1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LMV
42 #include <linux/slab.h>
43 #include <linux/module.h>
44 #include <linux/init.h>
45 #include <linux/slab.h>
46 #include <linux/pagemap.h>
48 #include <asm/div64.h>
49 #include <linux/seq_file.h>
50 #include <linux/namei.h>
52 #include <liblustre.h>
55 #include <lustre_log.h>
56 #include <obd_support.h>
57 #include <lustre_lib.h>
58 #include <lustre_net.h>
59 #include <obd_class.h>
60 #include <lprocfs_status.h>
61 #include <lustre_lite.h>
62 #include <lustre_fid.h>
63 #include "lmv_internal.h"
66 cfs_mem_cache_t *lmv_object_cache;
67 cfs_atomic_t lmv_object_count = CFS_ATOMIC_INIT(0);
69 static void lmv_activate_target(struct lmv_obd *lmv,
70 struct lmv_tgt_desc *tgt,
73 if (tgt->ltd_active == activate)
76 tgt->ltd_active = activate;
77 lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
83 * -EINVAL : UUID can't be found in the LMV's target list
84 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
85 * -EBADF : The UUID is found, but the OBD of the wrong type (!)
87 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
90 struct lmv_tgt_desc *tgt;
91 struct obd_device *obd;
96 CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
97 lmv, uuid->uuid, activate);
99 cfs_spin_lock(&lmv->lmv_lock);
100 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
101 if (tgt->ltd_exp == NULL)
104 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n",
105 i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
107 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
111 if (i == lmv->desc.ld_tgt_count)
112 GOTO(out_lmv_lock, rc = -EINVAL);
114 obd = class_exp2obd(tgt->ltd_exp);
116 GOTO(out_lmv_lock, rc = -ENOTCONN);
118 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
119 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
120 obd->obd_type->typ_name, i);
121 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
123 if (tgt->ltd_active == activate) {
124 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
125 activate ? "" : "in");
126 GOTO(out_lmv_lock, rc);
129 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
130 activate ? "" : "in");
131 lmv_activate_target(lmv, tgt, activate);
135 cfs_spin_unlock(&lmv->lmv_lock);
139 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
140 struct obd_connect_data *data)
142 struct lmv_tgt_desc *tgt;
146 LASSERT(data != NULL);
148 cfs_spin_lock(&lmv->lmv_lock);
149 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
150 if (tgt->ltd_exp == NULL)
153 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
154 lmv->datas[tgt->ltd_idx] = *data;
158 cfs_spin_unlock(&lmv->lmv_lock);
162 struct obd_uuid *lmv_get_uuid(struct obd_export *exp) {
163 struct obd_device *obd = exp->exp_obd;
164 struct lmv_obd *lmv = &obd->u.lmv;
165 return obd_get_uuid(lmv->tgts[0].ltd_exp);
168 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
169 enum obd_notify_event ev, void *data)
171 struct obd_connect_data *conn_data;
172 struct lmv_obd *lmv = &obd->u.lmv;
173 struct obd_uuid *uuid;
177 if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
178 CERROR("unexpected notification of %s %s!\n",
179 watched->obd_type->typ_name,
184 uuid = &watched->u.cli.cl_target_uuid;
185 if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
187 * Set MDC as active before notifying the observer, so the
188 * observer can use the MDC normally.
190 rc = lmv_set_mdc_active(lmv, uuid,
191 ev == OBD_NOTIFY_ACTIVE);
193 CERROR("%sactivation of %s failed: %d\n",
194 ev == OBD_NOTIFY_ACTIVE ? "" : "de",
198 } else if (ev == OBD_NOTIFY_OCD) {
199 conn_data = &watched->u.cli.cl_import->imp_connect_data;
202 * Set connect data to desired target, update exp_connect_flags.
204 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
206 CERROR("can't set connect data to target %s, rc %d\n",
212 * XXX: Make sure that ocd_connect_flags from all targets are
213 * the same. Otherwise one of MDTs runs wrong version or
214 * something like this. --umka
216 obd->obd_self_export->exp_connect_flags =
217 conn_data->ocd_connect_flags;
220 else if (ev == OBD_NOTIFY_DISCON) {
222 * For disconnect event, flush fld cache for failout MDS case.
224 fld_client_flush(&lmv->lmv_fld);
228 * Pass the notification up the chain.
230 if (obd->obd_observer)
231 rc = obd_notify(obd->obd_observer, watched, ev, data);
237 * This is fake connect function. Its purpose is to initialize lmv and say
238 * caller that everything is okay. Real connection will be performed later.
240 static int lmv_connect(const struct lu_env *env,
241 struct obd_export **exp, struct obd_device *obd,
242 struct obd_uuid *cluuid, struct obd_connect_data *data,
246 struct proc_dir_entry *lmv_proc_dir;
248 struct lmv_obd *lmv = &obd->u.lmv;
249 struct lustre_handle conn = { 0 };
254 * We don't want to actually do the underlying connections more than
255 * once, so keep track.
258 if (lmv->refcount > 1) {
263 rc = class_connect(&conn, obd, cluuid);
265 CERROR("class_connection() returned %d\n", rc);
269 *exp = class_conn2export(&conn);
270 class_export_get(*exp);
274 lmv->cluuid = *cluuid;
277 lmv->conn_data = *data;
280 lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
282 if (IS_ERR(lmv_proc_dir)) {
283 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
284 obd->obd_type->typ_name, obd->obd_name);
290 * All real clients should perform actual connection right away, because
291 * it is possible, that LMV will not have opportunity to connect targets
292 * and MDC stuff will be called directly, for instance while reading
293 * ../mdc/../kbytesfree procfs file, etc.
295 if (data->ocd_connect_flags & OBD_CONNECT_REAL)
296 rc = lmv_check_connect(obd);
301 lprocfs_remove(&lmv_proc_dir);
308 static void lmv_set_timeouts(struct obd_device *obd)
310 struct lmv_tgt_desc *tgts;
315 if (lmv->server_timeout == 0)
318 if (lmv->connected == 0)
321 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
322 if (tgts->ltd_exp == NULL)
325 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
326 KEY_INTERMDS, 0, NULL, NULL);
330 static int lmv_init_ea_size(struct obd_export *exp, int easize,
331 int def_easize, int cookiesize)
333 struct obd_device *obd = exp->exp_obd;
334 struct lmv_obd *lmv = &obd->u.lmv;
340 if (lmv->max_easize < easize) {
341 lmv->max_easize = easize;
344 if (lmv->max_def_easize < def_easize) {
345 lmv->max_def_easize = def_easize;
348 if (lmv->max_cookiesize < cookiesize) {
349 lmv->max_cookiesize = cookiesize;
355 if (lmv->connected == 0)
358 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
359 if (lmv->tgts[i].ltd_exp == NULL) {
360 CWARN("%s: NULL export for %d\n", obd->obd_name, i);
364 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
367 CERROR("obd_init_ea_size() failed on MDT target %d, "
368 "error %d.\n", i, rc);
375 #define MAX_STRING_SIZE 128
377 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
380 struct proc_dir_entry *lmv_proc_dir;
382 struct lmv_obd *lmv = &obd->u.lmv;
383 struct obd_uuid *cluuid = &lmv->cluuid;
384 struct obd_connect_data *mdc_data = NULL;
385 struct obd_uuid lmv_mdc_uuid = { "LMV_MDC_UUID" };
386 struct obd_device *mdc_obd;
387 struct obd_export *mdc_exp;
388 struct lu_fld_target target;
392 mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
395 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
399 CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
400 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
401 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
404 if (!mdc_obd->obd_set_up) {
405 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
409 rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
410 &lmv->conn_data, NULL);
412 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
417 * Init fid sequence client for this mdc and add new fld target.
419 rc = obd_fid_init(mdc_exp);
423 target.ft_srv = NULL;
424 target.ft_exp = mdc_exp;
425 target.ft_idx = tgt->ltd_idx;
427 fld_client_add_target(&lmv->lmv_fld, &target);
429 mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
431 rc = obd_register_observer(mdc_obd, obd);
433 obd_disconnect(mdc_exp);
434 CERROR("target %s register_observer error %d\n",
435 tgt->ltd_uuid.uuid, rc);
439 if (obd->obd_observer) {
441 * Tell the observer about the new target.
443 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
444 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
446 obd_disconnect(mdc_exp);
452 tgt->ltd_exp = mdc_exp;
453 lmv->desc.ld_active_tgt_count++;
456 * Copy connect data, it may be used later.
458 lmv->datas[tgt->ltd_idx] = *mdc_data;
460 md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
461 lmv->max_def_easize, lmv->max_cookiesize);
463 CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
464 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
465 cfs_atomic_read(&obd->obd_refcount));
468 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
470 struct proc_dir_entry *mdc_symlink;
472 LASSERT(mdc_obd->obd_type != NULL);
473 LASSERT(mdc_obd->obd_type->typ_name != NULL);
474 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
477 mdc_obd->obd_type->typ_name,
479 if (mdc_symlink == NULL) {
480 CERROR("Could not register LMV target "
481 "/proc/fs/lustre/%s/%s/target_obds/%s.",
482 obd->obd_type->typ_name, obd->obd_name,
484 lprocfs_remove(&lmv_proc_dir);
492 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
494 struct lmv_obd *lmv = &obd->u.lmv;
495 struct lmv_tgt_desc *tgt;
499 CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
503 if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
504 lmv_init_unlock(lmv);
505 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
506 "That many MDCs already configured.\n",
507 tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
510 if (lmv->desc.ld_tgt_count == 0) {
511 struct obd_device *mdc_obd;
513 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
516 lmv_init_unlock(lmv);
517 CERROR("Target %s not attached\n", tgt_uuid->uuid);
521 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, NULL);
523 lmv_init_unlock(lmv);
524 CERROR("lmv failed to setup llogging subsystems\n");
527 cfs_spin_lock(&lmv->lmv_lock);
528 tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
529 tgt->ltd_uuid = *tgt_uuid;
530 cfs_spin_unlock(&lmv->lmv_lock);
532 if (lmv->connected) {
533 rc = lmv_connect_mdc(obd, tgt);
535 cfs_spin_lock(&lmv->lmv_lock);
536 lmv->desc.ld_tgt_count--;
537 memset(tgt, 0, sizeof(*tgt));
538 cfs_spin_unlock(&lmv->lmv_lock);
540 int easize = sizeof(struct lmv_stripe_md) +
541 lmv->desc.ld_tgt_count *
542 sizeof(struct lu_fid);
543 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
547 lmv_init_unlock(lmv);
551 int lmv_check_connect(struct obd_device *obd)
553 struct lmv_obd *lmv = &obd->u.lmv;
554 struct lmv_tgt_desc *tgt;
564 if (lmv->connected) {
565 lmv_init_unlock(lmv);
569 if (lmv->desc.ld_tgt_count == 0) {
570 CERROR("%s: no targets configured.\n", obd->obd_name);
574 CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
575 lmv->cluuid.uuid, obd->obd_name);
577 LASSERT(lmv->tgts != NULL);
579 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
580 rc = lmv_connect_mdc(obd, tgt);
585 lmv_set_timeouts(obd);
586 class_export_put(lmv->exp);
588 easize = lmv_get_easize(lmv);
589 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
590 lmv_init_unlock(lmv);
599 --lmv->desc.ld_active_tgt_count;
600 rc2 = obd_disconnect(tgt->ltd_exp);
602 CERROR("LMV target %s disconnect on "
603 "MDC idx %d: error %d\n",
604 tgt->ltd_uuid.uuid, i, rc2);
608 class_disconnect(lmv->exp);
609 lmv_init_unlock(lmv);
613 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
616 struct proc_dir_entry *lmv_proc_dir;
618 struct lmv_obd *lmv = &obd->u.lmv;
619 struct obd_device *mdc_obd;
623 LASSERT(tgt != NULL);
624 LASSERT(obd != NULL);
626 mdc_obd = class_exp2obd(tgt->ltd_exp);
629 mdc_obd->obd_force = obd->obd_force;
630 mdc_obd->obd_fail = obd->obd_fail;
631 mdc_obd->obd_no_recov = obd->obd_no_recov;
635 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
637 struct proc_dir_entry *mdc_symlink;
639 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
641 lprocfs_remove(&mdc_symlink);
643 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
644 obd->obd_type->typ_name, obd->obd_name,
649 rc = obd_fid_fini(tgt->ltd_exp);
651 CERROR("Can't finanize fids factory\n");
653 CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
654 tgt->ltd_exp->exp_obd->obd_name,
655 tgt->ltd_exp->exp_obd->obd_uuid.uuid);
657 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
658 rc = obd_disconnect(tgt->ltd_exp);
660 if (tgt->ltd_active) {
661 CERROR("Target %s disconnect error %d\n",
662 tgt->ltd_uuid.uuid, rc);
666 lmv_activate_target(lmv, tgt, 0);
671 static int lmv_disconnect(struct obd_export *exp)
673 struct obd_device *obd = class_exp2obd(exp);
675 struct proc_dir_entry *lmv_proc_dir;
677 struct lmv_obd *lmv = &obd->u.lmv;
686 * Only disconnect the underlying layers on the final disconnect.
689 if (lmv->refcount != 0)
692 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
693 if (lmv->tgts[i].ltd_exp == NULL)
695 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
699 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
701 lprocfs_remove(&lmv_proc_dir);
703 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
704 obd->obd_type->typ_name, obd->obd_name);
710 * This is the case when no real connection is established by
711 * lmv_check_connect().
714 class_export_put(exp);
715 rc = class_disconnect(exp);
716 if (lmv->refcount == 0)
721 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
722 int len, void *karg, void *uarg)
724 struct obd_device *obddev = class_exp2obd(exp);
725 struct lmv_obd *lmv = &obddev->u.lmv;
729 int count = lmv->desc.ld_tgt_count;
736 case IOC_OBD_STATFS: {
737 struct obd_ioctl_data *data = karg;
738 struct obd_device *mdc_obd;
739 struct obd_statfs stat_buf = {0};
742 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
743 if ((index >= count))
746 if (!lmv->tgts[index].ltd_active)
749 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
754 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
755 min((int) data->ioc_plen2,
756 (int) sizeof(struct obd_uuid))))
759 rc = obd_statfs(mdc_obd, &stat_buf,
760 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
764 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
765 min((int) data->ioc_plen1,
766 (int) sizeof(stat_buf))))
770 case OBD_IOC_QUOTACTL: {
771 struct if_quotactl *qctl = karg;
772 struct lmv_tgt_desc *tgt = NULL;
773 struct obd_quotactl *oqctl;
775 if (qctl->qc_valid == QC_MDTIDX) {
776 if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
779 tgt = &lmv->tgts[qctl->qc_idx];
782 } else if (qctl->qc_valid == QC_UUID) {
783 for (i = 0; i < count; i++) {
785 if (!obd_uuid_equals(&tgt->ltd_uuid,
789 if (tgt->ltd_exp == NULL)
801 LASSERT(tgt && tgt->ltd_exp);
802 OBD_ALLOC_PTR(oqctl);
806 QCTL_COPY(oqctl, qctl);
807 rc = obd_quotactl(tgt->ltd_exp, oqctl);
809 QCTL_COPY(qctl, oqctl);
810 qctl->qc_valid = QC_MDTIDX;
811 qctl->obd_uuid = tgt->ltd_uuid;
816 case OBD_IOC_CHANGELOG_SEND:
817 case OBD_IOC_CHANGELOG_CLEAR: {
818 struct ioc_changelog *icc = karg;
820 if (icc->icc_mdtindex >= count)
823 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
824 sizeof(*icc), icc, NULL);
827 case LL_IOC_GET_CONNECT_FLAGS: {
828 rc = obd_iocontrol(cmd, lmv->tgts[0].ltd_exp, len, karg, uarg);
833 for (i = 0; i < count; i++) {
835 struct obd_device *mdc_obd;
837 if (lmv->tgts[i].ltd_exp == NULL)
839 /* ll_umount_begin() sets force flag but for lmv, not
840 * mdc. Let's pass it through */
841 mdc_obd = class_exp2obd(lmv->tgts[i].ltd_exp);
842 mdc_obd->obd_force = obddev->obd_force;
843 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
845 if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
848 if (lmv->tgts[i].ltd_active) {
849 CERROR("error: iocontrol MDC %s on MDT"
850 "idx %d cmd %x: err = %d\n",
851 lmv->tgts[i].ltd_uuid.uuid,
866 static int lmv_all_chars_policy(int count, const char *name,
877 static int lmv_nid_policy(struct lmv_obd *lmv)
879 struct obd_import *imp;
883 * XXX: To get nid we assume that underlying obd device is mdc.
885 imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
886 id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
887 return id % lmv->desc.ld_tgt_count;
890 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
891 placement_policy_t placement)
894 case PLACEMENT_CHAR_POLICY:
895 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
897 op_data->op_namelen);
898 case PLACEMENT_NID_POLICY:
899 return lmv_nid_policy(lmv);
905 CERROR("Unsupported placement policy %x\n", placement);
910 * This is _inode_ placement policy function (not name).
912 static int lmv_placement_policy(struct obd_device *obd,
913 struct md_op_data *op_data,
916 struct lmv_obd *lmv = &obd->u.lmv;
917 struct lmv_object *obj;
921 LASSERT(mds != NULL);
923 if (lmv->desc.ld_tgt_count == 1) {
929 * Allocate new fid on target according to operation type and parent
932 obj = lmv_object_find(obd, &op_data->op_fid1);
933 if (obj != NULL || op_data->op_name == NULL ||
934 op_data->op_opc != LUSTRE_OPC_MKDIR) {
936 * Allocate fid for non-dir or for null name or for case parent
943 * If we have this flag turned on, and we see that
944 * parent dir is split, this means, that caller did not
945 * notice split yet. This is race and we would like to
946 * let caller know that.
948 if (op_data->op_bias & MDS_CHECK_SPLIT)
953 * Allocate new fid on same mds where parent fid is located and
954 * where operation will be sent. In case of split dir, ->op_fid1
955 * and ->op_mds here will contain fid and mds of slave directory
956 * object (assigned by caller).
958 *mds = op_data->op_mds;
962 * Parent directory is not split and we want to create a
963 * directory in it. Let's calculate where to place it according
964 * to operation data @op_data.
966 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
971 CERROR("Can't choose MDS, err = %d\n", rc);
973 LASSERT(*mds < lmv->desc.ld_tgt_count);
979 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
982 struct lmv_tgt_desc *tgt;
986 tgt = lmv_get_target(lmv, mds);
989 * New seq alloc and FLD setup should be atomic. Otherwise we may find
990 * on server that seq in new allocated fid is not yet known.
992 cfs_down(&tgt->ltd_fid_sem);
994 if (!tgt->ltd_active)
995 GOTO(out, rc = -ENODEV);
998 * Asking underlaying tgt layer to allocate new fid.
1000 rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
1002 LASSERT(fid_is_sane(fid));
1008 cfs_up(&tgt->ltd_fid_sem);
1012 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1013 struct md_op_data *op_data)
1015 struct obd_device *obd = class_exp2obd(exp);
1016 struct lmv_obd *lmv = &obd->u.lmv;
1021 LASSERT(op_data != NULL);
1022 LASSERT(fid != NULL);
1024 rc = lmv_placement_policy(obd, op_data, &mds);
1026 CERROR("Can't get target for allocating fid, "
1031 rc = __lmv_fid_alloc(lmv, fid, mds);
1033 CERROR("Can't alloc new fid, rc %d\n", rc);
1040 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1043 LASSERT(exp != NULL && fid != NULL);
1044 if (lmv_object_delete(exp, fid)) {
1045 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1051 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1053 struct lmv_obd *lmv = &obd->u.lmv;
1054 struct lprocfs_static_vars lvars;
1055 struct lmv_desc *desc;
1060 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1061 CERROR("LMV setup requires a descriptor\n");
1065 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1066 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1067 CERROR("Lmv descriptor size wrong: %d > %d\n",
1068 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1072 lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1074 OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1075 if (lmv->tgts == NULL)
1078 for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1079 cfs_sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1080 lmv->tgts[i].ltd_idx = i;
1083 lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1085 OBD_ALLOC(lmv->datas, lmv->datas_size);
1086 if (lmv->datas == NULL)
1087 GOTO(out_free_tgts, rc = -ENOMEM);
1089 obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1090 lmv->desc.ld_tgt_count = 0;
1091 lmv->desc.ld_active_tgt_count = 0;
1092 lmv->max_cookiesize = 0;
1093 lmv->max_def_easize = 0;
1094 lmv->max_easize = 0;
1095 lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1097 cfs_spin_lock_init(&lmv->lmv_lock);
1098 cfs_sema_init(&lmv->init_sem, 1);
1100 rc = lmv_object_setup(obd);
1102 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1103 GOTO(out_free_datas, rc);
1106 lprocfs_lmv_init_vars(&lvars);
1107 lprocfs_obd_setup(obd, lvars.obd_vars);
1110 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1111 0444, &lmv_proc_target_fops, obd);
1113 CWARN("Error adding target_obd_stats file (%d)\n", rc);
1116 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1117 LUSTRE_CLI_FLD_HASH_DHT);
1119 CERROR("Can't init FLD, err %d\n", rc);
1120 GOTO(out_free_datas, rc);
1126 OBD_FREE(lmv->datas, lmv->datas_size);
1129 OBD_FREE(lmv->tgts, lmv->tgts_size);
1134 static int lmv_cleanup(struct obd_device *obd)
1136 struct lmv_obd *lmv = &obd->u.lmv;
1139 fld_client_fini(&lmv->lmv_fld);
1140 lprocfs_obd_cleanup(obd);
1141 lmv_object_cleanup(obd);
1142 OBD_FREE(lmv->datas, lmv->datas_size);
1143 OBD_FREE(lmv->tgts, lmv->tgts_size);
1148 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1150 struct lustre_cfg *lcfg = buf;
1151 struct obd_uuid tgt_uuid;
1155 switch(lcfg->lcfg_command) {
1157 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1158 GOTO(out, rc = -EINVAL);
1160 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1161 rc = lmv_add_target(obd, &tgt_uuid);
1164 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1165 GOTO(out, rc = -EINVAL);
1172 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1173 __u64 max_age, __u32 flags)
1175 struct lmv_obd *lmv = &obd->u.lmv;
1176 struct obd_statfs *temp;
1181 rc = lmv_check_connect(obd);
1185 OBD_ALLOC(temp, sizeof(*temp));
1189 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1190 if (lmv->tgts[i].ltd_exp == NULL)
1193 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1196 CERROR("can't stat MDS #%d (%s), error %d\n", i,
1197 lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1199 GOTO(out_free_temp, rc);
1204 osfs->os_bavail += temp->os_bavail;
1205 osfs->os_blocks += temp->os_blocks;
1206 osfs->os_ffree += temp->os_ffree;
1207 osfs->os_files += temp->os_files;
1213 OBD_FREE(temp, sizeof(*temp));
1217 static int lmv_getstatus(struct obd_export *exp,
1219 struct obd_capa **pc)
1221 struct obd_device *obd = exp->exp_obd;
1222 struct lmv_obd *lmv = &obd->u.lmv;
1226 rc = lmv_check_connect(obd);
1230 rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1234 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1235 struct obd_capa *oc, obd_valid valid, const char *name,
1236 const char *input, int input_size, int output_size,
1237 int flags, struct ptlrpc_request **request)
1239 struct obd_device *obd = exp->exp_obd;
1240 struct lmv_obd *lmv = &obd->u.lmv;
1241 struct lmv_tgt_desc *tgt;
1245 rc = lmv_check_connect(obd);
1249 tgt = lmv_find_target(lmv, fid);
1251 RETURN(PTR_ERR(tgt));
1253 rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1254 input_size, output_size, flags, request);
1259 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1260 struct obd_capa *oc, obd_valid valid, const char *name,
1261 const char *input, int input_size, int output_size,
1262 int flags, __u32 suppgid,
1263 struct ptlrpc_request **request)
1265 struct obd_device *obd = exp->exp_obd;
1266 struct lmv_obd *lmv = &obd->u.lmv;
1267 struct lmv_tgt_desc *tgt;
1271 rc = lmv_check_connect(obd);
1275 tgt = lmv_find_target(lmv, fid);
1277 RETURN(PTR_ERR(tgt));
1279 rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1280 input_size, output_size, flags, suppgid,
1286 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1287 struct ptlrpc_request **request)
1289 struct obd_device *obd = exp->exp_obd;
1290 struct lmv_obd *lmv = &obd->u.lmv;
1291 struct lmv_tgt_desc *tgt;
1292 struct lmv_object *obj;
1297 rc = lmv_check_connect(obd);
1301 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1303 RETURN(PTR_ERR(tgt));
1305 if (op_data->op_valid & OBD_MD_MDTIDX) {
1306 op_data->op_mds = tgt->ltd_idx;
1310 rc = md_getattr(tgt->ltd_exp, op_data, request);
1314 obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1316 CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1317 obj ? "(split)" : "");
1320 * If object is split, then we loop over all the slaves and gather size
1321 * attribute. In ideal world we would have to gather also mds field from
1322 * all slaves, as object is spread over the cluster and this is
1323 * definitely interesting information and it is not good to loss it,
1327 struct mdt_body *body;
1329 if (*request == NULL) {
1330 lmv_object_put(obj);
1334 body = req_capsule_server_get(&(*request)->rq_pill,
1336 LASSERT(body != NULL);
1338 for (i = 0; i < obj->lo_objcount; i++) {
1339 if (lmv->tgts[i].ltd_exp == NULL) {
1340 CWARN("%s: NULL export for %d\n",
1346 * Skip master object.
1348 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1351 body->size += obj->lo_stripes[i].ls_size;
1354 lmv_object_put_unlock(obj);
1360 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1361 ldlm_iterator_t it, void *data)
1363 struct obd_device *obd = exp->exp_obd;
1364 struct lmv_obd *lmv = &obd->u.lmv;
1369 rc = lmv_check_connect(obd);
1373 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1376 * With CMD every object can have two locks in different namespaces:
1377 * lookup lock in space of mds storing direntry and update/open lock in
1378 * space of mds storing inode.
1380 for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1381 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1386 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1387 ldlm_iterator_t it, void *data)
1389 struct obd_device *obd = exp->exp_obd;
1390 struct lmv_obd *lmv = &obd->u.lmv;
1395 rc = lmv_check_connect(obd);
1399 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1402 * With CMD every object can have two locks in different namespaces:
1403 * lookup lock in space of mds storing direntry and update/open lock in
1404 * space of mds storing inode.
1406 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1407 rc = md_find_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1416 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1417 struct md_open_data *mod, struct ptlrpc_request **request)
1419 struct obd_device *obd = exp->exp_obd;
1420 struct lmv_obd *lmv = &obd->u.lmv;
1421 struct lmv_tgt_desc *tgt;
1425 rc = lmv_check_connect(obd);
1429 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1431 RETURN(PTR_ERR(tgt));
1433 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1434 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1439 * Called in the case MDS returns -ERESTART on create on open, what means that
1440 * directory is split and its LMV presentation object has to be updated.
1442 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1444 struct obd_device *obd = exp->exp_obd;
1445 struct lmv_obd *lmv = &obd->u.lmv;
1446 struct ptlrpc_request *req = NULL;
1447 struct lmv_tgt_desc *tgt;
1448 struct lmv_object *obj;
1449 struct lustre_md md;
1450 struct md_op_data *op_data;
1457 mealen = lmv_get_easize(lmv);
1459 valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1461 tgt = lmv_find_target(lmv, fid);
1463 RETURN(PTR_ERR(tgt));
1466 * Time to update mea of parent fid.
1469 OBD_ALLOC_PTR(op_data);
1470 if (op_data == NULL)
1473 op_data->op_fid1 = *fid;
1474 op_data->op_mode = mealen;
1475 op_data->op_valid = valid;
1477 rc = md_getattr(tgt->ltd_exp, op_data, &req);
1478 OBD_FREE_PTR(op_data);
1480 CERROR("md_getattr() failed, error %d\n", rc);
1484 rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1486 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1491 GOTO(cleanup, rc = -ENODATA);
1493 obj = lmv_object_create(exp, fid, md.mea);
1497 lmv_object_put(obj);
1499 obd_free_memmd(exp, (void *)&md.mea);
1503 ptlrpc_req_finished(req);
1507 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1508 const void *data, int datalen, int mode, __u32 uid,
1509 __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1510 struct ptlrpc_request **request)
1512 struct obd_device *obd = exp->exp_obd;
1513 struct lmv_obd *lmv = &obd->u.lmv;
1514 struct lmv_tgt_desc *tgt;
1515 struct lmv_object *obj;
1521 rc = lmv_check_connect(obd);
1525 if (!lmv->desc.ld_active_tgt_count)
1531 obj = lmv_object_find(obd, &op_data->op_fid1);
1533 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1534 op_data->op_name, op_data->op_namelen);
1535 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1536 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1537 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1538 tgt = lmv_get_target(lmv, op_data->op_mds);
1539 lmv_object_put(obj);
1541 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1542 op_data->op_bias |= MDS_CHECK_SPLIT;
1543 op_data->op_mds = tgt->ltd_idx;
1547 RETURN(PTR_ERR(tgt));
1549 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1550 if (rc == -ERESTART)
1555 CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1556 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1559 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1560 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1561 cap_effective, rdev, request);
1563 if (*request == NULL)
1565 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1566 } else if (rc == -ERESTART) {
1567 LASSERT(*request != NULL);
1568 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1569 "Got -ERESTART during create!\n");
1570 ptlrpc_req_finished(*request);
1574 * Directory got split. Time to update local object and repeat
1575 * the request with proper MDS.
1577 rc = lmv_handle_split(exp, &op_data->op_fid1);
1579 rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1580 op_data, &op_data->op_fid2);
1589 static int lmv_done_writing(struct obd_export *exp,
1590 struct md_op_data *op_data,
1591 struct md_open_data *mod)
1593 struct obd_device *obd = exp->exp_obd;
1594 struct lmv_obd *lmv = &obd->u.lmv;
1595 struct lmv_tgt_desc *tgt;
1599 rc = lmv_check_connect(obd);
1603 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1605 RETURN(PTR_ERR(tgt));
1607 rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1612 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1613 struct lookup_intent *it, struct md_op_data *op_data,
1614 struct lustre_handle *lockh, void *lmm, int lmmsize)
1616 struct obd_device *obd = exp->exp_obd;
1617 struct lmv_obd *lmv = &obd->u.lmv;
1618 struct lmv_stripe_md *mea = op_data->op_mea1;
1619 struct md_op_data *op_data2;
1620 struct lmv_tgt_desc *tgt;
1625 OBD_ALLOC_PTR(op_data2);
1626 if (op_data2 == NULL)
1629 LASSERT(mea != NULL);
1630 for (i = 0; i < mea->mea_count; i++) {
1631 memset(op_data2, 0, sizeof(*op_data2));
1632 op_data2->op_fid1 = mea->mea_ids[i];
1633 op_data2->op_bias = 0;
1635 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1637 GOTO(cleanup, rc = PTR_ERR(tgt));
1639 if (tgt->ltd_exp == NULL)
1642 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1643 lockh + i, lmm, lmmsize, NULL, 0);
1645 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1646 PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1651 if (it->d.lustre.it_data) {
1652 struct ptlrpc_request *req;
1653 req = (struct ptlrpc_request *)it->d.lustre.it_data;
1654 ptlrpc_req_finished(req);
1657 if (it->d.lustre.it_status)
1658 GOTO(cleanup, rc = it->d.lustre.it_status);
1663 OBD_FREE_PTR(op_data2);
1667 * Drop all taken locks.
1670 if (lockh[i].cookie)
1671 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1672 lockh[i].cookie = 0;
1679 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1680 struct lookup_intent *it, struct md_op_data *op_data,
1681 struct lustre_handle *lockh, void *lmm, int lmmsize,
1682 int extra_lock_flags)
1684 struct ptlrpc_request *req = it->d.lustre.it_data;
1685 struct obd_device *obd = exp->exp_obd;
1686 struct lmv_obd *lmv = &obd->u.lmv;
1687 struct lustre_handle plock;
1688 struct lmv_tgt_desc *tgt;
1689 struct md_op_data *rdata;
1691 struct mdt_body *body;
1696 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1697 LASSERT(body != NULL);
1699 if (!(body->valid & OBD_MD_MDS))
1702 CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1703 LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1706 * We got LOOKUP lock, but we really need attrs.
1708 pmode = it->d.lustre.it_lock_mode;
1709 LASSERT(pmode != 0);
1710 memcpy(&plock, lockh, sizeof(plock));
1711 it->d.lustre.it_lock_mode = 0;
1712 it->d.lustre.it_data = NULL;
1715 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1716 ptlrpc_req_finished(req);
1718 tgt = lmv_find_target(lmv, &fid1);
1720 GOTO(out, rc = PTR_ERR(tgt));
1722 OBD_ALLOC_PTR(rdata);
1724 GOTO(out, rc = -ENOMEM);
1726 rdata->op_fid1 = fid1;
1727 rdata->op_bias = MDS_CROSS_REF;
1729 rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1730 lmm, lmmsize, NULL, extra_lock_flags);
1731 OBD_FREE_PTR(rdata);
1734 ldlm_lock_decref(&plock, pmode);
1739 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1740 struct lookup_intent *it, struct md_op_data *op_data,
1741 struct lustre_handle *lockh, void *lmm, int lmmsize,
1742 struct ptlrpc_request **req, int extra_lock_flags)
1744 struct obd_device *obd = exp->exp_obd;
1745 struct lmv_obd *lmv = &obd->u.lmv;
1746 struct lmv_tgt_desc *tgt;
1747 struct lmv_object *obj;
1752 rc = lmv_check_connect(obd);
1756 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1757 LL_IT2STR(it), PFID(&op_data->op_fid1));
1759 if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1760 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1761 lockh, lmm, lmmsize);
1765 obj = lmv_object_find(obd, &op_data->op_fid1);
1766 if (obj && op_data->op_namelen) {
1767 sidx = raw_name2idx(obj->lo_hashtype,
1769 (char *)op_data->op_name,
1770 op_data->op_namelen);
1771 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1772 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1774 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1777 lmv_object_put(obj);
1780 RETURN(PTR_ERR(tgt));
1782 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1783 LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1785 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1786 lmm, lmmsize, req, extra_lock_flags);
1788 if (rc == 0 && it && it->it_op == IT_OPEN) {
1789 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1790 lmm, lmmsize, extra_lock_flags);
1796 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1797 struct ptlrpc_request **request)
1799 struct ptlrpc_request *req = NULL;
1800 struct obd_device *obd = exp->exp_obd;
1801 struct lmv_obd *lmv = &obd->u.lmv;
1802 struct lu_fid rid = op_data->op_fid1;
1803 struct lmv_tgt_desc *tgt;
1804 struct mdt_body *body;
1805 struct lmv_object *obj;
1806 obd_valid valid = op_data->op_valid;
1812 rc = lmv_check_connect(obd);
1819 obj = lmv_object_find(obd, &rid);
1821 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1822 op_data->op_name, op_data->op_namelen);
1823 rid = obj->lo_stripes[sidx].ls_fid;
1824 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1825 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1826 valid &= ~OBD_MD_FLCKSPLIT;
1827 lmv_object_put(obj);
1829 tgt = lmv_find_target(lmv, &rid);
1830 valid |= OBD_MD_FLCKSPLIT;
1831 op_data->op_mds = tgt->ltd_idx;
1834 RETURN(PTR_ERR(tgt));
1836 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1837 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1838 PFID(&rid), tgt->ltd_idx);
1840 op_data->op_valid = valid;
1841 op_data->op_fid1 = rid;
1842 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1844 body = req_capsule_server_get(&(*request)->rq_pill,
1846 LASSERT(body != NULL);
1848 if (body->valid & OBD_MD_MDS) {
1850 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1853 tgt = lmv_find_target(lmv, &rid);
1855 ptlrpc_req_finished(*request);
1856 RETURN(PTR_ERR(tgt));
1859 op_data->op_fid1 = rid;
1860 op_data->op_valid |= OBD_MD_FLCROSSREF;
1861 op_data->op_namelen = 0;
1862 op_data->op_name = NULL;
1863 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1864 ptlrpc_req_finished(*request);
1867 } else if (rc == -ERESTART) {
1868 LASSERT(*request != NULL);
1869 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1870 "Got -ERESTART during getattr!\n");
1871 ptlrpc_req_finished(*request);
1875 * Directory got split. Time to update local object and repeat
1876 * the request with proper MDS.
1878 rc = lmv_handle_split(exp, &rid);
1885 #define md_op_data_fid(op_data, fl) \
1886 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1887 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1888 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1889 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1892 static int lmv_early_cancel_slaves(struct obd_export *exp,
1893 struct md_op_data *op_data, int op_tgt,
1894 ldlm_mode_t mode, int bits, int flag)
1896 struct obd_device *obd = exp->exp_obd;
1897 struct lmv_obd *lmv = &obd->u.lmv;
1898 ldlm_policy_data_t policy = {{0}};
1899 struct lu_fid *op_fid;
1900 struct lu_fid *st_fid;
1901 struct lmv_tgt_desc *tgt;
1902 struct lmv_object *obj;
1907 op_fid = md_op_data_fid(op_data, flag);
1908 if (!fid_is_sane(op_fid))
1911 obj = lmv_object_find(obd, op_fid);
1915 policy.l_inodebits.bits = bits;
1916 for (i = 0; i < obj->lo_objcount; i++) {
1917 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1918 st_fid = &obj->lo_stripes[i].ls_fid;
1919 if (op_tgt != tgt->ltd_idx) {
1920 CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1921 PFID(st_fid), tgt->ltd_idx);
1922 rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1923 mode, LCF_ASYNC, NULL);
1925 GOTO(out_put_obj, rc);
1928 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1929 op_tgt, PFID(st_fid));
1931 * Do not cancel locks for operation target, they will
1932 * be handled later in underlaying layer when calling
1933 * function we run on behalf of.
1936 op_data->op_flags |= flag;
1941 lmv_object_put(obj);
1945 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1946 int op_tgt, ldlm_mode_t mode, int bits, int flag)
1948 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1949 struct obd_device *obd = exp->exp_obd;
1950 struct lmv_obd *lmv = &obd->u.lmv;
1951 struct lmv_tgt_desc *tgt;
1952 ldlm_policy_data_t policy = {{0}};
1953 struct lmv_object *obj;
1957 if (!fid_is_sane(fid))
1960 obj = lmv_object_find(obd, fid);
1962 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1964 lmv_object_put(obj);
1966 tgt = lmv_find_target(lmv, fid);
1968 RETURN(PTR_ERR(tgt));
1970 if (tgt->ltd_idx != op_tgt) {
1971 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1972 policy.l_inodebits.bits = bits;
1973 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1974 mode, LCF_ASYNC, NULL);
1977 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1979 op_data->op_flags |= flag;
1988 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1991 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1992 struct ptlrpc_request **request)
1994 struct obd_device *obd = exp->exp_obd;
1995 struct lmv_obd *lmv = &obd->u.lmv;
1996 struct lmv_tgt_desc *tgt;
1997 struct lmv_object *obj;
2004 rc = lmv_check_connect(obd);
2011 LASSERT(op_data->op_namelen != 0);
2013 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2014 PFID(&op_data->op_fid2), op_data->op_namelen,
2015 op_data->op_name, PFID(&op_data->op_fid1));
2017 obj = lmv_object_find(obd, &op_data->op_fid2);
2019 sidx = raw_name2idx(obj->lo_hashtype,
2022 op_data->op_namelen);
2023 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2024 mds = obj->lo_stripes[sidx].ls_mds;
2025 lmv_object_put(obj);
2027 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
2032 CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
2033 mds, PFID(&op_data->op_fid1));
2035 op_data->op_fsuid = cfs_curproc_fsuid();
2036 op_data->op_fsgid = cfs_curproc_fsgid();
2037 op_data->op_cap = cfs_curproc_cap_pack();
2038 tgt = lmv_get_target(lmv, mds);
2041 * Cancel UPDATE lock on child (fid1).
2043 op_data->op_flags |= MF_MDC_CANCEL_FID2;
2044 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2045 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2047 rc = md_link(tgt->ltd_exp, op_data, request);
2048 if (rc == -ERESTART) {
2049 LASSERT(*request != NULL);
2050 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2051 "Got -ERESTART during link!\n");
2052 ptlrpc_req_finished(*request);
2056 * Directory got split. Time to update local object and repeat
2057 * the request with proper MDS.
2059 rc = lmv_handle_split(exp, &op_data->op_fid2);
2067 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2068 const char *old, int oldlen, const char *new, int newlen,
2069 struct ptlrpc_request **request)
2071 struct obd_device *obd = exp->exp_obd;
2072 struct lmv_obd *lmv = &obd->u.lmv;
2073 struct lmv_tgt_desc *src_tgt;
2074 struct lmv_tgt_desc *tgt_tgt;
2078 struct lmv_object *obj;
2083 LASSERT(oldlen != 0);
2085 CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2086 oldlen, old, PFID(&op_data->op_fid1),
2087 newlen, new, PFID(&op_data->op_fid2));
2089 rc = lmv_check_connect(obd);
2096 obj = lmv_object_find(obd, &op_data->op_fid1);
2098 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2099 (char *)old, oldlen);
2100 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2101 mds1 = obj->lo_stripes[sidx].ls_mds;
2102 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2103 lmv_object_put(obj);
2105 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2110 obj = lmv_object_find(obd, &op_data->op_fid2);
2113 * Directory is already split, so we have to forward request to
2116 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2117 (char *)new, newlen);
2119 mds2 = obj->lo_stripes[sidx].ls_mds;
2120 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2121 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2122 lmv_object_put(obj);
2124 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2129 op_data->op_fsuid = cfs_curproc_fsuid();
2130 op_data->op_fsgid = cfs_curproc_fsgid();
2131 op_data->op_cap = cfs_curproc_cap_pack();
2133 src_tgt = lmv_get_target(lmv, mds1);
2134 tgt_tgt = lmv_get_target(lmv, mds2);
2137 * LOOKUP lock on src child (fid3) should also be cancelled for
2138 * src_tgt in mdc_rename.
2140 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2143 * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2146 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2147 LCK_EX, MDS_INODELOCK_UPDATE,
2148 MF_MDC_CANCEL_FID2);
2151 * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2154 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2155 LCK_EX, MDS_INODELOCK_LOOKUP,
2156 MF_MDC_CANCEL_FID4);
2160 * Cancel all the locks on tgt child (fid4).
2163 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2164 LCK_EX, MDS_INODELOCK_FULL,
2165 MF_MDC_CANCEL_FID4);
2168 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2169 new, newlen, request);
2171 if (rc == -ERESTART) {
2172 LASSERT(*request != NULL);
2173 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2174 "Got -ERESTART during rename!\n");
2175 ptlrpc_req_finished(*request);
2179 * Directory got split. Time to update local object and repeat
2180 * the request with proper MDS.
2182 rc = lmv_handle_split(exp, &op_data->op_fid1);
2189 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2190 void *ea, int ealen, void *ea2, int ea2len,
2191 struct ptlrpc_request **request,
2192 struct md_open_data **mod)
2194 struct obd_device *obd = exp->exp_obd;
2195 struct lmv_obd *lmv = &obd->u.lmv;
2196 struct ptlrpc_request *req;
2197 struct lmv_tgt_desc *tgt;
2198 struct lmv_object *obj;
2203 rc = lmv_check_connect(obd);
2207 obj = lmv_object_find(obd, &op_data->op_fid1);
2209 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2210 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2211 obj ? ", split" : "");
2213 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2215 for (i = 0; i < obj->lo_objcount; i++) {
2216 op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2218 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2224 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2225 ea2, ea2len, &req, mod);
2227 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2229 * This is master object and this request should
2230 * be returned back to llite.
2234 ptlrpc_req_finished(req);
2240 lmv_object_put(obj);
2242 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2244 RETURN(PTR_ERR(tgt));
2246 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2247 ea2len, request, mod);
2252 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2253 struct obd_capa *oc, struct ptlrpc_request **request)
2255 struct obd_device *obd = exp->exp_obd;
2256 struct lmv_obd *lmv = &obd->u.lmv;
2257 struct lmv_tgt_desc *tgt;
2261 rc = lmv_check_connect(obd);
2265 tgt = lmv_find_target(lmv, fid);
2267 RETURN(PTR_ERR(tgt));
2269 rc = md_sync(tgt->ltd_exp, fid, oc, request);
2274 * Main purpose of LMV blocking ast is to remove split directory LMV
2275 * presentation object (struct lmv_object) attached to the lock being revoked.
2277 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2278 void *data, int flag)
2280 struct lustre_handle lockh;
2281 struct lmv_object *obj;
2286 case LDLM_CB_BLOCKING:
2287 ldlm_lock2handle(lock, &lockh);
2288 rc = ldlm_cli_cancel(&lockh);
2290 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2294 case LDLM_CB_CANCELING:
2296 * Time to drop cached attrs for split directory object
2298 obj = lock->l_ast_data;
2300 CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2301 ", master "DFID"\n",
2302 lock->l_resource->lr_name.name[3] == 1 ?
2303 "LOOKUP" : "UPDATE",
2304 lock->l_resource->lr_name.name[0],
2305 lock->l_resource->lr_name.name[1],
2306 PFID(&obj->lo_fid));
2307 lmv_object_put(obj);
2316 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2320 val = le64_to_cpu(*hash);
2322 val += MAX_HASH_SIZE;
2323 if (val != DIR_END_OFF)
2324 *hash = cpu_to_le64(val - hash_adj);
2327 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2330 struct obd_import *imp;
2333 * XXX: to get nid we assume that underlying obd device is mdc.
2335 imp = class_exp2cliimp(exp);
2336 id = imp->imp_connection->c_self + fid_flatten(fid);
2338 CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2339 imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2341 return id ^ (id >> 32);
2344 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2345 struct obd_capa *oc, __u64 offset64, struct page *page,
2346 struct ptlrpc_request **request)
2348 struct obd_device *obd = exp->exp_obd;
2349 struct lmv_obd *lmv = &obd->u.lmv;
2350 struct lu_fid rid = *fid;
2351 struct lmv_object *obj;
2361 struct lmv_stripe *los;
2362 struct lmv_tgt_desc *tgt;
2363 struct lu_dirpage *dp;
2364 struct lu_dirent *ent;
2369 rc = lmv_check_connect(obd);
2373 CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2376 * This case handle directory lookup in clustered metadata case (i.e.
2377 * split directory is located on multiple md servers.)
2378 * each server keeps directory entries for certain range of hashes.
2379 * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2380 * first server will keep records with hashes [ 0 ... MAX_HASH / N - 1],
2381 * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2383 * readdir can simply start reading entries from 0 - N server in
2384 * order but that will not scale well as all client will request dir in
2385 * to server in same order.
2386 * Following algorithm does optimization:
2387 * Instead of doing readdir in 1, 2, ...., N order, client with a
2388 * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2389 * (every client has rank R)
2390 * But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2391 * since client ask dir from MDS{R} client has pages with offsets
2392 * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2393 * on hash values that we get.
2395 obj = lmv_object_find_lock(obd, fid);
2397 nr = obj->lo_objcount;
2399 seg_size = MAX_HASH_SIZE;
2400 do_div(seg_size, nr);
2401 los = obj->lo_stripes;
2402 tgt = lmv_get_target(lmv, los[0].ls_mds);
2403 rank = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2405 do_div(tgt_tmp, seg_size);
2406 tgt0_idx = do_div(tgt_tmp, nr);
2407 tgt_idx = (tgt0_idx + rank) % nr;
2409 if (tgt_idx < tgt0_idx)
2413 * Last segment has unusual length due to division
2416 hash_adj = MAX_HASH_SIZE - seg_size * nr;
2420 hash_adj += rank * seg_size;
2422 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2423 LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2424 offset, tgt0_idx, offset + hash_adj, tgt_idx);
2426 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2427 rid = obj->lo_stripes[tgt_idx].ls_fid;
2428 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2430 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2431 PFID(&rid), (unsigned long)offset, tgt_idx);
2433 tgt = lmv_find_target(lmv, &rid);
2436 GOTO(cleanup, rc = PTR_ERR(tgt));
2438 rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2442 dp = cfs_kmap(page);
2444 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2445 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2446 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2448 for (ent = lu_dirent_start(dp); ent != NULL;
2449 ent = lu_dirent_next(ent))
2450 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2452 if (tgt0_idx != nr - 1) {
2455 end = le64_to_cpu(dp->ldp_hash_end);
2456 if (end == DIR_END_OFF) {
2457 dp->ldp_hash_end = cpu_to_le32(seg_size *
2460 ""DFID" reset end "LPX64" tgt %d\n",
2462 (__u64)le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2470 lmv_object_put_unlock(obj);
2474 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2475 struct ptlrpc_request **request)
2477 struct obd_device *obd = exp->exp_obd;
2478 struct lmv_obd *lmv = &obd->u.lmv;
2479 struct lmv_tgt_desc *tgt = NULL;
2480 struct lmv_object *obj;
2486 rc = lmv_check_connect(obd);
2493 LASSERT(op_data->op_namelen != 0);
2495 obj = lmv_object_find(obd, &op_data->op_fid1);
2497 sidx = raw_name2idx(obj->lo_hashtype,
2500 op_data->op_namelen);
2501 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2502 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2503 tgt = lmv_get_target(lmv,
2504 obj->lo_stripes[sidx].ls_mds);
2505 lmv_object_put(obj);
2506 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2507 op_data->op_namelen, op_data->op_name,
2508 PFID(&op_data->op_fid1), sidx);
2512 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2514 RETURN(PTR_ERR(tgt));
2515 op_data->op_bias |= MDS_CHECK_SPLIT;
2518 op_data->op_fsuid = cfs_curproc_fsuid();
2519 op_data->op_fsgid = cfs_curproc_fsgid();
2520 op_data->op_cap = cfs_curproc_cap_pack();
2523 * If child's fid is given, cancel unused locks for it if it is from
2524 * another export than parent.
2526 * LOOKUP lock for child (fid3) should also be cancelled on parent
2527 * tgt_tgt in mdc_unlink().
2529 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2532 * Cancel FULL locks on child (fid3).
2534 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2535 MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2538 rc = md_unlink(tgt->ltd_exp, op_data, request);
2540 if (rc == -ERESTART) {
2541 LASSERT(*request != NULL);
2542 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2543 "Got -ERESTART during unlink!\n");
2544 ptlrpc_req_finished(*request);
2548 * Directory got split. Time to update local object and repeat
2549 * the request with proper MDS.
2551 rc = lmv_handle_split(exp, &op_data->op_fid1);
2558 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2563 case OBD_CLEANUP_EARLY:
2564 /* XXX: here should be calling obd_precleanup() down to
2567 case OBD_CLEANUP_EXPORTS:
2568 rc = obd_llog_finish(obd, 0);
2570 CERROR("failed to cleanup llogging subsystems\n");
2578 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2579 void *key, __u32 *vallen, void *val,
2580 struct lov_stripe_md *lsm)
2582 struct obd_device *obd;
2583 struct lmv_obd *lmv;
2587 obd = class_exp2obd(exp);
2589 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2590 exp->exp_handle.h_cookie);
2595 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2596 struct lmv_tgt_desc *tgts;
2599 rc = lmv_check_connect(obd);
2603 LASSERT(*vallen == sizeof(__u32));
2604 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2608 * All tgts should be connected when this gets called.
2610 if (!tgts || !tgts->ltd_exp) {
2611 CERROR("target not setup?\n");
2615 if (!obd_get_info(tgts->ltd_exp, keylen, key,
2620 } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2621 rc = lmv_check_connect(obd);
2626 * Forwarding this request to first MDS, it should know LOV
2629 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2631 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2632 exp->exp_connect_flags =
2633 ((struct obd_connect_data *)val)->ocd_connect_flags;
2638 CDEBUG(D_IOCTL, "Invalid key\n");
2642 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2643 void *key, obd_count vallen, void *val,
2644 struct ptlrpc_request_set *set)
2646 struct lmv_tgt_desc *tgt;
2647 struct obd_device *obd;
2648 struct lmv_obd *lmv;
2652 obd = class_exp2obd(exp);
2654 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2655 exp->exp_handle.h_cookie);
2660 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX)) {
2663 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2664 tgt = &lmv->tgts[i];
2669 err = obd_set_info_async(tgt->ltd_exp,
2670 keylen, key, vallen, val, set);
2681 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2682 struct lov_stripe_md *lsm)
2684 struct obd_device *obd = class_exp2obd(exp);
2685 struct lmv_obd *lmv = &obd->u.lmv;
2686 struct lmv_stripe_md *meap;
2687 struct lmv_stripe_md *lsmp;
2692 mea_size = lmv_get_easize(lmv);
2696 if (*lmmp && !lsm) {
2697 OBD_FREE_LARGE(*lmmp, mea_size);
2702 if (*lmmp == NULL) {
2703 OBD_ALLOC_LARGE(*lmmp, mea_size);
2711 lsmp = (struct lmv_stripe_md *)lsm;
2712 meap = (struct lmv_stripe_md *)*lmmp;
2714 if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2715 lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2718 meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2719 meap->mea_count = cpu_to_le32(lsmp->mea_count);
2720 meap->mea_master = cpu_to_le32(lsmp->mea_master);
2722 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2723 meap->mea_ids[i] = meap->mea_ids[i];
2724 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2730 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2731 struct lov_mds_md *lmm, int lmm_size)
2733 struct obd_device *obd = class_exp2obd(exp);
2734 struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2735 struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2736 struct lmv_obd *lmv = &obd->u.lmv;
2742 mea_size = lmv_get_easize(lmv);
2746 if (*lsmp != NULL && lmm == NULL) {
2747 OBD_FREE_LARGE(*tmea, mea_size);
2752 LASSERT(mea_size == lmm_size);
2754 OBD_ALLOC_LARGE(*tmea, mea_size);
2761 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2762 mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2763 mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2765 magic = le32_to_cpu(mea->mea_magic);
2768 * Old mea is not handled here.
2770 CERROR("Old not supportable EA is found\n");
2774 (*tmea)->mea_magic = magic;
2775 (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2776 (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2778 for (i = 0; i < (*tmea)->mea_count; i++) {
2779 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2780 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2785 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2786 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2787 ldlm_cancel_flags_t flags, void *opaque)
2789 struct obd_device *obd = exp->exp_obd;
2790 struct lmv_obd *lmv = &obd->u.lmv;
2796 LASSERT(fid != NULL);
2798 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2799 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2802 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2803 policy, mode, flags, opaque);
2810 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2813 struct obd_device *obd = exp->exp_obd;
2814 struct lmv_obd *lmv = &obd->u.lmv;
2818 rc = md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2822 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2823 const struct lu_fid *fid, ldlm_type_t type,
2824 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2825 struct lustre_handle *lockh)
2827 struct obd_device *obd = exp->exp_obd;
2828 struct lmv_obd *lmv = &obd->u.lmv;
2833 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2836 * With CMD every object can have two locks in different namespaces:
2837 * lookup lock in space of mds storing direntry and update/open lock in
2838 * space of mds storing inode. Thus we check all targets, not only that
2839 * one fid was created in.
2841 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2842 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2843 type, policy, mode, lockh);
2851 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2852 struct obd_export *dt_exp, struct obd_export *md_exp,
2853 struct lustre_md *md)
2855 struct obd_device *obd = exp->exp_obd;
2856 struct lmv_obd *lmv = &obd->u.lmv;
2859 rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2863 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2865 struct obd_device *obd = exp->exp_obd;
2866 struct lmv_obd *lmv = &obd->u.lmv;
2870 obd_free_memmd(exp, (void *)&md->mea);
2871 RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2874 int lmv_set_open_replay_data(struct obd_export *exp,
2875 struct obd_client_handle *och,
2876 struct ptlrpc_request *open_req)
2878 struct obd_device *obd = exp->exp_obd;
2879 struct lmv_obd *lmv = &obd->u.lmv;
2880 struct lmv_tgt_desc *tgt;
2883 tgt = lmv_find_target(lmv, &och->och_fid);
2885 RETURN(PTR_ERR(tgt));
2887 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2890 int lmv_clear_open_replay_data(struct obd_export *exp,
2891 struct obd_client_handle *och)
2893 struct obd_device *obd = exp->exp_obd;
2894 struct lmv_obd *lmv = &obd->u.lmv;
2895 struct lmv_tgt_desc *tgt;
2898 tgt = lmv_find_target(lmv, &och->och_fid);
2900 RETURN(PTR_ERR(tgt));
2902 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2905 static int lmv_get_remote_perm(struct obd_export *exp,
2906 const struct lu_fid *fid,
2907 struct obd_capa *oc, __u32 suppgid,
2908 struct ptlrpc_request **request)
2910 struct obd_device *obd = exp->exp_obd;
2911 struct lmv_obd *lmv = &obd->u.lmv;
2912 struct lmv_tgt_desc *tgt;
2916 rc = lmv_check_connect(obd);
2920 tgt = lmv_find_target(lmv, fid);
2922 RETURN(PTR_ERR(tgt));
2924 rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2928 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2931 struct obd_device *obd = exp->exp_obd;
2932 struct lmv_obd *lmv = &obd->u.lmv;
2933 struct lmv_tgt_desc *tgt;
2937 rc = lmv_check_connect(obd);
2941 tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2943 RETURN(PTR_ERR(tgt));
2945 rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2949 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2950 const struct req_msg_field *field, struct obd_capa **oc)
2952 struct obd_device *obd = exp->exp_obd;
2953 struct lmv_obd *lmv = &obd->u.lmv;
2957 rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2961 int lmv_intent_getattr_async(struct obd_export *exp,
2962 struct md_enqueue_info *minfo,
2963 struct ldlm_enqueue_info *einfo)
2965 struct md_op_data *op_data = &minfo->mi_data;
2966 struct obd_device *obd = exp->exp_obd;
2967 struct lmv_obd *lmv = &obd->u.lmv;
2968 struct lmv_object *obj;
2969 struct lmv_tgt_desc *tgt;
2974 rc = lmv_check_connect(obd);
2978 if (!fid_is_sane(&op_data->op_fid2)) {
2979 obj = lmv_object_find(obd, &op_data->op_fid1);
2980 if (obj && op_data->op_namelen) {
2981 sidx = raw_name2idx(obj->lo_hashtype,
2983 (char *)op_data->op_name,
2984 op_data->op_namelen);
2985 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2986 tgt = lmv_get_target(lmv,
2987 obj->lo_stripes[sidx].ls_mds);
2989 "Choose slave dir ("DFID") -> mds #%d\n",
2990 PFID(&op_data->op_fid1), tgt->ltd_idx);
2992 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2995 lmv_object_put(obj);
2997 op_data->op_fid1 = op_data->op_fid2;
2998 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2999 op_data->op_bias = MDS_CROSS_REF;
3001 * Unfortunately, we have to lie to MDC/MDS to retrieve
3002 * attributes llite needs.
3004 if (minfo->mi_it.it_op & IT_LOOKUP)
3005 minfo->mi_it.it_op = IT_GETATTR;
3009 RETURN(PTR_ERR(tgt));
3011 rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
3015 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3018 struct obd_device *obd = exp->exp_obd;
3019 struct lmv_obd *lmv = &obd->u.lmv;
3020 struct lmv_tgt_desc *tgt;
3024 rc = lmv_check_connect(obd);
3028 tgt = lmv_find_target(lmv, fid);
3030 RETURN(PTR_ERR(tgt));
3032 rc = md_revalidate_lock(tgt->ltd_exp, it, fid);
3037 struct obd_ops lmv_obd_ops = {
3038 .o_owner = THIS_MODULE,
3039 .o_setup = lmv_setup,
3040 .o_cleanup = lmv_cleanup,
3041 .o_precleanup = lmv_precleanup,
3042 .o_process_config = lmv_process_config,
3043 .o_connect = lmv_connect,
3044 .o_disconnect = lmv_disconnect,
3045 .o_statfs = lmv_statfs,
3046 .o_get_info = lmv_get_info,
3047 .o_set_info_async = lmv_set_info_async,
3048 .o_packmd = lmv_packmd,
3049 .o_unpackmd = lmv_unpackmd,
3050 .o_notify = lmv_notify,
3051 .o_get_uuid = lmv_get_uuid,
3052 .o_iocontrol = lmv_iocontrol,
3053 .o_fid_delete = lmv_fid_delete
3056 struct md_ops lmv_md_ops = {
3057 .m_getstatus = lmv_getstatus,
3058 .m_change_cbdata = lmv_change_cbdata,
3059 .m_find_cbdata = lmv_find_cbdata,
3060 .m_close = lmv_close,
3061 .m_create = lmv_create,
3062 .m_done_writing = lmv_done_writing,
3063 .m_enqueue = lmv_enqueue,
3064 .m_getattr = lmv_getattr,
3065 .m_getxattr = lmv_getxattr,
3066 .m_getattr_name = lmv_getattr_name,
3067 .m_intent_lock = lmv_intent_lock,
3069 .m_rename = lmv_rename,
3070 .m_setattr = lmv_setattr,
3071 .m_setxattr = lmv_setxattr,
3073 .m_readpage = lmv_readpage,
3074 .m_unlink = lmv_unlink,
3075 .m_init_ea_size = lmv_init_ea_size,
3076 .m_cancel_unused = lmv_cancel_unused,
3077 .m_set_lock_data = lmv_set_lock_data,
3078 .m_lock_match = lmv_lock_match,
3079 .m_get_lustre_md = lmv_get_lustre_md,
3080 .m_free_lustre_md = lmv_free_lustre_md,
3081 .m_set_open_replay_data = lmv_set_open_replay_data,
3082 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3083 .m_renew_capa = lmv_renew_capa,
3084 .m_unpack_capa = lmv_unpack_capa,
3085 .m_get_remote_perm = lmv_get_remote_perm,
3086 .m_intent_getattr_async = lmv_intent_getattr_async,
3087 .m_revalidate_lock = lmv_revalidate_lock
3090 static quota_interface_t *quota_interface;
3091 extern quota_interface_t lmv_quota_interface;
3093 int __init lmv_init(void)
3095 struct lprocfs_static_vars lvars;
3098 lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3099 sizeof(struct lmv_object),
3101 if (!lmv_object_cache) {
3102 CERROR("Error allocating lmv objects cache\n");
3106 lprocfs_lmv_init_vars(&lvars);
3108 cfs_request_module("lquota");
3109 quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3110 init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3112 rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3113 lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3115 if (quota_interface)
3116 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3117 cfs_mem_cache_destroy(lmv_object_cache);
3124 static void lmv_exit(void)
3126 if (quota_interface)
3127 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3129 class_unregister_type(LUSTRE_LMV_NAME);
3131 LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3132 "Can't free lmv objects cache, %d object(s) busy\n",
3133 cfs_atomic_read(&lmv_object_count));
3134 cfs_mem_cache_destroy(lmv_object_cache);
3137 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3138 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3139 MODULE_LICENSE("GPL");
3141 module_init(lmv_init);
3142 module_exit(lmv_exit);