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 2008 Sun Microsystems, Inc. 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_no_recov = obd->obd_no_recov;
632 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
634 struct proc_dir_entry *mdc_symlink;
636 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
638 lprocfs_remove(&mdc_symlink);
640 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
641 obd->obd_type->typ_name, obd->obd_name,
646 rc = obd_fid_fini(tgt->ltd_exp);
648 CERROR("Can't finanize fids factory\n");
650 CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
651 tgt->ltd_exp->exp_obd->obd_name,
652 tgt->ltd_exp->exp_obd->obd_uuid.uuid);
654 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
655 rc = obd_disconnect(tgt->ltd_exp);
657 if (tgt->ltd_active) {
658 CERROR("Target %s disconnect error %d\n",
659 tgt->ltd_uuid.uuid, rc);
663 lmv_activate_target(lmv, tgt, 0);
668 static int lmv_disconnect(struct obd_export *exp)
670 struct obd_device *obd = class_exp2obd(exp);
672 struct proc_dir_entry *lmv_proc_dir;
674 struct lmv_obd *lmv = &obd->u.lmv;
683 * Only disconnect the underlying layers on the final disconnect.
686 if (lmv->refcount != 0)
689 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
690 if (lmv->tgts[i].ltd_exp == NULL)
692 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
696 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
698 lprocfs_remove(&lmv_proc_dir);
700 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
701 obd->obd_type->typ_name, obd->obd_name);
707 * This is the case when no real connection is established by
708 * lmv_check_connect().
711 class_export_put(exp);
712 rc = class_disconnect(exp);
713 if (lmv->refcount == 0)
718 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
719 int len, void *karg, void *uarg)
721 struct obd_device *obddev = class_exp2obd(exp);
722 struct lmv_obd *lmv = &obddev->u.lmv;
726 int count = lmv->desc.ld_tgt_count;
733 case IOC_OBD_STATFS: {
734 struct obd_ioctl_data *data = karg;
735 struct obd_device *mdc_obd;
736 struct obd_statfs stat_buf = {0};
739 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
740 if ((index >= count))
743 if (!lmv->tgts[index].ltd_active)
746 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
751 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
752 min((int) data->ioc_plen2,
753 (int) sizeof(struct obd_uuid))))
756 rc = obd_statfs(mdc_obd, &stat_buf,
757 cfs_time_current_64() - CFS_HZ, 0);
760 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
761 min((int) data->ioc_plen1,
762 (int) sizeof(stat_buf))))
766 case OBD_IOC_QUOTACTL: {
767 struct if_quotactl *qctl = karg;
768 struct lmv_tgt_desc *tgt = NULL;
769 struct obd_quotactl *oqctl;
771 if (qctl->qc_valid == QC_MDTIDX) {
772 if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
775 tgt = &lmv->tgts[qctl->qc_idx];
778 } else if (qctl->qc_valid == QC_UUID) {
779 for (i = 0; i < count; i++) {
781 if (!obd_uuid_equals(&tgt->ltd_uuid,
785 if (tgt->ltd_exp == NULL)
797 LASSERT(tgt && tgt->ltd_exp);
798 OBD_ALLOC_PTR(oqctl);
802 QCTL_COPY(oqctl, qctl);
803 rc = obd_quotactl(tgt->ltd_exp, oqctl);
805 QCTL_COPY(qctl, oqctl);
806 qctl->qc_valid = QC_MDTIDX;
807 qctl->obd_uuid = tgt->ltd_uuid;
812 case OBD_IOC_CHANGELOG_SEND:
813 case OBD_IOC_CHANGELOG_CLEAR: {
814 struct ioc_changelog *icc = karg;
816 if (icc->icc_mdtindex >= count)
819 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
820 sizeof(*icc), icc, NULL);
823 case LL_IOC_GET_CONNECT_FLAGS: {
824 rc = obd_iocontrol(cmd, lmv->tgts[0].ltd_exp, len, karg, uarg);
829 for (i = 0; i < count; i++) {
832 if (lmv->tgts[i].ltd_exp == NULL)
835 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
837 if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
840 if (lmv->tgts[i].ltd_active) {
841 CERROR("error: iocontrol MDC %s on MDT"
842 "idx %d cmd %x: err = %d\n",
843 lmv->tgts[i].ltd_uuid.uuid,
858 static int lmv_all_chars_policy(int count, const char *name,
869 static int lmv_nid_policy(struct lmv_obd *lmv)
871 struct obd_import *imp;
875 * XXX: To get nid we assume that underlying obd device is mdc.
877 imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
878 id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
879 return id % lmv->desc.ld_tgt_count;
882 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
883 placement_policy_t placement)
886 case PLACEMENT_CHAR_POLICY:
887 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
889 op_data->op_namelen);
890 case PLACEMENT_NID_POLICY:
891 return lmv_nid_policy(lmv);
897 CERROR("Unsupported placement policy %x\n", placement);
902 * This is _inode_ placement policy function (not name).
904 static int lmv_placement_policy(struct obd_device *obd,
905 struct md_op_data *op_data,
908 struct lmv_obd *lmv = &obd->u.lmv;
909 struct lmv_object *obj;
913 LASSERT(mds != NULL);
915 if (lmv->desc.ld_tgt_count == 1) {
921 * Allocate new fid on target according to operation type and parent
924 obj = lmv_object_find(obd, &op_data->op_fid1);
925 if (obj != NULL || op_data->op_name == NULL ||
926 op_data->op_opc != LUSTRE_OPC_MKDIR) {
928 * Allocate fid for non-dir or for null name or for case parent
935 * If we have this flag turned on, and we see that
936 * parent dir is split, this means, that caller did not
937 * notice split yet. This is race and we would like to
938 * let caller know that.
940 if (op_data->op_bias & MDS_CHECK_SPLIT)
945 * Allocate new fid on same mds where parent fid is located and
946 * where operation will be sent. In case of split dir, ->op_fid1
947 * and ->op_mds here will contain fid and mds of slave directory
948 * object (assigned by caller).
950 *mds = op_data->op_mds;
954 * Parent directory is not split and we want to create a
955 * directory in it. Let's calculate where to place it according
956 * to operation data @op_data.
958 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
963 CERROR("Can't choose MDS, err = %d\n", rc);
965 LASSERT(*mds < lmv->desc.ld_tgt_count);
971 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
974 struct lmv_tgt_desc *tgt;
978 tgt = lmv_get_target(lmv, mds);
981 * New seq alloc and FLD setup should be atomic. Otherwise we may find
982 * on server that seq in new allocated fid is not yet known.
984 cfs_down(&tgt->ltd_fid_sem);
986 if (!tgt->ltd_active)
987 GOTO(out, rc = -ENODEV);
990 * Asking underlaying tgt layer to allocate new fid.
992 rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
994 LASSERT(fid_is_sane(fid));
1000 cfs_up(&tgt->ltd_fid_sem);
1004 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1005 struct md_op_data *op_data)
1007 struct obd_device *obd = class_exp2obd(exp);
1008 struct lmv_obd *lmv = &obd->u.lmv;
1013 LASSERT(op_data != NULL);
1014 LASSERT(fid != NULL);
1016 rc = lmv_placement_policy(obd, op_data, &mds);
1018 CERROR("Can't get target for allocating fid, "
1023 rc = __lmv_fid_alloc(lmv, fid, mds);
1025 CERROR("Can't alloc new fid, rc %d\n", rc);
1032 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1035 LASSERT(exp != NULL && fid != NULL);
1036 if (lmv_object_delete(exp, fid)) {
1037 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1043 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1045 struct lmv_obd *lmv = &obd->u.lmv;
1046 struct lprocfs_static_vars lvars;
1047 struct lmv_desc *desc;
1052 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1053 CERROR("LMV setup requires a descriptor\n");
1057 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1058 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1059 CERROR("Lmv descriptor size wrong: %d > %d\n",
1060 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1064 lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1066 OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1067 if (lmv->tgts == NULL)
1070 for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1071 cfs_sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1072 lmv->tgts[i].ltd_idx = i;
1075 lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1077 OBD_ALLOC(lmv->datas, lmv->datas_size);
1078 if (lmv->datas == NULL)
1079 GOTO(out_free_tgts, rc = -ENOMEM);
1081 obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1082 lmv->desc.ld_tgt_count = 0;
1083 lmv->desc.ld_active_tgt_count = 0;
1084 lmv->max_cookiesize = 0;
1085 lmv->max_def_easize = 0;
1086 lmv->max_easize = 0;
1087 lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1089 cfs_spin_lock_init(&lmv->lmv_lock);
1090 cfs_sema_init(&lmv->init_sem, 1);
1092 rc = lmv_object_setup(obd);
1094 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1095 GOTO(out_free_datas, rc);
1098 lprocfs_lmv_init_vars(&lvars);
1099 lprocfs_obd_setup(obd, lvars.obd_vars);
1102 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1103 0444, &lmv_proc_target_fops, obd);
1105 CWARN("Error adding target_obd_stats file (%d)\n", rc);
1108 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1109 LUSTRE_CLI_FLD_HASH_DHT);
1111 CERROR("Can't init FLD, err %d\n", rc);
1112 GOTO(out_free_datas, rc);
1118 OBD_FREE(lmv->datas, lmv->datas_size);
1121 OBD_FREE(lmv->tgts, lmv->tgts_size);
1126 static int lmv_cleanup(struct obd_device *obd)
1128 struct lmv_obd *lmv = &obd->u.lmv;
1131 fld_client_fini(&lmv->lmv_fld);
1132 lprocfs_obd_cleanup(obd);
1133 lmv_object_cleanup(obd);
1134 OBD_FREE(lmv->datas, lmv->datas_size);
1135 OBD_FREE(lmv->tgts, lmv->tgts_size);
1140 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1142 struct lustre_cfg *lcfg = buf;
1143 struct obd_uuid tgt_uuid;
1147 switch(lcfg->lcfg_command) {
1149 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1150 GOTO(out, rc = -EINVAL);
1152 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1153 rc = lmv_add_target(obd, &tgt_uuid);
1156 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1157 GOTO(out, rc = -EINVAL);
1164 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1165 __u64 max_age, __u32 flags)
1167 struct lmv_obd *lmv = &obd->u.lmv;
1168 struct obd_statfs *temp;
1173 rc = lmv_check_connect(obd);
1177 OBD_ALLOC(temp, sizeof(*temp));
1181 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1182 if (lmv->tgts[i].ltd_exp == NULL)
1185 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1188 CERROR("can't stat MDS #%d (%s), error %d\n", i,
1189 lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1191 GOTO(out_free_temp, rc);
1196 osfs->os_bavail += temp->os_bavail;
1197 osfs->os_blocks += temp->os_blocks;
1198 osfs->os_ffree += temp->os_ffree;
1199 osfs->os_files += temp->os_files;
1205 OBD_FREE(temp, sizeof(*temp));
1209 static int lmv_getstatus(struct obd_export *exp,
1211 struct obd_capa **pc)
1213 struct obd_device *obd = exp->exp_obd;
1214 struct lmv_obd *lmv = &obd->u.lmv;
1218 rc = lmv_check_connect(obd);
1222 rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1226 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1227 struct obd_capa *oc, obd_valid valid, const char *name,
1228 const char *input, int input_size, int output_size,
1229 int flags, struct ptlrpc_request **request)
1231 struct obd_device *obd = exp->exp_obd;
1232 struct lmv_obd *lmv = &obd->u.lmv;
1233 struct lmv_tgt_desc *tgt;
1237 rc = lmv_check_connect(obd);
1241 tgt = lmv_find_target(lmv, fid);
1243 RETURN(PTR_ERR(tgt));
1245 rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1246 input_size, output_size, flags, request);
1251 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1252 struct obd_capa *oc, obd_valid valid, const char *name,
1253 const char *input, int input_size, int output_size,
1254 int flags, __u32 suppgid,
1255 struct ptlrpc_request **request)
1257 struct obd_device *obd = exp->exp_obd;
1258 struct lmv_obd *lmv = &obd->u.lmv;
1259 struct lmv_tgt_desc *tgt;
1263 rc = lmv_check_connect(obd);
1267 tgt = lmv_find_target(lmv, fid);
1269 RETURN(PTR_ERR(tgt));
1271 rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1272 input_size, output_size, flags, suppgid,
1278 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1279 struct ptlrpc_request **request)
1281 struct obd_device *obd = exp->exp_obd;
1282 struct lmv_obd *lmv = &obd->u.lmv;
1283 struct lmv_tgt_desc *tgt;
1284 struct lmv_object *obj;
1289 rc = lmv_check_connect(obd);
1293 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1295 RETURN(PTR_ERR(tgt));
1297 if (op_data->op_valid & OBD_MD_MDTIDX) {
1298 op_data->op_mds = tgt->ltd_idx;
1302 rc = md_getattr(tgt->ltd_exp, op_data, request);
1306 obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1308 CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1309 obj ? "(split)" : "");
1312 * If object is split, then we loop over all the slaves and gather size
1313 * attribute. In ideal world we would have to gather also mds field from
1314 * all slaves, as object is spread over the cluster and this is
1315 * definitely interesting information and it is not good to loss it,
1319 struct mdt_body *body;
1321 if (*request == NULL) {
1322 lmv_object_put(obj);
1326 body = req_capsule_server_get(&(*request)->rq_pill,
1328 LASSERT(body != NULL);
1330 for (i = 0; i < obj->lo_objcount; i++) {
1331 if (lmv->tgts[i].ltd_exp == NULL) {
1332 CWARN("%s: NULL export for %d\n",
1338 * Skip master object.
1340 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1343 body->size += obj->lo_stripes[i].ls_size;
1346 lmv_object_put_unlock(obj);
1352 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1353 ldlm_iterator_t it, void *data)
1355 struct obd_device *obd = exp->exp_obd;
1356 struct lmv_obd *lmv = &obd->u.lmv;
1361 rc = lmv_check_connect(obd);
1365 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1368 * With CMD every object can have two locks in different namespaces:
1369 * lookup lock in space of mds storing direntry and update/open lock in
1370 * space of mds storing inode.
1372 for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1373 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1378 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1379 ldlm_iterator_t it, void *data)
1381 struct obd_device *obd = exp->exp_obd;
1382 struct lmv_obd *lmv = &obd->u.lmv;
1387 rc = lmv_check_connect(obd);
1391 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1394 * With CMD every object can have two locks in different namespaces:
1395 * lookup lock in space of mds storing direntry and update/open lock in
1396 * space of mds storing inode.
1398 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1399 rc = md_find_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1408 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1409 struct md_open_data *mod, struct ptlrpc_request **request)
1411 struct obd_device *obd = exp->exp_obd;
1412 struct lmv_obd *lmv = &obd->u.lmv;
1413 struct lmv_tgt_desc *tgt;
1417 rc = lmv_check_connect(obd);
1421 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1423 RETURN(PTR_ERR(tgt));
1425 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1426 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1431 * Called in the case MDS returns -ERESTART on create on open, what means that
1432 * directory is split and its LMV presentation object has to be updated.
1434 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1436 struct obd_device *obd = exp->exp_obd;
1437 struct lmv_obd *lmv = &obd->u.lmv;
1438 struct ptlrpc_request *req = NULL;
1439 struct lmv_tgt_desc *tgt;
1440 struct lmv_object *obj;
1441 struct lustre_md md;
1442 struct md_op_data *op_data;
1449 mealen = lmv_get_easize(lmv);
1451 valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1453 tgt = lmv_find_target(lmv, fid);
1455 RETURN(PTR_ERR(tgt));
1458 * Time to update mea of parent fid.
1461 OBD_ALLOC_PTR(op_data);
1462 if (op_data == NULL)
1465 op_data->op_fid1 = *fid;
1466 op_data->op_mode = mealen;
1467 op_data->op_valid = valid;
1469 rc = md_getattr(tgt->ltd_exp, op_data, &req);
1470 OBD_FREE_PTR(op_data);
1472 CERROR("md_getattr() failed, error %d\n", rc);
1476 rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1478 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1483 GOTO(cleanup, rc = -ENODATA);
1485 obj = lmv_object_create(exp, fid, md.mea);
1489 lmv_object_put(obj);
1491 obd_free_memmd(exp, (void *)&md.mea);
1495 ptlrpc_req_finished(req);
1499 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1500 const void *data, int datalen, int mode, __u32 uid,
1501 __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1502 struct ptlrpc_request **request)
1504 struct obd_device *obd = exp->exp_obd;
1505 struct lmv_obd *lmv = &obd->u.lmv;
1506 struct lmv_tgt_desc *tgt;
1507 struct lmv_object *obj;
1513 rc = lmv_check_connect(obd);
1517 if (!lmv->desc.ld_active_tgt_count)
1523 obj = lmv_object_find(obd, &op_data->op_fid1);
1525 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1526 op_data->op_name, op_data->op_namelen);
1527 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1528 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1529 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1530 tgt = lmv_get_target(lmv, op_data->op_mds);
1531 lmv_object_put(obj);
1533 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1534 op_data->op_bias |= MDS_CHECK_SPLIT;
1535 op_data->op_mds = tgt->ltd_idx;
1539 RETURN(PTR_ERR(tgt));
1541 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1542 if (rc == -ERESTART)
1547 CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1548 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1551 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1552 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1553 cap_effective, rdev, request);
1555 if (*request == NULL)
1557 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1558 } else if (rc == -ERESTART) {
1559 LASSERT(*request != NULL);
1560 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1561 "Got -ERESTART during create!\n");
1562 ptlrpc_req_finished(*request);
1566 * Directory got split. Time to update local object and repeat
1567 * the request with proper MDS.
1569 rc = lmv_handle_split(exp, &op_data->op_fid1);
1571 rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1572 op_data, &op_data->op_fid2);
1581 static int lmv_done_writing(struct obd_export *exp,
1582 struct md_op_data *op_data,
1583 struct md_open_data *mod)
1585 struct obd_device *obd = exp->exp_obd;
1586 struct lmv_obd *lmv = &obd->u.lmv;
1587 struct lmv_tgt_desc *tgt;
1591 rc = lmv_check_connect(obd);
1595 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1597 RETURN(PTR_ERR(tgt));
1599 rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1604 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1605 struct lookup_intent *it, struct md_op_data *op_data,
1606 struct lustre_handle *lockh, void *lmm, int lmmsize)
1608 struct obd_device *obd = exp->exp_obd;
1609 struct lmv_obd *lmv = &obd->u.lmv;
1610 struct lmv_stripe_md *mea = op_data->op_mea1;
1611 struct md_op_data *op_data2;
1612 struct lmv_tgt_desc *tgt;
1617 OBD_ALLOC_PTR(op_data2);
1618 if (op_data2 == NULL)
1621 LASSERT(mea != NULL);
1622 for (i = 0; i < mea->mea_count; i++) {
1623 memset(op_data2, 0, sizeof(*op_data2));
1624 op_data2->op_fid1 = mea->mea_ids[i];
1625 op_data2->op_bias = 0;
1627 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1629 GOTO(cleanup, rc = PTR_ERR(tgt));
1631 if (tgt->ltd_exp == NULL)
1634 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1635 lockh + i, lmm, lmmsize, NULL, 0);
1637 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1638 PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1643 if (it->d.lustre.it_data) {
1644 struct ptlrpc_request *req;
1645 req = (struct ptlrpc_request *)it->d.lustre.it_data;
1646 ptlrpc_req_finished(req);
1649 if (it->d.lustre.it_status)
1650 GOTO(cleanup, rc = it->d.lustre.it_status);
1655 OBD_FREE_PTR(op_data2);
1659 * Drop all taken locks.
1662 if (lockh[i].cookie)
1663 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1664 lockh[i].cookie = 0;
1671 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1672 struct lookup_intent *it, struct md_op_data *op_data,
1673 struct lustre_handle *lockh, void *lmm, int lmmsize,
1674 int extra_lock_flags)
1676 struct ptlrpc_request *req = it->d.lustre.it_data;
1677 struct obd_device *obd = exp->exp_obd;
1678 struct lmv_obd *lmv = &obd->u.lmv;
1679 struct lustre_handle plock;
1680 struct lmv_tgt_desc *tgt;
1681 struct md_op_data *rdata;
1683 struct mdt_body *body;
1688 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1689 LASSERT(body != NULL);
1691 if (!(body->valid & OBD_MD_MDS))
1694 CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1695 LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1698 * We got LOOKUP lock, but we really need attrs.
1700 pmode = it->d.lustre.it_lock_mode;
1701 LASSERT(pmode != 0);
1702 memcpy(&plock, lockh, sizeof(plock));
1703 it->d.lustre.it_lock_mode = 0;
1704 it->d.lustre.it_data = NULL;
1707 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1708 ptlrpc_req_finished(req);
1710 tgt = lmv_find_target(lmv, &fid1);
1712 GOTO(out, rc = PTR_ERR(tgt));
1714 OBD_ALLOC_PTR(rdata);
1716 GOTO(out, rc = -ENOMEM);
1718 rdata->op_fid1 = fid1;
1719 rdata->op_bias = MDS_CROSS_REF;
1721 rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1722 lmm, lmmsize, NULL, extra_lock_flags);
1723 OBD_FREE_PTR(rdata);
1726 ldlm_lock_decref(&plock, pmode);
1731 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1732 struct lookup_intent *it, struct md_op_data *op_data,
1733 struct lustre_handle *lockh, void *lmm, int lmmsize,
1734 struct ptlrpc_request **req, int extra_lock_flags)
1736 struct obd_device *obd = exp->exp_obd;
1737 struct lmv_obd *lmv = &obd->u.lmv;
1738 struct lmv_tgt_desc *tgt;
1739 struct lmv_object *obj;
1744 rc = lmv_check_connect(obd);
1748 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1749 LL_IT2STR(it), PFID(&op_data->op_fid1));
1751 if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1752 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1753 lockh, lmm, lmmsize);
1757 obj = lmv_object_find(obd, &op_data->op_fid1);
1758 if (obj && op_data->op_namelen) {
1759 sidx = raw_name2idx(obj->lo_hashtype,
1761 (char *)op_data->op_name,
1762 op_data->op_namelen);
1763 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1764 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1766 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1769 lmv_object_put(obj);
1772 RETURN(PTR_ERR(tgt));
1774 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1775 LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1777 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1778 lmm, lmmsize, req, extra_lock_flags);
1780 if (rc == 0 && it && it->it_op == IT_OPEN) {
1781 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1782 lmm, lmmsize, extra_lock_flags);
1788 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1789 struct ptlrpc_request **request)
1791 struct ptlrpc_request *req = NULL;
1792 struct obd_device *obd = exp->exp_obd;
1793 struct lmv_obd *lmv = &obd->u.lmv;
1794 struct lu_fid rid = op_data->op_fid1;
1795 struct lmv_tgt_desc *tgt;
1796 struct mdt_body *body;
1797 struct lmv_object *obj;
1798 obd_valid valid = op_data->op_valid;
1804 rc = lmv_check_connect(obd);
1811 obj = lmv_object_find(obd, &rid);
1813 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1814 op_data->op_name, op_data->op_namelen);
1815 rid = obj->lo_stripes[sidx].ls_fid;
1816 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1817 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1818 valid &= ~OBD_MD_FLCKSPLIT;
1819 lmv_object_put(obj);
1821 tgt = lmv_find_target(lmv, &rid);
1822 valid |= OBD_MD_FLCKSPLIT;
1823 op_data->op_mds = tgt->ltd_idx;
1826 RETURN(PTR_ERR(tgt));
1828 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1829 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1830 PFID(&rid), tgt->ltd_idx);
1832 op_data->op_valid = valid;
1833 op_data->op_fid1 = rid;
1834 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1836 body = req_capsule_server_get(&(*request)->rq_pill,
1838 LASSERT(body != NULL);
1840 if (body->valid & OBD_MD_MDS) {
1842 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1845 tgt = lmv_find_target(lmv, &rid);
1847 ptlrpc_req_finished(*request);
1848 RETURN(PTR_ERR(tgt));
1851 op_data->op_fid1 = rid;
1852 op_data->op_valid |= OBD_MD_FLCROSSREF;
1853 op_data->op_namelen = 0;
1854 op_data->op_name = NULL;
1855 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1856 ptlrpc_req_finished(*request);
1859 } else if (rc == -ERESTART) {
1860 LASSERT(*request != NULL);
1861 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1862 "Got -ERESTART during getattr!\n");
1863 ptlrpc_req_finished(*request);
1867 * Directory got split. Time to update local object and repeat
1868 * the request with proper MDS.
1870 rc = lmv_handle_split(exp, &rid);
1877 #define md_op_data_fid(op_data, fl) \
1878 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1879 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1880 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1881 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1884 static int lmv_early_cancel_slaves(struct obd_export *exp,
1885 struct md_op_data *op_data, int op_tgt,
1886 ldlm_mode_t mode, int bits, int flag)
1888 struct obd_device *obd = exp->exp_obd;
1889 struct lmv_obd *lmv = &obd->u.lmv;
1890 ldlm_policy_data_t policy = {{0}};
1891 struct lu_fid *op_fid;
1892 struct lu_fid *st_fid;
1893 struct lmv_tgt_desc *tgt;
1894 struct lmv_object *obj;
1899 op_fid = md_op_data_fid(op_data, flag);
1900 if (!fid_is_sane(op_fid))
1903 obj = lmv_object_find(obd, op_fid);
1907 policy.l_inodebits.bits = bits;
1908 for (i = 0; i < obj->lo_objcount; i++) {
1909 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1910 st_fid = &obj->lo_stripes[i].ls_fid;
1911 if (op_tgt != tgt->ltd_idx) {
1912 CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1913 PFID(st_fid), tgt->ltd_idx);
1914 rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1915 mode, LCF_ASYNC, NULL);
1917 GOTO(out_put_obj, rc);
1920 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1921 op_tgt, PFID(st_fid));
1923 * Do not cancel locks for operation target, they will
1924 * be handled later in underlaying layer when calling
1925 * function we run on behalf of.
1928 op_data->op_flags |= flag;
1933 lmv_object_put(obj);
1937 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1938 int op_tgt, ldlm_mode_t mode, int bits, int flag)
1940 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1941 struct obd_device *obd = exp->exp_obd;
1942 struct lmv_obd *lmv = &obd->u.lmv;
1943 struct lmv_tgt_desc *tgt;
1944 ldlm_policy_data_t policy = {{0}};
1945 struct lmv_object *obj;
1949 if (!fid_is_sane(fid))
1952 obj = lmv_object_find(obd, fid);
1954 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1956 lmv_object_put(obj);
1958 tgt = lmv_find_target(lmv, fid);
1960 RETURN(PTR_ERR(tgt));
1962 if (tgt->ltd_idx != op_tgt) {
1963 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1964 policy.l_inodebits.bits = bits;
1965 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1966 mode, LCF_ASYNC, NULL);
1969 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1971 op_data->op_flags |= flag;
1980 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1983 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1984 struct ptlrpc_request **request)
1986 struct obd_device *obd = exp->exp_obd;
1987 struct lmv_obd *lmv = &obd->u.lmv;
1988 struct lmv_tgt_desc *tgt;
1989 struct lmv_object *obj;
1996 rc = lmv_check_connect(obd);
2003 LASSERT(op_data->op_namelen != 0);
2005 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2006 PFID(&op_data->op_fid2), op_data->op_namelen,
2007 op_data->op_name, PFID(&op_data->op_fid1));
2009 obj = lmv_object_find(obd, &op_data->op_fid2);
2011 sidx = raw_name2idx(obj->lo_hashtype,
2014 op_data->op_namelen);
2015 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2016 mds = obj->lo_stripes[sidx].ls_mds;
2017 lmv_object_put(obj);
2019 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
2024 CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
2025 mds, PFID(&op_data->op_fid1));
2027 op_data->op_fsuid = cfs_curproc_fsuid();
2028 op_data->op_fsgid = cfs_curproc_fsgid();
2029 op_data->op_cap = cfs_curproc_cap_pack();
2030 tgt = lmv_get_target(lmv, mds);
2033 * Cancel UPDATE lock on child (fid1).
2035 op_data->op_flags |= MF_MDC_CANCEL_FID2;
2036 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2037 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2039 rc = md_link(tgt->ltd_exp, op_data, request);
2040 if (rc == -ERESTART) {
2041 LASSERT(*request != NULL);
2042 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2043 "Got -ERESTART during link!\n");
2044 ptlrpc_req_finished(*request);
2048 * Directory got split. Time to update local object and repeat
2049 * the request with proper MDS.
2051 rc = lmv_handle_split(exp, &op_data->op_fid2);
2059 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2060 const char *old, int oldlen, const char *new, int newlen,
2061 struct ptlrpc_request **request)
2063 struct obd_device *obd = exp->exp_obd;
2064 struct lmv_obd *lmv = &obd->u.lmv;
2065 struct lmv_tgt_desc *src_tgt;
2066 struct lmv_tgt_desc *tgt_tgt;
2070 struct lmv_object *obj;
2075 LASSERT(oldlen != 0);
2077 CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2078 oldlen, old, PFID(&op_data->op_fid1),
2079 newlen, new, PFID(&op_data->op_fid2));
2081 rc = lmv_check_connect(obd);
2088 obj = lmv_object_find(obd, &op_data->op_fid1);
2090 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2091 (char *)old, oldlen);
2092 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2093 mds1 = obj->lo_stripes[sidx].ls_mds;
2094 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2095 lmv_object_put(obj);
2097 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2102 obj = lmv_object_find(obd, &op_data->op_fid2);
2105 * Directory is already split, so we have to forward request to
2108 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2109 (char *)new, newlen);
2111 mds2 = obj->lo_stripes[sidx].ls_mds;
2112 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2113 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2114 lmv_object_put(obj);
2116 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2121 op_data->op_fsuid = cfs_curproc_fsuid();
2122 op_data->op_fsgid = cfs_curproc_fsgid();
2123 op_data->op_cap = cfs_curproc_cap_pack();
2125 src_tgt = lmv_get_target(lmv, mds1);
2126 tgt_tgt = lmv_get_target(lmv, mds2);
2129 * LOOKUP lock on src child (fid3) should also be cancelled for
2130 * src_tgt in mdc_rename.
2132 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2135 * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2138 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2139 LCK_EX, MDS_INODELOCK_UPDATE,
2140 MF_MDC_CANCEL_FID2);
2143 * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2146 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2147 LCK_EX, MDS_INODELOCK_LOOKUP,
2148 MF_MDC_CANCEL_FID4);
2152 * Cancel all the locks on tgt child (fid4).
2155 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2156 LCK_EX, MDS_INODELOCK_FULL,
2157 MF_MDC_CANCEL_FID4);
2160 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2161 new, newlen, request);
2163 if (rc == -ERESTART) {
2164 LASSERT(*request != NULL);
2165 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2166 "Got -ERESTART during rename!\n");
2167 ptlrpc_req_finished(*request);
2171 * Directory got split. Time to update local object and repeat
2172 * the request with proper MDS.
2174 rc = lmv_handle_split(exp, &op_data->op_fid1);
2181 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2182 void *ea, int ealen, void *ea2, int ea2len,
2183 struct ptlrpc_request **request,
2184 struct md_open_data **mod)
2186 struct obd_device *obd = exp->exp_obd;
2187 struct lmv_obd *lmv = &obd->u.lmv;
2188 struct ptlrpc_request *req;
2189 struct lmv_tgt_desc *tgt;
2190 struct lmv_object *obj;
2195 rc = lmv_check_connect(obd);
2199 obj = lmv_object_find(obd, &op_data->op_fid1);
2201 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2202 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2203 obj ? ", split" : "");
2205 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2207 for (i = 0; i < obj->lo_objcount; i++) {
2208 op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2210 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2216 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2217 ea2, ea2len, &req, mod);
2219 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2221 * This is master object and this request should
2222 * be returned back to llite.
2226 ptlrpc_req_finished(req);
2232 lmv_object_put(obj);
2234 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2236 RETURN(PTR_ERR(tgt));
2238 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2239 ea2len, request, mod);
2244 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2245 struct obd_capa *oc, struct ptlrpc_request **request)
2247 struct obd_device *obd = exp->exp_obd;
2248 struct lmv_obd *lmv = &obd->u.lmv;
2249 struct lmv_tgt_desc *tgt;
2253 rc = lmv_check_connect(obd);
2257 tgt = lmv_find_target(lmv, fid);
2259 RETURN(PTR_ERR(tgt));
2261 rc = md_sync(tgt->ltd_exp, fid, oc, request);
2266 * Main purpose of LMV blocking ast is to remove split directory LMV
2267 * presentation object (struct lmv_object) attached to the lock being revoked.
2269 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2270 void *data, int flag)
2272 struct lustre_handle lockh;
2273 struct lmv_object *obj;
2278 case LDLM_CB_BLOCKING:
2279 ldlm_lock2handle(lock, &lockh);
2280 rc = ldlm_cli_cancel(&lockh);
2282 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2286 case LDLM_CB_CANCELING:
2288 * Time to drop cached attrs for split directory object
2290 obj = lock->l_ast_data;
2292 CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2293 ", master "DFID"\n",
2294 lock->l_resource->lr_name.name[3] == 1 ?
2295 "LOOKUP" : "UPDATE",
2296 lock->l_resource->lr_name.name[0],
2297 lock->l_resource->lr_name.name[1],
2298 PFID(&obj->lo_fid));
2299 lmv_object_put(obj);
2308 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2312 val = le64_to_cpu(*hash);
2314 val += MAX_HASH_SIZE;
2315 if (val != DIR_END_OFF)
2316 *hash = cpu_to_le64(val - hash_adj);
2319 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2322 struct obd_import *imp;
2325 * XXX: to get nid we assume that underlying obd device is mdc.
2327 imp = class_exp2cliimp(exp);
2328 id = imp->imp_connection->c_self + fid_flatten(fid);
2330 CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2331 imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2333 return id ^ (id >> 32);
2336 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2337 struct obd_capa *oc, __u64 offset64, struct page *page,
2338 struct ptlrpc_request **request)
2340 struct obd_device *obd = exp->exp_obd;
2341 struct lmv_obd *lmv = &obd->u.lmv;
2342 struct lu_fid rid = *fid;
2343 struct lmv_object *obj;
2353 struct lmv_stripe *los;
2354 struct lmv_tgt_desc *tgt;
2355 struct lu_dirpage *dp;
2356 struct lu_dirent *ent;
2361 rc = lmv_check_connect(obd);
2365 CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2368 * This case handle directory lookup in clustered metadata case (i.e.
2369 * split directory is located on multiple md servers.)
2370 * each server keeps directory entries for certain range of hashes.
2371 * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2372 * first server will keep records with hashes [ 0 ... MAX_HASH / N - 1],
2373 * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2375 * readdir can simply start reading entries from 0 - N server in
2376 * order but that will not scale well as all client will request dir in
2377 * to server in same order.
2378 * Following algorithm does optimization:
2379 * Instead of doing readdir in 1, 2, ...., N order, client with a
2380 * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2381 * (every client has rank R)
2382 * But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2383 * since client ask dir from MDS{R} client has pages with offsets
2384 * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2385 * on hash values that we get.
2387 obj = lmv_object_find_lock(obd, fid);
2389 nr = obj->lo_objcount;
2391 seg_size = MAX_HASH_SIZE;
2392 do_div(seg_size, nr);
2393 los = obj->lo_stripes;
2394 tgt = lmv_get_target(lmv, los[0].ls_mds);
2395 rank = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2397 do_div(tgt_tmp, seg_size);
2398 tgt0_idx = do_div(tgt_tmp, nr);
2399 tgt_idx = (tgt0_idx + rank) % nr;
2401 if (tgt_idx < tgt0_idx)
2405 * Last segment has unusual length due to division
2408 hash_adj = MAX_HASH_SIZE - seg_size * nr;
2412 hash_adj += rank * seg_size;
2414 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2415 LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2416 offset, tgt0_idx, offset + hash_adj, tgt_idx);
2418 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2419 rid = obj->lo_stripes[tgt_idx].ls_fid;
2420 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2422 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2423 PFID(&rid), (unsigned long)offset, tgt_idx);
2425 tgt = lmv_find_target(lmv, &rid);
2428 GOTO(cleanup, rc = PTR_ERR(tgt));
2430 rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2434 dp = cfs_kmap(page);
2436 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2437 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2438 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2440 for (ent = lu_dirent_start(dp); ent != NULL;
2441 ent = lu_dirent_next(ent))
2442 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2444 if (tgt0_idx != nr - 1) {
2447 end = le64_to_cpu(dp->ldp_hash_end);
2448 if (end == DIR_END_OFF) {
2449 dp->ldp_hash_end = cpu_to_le32(seg_size *
2452 ""DFID" reset end "LPX64" tgt %d\n",
2454 (__u64)le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2462 lmv_object_put_unlock(obj);
2466 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2467 struct ptlrpc_request **request)
2469 struct obd_device *obd = exp->exp_obd;
2470 struct lmv_obd *lmv = &obd->u.lmv;
2471 struct lmv_tgt_desc *tgt = NULL;
2472 struct lmv_object *obj;
2478 rc = lmv_check_connect(obd);
2485 LASSERT(op_data->op_namelen != 0);
2487 obj = lmv_object_find(obd, &op_data->op_fid1);
2489 sidx = raw_name2idx(obj->lo_hashtype,
2492 op_data->op_namelen);
2493 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2494 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2495 tgt = lmv_get_target(lmv,
2496 obj->lo_stripes[sidx].ls_mds);
2497 lmv_object_put(obj);
2498 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2499 op_data->op_namelen, op_data->op_name,
2500 PFID(&op_data->op_fid1), sidx);
2504 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2506 RETURN(PTR_ERR(tgt));
2507 op_data->op_bias |= MDS_CHECK_SPLIT;
2510 op_data->op_fsuid = cfs_curproc_fsuid();
2511 op_data->op_fsgid = cfs_curproc_fsgid();
2512 op_data->op_cap = cfs_curproc_cap_pack();
2515 * If child's fid is given, cancel unused locks for it if it is from
2516 * another export than parent.
2518 * LOOKUP lock for child (fid3) should also be cancelled on parent
2519 * tgt_tgt in mdc_unlink().
2521 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2524 * Cancel FULL locks on child (fid3).
2526 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2527 MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2530 rc = md_unlink(tgt->ltd_exp, op_data, request);
2532 if (rc == -ERESTART) {
2533 LASSERT(*request != NULL);
2534 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2535 "Got -ERESTART during unlink!\n");
2536 ptlrpc_req_finished(*request);
2540 * Directory got split. Time to update local object and repeat
2541 * the request with proper MDS.
2543 rc = lmv_handle_split(exp, &op_data->op_fid1);
2550 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2555 case OBD_CLEANUP_EARLY:
2556 /* XXX: here should be calling obd_precleanup() down to
2559 case OBD_CLEANUP_EXPORTS:
2560 rc = obd_llog_finish(obd, 0);
2562 CERROR("failed to cleanup llogging subsystems\n");
2570 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2571 void *key, __u32 *vallen, void *val,
2572 struct lov_stripe_md *lsm)
2574 struct obd_device *obd;
2575 struct lmv_obd *lmv;
2579 obd = class_exp2obd(exp);
2581 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2582 exp->exp_handle.h_cookie);
2587 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2588 struct lmv_tgt_desc *tgts;
2591 rc = lmv_check_connect(obd);
2595 LASSERT(*vallen == sizeof(__u32));
2596 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2600 * All tgts should be connected when this gets called.
2602 if (!tgts || !tgts->ltd_exp) {
2603 CERROR("target not setup?\n");
2607 if (!obd_get_info(tgts->ltd_exp, keylen, key,
2612 } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2613 rc = lmv_check_connect(obd);
2618 * Forwarding this request to first MDS, it should know LOV
2621 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2623 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2624 exp->exp_connect_flags =
2625 ((struct obd_connect_data *)val)->ocd_connect_flags;
2630 CDEBUG(D_IOCTL, "Invalid key\n");
2634 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2635 void *key, obd_count vallen, void *val,
2636 struct ptlrpc_request_set *set)
2638 struct lmv_tgt_desc *tgt;
2639 struct obd_device *obd;
2640 struct lmv_obd *lmv;
2644 obd = class_exp2obd(exp);
2646 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2647 exp->exp_handle.h_cookie);
2652 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX)) {
2655 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2656 tgt = &lmv->tgts[i];
2661 err = obd_set_info_async(tgt->ltd_exp,
2662 keylen, key, vallen, val, set);
2673 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2674 struct lov_stripe_md *lsm)
2676 struct obd_device *obd = class_exp2obd(exp);
2677 struct lmv_obd *lmv = &obd->u.lmv;
2678 struct lmv_stripe_md *meap;
2679 struct lmv_stripe_md *lsmp;
2684 mea_size = lmv_get_easize(lmv);
2688 if (*lmmp && !lsm) {
2689 OBD_FREE(*lmmp, mea_size);
2694 if (*lmmp == NULL) {
2695 OBD_ALLOC(*lmmp, mea_size);
2703 lsmp = (struct lmv_stripe_md *)lsm;
2704 meap = (struct lmv_stripe_md *)*lmmp;
2706 if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2707 lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2710 meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2711 meap->mea_count = cpu_to_le32(lsmp->mea_count);
2712 meap->mea_master = cpu_to_le32(lsmp->mea_master);
2714 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2715 meap->mea_ids[i] = meap->mea_ids[i];
2716 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2722 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2723 struct lov_mds_md *lmm, int lmm_size)
2725 struct obd_device *obd = class_exp2obd(exp);
2726 struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2727 struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2728 struct lmv_obd *lmv = &obd->u.lmv;
2734 mea_size = lmv_get_easize(lmv);
2738 if (*lsmp != NULL && lmm == NULL) {
2739 OBD_FREE(*tmea, mea_size);
2744 LASSERT(mea_size == lmm_size);
2746 OBD_ALLOC(*tmea, mea_size);
2753 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2754 mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2755 mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2757 magic = le32_to_cpu(mea->mea_magic);
2760 * Old mea is not handled here.
2762 CERROR("Old not supportable EA is found\n");
2766 (*tmea)->mea_magic = magic;
2767 (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2768 (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2770 for (i = 0; i < (*tmea)->mea_count; i++) {
2771 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2772 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2777 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2778 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2779 ldlm_cancel_flags_t flags, void *opaque)
2781 struct obd_device *obd = exp->exp_obd;
2782 struct lmv_obd *lmv = &obd->u.lmv;
2788 LASSERT(fid != NULL);
2790 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2791 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2794 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2795 policy, mode, flags, opaque);
2802 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2805 struct obd_device *obd = exp->exp_obd;
2806 struct lmv_obd *lmv = &obd->u.lmv;
2810 rc = md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2814 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2815 const struct lu_fid *fid, ldlm_type_t type,
2816 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2817 struct lustre_handle *lockh)
2819 struct obd_device *obd = exp->exp_obd;
2820 struct lmv_obd *lmv = &obd->u.lmv;
2825 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2828 * With CMD every object can have two locks in different namespaces:
2829 * lookup lock in space of mds storing direntry and update/open lock in
2830 * space of mds storing inode. Thus we check all targets, not only that
2831 * one fid was created in.
2833 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2834 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2835 type, policy, mode, lockh);
2843 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2844 struct obd_export *dt_exp, struct obd_export *md_exp,
2845 struct lustre_md *md)
2847 struct obd_device *obd = exp->exp_obd;
2848 struct lmv_obd *lmv = &obd->u.lmv;
2851 rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2855 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2857 struct obd_device *obd = exp->exp_obd;
2858 struct lmv_obd *lmv = &obd->u.lmv;
2862 obd_free_memmd(exp, (void *)&md->mea);
2863 RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2866 int lmv_set_open_replay_data(struct obd_export *exp,
2867 struct obd_client_handle *och,
2868 struct ptlrpc_request *open_req)
2870 struct obd_device *obd = exp->exp_obd;
2871 struct lmv_obd *lmv = &obd->u.lmv;
2872 struct lmv_tgt_desc *tgt;
2875 tgt = lmv_find_target(lmv, &och->och_fid);
2877 RETURN(PTR_ERR(tgt));
2879 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2882 int lmv_clear_open_replay_data(struct obd_export *exp,
2883 struct obd_client_handle *och)
2885 struct obd_device *obd = exp->exp_obd;
2886 struct lmv_obd *lmv = &obd->u.lmv;
2887 struct lmv_tgt_desc *tgt;
2890 tgt = lmv_find_target(lmv, &och->och_fid);
2892 RETURN(PTR_ERR(tgt));
2894 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2897 static int lmv_get_remote_perm(struct obd_export *exp,
2898 const struct lu_fid *fid,
2899 struct obd_capa *oc, __u32 suppgid,
2900 struct ptlrpc_request **request)
2902 struct obd_device *obd = exp->exp_obd;
2903 struct lmv_obd *lmv = &obd->u.lmv;
2904 struct lmv_tgt_desc *tgt;
2908 rc = lmv_check_connect(obd);
2912 tgt = lmv_find_target(lmv, fid);
2914 RETURN(PTR_ERR(tgt));
2916 rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2920 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2923 struct obd_device *obd = exp->exp_obd;
2924 struct lmv_obd *lmv = &obd->u.lmv;
2925 struct lmv_tgt_desc *tgt;
2929 rc = lmv_check_connect(obd);
2933 tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2935 RETURN(PTR_ERR(tgt));
2937 rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2941 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2942 const struct req_msg_field *field, struct obd_capa **oc)
2944 struct obd_device *obd = exp->exp_obd;
2945 struct lmv_obd *lmv = &obd->u.lmv;
2949 rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2953 int lmv_intent_getattr_async(struct obd_export *exp,
2954 struct md_enqueue_info *minfo,
2955 struct ldlm_enqueue_info *einfo)
2957 struct md_op_data *op_data = &minfo->mi_data;
2958 struct obd_device *obd = exp->exp_obd;
2959 struct lmv_obd *lmv = &obd->u.lmv;
2960 struct lmv_object *obj;
2961 struct lmv_tgt_desc *tgt;
2966 rc = lmv_check_connect(obd);
2970 if (!fid_is_sane(&op_data->op_fid2)) {
2971 obj = lmv_object_find(obd, &op_data->op_fid1);
2972 if (obj && op_data->op_namelen) {
2973 sidx = raw_name2idx(obj->lo_hashtype,
2975 (char *)op_data->op_name,
2976 op_data->op_namelen);
2977 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2978 tgt = lmv_get_target(lmv,
2979 obj->lo_stripes[sidx].ls_mds);
2981 "Choose slave dir ("DFID") -> mds #%d\n",
2982 PFID(&op_data->op_fid1), tgt->ltd_idx);
2984 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2987 lmv_object_put(obj);
2989 op_data->op_fid1 = op_data->op_fid2;
2990 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2991 op_data->op_bias = MDS_CROSS_REF;
2993 * Unfortunately, we have to lie to MDC/MDS to retrieve
2994 * attributes llite needs.
2996 if (minfo->mi_it.it_op & IT_LOOKUP)
2997 minfo->mi_it.it_op = IT_GETATTR;
3001 RETURN(PTR_ERR(tgt));
3003 rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
3007 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3008 struct lu_fid *fid, __u32 *bits)
3010 struct obd_device *obd = exp->exp_obd;
3011 struct lmv_obd *lmv = &obd->u.lmv;
3012 struct lmv_tgt_desc *tgt;
3016 rc = lmv_check_connect(obd);
3020 tgt = lmv_find_target(lmv, fid);
3022 RETURN(PTR_ERR(tgt));
3024 rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3029 struct obd_ops lmv_obd_ops = {
3030 .o_owner = THIS_MODULE,
3031 .o_setup = lmv_setup,
3032 .o_cleanup = lmv_cleanup,
3033 .o_precleanup = lmv_precleanup,
3034 .o_process_config = lmv_process_config,
3035 .o_connect = lmv_connect,
3036 .o_disconnect = lmv_disconnect,
3037 .o_statfs = lmv_statfs,
3038 .o_get_info = lmv_get_info,
3039 .o_set_info_async = lmv_set_info_async,
3040 .o_packmd = lmv_packmd,
3041 .o_unpackmd = lmv_unpackmd,
3042 .o_notify = lmv_notify,
3043 .o_get_uuid = lmv_get_uuid,
3044 .o_iocontrol = lmv_iocontrol,
3045 .o_fid_delete = lmv_fid_delete
3048 struct md_ops lmv_md_ops = {
3049 .m_getstatus = lmv_getstatus,
3050 .m_change_cbdata = lmv_change_cbdata,
3051 .m_find_cbdata = lmv_find_cbdata,
3052 .m_close = lmv_close,
3053 .m_create = lmv_create,
3054 .m_done_writing = lmv_done_writing,
3055 .m_enqueue = lmv_enqueue,
3056 .m_getattr = lmv_getattr,
3057 .m_getxattr = lmv_getxattr,
3058 .m_getattr_name = lmv_getattr_name,
3059 .m_intent_lock = lmv_intent_lock,
3061 .m_rename = lmv_rename,
3062 .m_setattr = lmv_setattr,
3063 .m_setxattr = lmv_setxattr,
3065 .m_readpage = lmv_readpage,
3066 .m_unlink = lmv_unlink,
3067 .m_init_ea_size = lmv_init_ea_size,
3068 .m_cancel_unused = lmv_cancel_unused,
3069 .m_set_lock_data = lmv_set_lock_data,
3070 .m_lock_match = lmv_lock_match,
3071 .m_get_lustre_md = lmv_get_lustre_md,
3072 .m_free_lustre_md = lmv_free_lustre_md,
3073 .m_set_open_replay_data = lmv_set_open_replay_data,
3074 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3075 .m_renew_capa = lmv_renew_capa,
3076 .m_unpack_capa = lmv_unpack_capa,
3077 .m_get_remote_perm = lmv_get_remote_perm,
3078 .m_intent_getattr_async = lmv_intent_getattr_async,
3079 .m_revalidate_lock = lmv_revalidate_lock
3082 static quota_interface_t *quota_interface;
3083 extern quota_interface_t lmv_quota_interface;
3085 int __init lmv_init(void)
3087 struct lprocfs_static_vars lvars;
3090 lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3091 sizeof(struct lmv_object),
3093 if (!lmv_object_cache) {
3094 CERROR("Error allocating lmv objects cache\n");
3098 lprocfs_lmv_init_vars(&lvars);
3100 cfs_request_module("lquota");
3101 quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3102 init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3104 rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3105 lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3107 if (quota_interface)
3108 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3109 cfs_mem_cache_destroy(lmv_object_cache);
3116 static void lmv_exit(void)
3118 if (quota_interface)
3119 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3121 class_unregister_type(LUSTRE_LMV_NAME);
3123 LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3124 "Can't free lmv objects cache, %d object(s) busy\n",
3125 cfs_atomic_read(&lmv_object_count));
3126 cfs_mem_cache_destroy(lmv_object_cache);
3129 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3130 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3131 MODULE_LICENSE("GPL");
3133 module_init(lmv_init);
3134 module_exit(lmv_exit);