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 static int lov_connect(struct lustre_handle *conn, struct obd_device *obd,
118 struct obd_uuid *cluuid)
120 struct ptlrpc_request *req = NULL;
121 struct lov_obd *lov = &obd->u.lov;
122 struct lov_desc *desc = &lov->desc;
123 struct lov_tgt_desc *tgts;
124 struct obd_export *exp;
128 rc = class_connect(conn, obd, cluuid);
132 exp = class_conn2export(conn);
134 /* We don't want to actually do the underlying connections more than
135 * once, so keep track. */
137 if (lov->refcount > 1) {
138 class_export_put(exp);
142 for (i = 0, tgts = lov->tgts; i < desc->ld_tgt_count; i++, tgts++) {
143 struct obd_uuid *tgt_uuid = &tgts->uuid;
144 struct obd_device *tgt_obd;
145 struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
146 struct lustre_handle conn = {0, };
148 LASSERT( tgt_uuid != NULL);
150 tgt_obd = class_find_client_obd(tgt_uuid, LUSTRE_OSC_NAME,
154 CERROR("Target %s not attached\n", tgt_uuid->uuid);
155 GOTO(out_disc, rc = -EINVAL);
158 if (!tgt_obd->obd_set_up) {
159 CERROR("Target %s not set up\n", tgt_uuid->uuid);
160 GOTO(out_disc, rc = -EINVAL);
163 if (tgt_obd->u.cli.cl_import->imp_invalid) {
164 CERROR("not connecting OSC %s; administratively "
165 "disabled\n", tgt_uuid->uuid);
166 rc = obd_register_observer(tgt_obd, obd);
168 CERROR("Target %s register_observer error %d; "
169 "will not be able to reactivate\n",
175 rc = obd_connect(&conn, tgt_obd, &lov_osc_uuid);
177 CERROR("Target %s connect error %d\n", tgt_uuid->uuid,
181 tgts->ltd_exp = class_conn2export(&conn);
183 rc = obd_register_observer(tgt_obd, obd);
185 CERROR("Target %s register_observer error %d\n",
187 obd_disconnect(tgts->ltd_exp, 0);
191 desc->ld_active_tgt_count++;
195 ptlrpc_req_finished(req);
196 class_export_put(exp);
201 struct obd_uuid uuid;
203 --desc->ld_active_tgt_count;
205 /* save for CERROR below; (we know it's terminated) */
207 rc2 = obd_disconnect(tgts->ltd_exp, 0);
209 CERROR("error: LOV target %s disconnect on OST idx %d: "
210 "rc = %d\n", uuid.uuid, i, rc2);
212 class_disconnect(exp, 0);
216 static int lov_disconnect(struct obd_export *exp, int flags)
218 struct obd_device *obd = class_exp2obd(exp);
219 struct lov_obd *lov = &obd->u.lov;
226 /* Only disconnect the underlying layers on the final disconnect. */
228 if (lov->refcount != 0)
231 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
232 if (lov->tgts[i].ltd_exp == NULL)
235 if (obd->obd_no_recov) {
236 /* Pass it on to our clients.
237 * XXX This should be an argument to disconnect,
238 * XXX not a back-door flag on the OBD. Ah well.
240 struct obd_device *osc_obd;
241 osc_obd = class_exp2obd(lov->tgts[i].ltd_exp);
243 osc_obd->obd_no_recov = 1;
246 obd_register_observer(lov->tgts[i].ltd_exp->exp_obd, NULL);
248 rc = obd_disconnect(lov->tgts[i].ltd_exp, flags);
250 if (lov->tgts[i].active) {
251 CERROR("Target %s disconnect error %d\n",
252 lov->tgts[i].uuid.uuid, rc);
256 if (lov->tgts[i].active) {
257 lov->desc.ld_active_tgt_count--;
258 lov->tgts[i].active = 0;
260 lov->tgts[i].ltd_exp = NULL;
264 rc = class_disconnect(exp, 0);
270 * -EINVAL : UUID can't be found in the LOV's target list
271 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
272 * -EBADF : The UUID is found, but the OBD is the wrong type (!)
274 static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid,
277 struct obd_device *obd;
278 struct lov_tgt_desc *tgt;
282 CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
283 lov, uuid->uuid, activate);
285 spin_lock(&lov->lov_lock);
286 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
287 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
288 i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
289 if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0)
293 if (i == lov->desc.ld_tgt_count)
294 GOTO(out, rc = -EINVAL);
296 obd = class_exp2obd(tgt->ltd_exp);
298 /* This can happen if OST failure races with node shutdown */
299 GOTO(out, rc = -ENOTCONN);
302 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LOV idx %d\n",
303 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
304 obd->obd_type->typ_name, i);
305 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_OSC_NAME) == 0);
307 if (tgt->active == activate) {
308 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
309 activate ? "" : "in");
313 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd, activate ? "" : "in");
315 tgt->active = activate;
317 lov->desc.ld_active_tgt_count++;
319 lov->desc.ld_active_tgt_count--;
323 spin_unlock(&lov->lov_lock);
327 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
331 struct obd_uuid *uuid;
333 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
334 CERROR("unexpected notification of %s %s!\n",
335 watched->obd_type->typ_name,
339 uuid = &watched->u.cli.cl_import->imp_target_uuid;
341 /* Set OSC as active before notifying the observer, so the
342 * observer can use the OSC normally.
344 rc = lov_set_osc_active(&obd->u.lov, uuid, active);
346 CERROR("%sactivation of %s failed: %d\n",
347 active ? "" : "de", uuid->uuid, rc);
351 if (obd->obd_observer)
352 /* Pass the notification up the chain. */
353 rc = obd_notify(obd->obd_observer, watched, active);
358 int lov_attach(struct obd_device *dev, obd_count len, void *data)
360 struct lprocfs_static_vars lvars;
363 lprocfs_init_vars(lov, &lvars);
364 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
367 struct proc_dir_entry *entry;
369 entry = create_proc_entry("target_obd", 0444,
370 dev->obd_proc_entry);
374 entry->proc_fops = &lov_proc_target_fops;
382 int lov_detach(struct obd_device *dev)
384 return lprocfs_obd_detach(dev);
387 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
389 struct lustre_cfg *lcfg = buf;
390 struct lov_desc *desc;
391 struct lov_obd *lov = &obd->u.lov;
392 struct obd_uuid *uuids;
393 struct lov_tgt_desc *tgts;
398 if (lcfg->lcfg_inllen1 < 1) {
399 CERROR("LOV setup requires a descriptor\n");
403 if (lcfg->lcfg_inllen2 < 1) {
404 CERROR("LOV setup requires an OST UUID list\n");
408 desc = (struct lov_desc *)lcfg->lcfg_inlbuf1;
409 if (sizeof(*desc) > lcfg->lcfg_inllen1) {
410 CERROR("descriptor size wrong: %d > %d\n",
411 (int)sizeof(*desc), lcfg->lcfg_inllen1);
415 count = desc->ld_tgt_count;
416 uuids = (struct obd_uuid *)lcfg->lcfg_inlbuf2;
417 if (sizeof(*uuids) * count != lcfg->lcfg_inllen2) {
418 CERROR("UUID array size wrong: %u * %u != %u\n",
419 (int)sizeof(*uuids), count, lcfg->lcfg_inllen2);
423 /* Because of 64-bit divide/mod operations only work with a 32-bit
424 * divisor in a 32-bit kernel, we cannot support a stripe width
425 * of 4GB or larger on 32-bit CPUs.
427 if ((desc->ld_default_stripe_count ?
428 desc->ld_default_stripe_count : desc->ld_tgt_count) *
429 desc->ld_default_stripe_size > ~0UL) {
430 CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n",
431 desc->ld_default_stripe_size,
432 desc->ld_default_stripe_count ?
433 desc->ld_default_stripe_count : desc->ld_tgt_count,~0UL);
437 lov->bufsize = sizeof(struct lov_tgt_desc) * count;
438 OBD_ALLOC(lov->tgts, lov->bufsize);
439 if (lov->tgts == NULL) {
440 CERROR("Out of memory\n");
445 spin_lock_init(&lov->lov_lock);
447 for (i = 0, tgts = lov->tgts; i < desc->ld_tgt_count; i++, tgts++) {
448 struct obd_uuid *uuid = &tgts->uuid;
450 /* NULL termination already checked */
457 static int lov_cleanup(struct obd_device *obd, int flags)
459 struct lov_obd *lov = &obd->u.lov;
461 OBD_FREE(lov->tgts, lov->bufsize);
465 /* compute object size given "stripeno" and the ost size */
466 static obd_size lov_stripe_size(struct lov_stripe_md *lsm, obd_size ost_size,
469 unsigned long ssize = lsm->lsm_stripe_size;
470 unsigned long swidth = ssize * lsm->lsm_stripe_count;
471 unsigned long stripe_size;
477 /* do_div(a, b) returns a % b, and a = a / b */
478 stripe_size = do_div(ost_size, ssize);
480 lov_size = ost_size * swidth + stripeno * ssize + stripe_size;
482 lov_size = (ost_size - 1) * swidth + (stripeno + 1) * ssize;
487 static void lov_merge_attrs(struct obdo *tgt, struct obdo *src, obd_flag valid,
488 struct lov_stripe_md *lsm, int stripeno, int *set)
490 valid &= src->o_valid;
493 if (valid & OBD_MD_FLSIZE) {
494 /* this handles sparse files properly */
497 lov_size = lov_stripe_size(lsm, src->o_size, stripeno);
498 if (lov_size > tgt->o_size)
499 tgt->o_size = lov_size;
501 if (valid & OBD_MD_FLBLOCKS)
502 tgt->o_blocks += src->o_blocks;
503 if (valid & OBD_MD_FLBLKSZ)
504 tgt->o_blksize += src->o_blksize;
505 if (valid & OBD_MD_FLCTIME && tgt->o_ctime < src->o_ctime)
506 tgt->o_ctime = src->o_ctime;
507 if (valid & OBD_MD_FLMTIME && tgt->o_mtime < src->o_mtime)
508 tgt->o_mtime = src->o_mtime;
510 memcpy(tgt, src, sizeof(*tgt));
511 tgt->o_id = lsm->lsm_object_id;
512 if (valid & OBD_MD_FLSIZE)
513 tgt->o_size = lov_stripe_size(lsm,src->o_size,stripeno);
519 #define log2(n) ffz(~(n))
522 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
523 struct lov_stripe_md **ea,
524 struct obd_trans_info *oti)
528 struct obd_uuid *ost_uuid = NULL;
532 LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
533 src_oa->o_flags == OBD_FL_DELORPHAN);
535 lov = &export->exp_obd->u.lov;
537 tmp_oa = obdo_alloc();
541 if (src_oa->o_valid & OBD_MD_FLINLINE) {
542 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
543 CDEBUG(D_HA, "clearing orphans only for %s\n",
547 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
548 struct lov_stripe_md obj_md;
549 struct lov_stripe_md *obj_mdp = &obj_md;
552 /* if called for a specific target, we don't
553 care if it is not active. */
554 if (lov->tgts[i].active == 0 && ost_uuid == NULL) {
555 CDEBUG(D_HA, "lov idx %d inactive\n", i);
559 if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid))
562 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
564 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
565 err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, &obj_mdp, oti);
567 /* This export will be disabled until it is recovered,
568 and then orphan recovery will be completed. */
569 CERROR("error in orphan recovery on OST idx %d/%d: "
570 "rc = %d\n", i, lov->desc.ld_tgt_count, err);
579 #define LOV_CREATE_RESEED_INTERVAL 1000
581 /* the LOV expects oa->o_id to be set to the LOV object id */
582 static int lov_create(struct obd_export *exp, struct obdo *src_oa,
583 struct lov_stripe_md **ea, struct obd_trans_info *oti)
585 static int ost_start_idx, ost_start_count;
587 struct lov_stripe_md *lsm;
588 struct lov_oinfo *loi = NULL;
589 struct obdo *tmp_oa, *ret_oa;
590 struct llog_cookie *cookies = NULL;
591 unsigned ost_count, ost_idx;
592 int set = 0, obj_alloc = 0, cookie_sent = 0, rc = 0, i;
597 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
598 src_oa->o_flags == OBD_FL_DELORPHAN) {
599 rc = lov_clear_orphans(exp, src_oa, ea, oti);
606 lov = &exp->exp_obd->u.lov;
608 if (!lov->desc.ld_active_tgt_count)
611 /* Recreate a specific object id at the given OST index */
612 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
613 (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) {
614 struct lov_stripe_md obj_md;
615 struct lov_stripe_md *obj_mdp = &obj_md;
617 ost_idx = src_oa->o_nlink;
621 if (ost_idx >= lov->desc.ld_tgt_count)
623 for (i = 0; i < lsm->lsm_stripe_count; i++) {
624 if (lsm->lsm_oinfo[i].loi_ost_idx == ost_idx) {
625 if (lsm->lsm_oinfo[i].loi_id != src_oa->o_id ||
626 lsm->lsm_oinfo[i].loi_gr != src_oa->o_gr) {
632 if (i == lsm->lsm_stripe_count)
635 rc = obd_create(lov->tgts[ost_idx].ltd_exp, src_oa,
640 ret_oa = obdo_alloc();
644 tmp_oa = obdo_alloc();
646 GOTO(out_oa, rc = -ENOMEM);
651 ost_count = lov_get_stripecnt(lov, 0);
653 /* If the MDS file was truncated up to some size, stripe over
654 * enough OSTs to allow the file to be created at that size. */
655 if (src_oa->o_valid & OBD_MD_FLSIZE) {
656 stripes=((src_oa->o_size+LUSTRE_STRIPE_MAXBYTES)>>12)-1;
657 do_div(stripes, (__u32)(LUSTRE_STRIPE_MAXBYTES >> 12));
659 if (stripes > lov->desc.ld_active_tgt_count)
661 if (stripes < ost_count)
667 rc = lov_alloc_memmd(&lsm, stripes, lov->desc.ld_pattern ?
668 lov->desc.ld_pattern : LOV_PATTERN_RAID0);
675 ost_count = lov->desc.ld_tgt_count;
677 LASSERT(src_oa->o_gr > 0);
678 LASSERT(src_oa->o_valid & OBD_MD_FLID);
679 lsm->lsm_object_id = src_oa->o_id;
680 lsm->lsm_object_gr = src_oa->o_gr;
681 if (!lsm->lsm_stripe_size)
682 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
683 if (!lsm->lsm_pattern) {
684 lsm->lsm_pattern = lov->desc.ld_pattern ?
685 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
688 if (*ea == NULL || lsm->lsm_oinfo[0].loi_ost_idx >= ost_count) {
689 if (--ost_start_count <= 0) {
690 ost_start_idx = ll_insecure_random_int();
691 ost_start_count = LOV_CREATE_RESEED_INTERVAL;
692 } else if (lsm->lsm_stripe_count >=
693 lov->desc.ld_active_tgt_count) {
694 /* If we allocate from all of the stripes, make the
695 * next file start on the next OST. */
698 ost_idx = ost_start_idx % ost_count;
700 ost_idx = lsm->lsm_oinfo[0].loi_ost_idx;
703 CDEBUG(D_INODE, "allocating %d subobjs for objid "LPX64" at idx %d\n",
704 lsm->lsm_stripe_count, lsm->lsm_object_id, ost_idx);
706 /* XXX LOV STACKING: need to figure out how many real OSCs */
707 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
708 oti_alloc_cookies(oti, lsm->lsm_stripe_count);
709 if (!oti->oti_logcookies)
710 GOTO(out_cleanup, rc = -ENOMEM);
711 cookies = oti->oti_logcookies;
714 loi = lsm->lsm_oinfo;
715 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
716 struct lov_stripe_md obj_md;
717 struct lov_stripe_md *obj_mdp = &obj_md;
721 if (lov->tgts[ost_idx].active == 0) {
722 CDEBUG(D_HA, "lov idx %d inactive\n", ost_idx);
726 /* create data objects with "parent" OA */
727 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
729 /* XXX When we start creating objects on demand, we need to
730 * make sure that we always create the object on the
731 * stripe which holds the existing file size.
733 if (src_oa->o_valid & OBD_MD_FLSIZE) {
734 if (lov_stripe_offset(lsm, src_oa->o_size, i,
735 &tmp_oa->o_size) < 0 &&
739 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
740 i, tmp_oa->o_size, src_oa->o_size);
744 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
745 err = obd_create(lov->tgts[ost_idx].ltd_exp, tmp_oa, &obj_mdp,
748 if (lov->tgts[ost_idx].active) {
749 CERROR("error creating objid "LPX64" sub-object"
750 " on OST idx %d/%d: rc = %d\n",
751 src_oa->o_id, ost_idx,
752 lsm->lsm_stripe_count, err);
754 CERROR("obd_create returned invalid "
764 oti->oti_objid[ost_idx] = tmp_oa->o_id;
765 loi->loi_id = tmp_oa->o_id;
766 loi->loi_gr = tmp_oa->o_gr;
767 loi->loi_ost_idx = ost_idx;
769 "objid "LPX64" has subobj "LPX64"/"LPX64" at idx %d\n",
770 lsm->lsm_object_id, loi->loi_id, loi->loi_id, ost_idx);
772 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
777 ++oti->oti_logcookies;
778 if (tmp_oa->o_valid & OBD_MD_FLCOOKIE)
783 /* If we have allocated enough objects, we are OK */
784 if (obj_alloc == lsm->lsm_stripe_count)
785 GOTO(out_done, rc = 0);
788 if (obj_alloc == 0) {
791 GOTO(out_cleanup, rc);
794 /* If we were passed specific striping params, then a failure to
795 * meet those requirements is an error, since we can't reallocate
796 * that memory (it might be part of a larger array or something).
798 * We can only get here if lsm_stripe_count was originally > 1.
801 CERROR("can't lstripe objid "LPX64": have %u want %u, rc %d\n",
802 lsm->lsm_object_id, obj_alloc, lsm->lsm_stripe_count,rc);
805 GOTO(out_cleanup, rc);
807 struct lov_stripe_md *lsm_new;
808 /* XXX LOV STACKING call into osc for sizes */
809 unsigned oldsize, newsize;
811 if (oti && cookies && cookie_sent) {
812 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
813 newsize = obj_alloc * sizeof(*cookies);
815 oti_alloc_cookies(oti, obj_alloc);
816 if (oti->oti_logcookies) {
817 memcpy(oti->oti_logcookies, cookies, newsize);
818 OBD_FREE(cookies, oldsize);
819 cookies = oti->oti_logcookies;
821 CWARN("'leaking' %d bytes\n", oldsize-newsize);
825 CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
826 lsm->lsm_object_id, lsm->lsm_stripe_count, obj_alloc);
827 oldsize = lov_stripe_md_size(lsm->lsm_stripe_count);
828 newsize = lov_stripe_md_size(obj_alloc);
829 OBD_ALLOC(lsm_new, newsize);
830 if (lsm_new != NULL) {
831 memcpy(lsm_new, lsm, newsize);
832 lsm_new->lsm_stripe_count = obj_alloc;
833 OBD_FREE(lsm, oldsize);
836 CWARN("'leaking' %d bytes\n", oldsize - newsize);
843 if (src_oa->o_valid & OBD_MD_FLSIZE &&
844 ret_oa->o_size != src_oa->o_size) {
845 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
846 src_oa->o_size, ret_oa->o_size);
849 ret_oa->o_id = src_oa->o_id;
850 ret_oa->o_gr = src_oa->o_gr;
851 ret_oa->o_valid |= OBD_MD_FLGROUP;
852 memcpy(src_oa, ret_oa, sizeof(*src_oa));
858 if (oti && cookies) {
859 oti->oti_logcookies = cookies;
861 oti_free_cookies(oti);
862 src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
864 src_oa->o_valid |= OBD_MD_FLCOOKIE;
870 while (obj_alloc-- > 0) {
871 struct obd_export *sub_exp;
875 sub_exp = lov->tgts[loi->loi_ost_idx].ltd_exp;
876 /* destroy already created objects here */
877 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
878 tmp_oa->o_id = loi->loi_id;
880 err = obd_destroy(sub_exp, tmp_oa, NULL, oti);
882 CERROR("Failed to uncreate objid "LPX64" subobj "LPX64
883 " on OST idx %d: rc = %d\n", src_oa->o_id,
884 loi->loi_id, loi->loi_ost_idx, err);
887 obd_free_memmd(exp, &lsm);
891 #define lsm_bad_magic(LSMP) \
893 struct lov_stripe_md *_lsm__ = (LSMP); \
896 CERROR("LOV requires striping ea\n"); \
898 } else if (_lsm__->lsm_magic != LOV_MAGIC) { \
899 CERROR("LOV striping magic bad %#x != %#x\n", \
900 _lsm__->lsm_magic, LOV_MAGIC); \
906 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
907 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
911 struct lov_oinfo *loi;
915 if (lsm_bad_magic(lsm))
918 if (!exp || !exp->exp_obd)
921 lov = &exp->exp_obd->u.lov;
922 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
924 if (lov->tgts[loi->loi_ost_idx].active == 0) {
925 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
926 /* Orphan clean up will (someday) fix this up. */
927 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
928 oti->oti_logcookies++;
932 memcpy(&tmp, oa, sizeof(tmp));
933 tmp.o_id = loi->loi_id;
934 err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
936 if (err && lov->tgts[loi->loi_ost_idx].active) {
937 CDEBUG(D_INODE, "error: destroying objid "LPX64" subobj "
938 LPX64" on OST idx %d: rc = %d\n",
939 oa->o_id, loi->loi_id, loi->loi_ost_idx, err);
947 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
948 struct lov_stripe_md *lsm)
952 struct lov_oinfo *loi;
953 int i, rc = 0, set = 0;
956 if (lsm_bad_magic(lsm))
959 if (!exp || !exp->exp_obd)
962 lov = &exp->exp_obd->u.lov;
964 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
965 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
966 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
969 if (lov->tgts[loi->loi_ost_idx].active == 0) {
970 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
974 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
975 "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
976 /* create data objects with "parent" OA */
977 memcpy(&tmp, oa, sizeof(tmp));
978 tmp.o_id = loi->loi_id;
980 err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
983 if (lov->tgts[loi->loi_ost_idx].active) {
984 CERROR("error: getattr objid "LPX64" subobj "
985 LPX64" on OST idx %d: rc = %d\n",
986 oa->o_id, loi->loi_id, loi->loi_ost_idx,
991 lov_merge_attrs(oa, &tmp, tmp.o_valid, lsm, i, &set);
999 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
1002 struct lov_getattr_async_args *aa = data;
1003 struct lov_stripe_md *lsm = aa->aa_lsm;
1004 struct obdo *oa = aa->aa_oa;
1005 struct obdo *obdos = aa->aa_obdos;
1006 struct lov_oinfo *loi;
1012 /* NB all stripe requests succeeded to get here */
1014 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1016 if (obdos[i].o_valid == 0) /* inactive stripe */
1019 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1024 CERROR ("No stripes had valid attrs\n");
1029 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1033 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
1034 struct lov_stripe_md *lsm,
1035 struct ptlrpc_request_set *rqset)
1038 struct lov_obd *lov;
1039 struct lov_oinfo *loi;
1040 struct lov_getattr_async_args *aa;
1041 int i, rc = 0, set = 0;
1044 if (lsm_bad_magic(lsm))
1047 if (!exp || !exp->exp_obd)
1050 lov = &exp->exp_obd->u.lov;
1052 OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1056 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1057 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1058 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1061 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1062 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1063 /* leaves obdos[i].obd_valid unset */
1067 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1068 "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
1069 /* create data objects with "parent" OA */
1070 memcpy(&obdos[i], oa, sizeof(obdos[i]));
1071 obdos[i].o_id = loi->loi_id;
1073 err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp,
1074 &obdos[i], NULL, rqset);
1076 CERROR("error: getattr objid "LPX64" subobj "
1077 LPX64" on OST idx %d: rc = %d\n",
1078 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1080 GOTO(out_obdos, rc = err);
1085 GOTO (out_obdos, rc = -EIO);
1087 LASSERT (rqset->set_interpret == NULL);
1088 rqset->set_interpret = lov_getattr_interpret;
1089 LASSERT (sizeof (rqset->set_args) >= sizeof (*aa));
1090 aa = (struct lov_getattr_async_args *)&rqset->set_args;
1093 aa->aa_obdos = obdos;
1098 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1104 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
1105 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1107 struct obdo *tmp_oa, *ret_oa;
1108 struct lov_obd *lov;
1109 struct lov_oinfo *loi;
1110 int rc = 0, i, set = 0;
1113 if (lsm_bad_magic(lsm))
1116 if (!exp || !exp->exp_obd)
1119 /* for now, we only expect time updates here */
1120 LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
1121 OBD_MD_FLATIME | OBD_MD_FLMTIME |
1122 OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
1123 OBD_MD_FLSIZE | OBD_MD_FLGROUP)));
1125 LASSERT(!(src_oa->o_valid & OBD_MD_FLGROUP) || src_oa->o_gr > 0);
1127 ret_oa = obdo_alloc();
1131 tmp_oa = obdo_alloc();
1133 GOTO(out_oa, rc = -ENOMEM);
1135 lov = &exp->exp_obd->u.lov;
1136 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1139 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1140 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1144 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1145 tmp_oa->o_id = loi->loi_id;
1146 LASSERT(!(tmp_oa->o_valid & OBD_MD_FLGROUP) || tmp_oa->o_gr>0);
1148 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1149 if (lov_stripe_offset(lsm, src_oa->o_size, i,
1150 &tmp_oa->o_size) < 0 &&
1154 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1155 i, tmp_oa->o_size, src_oa->o_size);
1158 err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa,
1161 if (lov->tgts[loi->loi_ost_idx].active) {
1162 CERROR("error: setattr objid "LPX64" subobj "
1163 LPX64" on OST idx %d: rc = %d\n",
1164 src_oa->o_id, loi->loi_id,
1165 loi->loi_ost_idx, err);
1171 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set);
1176 ret_oa->o_id = src_oa->o_id;
1177 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1186 /* we have an offset in file backed by an lov and want to find out where
1187 * that offset lands in our given stripe of the file. for the easy
1188 * case where the offset is within the stripe, we just have to scale the
1189 * offset down to make it relative to the stripe instead of the lov.
1191 * the harder case is what to do when the offset doesn't intersect the
1192 * stripe. callers will want start offsets clamped ahead to the start
1193 * of the nearest stripe in the file. end offsets similarly clamped to the
1194 * nearest ending byte of a stripe in the file:
1196 * all this function does is move offsets to the nearest region of the
1197 * stripe, and it does its work "mod" the full length of all the stripes.
1198 * consider a file with 3 stripes:
1201 * ---------------------------------------------------------------------
1202 * | 0 | 1 | 2 | 0 | 1 | 2 |
1203 * ---------------------------------------------------------------------
1205 * to find stripe 1's offsets for S and E, it divides by the full stripe
1206 * width and does its math in the context of a single set of stripes:
1209 * -----------------------------------
1211 * -----------------------------------
1213 * it'll notice that E is outside stripe 1 and clamp it to the end of the
1214 * stripe, then multiply it back out by lov_off to give the real offsets in
1218 * ---------------------------------------------------------------------
1219 * | 1 | 1 | 1 | 1 | 1 | 1 |
1220 * ---------------------------------------------------------------------
1222 * it would have done similarly and pulled S forward to the start of a 1
1223 * stripe if, say, S had landed in a 0 stripe.
1225 * this rounding isn't always correct. consider an E lov offset that lands
1226 * on a 0 stripe, the "mod stripe width" math will pull it forward to the
1227 * start of a 1 stripe, when in fact it wanted to be rounded back to the end
1228 * of a previous 1 stripe. this logic is handled by callers and this is why:
1230 * this function returns < 0 when the offset was "before" the stripe and
1231 * was moved forward to the start of the stripe in question; 0 when it
1232 * falls in the stripe and no shifting was done; > 0 when the offset
1233 * was outside the stripe and was pulled back to its final byte. */
1234 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
1235 int stripeno, obd_off *obd_off)
1237 unsigned long ssize = lsm->lsm_stripe_size;
1238 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1239 unsigned long stripe_off, this_stripe;
1242 if (lov_off == OBD_OBJECT_EOF) {
1243 *obd_off = OBD_OBJECT_EOF;
1247 /* do_div(a, b) returns a % b, and a = a / b */
1248 stripe_off = do_div(lov_off, swidth);
1250 this_stripe = stripeno * ssize;
1251 if (stripe_off < this_stripe) {
1255 stripe_off -= this_stripe;
1257 if (stripe_off >= ssize) {
1263 *obd_off = lov_off * ssize + stripe_off;
1267 /* Given a whole-file size and a stripe number, give the file size which
1268 * corresponds to the individual object of that stripe.
1270 * This behaves basically in the same was as lov_stripe_offset, except that
1271 * file sizes falling before the beginning of a stripe are clamped to the end
1272 * of the previous stripe, not the beginning of the next:
1275 * ---------------------------------------------------------------------
1276 * | 0 | 1 | 2 | 0 | 1 | 2 |
1277 * ---------------------------------------------------------------------
1279 * if clamped to stripe 2 becomes:
1282 * ---------------------------------------------------------------------
1283 * | 0 | 1 | 2 | 0 | 1 | 2 |
1284 * ---------------------------------------------------------------------
1286 static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size,
1289 unsigned long ssize = lsm->lsm_stripe_size;
1290 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1291 unsigned long stripe_off, this_stripe;
1293 if (file_size == OBD_OBJECT_EOF)
1294 return OBD_OBJECT_EOF;
1296 /* do_div(a, b) returns a % b, and a = a / b */
1297 stripe_off = do_div(file_size, swidth);
1299 this_stripe = stripeno * ssize;
1300 if (stripe_off < this_stripe) {
1301 /* Move to end of previous stripe, or zero */
1302 if (file_size > 0) {
1309 stripe_off -= this_stripe;
1311 if (stripe_off >= ssize) {
1312 /* Clamp to end of this stripe */
1317 return (file_size * ssize + stripe_off);
1320 /* given an extent in an lov and a stripe, calculate the extent of the stripe
1321 * that is contained within the lov extent. this returns true if the given
1322 * stripe does intersect with the lov extent. */
1323 static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno,
1324 obd_off start, obd_off end,
1325 obd_off *obd_start, obd_off *obd_end)
1327 int start_side = 0, end_side = 0;
1329 switch (lsm->lsm_pattern) {
1330 case LOV_PATTERN_RAID0:
1331 start_side = lov_stripe_offset(lsm, start, stripeno, obd_start);
1332 end_side = lov_stripe_offset(lsm, end, stripeno, obd_end);
1334 case LOV_PATTERN_CMOBD:
1337 start_side = end_side = 0;
1343 CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n",
1344 start, end, start_side, *obd_start, *obd_end, end_side);
1346 /* this stripe doesn't intersect the file extent when neither
1347 * start or the end intersected the stripe and obd_start and
1348 * obd_end got rounded up to the save value. */
1349 if (start_side != 0 && end_side != 0 && *obd_start == *obd_end)
1352 /* as mentioned in the lov_stripe_offset commentary, end
1353 * might have been shifted in the wrong direction. This
1354 * happens when an end offset is before the stripe when viewed
1355 * through the "mod stripe size" math. we detect it being shifted
1356 * in the wrong direction and touch it up.
1357 * interestingly, this can't underflow since end must be > start
1358 * if we passed through the previous check.
1359 * (should we assert for that somewhere?) */
1366 /* compute which stripe number "lov_off" will be written into */
1367 static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off)
1369 unsigned long ssize = lsm->lsm_stripe_size;
1370 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1371 unsigned long stripe_off;
1373 stripe_off = do_div(lov_off, swidth);
1375 return stripe_off / ssize;
1378 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1379 * we can send this 'punch' to just the authoritative node and the nodes
1380 * that the punch will affect. */
1381 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1382 struct lov_stripe_md *lsm,
1383 obd_off start, obd_off end, struct obd_trans_info *oti)
1386 struct lov_obd *lov;
1387 struct lov_oinfo *loi;
1391 if (lsm_bad_magic(lsm))
1394 if (!exp || !exp->exp_obd)
1397 lov = &exp->exp_obd->u.lov;
1398 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1399 obd_off starti, endi;
1402 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1403 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1407 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1410 /* create data objects with "parent" OA */
1411 memcpy(&tmp, oa, sizeof(tmp));
1412 tmp.o_id = loi->loi_id;
1414 err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, NULL,
1415 starti, endi, NULL);
1417 if (lov->tgts[loi->loi_ost_idx].active) {
1418 CERROR("error: punch objid "LPX64" subobj "LPX64
1419 " on OST idx %d: rc = %d\n", oa->o_id,
1420 loi->loi_id, loi->loi_ost_idx, err);
1425 loi->loi_kms = loi->loi_rss = starti;
1431 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1432 struct lov_stripe_md *lsm, obd_off start, obd_off end)
1435 struct lov_obd *lov;
1436 struct lov_oinfo *loi;
1440 if (lsm_bad_magic(lsm))
1450 lov = &exp->exp_obd->u.lov;
1451 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1452 obd_off starti, endi;
1455 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1456 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1460 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1463 memcpy(tmp, oa, sizeof(*tmp));
1464 tmp->o_id = loi->loi_id;
1466 err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL,
1469 if (lov->tgts[loi->loi_ost_idx].active) {
1470 CERROR("error: fsync objid "LPX64" subobj "LPX64
1471 " on OST idx %d: rc = %d\n", oa->o_id,
1472 loi->loi_id, loi->loi_ost_idx, err);
1483 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1484 struct lov_stripe_md *lsm,
1485 obd_count oa_bufs, struct brw_page *pga)
1489 /* The caller just wants to know if there's a chance that this
1490 * I/O can succeed */
1491 for (i = 0; i < oa_bufs; i++) {
1492 int stripe = lov_stripe_number(lsm, pga[i].off);
1493 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1496 if (!lov_stripe_intersects(lsm, i, pga[i].off,
1497 pga[i].off + pga[i].count,
1501 if (lov->tgts[ost].active == 0) {
1502 CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1505 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[stripe].ltd_exp, oa,
1506 NULL, 1, &pga[i], NULL);
1513 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1514 struct lov_stripe_md *lsm, obd_count oa_bufs,
1515 struct brw_page *pga, struct obd_trans_info *oti)
1521 struct lov_stripe_md lsm;
1523 } *stripeinfo, *si, *si_last;
1524 struct obdo *ret_oa = NULL, *tmp_oa = NULL;
1525 struct lov_obd *lov;
1526 struct brw_page *ioarr;
1527 struct lov_oinfo *loi;
1528 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0;
1531 if (lsm_bad_magic(lsm))
1534 lov = &exp->exp_obd->u.lov;
1536 if (cmd == OBD_BRW_CHECK) {
1537 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1541 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1545 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1547 GOTO(out_sinfo, rc = -ENOMEM);
1549 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1551 GOTO(out_where, rc = -ENOMEM);
1554 ret_oa = obdo_alloc();
1556 GOTO(out_ioarr, rc = -ENOMEM);
1558 tmp_oa = obdo_alloc();
1560 GOTO(out_oa, rc = -ENOMEM);
1563 for (i = 0; i < oa_bufs; i++) {
1564 where[i] = lov_stripe_number(lsm, pga[i].off);
1565 stripeinfo[where[i]].bufct++;
1568 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1569 i < stripe_count; i++, loi++, si_last = si, si++) {
1571 si->index = si_last->index + si_last->bufct;
1572 si->lsm.lsm_object_id = loi->loi_id;
1573 si->lsm.lsm_object_gr = lsm->lsm_object_gr;
1574 si->ost_idx = loi->loi_ost_idx;
1577 for (i = 0; i < oa_bufs; i++) {
1578 int which = where[i];
1581 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1582 LASSERT(shift < oa_bufs);
1583 ioarr[shift] = pga[i];
1584 lov_stripe_offset(lsm, pga[i].off, which,
1586 stripeinfo[which].subcount++;
1589 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1590 int shift = si->index;
1592 if (lov->tgts[si->ost_idx].active == 0) {
1593 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1594 GOTO(out_oa, rc = -EIO);
1598 LASSERT(shift < oa_bufs);
1600 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1602 tmp_oa->o_id = si->lsm.lsm_object_id;
1603 rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp,
1604 tmp_oa, &si->lsm, si->bufct,
1605 &ioarr[shift], oti);
1607 GOTO(out_ioarr, rc);
1609 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1614 ret_oa->o_id = src_oa->o_id;
1615 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1624 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1626 OBD_FREE(where, sizeof(*where) * oa_bufs);
1628 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1632 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1635 struct lov_brw_async_args *aa = data;
1636 struct lov_stripe_md *lsm = aa->aa_lsm;
1637 obd_count oa_bufs = aa->aa_oa_bufs;
1638 struct obdo *oa = aa->aa_oa;
1639 struct obdo *obdos = aa->aa_obdos;
1640 struct brw_page *ioarr = aa->aa_ioarr;
1641 struct lov_oinfo *loi;
1646 /* NB all stripe requests succeeded to get here */
1648 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1650 if (obdos[i].o_valid == 0) /* inactive stripe */
1653 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1658 CERROR("No stripes had valid attrs\n");
1662 oa->o_id = lsm->lsm_object_id;
1664 OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos));
1665 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1669 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1670 struct lov_stripe_md *lsm, obd_count oa_bufs,
1671 struct brw_page *pga, struct ptlrpc_request_set *set,
1672 struct obd_trans_info *oti)
1678 struct lov_stripe_md lsm;
1680 } *stripeinfo, *si, *si_last;
1681 struct lov_obd *lov;
1682 struct brw_page *ioarr;
1683 struct obdo *obdos = NULL;
1684 struct lov_oinfo *loi;
1685 struct lov_brw_async_args *aa;
1686 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count;
1689 if (lsm_bad_magic(lsm))
1692 lov = &exp->exp_obd->u.lov;
1694 if (cmd == OBD_BRW_CHECK) {
1695 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
1699 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1703 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1705 GOTO(out_sinfo, rc = -ENOMEM);
1708 OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count);
1710 GOTO(out_where, rc = -ENOMEM);
1713 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1715 GOTO(out_obdos, rc = -ENOMEM);
1717 for (i = 0; i < oa_bufs; i++) {
1718 where[i] = lov_stripe_number(lsm, pga[i].off);
1719 stripeinfo[where[i]].bufct++;
1722 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1723 i < stripe_count; i++, loi++, si_last = si, si++) {
1725 si->index = si_last->index + si_last->bufct;
1726 si->lsm.lsm_object_id = loi->loi_id;
1727 si->ost_idx = loi->loi_ost_idx;
1730 memcpy(&obdos[i], oa, sizeof(*obdos));
1731 obdos[i].o_id = si->lsm.lsm_object_id;
1735 for (i = 0; i < oa_bufs; i++) {
1736 int which = where[i];
1739 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1740 LASSERT(shift < oa_bufs);
1741 ioarr[shift] = pga[i];
1742 lov_stripe_offset(lsm, pga[i].off, which,
1744 stripeinfo[which].subcount++;
1747 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1748 int shift = si->index;
1753 if (lov->tgts[si->ost_idx].active == 0) {
1754 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1755 GOTO(out_ioarr, rc = -EIO);
1758 LASSERT(shift < oa_bufs);
1760 rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp,
1761 &obdos[i], &si->lsm, si->bufct,
1762 &ioarr[shift], set, oti);
1764 GOTO(out_ioarr, rc);
1767 LASSERT(set->set_interpret == NULL);
1768 set->set_interpret = (set_interpreter_func)lov_brw_interpret;
1769 LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args));
1770 aa = (struct lov_brw_async_args *)&set->set_args;
1772 aa->aa_obdos = obdos;
1774 aa->aa_ioarr = ioarr;
1775 aa->aa_oa_bufs = oa_bufs;
1777 /* Don't free ioarr or obdos - that's done in lov_brw_interpret */
1778 GOTO(out_where, rc);
1781 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1783 OBD_FREE(obdos, stripe_count * sizeof(*obdos));
1785 OBD_FREE(where, sizeof(*where) * oa_bufs);
1787 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1791 struct lov_async_page *lap_from_cookie(void *cookie)
1793 struct lov_async_page *lap = cookie;
1794 if (lap->lap_magic != LAP_MAGIC)
1795 return ERR_PTR(-EINVAL);
1799 static int lov_ap_make_ready(void *data, int cmd)
1801 struct lov_async_page *lap = lap_from_cookie(data);
1802 /* XXX should these assert? */
1806 return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
1808 static int lov_ap_refresh_count(void *data, int cmd)
1810 struct lov_async_page *lap = lap_from_cookie(data);
1814 return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
1817 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
1819 struct lov_async_page *lap = lap_from_cookie(data);
1820 /* XXX should these assert? */
1824 lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
1825 /* XXX woah, shouldn't we be altering more here? size? */
1826 oa->o_id = lap->lap_loi_id;
1829 static void lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
1831 struct lov_async_page *lap = lap_from_cookie(data);
1835 /* in a raid1 regime this would down a count of many ios
1836 * in flight, onl calling the caller_ops completion when all
1837 * the raid1 ios are complete */
1838 lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, oa, rc);
1841 static struct obd_async_page_ops lov_async_page_ops = {
1842 .ap_make_ready = lov_ap_make_ready,
1843 .ap_refresh_count = lov_ap_refresh_count,
1844 .ap_fill_obdo = lov_ap_fill_obdo,
1845 .ap_completion = lov_ap_completion,
1848 int lov_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1849 struct lov_oinfo *loi, struct page *page,
1850 obd_off offset, struct obd_async_page_ops *ops,
1851 void *data, void **res)
1853 struct lov_obd *lov = &exp->exp_obd->u.lov;
1854 struct lov_async_page *lap;
1858 if (lsm_bad_magic(lsm))
1860 LASSERT(loi == NULL);
1862 OBD_ALLOC(lap, sizeof(*lap));
1866 lap->lap_magic = LAP_MAGIC;
1867 lap->lap_caller_ops = ops;
1868 lap->lap_caller_data = data;
1870 /* FIXME handle multiple oscs after landing b_raid1 */
1871 switch (lsm->lsm_pattern) {
1872 case LOV_PATTERN_RAID0:
1873 lap->lap_stripe = lov_stripe_number(lsm, offset);
1874 lov_stripe_offset(lsm, offset, lap->lap_stripe,
1875 &lap->lap_sub_offset);
1877 case LOV_PATTERN_CMOBD:
1878 lap->lap_stripe = 0;
1879 lap->lap_sub_offset = offset;
1884 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1886 /* so the callback doesn't need the lsm */
1887 lap->lap_loi_id = loi->loi_id;
1889 rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
1890 lsm, loi, page, lap->lap_sub_offset,
1891 &lov_async_page_ops, lap,
1892 &lap->lap_sub_cookie);
1894 OBD_FREE(lap, sizeof(*lap));
1897 CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
1898 lap->lap_sub_cookie, offset);
1903 static int lov_queue_async_io(struct obd_export *exp,
1904 struct lov_stripe_md *lsm,
1905 struct lov_oinfo *loi, void *cookie,
1906 int cmd, obd_off off, int count,
1907 obd_flag brw_flags, obd_flag async_flags)
1909 struct lov_obd *lov = &exp->exp_obd->u.lov;
1910 struct lov_async_page *lap;
1913 LASSERT(loi == NULL);
1915 if (lsm_bad_magic(lsm))
1918 lap = lap_from_cookie(cookie);
1920 RETURN(PTR_ERR(lap));
1922 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1923 rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
1924 loi, lap->lap_sub_cookie, cmd, off, count,
1925 brw_flags, async_flags);
1929 static int lov_set_async_flags(struct obd_export *exp,
1930 struct lov_stripe_md *lsm,
1931 struct lov_oinfo *loi, void *cookie,
1932 obd_flag async_flags)
1934 struct lov_obd *lov = &exp->exp_obd->u.lov;
1935 struct lov_async_page *lap;
1938 LASSERT(loi == NULL);
1940 if (lsm_bad_magic(lsm))
1943 lap = lap_from_cookie(cookie);
1945 RETURN(PTR_ERR(lap));
1947 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1948 rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
1949 lsm, loi, lap->lap_sub_cookie, async_flags);
1953 static int lov_queue_group_io(struct obd_export *exp,
1954 struct lov_stripe_md *lsm,
1955 struct lov_oinfo *loi,
1956 struct obd_io_group *oig, void *cookie,
1957 int cmd, obd_off off, int count,
1958 obd_flag brw_flags, obd_flag async_flags)
1960 struct lov_obd *lov = &exp->exp_obd->u.lov;
1961 struct lov_async_page *lap;
1964 LASSERT(loi == NULL);
1966 if (lsm_bad_magic(lsm))
1969 lap = lap_from_cookie(cookie);
1971 RETURN(PTR_ERR(lap));
1973 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1974 rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
1975 oig, lap->lap_sub_cookie, cmd, off, count,
1976 brw_flags, async_flags);
1980 /* this isn't exactly optimal. we may have queued sync io in oscs on
1981 * all stripes, but we don't record that fact at queue time. so we
1982 * trigger sync io on all stripes. */
1983 static int lov_trigger_group_io(struct obd_export *exp,
1984 struct lov_stripe_md *lsm,
1985 struct lov_oinfo *loi,
1986 struct obd_io_group *oig)
1988 struct lov_obd *lov = &exp->exp_obd->u.lov;
1991 LASSERT(loi == NULL);
1993 if (lsm_bad_magic(lsm))
1996 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1998 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
2000 if (rc == 0 && err != 0)
2006 static int lov_teardown_async_page(struct obd_export *exp,
2007 struct lov_stripe_md *lsm,
2008 struct lov_oinfo *loi, void *cookie)
2010 struct lov_obd *lov = &exp->exp_obd->u.lov;
2011 struct lov_async_page *lap;
2014 LASSERT(loi == NULL);
2016 if (lsm_bad_magic(lsm))
2019 lap = lap_from_cookie(cookie);
2021 RETURN(PTR_ERR(lap));
2023 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2024 rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2025 lsm, loi, lap->lap_sub_cookie);
2027 CERROR("unable to teardown sub cookie %p: %d\n",
2028 lap->lap_sub_cookie, rc);
2031 OBD_FREE(lap, sizeof(*lap));
2035 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2036 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2037 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2038 void *data,__u32 lvb_len, void *lvb_swabber,
2039 struct lustre_handle *lockh)
2041 struct lov_lock_handles *lov_lockh = NULL;
2042 struct lustre_handle *lov_lockhp;
2043 struct lov_obd *lov;
2044 struct lov_oinfo *loi;
2045 char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)];
2046 struct lov_stripe_md *submd = (void *)submd_buf;
2048 int i, save_flags = *flags;
2051 if (lsm_bad_magic(lsm))
2054 /* we should never be asked to replay a lock this way. */
2055 LASSERT((*flags & LDLM_FL_REPLAY) == 0);
2057 if (!exp || !exp->exp_obd)
2060 if (lsm->lsm_stripe_count > 1) {
2061 lov_lockh = lov_llh_new(lsm);
2062 if (lov_lockh == NULL)
2065 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2066 lov_lockhp = lov_lockh->llh_handles;
2071 lov = &exp->exp_obd->u.lov;
2072 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2073 i++, loi++, lov_lockhp++) {
2074 ldlm_policy_data_t sub_ext;
2077 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2078 policy->l_extent.end, &start,
2082 sub_ext.l_extent.start = start;
2083 sub_ext.l_extent.end = end;
2084 sub_ext.l_extent.gid = policy->l_extent.gid;
2086 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2087 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2091 /* XXX LOV STACKING: submd should be from the subobj */
2092 submd->lsm_object_id = loi->loi_id;
2093 submd->lsm_object_gr = lsm->lsm_object_gr;
2094 submd->lsm_stripe_count = 0;
2095 submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid;
2096 submd->lsm_oinfo->loi_rss = loi->loi_rss;
2097 submd->lsm_oinfo->loi_kms = loi->loi_kms;
2098 submd->lsm_oinfo->loi_blocks = loi->loi_blocks;
2099 loi->loi_mtime = submd->lsm_oinfo->loi_mtime;
2100 /* XXX submd is not fully initialized here */
2101 *flags = save_flags;
2102 rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd,
2103 type, &sub_ext, mode, flags, bl_cb, cp_cb,
2104 gl_cb, data, lvb_len, lvb_swabber, lov_lockhp);
2106 /* XXX FIXME: This unpleasantness doesn't belong here at *all*.
2107 * It belongs in the OSC, except that the OSC doesn't have
2108 * access to the real LOI -- it gets a copy, that we created
2109 * above, and that copy can be arbitrarily out of date.
2111 * The LOV API is due for a serious rewriting anyways, and this
2112 * can be addressed then. */
2113 if (rc == ELDLM_OK) {
2114 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2115 __u64 tmp = submd->lsm_oinfo->loi_rss;
2117 LASSERT(lock != NULL);
2119 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2120 /* Extend KMS up to the end of this lock and no further
2121 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2122 if (tmp > lock->l_policy_data.l_extent.end)
2123 tmp = lock->l_policy_data.l_extent.end + 1;
2124 if (tmp >= loi->loi_kms) {
2125 CDEBUG(D_INODE, "lock acquired, setting rss="
2126 LPU64", kms="LPU64"\n", loi->loi_rss,
2129 loi->loi_kms_valid = 1;
2131 CDEBUG(D_INODE, "lock acquired, setting rss="
2132 LPU64"; leaving kms="LPU64", end="LPU64
2133 "\n", loi->loi_rss, loi->loi_kms,
2134 lock->l_policy_data.l_extent.end);
2136 ldlm_lock_allow_match(lock);
2137 LDLM_LOCK_PUT(lock);
2138 } else if (rc == ELDLM_LOCK_ABORTED &&
2139 save_flags & LDLM_FL_HAS_INTENT) {
2140 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2141 loi->loi_rss = submd->lsm_oinfo->loi_rss;
2142 loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2143 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2144 " kms="LPU64"\n", loi->loi_rss, loi->loi_kms);
2146 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2147 if (lov->tgts[loi->loi_ost_idx].active) {
2148 CERROR("error: enqueue objid "LPX64" subobj "
2149 LPX64" on OST idx %d: rc = %d\n",
2150 lsm->lsm_object_id, loi->loi_id,
2151 loi->loi_ost_idx, rc);
2152 GOTO(out_locks, rc);
2156 if (lsm->lsm_stripe_count > 1)
2157 lov_llh_put(lov_lockh);
2161 while (loi--, lov_lockhp--, i-- > 0) {
2162 struct lov_stripe_md submd;
2165 if (lov_lockhp->cookie == 0)
2168 /* XXX LOV STACKING: submd should be from the subobj */
2169 submd.lsm_object_id = loi->loi_id;
2170 submd.lsm_object_gr = lsm->lsm_object_gr;
2171 submd.lsm_stripe_count = 0;
2172 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2174 if (err && lov->tgts[loi->loi_ost_idx].active) {
2175 CERROR("error: cancelling objid "LPX64" on OST "
2176 "idx %d after enqueue error: rc = %d\n",
2177 loi->loi_id, loi->loi_ost_idx, err);
2181 if (lsm->lsm_stripe_count > 1) {
2182 lov_llh_destroy(lov_lockh);
2183 lov_llh_put(lov_lockh);
2188 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2189 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2190 int *flags, void *data, struct lustre_handle *lockh)
2192 struct lov_lock_handles *lov_lockh = NULL;
2193 struct lustre_handle *lov_lockhp;
2194 struct lov_obd *lov;
2195 struct lov_oinfo *loi;
2196 struct lov_stripe_md submd;
2197 ldlm_error_t rc = 0;
2201 if (lsm_bad_magic(lsm))
2204 if (!exp || !exp->exp_obd)
2207 if (lsm->lsm_stripe_count > 1) {
2208 lov_lockh = lov_llh_new(lsm);
2209 if (lov_lockh == NULL)
2212 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2213 lov_lockhp = lov_lockh->llh_handles;
2218 lov = &exp->exp_obd->u.lov;
2219 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2220 i++, loi++, lov_lockhp++) {
2221 ldlm_policy_data_t sub_ext;
2225 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2226 policy->l_extent.end, &start, &end))
2229 sub_ext.l_extent.start = start;
2230 sub_ext.l_extent.end = end;
2232 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2233 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2238 /* XXX LOV STACKING: submd should be from the subobj */
2239 submd.lsm_object_id = loi->loi_id;
2240 submd.lsm_object_gr = lsm->lsm_object_gr;
2241 submd.lsm_stripe_count = 0;
2243 /* XXX submd is not fully initialized here */
2244 rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2245 type, &sub_ext, mode, &lov_flags, data,
2251 if (lsm->lsm_stripe_count > 1) {
2252 if (*flags & LDLM_FL_TEST_LOCK)
2253 lov_llh_destroy(lov_lockh);
2254 lov_llh_put(lov_lockh);
2259 while (loi--, lov_lockhp--, i-- > 0) {
2260 struct lov_stripe_md submd;
2263 if (lov_lockhp->cookie == 0)
2266 /* XXX LOV STACKING: submd should be from the subobj */
2267 submd.lsm_object_id = loi->loi_id;
2268 submd.lsm_object_gr = lsm->lsm_object_gr;
2269 submd.lsm_stripe_count = 0;
2270 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2272 if (err && lov->tgts[loi->loi_ost_idx].active) {
2273 CERROR("error: cancelling objid "LPX64" on OST "
2274 "idx %d after match failure: rc = %d\n",
2275 loi->loi_id, loi->loi_ost_idx, err);
2279 if (lsm->lsm_stripe_count > 1) {
2280 lov_llh_destroy(lov_lockh);
2281 lov_llh_put(lov_lockh);
2286 static int lov_change_cbdata(struct obd_export *exp,
2287 struct lov_stripe_md *lsm, ldlm_iterator_t it,
2290 struct lov_obd *lov;
2291 struct lov_oinfo *loi;
2295 if (lsm_bad_magic(lsm))
2298 if (!exp || !exp->exp_obd)
2301 LASSERT(lsm->lsm_object_gr > 0);
2303 lov = &exp->exp_obd->u.lov;
2304 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2305 struct lov_stripe_md submd;
2306 if (lov->tgts[loi->loi_ost_idx].active == 0)
2307 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2309 submd.lsm_object_id = loi->loi_id;
2310 submd.lsm_object_gr = lsm->lsm_object_gr;
2311 submd.lsm_stripe_count = 0;
2312 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
2318 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
2319 __u32 mode, struct lustre_handle *lockh)
2321 struct lov_lock_handles *lov_lockh = NULL;
2322 struct lustre_handle *lov_lockhp;
2323 struct lov_obd *lov;
2324 struct lov_oinfo *loi;
2328 if (lsm_bad_magic(lsm))
2331 if (!exp || !exp->exp_obd)
2334 LASSERT(lsm->lsm_object_gr > 0);
2337 if (lsm->lsm_stripe_count > 1) {
2338 lov_lockh = lov_handle2llh(lockh);
2340 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2344 lov_lockhp = lov_lockh->llh_handles;
2349 lov = &exp->exp_obd->u.lov;
2350 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2351 i++, loi++, lov_lockhp++) {
2352 struct lov_stripe_md submd;
2355 if (lov_lockhp->cookie == 0) {
2356 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2357 loi->loi_ost_idx, loi->loi_id);
2361 /* XXX LOV STACKING: submd should be from the subobj */
2362 submd.lsm_object_id = loi->loi_id;
2363 submd.lsm_object_gr = lsm->lsm_object_gr;
2364 submd.lsm_stripe_count = 0;
2365 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2368 if (lov->tgts[loi->loi_ost_idx].active) {
2369 CERROR("error: cancel objid "LPX64" subobj "
2370 LPX64" on OST idx %d: rc = %d\n",
2372 loi->loi_id, loi->loi_ost_idx, err);
2379 if (lsm->lsm_stripe_count > 1)
2380 lov_llh_destroy(lov_lockh);
2381 if (lov_lockh != NULL)
2382 lov_llh_put(lov_lockh);
2386 static int lov_cancel_unused(struct obd_export *exp,
2387 struct lov_stripe_md *lsm, int flags, void *opaque)
2389 struct lov_obd *lov;
2390 struct lov_oinfo *loi;
2394 if (lsm_bad_magic(lsm))
2397 if (!exp || !exp->exp_obd)
2400 LASSERT(lsm->lsm_object_gr > 0);
2402 lov = &exp->exp_obd->u.lov;
2403 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2404 struct lov_stripe_md submd;
2407 if (lov->tgts[loi->loi_ost_idx].active == 0)
2408 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2410 submd.lsm_object_id = loi->loi_id;
2411 submd.lsm_object_gr = lsm->lsm_object_gr;
2412 submd.lsm_stripe_count = 0;
2413 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
2414 &submd, flags, opaque);
2415 if (err && lov->tgts[loi->loi_ost_idx].active) {
2416 CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2417 " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2418 loi->loi_id, loi->loi_ost_idx, err);
2426 #define LOV_U64_MAX ((__u64)~0ULL)
2427 #define LOV_SUM_MAX(tot, add) \
2429 if ((tot) + (add) < (tot)) \
2430 (tot) = LOV_U64_MAX; \
2435 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2436 unsigned long max_age)
2438 struct lov_obd *lov = &obd->u.lov;
2439 struct obd_statfs lov_sfs;
2446 /* We only get block data from the OBD */
2447 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2450 if (!lov->tgts[i].active) {
2451 CDEBUG(D_HA, "lov idx %d inactive\n", i);
2455 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
2458 if (lov->tgts[i].active && !rc)
2464 memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
2467 osfs->os_bfree += lov_sfs.os_bfree;
2468 osfs->os_bavail += lov_sfs.os_bavail;
2469 osfs->os_blocks += lov_sfs.os_blocks;
2470 /* XXX not sure about this one - depends on policy.
2471 * - could be minimum if we always stripe on all OBDs
2472 * (but that would be wrong for any other policy,
2473 * if one of the OBDs has no more objects left)
2474 * - could be sum if we stripe whole objects
2475 * - could be average, just to give a nice number
2477 * To give a "reasonable" (if not wholly accurate)
2478 * number, we divide the total number of free objects
2479 * by expected stripe count (watch out for overflow).
2481 LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
2482 LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
2487 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
2488 lov->desc.ld_default_stripe_count :
2489 lov->desc.ld_active_tgt_count;
2491 if (osfs->os_files != LOV_U64_MAX)
2492 do_div(osfs->os_files, expected_stripes);
2493 if (osfs->os_ffree != LOV_U64_MAX)
2494 do_div(osfs->os_ffree, expected_stripes);
2501 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2502 void *karg, void *uarg)
2504 struct obd_device *obddev = class_exp2obd(exp);
2505 struct lov_obd *lov = &obddev->u.lov;
2506 int i, count = lov->desc.ld_tgt_count;
2507 struct obd_uuid *uuidp;
2513 case OBD_IOC_LOV_GET_CONFIG: {
2514 struct obd_ioctl_data *data = karg;
2515 struct lov_tgt_desc *tgtdesc;
2516 struct lov_desc *desc;
2521 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2524 data = (struct obd_ioctl_data *)buf;
2526 if (sizeof(*desc) > data->ioc_inllen1) {
2531 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2536 desc = (struct lov_desc *)data->ioc_inlbuf1;
2537 memcpy(desc, &(lov->desc), sizeof(*desc));
2539 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2540 tgtdesc = lov->tgts;
2541 for (i = 0; i < count; i++, uuidp++, tgtdesc++)
2542 obd_str2uuid(uuidp, tgtdesc->uuid.uuid);
2544 rc = copy_to_user((void *)uarg, buf, len);
2547 obd_ioctl_freedata(buf, len);
2550 case LL_IOC_LOV_SETSTRIPE:
2551 rc = lov_setstripe(exp, karg, uarg);
2553 case LL_IOC_LOV_GETSTRIPE:
2554 rc = lov_getstripe(exp, karg, uarg);
2556 case LL_IOC_LOV_SETEA:
2557 rc = lov_setea(exp, karg, uarg);
2564 for (i = 0; i < count; i++) {
2567 err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
2570 if (lov->tgts[i].active) {
2571 CERROR("error: iocontrol OSC %s on OST"
2572 "idx %d: err = %d\n",
2573 lov->tgts[i].uuid.uuid, i, err);
2588 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2589 void *key, __u32 *vallen, void *val)
2591 struct obd_device *obddev = class_exp2obd(exp);
2592 struct lov_obd *lov = &obddev->u.lov;
2596 if (!vallen || !val)
2599 if (keylen > strlen("lock_to_stripe") &&
2600 strcmp(key, "lock_to_stripe") == 0) {
2603 struct ldlm_lock *lock;
2604 struct lov_stripe_md *lsm;
2606 struct lov_oinfo *loi;
2607 __u32 *stripe = val;
2609 if (*vallen < sizeof(*stripe))
2611 *vallen = sizeof(*stripe);
2613 /* XXX This is another one of those bits that will need to
2614 * change if we ever actually support nested LOVs. It uses
2615 * the lock's export to find out which stripe it is. */
2616 for (i = 0, loi = data->lsm->lsm_oinfo;
2617 i < data->lsm->lsm_stripe_count;
2619 if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
2620 data->lock->l_conn_export) {
2626 } else if (keylen >= strlen("size_to_stripe") &&
2627 strcmp(key, "size_to_stripe") == 0) {
2631 struct lov_stripe_md *lsm;
2634 if (*vallen < sizeof(*data))
2637 data->size = lov_size_to_stripe(data->lsm, data->size,
2638 data->stripe_number);
2640 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2642 int rc, size = sizeof(obd_id);
2643 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2644 if (!lov->tgts[i].active)
2646 rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key,
2652 } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
2653 struct lov_desc *desc_ret = val;
2654 *desc_ret = lov->desc;
2662 static int lov_set_info(struct obd_export *exp, obd_count keylen,
2663 void *key, obd_count vallen, void *val)
2665 struct obd_device *obddev = class_exp2obd(exp);
2666 struct lov_obd *lov = &obddev->u.lov;
2670 #define KEY_IS(str) \
2671 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
2673 if (KEY_IS("next_id")) {
2674 if (vallen != lov->desc.ld_tgt_count)
2676 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2679 /* initialize all OSCs, even inactive ones */
2681 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key,
2682 sizeof(obd_id), ((obd_id*)val) + i);
2688 if (KEY_IS("growth_count")) {
2689 if (vallen != sizeof(int))
2691 } else if (KEY_IS("mds_conn")) {
2692 if (vallen != sizeof(__u32))
2694 } else if (KEY_IS("unlinked") || KEY_IS("unrecovery")) {
2701 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2704 if (!val && !lov->tgts[i].active)
2707 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen,
2717 /* Merge rss if kms == 0
2719 * Even when merging RSS, we will take the KMS value if it's larger.
2720 * This prevents getattr from stomping on dirty cached pages which
2721 * extend the file size. */
2722 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms)
2724 struct lov_oinfo *loi;
2728 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2730 obd_size lov_size, tmpsize;
2732 tmpsize = loi->loi_kms;
2733 if (kms == 0 && loi->loi_rss > tmpsize)
2734 tmpsize = loi->loi_rss;
2736 lov_size = lov_stripe_size(lsm, tmpsize, i);
2737 if (lov_size > size)
2743 EXPORT_SYMBOL(lov_merge_size);
2746 __u64 lov_merge_blocks(struct lov_stripe_md *lsm)
2748 struct lov_oinfo *loi;
2752 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2754 blocks += loi->loi_blocks;
2758 EXPORT_SYMBOL(lov_merge_blocks);
2760 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time)
2762 struct lov_oinfo *loi;
2765 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2767 if (loi->loi_mtime > current_time)
2768 current_time = loi->loi_mtime;
2770 return current_time;
2772 EXPORT_SYMBOL(lov_merge_mtime);
2775 struct lov_multi_wait {
2776 struct ldlm_lock *lock;
2782 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
2783 struct lustre_handle *lockh)
2785 struct lov_lock_handles *lov_lockh = NULL;
2786 struct lustre_handle *lov_lockhp;
2787 struct lov_obd *lov;
2788 struct lov_oinfo *loi;
2789 struct lov_multi_wait *queues;
2793 if (lsm_bad_magic(lsm))
2796 if (!exp || !exp->exp_obd)
2799 LASSERT(lockh != NULL);
2800 if (lsm->lsm_stripe_count > 1) {
2801 lov_lockh = lov_handle2llh(lockh);
2802 if (lov_lockh == NULL) {
2803 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2807 lov_lockhp = lov_lockh->llh_handles;
2812 OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
2814 GOTO(out, rc = -ENOMEM);
2816 lov = &exp->exp_obd->u.lov;
2817 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2818 i++, loi++, lov_lockhp++) {
2819 struct ldlm_lock *lock;
2820 struct obd_device *obd;
2821 unsigned long irqflags;
2823 lock = ldlm_handle2lock(lov_lockhp);
2825 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2826 loi->loi_ost_idx, loi->loi_id);
2827 queues[i].completed = 1;
2831 queues[i].lock = lock;
2832 init_waitqueue_entry(&(queues[i].wait), current);
2833 add_wait_queue(lock->l_waitq, &(queues[i].wait));
2835 obd = class_exp2obd(lock->l_conn_export);
2837 imp = obd->u.cli.cl_import;
2839 spin_lock_irqsave(&imp->imp_lock, irqflags);
2840 queues[i].generation = imp->imp_generation;
2841 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
2845 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
2846 interrupted_completion_wait, &lwd);
2847 rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
2849 for (i = 0; i < lsm->lsm_stripe_count; i++)
2850 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
2852 if (rc == -EINTR || rc == -ETIMEDOUT) {
2858 if (lov_lockh != NULL)
2859 lov_llh_put(lov_lockh);
2864 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
2867 struct lov_oinfo *loi;
2873 stripe = lov_stripe_number(lsm, size - 1);
2874 kms = lov_size_to_stripe(lsm, size, stripe);
2875 loi = &(lsm->lsm_oinfo[stripe]);
2877 CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n",
2878 stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms);
2879 if (kms > loi->loi_kms)
2883 EXPORT_SYMBOL(lov_increase_kms);
2885 struct obd_ops lov_obd_ops = {
2886 .o_owner = THIS_MODULE,
2887 .o_attach = lov_attach,
2888 .o_detach = lov_detach,
2889 .o_setup = lov_setup,
2890 .o_cleanup = lov_cleanup,
2891 .o_connect = lov_connect,
2892 .o_disconnect = lov_disconnect,
2893 .o_statfs = lov_statfs,
2894 .o_packmd = lov_packmd,
2895 .o_unpackmd = lov_unpackmd,
2896 .o_create = lov_create,
2897 .o_destroy = lov_destroy,
2898 .o_getattr = lov_getattr,
2899 .o_getattr_async = lov_getattr_async,
2900 .o_setattr = lov_setattr,
2902 .o_brw_async = lov_brw_async,
2903 .o_prep_async_page = lov_prep_async_page,
2904 .o_queue_async_io = lov_queue_async_io,
2905 .o_set_async_flags = lov_set_async_flags,
2906 .o_queue_group_io = lov_queue_group_io,
2907 .o_trigger_group_io = lov_trigger_group_io,
2908 .o_teardown_async_page = lov_teardown_async_page,
2909 .o_punch = lov_punch,
2911 .o_enqueue = lov_enqueue,
2912 .o_match = lov_match,
2913 .o_change_cbdata = lov_change_cbdata,
2914 .o_cancel = lov_cancel,
2915 .o_cancel_unused = lov_cancel_unused,
2916 .o_iocontrol = lov_iocontrol,
2917 .o_get_info = lov_get_info,
2918 .o_set_info = lov_set_info,
2919 .o_llog_init = lov_llog_init,
2920 .o_llog_finish = lov_llog_finish,
2921 .o_notify = lov_notify,
2924 int __init lov_init(void)
2926 struct lprocfs_static_vars lvars;
2929 lprocfs_init_vars(lov, &lvars);
2930 rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars,
2931 OBD_LOV_DEVICENAME);
2936 static void /*__exit*/ lov_exit(void)
2938 class_unregister_type(OBD_LOV_DEVICENAME);
2941 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2942 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
2943 MODULE_LICENSE("GPL");
2945 module_init(lov_init);
2946 module_exit(lov_exit);