_io_group *oig, void *cookie,
2307 int cmd, obd_off off, int count,
2308 obd_flag brw_flags, obd_flag async_flags)
2310 struct lov_obd *lov = &exp->exp_obd->u.lov;
2311 struct lov_async_page *lap;
2314 LASSERT(loi == NULL);
2316 if (lsm_bad_magic(lsm))
2319 lap = lap_from_cookie(cookie);
2321 RETURN(PTR_ERR(lap));
2323 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2325 rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
2326 oig, lap->lap_sub_cookie, cmd, off, count,
2327 brw_flags, async_flags);
2331 /* this isn't exactly optimal. we may have queued sync io in oscs on
2332 * all stripes, but we don't record that fact at queue time. so we
2333 * trigger sync io on all stripes. */
2334 static int lov_trigger_group_io(struct obd_export *exp,
2335 struct lov_stripe_md *lsm,
2336 struct lov_oinfo *loi,
2337 struct obd_io_group *oig)
2339 struct lov_obd *lov = &exp->exp_obd->u.lov;
2342 LASSERT(loi == NULL);
2344 if (lsm_bad_magic(lsm))
2347 loi = lsm->lsm_oinfo;
2348 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
2349 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2350 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2354 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
2356 if (rc == 0 && err != 0)
2362 static int lov_teardown_async_page(struct obd_export *exp,
2363 struct lov_stripe_md *lsm,
2364 struct lov_oinfo *loi, void *cookie)
2366 struct lov_obd *lov = &exp->exp_obd->u.lov;
2367 struct lov_async_page *lap;
2370 LASSERT(loi == NULL);
2372 if (lsm_bad_magic(lsm))
2375 lap = lap_from_cookie(cookie);
2377 RETURN(PTR_ERR(lap));
2379 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2381 rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2382 lsm, loi, lap->lap_sub_cookie);
2384 CERROR("unable to teardown sub cookie %p: %d\n",
2385 lap->lap_sub_cookie, rc);
2388 OBD_FREE(lap, sizeof(*lap));
2392 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2393 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2394 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2395 void *data,__u32 lvb_len, void *lvb_swabber,
2396 struct lustre_handle *lockh)
2398 struct lov_lock_handles *lov_lockh = NULL;
2399 struct lustre_handle *lov_lockhp;
2400 struct lov_obd *lov;
2401 struct lov_oinfo *loi;
2402 char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)];
2403 struct lov_stripe_md *submd = (void *)submd_buf;
2405 int i, save_flags = *flags;
2408 if (lsm_bad_magic(lsm))
2411 /* we should never be asked to replay a lock this way. */
2412 LASSERT((*flags & LDLM_FL_REPLAY) == 0);
2414 if (!exp || !exp->exp_obd)
2417 if (lsm->lsm_stripe_count > 1) {
2418 lov_lockh = lov_llh_new(lsm);
2419 if (lov_lockh == NULL)
2422 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2423 lov_lockhp = lov_lockh->llh_handles;
2428 lov = &exp->exp_obd->u.lov;
2429 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2430 i++, loi++, lov_lockhp++) {
2431 ldlm_policy_data_t sub_ext;
2434 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2435 policy->l_extent.end, &start,
2439 sub_ext.l_extent.start = start;
2440 sub_ext.l_extent.end = end;
2441 sub_ext.l_extent.gid = policy->l_extent.gid;
2443 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2444 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2448 /* XXX LOV STACKING: submd should be from the subobj */
2449 submd->lsm_object_id = loi->loi_id;
2450 submd->lsm_object_gr = lsm->lsm_object_gr;
2451 submd->lsm_stripe_count = 0;
2452 submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid;
2453 submd->lsm_oinfo->loi_rss = loi->loi_rss;
2454 submd->lsm_oinfo->loi_kms = loi->loi_kms;
2455 submd->lsm_oinfo->loi_blocks = loi->loi_blocks;
2456 loi->loi_mtime = submd->lsm_oinfo->loi_mtime;
2457 /* XXX submd is not fully initialized here */
2458 *flags = save_flags;
2459 rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd,
2460 type, &sub_ext, mode, flags, bl_cb, cp_cb,
2461 gl_cb, data, lvb_len, lvb_swabber, lov_lockhp);
2463 /* XXX FIXME: This unpleasantness doesn't belong here at *all*.
2464 * It belongs in the OSC, except that the OSC doesn't have
2465 * access to the real LOI -- it gets a copy, that we created
2466 * above, and that copy can be arbitrarily out of date.
2468 * The LOV API is due for a serious rewriting anyways, and this
2469 * can be addressed then. */
2470 if (rc == ELDLM_OK) {
2471 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2472 __u64 tmp = submd->lsm_oinfo->loi_rss;
2474 LASSERT(lock != NULL);
2476 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2477 /* Extend KMS up to the end of this lock and no further
2478 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2479 if (tmp > lock->l_policy_data.l_extent.end)
2480 tmp = lock->l_policy_data.l_extent.end + 1;
2481 if (tmp >= loi->loi_kms) {
2482 CDEBUG(D_INODE, "lock acquired, setting rss="
2483 LPU64", kms="LPU64"\n", loi->loi_rss,
2486 loi->loi_kms_valid = 1;
2488 CDEBUG(D_INODE, "lock acquired, setting rss="
2489 LPU64"; leaving kms="LPU64", end="LPU64
2490 "\n", loi->loi_rss, loi->loi_kms,
2491 lock->l_policy_data.l_extent.end);
2493 ldlm_lock_allow_match(lock);
2494 LDLM_LOCK_PUT(lock);
2495 } else if (rc == ELDLM_LOCK_ABORTED &&
2496 save_flags & LDLM_FL_HAS_INTENT) {
2497 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2498 loi->loi_rss = submd->lsm_oinfo->loi_rss;
2499 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2500 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2501 " kms="LPU64"\n", loi->loi_rss, loi->loi_kms);
2503 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2504 if (lov->tgts[loi->loi_ost_idx].active) {
2505 CERROR("error: enqueue objid "LPX64" subobj "
2506 LPX64" on OST idx %d: rc = %d\n",
2507 lsm->lsm_object_id, loi->loi_id,
2508 loi->loi_ost_idx, rc);
2509 GOTO(out_locks, rc);
2513 if (lsm->lsm_stripe_count > 1)
2514 lov_llh_put(lov_lockh);
2518 while (loi--, lov_lockhp--, i-- > 0) {
2519 struct lov_stripe_md submd;
2522 if (lov_lockhp->cookie == 0)
2525 /* XXX LOV STACKING: submd should be from the subobj */
2526 submd.lsm_object_id = loi->loi_id;
2527 submd.lsm_object_gr = lsm->lsm_object_gr;
2528 submd.lsm_stripe_count = 0;
2529 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2531 if (err && lov->tgts[loi->loi_ost_idx].active) {
2532 CERROR("error: cancelling objid "LPX64" on OST "
2533 "idx %d after enqueue error: rc = %d\n",
2534 loi->loi_id, loi->loi_ost_idx, err);
2538 if (lsm->lsm_stripe_count > 1) {
2539 lov_llh_destroy(lov_lockh);
2540 lov_llh_put(lov_lockh);
2545 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2546 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2547 int *flags, void *data, struct lustre_handle *lockh)
2549 struct lov_lock_handles *lov_lockh = NULL;
2550 struct lustre_handle *lov_lockhp;
2551 struct lov_obd *lov;
2552 struct lov_oinfo *loi;
2553 struct lov_stripe_md submd;
2554 ldlm_error_t rc = 0;
2558 if (lsm_bad_magic(lsm))
2561 if (!exp || !exp->exp_obd)
2564 if (lsm->lsm_stripe_count > 1) {
2565 lov_lockh = lov_llh_new(lsm);
2566 if (lov_lockh == NULL)
2569 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2570 lov_lockhp = lov_lockh->llh_handles;
2575 lov = &exp->exp_obd->u.lov;
2576 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2577 i++, loi++, lov_lockhp++) {
2578 ldlm_policy_data_t sub_ext;
2582 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2583 policy->l_extent.end, &start, &end))
2586 sub_ext.l_extent.start = start;
2587 sub_ext.l_extent.end = end;
2589 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid)) {
2590 CDEBUG(D_HA, "lov idx %d deleted\n", loi->loi_ost_idx);
2593 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2594 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2599 /* XXX LOV STACKING: submd should be from the subobj */
2600 submd.lsm_object_id = loi->loi_id;
2601 submd.lsm_object_gr = lsm->lsm_object_gr;
2602 submd.lsm_stripe_count = 0;
2604 /* XXX submd is not fully initialized here */
2605 rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2606 type, &sub_ext, mode, &lov_flags, data,
2612 if (lsm->lsm_stripe_count > 1) {
2613 if (*flags & LDLM_FL_TEST_LOCK)
2614 lov_llh_destroy(lov_lockh);
2615 lov_llh_put(lov_lockh);
2620 while (loi--, lov_lockhp--, i-- > 0) {
2621 struct lov_stripe_md submd;
2624 if (lov_lockhp->cookie == 0)
2627 /* XXX LOV STACKING: submd should be from the subobj */
2628 submd.lsm_object_id = loi->loi_id;
2629 submd.lsm_object_gr = lsm->lsm_object_gr;
2630 submd.lsm_stripe_count = 0;
2631 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2633 if (err && lov->tgts[loi->loi_ost_idx].active) {
2634 CERROR("error: cancelling objid "LPX64" on OST "
2635 "idx %d after match failure: rc = %d\n",
2636 loi->loi_id, loi->loi_ost_idx, err);
2640 if (lsm->lsm_stripe_count > 1) {
2641 lov_llh_destroy(lov_lockh);
2642 lov_llh_put(lov_lockh);
2647 static int lov_change_cbdata(struct obd_export *exp,
2648 struct lov_stripe_md *lsm, ldlm_iterator_t it,
2651 struct lov_obd *lov;
2652 struct lov_oinfo *loi;
2656 if (lsm_bad_magic(lsm))
2659 if (!exp || !exp->exp_obd)
2662 LASSERT(lsm->lsm_object_gr > 0);
2664 lov = &exp->exp_obd->u.lov;
2665 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2666 struct lov_stripe_md submd;
2667 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2668 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2672 submd.lsm_object_id = loi->loi_id;
2673 submd.lsm_object_gr = lsm->lsm_object_gr;
2674 submd.lsm_stripe_count = 0;
2675 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
2681 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
2682 __u32 mode, struct lustre_handle *lockh)
2684 struct lov_lock_handles *lov_lockh = NULL;
2685 struct lustre_handle *lov_lockhp;
2686 struct lov_obd *lov;
2687 struct lov_oinfo *loi;
2691 if (lsm_bad_magic(lsm))
2694 if (!exp || !exp->exp_obd)
2697 LASSERT(lsm->lsm_object_gr > 0);
2700 if (lsm->lsm_stripe_count > 1) {
2701 lov_lockh = lov_handle2llh(lockh);
2703 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2707 lov_lockhp = lov_lockh->llh_handles;
2712 lov = &exp->exp_obd->u.lov;
2713 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2714 i++, loi++, lov_lockhp++) {
2715 struct lov_stripe_md submd;
2718 if (lov_lockhp->cookie == 0) {
2719 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2720 loi->loi_ost_idx, loi->loi_id);
2724 /* XXX LOV STACKING: submd should be from the subobj */
2725 submd.lsm_object_id = loi->loi_id;
2726 submd.lsm_object_gr = lsm->lsm_object_gr;
2727 submd.lsm_stripe_count = 0;
2728 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2731 if (lov->tgts[loi->loi_ost_idx].active) {
2732 CERROR("error: cancel objid "LPX64" subobj "
2733 LPX64" on OST idx %d: rc = %d\n",
2735 loi->loi_id, loi->loi_ost_idx, err);
2742 if (lsm->lsm_stripe_count > 1)
2743 lov_llh_destroy(lov_lockh);
2744 if (lov_lockh != NULL)
2745 lov_llh_put(lov_lockh);
2749 static int lov_cancel_unused(struct obd_export *exp,
2750 struct lov_stripe_md *lsm, int flags, void *opaque)
2752 struct lov_obd *lov;
2753 struct lov_oinfo *loi;
2757 lov = &exp->exp_obd->u.lov;
2759 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2760 int err = obd_cancel_unused(lov->tgts[i].ltd_exp, NULL,
2768 if (lsm_bad_magic(lsm))
2771 if (!exp || !exp->exp_obd)
2774 LASSERT(lsm->lsm_object_gr > 0);
2776 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2777 struct lov_stripe_md submd;
2780 if (lov->tgts[loi->loi_ost_idx].active == 0)
2781 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2783 submd.lsm_object_id = loi->loi_id;
2784 submd.lsm_object_gr = lsm->lsm_object_gr;
2785 submd.lsm_stripe_count = 0;
2786 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
2787 &submd, flags, opaque);
2788 if (err && lov->tgts[loi->loi_ost_idx].active) {
2789 CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2790 " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2791 loi->loi_id, loi->loi_ost_idx, err);
2799 #define LOV_U64_MAX ((__u64)~0ULL)
2800 #define LOV_SUM_MAX(tot, add) \
2802 if ((tot) + (add) < (tot)) \
2803 (tot) = LOV_U64_MAX; \
2808 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2809 unsigned long max_age)
2811 struct lov_obd *lov = &obd->u.lov;
2812 struct obd_statfs lov_sfs;
2819 /* We only get block data from the OBD */
2820 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2822 if (!lov->tgts[i].active) {
2823 CDEBUG(D_HA, "lov idx %d inactive\n", i);
2827 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
2830 if (lov->tgts[i].active && !rc)
2836 memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
2839 osfs->os_bfree += lov_sfs.os_bfree;
2840 osfs->os_bavail += lov_sfs.os_bavail;
2841 osfs->os_blocks += lov_sfs.os_blocks;
2842 /* XXX not sure about this one - depends on policy.
2843 * - could be minimum if we always stripe on all OBDs
2844 * (but that would be wrong for any other policy,
2845 * if one of the OBDs has no more objects left)
2846 * - could be sum if we stripe whole objects
2847 * - could be average, just to give a nice number
2849 * To give a "reasonable" (if not wholly accurate)
2850 * number, we divide the total number of free objects
2851 * by expected stripe count (watch out for overflow).
2853 LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
2854 LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
2859 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
2860 lov->desc.ld_default_stripe_count :
2861 lov->desc.ld_active_tgt_count;
2863 if (osfs->os_files != LOV_U64_MAX)
2864 do_div(osfs->os_files, expected_stripes);
2865 if (osfs->os_ffree != LOV_U64_MAX)
2866 do_div(osfs->os_ffree, expected_stripes);
2873 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2874 void *karg, void *uarg)
2876 struct obd_device *obddev = class_exp2obd(exp);
2877 struct lov_obd *lov = &obddev->u.lov;
2878 int i, rc, count = lov->desc.ld_tgt_count;
2879 struct obd_uuid *uuidp;
2883 case OBD_IOC_LOV_GET_CONFIG: {
2884 struct obd_ioctl_data *data = karg;
2885 struct lov_tgt_desc *tgtdesc;
2886 struct lov_desc *desc;
2892 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2895 data = (struct obd_ioctl_data *)buf;
2897 if (sizeof(*desc) > data->ioc_inllen1) {
2898 obd_ioctl_freedata(buf, len);
2902 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2903 obd_ioctl_freedata(buf, len);
2907 if (sizeof(__u32) * count > data->ioc_inllen3) {
2908 obd_ioctl_freedata(buf, len);
2912 desc = (struct lov_desc *)data->ioc_inlbuf1;
2913 memcpy(desc, &(lov->desc), sizeof(*desc));
2915 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2916 genp = (__u32 *)data->ioc_inlbuf3;
2917 tgtdesc = lov->tgts;
2918 /* the uuid will be empty for deleted OSTs */
2919 for (i = 0; i < count; i++, uuidp++, genp++, tgtdesc++) {
2920 obd_str2uuid(uuidp, tgtdesc->uuid.uuid);
2921 *genp = tgtdesc->ltd_gen;
2924 rc = copy_to_user((void *)uarg, buf, len);
2927 obd_ioctl_freedata(buf, len);
2930 case LL_IOC_LOV_SETSTRIPE:
2931 rc = lov_setstripe(exp, karg, uarg);
2933 case LL_IOC_LOV_GETSTRIPE:
2934 rc = lov_getstripe(exp, karg, uarg);
2936 case LL_IOC_LOV_SETEA:
2937 rc = lov_setea(exp, karg, uarg);
2944 for (i = 0; i < count; i++) {
2947 /* OST was deleted */
2948 if (obd_uuid_empty(&lov->tgts[i].uuid))
2951 err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
2954 if (lov->tgts[i].active) {
2955 CERROR("error: iocontrol OSC %s on OST"
2956 "idx %d: err = %d\n",
2957 lov->tgts[i].uuid.uuid, i, err);
2972 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2973 void *key, __u32 *vallen, void *val)
2975 struct obd_device *obddev = class_exp2obd(exp);
2976 struct lov_obd *lov = &obddev->u.lov;
2980 if (!vallen || !val)
2983 if (keylen > strlen("lock_to_stripe") &&
2984 strcmp(key, "lock_to_stripe") == 0) {
2987 struct ldlm_lock *lock;
2988 struct lov_stripe_md *lsm;
2990 struct lov_oinfo *loi;
2991 struct ldlm_res_id *res_id = &data->lock->l_resource->lr_name;
2992 __u32 *stripe = val;
2994 if (*vallen < sizeof(*stripe))
2996 *vallen = sizeof(*stripe);
2998 /* XXX This is another one of those bits that will need to
2999 * change if we ever actually support nested LOVs. It uses
3000 * the lock's export to find out which stripe it is. */
3001 /* XXX - it's assumed all the locks for deleted OSTs have
3002 * been cancelled. Also, the export for deleted OSTs will
3003 * be NULL and won't match the lock's export. */
3004 for (i = 0, loi = data->lsm->lsm_oinfo;
3005 i < data->lsm->lsm_stripe_count;
3007 if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
3008 data->lock->l_conn_export &&
3009 loi->loi_id == res_id->name[0] &&
3010 loi->loi_gr == res_id->name[2]) {
3015 LDLM_ERROR(data->lock, "lock on inode without such object\n");
3016 dump_lsm(D_ERROR, data->lsm);
3018 } else if (keylen >= strlen("size_to_stripe") &&
3019 strcmp(key, "size_to_stripe") == 0) {
3023 struct lov_stripe_md *lsm;
3026 if (*vallen < sizeof(*data))
3029 data->size = lov_size_to_stripe(data->lsm, data->size,
3030 data->stripe_number);
3032 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
3034 int rc, size = sizeof(obd_id);
3035 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3036 if (!lov->tgts[i].active)
3038 rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key,
3044 } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
3045 struct lov_desc *desc_ret = val;
3046 *desc_ret = lov->desc;
3054 static int lov_set_info(struct obd_export *exp, obd_count keylen,
3055 void *key, obd_count vallen, void *val)
3057 struct obd_device *obddev = class_exp2obd(exp);
3058 struct lov_obd *lov = &obddev->u.lov;
3062 #define KEY_IS(str) \
3063 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
3065 if (KEY_IS("next_id")) {
3066 if (vallen != lov->desc.ld_tgt_count)
3068 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3071 /* OST was deleted */
3072 if (obd_uuid_empty(&lov->tgts[i].uuid))
3075 /* initialize all OSCs, even inactive ones */
3077 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key,
3078 sizeof(obd_id), ((obd_id*)val) + i);
3084 if (KEY_IS("growth_count")) {
3085 if (vallen != sizeof(int))
3087 } else if (KEY_IS("mds_conn")) {
3088 if (vallen != sizeof(__u32))
3090 } else if (KEY_IS("unlinked") || KEY_IS("unrecovery")) {
3097 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3100 /* OST was deleted */
3101 if (obd_uuid_empty(&lov->tgts[i].uuid))
3104 if (!val && !lov->tgts[i].active)
3107 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen,
3117 /* Merge rss if kms == 0
3119 * Even when merging RSS, we will take the KMS value if it's larger.
3120 * This prevents getattr from stomping on dirty cached pages which
3121 * extend the file size. */
3122 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms)
3124 struct lov_oinfo *loi;
3128 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3130 obd_size lov_size, tmpsize;
3132 tmpsize = loi->loi_kms;
3133 if (kms == 0 && loi->loi_rss > tmpsize)
3134 tmpsize = loi->loi_rss;
3136 lov_size = lov_stripe_size(lsm, tmpsize, i);
3137 if (lov_size > size)
3143 EXPORT_SYMBOL(lov_merge_size);
3146 __u64 lov_merge_blocks(struct lov_stripe_md *lsm)
3148 struct lov_oinfo *loi;
3152 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3154 blocks += loi->loi_blocks;
3158 EXPORT_SYMBOL(lov_merge_blocks);
3160 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time)
3162 struct lov_oinfo *loi;
3165 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3167 if (loi->loi_mtime > current_time)
3168 current_time = loi->loi_mtime;
3170 return current_time;
3172 EXPORT_SYMBOL(lov_merge_mtime);
3175 struct lov_multi_wait {
3176 struct ldlm_lock *lock;
3182 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
3183 struct lustre_handle *lockh)
3185 struct lov_lock_handles *lov_lockh = NULL;
3186 struct lustre_handle *lov_lockhp;
3187 struct lov_obd *lov;
3188 struct lov_oinfo *loi;
3189 struct lov_multi_wait *queues;
3193 if (lsm_bad_magic(lsm))
3196 if (!exp || !exp->exp_obd)
3199 LASSERT(lockh != NULL);
3200 if (lsm->lsm_stripe_count > 1) {
3201 lov_lockh = lov_handle2llh(lockh);
3202 if (lov_lockh == NULL) {
3203 CERROR("LOV: invalid lov lock handle %p\n", lockh);
3207 lov_lockhp = lov_lockh->llh_handles;
3212 OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
3214 GOTO(out, rc = -ENOMEM);
3216 lov = &exp->exp_obd->u.lov;
3217 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3218 i++, loi++, lov_lockhp++) {
3219 struct ldlm_lock *lock;
3220 struct obd_device *obd;
3221 unsigned long irqflags;
3223 lock = ldlm_handle2lock(lov_lockhp);
3225 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
3226 loi->loi_ost_idx, loi->loi_id);
3227 queues[i].completed = 1;
3231 queues[i].lock = lock;
3232 init_waitqueue_entry(&(queues[i].wait), current);
3233 add_wait_queue(lock->l_waitq, &(queues[i].wait));
3235 obd = class_exp2obd(lock->l_conn_export);
3237 imp = obd->u.cli.cl_import;
3239 spin_lock_irqsave(&imp->imp_lock, irqflags);
3240 queues[i].generation = imp->imp_generation;
3241 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
3245 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
3246 interrupted_completion_wait, &lwd);
3247 rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
3249 for (i = 0; i < lsm->lsm_stripe_count; i++)
3250 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
3252 if (rc == -EINTR || rc == -ETIMEDOUT) {
3258 if (lov_lockh != NULL)
3259 lov_llh_put(lov_lockh);
3264 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
3267 struct lov_oinfo *loi;
3273 stripe = lov_stripe_number(lsm, size - 1);
3274 kms = lov_size_to_stripe(lsm, size, stripe);
3275 loi = &(lsm->lsm_oinfo[stripe]);
3277 CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n",
3278 stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms);
3279 if (kms > loi->loi_kms)
3283 EXPORT_SYMBOL(lov_increase_kms);
3285 struct obd_ops lov_obd_ops = {
3286 .o_owner = THIS_MODULE,
3287 .o_attach = lov_attach,
3288 .o_detach = lov_detach,
3289 .o_setup = lov_setup,
3290 .o_cleanup = lov_cleanup,
3291 .o_process_config = lov_process_config,
3292 .o_connect = lov_connect,
3293 .o_disconnect = lov_disconnect,
3294 .o_statfs = lov_statfs,
3295 .o_packmd = lov_packmd,
3296 .o_unpackmd = lov_unpackmd,
3297 .o_revalidate_md = lov_revalidate_md,
3298 .o_create = lov_create,
3299 .o_destroy = lov_destroy,
3300 .o_getattr = lov_getattr,
3301 .o_getattr_async = lov_getattr_async,
3302 .o_setattr = lov_setattr,
3304 .o_brw_async = lov_brw_async,
3305 .o_prep_async_page = lov_prep_async_page,
3306 .o_queue_async_io = lov_queue_async_io,
3307 .o_set_async_flags = lov_set_async_flags,
3308 .o_queue_group_io = lov_queue_group_io,
3309 .o_trigger_group_io = lov_trigger_group_io,
3310 .o_teardown_async_page = lov_teardown_async_page,
3311 .o_punch = lov_punch,
3313 .o_enqueue = lov_enqueue,
3314 .o_match = lov_match,
3315 .o_change_cbdata = lov_change_cbdata,
3316 .o_cancel = lov_cancel,
3317 .o_cancel_unused = lov_cancel_unused,
3318 .o_iocontrol = lov_iocontrol,
3319 .o_get_info = lov_get_info,
3320 .o_set_info = lov_set_info,
3321 .o_llog_init = lov_llog_init,
3322 .o_llog_finish = lov_llog_finish,
3323 .o_notify = lov_notify,
3326 int __init lov_init(void)
3328 struct lprocfs_static_vars lvars;
3332 lprocfs_init_vars(lov, &lvars);
3333 rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars,
3334 OBD_LOV_DEVICENAME);
3339 static void /*__exit*/ lov_exit(void)
3341 class_unregister_type(OBD_LOV_DEVICENAME);
3344 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3345 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
3346 MODULE_LICENSE("GPL");
3348 module_init(lov_init);
3349 module_exit(lov_exit);