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 int lov_attach(struct obd_device *dev, obd_count len, void *data)
119 struct lprocfs_static_vars lvars;
122 lprocfs_init_vars(lov, &lvars);
123 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
126 struct proc_dir_entry *entry;
128 entry = create_proc_entry("target_obd", 0444,
129 dev->obd_proc_entry);
133 entry->proc_fops = &lov_proc_target_fops;
141 int lov_detach(struct obd_device *dev)
143 return lprocfs_obd_detach(dev);
146 static int lov_connect(struct lustre_handle *conn, struct obd_device *obd,
147 struct obd_uuid *cluuid)
149 struct ptlrpc_request *req = NULL;
150 struct lov_obd *lov = &obd->u.lov;
151 struct lov_desc *desc = &lov->desc;
152 struct lov_tgt_desc *tgts;
153 struct obd_export *exp;
157 rc = class_connect(conn, obd, cluuid);
161 exp = class_conn2export(conn);
163 /* We don't want to actually do the underlying connections more than
164 * once, so keep track. */
166 if (lov->refcount > 1) {
167 class_export_put(exp);
171 for (i = 0, tgts = lov->tgts; i < desc->ld_tgt_count; i++, tgts++) {
172 struct obd_uuid *tgt_uuid = &tgts->uuid;
173 struct obd_device *tgt_obd;
174 struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
175 struct lustre_handle conn = {0, };
177 LASSERT( tgt_uuid != NULL);
179 tgt_obd = class_find_client_obd(tgt_uuid, LUSTRE_OSC_NAME,
183 CERROR("Target %s not attached\n", tgt_uuid->uuid);
184 GOTO(out_disc, rc = -EINVAL);
187 if (!tgt_obd->obd_set_up) {
188 CERROR("Target %s not set up\n", tgt_uuid->uuid);
189 GOTO(out_disc, rc = -EINVAL);
192 if (tgt_obd->u.cli.cl_import->imp_invalid) {
193 CERROR("not connecting OSC %s; administratively "
194 "disabled\n", tgt_uuid->uuid);
195 rc = obd_register_observer(tgt_obd, obd);
197 CERROR("Target %s register_observer error %d; "
198 "will not be able to reactivate\n",
204 rc = obd_connect(&conn, tgt_obd, &lov_osc_uuid);
206 CERROR("Target %s connect error %d\n", tgt_uuid->uuid,
210 tgts->ltd_exp = class_conn2export(&conn);
212 rc = obd_register_observer(tgt_obd, obd);
214 CERROR("Target %s register_observer error %d\n",
216 obd_disconnect(tgts->ltd_exp, 0);
220 desc->ld_active_tgt_count++;
224 ptlrpc_req_finished(req);
225 class_export_put(exp);
230 struct obd_uuid uuid;
232 --desc->ld_active_tgt_count;
234 /* save for CERROR below; (we know it's terminated) */
236 rc2 = obd_disconnect(tgts->ltd_exp, 0);
238 CERROR("error: LOV target %s disconnect on OST idx %d: "
239 "rc = %d\n", uuid.uuid, i, rc2);
241 class_disconnect(exp, 0);
245 static int lov_disconnect(struct obd_export *exp, int flags)
247 struct obd_device *obd = class_exp2obd(exp);
248 struct lov_obd *lov = &obd->u.lov;
255 /* Only disconnect the underlying layers on the final disconnect. */
257 if (lov->refcount != 0)
260 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
261 if (lov->tgts[i].ltd_exp == NULL)
264 if (obd->obd_no_recov) {
265 /* Pass it on to our clients.
266 * XXX This should be an argument to disconnect,
267 * XXX not a back-door flag on the OBD. Ah well.
269 struct obd_device *osc_obd;
270 osc_obd = class_exp2obd(lov->tgts[i].ltd_exp);
272 osc_obd->obd_no_recov = 1;
275 obd_register_observer(lov->tgts[i].ltd_exp->exp_obd, NULL);
277 rc = obd_disconnect(lov->tgts[i].ltd_exp, flags);
279 if (lov->tgts[i].active) {
280 CERROR("Target %s disconnect error %d\n",
281 lov->tgts[i].uuid.uuid, rc);
285 if (lov->tgts[i].active) {
286 lov->desc.ld_active_tgt_count--;
287 lov->tgts[i].active = 0;
289 lov->tgts[i].ltd_exp = NULL;
293 rc = class_disconnect(exp, 0);
299 * -EINVAL : UUID can't be found in the LOV's target list
300 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
301 * -EBADF : The UUID is found, but the OBD is the wrong type (!)
303 static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid,
306 struct obd_device *obd;
307 struct lov_tgt_desc *tgt;
311 CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
312 lov, uuid->uuid, activate);
314 spin_lock(&lov->lov_lock);
315 for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
316 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
317 i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
318 if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0)
322 if (i == lov->desc.ld_tgt_count)
323 GOTO(out, rc = -EINVAL);
325 obd = class_exp2obd(tgt->ltd_exp);
327 /* This can happen if OST failure races with node shutdown */
328 GOTO(out, rc = -ENOTCONN);
331 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LOV idx %d\n",
332 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
333 obd->obd_type->typ_name, i);
334 LASSERT(strcmp(obd->obd_type->typ_name, "osc") == 0);
336 if (tgt->active == activate) {
337 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
338 activate ? "" : "in");
342 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd, activate ? "" : "in");
344 tgt->active = activate;
346 lov->desc.ld_active_tgt_count++;
348 lov->desc.ld_active_tgt_count--;
352 spin_unlock(&lov->lov_lock);
356 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
360 struct obd_uuid *uuid;
362 if (strcmp(watched->obd_type->typ_name, "osc")) {
363 CERROR("unexpected notification of %s %s!\n",
364 watched->obd_type->typ_name,
368 uuid = &watched->u.cli.cl_import->imp_target_uuid;
371 * Must notify (MDS) before we mark the OSC as active, so that
372 * the orphan deletion happens without interference from racing
375 if (obd->obd_observer) {
376 /* Pass the notification up the chain. */
377 rc = obd_notify(obd->obd_observer, watched, active);
382 rc = lov_set_osc_active(&obd->u.lov, uuid, active);
385 CERROR("%sactivation of %s failed: %d\n",
386 active ? "" : "de", uuid->uuid, rc);
391 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
393 struct lustre_cfg *lcfg = buf;
394 struct lov_desc *desc;
395 struct lov_obd *lov = &obd->u.lov;
396 struct obd_uuid *uuids;
397 struct lov_tgt_desc *tgts;
403 if (lcfg->lcfg_inllen1 < 1) {
404 CERROR("LOV setup requires a descriptor\n");
408 if (lcfg->lcfg_inllen2 < 1) {
409 CERROR("LOV setup requires an OST UUID list\n");
413 desc = (struct lov_desc *)lcfg->lcfg_inlbuf1;
414 if (sizeof(*desc) > lcfg->lcfg_inllen1) {
415 CERROR("descriptor size wrong: %d > %d\n",
416 (int)sizeof(*desc), lcfg->lcfg_inllen1);
420 count = desc->ld_tgt_count;
421 uuids = (struct obd_uuid *)lcfg->lcfg_inlbuf2;
422 if (sizeof(*uuids) * count != lcfg->lcfg_inllen2) {
423 CERROR("UUID array size wrong: %u * %u != %u\n",
424 (int)sizeof(*uuids), count, lcfg->lcfg_inllen2);
428 /* Because of 64-bit divide/mod operations only work with a 32-bit
429 * divisor in a 32-bit kernel, we cannot support a stripe width
430 * of 4GB or larger on 32-bit CPUs.
432 if ((desc->ld_default_stripe_count ?
433 desc->ld_default_stripe_count : desc->ld_tgt_count) *
434 desc->ld_default_stripe_size > ~0UL) {
435 CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n",
436 desc->ld_default_stripe_size,
437 desc->ld_default_stripe_count ?
438 desc->ld_default_stripe_count : desc->ld_tgt_count,~0UL);
442 lov->bufsize = sizeof(struct lov_tgt_desc) * count;
443 OBD_ALLOC(lov->tgts, lov->bufsize);
444 if (lov->tgts == NULL) {
445 CERROR("Out of memory\n");
450 spin_lock_init(&lov->lov_lock);
452 for (i = 0, tgts = lov->tgts; i < desc->ld_tgt_count; i++, tgts++) {
453 struct obd_uuid *uuid = &tgts->uuid;
455 /* NULL termination already checked */
463 static int lov_cleanup(struct obd_device *obd, int flags)
465 struct lov_obd *lov = &obd->u.lov;
467 OBD_FREE(lov->tgts, lov->bufsize);
472 /* compute object size given "stripeno" and the ost size */
473 static obd_size lov_stripe_size(struct lov_stripe_md *lsm, obd_size ost_size,
476 unsigned long ssize = lsm->lsm_stripe_size;
477 unsigned long swidth = ssize * lsm->lsm_stripe_count;
478 unsigned long stripe_size;
484 /* do_div(a, b) returns a % b, and a = a / b */
485 stripe_size = do_div(ost_size, ssize);
488 lov_size = ost_size * swidth + stripeno * ssize + stripe_size;
490 lov_size = (ost_size - 1) * swidth + (stripeno + 1) * ssize;
495 static void lov_merge_attrs(struct obdo *tgt, struct obdo *src, obd_flag valid,
496 struct lov_stripe_md *lsm, int stripeno, int *set)
498 valid &= src->o_valid;
501 if (valid & OBD_MD_FLSIZE) {
502 /* this handles sparse files properly */
505 lov_size = lov_stripe_size(lsm, src->o_size, stripeno);
506 if (lov_size > tgt->o_size)
507 tgt->o_size = lov_size;
509 if (valid & OBD_MD_FLBLOCKS)
510 tgt->o_blocks += src->o_blocks;
511 if (valid & OBD_MD_FLBLKSZ)
512 tgt->o_blksize += src->o_blksize;
513 if (valid & OBD_MD_FLCTIME && tgt->o_ctime < src->o_ctime)
514 tgt->o_ctime = src->o_ctime;
515 if (valid & OBD_MD_FLMTIME && tgt->o_mtime < src->o_mtime)
516 tgt->o_mtime = src->o_mtime;
518 memcpy(tgt, src, sizeof(*tgt));
519 tgt->o_id = lsm->lsm_object_id;
520 if (valid & OBD_MD_FLSIZE)
521 tgt->o_size = lov_stripe_size(lsm,src->o_size,stripeno);
527 #define log2(n) ffz(~(n))
530 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
531 struct lov_stripe_md **ea,
532 struct obd_trans_info *oti)
536 struct obd_uuid *ost_uuid = NULL;
540 LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
541 src_oa->o_flags == OBD_FL_DELORPHAN);
543 lov = &export->exp_obd->u.lov;
545 tmp_oa = obdo_alloc();
549 if (src_oa->o_valid & OBD_MD_FLINLINE) {
550 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
551 CDEBUG(D_HA, "clearing orphans only for %s\n",
555 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
556 struct lov_stripe_md obj_md;
557 struct lov_stripe_md *obj_mdp = &obj_md;
560 /* if called for a specific target, we don't
561 care if it is not active. */
562 if (lov->tgts[i].active == 0 && ost_uuid == NULL) {
563 CDEBUG(D_HA, "lov idx %d inactive\n", i);
567 if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid))
570 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
572 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
573 err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, &obj_mdp, oti);
575 /* This export will be disabled until it is recovered,
576 and then orphan recovery will be completed. */
577 CERROR("error in orphan recovery on OST idx %d/%d: "
578 "rc = %d\n", i, lov->desc.ld_tgt_count, err);
587 #define LOV_CREATE_RESEED_INTERVAL 1000
589 /* the LOV expects oa->o_id to be set to the LOV object id */
590 static int lov_create(struct obd_export *exp, struct obdo *src_oa,
591 struct lov_stripe_md **ea, struct obd_trans_info *oti)
593 static int ost_start_idx, ost_start_count;
595 struct lov_stripe_md *lsm;
596 struct lov_oinfo *loi = NULL;
597 struct obdo *tmp_oa, *ret_oa;
598 struct llog_cookie *cookies = NULL;
599 unsigned ost_count, ost_idx;
600 int set = 0, obj_alloc = 0, cookie_sent = 0, rc = 0, i;
605 if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
606 src_oa->o_flags == OBD_FL_DELORPHAN) {
607 rc = lov_clear_orphans(exp, src_oa, ea, oti);
614 lov = &exp->exp_obd->u.lov;
616 if (!lov->desc.ld_active_tgt_count)
619 /* Recreate a specific object id at the given OST index */
620 if (src_oa->o_valid & OBD_MD_FLFLAGS && src_oa->o_flags &
621 OBD_FL_RECREATE_OBJS) {
622 struct lov_stripe_md obj_md;
623 struct lov_stripe_md *obj_mdp = &obj_md;
625 ost_idx = src_oa->o_nlink;
629 if (ost_idx >= lov->desc.ld_tgt_count)
631 for (i = 0; i < lsm->lsm_stripe_count; i++) {
632 if (lsm->lsm_oinfo[i].loi_ost_idx == ost_idx) {
633 if (lsm->lsm_oinfo[i].loi_id != src_oa->o_id)
638 if (i == lsm->lsm_stripe_count)
641 rc = obd_create(lov->tgts[ost_idx].ltd_exp, src_oa,
646 ret_oa = obdo_alloc();
650 tmp_oa = obdo_alloc();
652 GOTO(out_oa, rc = -ENOMEM);
657 ost_count = lov_get_stripecnt(lov, 0);
659 /* If the MDS file was truncated up to some size, stripe over
660 * enough OSTs to allow the file to be created at that size. */
661 if (src_oa->o_valid & OBD_MD_FLSIZE) {
662 stripes=((src_oa->o_size+LUSTRE_STRIPE_MAXBYTES)>>12)-1;
663 do_div(stripes, (__u32)(LUSTRE_STRIPE_MAXBYTES >> 12));
665 if (stripes > lov->desc.ld_active_tgt_count)
667 if (stripes > ost_count)
673 rc = lov_alloc_memmd(&lsm, stripes, lov->desc.ld_pattern ?
674 lov->desc.ld_pattern : LOV_PATTERN_RAID0);
681 ost_count = lov->desc.ld_tgt_count;
683 LASSERT(src_oa->o_gr > 0);
684 LASSERT(src_oa->o_valid & OBD_MD_FLID);
685 lsm->lsm_object_id = src_oa->o_id;
686 lsm->lsm_object_gr = src_oa->o_gr;
687 if (!lsm->lsm_stripe_size)
688 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
689 if (!lsm->lsm_pattern) {
690 lsm->lsm_pattern = lov->desc.ld_pattern ?
691 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
694 if (*ea == NULL || lsm->lsm_oinfo[0].loi_ost_idx >= ost_count) {
695 if (--ost_start_count <= 0) {
696 ost_start_idx = ll_insecure_random_int();
697 ost_start_count = LOV_CREATE_RESEED_INTERVAL;
698 } else if (lsm->lsm_stripe_count >=
699 lov->desc.ld_active_tgt_count) {
700 /* If we allocate from all of the stripes, make the
701 * next file start on the next OST. */
704 ost_idx = ost_start_idx % ost_count;
706 ost_idx = lsm->lsm_oinfo[0].loi_ost_idx;
709 CDEBUG(D_INODE, "allocating %d subobjs for objid "LPX64" at idx %d\n",
710 lsm->lsm_stripe_count, lsm->lsm_object_id, ost_idx);
712 /* XXX LOV STACKING: need to figure out how many real OSCs */
713 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
714 oti_alloc_cookies(oti, lsm->lsm_stripe_count);
715 if (!oti->oti_logcookies)
716 GOTO(out_cleanup, rc = -ENOMEM);
717 cookies = oti->oti_logcookies;
720 loi = lsm->lsm_oinfo;
721 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
722 struct lov_stripe_md obj_md;
723 struct lov_stripe_md *obj_mdp = &obj_md;
727 if (lov->tgts[ost_idx].active == 0) {
728 CDEBUG(D_HA, "lov idx %d inactive\n", ost_idx);
732 /* create data objects with "parent" OA */
733 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
735 /* XXX When we start creating objects on demand, we need to
736 * make sure that we always create the object on the
737 * stripe which holds the existing file size.
739 if (src_oa->o_valid & OBD_MD_FLSIZE) {
740 if (lov_stripe_offset(lsm, src_oa->o_size, i,
741 &tmp_oa->o_size) < 0 &&
745 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
746 i, tmp_oa->o_size, src_oa->o_size);
750 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
751 err = obd_create(lov->tgts[ost_idx].ltd_exp, tmp_oa, &obj_mdp,
754 if (lov->tgts[ost_idx].active) {
755 CERROR("error creating objid "LPX64" sub-object"
756 " on OST idx %d/%d: rc = %d\n",
757 src_oa->o_id, ost_idx,
758 lsm->lsm_stripe_count, err);
760 CERROR("obd_create returned invalid "
770 oti->oti_objid[ost_idx] = tmp_oa->o_id;
771 loi->loi_id = tmp_oa->o_id;
772 loi->loi_gr = tmp_oa->o_gr;
773 loi->loi_ost_idx = ost_idx;
775 "objid "LPX64" has subobj "LPX64"/"LPX64" at idx %d\n",
776 lsm->lsm_object_id, loi->loi_id, loi->loi_id, ost_idx);
778 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
783 ++oti->oti_logcookies;
784 if (tmp_oa->o_valid & OBD_MD_FLCOOKIE)
789 /* If we have allocated enough objects, we are OK */
790 if (obj_alloc == lsm->lsm_stripe_count)
791 GOTO(out_done, rc = 0);
794 if (obj_alloc == 0) {
797 GOTO(out_cleanup, rc);
800 /* If we were passed specific striping params, then a failure to
801 * meet those requirements is an error, since we can't reallocate
802 * that memory (it might be part of a larger array or something).
804 * We can only get here if lsm_stripe_count was originally > 1.
807 CERROR("can't lstripe objid "LPX64": have %u want %u, rc %d\n",
808 lsm->lsm_object_id, obj_alloc, lsm->lsm_stripe_count,rc);
811 GOTO(out_cleanup, rc);
813 struct lov_stripe_md *lsm_new;
814 /* XXX LOV STACKING call into osc for sizes */
815 unsigned oldsize, newsize;
817 if (oti && cookies && cookie_sent) {
818 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
819 newsize = obj_alloc * sizeof(*cookies);
821 oti_alloc_cookies(oti, obj_alloc);
822 if (oti->oti_logcookies) {
823 memcpy(oti->oti_logcookies, cookies, newsize);
824 OBD_FREE(cookies, oldsize);
825 cookies = oti->oti_logcookies;
827 CWARN("'leaking' %d bytes\n", oldsize-newsize);
831 CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
832 lsm->lsm_object_id, lsm->lsm_stripe_count, obj_alloc);
833 oldsize = lov_stripe_md_size(lsm->lsm_stripe_count);
834 newsize = lov_stripe_md_size(obj_alloc);
835 OBD_ALLOC(lsm_new, newsize);
836 if (lsm_new != NULL) {
837 memcpy(lsm_new, lsm, newsize);
838 lsm_new->lsm_stripe_count = obj_alloc;
839 OBD_FREE(lsm, newsize);
842 CWARN("'leaking' %d bytes\n", oldsize - newsize);
849 if (src_oa->o_valid & OBD_MD_FLSIZE &&
850 ret_oa->o_size != src_oa->o_size) {
851 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
852 src_oa->o_size, ret_oa->o_size);
855 ret_oa->o_id = src_oa->o_id;
856 ret_oa->o_gr = src_oa->o_gr;
857 ret_oa->o_valid |= OBD_MD_FLGROUP;
858 memcpy(src_oa, ret_oa, sizeof(*src_oa));
864 if (oti && cookies) {
865 oti->oti_logcookies = cookies;
867 oti_free_cookies(oti);
868 src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
870 src_oa->o_valid |= OBD_MD_FLCOOKIE;
876 while (obj_alloc-- > 0) {
877 struct obd_export *sub_exp;
881 sub_exp = lov->tgts[loi->loi_ost_idx].ltd_exp;
882 /* destroy already created objects here */
883 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
884 tmp_oa->o_id = loi->loi_id;
886 err = obd_destroy(sub_exp, tmp_oa, NULL, oti);
888 CERROR("Failed to uncreate objid "LPX64" subobj "LPX64
889 " on OST idx %d: rc = %d\n", src_oa->o_id,
890 loi->loi_id, loi->loi_ost_idx, err);
893 obd_free_memmd(exp, &lsm);
897 #define lsm_bad_magic(LSMP) \
899 struct lov_stripe_md *_lsm__ = (LSMP); \
902 CERROR("LOV requires striping ea\n"); \
904 } else if (_lsm__->lsm_magic != LOV_MAGIC) { \
905 CERROR("LOV striping magic bad %#x != %#x\n", \
906 _lsm__->lsm_magic, LOV_MAGIC); \
912 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
913 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
917 struct lov_oinfo *loi;
921 if (lsm_bad_magic(lsm))
924 if (!exp || !exp->exp_obd)
927 lov = &exp->exp_obd->u.lov;
928 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
930 if (lov->tgts[loi->loi_ost_idx].active == 0) {
931 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
932 /* Orphan clean up will (someday) fix this up. */
936 memcpy(&tmp, oa, sizeof(tmp));
937 tmp.o_id = loi->loi_id;
938 err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
940 if (err && lov->tgts[loi->loi_ost_idx].active) {
941 CERROR("error: destroying objid "LPX64" subobj "
942 LPX64" on OST idx %d: rc = %d\n",
943 oa->o_id, loi->loi_id, loi->loi_ost_idx, err);
951 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
952 struct lov_stripe_md *lsm)
956 struct lov_oinfo *loi;
957 int i, rc = 0, set = 0;
960 if (lsm_bad_magic(lsm))
963 if (!exp || !exp->exp_obd)
966 lov = &exp->exp_obd->u.lov;
968 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
969 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
970 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
973 if (lov->tgts[loi->loi_ost_idx].active == 0) {
974 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
978 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
979 "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
980 /* create data objects with "parent" OA */
981 memcpy(&tmp, oa, sizeof(tmp));
982 tmp.o_id = loi->loi_id;
984 err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
987 if (lov->tgts[loi->loi_ost_idx].active) {
988 CERROR("error: getattr objid "LPX64" subobj "
989 LPX64" on OST idx %d: rc = %d\n",
990 oa->o_id, loi->loi_id, loi->loi_ost_idx,
995 lov_merge_attrs(oa, &tmp, tmp.o_valid, lsm, i, &set);
1003 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
1006 struct lov_getattr_async_args *aa = data;
1007 struct lov_stripe_md *lsm = aa->aa_lsm;
1008 struct obdo *oa = aa->aa_oa;
1009 struct obdo *obdos = aa->aa_obdos;
1010 struct lov_oinfo *loi;
1016 /* NB all stripe requests succeeded to get here */
1018 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1020 if (obdos[i].o_valid == 0) /* inactive stripe */
1023 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1028 CERROR ("No stripes had valid attrs\n");
1033 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1037 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
1038 struct lov_stripe_md *lsm,
1039 struct ptlrpc_request_set *rqset)
1042 struct lov_obd *lov;
1043 struct lov_oinfo *loi;
1044 struct lov_getattr_async_args *aa;
1045 int i, rc = 0, set = 0;
1048 if (lsm_bad_magic(lsm))
1051 if (!exp || !exp->exp_obd)
1054 lov = &exp->exp_obd->u.lov;
1056 OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1060 CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1061 lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1062 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1065 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1066 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1067 /* leaves obdos[i].obd_valid unset */
1071 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1072 "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
1073 /* create data objects with "parent" OA */
1074 memcpy(&obdos[i], oa, sizeof(obdos[i]));
1075 obdos[i].o_id = loi->loi_id;
1077 err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp,
1078 &obdos[i], NULL, rqset);
1080 CERROR("error: getattr objid "LPX64" subobj "
1081 LPX64" on OST idx %d: rc = %d\n",
1082 oa->o_id, loi->loi_id, loi->loi_ost_idx,
1084 GOTO(out_obdos, rc = err);
1089 GOTO (out_obdos, rc = -EIO);
1091 LASSERT (rqset->set_interpret == NULL);
1092 rqset->set_interpret = lov_getattr_interpret;
1093 LASSERT (sizeof (rqset->set_args) >= sizeof (*aa));
1094 aa = (struct lov_getattr_async_args *)&rqset->set_args;
1097 aa->aa_obdos = obdos;
1102 OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1108 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
1109 struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1111 struct obdo *tmp_oa, *ret_oa;
1112 struct lov_obd *lov;
1113 struct lov_oinfo *loi;
1114 int rc = 0, i, set = 0;
1117 if (lsm_bad_magic(lsm))
1120 if (!exp || !exp->exp_obd)
1123 /* for now, we only expect time updates here */
1124 LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
1125 OBD_MD_FLATIME | OBD_MD_FLMTIME |
1126 OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
1127 OBD_MD_FLSIZE | OBD_MD_FLGROUP)));
1129 LASSERT(!(src_oa->o_valid & OBD_MD_FLGROUP) || src_oa->o_gr > 0);
1131 ret_oa = obdo_alloc();
1135 tmp_oa = obdo_alloc();
1137 GOTO(out_oa, rc = -ENOMEM);
1139 lov = &exp->exp_obd->u.lov;
1140 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1143 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1144 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1148 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1149 tmp_oa->o_id = loi->loi_id;
1150 LASSERT(!(tmp_oa->o_valid & OBD_MD_FLGROUP) || tmp_oa->o_gr>0);
1152 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1153 if (lov_stripe_offset(lsm, src_oa->o_size, i,
1154 &tmp_oa->o_size) < 0 &&
1158 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1159 i, tmp_oa->o_size, src_oa->o_size);
1162 err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa,
1165 if (lov->tgts[loi->loi_ost_idx].active) {
1166 CERROR("error: setattr objid "LPX64" subobj "
1167 LPX64" on OST idx %d: rc = %d\n",
1168 src_oa->o_id, loi->loi_id,
1169 loi->loi_ost_idx, err);
1175 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set);
1180 ret_oa->o_id = src_oa->o_id;
1181 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1190 /* we have an offset in file backed by an lov and want to find out where
1191 * that offset lands in our given stripe of the file. for the easy
1192 * case where the offset is within the stripe, we just have to scale the
1193 * offset down to make it relative to the stripe instead of the lov.
1195 * the harder case is what to do when the offset doesn't intersect the
1196 * stripe. callers will want start offsets clamped ahead to the start
1197 * of the nearest stripe in the file. end offsets similarly clamped to the
1198 * nearest ending byte of a stripe in the file:
1200 * all this function does is move offsets to the nearest region of the
1201 * stripe, and it does its work "mod" the full length of all the stripes.
1202 * consider a file with 3 stripes:
1205 * ---------------------------------------------------------------------
1206 * | 0 | 1 | 2 | 0 | 1 | 2 |
1207 * ---------------------------------------------------------------------
1209 * to find stripe 1's offsets for S and E, it divides by the full stripe
1210 * width and does its math in the context of a single set of stripes:
1213 * -----------------------------------
1215 * -----------------------------------
1217 * it'll notice that E is outside stripe 1 and clamp it to the end of the
1218 * stripe, then multiply it back out by lov_off to give the real offsets in
1222 * ---------------------------------------------------------------------
1223 * | 1 | 1 | 1 | 1 | 1 | 1 |
1224 * ---------------------------------------------------------------------
1226 * it would have done similarly and pulled S forward to the start of a 1
1227 * stripe if, say, S had landed in a 0 stripe.
1229 * this rounding isn't always correct. consider an E lov offset that lands
1230 * on a 0 stripe, the "mod stripe width" math will pull it forward to the
1231 * start of a 1 stripe, when in fact it wanted to be rounded back to the end
1232 * of a previous 1 stripe. this logic is handled by callers and this is why:
1234 * this function returns < 0 when the offset was "before" the stripe and
1235 * was moved forward to the start of the stripe in question; 0 when it
1236 * falls in the stripe and no shifting was done; > 0 when the offset
1237 * was outside the stripe and was pulled back to its final byte. */
1238 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
1239 int stripeno, obd_off *obd_off)
1241 unsigned long ssize = lsm->lsm_stripe_size;
1242 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1243 unsigned long stripe_off, this_stripe;
1246 if (lov_off == OBD_OBJECT_EOF) {
1247 *obd_off = OBD_OBJECT_EOF;
1251 /* do_div(a, b) returns a % b, and a = a / b */
1252 stripe_off = do_div(lov_off, swidth);
1254 this_stripe = stripeno * ssize;
1255 if (stripe_off < this_stripe) {
1259 stripe_off -= this_stripe;
1261 if (stripe_off >= ssize) {
1267 *obd_off = lov_off * ssize + stripe_off;
1271 /* Given a whole-file size and a stripe number, give the file size which
1272 * corresponds to the individual object of that stripe.
1274 * This behaves basically in the same was as lov_stripe_offset, except that
1275 * file sizes falling before the beginning of a stripe are clamped to the end
1276 * of the previous stripe, not the beginning of the next:
1279 * ---------------------------------------------------------------------
1280 * | 0 | 1 | 2 | 0 | 1 | 2 |
1281 * ---------------------------------------------------------------------
1283 * if clamped to stripe 2 becomes:
1286 * ---------------------------------------------------------------------
1287 * | 0 | 1 | 2 | 0 | 1 | 2 |
1288 * ---------------------------------------------------------------------
1290 static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size,
1293 unsigned long ssize = lsm->lsm_stripe_size;
1294 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1295 unsigned long stripe_off, this_stripe;
1297 if (file_size == OBD_OBJECT_EOF)
1298 return OBD_OBJECT_EOF;
1300 /* do_div(a, b) returns a % b, and a = a / b */
1301 stripe_off = do_div(file_size, swidth);
1303 this_stripe = stripeno * ssize;
1304 if (stripe_off < this_stripe) {
1305 /* Move to end of previous stripe, or zero */
1306 if (file_size > 0) {
1313 stripe_off -= this_stripe;
1315 if (stripe_off >= ssize) {
1316 /* Clamp to end of this stripe */
1321 return (file_size * ssize + stripe_off);
1324 /* given an extent in an lov and a stripe, calculate the extent of the stripe
1325 * that is contained within the lov extent. this returns true if the given
1326 * stripe does intersect with the lov extent. */
1327 static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno,
1328 obd_off start, obd_off end,
1329 obd_off *obd_start, obd_off *obd_end)
1331 int start_side, end_side;
1333 start_side = lov_stripe_offset(lsm, start, stripeno, obd_start);
1334 end_side = lov_stripe_offset(lsm, end, stripeno, obd_end);
1336 CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n",
1337 start, end, start_side, *obd_start, *obd_end, end_side);
1339 /* this stripe doesn't intersect the file extent when neither
1340 * start or the end intersected the stripe and obd_start and
1341 * obd_end got rounded up to the save value. */
1342 if (start_side != 0 && end_side != 0 && *obd_start == *obd_end)
1345 /* as mentioned in the lov_stripe_offset commentary, end
1346 * might have been shifted in the wrong direction. This
1347 * happens when an end offset is before the stripe when viewed
1348 * through the "mod stripe size" math. we detect it being shifted
1349 * in the wrong direction and touch it up.
1350 * interestingly, this can't underflow since end must be > start
1351 * if we passed through the previous check.
1352 * (should we assert for that somewhere?) */
1359 /* compute which stripe number "lov_off" will be written into */
1360 static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off)
1362 unsigned long ssize = lsm->lsm_stripe_size;
1363 unsigned long swidth = ssize * lsm->lsm_stripe_count;
1364 unsigned long stripe_off;
1366 stripe_off = do_div(lov_off, swidth);
1368 return stripe_off / ssize;
1371 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1372 * we can send this 'punch' to just the authoritative node and the nodes
1373 * that the punch will affect. */
1374 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1375 struct lov_stripe_md *lsm,
1376 obd_off start, obd_off end, struct obd_trans_info *oti)
1379 struct lov_obd *lov;
1380 struct lov_oinfo *loi;
1384 if (lsm_bad_magic(lsm))
1387 if (!exp || !exp->exp_obd)
1390 lov = &exp->exp_obd->u.lov;
1391 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1392 obd_off starti, endi;
1395 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1396 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1400 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1403 /* create data objects with "parent" OA */
1404 memcpy(&tmp, oa, sizeof(tmp));
1405 tmp.o_id = loi->loi_id;
1407 err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, NULL,
1408 starti, endi, NULL);
1410 if (lov->tgts[loi->loi_ost_idx].active) {
1411 CERROR("error: punch objid "LPX64" subobj "LPX64
1412 " on OST idx %d: rc = %d\n", oa->o_id,
1413 loi->loi_id, loi->loi_ost_idx, err);
1418 loi->loi_kms = loi->loi_rss = starti;
1424 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1425 struct lov_stripe_md *lsm, obd_off start, obd_off end)
1428 struct lov_obd *lov;
1429 struct lov_oinfo *loi;
1433 if (lsm_bad_magic(lsm))
1443 lov = &exp->exp_obd->u.lov;
1444 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1445 obd_off starti, endi;
1448 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1449 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1453 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1456 memcpy(tmp, oa, sizeof(*tmp));
1457 tmp->o_id = loi->loi_id;
1459 err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL,
1462 if (lov->tgts[loi->loi_ost_idx].active) {
1463 CERROR("error: fsync objid "LPX64" subobj "LPX64
1464 " on OST idx %d: rc = %d\n", oa->o_id,
1465 loi->loi_id, loi->loi_ost_idx, err);
1476 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1477 struct lov_stripe_md *lsm,
1478 obd_count oa_bufs, struct brw_page *pga)
1482 /* The caller just wants to know if there's a chance that this
1483 * I/O can succeed */
1484 for (i = 0; i < oa_bufs; i++) {
1485 int stripe = lov_stripe_number(lsm, pga[i].off);
1486 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1487 struct ldlm_extent ext, subext;
1488 ext.start = pga[i].off;
1489 ext.end = pga[i].off + pga[i].count;
1491 if (!lov_stripe_intersects(lsm, i, ext.start, ext.end,
1492 &subext.start, &subext.end))
1495 if (lov->tgts[ost].active == 0) {
1496 CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1499 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[stripe].ltd_exp, oa,
1500 NULL, 1, &pga[i], NULL);
1507 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1508 struct lov_stripe_md *lsm, obd_count oa_bufs,
1509 struct brw_page *pga, struct obd_trans_info *oti)
1515 struct lov_stripe_md lsm;
1517 } *stripeinfo, *si, *si_last;
1518 struct obdo *ret_oa = NULL, *tmp_oa = NULL;
1519 struct lov_obd *lov;
1520 struct brw_page *ioarr;
1521 struct lov_oinfo *loi;
1522 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0;
1525 if (lsm_bad_magic(lsm))
1528 lov = &exp->exp_obd->u.lov;
1530 if (cmd == OBD_BRW_CHECK) {
1531 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1535 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1539 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1541 GOTO(out_sinfo, rc = -ENOMEM);
1543 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1545 GOTO(out_where, rc = -ENOMEM);
1548 ret_oa = obdo_alloc();
1550 GOTO(out_ioarr, rc = -ENOMEM);
1552 tmp_oa = obdo_alloc();
1554 GOTO(out_oa, rc = -ENOMEM);
1557 for (i = 0; i < oa_bufs; i++) {
1558 where[i] = lov_stripe_number(lsm, pga[i].off);
1559 stripeinfo[where[i]].bufct++;
1562 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1563 i < stripe_count; i++, loi++, si_last = si, si++) {
1565 si->index = si_last->index + si_last->bufct;
1566 si->lsm.lsm_object_id = loi->loi_id;
1567 si->lsm.lsm_object_gr = lsm->lsm_object_gr;
1568 si->ost_idx = loi->loi_ost_idx;
1571 for (i = 0; i < oa_bufs; i++) {
1572 int which = where[i];
1575 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1576 LASSERT(shift < oa_bufs);
1577 ioarr[shift] = pga[i];
1578 lov_stripe_offset(lsm, pga[i].off, which, &ioarr[shift].off);
1579 stripeinfo[which].subcount++;
1582 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1583 int shift = si->index;
1585 if (lov->tgts[si->ost_idx].active == 0) {
1586 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1587 GOTO(out_oa, rc = -EIO);
1591 LASSERT(shift < oa_bufs);
1593 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1595 tmp_oa->o_id = si->lsm.lsm_object_id;
1596 rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp,
1597 tmp_oa, &si->lsm, si->bufct,
1598 &ioarr[shift], oti);
1600 GOTO(out_ioarr, rc);
1602 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1607 ret_oa->o_id = src_oa->o_id;
1608 memcpy(src_oa, ret_oa, sizeof(*src_oa));
1617 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1619 OBD_FREE(where, sizeof(*where) * oa_bufs);
1621 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1625 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1628 struct lov_brw_async_args *aa = data;
1629 struct lov_stripe_md *lsm = aa->aa_lsm;
1630 obd_count oa_bufs = aa->aa_oa_bufs;
1631 struct obdo *oa = aa->aa_oa;
1632 struct obdo *obdos = aa->aa_obdos;
1633 struct brw_page *ioarr = aa->aa_ioarr;
1634 struct lov_oinfo *loi;
1639 /* NB all stripe requests succeeded to get here */
1641 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1643 if (obdos[i].o_valid == 0) /* inactive stripe */
1646 lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1651 CERROR("No stripes had valid attrs\n");
1655 oa->o_id = lsm->lsm_object_id;
1657 OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos));
1658 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1662 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1663 struct lov_stripe_md *lsm, obd_count oa_bufs,
1664 struct brw_page *pga, struct ptlrpc_request_set *set,
1665 struct obd_trans_info *oti)
1671 struct lov_stripe_md lsm;
1673 } *stripeinfo, *si, *si_last;
1674 struct lov_obd *lov;
1675 struct brw_page *ioarr;
1676 struct obdo *obdos = NULL;
1677 struct lov_oinfo *loi;
1678 struct lov_brw_async_args *aa;
1679 int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count;
1682 if (lsm_bad_magic(lsm))
1685 lov = &exp->exp_obd->u.lov;
1687 if (cmd == OBD_BRW_CHECK) {
1688 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
1692 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1696 OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1698 GOTO(out_sinfo, rc = -ENOMEM);
1701 OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count);
1703 GOTO(out_where, rc = -ENOMEM);
1706 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1708 GOTO(out_obdos, rc = -ENOMEM);
1710 for (i = 0; i < oa_bufs; i++) {
1711 where[i] = lov_stripe_number(lsm, pga[i].off);
1712 stripeinfo[where[i]].bufct++;
1715 for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1716 i < stripe_count; i++, loi++, si_last = si, si++) {
1718 si->index = si_last->index + si_last->bufct;
1719 si->lsm.lsm_object_id = loi->loi_id;
1720 si->ost_idx = loi->loi_ost_idx;
1723 memcpy(&obdos[i], oa, sizeof(*obdos));
1724 obdos[i].o_id = si->lsm.lsm_object_id;
1728 for (i = 0; i < oa_bufs; i++) {
1729 int which = where[i];
1732 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1733 LASSERT(shift < oa_bufs);
1734 ioarr[shift] = pga[i];
1735 lov_stripe_offset(lsm, pga[i].off, which, &ioarr[shift].off);
1736 stripeinfo[which].subcount++;
1739 for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1740 int shift = si->index;
1745 if (lov->tgts[si->ost_idx].active == 0) {
1746 CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1747 GOTO(out_ioarr, rc = -EIO);
1750 LASSERT(shift < oa_bufs);
1752 rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp,
1753 &obdos[i], &si->lsm, si->bufct,
1754 &ioarr[shift], set, oti);
1756 GOTO(out_ioarr, rc);
1759 LASSERT(set->set_interpret == NULL);
1760 set->set_interpret = (set_interpreter_func)lov_brw_interpret;
1761 LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args));
1762 aa = (struct lov_brw_async_args *)&set->set_args;
1764 aa->aa_obdos = obdos;
1766 aa->aa_ioarr = ioarr;
1767 aa->aa_oa_bufs = oa_bufs;
1769 /* Don't free ioarr or obdos - that's done in lov_brw_interpret */
1770 GOTO(out_where, rc);
1773 OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1775 OBD_FREE(obdos, stripe_count * sizeof(*obdos));
1777 OBD_FREE(where, sizeof(*where) * oa_bufs);
1779 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1783 struct lov_async_page *lap_from_cookie(void *cookie)
1785 struct lov_async_page *lap = cookie;
1786 if (lap->lap_magic != LAP_MAGIC)
1787 return ERR_PTR(-EINVAL);
1791 static int lov_ap_make_ready(void *data, int cmd)
1793 struct lov_async_page *lap = lap_from_cookie(data);
1794 /* XXX should these assert? */
1798 return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
1800 static int lov_ap_refresh_count(void *data, int cmd)
1802 struct lov_async_page *lap = lap_from_cookie(data);
1806 return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
1809 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
1811 struct lov_async_page *lap = lap_from_cookie(data);
1812 /* XXX should these assert? */
1816 lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
1817 /* XXX woah, shouldn't we be altering more here? size? */
1818 oa->o_id = lap->lap_loi_id;
1820 static void lov_ap_completion(void *data, int cmd, int rc)
1822 struct lov_async_page *lap = lap_from_cookie(data);
1826 /* in a raid1 regime this would down a count of many ios
1827 * in flight, onl calling the caller_ops completion when all
1828 * the raid1 ios are complete */
1829 lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, rc);
1832 static struct obd_async_page_ops lov_async_page_ops = {
1833 .ap_make_ready = lov_ap_make_ready,
1834 .ap_refresh_count = lov_ap_refresh_count,
1835 .ap_fill_obdo = lov_ap_fill_obdo,
1836 .ap_completion = lov_ap_completion,
1839 int lov_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1840 struct lov_oinfo *loi, struct page *page,
1841 obd_off offset, struct obd_async_page_ops *ops,
1842 void *data, void **res)
1844 struct lov_obd *lov = &exp->exp_obd->u.lov;
1845 struct lov_async_page *lap;
1849 if (lsm_bad_magic(lsm))
1851 LASSERT(loi == NULL);
1853 OBD_ALLOC(lap, sizeof(*lap));
1857 lap->lap_magic = LAP_MAGIC;
1858 lap->lap_caller_ops = ops;
1859 lap->lap_caller_data = data;
1861 /* for now only raid 0 which passes through */
1862 lap->lap_stripe = lov_stripe_number(lsm, offset);
1863 lov_stripe_offset(lsm, offset, lap->lap_stripe, &lap->lap_sub_offset);
1864 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1866 /* so the callback doesn't need the lsm */
1867 lap->lap_loi_id = loi->loi_id;
1869 rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
1870 lsm, loi, page, lap->lap_sub_offset,
1871 &lov_async_page_ops, lap,
1872 &lap->lap_sub_cookie);
1874 OBD_FREE(lap, sizeof(*lap));
1877 CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
1878 lap->lap_sub_cookie, offset);
1883 static int lov_queue_async_io(struct obd_export *exp,
1884 struct lov_stripe_md *lsm,
1885 struct lov_oinfo *loi, void *cookie,
1886 int cmd, obd_off off, int count,
1887 obd_flag brw_flags, obd_flag async_flags)
1889 struct lov_obd *lov = &exp->exp_obd->u.lov;
1890 struct lov_async_page *lap;
1893 LASSERT(loi == NULL);
1895 if (lsm_bad_magic(lsm))
1898 lap = lap_from_cookie(cookie);
1900 RETURN(PTR_ERR(lap));
1902 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1903 rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
1904 loi, lap->lap_sub_cookie, cmd, off, count,
1905 brw_flags, async_flags);
1909 static int lov_set_async_flags(struct obd_export *exp,
1910 struct lov_stripe_md *lsm,
1911 struct lov_oinfo *loi, void *cookie,
1912 obd_flag async_flags)
1914 struct lov_obd *lov = &exp->exp_obd->u.lov;
1915 struct lov_async_page *lap;
1918 LASSERT(loi == NULL);
1920 if (lsm_bad_magic(lsm))
1923 lap = lap_from_cookie(cookie);
1925 RETURN(PTR_ERR(lap));
1927 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1928 rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
1929 lsm, loi, lap->lap_sub_cookie, async_flags);
1933 static int lov_queue_group_io(struct obd_export *exp,
1934 struct lov_stripe_md *lsm,
1935 struct lov_oinfo *loi,
1936 struct obd_io_group *oig, void *cookie,
1937 int cmd, obd_off off, int count,
1938 obd_flag brw_flags, obd_flag async_flags)
1940 struct lov_obd *lov = &exp->exp_obd->u.lov;
1941 struct lov_async_page *lap;
1944 LASSERT(loi == NULL);
1946 if (lsm_bad_magic(lsm))
1949 lap = lap_from_cookie(cookie);
1951 RETURN(PTR_ERR(lap));
1953 loi = &lsm->lsm_oinfo[lap->lap_stripe];
1954 rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
1955 oig, lap->lap_sub_cookie, cmd, off, count,
1956 brw_flags, async_flags);
1960 /* this isn't exactly optimal. we may have queued sync io in oscs on
1961 * all stripes, but we don't record that fact at queue time. so we
1962 * trigger sync io on all stripes. */
1963 static int lov_trigger_group_io(struct obd_export *exp,
1964 struct lov_stripe_md *lsm,
1965 struct lov_oinfo *loi,
1966 struct obd_io_group *oig)
1968 struct lov_obd *lov = &exp->exp_obd->u.lov;
1971 LASSERT(loi == NULL);
1973 if (lsm_bad_magic(lsm))
1976 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1978 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
1980 if (rc == 0 && err != 0)
1986 static int lov_teardown_async_page(struct obd_export *exp,
1987 struct lov_stripe_md *lsm,
1988 struct lov_oinfo *loi, void *cookie)
1990 struct lov_obd *lov = &exp->exp_obd->u.lov;
1991 struct lov_async_page *lap;
1994 LASSERT(loi == NULL);
1996 if (lsm_bad_magic(lsm))
1999 lap = lap_from_cookie(cookie);
2001 RETURN(PTR_ERR(lap));
2003 loi = &lsm->lsm_oinfo[lap->lap_stripe];
2004 rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
2005 lsm, loi, lap->lap_sub_cookie);
2007 CERROR("unable to teardown sub cookie %p: %d\n",
2008 lap->lap_sub_cookie, rc);
2011 OBD_FREE(lap, sizeof(*lap));
2015 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2016 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2017 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2018 void *data,__u32 lvb_len, void *lvb_swabber,
2019 struct lustre_handle *lockh)
2021 struct lov_lock_handles *lov_lockh = NULL;
2022 struct lustre_handle *lov_lockhp;
2023 struct lov_obd *lov;
2024 struct lov_oinfo *loi;
2025 char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)];
2026 struct lov_stripe_md *submd = (void *)submd_buf;
2028 int i, save_flags = *flags;
2031 if (lsm_bad_magic(lsm))
2034 /* we should never be asked to replay a lock this way. */
2035 LASSERT((*flags & LDLM_FL_REPLAY) == 0);
2037 if (!exp || !exp->exp_obd)
2040 if (lsm->lsm_stripe_count > 1) {
2041 lov_lockh = lov_llh_new(lsm);
2042 if (lov_lockh == NULL)
2045 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2046 lov_lockhp = lov_lockh->llh_handles;
2051 lov = &exp->exp_obd->u.lov;
2052 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2053 i++, loi++, lov_lockhp++) {
2054 ldlm_policy_data_t sub_ext;
2056 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2057 policy->l_extent.end,
2058 &sub_ext.l_extent.start,
2059 &sub_ext.l_extent.end))
2062 sub_ext.l_extent.gid = policy->l_extent.gid;
2064 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2065 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2069 /* XXX LOV STACKING: submd should be from the subobj */
2070 submd->lsm_object_id = loi->loi_id;
2071 submd->lsm_object_gr = lsm->lsm_object_gr;
2072 submd->lsm_stripe_count = 0;
2073 submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid;
2074 submd->lsm_oinfo->loi_rss = loi->loi_rss;
2075 submd->lsm_oinfo->loi_kms = loi->loi_kms;
2076 loi->loi_mtime = submd->lsm_oinfo->loi_mtime;
2077 /* XXX submd is not fully initialized here */
2078 *flags = save_flags;
2079 rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd,
2080 type, &sub_ext, mode, flags, bl_cb, cp_cb,
2081 gl_cb, data, lvb_len, lvb_swabber, lov_lockhp);
2083 /* XXX FIXME: This unpleasantness doesn't belong here at *all*.
2084 * It belongs in the OSC, except that the OSC doesn't have
2085 * access to the real LOI -- it gets a copy, that we created
2086 * above, and that copy can be arbitrarily out of date.
2088 * The LOV API is due for a serious rewriting anyways, and this
2089 * can be addressed then. */
2090 if (rc == ELDLM_OK) {
2091 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2092 __u64 tmp = submd->lsm_oinfo->loi_rss;
2094 LASSERT(lock != NULL);
2096 /* Extend KMS up to the end of this lock and no further
2097 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2098 if (tmp > lock->l_policy_data.l_extent.end)
2099 tmp = lock->l_policy_data.l_extent.end + 1;
2100 if (tmp >= loi->loi_kms) {
2101 CDEBUG(D_INODE, "lock acquired, setting rss="
2102 LPU64", kms="LPU64"\n", loi->loi_rss,
2105 loi->loi_kms_valid = 1;
2107 CDEBUG(D_INODE, "lock acquired, setting rss="
2108 LPU64"; leaving kms="LPU64", end="LPU64
2109 "\n", loi->loi_rss, loi->loi_kms,
2110 lock->l_policy_data.l_extent.end);
2112 ldlm_lock_allow_match(lock);
2113 LDLM_LOCK_PUT(lock);
2114 } else if (rc == ELDLM_LOCK_ABORTED &&
2115 save_flags & LDLM_FL_HAS_INTENT) {
2116 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2117 loi->loi_rss = submd->lsm_oinfo->loi_rss;
2118 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2119 " kms="LPU64"\n", loi->loi_rss, loi->loi_kms);
2121 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2122 if (lov->tgts[loi->loi_ost_idx].active) {
2123 CERROR("error: enqueue objid "LPX64" subobj "
2124 LPX64" on OST idx %d: rc = %d\n",
2125 lsm->lsm_object_id, loi->loi_id,
2126 loi->loi_ost_idx, rc);
2127 GOTO(out_locks, rc);
2131 if (lsm->lsm_stripe_count > 1)
2132 lov_llh_put(lov_lockh);
2136 while (loi--, lov_lockhp--, i-- > 0) {
2137 struct lov_stripe_md submd;
2140 if (lov_lockhp->cookie == 0)
2143 /* XXX LOV STACKING: submd should be from the subobj */
2144 submd.lsm_object_id = loi->loi_id;
2145 submd.lsm_object_gr = lsm->lsm_object_gr;
2146 submd.lsm_stripe_count = 0;
2147 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2149 if (err && lov->tgts[loi->loi_ost_idx].active) {
2150 CERROR("error: cancelling objid "LPX64" on OST "
2151 "idx %d after enqueue error: rc = %d\n",
2152 loi->loi_id, loi->loi_ost_idx, err);
2156 if (lsm->lsm_stripe_count > 1) {
2157 lov_llh_destroy(lov_lockh);
2158 lov_llh_put(lov_lockh);
2163 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2164 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2165 int *flags, void *data, struct lustre_handle *lockh)
2167 struct lov_lock_handles *lov_lockh = NULL;
2168 struct lustre_handle *lov_lockhp;
2169 struct lov_obd *lov;
2170 struct lov_oinfo *loi;
2171 struct lov_stripe_md submd;
2172 ldlm_error_t rc = 0;
2176 if (lsm_bad_magic(lsm))
2179 if (!exp || !exp->exp_obd)
2182 if (lsm->lsm_stripe_count > 1) {
2183 lov_lockh = lov_llh_new(lsm);
2184 if (lov_lockh == NULL)
2187 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2188 lov_lockhp = lov_lockh->llh_handles;
2193 lov = &exp->exp_obd->u.lov;
2194 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2195 i++, loi++, lov_lockhp++) {
2196 ldlm_policy_data_t sub_ext;
2199 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2200 policy->l_extent.end,
2201 &sub_ext.l_extent.start,
2202 &sub_ext.l_extent.end))
2205 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2206 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2211 /* XXX LOV STACKING: submd should be from the subobj */
2212 submd.lsm_object_id = loi->loi_id;
2213 submd.lsm_object_gr = lsm->lsm_object_gr;
2214 submd.lsm_stripe_count = 0;
2216 /* XXX submd is not fully initialized here */
2217 rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2218 type, &sub_ext, mode, &lov_flags, data,
2224 if (lsm->lsm_stripe_count > 1) {
2225 if (*flags & LDLM_FL_TEST_LOCK)
2226 lov_llh_destroy(lov_lockh);
2227 lov_llh_put(lov_lockh);
2232 while (loi--, lov_lockhp--, i-- > 0) {
2233 struct lov_stripe_md submd;
2236 if (lov_lockhp->cookie == 0)
2239 /* XXX LOV STACKING: submd should be from the subobj */
2240 submd.lsm_object_id = loi->loi_id;
2241 submd.lsm_object_gr = lsm->lsm_object_gr;
2242 submd.lsm_stripe_count = 0;
2243 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2245 if (err && lov->tgts[loi->loi_ost_idx].active) {
2246 CERROR("error: cancelling objid "LPX64" on OST "
2247 "idx %d after match failure: rc = %d\n",
2248 loi->loi_id, loi->loi_ost_idx, err);
2252 if (lsm->lsm_stripe_count > 1) {
2253 lov_llh_destroy(lov_lockh);
2254 lov_llh_put(lov_lockh);
2259 static int lov_change_cbdata(struct obd_export *exp,
2260 struct lov_stripe_md *lsm, ldlm_iterator_t it,
2263 struct lov_obd *lov;
2264 struct lov_oinfo *loi;
2268 if (lsm_bad_magic(lsm))
2271 if (!exp || !exp->exp_obd)
2274 LASSERT(lsm->lsm_object_gr > 0);
2276 lov = &exp->exp_obd->u.lov;
2277 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2278 struct lov_stripe_md submd;
2279 if (lov->tgts[loi->loi_ost_idx].active == 0)
2280 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2282 submd.lsm_object_id = loi->loi_id;
2283 submd.lsm_object_gr = lsm->lsm_object_gr;
2284 submd.lsm_stripe_count = 0;
2285 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
2291 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
2292 __u32 mode, struct lustre_handle *lockh)
2294 struct lov_lock_handles *lov_lockh = NULL;
2295 struct lustre_handle *lov_lockhp;
2296 struct lov_obd *lov;
2297 struct lov_oinfo *loi;
2301 if (lsm_bad_magic(lsm))
2304 if (!exp || !exp->exp_obd)
2307 LASSERT(lsm->lsm_object_gr > 0);
2310 if (lsm->lsm_stripe_count > 1) {
2311 lov_lockh = lov_handle2llh(lockh);
2313 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2317 lov_lockhp = lov_lockh->llh_handles;
2322 lov = &exp->exp_obd->u.lov;
2323 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2324 i++, loi++, lov_lockhp++) {
2325 struct lov_stripe_md submd;
2328 if (lov_lockhp->cookie == 0) {
2329 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2330 loi->loi_ost_idx, loi->loi_id);
2334 /* XXX LOV STACKING: submd should be from the subobj */
2335 submd.lsm_object_id = loi->loi_id;
2336 submd.lsm_object_gr = lsm->lsm_object_gr;
2337 submd.lsm_stripe_count = 0;
2338 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2341 if (lov->tgts[loi->loi_ost_idx].active) {
2342 CERROR("error: cancel objid "LPX64" subobj "
2343 LPX64" on OST idx %d: rc = %d\n",
2345 loi->loi_id, loi->loi_ost_idx, err);
2352 if (lsm->lsm_stripe_count > 1)
2353 lov_llh_destroy(lov_lockh);
2354 if (lov_lockh != NULL)
2355 lov_llh_put(lov_lockh);
2359 static int lov_cancel_unused(struct obd_export *exp,
2360 struct lov_stripe_md *lsm, int flags, void *opaque)
2362 struct lov_obd *lov;
2363 struct lov_oinfo *loi;
2367 if (lsm_bad_magic(lsm))
2370 if (!exp || !exp->exp_obd)
2373 LASSERT(lsm->lsm_object_gr > 0);
2375 lov = &exp->exp_obd->u.lov;
2376 for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2377 struct lov_stripe_md submd;
2380 if (lov->tgts[loi->loi_ost_idx].active == 0)
2381 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2383 submd.lsm_object_id = loi->loi_id;
2384 submd.lsm_object_gr = lsm->lsm_object_gr;
2385 submd.lsm_stripe_count = 0;
2386 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
2387 &submd, flags, opaque);
2388 if (err && lov->tgts[loi->loi_ost_idx].active) {
2389 CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2390 " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2391 loi->loi_id, loi->loi_ost_idx, err);
2399 #define LOV_U64_MAX ((__u64)~0ULL)
2400 #define LOV_SUM_MAX(tot, add) \
2402 if ((tot) + (add) < (tot)) \
2403 (tot) = LOV_U64_MAX; \
2408 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2409 unsigned long max_age)
2411 struct lov_obd *lov = &obd->u.lov;
2412 struct obd_statfs lov_sfs;
2419 /* We only get block data from the OBD */
2420 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2423 if (!lov->tgts[i].active) {
2424 CDEBUG(D_HA, "lov idx %d inactive\n", i);
2428 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
2431 if (lov->tgts[i].active && !rc)
2437 memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
2440 osfs->os_bfree += lov_sfs.os_bfree;
2441 osfs->os_bavail += lov_sfs.os_bavail;
2442 osfs->os_blocks += lov_sfs.os_blocks;
2443 /* XXX not sure about this one - depends on policy.
2444 * - could be minimum if we always stripe on all OBDs
2445 * (but that would be wrong for any other policy,
2446 * if one of the OBDs has no more objects left)
2447 * - could be sum if we stripe whole objects
2448 * - could be average, just to give a nice number
2450 * To give a "reasonable" (if not wholly accurate)
2451 * number, we divide the total number of free objects
2452 * by expected stripe count (watch out for overflow).
2454 LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
2455 LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
2460 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
2461 lov->desc.ld_default_stripe_count :
2462 lov->desc.ld_active_tgt_count;
2464 if (osfs->os_files != LOV_U64_MAX)
2465 do_div(osfs->os_files, expected_stripes);
2466 if (osfs->os_ffree != LOV_U64_MAX)
2467 do_div(osfs->os_ffree, expected_stripes);
2474 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2475 void *karg, void *uarg)
2477 struct obd_device *obddev = class_exp2obd(exp);
2478 struct lov_obd *lov = &obddev->u.lov;
2479 int i, count = lov->desc.ld_tgt_count;
2480 struct obd_uuid *uuidp;
2486 case OBD_IOC_LOV_GET_CONFIG: {
2487 struct obd_ioctl_data *data = karg;
2488 struct lov_tgt_desc *tgtdesc;
2489 struct lov_desc *desc;
2494 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2497 data = (struct obd_ioctl_data *)buf;
2499 if (sizeof(*desc) > data->ioc_inllen1) {
2504 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2509 desc = (struct lov_desc *)data->ioc_inlbuf1;
2510 memcpy(desc, &(lov->desc), sizeof(*desc));
2512 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2513 tgtdesc = lov->tgts;
2514 for (i = 0; i < count; i++, uuidp++, tgtdesc++)
2515 obd_str2uuid(uuidp, tgtdesc->uuid.uuid);
2517 rc = copy_to_user((void *)uarg, buf, len);
2520 obd_ioctl_freedata(buf, len);
2523 case LL_IOC_LOV_SETSTRIPE:
2524 rc = lov_setstripe(exp, karg, uarg);
2526 case LL_IOC_LOV_GETSTRIPE:
2527 rc = lov_getstripe(exp, karg, uarg);
2529 case LL_IOC_LOV_SETEA:
2530 rc = lov_setea(exp, karg, uarg);
2537 for (i = 0; i < count; i++) {
2540 err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
2543 if (lov->tgts[i].active) {
2544 CERROR("error: iocontrol OSC %s on OST"
2545 "idx %d: err = %d\n",
2546 lov->tgts[i].uuid.uuid, i, err);
2561 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2562 void *key, __u32 *vallen, void *val)
2564 struct obd_device *obddev = class_exp2obd(exp);
2565 struct lov_obd *lov = &obddev->u.lov;
2569 if (!vallen || !val)
2572 if (keylen > strlen("lock_to_stripe") &&
2573 strcmp(key, "lock_to_stripe") == 0) {
2576 struct ldlm_lock *lock;
2577 struct lov_stripe_md *lsm;
2579 struct lov_oinfo *loi;
2580 __u32 *stripe = val;
2582 if (*vallen < sizeof(*stripe))
2584 *vallen = sizeof(*stripe);
2586 /* XXX This is another one of those bits that will need to
2587 * change if we ever actually support nested LOVs. It uses
2588 * the lock's export to find out which stripe it is. */
2589 for (i = 0, loi = data->lsm->lsm_oinfo;
2590 i < data->lsm->lsm_stripe_count;
2592 if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
2593 data->lock->l_conn_export) {
2599 } else if (keylen >= strlen("size_to_stripe") &&
2600 strcmp(key, "size_to_stripe") == 0) {
2604 struct lov_stripe_md *lsm;
2607 if (*vallen < sizeof(*data))
2610 data->size = lov_size_to_stripe(data->lsm, data->size,
2611 data->stripe_number);
2613 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2615 int rc, size = sizeof(obd_id);
2616 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2617 if (!lov->tgts[i].active)
2619 rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key,
2625 } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
2626 struct lov_desc *desc_ret = val;
2627 *desc_ret = lov->desc;
2635 static int lov_set_info(struct obd_export *exp, obd_count keylen,
2636 void *key, obd_count vallen, void *val)
2638 struct obd_device *obddev = class_exp2obd(exp);
2639 struct lov_obd *lov = &obddev->u.lov;
2643 #define KEY_IS(str) \
2644 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
2646 if (KEY_IS("next_id")) {
2647 if (vallen != lov->desc.ld_tgt_count)
2649 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2652 /* initialize all OSCs, even inactive ones */
2654 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key,
2655 sizeof(obd_id), ((obd_id*)val) + i);
2662 if (KEY_IS("growth_count")) {
2663 if (vallen != sizeof(int))
2665 } else if (KEY_IS("mds_conn")) {
2666 if (vallen != sizeof(__u32))
2668 } else if (KEY_IS("unlinked")) {
2675 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2678 if (!lov->tgts[i].active)
2681 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen,
2691 /* Merge rss if kms == 0
2693 * Even when merging RSS, we will take the KMS value if it's larger.
2694 * This prevents getattr from stomping on dirty cached pages which
2695 * extend the file size. */
2696 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms)
2698 struct lov_oinfo *loi;
2702 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2704 obd_size lov_size, tmpsize;
2706 tmpsize = loi->loi_kms;
2707 if (kms == 0 && loi->loi_rss > tmpsize)
2708 tmpsize = loi->loi_rss;
2710 lov_size = lov_stripe_size(lsm, tmpsize, i);
2711 if (lov_size > size)
2716 EXPORT_SYMBOL(lov_merge_size);
2718 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time)
2720 struct lov_oinfo *loi;
2723 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2725 if (loi->loi_mtime > current_time)
2726 current_time = loi->loi_mtime;
2728 return current_time;
2730 EXPORT_SYMBOL(lov_merge_mtime);
2733 struct lov_multi_wait {
2734 struct ldlm_lock *lock;
2740 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
2741 struct lustre_handle *lockh)
2743 struct lov_lock_handles *lov_lockh = NULL;
2744 struct lustre_handle *lov_lockhp;
2745 struct lov_obd *lov;
2746 struct lov_oinfo *loi;
2747 struct lov_multi_wait *queues;
2751 if (lsm_bad_magic(lsm))
2754 if (!exp || !exp->exp_obd)
2757 LASSERT(lockh != NULL);
2758 if (lsm->lsm_stripe_count > 1) {
2759 lov_lockh = lov_handle2llh(lockh);
2760 if (lov_lockh == NULL) {
2761 CERROR("LOV: invalid lov lock handle %p\n", lockh);
2765 lov_lockhp = lov_lockh->llh_handles;
2770 OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
2772 GOTO(out, rc = -ENOMEM);
2774 lov = &exp->exp_obd->u.lov;
2775 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2776 i++, loi++, lov_lockhp++) {
2777 struct ldlm_lock *lock;
2778 struct obd_device *obd;
2779 unsigned long irqflags;
2781 lock = ldlm_handle2lock(lov_lockhp);
2783 CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2784 loi->loi_ost_idx, loi->loi_id);
2785 queues[i].completed = 1;
2789 queues[i].lock = lock;
2790 init_waitqueue_entry(&(queues[i].wait), current);
2791 add_wait_queue(lock->l_waitq, &(queues[i].wait));
2793 obd = class_exp2obd(lock->l_conn_export);
2795 imp = obd->u.cli.cl_import;
2797 spin_lock_irqsave(&imp->imp_lock, irqflags);
2798 queues[i].generation = imp->imp_generation;
2799 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
2803 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
2804 interrupted_completion_wait, &lwd);
2805 rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
2807 for (i = 0; i < lsm->lsm_stripe_count; i++)
2808 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
2810 if (rc == -EINTR || rc == -ETIMEDOUT) {
2816 if (lov_lockh != NULL)
2817 lov_llh_put(lov_lockh);
2822 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
2825 struct lov_oinfo *loi;
2831 stripe = lov_stripe_number(lsm, size - 1);
2832 kms = lov_size_to_stripe(lsm, size, stripe);
2833 loi = &(lsm->lsm_oinfo[stripe]);
2835 CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n",
2836 stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms);
2837 if (kms > loi->loi_kms)
2841 EXPORT_SYMBOL(lov_increase_kms);
2843 struct obd_ops lov_obd_ops = {
2844 o_owner: THIS_MODULE,
2845 o_attach: lov_attach,
2846 o_detach: lov_detach,
2848 o_cleanup: lov_cleanup,
2849 o_connect: lov_connect,
2850 o_disconnect: lov_disconnect,
2851 o_statfs: lov_statfs,
2852 o_packmd: lov_packmd,
2853 o_unpackmd: lov_unpackmd,
2854 o_create: lov_create,
2855 o_destroy: lov_destroy,
2856 o_getattr: lov_getattr,
2857 o_getattr_async: lov_getattr_async,
2858 o_setattr: lov_setattr,
2860 o_brw_async: lov_brw_async,
2861 .o_prep_async_page = lov_prep_async_page,
2862 .o_queue_async_io = lov_queue_async_io,
2863 .o_set_async_flags = lov_set_async_flags,
2864 .o_queue_group_io = lov_queue_group_io,
2865 .o_trigger_group_io = lov_trigger_group_io,
2866 .o_teardown_async_page lov_teardown_async_page,
2869 o_enqueue: lov_enqueue,
2871 o_change_cbdata: lov_change_cbdata,
2872 o_cancel: lov_cancel,
2873 o_cancel_unused: lov_cancel_unused,
2874 o_iocontrol: lov_iocontrol,
2875 o_get_info: lov_get_info,
2876 o_set_info: lov_set_info,
2877 o_llog_init: lov_llog_init,
2878 o_llog_finish: lov_llog_finish,
2879 o_notify: lov_notify,
2882 int __init lov_init(void)
2884 struct lprocfs_static_vars lvars;
2887 lprocfs_init_vars(lov, &lvars);
2888 rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars,
2889 OBD_LOV_DEVICENAME);
2894 static void /*__exit*/ lov_exit(void)
2896 class_unregister_type(OBD_LOV_DEVICENAME);
2899 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2900 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
2901 MODULE_LICENSE("GPL");
2903 module_init(lov_init);
2904 module_exit(lov_exit);