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;
216 rc = class_connect(conn, obd, cluuid);
220 exp = class_conn2export(conn);
222 /* We don't want to actually do the underlying connections more than
223 * once, so keep track. */
225 if (lov->refcount > 1) {
226 class_export_put(exp);
230 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
231 if (obd_uuid_empty(&tgt->uuid))
233 rc = lov_connect_obd(obd, tgt, 0, connect_flags);
238 class_export_put(exp);
243 struct obd_uuid uuid;
245 --lov->desc.ld_active_tgt_count;
247 /* save for CERROR below; (we know it's terminated) */
249 rc2 = obd_disconnect(tgt->ltd_exp, 0);
251 CERROR("error: LOV target %s disconnect on OST idx %d: "
252 "rc = %d\n", uuid.uuid, i, rc2);
254 class_disconnect(exp, 0);
258 static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt,
261 struct proc_dir_entry *lov_proc_dir;
262 struct obd_device *osc_obd = class_exp2obd(tgt->ltd_exp);
263 struct lov_obd *lov = &obd->u.lov;
267 lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
269 struct proc_dir_entry *osc_symlink;
271 osc_symlink = lprocfs_srch(lov_proc_dir, osc_obd->obd_name);
273 lprocfs_remove(osc_symlink);
275 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing.",
276 obd->obd_type->typ_name, obd->obd_name,
281 if (obd->obd_no_recov) {
282 /* Pass it on to our clients.
283 * XXX This should be an argument to disconnect,
284 * XXX not a back-door flag on the OBD. Ah well.
287 osc_obd->obd_no_recov = 1;
290 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
292 rc = obd_disconnect(tgt->ltd_exp, flags);
295 CERROR("Target %s disconnect error %d\n",
303 lov->desc.ld_active_tgt_count--;
310 static int lov_disconnect(struct obd_export *exp, int flags)
312 struct obd_device *obd = class_exp2obd(exp);
313 struct lov_obd *lov = &obd->u.lov;
314 struct lov_tgt_desc *tgt;
321 /* Only disconnect the underlying layers on the final disconnect. */
323 if (lov->refcount != 0)
326 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
328 lov_disconnect_obd(obd, tgt, flags);
332 rc = class_disconnect(exp, 0);
338 * -EINVAL : UUID can't be found in the LOV's target list
339 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
340 * -EBADF : The UUID is found, but the OBD is the wrong type (!)
342 static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid,
345 struct obd_device *obd;
346 struct lov_tgt_desc *tgt;
350 CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
351 lov, uuid->uuid, activate);
353 spin_lock(&lov->lov_lock);
354 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
355 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
356 i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
357 if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0)
361 if (i == lov->desc.ld_tgt_count)
362 GOTO(out, rc = -EINVAL);
364 obd = class_exp2obd(tgt->ltd_exp);
366 /* This can happen if OST failure races with node shutdown */
367 GOTO(out, rc = -ENOTCONN);
370 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LOV idx %d\n",
371 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
372 obd->obd_type->typ_name, i);
373 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_OSC_NAME) == 0);
375 if (tgt->active == activate) {
376 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
377 activate ? "" : "in");
381 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd, activate ? "" : "in");
383 tgt->active = activate;
385 lov->desc.ld_active_tgt_count++;
387 lov->desc.ld_active_tgt_count--;
391 spin_unlock(&lov->lov_lock);
395 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
399 struct obd_uuid *uuid;
401 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
402 CERROR("unexpected notification of %s %s!\n",
403 watched->obd_type->typ_name,
407 uuid = &watched->u.cli.cl_import->imp_target_uuid;
409 /* Set OSC as active before notifying the observer, so the
410 * observer can use the OSC normally.
412 rc = lov_set_osc_active(&obd->u.lov, uuid, active);
414 CERROR("%sactivation of %s failed: %d\n",
415 active ? "" : "de", uuid->uuid, rc);
419 if (obd->obd_observer)
420 /* Pass the notification up the chain. */
421 rc = obd_notify(obd->obd_observer, watched, active);
426 int lov_attach(struct obd_device *dev, obd_count len, void *data)
428 struct lprocfs_static_vars lvars;
431 lprocfs_init_vars(lov, &lvars);
432 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
435 struct proc_dir_entry *entry;
437 entry = create_proc_entry("target_obd", 0444,
438 dev->obd_proc_entry);
442 entry->proc_fops = &lov_proc_target_fops;
450 int lov_detach(struct obd_device *dev)
452 return lprocfs_obd_detach(dev);
455 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
457 struct lustre_cfg *lcfg = buf;
458 struct lov_desc *desc;
459 struct lov_obd *lov = &obd->u.lov;
463 if (lcfg->lcfg_inllen1 < 1) {
464 CERROR("LOV setup requires a descriptor\n");
468 desc = (struct lov_desc *)lcfg->lcfg_inlbuf1;
469 if (sizeof(*desc) > lcfg->lcfg_inllen1) {
470 CERROR("descriptor size wrong: %d > %d\n",
471 (int)sizeof(*desc), lcfg->lcfg_inllen1);
475 /* Because of 64-bit divide/mod operations only work with a 32-bit
476 * divisor in a 32-bit kernel, we cannot support a stripe width
477 * of 4GB or larger on 32-bit CPUs.
480 count = desc->ld_default_stripe_count;
481 if (count && (count * desc->ld_default_stripe_size) > ~0UL) {
482 CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n",
483 desc->ld_default_stripe_size, count, ~0UL);
487 if (desc->ld_tgt_count) {
488 lov->bufsize = desc->ld_tgt_count * sizeof(struct lov_tgt_desc);
489 OBD_ALLOC(lov->tgts, lov->bufsize);
490 if (lov->tgts == NULL) {
492 CERROR("couldn't allocate %d bytes for target table.\n",
496 desc->ld_tgt_count = 0;
499 desc->ld_active_tgt_count = 0;
500 CDEBUG(D_CONFIG, "tgts: %p bufsize: 0x%x\n", lov->tgts, lov->bufsize);
502 spin_lock_init(&lov->lov_lock);
503 sema_init(&lov->lov_llog_sem, 1);
508 static int lov_cleanup(struct obd_device *obd, int flags)
510 struct lov_obd *lov = &obd->u.lov;
512 OBD_FREE(lov->tgts, lov->bufsize);
517 lov_add_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
519 struct lov_obd *lov = &obd->u.lov;
520 struct lov_tgt_desc *tgt;
521 struct obd_export *exp_observer;
525 char name[32] = "CATLIST";
529 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
530 uuidp->uuid, index, gen);
533 CERROR("request to add OBD %s at invalid index: %d\n",
539 CERROR("request to add OBD %s with invalid generation: %d\n",
544 bufsize = sizeof(struct lov_tgt_desc) * (index + 1);
545 if (bufsize > lov->bufsize) {
546 OBD_ALLOC(tgt, bufsize);
548 CERROR("couldn't allocate %d bytes for new table.\n",
553 memset(tgt, 0, bufsize);
555 memcpy(tgt, lov->tgts, lov->bufsize);
556 OBD_FREE(lov->tgts, lov->bufsize);
560 lov->bufsize = bufsize;
561 CDEBUG(D_CONFIG, "tgts: %p bufsize: %d\n",
562 lov->tgts, lov->bufsize);
565 tgt = lov->tgts + index;
566 if (!obd_uuid_empty(&tgt->uuid)) {
567 CERROR("OBD already assigned at LOV target index %d\n",
573 /* XXX - add a sanity check on the generation number. */
576 old_count = lov->desc.ld_tgt_count;
577 if (index >= lov->desc.ld_tgt_count)
578 lov->desc.ld_tgt_count = index + 1;
580 CDEBUG(D_CONFIG, "idx: %d ltd_gen: %d ld_tgt_count: %d\n",
581 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
583 if (lov->refcount == 0)
587 struct obd_device *osc_obd;
589 osc_obd = class_exp2obd(tgt->ltd_exp);
591 osc_obd->obd_no_recov = 0;
594 rc = lov_connect_obd(obd, tgt, 1, 0);
595 if (rc || !obd->obd_observer)
598 /* tell the mds_lov about the new target */
599 obd_llog_finish(obd->obd_observer, &obd->obd_observer->obd_llogs,
601 obd_llog_cat_initialize(obd->obd_observer,
602 &obd->obd_observer->obd_llogs,
603 lov->desc.ld_tgt_count, name);
606 rc = obd_get_info(tgt->ltd_exp, strlen("last_id"), "last_id", &size,
611 exp_observer = obd->obd_observer->obd_self_export;
612 rc = obd_set_info(exp_observer, strlen("next_id"),"next_id", 2, params);
616 rc = lov_notify(obd, tgt->ltd_exp->exp_obd, 1);
619 if (rc && tgt->ltd_exp != NULL)
620 lov_disconnect_obd(obd, tgt, 0);
625 lov_del_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen)
627 struct lov_obd *lov = &obd->u.lov;
628 struct lov_tgt_desc *tgt;
629 int count = lov->desc.ld_tgt_count;
633 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n",
634 uuidp->uuid, index, gen);
636 if (index >= count) {
637 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
642 tgt = lov->tgts + index;
644 if (obd_uuid_empty(&tgt->uuid)) {
645 CERROR("LOV target at index %d is not setup.\n", index);
649 if (strncmp(uuidp->uuid, tgt->uuid.uuid, sizeof uuidp->uuid) != 0) {
650 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
651 tgt->uuid.uuid, index, uuidp->uuid);
656 struct obd_device *osc_obd;
658 osc_obd = class_exp2obd(tgt->ltd_exp);
660 osc_obd->obd_no_recov = 1;
661 rc = obd_llog_finish(osc_obd, &osc_obd->obd_llogs, 1);
663 CERROR("osc_llog_finish error: %d\n", rc);
665 lov_disconnect_obd(obd, tgt, 0);
668 /* XXX - right now there is a dependency on ld_tgt_count being the
669 * maximum tgt index for computing the mds_max_easize. So we can't
672 /* lt_gen = 0 will mean it will not match the gen of any valid loi */
673 memset(tgt, 0, sizeof(*tgt));
675 CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
676 tgt->uuid.uuid, index, tgt->ltd_gen, tgt->ltd_exp, tgt->active);
681 static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
683 struct lustre_cfg *lcfg = buf;
684 struct obd_uuid obd_uuid;
691 switch(cmd = lcfg->lcfg_command) {
692 case LCFG_LOV_ADD_OBD:
693 case LCFG_LOV_DEL_OBD: {
694 if (lcfg->lcfg_inllen1 > sizeof(obd_uuid.uuid))
695 GOTO(out, rc = -EINVAL);
697 obd_str2uuid(&obd_uuid, lcfg->lcfg_inlbuf1);
699 if (sscanf(lcfg->lcfg_inlbuf2, "%d", &index) != 1)
700 GOTO(out, rc = -EINVAL);
701 if (sscanf(lcfg->lcfg_inlbuf3, "%d", &gen) != 1)
702 GOTO(out, rc = -EINVAL);
703 if (cmd == LCFG_LOV_ADD_OBD)
704 rc = lov_add_obd(obd, &obd_uuid, index, gen);
706 rc = lov_del_obd(obd, &obd_uuid, index, gen);
710 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
711 GOTO(out, rc = -EINVAL);
719 /* compute object size given "stripeno" and the ost size */
720 static obd_size lov_stripe_size(struct lov_stripe_md *lsm, obd_size ost_size,
723 unsigned long ssize = lsm->lsm_stripe_size;
724 unsigned long swidth = ssize * lsm->lsm_stripe_count;
725 unsigned long stripe_size;
731 /* do_div(a, b) returns a % b, and a = a / b */
732 stripe_size = do_div(ost_size, ssize);
734 lov_size = ost_size * swidth + stripeno * ssize + stripe_size;
736 lov_size = (ost_size - 1) * swidth + (stripeno + 1) * ssize;
741 static void lov_merge_attrs(struct obdo *tgt, struct obdo *src, obd_flag valid,
742 struct lov_stripe_md *lsm, int stripeno, int *set)
744 valid &= src->o_valid;
747 if (valid & OBD_MD_FLSIZE) {
748 /* this handles sparse files properly */
751 lov_size = lov_stripe_size(lsm, src->o_size, stripeno);
752 if (lov_size > tgt->o_size)
753 tgt->o_size = lov_size;
755 if (valid & OBD_MD_FLBLOCKS)
756 tgt->o_blocks += src->o_blocks;
757 if (valid & OBD_MD_FLBLKSZ)
758 tgt->o_blksize += src->o_blksize;
759 if (valid & OBD_MD_FLCTIME && tgt->o_ctime < src->o_ctime)
760 tgt->o_ctime = src->o_ctime;
761 if (valid & OBD_MD_FLMTIME && tgt->o_mtime < src->o_mtime)
762 tgt->o_mtime = src->o_mtime;
764 memcpy(tgt, src, sizeof(*tgt));
765 tgt->o_id = lsm->lsm_object_id;
766 if (valid & OBD_MD_FLSIZE)
767 tgt->o_size = lov_stripe_size(lsm,src->o_size,stripeno);
773 #define log2(n) ffz(~(n))
776 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
777 struct lov_stripe_md **ea,
778 struct obd_trans_info *oti)
782 struct obd_uuid *ost_uuid = NULL;
786 LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
787 src_oa->o_flags == OBD_FL_DELORPHAN);
789 lov = &export->exp_obd->u.lov;
791 tmp_oa = obdo_alloc();
795 if (src_oa->o_valid & OBD_MD_FLINLINE) {
796 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
797 CDEBUG(D_HA, "clearing orphans only for %s\n",
801 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
802 struct lov_stripe_md obj_md;
803 struct lov_stripe_md *obj_mdp = &obj_md;
806 /* if called for a specific target, we don't
807 care if it is not active. */
808 if (lov->tgts[i].active == 0 && ost_uuid == NULL) {
809 CDEBUG(D_HA, "lov idx %d inactive\n", i);
813 if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid))
816 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
818 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
819 err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, &obj_mdp, oti);
821 /* This export will be disabled until it is recovered,
822 and then orphan recovery will be completed. */
823 CERROR("error in orphan recovery on OST idx %d/%d: "
824 "rc = %d\n", i, lov->desc.ld_tgt_count, err);
833 #define LOV_CREATE_RESEED_INTERVAL 1000
835 /* the LOV expects oa->o_id to be set to the LOV object id */
836 static int lov_create(struct obd_export *exp, struct obdo *src_oa,
837 struct lov_stripe_md **ea, struct obd_trans_info *oti)
839 static int ost_start_idx, ost_start_count;
841 struct lov_stripe_md *lsm;
842 struct lov_oinfo *loi = NULL;
843 struct obdo *tmp_oa, *ret_oa;
844 struct llog_cookie *cookies = NULL;
845 unsigned ost_count, ost_idx;
846 int set = 0, obj_alloc = 0, cookie_sent = 0, rc = 0, i;
851 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
852 src_oa->o_flags == OBD_FL_DELORPHAN) {
853 rc = lov_clear_orphans(exp, src_oa, ea, oti);
860 lov = &exp->exp_obd->u.lov;
862 if (!lov->desc.ld_active_tgt_count)
865 /* Recreate a specific object id at the given OST index */
866 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
867 (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) {
868 struct lov_stripe_md obj_md;
869 struct lov_stripe_md *obj_mdp = &obj_md;
871 ost_idx = src_oa->o_nlink;
875 if (ost_idx >= lov->desc.ld_tgt_count)
877 for (i = 0; i < lsm->lsm_stripe_count; i++) {
878 if (lsm->lsm_oinfo[i].loi_ost_idx == ost_idx) {
879 if (lsm->lsm_oinfo[i].loi_id != src_oa->o_id ||
880 lsm->lsm_oinfo[i].loi_gr != src_oa->o_gr) {
886 if (i == lsm->lsm_stripe_count)
889 rc = obd_create(lov->tgts[ost_idx].ltd_exp, src_oa,
894 ret_oa = obdo_alloc();
898 tmp_oa = obdo_alloc();
900 GOTO(out_oa, rc = -ENOMEM);
905 ost_count = lov_get_stripecnt(lov, 0);
907 /* If the MDS file was truncated up to some size, stripe over
908 * enough OSTs to allow the file to be created at that size. */
909 if (src_oa->o_valid & OBD_MD_FLSIZE) {
910 stripes=((src_oa->o_size+LUSTRE_STRIPE_MAXBYTES)>>12)-1;
911 do_div(stripes, (__u32)(LUSTRE_STRIPE_MAXBYTES >> 12));
913 if (stripes > lov->desc.ld_active_tgt_count)
915 if (stripes < ost_count)
921 rc = lov_alloc_memmd(&lsm, stripes, lov->desc.ld_pattern ?
922 lov->desc.ld_pattern : LOV_PATTERN_RAID0);
929 ost_count = lov->desc.ld_tgt_count;
931 LASSERT(src_oa->o_gr > 0);
932 LASSERT(src_oa->o_valid & OBD_MD_FLID);
933 lsm->lsm_object_id = src_oa->o_id;
934 lsm->lsm_object_gr = src_oa->o_gr;
935 if (!lsm->lsm_stripe_size)
936 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
937 if (!lsm->lsm_pattern) {
938 lsm->lsm_pattern = lov->desc.ld_pattern ?
939 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
942 if (*ea == NULL || lsm->lsm_oinfo[0].loi_ost_idx >= ost_count) {
943 if (--ost_start_count <= 0) {
944 ost_start_idx = ll_insecure_random_int();
945 ost_start_count = LOV_CREATE_RESEED_INTERVAL;
946 } else if (lsm->lsm_stripe_count >=
947 lov->desc.ld_active_tgt_count) {
948 /* If we allocate from all of the stripes, make the
949 * next file start on the next OST. */
952 ost_idx = ost_start_idx % ost_count;
954 ost_idx = lsm->lsm_oinfo[0].loi_ost_idx;
957 CDEBUG(D_INODE, "allocating %d subobjs for objid "LPX64" at idx %d\n",
958 lsm->lsm_stripe_count, lsm->lsm_object_id, ost_idx);
960 /* XXX LOV STACKING: need to figure out how many real OSCs */
961 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
962 oti_alloc_cookies(oti, lsm->lsm_stripe_count);
963 if (!oti->oti_logcookies)
964 GOTO(out_cleanup, rc = -ENOMEM);
965 cookies = oti->oti_logcookies;
968 loi = lsm->lsm_oinfo;
969 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
970 struct lov_stripe_md obj_md;
971 struct lov_stripe_md *obj_mdp = &obj_md;
975 if (lov->tgts[ost_idx].active == 0) {
976 if (!obd_uuid_empty(&lov->tgts[ost_idx].uuid))
977 CDEBUG(D_HA, "lov idx %d inactive\n", ost_idx);
981 /* create data objects with "parent" OA */
982 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
984 /* XXX When we start creating objects on demand, we need to
985 * make sure that we always create the object on the
986 * stripe which holds the existing file size.
988 if (src_oa->o_valid & OBD_MD_FLSIZE) {
989 if (lov_stripe_offset(lsm, src_oa->o_size, i,
990 &tmp_oa->o_size) < 0 &&
994 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
995 i, tmp_oa->o_size, src_oa->o_size);
999 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
1000 err = obd_create(lov->tgts[ost_idx].ltd_exp, tmp_oa, &obj_mdp,
1003 if (lov->tgts[ost_idx].active) {
1004 CERROR("error creating objid "LPX64" sub-object"
1005 " on OST idx %d/%d: rc = %d\n",
1006 src_oa->o_id, ost_idx,
1007 lsm->lsm_stripe_count, err);
1009 CERROR("obd_create returned invalid "
1019 oti->oti_objid[ost_idx] = tmp_oa->o_id;
1020 loi->loi_id = tmp_oa->o_id;
1021 loi->loi_gr = tmp_oa->o_gr;
1022 loi->loi_ost_idx = ost_idx;
1023 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
1024 CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64" at "
1025 "idx %d gen %d\n", lsm->lsm_object_id, loi->loi_id,
1026 ost_idx, loi->loi_ost_gen);
1028 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1033 ++oti->oti_logcookies;
1034 if (tmp_oa->o_valid & OBD_MD_FLCOOKIE)
1039 /* If we have allocated enough objects, we are OK */
1040 if (obj_alloc == lsm->lsm_stripe_count)
1041 GOTO(out_done, rc = 0);
1044 if (obj_alloc == 0) {
1047 GOTO(out_cleanup, rc);
1050 /* If we were passed specific striping params, then a failure to
1051 * meet those requirements is an error, since we can't reallocate
1052 * that memory (it might be part of a larger array or something).
1054 * We can only get here if lsm_stripe_count was originally > 1.
1057 CERROR("can't lstripe objid "LPX64": have %u want %u, rc %d\n",
1058 lsm->lsm_object_id, obj_alloc, lsm->lsm_stripe_count,rc);
1061 GOTO(out_cleanup, rc);
1063 struct lov_stripe_md *lsm_new;
1064 /* XXX LOV STACKING call into osc for sizes */
1065 unsigned oldsize, newsize;
1067 if (oti && cookies && cookie_sent) {
1068 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
1069 newsize = obj_alloc * sizeof(*cookies);
1071 oti_alloc_cookies(oti, obj_alloc);
1072 if (oti->oti_logcookies) {
1073 memcpy(oti->oti_logcookies, cookies, newsize);
1074 OBD_FREE(cookies, oldsize);
1075 cookies = oti->oti_logcookies;
1077 CWARN("'leaking' %d bytes\n", oldsize-newsize);
1081 CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
1082 lsm->lsm_object_id, lsm->lsm_stripe_count, obj_alloc);
1083 oldsize = lov_stripe_md_size(lsm->lsm_stripe_count);
1084 newsize = lov_stripe_md_size(obj_alloc);
1085 OBD_ALLOC(lsm_new, newsize);
1086 if (lsm_new != NULL) {
1087 memcpy(lsm_new, lsm, newsize);
1088 lsm_new->lsm_stripe_count = obj_alloc;
1089 OBD_FREE(lsm, oldsize);
1092 CWARN("'leaking' %d bytes\n", oldsize - newsize);
1099 if (src_oa->o_valid & OBD_MD_FLSIZE &&
1100 ret_oa->o_size != src_oa->o_size) {
1101 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
1102 src_oa->o_size, ret_oa->o_size);
1105 ret_oa->o_id = src_oa->o_id;
1106 ret_oa->o_gr = src_oa->o_gr;
1107 ret_oa->o_valid |= OBD_MD_FLGROUP;
1108 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1114 if (oti && cookies) {
1115 oti->oti_logcookies = cookies;
1117 oti_free_cookies(oti);
1118 src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
1120 src_oa->o_valid |= OBD_MD_FLCOOKIE;
1126 while (obj_alloc-- > 0) {
1127 struct obd_export *sub_exp;
1131 sub_exp = lov->tgts[loi->loi_ost_idx].ltd_exp;
1132 /* destroy already created objects here */
1133 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1134 tmp_oa->o_id = loi->loi_id;
1136 err = obd_destroy(sub_exp, tmp_oa, NULL, oti);
1138 CERROR("Failed to uncreate objid "LPX64" subobj "LPX64
1139 " on OST idx %d: rc = %d\n", src_oa->o_id,
1140 loi->loi_id, loi->loi_ost_idx, err);
1143 obd_free_memmd(exp, &lsm);
1147 static int lov_revalidate_policy(struct lov_obd *lov, struct lov_stripe_md *lsm)
1149 static int last_idx = 0;
1150 struct lov_tgt_desc *tgt;
1153 /* XXX - we should do something clever and take lsm
1154 * into account but just do round robin for now. */
1156 /* last_idx must always be less that count because
1157 * ld_tgt_count currently cannot shrink. */
1158 count = lov->desc.ld_tgt_count;
1160 for (i = last_idx, tgt = lov->tgts + i; i < count; i++, tgt++)
1162 RETURN(last_idx = i);
1164 for (i = 0, tgt = lov->tgts; i < last_idx; i++, tgt++)
1166 RETURN(last_idx = i);
1171 #define lsm_bad_magic(LSMP) \
1173 struct lov_stripe_md *_lsm__ = (LSMP); \
1176 CERROR("LOV requires striping ea\n"); \
1178 } else if (_lsm__->lsm_magic != LOV_MAGIC) { \
1179 CERROR("LOV striping magic bad %#x != %#x\n", \
1180 _lsm__->lsm_magic, LOV_MAGIC); \
1186 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
1187 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1190 struct lov_obd *lov;
1191 struct lov_oinfo *loi;
1195 if (lsm_bad_magic(lsm))
1198 if (!exp || !exp->exp_obd)
1201 lov = &exp->exp_obd->u.lov;
1202 loi = lsm->lsm_oinfo;
1203 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1206 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1207 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1208 /* Orphan clean up will (someday) fix this up. */
1209 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
1210 oti->oti_logcookies++;
1214 memcpy(&tmp, oa, sizeof(tmp));
1215 tmp.o_id = loi->loi_id;
1216 err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
1218 if (err && lov->tgts[loi->loi_ost_idx].active) {
1219 CDEBUG(D_INODE, "error: destroying objid "LPX64" subobj "
1220 LPX64" on OST idx %d: rc = %d\n",
1221 oa->o_id, loi->loi_id, loi->loi_ost_idx, err);
1229 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
1230 struct lov_stripe_md *lsm)
1233 struct lov_obd *lov;
1234 struct lov_oinfo *loi;
1235 int i, rc = 0, set = 0;
1238 if (lsm_bad_magic(lsm))
1241 if (!exp || !exp->exp_obd)
1244 lov = &exp->exp_obd->u.lov;
1246 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1247 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1248 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1251 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1252 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1256 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1257 "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
1258 /* create data objects with "parent" OA */
1259 memcpy(&tmp, oa, sizeof(tmp));
1260 tmp.o_id = loi->loi_id;
1262 err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
1265 if (lov->tgts[loi->loi_ost_idx].active) {
1266 CERROR("error: getattr objid "LPX64" subobj "
1267 LPX64" on OST idx %d: rc = %d\n",
1268 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1273 lov_merge_attrs(oa, &tmp, tmp.o_valid, lsm, i, &set);
1281 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
1284 struct lov_getattr_async_args *aa = data;
1285 struct lov_stripe_md *lsm = aa->aa_lsm;
1286 struct obdo *oa = aa->aa_oa;
1287 struct obdo *obdos = aa->aa_obdos;
1288 struct lov_oinfo *loi;
1294 /* NB all stripe requests succeeded to get here */
1296 loi = lsm->lsm_oinfo;
1297 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1298 if (obdos[i].o_valid == 0) /* inactive stripe */
1301 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1306 CERROR ("No stripes had valid attrs\n");
1311 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1315 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
1316 struct lov_stripe_md *lsm,
1317 struct ptlrpc_request_set *rqset)
1320 struct lov_obd *lov;
1321 struct lov_oinfo *loi;
1322 struct lov_getattr_async_args *aa;
1323 int i, rc = 0, set = 0;
1326 if (lsm_bad_magic(lsm))
1329 if (!exp || !exp->exp_obd)
1332 lov = &exp->exp_obd->u.lov;
1334 OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1338 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1339 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1340 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1343 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1344 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1345 /* leaves obdos[i].obd_valid unset */
1349 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at "
1350 "idx %u gen %d\n", oa->o_id, i, loi->loi_id,
1351 loi->loi_ost_idx, loi->loi_ost_gen);
1353 /* create data objects with "parent" OA */
1354 memcpy(&obdos[i], oa, sizeof(obdos[i]));
1355 obdos[i].o_id = loi->loi_id;
1357 err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp,
1358 &obdos[i], NULL, rqset);
1360 CERROR("error: getattr objid "LPX64" subobj "
1361 LPX64" on OST idx %d: rc = %d\n",
1362 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1364 GOTO(out_obdos, rc = err);
1369 GOTO (out_obdos, rc = -EIO);
1371 LASSERT (rqset->set_interpret == NULL);
1372 rqset->set_interpret = lov_getattr_interpret;
1373 LASSERT (sizeof (rqset->set_args) >= sizeof (*aa));
1374 aa = (struct lov_getattr_async_args *)&rqset->set_args;
1377 aa->aa_obdos = obdos;
1382 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1388 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
1389 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1391 struct obdo *tmp_oa, *ret_oa;
1392 struct lov_obd *lov;
1393 struct lov_oinfo *loi;
1394 int rc = 0, i, set = 0;
1397 if (lsm_bad_magic(lsm))
1400 if (!exp || !exp->exp_obd)
1403 /* for now, we only expect time updates here */
1404 LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
1405 OBD_MD_FLATIME | OBD_MD_FLMTIME |
1406 OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
1407 OBD_MD_FLSIZE | OBD_MD_FLGROUP)));
1409 LASSERT(!(src_oa->o_valid & OBD_MD_FLGROUP) || src_oa->o_gr > 0);
1411 ret_oa = obdo_alloc();
1415 tmp_oa = obdo_alloc();
1417 GOTO(out_oa, rc = -ENOMEM);
1419 lov = &exp->exp_obd->u.lov;
1420 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1423 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1424 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1428 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1429 tmp_oa->o_id = loi->loi_id;
1430 LASSERT(!(tmp_oa->o_valid & OBD_MD_FLGROUP) || tmp_oa->o_gr>0);
1432 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1433 if (lov_stripe_offset(lsm, src_oa->o_size, i,
1434 &tmp_oa->o_size) < 0 &&
1438 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1439 i, tmp_oa->o_size, src_oa->o_size);
1442 err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa,
1445 if (lov->tgts[loi->loi_ost_idx].active) {
1446 CERROR("error: setattr objid "LPX64" subobj "
1447 LPX64" on OST idx %d: rc = %d\n",
1448 src_oa->o_id, loi->loi_id,
1449 loi->loi_ost_idx, err);
1455 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set);
1460 ret_oa->o_id = src_oa->o_id;
1461 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1470 /* we have an offset in file backed by an lov and want to find out where
1471 * that offset lands in our given stripe of the file. for the easy
1472 * case where the offset is within the stripe, we just have to scale the
1473 * offset down to make it relative to the stripe instead of the lov.
1475 * the harder case is what to do when the offset doesn't intersect the
1476 * stripe. callers will want start offsets clamped ahead to the start
1477 * of the nearest stripe in the file. end offsets similarly clamped to the
1478 * nearest ending byte of a stripe in the file:
1480 * all this function does is move offsets to the nearest region of the
1481 * stripe, and it does its work "mod" the full length of all the stripes.
1482 * consider a file with 3 stripes:
1485 * ---------------------------------------------------------------------
1486 * | 0 | 1 | 2 | 0 | 1 | 2 |
1487 * ---------------------------------------------------------------------
1489 * to find stripe 1's offsets for S and E, it divides by the full stripe
1490 * width and does its math in the context of a single set of stripes:
1493 * -----------------------------------
1495 * -----------------------------------
1497 * it'll notice that E is outside stripe 1 and clamp it to the end of the
1498 * stripe, then multiply it back out by lov_off to give the real offsets in
1502 * ---------------------------------------------------------------------
1503 * | 1 | 1 | 1 | 1 | 1 | 1 |
1504 * ---------------------------------------------------------------------
1506 * it would have done similarly and pulled S forward to the start of a 1
1507 * stripe if, say, S had landed in a 0 stripe.
1509 * this rounding isn't always correct. consider an E lov offset that lands
1510 * on a 0 stripe, the "mod stripe width" math will pull it forward to the
1511 * start of a 1 stripe, when in fact it wanted to be rounded back to the end
1512 * of a previous 1 stripe. this logic is handled by callers and this is why:
1514 * this function returns < 0 when the offset was "before" the stripe and
1515 * was moved forward to the start of the stripe in question; 0 when it
1516 * falls in the stripe and no shifting was done; > 0 when the offset
1517 * was outside the stripe and was pulled back to its final byte. */
1518 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
1519 int stripeno, obd_off *obd_off)
1521 unsigned long ssize = lsm->lsm_stripe_size;
1522 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1523 unsigned long stripe_off, this_stripe;
1526 if (lov_off == OBD_OBJECT_EOF) {
1527 *obd_off = OBD_OBJECT_EOF;
1531 /* do_div(a, b) returns a % b, and a = a / b */
1532 stripe_off = do_div(lov_off, swidth);
1534 this_stripe = stripeno * ssize;
1535 if (stripe_off < this_stripe) {
1539 stripe_off -= this_stripe;
1541 if (stripe_off >= ssize) {
1547 *obd_off = lov_off * ssize + stripe_off;
1551 /* Given a whole-file size and a stripe number, give the file size which
1552 * corresponds to the individual object of that stripe.
1554 * This behaves basically in the same was as lov_stripe_offset, except that
1555 * file sizes falling before the beginning of a stripe are clamped to the end
1556 * of the previous stripe, not the beginning of the next:
1559 * ---------------------------------------------------------------------
1560 * | 0 | 1 | 2 | 0 | 1 | 2 |
1561 * ---------------------------------------------------------------------
1563 * if clamped to stripe 2 becomes:
1566 * ---------------------------------------------------------------------
1567 * | 0 | 1 | 2 | 0 | 1 | 2 |
1568 * ---------------------------------------------------------------------
1570 static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size,
1573 unsigned long ssize = lsm->lsm_stripe_size;
1574 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1575 unsigned long stripe_off, this_stripe;
1577 if (file_size == OBD_OBJECT_EOF)
1578 return OBD_OBJECT_EOF;
1580 /* do_div(a, b) returns a % b, and a = a / b */
1581 stripe_off = do_div(file_size, swidth);
1583 this_stripe = stripeno * ssize;
1584 if (stripe_off < this_stripe) {
1585 /* Move to end of previous stripe, or zero */
1586 if (file_size > 0) {
1593 stripe_off -= this_stripe;
1595 if (stripe_off >= ssize) {
1596 /* Clamp to end of this stripe */
1601 return (file_size * ssize + stripe_off);
1604 /* given an extent in an lov and a stripe, calculate the extent of the stripe
1605 * that is contained within the lov extent. this returns true if the given
1606 * stripe does intersect with the lov extent. */
1607 static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno,
1608 obd_off start, obd_off end,
1609 obd_off *obd_start, obd_off *obd_end)
1611 int start_side = 0, end_side = 0;
1613 switch (lsm->lsm_pattern) {
1614 case LOV_PATTERN_RAID0:
1615 start_side = lov_stripe_offset(lsm, start, stripeno, obd_start);
1616 end_side = lov_stripe_offset(lsm, end, stripeno, obd_end);
1618 case LOV_PATTERN_CMOBD:
1621 start_side = end_side = 0;
1627 CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n",
1628 start, end, start_side, *obd_start, *obd_end, end_side);
1630 /* this stripe doesn't intersect the file extent when neither
1631 * start or the end intersected the stripe and obd_start and
1632 * obd_end got rounded up to the save value. */
1633 if (start_side != 0 && end_side != 0 && *obd_start == *obd_end)
1636 /* as mentioned in the lov_stripe_offset commentary, end
1637 * might have been shifted in the wrong direction. This
1638 * happens when an end offset is before the stripe when viewed
1639 * through the "mod stripe size" math. we detect it being shifted
1640 * in the wrong direction and touch it up.
1641 * interestingly, this can't underflow since end must be > start
1642 * if we passed through the previous check.
1643 * (should we assert for that somewhere?) */
1650 /* compute which stripe number "lov_off" will be written into */
1651 static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off)
1653 unsigned long ssize = lsm->lsm_stripe_size;
1654 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1655 unsigned long stripe_off;
1657 if (lsm->lsm_pattern == LOV_PATTERN_CMOBD)
1660 stripe_off = do_div(lov_off, swidth);
1662 return stripe_off / ssize;
1665 static int lov_revalidate_md(struct obd_export *exp, struct obdo *src_oa,
1666 struct lov_stripe_md *ea,
1667 struct obd_trans_info *oti)
1669 struct obd_export *osc_exp;
1670 struct lov_obd *lov = &exp->exp_obd->u.lov;
1671 struct lov_stripe_md *lsm = ea;
1672 struct lov_stripe_md obj_md;
1673 struct lov_stripe_md *obj_mdp = &obj_md;
1674 struct lov_oinfo *loi;
1675 struct obdo *tmp_oa;
1676 int ost_idx, updates = 0, i;
1679 tmp_oa = obdo_alloc();
1683 loi = lsm->lsm_oinfo;
1684 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
1686 if (!obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
1689 ost_idx = lov_revalidate_policy(lov, lsm);
1691 /* FIXME: punt for now. */
1692 CERROR("lov_revalidate_policy failed; no active "
1697 /* create a new object */
1698 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1699 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
1700 osc_exp = lov->tgts[ost_idx].ltd_exp;
1701 rc = obd_create(osc_exp, tmp_oa, &obj_mdp, oti);
1703 CERROR("error creating new subobj at idx %d; "
1704 "rc = %d\n", ost_idx, rc);
1708 oti->oti_objid[ost_idx] = tmp_oa->o_id;
1709 loi->loi_id = tmp_oa->o_id;
1710 loi->loi_ost_idx = ost_idx;
1711 loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen;
1712 CDEBUG(D_INODE, "replacing objid "LPX64" subobj "LPX64
1713 " with idx %d gen %d.\n", lsm->lsm_object_id,
1714 loi->loi_id, ost_idx, loi->loi_ost_gen);
1718 /* If we got an error revalidating an entry there's no need to
1719 * cleanup up objects we allocated here because the bad entry
1720 * still points to a deleted OST. */
1726 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1727 * we can send this 'punch' to just the authoritative node and the nodes
1728 * that the punch will affect. */
1729 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1730 struct lov_stripe_md *lsm,
1731 obd_off start, obd_off end, struct obd_trans_info *oti)
1734 struct lov_obd *lov;
1735 struct lov_oinfo *loi;
1739 if (lsm_bad_magic(lsm))
1742 if (!exp || !exp->exp_obd)
1745 lov = &exp->exp_obd->u.lov;
1746 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1747 obd_off starti, endi;
1750 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1751 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1755 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1758 /* create data objects with "parent" OA */
1759 memcpy(&tmp, oa, sizeof(tmp));
1760 tmp.o_id = loi->loi_id;
1762 err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, NULL,
1763 starti, endi, NULL);
1765 if (lov->tgts[loi->loi_ost_idx].active) {
1766 CERROR("error: punch objid "LPX64" subobj "LPX64
1767 " on OST idx %d: rc = %d\n", oa->o_id,
1768 loi->loi_id, loi->loi_ost_idx, err);
1773 loi->loi_kms = loi->loi_rss = starti;
1779 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1780 struct lov_stripe_md *lsm, obd_off start, obd_off end)
1783 struct lov_obd *lov;
1784 struct lov_oinfo *loi;
1788 if (lsm_bad_magic(lsm))
1798 lov = &exp->exp_obd->u.lov;
1799 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1800 obd_off starti, endi;
1803 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1804 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1808 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1811 memcpy(tmp, oa, sizeof(*tmp));
1812 tmp->o_id = loi->loi_id;
1814 err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL,
1817 if (lov->tgts[loi->loi_ost_idx].active) {
1818 CERROR("error: fsync objid "LPX64" subobj "LPX64
1819 " on OST idx %d: rc = %d\n", oa->o_id,
1820 loi->loi_id, loi->loi_ost_idx, err);
1831 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1832 struct lov_stripe_md *lsm,
1833 obd_count oa_bufs, struct brw_page *pga)
1838 /* The caller just wants to know if there's a chance that this
1839 * I/O can succeed */
1840 for (i = 0; i < oa_bufs; i++) {
1841 int stripe = lov_stripe_number(lsm, pga[i].off);
1842 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1845 if (!lov_stripe_intersects(lsm, i, pga[i].off,
1846 pga[i].off + pga[i].count,
1850 if (lov->tgts[ost].active == 0) {
1851 CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1854 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[ost].ltd_exp, oa,
1855 NULL, 1, &pga[i], NULL);
1862 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1863 struct lov_stripe_md *lsm, obd_count oa_bufs,
1864 struct brw_page *pga, struct obd_trans_info *oti)
1870 struct lov_stripe_md lsm;
1872 } *stripeinfo, *si, *si_last;
1873 struct obdo *ret_oa = NULL, *tmp_oa = NULL;
1874 struct lov_obd *lov;
1875 struct brw_page *ioarr;
1876 struct lov_oinfo *loi;
1877 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0;
1880 if (lsm_bad_magic(lsm))
1883 lov = &exp->exp_obd->u.lov;
1885 if (cmd == OBD_BRW_CHECK) {
1886 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1890 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1894 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1896 GOTO(out_sinfo, rc = -ENOMEM);
1898 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1900 GOTO(out_where, rc = -ENOMEM);
1903 ret_oa = obdo_alloc();
1905 GOTO(out_ioarr, rc = -ENOMEM);
1907 tmp_oa = obdo_alloc();
1909 GOTO(out_oa, rc = -ENOMEM);
1912 for (i = 0; i < oa_bufs; i++) {
1913 where[i] = lov_stripe_number(lsm, pga[i].off);
1914 stripeinfo[where[i]].bufct++;
1917 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1918 i < stripe_count; i++, loi++, si_last = si, si++) {
1920 si->index = si_last->index + si_last->bufct;
1921 si->lsm.lsm_object_id = loi->loi_id;
1922 si->lsm.lsm_object_gr = lsm->lsm_object_gr;
1923 si->ost_idx = loi->loi_ost_idx;
1926 for (i = 0; i < oa_bufs; i++) {
1927 int which = where[i];
1930 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1931 LASSERT(shift < oa_bufs);
1932 ioarr[shift] = pga[i];
1933 lov_stripe_offset(lsm, pga[i].off, which,
1935 stripeinfo[which].subcount++;
1938 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1939 int shift = si->index;
1941 if (lov->tgts[si->ost_idx].active == 0) {
1942 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1943 GOTO(out_oa, rc = -EIO);
1947 LASSERT(shift < oa_bufs);
1949 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1951 tmp_oa->o_id = si->lsm.lsm_object_id;
1952 rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp,
1953 tmp_oa, &si->lsm, si->bufct,
1954 &ioarr[shift], oti);
1956 GOTO(out_ioarr, rc);
1958 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1963 ret_oa->o_id = src_oa->o_id;
1964 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1973 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1975 OBD_FREE(where, sizeof(*where) * oa_bufs);
1977 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1981 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1984 struct lov_brw_async_args *aa = data;
1985 struct lov_stripe_md *lsm = aa->aa_lsm;
1986 obd_count oa_bufs = aa->aa_oa_bufs;
1987 struct obdo *oa = aa->aa_oa;
1988 struct obdo *obdos = aa->aa_obdos;
1989 struct brw_page *ioarr = aa->aa_ioarr;
1990 struct lov_oinfo *loi;
1995 /* NB all stripe requests succeeded to get here */
1997 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1999 if (obdos[i].o_valid == 0) /* inactive stripe */
2002 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
2007 CERROR("No stripes had valid attrs\n");
2011 oa->o_id = lsm->lsm_object_id;
2013 OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos));
2014 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
2018 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
2019 struct lov_stripe_md *lsm, obd_count oa_bufs,
2020 struct brw_page *pga, struct ptlrpc_request_set *set,
2021 struct obd_trans_info *oti)
2027 struct lov_stripe_md lsm;
2029 } *stripeinfo, *si, *si_last;
2030 struct lov_obd *lov;
2031 struct brw_page *ioarr;
2032 struct obdo *obdos = NULL;
2033 struct lov_oinfo *loi;
2034 struct lov_brw_async_args *aa;
2035 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count;
2038 if (lsm_bad_magic(lsm))
2041 lov = &exp->exp_obd->u.lov;
2043 if (cmd == OBD_BRW_CHECK) {
2044 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
2048 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
2052 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
2054 GOTO(out_sinfo, rc = -ENOMEM);
2057 OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count);
2059 GOTO(out_where, rc = -ENOMEM);
2062 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
2064 GOTO(out_obdos, rc = -ENOMEM);
2066 for (i = 0; i < oa_bufs; i++) {
2067 where[i] = lov_stripe_number(lsm, pga[i].off);
2068 stripeinfo[where[i]].bufct++;
2071 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
2072 i < stripe_count; i++, loi++, si_last = si, si++) {
2074 si->index = si_last->index + si_last->bufct;
2075 si->lsm.lsm_object_id = loi->loi_id;
2076 si->ost_idx = loi->loi_ost_idx;
2079 memcpy(&obdos[i], oa, sizeof(*obdos));
2080 obdos[i].o_id = si->lsm.lsm_object_id;
2084 for (i = 0; i < oa_bufs; i++) {
2085 int which = where[i];
2088 shift = stripeinfo[which].index + stripeinfo[which].subcount;
2089 LASSERT(shift < oa_bufs);
2090 ioarr[shift] = pga[i];
2091 lov_stripe_offset(lsm, pga[i].off, which,
2093 stripeinfo[which].subcount++;
2096 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
2097 int shift = si->index;
2102 if (lov->tgts[si->ost_idx].active == 0) {
2103 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
2104 GOTO(out_ioarr, rc = -EIO);
2107 LASSERT(shift < oa_bufs);
2109 rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp,
2110 &obdos[i], &si->lsm, si->bufct,
2111 &ioarr[shift], set, oti);
2113 GOTO(out_ioarr, rc);
2116 LASSERT(set->set_interpret == NULL);
2117 set->set_interpret = (set_interpreter_func)lov_brw_interpret;
2118 LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args));
2119 aa = (struct lov_brw_async_args *)&set->set_args;
2121 aa->aa_obdos = obdos;
2123 aa->aa_ioarr = ioarr;
2124 aa->aa_oa_bufs = oa_bufs;
2126 /* Don't free ioarr or obdos - that's done in lov_brw_interpret */
2127 GOTO(out_where, rc);
2130 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
2132 OBD_FREE(obdos, stripe_count * sizeof(*obdos));
2134 OBD_FREE(where, sizeof(*where) * oa_bufs);
2136 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
2140 struct lov_async_page *lap_from_cookie(void *cookie)
2142 struct lov_async_page *lap = cookie;
2143 if (lap->lap_magic != LAP_MAGIC)
2144 return ERR_PTR(-EINVAL);
2148 static int lov_ap_make_ready(void *data, int cmd)
2150 struct lov_async_page *lap = lap_from_cookie(data);
2151 /* XXX should these assert? */
2155 return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
2157 static int lov_ap_refresh_count(void *data, int cmd)
2159 struct lov_async_page *lap = lap_from_cookie(data);
2163 return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
2166 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
2168 struct lov_async_page *lap = lap_from_cookie(data);
2169 /* XXX should these assert? */
2173 lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
2174 /* XXX woah, shouldn't we be altering more here? size? */
2175 oa->o_id = lap->lap_loi_id;
2178 static void lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
2180 struct lov_async_page *lap = lap_from_cookie(data);
2184 /* in a raid1 regime this would down a count of many ios
2185 * in flight, onl calling the caller_ops completion when all
2186 * the raid1 ios are complete */
2187 lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, oa, rc);
2190 static struct obd_async_page_ops lov_async_page_ops = {
2191 .ap_make_ready = lov_ap_make_ready,
2192 .ap_refresh_count = lov_ap_refresh_count,
2193 .ap_fill_obdo = lov_ap_fill_obdo,
2194 .ap_completion = lov_ap_completion,
2197 static int lov_prep_async_page(struct obd_export *exp,
2198 struct lov_stripe_md *lsm,
2199 struct lov_oinfo *loi, struct page *page,
2200 obd_off offset, struct obd_async_page_ops *ops,
2201 void *data, void **res)
2203 struct lov_obd *lov = &exp->exp_obd->u.lov;
2204 struct lov_async_page *lap;
2208 if (lsm_bad_magic(lsm))
2210 LASSERT(loi == NULL);
2212 stripe = lov_stripe_number(lsm, offset);
2213 loi = &lsm->lsm_oinfo[stripe];
2215 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid))
2217 if (lov->tgts[loi->loi_ost_idx].active == 0)
2219 if (lov->tgts[loi->loi_ost_idx].ltd_exp == NULL) {
2220 CERROR("ltd_exp == NULL, but OST idx %d doesn't appear to be "
2221 "deleted or inactive.\n", loi->loi_ost_idx);
2225 OBD_ALLOC(lap, sizeof(*lap));
2229 lap->lap_magic = LAP_MAGIC;
2230 lap->lap_caller_ops = ops;
2231 lap->lap_caller_data = data;
2233 /* FIXME handle multiple oscs after landing b_raid1 */
2234 lap->lap_stripe = stripe;
2235 switch (lsm->lsm_pattern) {
2236 case LOV_PATTERN_RAID0:
2237 lov_stripe_offset(lsm, offset, lap->lap_stripe,
2238 &lap->lap_sub_offset);
2240 case LOV_PATTERN_CMOBD:
2241 lap->lap_sub_offset = offset;
2247 /* so the callback doesn't need the lsm */
2248 lap->lap_loi_id = loi->loi_id;
2250 rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2251 lsm, loi, page, lap->lap_sub_offset,
2252 &lov_async_page_ops, lap,
2253 &lap->lap_sub_cookie);
2255 OBD_FREE(lap, sizeof(*lap));
2258 CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
2259 lap->lap_sub_cookie, offset);
2264 static int lov_queue_async_io(struct obd_export *exp,
2265 struct lov_stripe_md *lsm,
2266 struct lov_oinfo *loi, void *cookie,
2267 int cmd, obd_off off, int count,
2268 obd_flag brw_flags, obd_flag async_flags)
2270 struct lov_obd *lov = &exp->exp_obd->u.lov;
2271 struct lov_async_page *lap;
2274 LASSERT(loi == NULL);
2276 if (lsm_bad_magic(lsm))
2279 lap = lap_from_cookie(cookie);
2281 RETURN(PTR_ERR(lap));
2283 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2285 rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
2286 loi, lap->lap_sub_cookie, cmd, off, count,
2287 brw_flags, async_flags);
2291 static int lov_set_async_flags(struct obd_export *exp,
2292 struct lov_stripe_md *lsm,
2293 struct lov_oinfo *loi, void *cookie,
2294 obd_flag async_flags)
2296 struct lov_obd *lov = &exp->exp_obd->u.lov;
2297 struct lov_async_page *lap;
2300 LASSERT(loi == NULL);
2302 if (lsm_bad_magic(lsm))
2305 lap = lap_from_cookie(cookie);
2307 RETURN(PTR_ERR(lap));
2309 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2311 rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
2312 lsm, loi, lap->lap_sub_cookie, async_flags);
2316 static int lov_queue_group_io(struct obd_export *exp,
2317 struct lov_stripe_md *lsm,
2318 struct lov_oinfo *loi,
2319 struct obd_io_group *oig, void *cookie,
2320 int cmd, obd_off off, int count,
2321 obd_flag brw_flags, obd_flag async_flags)
2323 struct lov_obd *lov = &exp->exp_obd->u.lov;
2324 struct lov_async_page *lap;
2327 LASSERT(loi == NULL);
2329 if (lsm_bad_magic(lsm))
2332 lap = lap_from_cookie(cookie);
2334 RETURN(PTR_ERR(lap));
2336 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2338 rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
2339 oig, lap->lap_sub_cookie, cmd, off, count,
2340 brw_flags, async_flags);
2344 /* this isn't exactly optimal. we may have queued sync io in oscs on
2345 * all stripes, but we don't record that fact at queue time. so we
2346 * trigger sync io on all stripes. */
2347 static int lov_trigger_group_io(struct obd_export *exp,
2348 struct lov_stripe_md *lsm,
2349 struct lov_oinfo *loi,
2350 struct obd_io_group *oig)
2352 struct lov_obd *lov = &exp->exp_obd->u.lov;
2355 LASSERT(loi == NULL);
2357 if (lsm_bad_magic(lsm))
2360 loi = lsm->lsm_oinfo;
2361 for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) {
2362 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2363 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2367 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
2369 if (rc == 0 && err != 0)
2375 static int lov_teardown_async_page(struct obd_export *exp,
2376 struct lov_stripe_md *lsm,
2377 struct lov_oinfo *loi, void *cookie)
2379 struct lov_obd *lov = &exp->exp_obd->u.lov;
2380 struct lov_async_page *lap;
2383 LASSERT(loi == NULL);
2385 if (lsm_bad_magic(lsm))
2388 lap = lap_from_cookie(cookie);
2390 RETURN(PTR_ERR(lap));
2392 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2394 rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2395 lsm, loi, lap->lap_sub_cookie);
2397 CERROR("unable to teardown sub cookie %p: %d\n",
2398 lap->lap_sub_cookie, rc);
2401 OBD_FREE(lap, sizeof(*lap));
2405 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2406 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2407 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2408 void *data,__u32 lvb_len, void *lvb_swabber,
2409 struct lustre_handle *lockh)
2411 struct lov_lock_handles *lov_lockh = NULL;
2412 struct lustre_handle *lov_lockhp;
2413 struct lov_obd *lov;
2414 struct lov_oinfo *loi;
2415 char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)];
2416 struct lov_stripe_md *submd = (void *)submd_buf;
2418 int i, save_flags = *flags;
2421 if (lsm_bad_magic(lsm))
2424 /* we should never be asked to replay a lock this way. */
2425 LASSERT((*flags & LDLM_FL_REPLAY) == 0);
2427 if (!exp || !exp->exp_obd)
2430 if (lsm->lsm_stripe_count > 1) {
2431 lov_lockh = lov_llh_new(lsm);
2432 if (lov_lockh == NULL)
2435 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2436 lov_lockhp = lov_lockh->llh_handles;
2441 lov = &exp->exp_obd->u.lov;
2442 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2443 i++, loi++, lov_lockhp++) {
2444 ldlm_policy_data_t sub_ext;
2447 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2448 policy->l_extent.end, &start,
2452 sub_ext.l_extent.start = start;
2453 sub_ext.l_extent.end = end;
2454 sub_ext.l_extent.gid = policy->l_extent.gid;
2456 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2457 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2461 /* XXX LOV STACKING: submd should be from the subobj */
2462 submd->lsm_object_id = loi->loi_id;
2463 submd->lsm_object_gr = lsm->lsm_object_gr;
2464 submd->lsm_stripe_count = 0;
2465 submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid;
2466 submd->lsm_oinfo->loi_rss = loi->loi_rss;
2467 submd->lsm_oinfo->loi_kms = loi->loi_kms;
2468 submd->lsm_oinfo->loi_blocks = loi->loi_blocks;
2469 loi->loi_mtime = submd->lsm_oinfo->loi_mtime;
2470 /* XXX submd is not fully initialized here */
2471 *flags = save_flags;
2472 rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd,
2473 type, &sub_ext, mode, flags, bl_cb, cp_cb,
2474 gl_cb, data, lvb_len, lvb_swabber, lov_lockhp);
2476 /* XXX FIXME: This unpleasantness doesn't belong here at *all*.
2477 * It belongs in the OSC, except that the OSC doesn't have
2478 * access to the real LOI -- it gets a copy, that we created
2479 * above, and that copy can be arbitrarily out of date.
2481 * The LOV API is due for a serious rewriting anyways, and this
2482 * can be addressed then. */
2483 if (rc == ELDLM_OK) {
2484 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2485 __u64 tmp = submd->lsm_oinfo->loi_rss;
2487 LASSERT(lock != NULL);
2489 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2490 /* Extend KMS up to the end of this lock and no further
2491 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2492 if (tmp > lock->l_policy_data.l_extent.end)
2493 tmp = lock->l_policy_data.l_extent.end + 1;
2494 if (tmp >= loi->loi_kms) {
2495 CDEBUG(D_INODE, "lock acquired, setting rss="
2496 LPU64", kms="LPU64"\n", loi->loi_rss,
2499 loi->loi_kms_valid = 1;
2501 CDEBUG(D_INODE, "lock acquired, setting rss="
2502 LPU64"; leaving kms="LPU64", end="LPU64
2503 "\n", loi->loi_rss, loi->loi_kms,
2504 lock->l_policy_data.l_extent.end);
2506 ldlm_lock_allow_match(lock);
2507 LDLM_LOCK_PUT(lock);
2508 } else if (rc == ELDLM_LOCK_ABORTED &&
2509 save_flags & LDLM_FL_HAS_INTENT) {
2510 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2511 loi->loi_rss = submd->lsm_oinfo->loi_rss;
2512 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2513 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2514 " kms="LPU64"\n", loi->loi_rss, loi->loi_kms);
2516 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2517 if (lov->tgts[loi->loi_ost_idx].active) {
2518 CERROR("error: enqueue objid "LPX64" subobj "
2519 LPX64" on OST idx %d: rc = %d\n",
2520 lsm->lsm_object_id, loi->loi_id,
2521 loi->loi_ost_idx, rc);
2522 GOTO(out_locks, rc);
2526 if (lsm->lsm_stripe_count > 1)
2527 lov_llh_put(lov_lockh);
2531 while (loi--, lov_lockhp--, i-- > 0) {
2532 struct lov_stripe_md submd;
2535 if (lov_lockhp->cookie == 0)
2538 /* XXX LOV STACKING: submd should be from the subobj */
2539 submd.lsm_object_id = loi->loi_id;
2540 submd.lsm_object_gr = lsm->lsm_object_gr;
2541 submd.lsm_stripe_count = 0;
2542 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2544 if (err && lov->tgts[loi->loi_ost_idx].active) {
2545 CERROR("error: cancelling objid "LPX64" on OST "
2546 "idx %d after enqueue error: rc = %d\n",
2547 loi->loi_id, loi->loi_ost_idx, err);
2551 if (lsm->lsm_stripe_count > 1) {
2552 lov_llh_destroy(lov_lockh);
2553 lov_llh_put(lov_lockh);
2558 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2559 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2560 int *flags, void *data, struct lustre_handle *lockh)
2562 struct lov_lock_handles *lov_lockh = NULL;
2563 struct lustre_handle *lov_lockhp;
2564 struct lov_obd *lov;
2565 struct lov_oinfo *loi;
2566 struct lov_stripe_md submd;
2567 ldlm_error_t rc = 0;
2571 if (lsm_bad_magic(lsm))
2574 if (!exp || !exp->exp_obd)
2577 if (lsm->lsm_stripe_count > 1) {
2578 lov_lockh = lov_llh_new(lsm);
2579 if (lov_lockh == NULL)
2582 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2583 lov_lockhp = lov_lockh->llh_handles;
2588 lov = &exp->exp_obd->u.lov;
2589 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2590 i++, loi++, lov_lockhp++) {
2591 ldlm_policy_data_t sub_ext;
2595 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2596 policy->l_extent.end, &start, &end))
2599 sub_ext.l_extent.start = start;
2600 sub_ext.l_extent.end = end;
2602 if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid)) {
2603 CDEBUG(D_HA, "lov idx %d deleted\n", loi->loi_ost_idx);
2606 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2607 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2612 /* XXX LOV STACKING: submd should be from the subobj */
2613 submd.lsm_object_id = loi->loi_id;
2614 submd.lsm_object_gr = lsm->lsm_object_gr;
2615 submd.lsm_stripe_count = 0;
2617 /* XXX submd is not fully initialized here */
2618 rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2619 type, &sub_ext, mode, &lov_flags, data,
2625 if (lsm->lsm_stripe_count > 1) {
2626 if (*flags & LDLM_FL_TEST_LOCK)
2627 lov_llh_destroy(lov_lockh);
2628 lov_llh_put(lov_lockh);
2633 while (loi--, lov_lockhp--, i-- > 0) {
2634 struct lov_stripe_md submd;
2637 if (lov_lockhp->cookie == 0)
2640 /* XXX LOV STACKING: submd should be from the subobj */
2641 submd.lsm_object_id = loi->loi_id;
2642 submd.lsm_object_gr = lsm->lsm_object_gr;
2643 submd.lsm_stripe_count = 0;
2644 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2646 if (err && lov->tgts[loi->loi_ost_idx].active) {
2647 CERROR("error: cancelling objid "LPX64" on OST "
2648 "idx %d after match failure: rc = %d\n",
2649 loi->loi_id, loi->loi_ost_idx, err);
2653 if (lsm->lsm_stripe_count > 1) {
2654 lov_llh_destroy(lov_lockh);
2655 lov_llh_put(lov_lockh);
2660 static int lov_change_cbdata(struct obd_export *exp,
2661 struct lov_stripe_md *lsm, ldlm_iterator_t it,
2664 struct lov_obd *lov;
2665 struct lov_oinfo *loi;
2669 if (lsm_bad_magic(lsm))
2672 if (!exp || !exp->exp_obd)
2675 LASSERT(lsm->lsm_object_gr > 0);
2677 lov = &exp->exp_obd->u.lov;
2678 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2679 struct lov_stripe_md submd;
2680 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2681 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2685 submd.lsm_object_id = loi->loi_id;
2686 submd.lsm_object_gr = lsm->lsm_object_gr;
2687 submd.lsm_stripe_count = 0;
2688 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
2694 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
2695 __u32 mode, struct lustre_handle *lockh)
2697 struct lov_lock_handles *lov_lockh = NULL;
2698 struct lustre_handle *lov_lockhp;
2699 struct lov_obd *lov;
2700 struct lov_oinfo *loi;
2704 if (lsm_bad_magic(lsm))
2707 if (!exp || !exp->exp_obd)
2710 LASSERT(lsm->lsm_object_gr > 0);
2713 if (lsm->lsm_stripe_count > 1) {
2714 lov_lockh = lov_handle2llh(lockh);
2716 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2720 lov_lockhp = lov_lockh->llh_handles;
2725 lov = &exp->exp_obd->u.lov;
2726 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2727 i++, loi++, lov_lockhp++) {
2728 struct lov_stripe_md submd;
2731 if (lov_lockhp->cookie == 0) {
2732 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2733 loi->loi_ost_idx, loi->loi_id);
2737 /* XXX LOV STACKING: submd should be from the subobj */
2738 submd.lsm_object_id = loi->loi_id;
2739 submd.lsm_object_gr = lsm->lsm_object_gr;
2740 submd.lsm_stripe_count = 0;
2741 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2744 if (lov->tgts[loi->loi_ost_idx].active) {
2745 CERROR("error: cancel objid "LPX64" subobj "
2746 LPX64" on OST idx %d: rc = %d\n",
2748 loi->loi_id, loi->loi_ost_idx, err);
2755 if (lsm->lsm_stripe_count > 1)
2756 lov_llh_destroy(lov_lockh);
2757 if (lov_lockh != NULL)
2758 lov_llh_put(lov_lockh);
2762 static int lov_cancel_unused(struct obd_export *exp,
2763 struct lov_stripe_md *lsm, int flags, void *opaque)
2765 struct lov_obd *lov;
2766 struct lov_oinfo *loi;
2770 lov = &exp->exp_obd->u.lov;
2772 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2773 int err = obd_cancel_unused(lov->tgts[i].ltd_exp, NULL,
2781 if (lsm_bad_magic(lsm))
2784 if (!exp || !exp->exp_obd)
2787 LASSERT(lsm->lsm_object_gr > 0);
2789 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2790 struct lov_stripe_md submd;
2793 if (lov->tgts[loi->loi_ost_idx].active == 0)
2794 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2796 submd.lsm_object_id = loi->loi_id;
2797 submd.lsm_object_gr = lsm->lsm_object_gr;
2798 submd.lsm_stripe_count = 0;
2799 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
2800 &submd, flags, opaque);
2801 if (err && lov->tgts[loi->loi_ost_idx].active) {
2802 CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2803 " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2804 loi->loi_id, loi->loi_ost_idx, err);
2812 #define LOV_U64_MAX ((__u64)~0ULL)
2813 #define LOV_SUM_MAX(tot, add) \
2815 if ((tot) + (add) < (tot)) \
2816 (tot) = LOV_U64_MAX; \
2821 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2822 unsigned long max_age)
2824 struct lov_obd *lov = &obd->u.lov;
2825 struct obd_statfs lov_sfs;
2832 /* We only get block data from the OBD */
2833 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2835 if (!lov->tgts[i].active) {
2836 CDEBUG(D_HA, "lov idx %d inactive\n", i);
2840 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
2843 if (lov->tgts[i].active && !rc)
2849 memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
2852 osfs->os_bfree += lov_sfs.os_bfree;
2853 osfs->os_bavail += lov_sfs.os_bavail;
2854 osfs->os_blocks += lov_sfs.os_blocks;
2855 /* XXX not sure about this one - depends on policy.
2856 * - could be minimum if we always stripe on all OBDs
2857 * (but that would be wrong for any other policy,
2858 * if one of the OBDs has no more objects left)
2859 * - could be sum if we stripe whole objects
2860 * - could be average, just to give a nice number
2862 * To give a "reasonable" (if not wholly accurate)
2863 * number, we divide the total number of free objects
2864 * by expected stripe count (watch out for overflow).
2866 LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
2867 LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
2872 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
2873 lov->desc.ld_default_stripe_count :
2874 lov->desc.ld_active_tgt_count;
2876 if (osfs->os_files != LOV_U64_MAX)
2877 do_div(osfs->os_files, expected_stripes);
2878 if (osfs->os_ffree != LOV_U64_MAX)
2879 do_div(osfs->os_ffree, expected_stripes);
2886 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2887 void *karg, void *uarg)
2889 struct obd_device *obddev = class_exp2obd(exp);
2890 struct lov_obd *lov = &obddev->u.lov;
2891 int i, rc, count = lov->desc.ld_tgt_count;
2892 struct obd_uuid *uuidp;
2896 case OBD_IOC_LOV_GET_CONFIG: {
2897 struct obd_ioctl_data *data = karg;
2898 struct lov_tgt_desc *tgtdesc;
2899 struct lov_desc *desc;
2905 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2908 data = (struct obd_ioctl_data *)buf;
2910 if (sizeof(*desc) > data->ioc_inllen1) {
2915 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2920 if (sizeof(__u32) * count > data->ioc_inllen3) {
2925 desc = (struct lov_desc *)data->ioc_inlbuf1;
2926 memcpy(desc, &(lov->desc), sizeof(*desc));
2928 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2929 genp = (__u32 *)data->ioc_inlbuf3;
2930 tgtdesc = lov->tgts;
2931 /* the uuid will be empty for deleted OSTs */
2932 for (i = 0; i < count; i++, uuidp++, genp++, tgtdesc++) {
2933 obd_str2uuid(uuidp, tgtdesc->uuid.uuid);
2934 *genp = tgtdesc->ltd_gen;
2937 rc = copy_to_user((void *)uarg, buf, len);
2940 obd_ioctl_freedata(buf, len);
2943 case LL_IOC_LOV_SETSTRIPE:
2944 rc = lov_setstripe(exp, karg, uarg);
2946 case LL_IOC_LOV_GETSTRIPE:
2947 rc = lov_getstripe(exp, karg, uarg);
2949 case LL_IOC_LOV_SETEA:
2950 rc = lov_setea(exp, karg, uarg);
2957 for (i = 0; i < count; i++) {
2960 /* OST was deleted */
2961 if (obd_uuid_empty(&lov->tgts[i].uuid))
2964 err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
2967 if (lov->tgts[i].active) {
2968 CERROR("error: iocontrol OSC %s on OST"
2969 "idx %d: err = %d\n",
2970 lov->tgts[i].uuid.uuid, i, err);
2985 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2986 void *key, __u32 *vallen, void *val)
2988 struct obd_device *obddev = class_exp2obd(exp);
2989 struct lov_obd *lov = &obddev->u.lov;
2993 if (!vallen || !val)
2996 if (keylen > strlen("lock_to_stripe") &&
2997 strcmp(key, "lock_to_stripe") == 0) {
3000 struct ldlm_lock *lock;
3001 struct lov_stripe_md *lsm;
3003 struct lov_oinfo *loi;
3004 __u32 *stripe = val;
3006 if (*vallen < sizeof(*stripe))
3008 *vallen = sizeof(*stripe);
3010 /* XXX This is another one of those bits that will need to
3011 * change if we ever actually support nested LOVs. It uses
3012 * the lock's export to find out which stripe it is. */
3013 /* XXX - it's assumed all the locks for deleted OSTs have
3014 * been cancelled. Also, the export for deleted OSTs will
3015 * be NULL and won't match the lock's export. */
3016 for (i = 0, loi = data->lsm->lsm_oinfo;
3017 i < data->lsm->lsm_stripe_count;
3019 if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
3020 data->lock->l_conn_export) {
3026 } else if (keylen >= strlen("size_to_stripe") &&
3027 strcmp(key, "size_to_stripe") == 0) {
3031 struct lov_stripe_md *lsm;
3034 if (*vallen < sizeof(*data))
3037 data->size = lov_size_to_stripe(data->lsm, data->size,
3038 data->stripe_number);
3040 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
3042 int rc, size = sizeof(obd_id);
3043 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3044 if (!lov->tgts[i].active)
3046 rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key,
3052 } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
3053 struct lov_desc *desc_ret = val;
3054 *desc_ret = lov->desc;
3062 static int lov_set_info(struct obd_export *exp, obd_count keylen,
3063 void *key, obd_count vallen, void *val)
3065 struct obd_device *obddev = class_exp2obd(exp);
3066 struct lov_obd *lov = &obddev->u.lov;
3070 #define KEY_IS(str) \
3071 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
3073 if (KEY_IS("next_id")) {
3074 if (vallen != lov->desc.ld_tgt_count)
3076 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3079 /* OST was deleted */
3080 if (obd_uuid_empty(&lov->tgts[i].uuid))
3083 /* initialize all OSCs, even inactive ones */
3085 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key,
3086 sizeof(obd_id), ((obd_id*)val) + i);
3092 if (KEY_IS("growth_count")) {
3093 if (vallen != sizeof(int))
3095 } else if (KEY_IS("mds_conn")) {
3096 if (vallen != sizeof(__u32))
3098 } else if (KEY_IS("unlinked") || KEY_IS("unrecovery")) {
3105 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
3108 /* OST was deleted */
3109 if (obd_uuid_empty(&lov->tgts[i].uuid))
3112 if (!val && !lov->tgts[i].active)
3115 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen,
3125 /* Merge rss if kms == 0
3127 * Even when merging RSS, we will take the KMS value if it's larger.
3128 * This prevents getattr from stomping on dirty cached pages which
3129 * extend the file size. */
3130 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms)
3132 struct lov_oinfo *loi;
3136 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3138 obd_size lov_size, tmpsize;
3140 tmpsize = loi->loi_kms;
3141 if (kms == 0 && loi->loi_rss > tmpsize)
3142 tmpsize = loi->loi_rss;
3144 lov_size = lov_stripe_size(lsm, tmpsize, i);
3145 if (lov_size > size)
3151 EXPORT_SYMBOL(lov_merge_size);
3154 __u64 lov_merge_blocks(struct lov_stripe_md *lsm)
3156 struct lov_oinfo *loi;
3160 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3162 blocks += loi->loi_blocks;
3166 EXPORT_SYMBOL(lov_merge_blocks);
3168 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time)
3170 struct lov_oinfo *loi;
3173 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3175 if (loi->loi_mtime > current_time)
3176 current_time = loi->loi_mtime;
3178 return current_time;
3180 EXPORT_SYMBOL(lov_merge_mtime);
3183 struct lov_multi_wait {
3184 struct ldlm_lock *lock;
3190 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
3191 struct lustre_handle *lockh)
3193 struct lov_lock_handles *lov_lockh = NULL;
3194 struct lustre_handle *lov_lockhp;
3195 struct lov_obd *lov;
3196 struct lov_oinfo *loi;
3197 struct lov_multi_wait *queues;
3201 if (lsm_bad_magic(lsm))
3204 if (!exp || !exp->exp_obd)
3207 LASSERT(lockh != NULL);
3208 if (lsm->lsm_stripe_count > 1) {
3209 lov_lockh = lov_handle2llh(lockh);
3210 if (lov_lockh == NULL) {
3211 CERROR("LOV: invalid lov lock handle %p\n", lockh);
3215 lov_lockhp = lov_lockh->llh_handles;
3220 OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
3222 GOTO(out, rc = -ENOMEM);
3224 lov = &exp->exp_obd->u.lov;
3225 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
3226 i++, loi++, lov_lockhp++) {
3227 struct ldlm_lock *lock;
3228 struct obd_device *obd;
3229 unsigned long irqflags;
3231 lock = ldlm_handle2lock(lov_lockhp);
3233 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
3234 loi->loi_ost_idx, loi->loi_id);
3235 queues[i].completed = 1;
3239 queues[i].lock = lock;
3240 init_waitqueue_entry(&(queues[i].wait), current);
3241 add_wait_queue(lock->l_waitq, &(queues[i].wait));
3243 obd = class_exp2obd(lock->l_conn_export);
3245 imp = obd->u.cli.cl_import;
3247 spin_lock_irqsave(&imp->imp_lock, irqflags);
3248 queues[i].generation = imp->imp_generation;
3249 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
3253 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
3254 interrupted_completion_wait, &lwd);
3255 rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
3257 for (i = 0; i < lsm->lsm_stripe_count; i++)
3258 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
3260 if (rc == -EINTR || rc == -ETIMEDOUT) {
3266 if (lov_lockh != NULL)
3267 lov_llh_put(lov_lockh);
3272 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
3275 struct lov_oinfo *loi;
3281 stripe = lov_stripe_number(lsm, size - 1);
3282 kms = lov_size_to_stripe(lsm, size, stripe);
3283 loi = &(lsm->lsm_oinfo[stripe]);
3285 CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n",
3286 stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms);
3287 if (kms > loi->loi_kms)
3291 EXPORT_SYMBOL(lov_increase_kms);
3293 struct obd_ops lov_obd_ops = {
3294 .o_owner = THIS_MODULE,
3295 .o_attach = lov_attach,
3296 .o_detach = lov_detach,
3297 .o_setup = lov_setup,
3298 .o_cleanup = lov_cleanup,
3299 o_process_config: lov_process_config,
3300 .o_connect = lov_connect,
3301 .o_disconnect = lov_disconnect,
3302 .o_statfs = lov_statfs,
3303 .o_packmd = lov_packmd,
3304 .o_unpackmd = lov_unpackmd,
3305 o_revalidate_md: lov_revalidate_md,
3306 .o_create = lov_create,
3307 .o_destroy = lov_destroy,
3308 .o_getattr = lov_getattr,
3309 .o_getattr_async = lov_getattr_async,
3310 .o_setattr = lov_setattr,
3312 .o_brw_async = lov_brw_async,
3313 .o_prep_async_page = lov_prep_async_page,
3314 .o_queue_async_io = lov_queue_async_io,
3315 .o_set_async_flags = lov_set_async_flags,
3316 .o_queue_group_io = lov_queue_group_io,
3317 .o_trigger_group_io = lov_trigger_group_io,
3318 .o_teardown_async_page = lov_teardown_async_page,
3319 .o_punch = lov_punch,
3321 .o_enqueue = lov_enqueue,
3322 .o_match = lov_match,
3323 .o_change_cbdata = lov_change_cbdata,
3324 .o_cancel = lov_cancel,
3325 .o_cancel_unused = lov_cancel_unused,
3326 .o_iocontrol = lov_iocontrol,
3327 .o_get_info = lov_get_info,
3328 .o_set_info = lov_set_info,
3329 .o_llog_init = lov_llog_init,
3330 .o_llog_finish = lov_llog_finish,
3331 .o_notify = lov_notify,
3334 int __init lov_init(void)
3336 struct lprocfs_static_vars lvars;
3340 lprocfs_init_vars(lov, &lvars);
3341 rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars,
3342 OBD_LOV_DEVICENAME);
3347 static void /*__exit*/ lov_exit(void)
3349 class_unregister_type(OBD_LOV_DEVICENAME);
3352 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3353 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
3354 MODULE_LICENSE("GPL");
3356 module_init(lov_init);
3357 module_exit(lov_exit);