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/lustre_idl.h>
56 #include <lustre_log.h>
57 #include <obd_support.h>
58 #include <lustre_lib.h>
59 #include <lustre_net.h>
60 #include <obd_class.h>
61 #include <lprocfs_status.h>
62 #include <lustre_lite.h>
63 #include <lustre_fid.h>
64 #include "lmv_internal.h"
67 cfs_mem_cache_t *lmv_object_cache;
68 cfs_atomic_t lmv_object_count = CFS_ATOMIC_INIT(0);
70 static void lmv_activate_target(struct lmv_obd *lmv,
71 struct lmv_tgt_desc *tgt,
74 if (tgt->ltd_active == activate)
77 tgt->ltd_active = activate;
78 lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
84 * -EINVAL : UUID can't be found in the LMV's target list
85 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
86 * -EBADF : The UUID is found, but the OBD of the wrong type (!)
88 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
91 struct lmv_tgt_desc *tgt;
92 struct obd_device *obd;
97 CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
98 lmv, uuid->uuid, activate);
100 cfs_spin_lock(&lmv->lmv_lock);
101 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
102 if (tgt->ltd_exp == NULL)
105 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n",
106 i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
108 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
112 if (i == lmv->desc.ld_tgt_count)
113 GOTO(out_lmv_lock, rc = -EINVAL);
115 obd = class_exp2obd(tgt->ltd_exp);
117 GOTO(out_lmv_lock, rc = -ENOTCONN);
119 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
120 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
121 obd->obd_type->typ_name, i);
122 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
124 if (tgt->ltd_active == activate) {
125 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
126 activate ? "" : "in");
127 GOTO(out_lmv_lock, rc);
130 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
131 activate ? "" : "in");
132 lmv_activate_target(lmv, tgt, activate);
136 cfs_spin_unlock(&lmv->lmv_lock);
140 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
141 struct obd_connect_data *data)
143 struct lmv_tgt_desc *tgt;
147 LASSERT(data != NULL);
149 cfs_spin_lock(&lmv->lmv_lock);
150 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
151 if (tgt->ltd_exp == NULL)
154 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
155 lmv->datas[tgt->ltd_idx] = *data;
159 cfs_spin_unlock(&lmv->lmv_lock);
163 struct obd_uuid *lmv_get_uuid(struct obd_export *exp) {
164 struct obd_device *obd = exp->exp_obd;
165 struct lmv_obd *lmv = &obd->u.lmv;
166 return obd_get_uuid(lmv->tgts[0].ltd_exp);
169 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
170 enum obd_notify_event ev, void *data)
172 struct obd_connect_data *conn_data;
173 struct lmv_obd *lmv = &obd->u.lmv;
174 struct obd_uuid *uuid;
178 if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
179 CERROR("unexpected notification of %s %s!\n",
180 watched->obd_type->typ_name,
185 uuid = &watched->u.cli.cl_target_uuid;
186 if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
188 * Set MDC as active before notifying the observer, so the
189 * observer can use the MDC normally.
191 rc = lmv_set_mdc_active(lmv, uuid,
192 ev == OBD_NOTIFY_ACTIVE);
194 CERROR("%sactivation of %s failed: %d\n",
195 ev == OBD_NOTIFY_ACTIVE ? "" : "de",
199 } else if (ev == OBD_NOTIFY_OCD) {
200 conn_data = &watched->u.cli.cl_import->imp_connect_data;
203 * Set connect data to desired target, update exp_connect_flags.
205 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
207 CERROR("can't set connect data to target %s, rc %d\n",
213 * XXX: Make sure that ocd_connect_flags from all targets are
214 * the same. Otherwise one of MDTs runs wrong version or
215 * something like this. --umka
217 obd->obd_self_export->exp_connect_flags =
218 conn_data->ocd_connect_flags;
221 else if (ev == OBD_NOTIFY_DISCON) {
223 * For disconnect event, flush fld cache for failout MDS case.
225 fld_client_flush(&lmv->lmv_fld);
229 * Pass the notification up the chain.
231 if (obd->obd_observer)
232 rc = obd_notify(obd->obd_observer, watched, ev, data);
238 * This is fake connect function. Its purpose is to initialize lmv and say
239 * caller that everything is okay. Real connection will be performed later.
241 static int lmv_connect(const struct lu_env *env,
242 struct obd_export **exp, struct obd_device *obd,
243 struct obd_uuid *cluuid, struct obd_connect_data *data,
247 struct proc_dir_entry *lmv_proc_dir;
249 struct lmv_obd *lmv = &obd->u.lmv;
250 struct lustre_handle conn = { 0 };
255 * We don't want to actually do the underlying connections more than
256 * once, so keep track.
259 if (lmv->refcount > 1) {
264 rc = class_connect(&conn, obd, cluuid);
266 CERROR("class_connection() returned %d\n", rc);
270 *exp = class_conn2export(&conn);
271 class_export_get(*exp);
275 lmv->cluuid = *cluuid;
278 lmv->conn_data = *data;
281 lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
283 if (IS_ERR(lmv_proc_dir)) {
284 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
285 obd->obd_type->typ_name, obd->obd_name);
291 * All real clients should perform actual connection right away, because
292 * it is possible, that LMV will not have opportunity to connect targets
293 * and MDC stuff will be called directly, for instance while reading
294 * ../mdc/../kbytesfree procfs file, etc.
296 if (data->ocd_connect_flags & OBD_CONNECT_REAL)
297 rc = lmv_check_connect(obd);
302 lprocfs_remove(&lmv_proc_dir);
309 static void lmv_set_timeouts(struct obd_device *obd)
311 struct lmv_tgt_desc *tgts;
316 if (lmv->server_timeout == 0)
319 if (lmv->connected == 0)
322 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
323 if (tgts->ltd_exp == NULL)
326 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
327 KEY_INTERMDS, 0, NULL, NULL);
331 static int lmv_init_ea_size(struct obd_export *exp, int easize,
332 int def_easize, int cookiesize)
334 struct obd_device *obd = exp->exp_obd;
335 struct lmv_obd *lmv = &obd->u.lmv;
341 if (lmv->max_easize < easize) {
342 lmv->max_easize = easize;
345 if (lmv->max_def_easize < def_easize) {
346 lmv->max_def_easize = def_easize;
349 if (lmv->max_cookiesize < cookiesize) {
350 lmv->max_cookiesize = cookiesize;
356 if (lmv->connected == 0)
359 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
360 if (lmv->tgts[i].ltd_exp == NULL) {
361 CWARN("%s: NULL export for %d\n", obd->obd_name, i);
365 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
368 CERROR("obd_init_ea_size() failed on MDT target %d, "
369 "error %d.\n", i, rc);
376 #define MAX_STRING_SIZE 128
378 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
381 struct proc_dir_entry *lmv_proc_dir;
383 struct lmv_obd *lmv = &obd->u.lmv;
384 struct obd_uuid *cluuid = &lmv->cluuid;
385 struct obd_connect_data *mdc_data = NULL;
386 struct obd_uuid lmv_mdc_uuid = { "LMV_MDC_UUID" };
387 struct obd_device *mdc_obd;
388 struct obd_export *mdc_exp;
389 struct lu_fld_target target;
393 mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
396 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
400 CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
401 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
402 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
405 if (!mdc_obd->obd_set_up) {
406 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
410 rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
411 &lmv->conn_data, NULL);
413 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
418 * Init fid sequence client for this mdc and add new fld target.
420 rc = obd_fid_init(mdc_exp);
424 target.ft_srv = NULL;
425 target.ft_exp = mdc_exp;
426 target.ft_idx = tgt->ltd_idx;
428 fld_client_add_target(&lmv->lmv_fld, &target);
430 mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
432 rc = obd_register_observer(mdc_obd, obd);
434 obd_disconnect(mdc_exp);
435 CERROR("target %s register_observer error %d\n",
436 tgt->ltd_uuid.uuid, rc);
440 if (obd->obd_observer) {
442 * Tell the observer about the new target.
444 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
445 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
447 obd_disconnect(mdc_exp);
453 tgt->ltd_exp = mdc_exp;
454 lmv->desc.ld_active_tgt_count++;
457 * Copy connect data, it may be used later.
459 lmv->datas[tgt->ltd_idx] = *mdc_data;
461 md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
462 lmv->max_def_easize, lmv->max_cookiesize);
464 CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
465 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
466 cfs_atomic_read(&obd->obd_refcount));
469 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
471 struct proc_dir_entry *mdc_symlink;
473 LASSERT(mdc_obd->obd_type != NULL);
474 LASSERT(mdc_obd->obd_type->typ_name != NULL);
475 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
478 mdc_obd->obd_type->typ_name,
480 if (mdc_symlink == NULL) {
481 CERROR("Could not register LMV target "
482 "/proc/fs/lustre/%s/%s/target_obds/%s.",
483 obd->obd_type->typ_name, obd->obd_name,
485 lprocfs_remove(&lmv_proc_dir);
493 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
495 struct lmv_obd *lmv = &obd->u.lmv;
496 struct lmv_tgt_desc *tgt;
500 CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
504 if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
505 lmv_init_unlock(lmv);
506 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
507 "That many MDCs already configured.\n",
508 tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
511 if (lmv->desc.ld_tgt_count == 0) {
512 struct obd_device *mdc_obd;
514 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
517 lmv_init_unlock(lmv);
518 CERROR("Target %s not attached\n", tgt_uuid->uuid);
522 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, NULL);
524 lmv_init_unlock(lmv);
525 CERROR("lmv failed to setup llogging subsystems\n");
528 cfs_spin_lock(&lmv->lmv_lock);
529 tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
530 tgt->ltd_uuid = *tgt_uuid;
531 cfs_spin_unlock(&lmv->lmv_lock);
533 if (lmv->connected) {
534 rc = lmv_connect_mdc(obd, tgt);
536 cfs_spin_lock(&lmv->lmv_lock);
537 lmv->desc.ld_tgt_count--;
538 memset(tgt, 0, sizeof(*tgt));
539 cfs_spin_unlock(&lmv->lmv_lock);
541 int easize = sizeof(struct lmv_stripe_md) +
542 lmv->desc.ld_tgt_count *
543 sizeof(struct lu_fid);
544 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
548 lmv_init_unlock(lmv);
552 int lmv_check_connect(struct obd_device *obd)
554 struct lmv_obd *lmv = &obd->u.lmv;
555 struct lmv_tgt_desc *tgt;
565 if (lmv->connected) {
566 lmv_init_unlock(lmv);
570 if (lmv->desc.ld_tgt_count == 0) {
571 CERROR("%s: no targets configured.\n", obd->obd_name);
575 CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
576 lmv->cluuid.uuid, obd->obd_name);
578 LASSERT(lmv->tgts != NULL);
580 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
581 rc = lmv_connect_mdc(obd, tgt);
586 lmv_set_timeouts(obd);
587 class_export_put(lmv->exp);
589 easize = lmv_get_easize(lmv);
590 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
591 lmv_init_unlock(lmv);
600 --lmv->desc.ld_active_tgt_count;
601 rc2 = obd_disconnect(tgt->ltd_exp);
603 CERROR("LMV target %s disconnect on "
604 "MDC idx %d: error %d\n",
605 tgt->ltd_uuid.uuid, i, rc2);
609 class_disconnect(lmv->exp);
610 lmv_init_unlock(lmv);
614 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
617 struct proc_dir_entry *lmv_proc_dir;
619 struct lmv_obd *lmv = &obd->u.lmv;
620 struct obd_device *mdc_obd;
624 LASSERT(tgt != NULL);
625 LASSERT(obd != NULL);
627 mdc_obd = class_exp2obd(tgt->ltd_exp);
630 mdc_obd->obd_no_recov = obd->obd_no_recov;
633 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
635 struct proc_dir_entry *mdc_symlink;
637 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
639 lprocfs_remove(&mdc_symlink);
641 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
642 obd->obd_type->typ_name, obd->obd_name,
647 rc = obd_fid_fini(tgt->ltd_exp);
649 CERROR("Can't finanize fids factory\n");
651 CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
652 tgt->ltd_exp->exp_obd->obd_name,
653 tgt->ltd_exp->exp_obd->obd_uuid.uuid);
655 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
656 rc = obd_disconnect(tgt->ltd_exp);
658 if (tgt->ltd_active) {
659 CERROR("Target %s disconnect error %d\n",
660 tgt->ltd_uuid.uuid, rc);
664 lmv_activate_target(lmv, tgt, 0);
669 static int lmv_disconnect(struct obd_export *exp)
671 struct obd_device *obd = class_exp2obd(exp);
673 struct proc_dir_entry *lmv_proc_dir;
675 struct lmv_obd *lmv = &obd->u.lmv;
684 * Only disconnect the underlying layers on the final disconnect.
687 if (lmv->refcount != 0)
690 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
691 if (lmv->tgts[i].ltd_exp == NULL)
693 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
697 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
699 lprocfs_remove(&lmv_proc_dir);
701 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
702 obd->obd_type->typ_name, obd->obd_name);
708 * This is the case when no real connection is established by
709 * lmv_check_connect().
712 class_export_put(exp);
713 rc = class_disconnect(exp);
714 if (lmv->refcount == 0)
719 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
720 int len, void *karg, void *uarg)
722 struct obd_device *obddev = class_exp2obd(exp);
723 struct lmv_obd *lmv = &obddev->u.lmv;
727 int count = lmv->desc.ld_tgt_count;
734 case IOC_OBD_STATFS: {
735 struct obd_ioctl_data *data = karg;
736 struct obd_device *mdc_obd;
737 struct obd_statfs stat_buf = {0};
740 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
741 if ((index >= count))
744 if (!lmv->tgts[index].ltd_active)
747 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
751 rc = obd_statfs(mdc_obd, &stat_buf,
752 cfs_time_current_64() - CFS_HZ, 0);
755 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
758 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
763 case OBD_IOC_QUOTACTL: {
764 struct if_quotactl *qctl = karg;
765 struct lmv_tgt_desc *tgt = NULL;
766 struct obd_quotactl *oqctl;
768 if (qctl->qc_valid == QC_MDTIDX) {
769 if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
772 tgt = &lmv->tgts[qctl->qc_idx];
775 } else if (qctl->qc_valid == QC_UUID) {
776 for (i = 0; i < count; i++) {
778 if (!obd_uuid_equals(&tgt->ltd_uuid,
782 if (tgt->ltd_exp == NULL)
794 LASSERT(tgt && tgt->ltd_exp);
795 OBD_ALLOC_PTR(oqctl);
799 QCTL_COPY(oqctl, qctl);
800 rc = obd_quotactl(tgt->ltd_exp, oqctl);
802 QCTL_COPY(qctl, oqctl);
803 qctl->qc_valid = QC_MDTIDX;
804 qctl->obd_uuid = tgt->ltd_uuid;
809 case OBD_IOC_CHANGELOG_CLEAR: {
810 struct ioc_changelog_clear *icc = karg;
812 if (icc->icc_mdtindex >= count)
815 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
816 sizeof(*icc), icc, NULL);
821 for (i = 0; i < count; i++) {
824 if (lmv->tgts[i].ltd_exp == NULL)
827 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
829 if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
832 if (lmv->tgts[i].ltd_active) {
833 CERROR("error: iocontrol MDC %s on MDT"
834 "idx %d cmd %x: err = %d\n",
835 lmv->tgts[i].ltd_uuid.uuid,
850 static int lmv_all_chars_policy(int count, const char *name,
861 static int lmv_nid_policy(struct lmv_obd *lmv)
863 struct obd_import *imp;
867 * XXX: To get nid we assume that underlying obd device is mdc.
869 imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
870 id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
871 return id % lmv->desc.ld_tgt_count;
874 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
875 placement_policy_t placement)
878 case PLACEMENT_CHAR_POLICY:
879 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
881 op_data->op_namelen);
882 case PLACEMENT_NID_POLICY:
883 return lmv_nid_policy(lmv);
889 CERROR("Unsupported placement policy %x\n", placement);
894 * This is _inode_ placement policy function (not name).
896 static int lmv_placement_policy(struct obd_device *obd,
897 struct md_op_data *op_data,
900 struct lmv_obd *lmv = &obd->u.lmv;
901 struct lmv_object *obj;
905 LASSERT(mds != NULL);
907 if (lmv->desc.ld_tgt_count == 1) {
913 * Allocate new fid on target according to operation type and parent
916 obj = lmv_object_find(obd, &op_data->op_fid1);
917 if (obj != NULL || op_data->op_name == NULL ||
918 op_data->op_opc != LUSTRE_OPC_MKDIR) {
920 * Allocate fid for non-dir or for null name or for case parent
927 * If we have this flag turned on, and we see that
928 * parent dir is split, this means, that caller did not
929 * notice split yet. This is race and we would like to
930 * let caller know that.
932 if (op_data->op_bias & MDS_CHECK_SPLIT)
937 * Allocate new fid on same mds where parent fid is located and
938 * where operation will be sent. In case of split dir, ->op_fid1
939 * and ->op_mds here will contain fid and mds of slave directory
940 * object (assigned by caller).
942 *mds = op_data->op_mds;
946 * Parent directory is not split and we want to create a
947 * directory in it. Let's calculate where to place it according
948 * to operation data @op_data.
950 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
955 CERROR("Can't choose MDS, err = %d\n", rc);
957 LASSERT(*mds < lmv->desc.ld_tgt_count);
963 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
966 struct lmv_tgt_desc *tgt;
970 tgt = lmv_get_target(lmv, mds);
973 * New seq alloc and FLD setup should be atomic. Otherwise we may find
974 * on server that seq in new allocated fid is not yet known.
976 cfs_down(&tgt->ltd_fid_sem);
978 if (!tgt->ltd_active)
979 GOTO(out, rc = -ENODEV);
982 * Asking underlaying tgt layer to allocate new fid.
984 rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
986 LASSERT(fid_is_sane(fid));
992 cfs_up(&tgt->ltd_fid_sem);
996 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
997 struct md_op_data *op_data)
999 struct obd_device *obd = class_exp2obd(exp);
1000 struct lmv_obd *lmv = &obd->u.lmv;
1005 LASSERT(op_data != NULL);
1006 LASSERT(fid != NULL);
1008 rc = lmv_placement_policy(obd, op_data, &mds);
1010 CERROR("Can't get target for allocating fid, "
1015 rc = __lmv_fid_alloc(lmv, fid, mds);
1017 CERROR("Can't alloc new fid, rc %d\n", rc);
1024 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1027 LASSERT(exp != NULL && fid != NULL);
1028 if (lmv_object_delete(exp, fid)) {
1029 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1035 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1037 struct lmv_obd *lmv = &obd->u.lmv;
1038 struct lprocfs_static_vars lvars;
1039 struct lmv_desc *desc;
1044 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1045 CERROR("LMV setup requires a descriptor\n");
1049 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1050 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1051 CERROR("Lmv descriptor size wrong: %d > %d\n",
1052 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1056 lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1058 OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1059 if (lmv->tgts == NULL)
1062 for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1063 cfs_sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1064 lmv->tgts[i].ltd_idx = i;
1067 lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1069 OBD_ALLOC(lmv->datas, lmv->datas_size);
1070 if (lmv->datas == NULL)
1071 GOTO(out_free_tgts, rc = -ENOMEM);
1073 obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1074 lmv->desc.ld_tgt_count = 0;
1075 lmv->desc.ld_active_tgt_count = 0;
1076 lmv->max_cookiesize = 0;
1077 lmv->max_def_easize = 0;
1078 lmv->max_easize = 0;
1079 lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1081 cfs_spin_lock_init(&lmv->lmv_lock);
1082 cfs_sema_init(&lmv->init_sem, 1);
1084 rc = lmv_object_setup(obd);
1086 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1087 GOTO(out_free_datas, rc);
1090 lprocfs_lmv_init_vars(&lvars);
1091 lprocfs_obd_setup(obd, lvars.obd_vars);
1094 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1095 0444, &lmv_proc_target_fops, obd);
1097 CWARN("Error adding target_obd_stats file (%d)\n", rc);
1100 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1101 LUSTRE_CLI_FLD_HASH_DHT);
1103 CERROR("Can't init FLD, err %d\n", rc);
1104 GOTO(out_free_datas, rc);
1110 OBD_FREE(lmv->datas, lmv->datas_size);
1113 OBD_FREE(lmv->tgts, lmv->tgts_size);
1118 static int lmv_cleanup(struct obd_device *obd)
1120 struct lmv_obd *lmv = &obd->u.lmv;
1123 fld_client_fini(&lmv->lmv_fld);
1124 lprocfs_obd_cleanup(obd);
1125 lmv_object_cleanup(obd);
1126 OBD_FREE(lmv->datas, lmv->datas_size);
1127 OBD_FREE(lmv->tgts, lmv->tgts_size);
1132 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1134 struct lustre_cfg *lcfg = buf;
1135 struct obd_uuid tgt_uuid;
1139 switch(lcfg->lcfg_command) {
1141 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1142 GOTO(out, rc = -EINVAL);
1144 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1145 rc = lmv_add_target(obd, &tgt_uuid);
1148 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1149 GOTO(out, rc = -EINVAL);
1156 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1157 __u64 max_age, __u32 flags)
1159 struct lmv_obd *lmv = &obd->u.lmv;
1160 struct obd_statfs *temp;
1165 rc = lmv_check_connect(obd);
1169 OBD_ALLOC(temp, sizeof(*temp));
1173 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1174 if (lmv->tgts[i].ltd_exp == NULL)
1177 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1180 CERROR("can't stat MDS #%d (%s), error %d\n", i,
1181 lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1183 GOTO(out_free_temp, rc);
1188 osfs->os_bavail += temp->os_bavail;
1189 osfs->os_blocks += temp->os_blocks;
1190 osfs->os_ffree += temp->os_ffree;
1191 osfs->os_files += temp->os_files;
1197 OBD_FREE(temp, sizeof(*temp));
1201 static int lmv_getstatus(struct obd_export *exp,
1203 struct obd_capa **pc)
1205 struct obd_device *obd = exp->exp_obd;
1206 struct lmv_obd *lmv = &obd->u.lmv;
1210 rc = lmv_check_connect(obd);
1214 rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1218 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1219 struct obd_capa *oc, obd_valid valid, const char *name,
1220 const char *input, int input_size, int output_size,
1221 int flags, struct ptlrpc_request **request)
1223 struct obd_device *obd = exp->exp_obd;
1224 struct lmv_obd *lmv = &obd->u.lmv;
1225 struct lmv_tgt_desc *tgt;
1229 rc = lmv_check_connect(obd);
1233 tgt = lmv_find_target(lmv, fid);
1235 RETURN(PTR_ERR(tgt));
1237 rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1238 input_size, output_size, flags, request);
1243 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1244 struct obd_capa *oc, obd_valid valid, const char *name,
1245 const char *input, int input_size, int output_size,
1246 int flags, __u32 suppgid,
1247 struct ptlrpc_request **request)
1249 struct obd_device *obd = exp->exp_obd;
1250 struct lmv_obd *lmv = &obd->u.lmv;
1251 struct lmv_tgt_desc *tgt;
1255 rc = lmv_check_connect(obd);
1259 tgt = lmv_find_target(lmv, fid);
1261 RETURN(PTR_ERR(tgt));
1263 rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1264 input_size, output_size, flags, suppgid,
1270 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1271 struct ptlrpc_request **request)
1273 struct obd_device *obd = exp->exp_obd;
1274 struct lmv_obd *lmv = &obd->u.lmv;
1275 struct lmv_tgt_desc *tgt;
1276 struct lmv_object *obj;
1281 rc = lmv_check_connect(obd);
1285 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1287 RETURN(PTR_ERR(tgt));
1289 if (op_data->op_valid & OBD_MD_MDTIDX) {
1290 op_data->op_mds = tgt->ltd_idx;
1294 rc = md_getattr(tgt->ltd_exp, op_data, request);
1298 obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1300 CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1301 obj ? "(split)" : "");
1304 * If object is split, then we loop over all the slaves and gather size
1305 * attribute. In ideal world we would have to gather also mds field from
1306 * all slaves, as object is spread over the cluster and this is
1307 * definitely interesting information and it is not good to loss it,
1311 struct mdt_body *body;
1313 if (*request == NULL) {
1314 lmv_object_put(obj);
1318 body = req_capsule_server_get(&(*request)->rq_pill,
1320 LASSERT(body != NULL);
1322 for (i = 0; i < obj->lo_objcount; i++) {
1323 if (lmv->tgts[i].ltd_exp == NULL) {
1324 CWARN("%s: NULL export for %d\n",
1330 * Skip master object.
1332 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1335 body->size += obj->lo_stripes[i].ls_size;
1338 lmv_object_put_unlock(obj);
1344 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1345 ldlm_iterator_t it, void *data)
1347 struct obd_device *obd = exp->exp_obd;
1348 struct lmv_obd *lmv = &obd->u.lmv;
1353 rc = lmv_check_connect(obd);
1357 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1360 * With CMD every object can have two locks in different namespaces:
1361 * lookup lock in space of mds storing direntry and update/open lock in
1362 * space of mds storing inode.
1364 for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1365 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1370 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1371 struct md_open_data *mod, struct ptlrpc_request **request)
1373 struct obd_device *obd = exp->exp_obd;
1374 struct lmv_obd *lmv = &obd->u.lmv;
1375 struct lmv_tgt_desc *tgt;
1379 rc = lmv_check_connect(obd);
1383 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1385 RETURN(PTR_ERR(tgt));
1387 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1388 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1393 * Called in the case MDS returns -ERESTART on create on open, what means that
1394 * directory is split and its LMV presentation object has to be updated.
1396 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1398 struct obd_device *obd = exp->exp_obd;
1399 struct lmv_obd *lmv = &obd->u.lmv;
1400 struct ptlrpc_request *req = NULL;
1401 struct lmv_tgt_desc *tgt;
1402 struct lmv_object *obj;
1403 struct lustre_md md;
1404 struct md_op_data *op_data;
1411 mealen = lmv_get_easize(lmv);
1413 valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1415 tgt = lmv_find_target(lmv, fid);
1417 RETURN(PTR_ERR(tgt));
1420 * Time to update mea of parent fid.
1423 OBD_ALLOC_PTR(op_data);
1424 if (op_data == NULL)
1427 op_data->op_fid1 = *fid;
1428 op_data->op_mode = mealen;
1429 op_data->op_valid = valid;
1431 rc = md_getattr(tgt->ltd_exp, op_data, &req);
1432 OBD_FREE_PTR(op_data);
1434 CERROR("md_getattr() failed, error %d\n", rc);
1438 rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1440 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1445 GOTO(cleanup, rc = -ENODATA);
1447 obj = lmv_object_create(exp, fid, md.mea);
1451 lmv_object_put(obj);
1453 obd_free_memmd(exp, (void *)&md.mea);
1457 ptlrpc_req_finished(req);
1461 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1462 const void *data, int datalen, int mode, __u32 uid,
1463 __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1464 struct ptlrpc_request **request)
1466 struct obd_device *obd = exp->exp_obd;
1467 struct lmv_obd *lmv = &obd->u.lmv;
1468 struct lmv_tgt_desc *tgt;
1469 struct lmv_object *obj;
1475 rc = lmv_check_connect(obd);
1479 if (!lmv->desc.ld_active_tgt_count)
1485 obj = lmv_object_find(obd, &op_data->op_fid1);
1487 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1488 op_data->op_name, op_data->op_namelen);
1489 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1490 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1491 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1492 tgt = lmv_get_target(lmv, op_data->op_mds);
1493 lmv_object_put(obj);
1495 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1496 op_data->op_bias |= MDS_CHECK_SPLIT;
1497 op_data->op_mds = tgt->ltd_idx;
1501 RETURN(PTR_ERR(tgt));
1503 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1504 if (rc == -ERESTART)
1509 CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1510 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1513 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1514 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1515 cap_effective, rdev, request);
1517 if (*request == NULL)
1519 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1520 } else if (rc == -ERESTART) {
1521 LASSERT(*request != NULL);
1522 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1523 "Got -ERESTART during create!\n");
1524 ptlrpc_req_finished(*request);
1528 * Directory got split. Time to update local object and repeat
1529 * the request with proper MDS.
1531 rc = lmv_handle_split(exp, &op_data->op_fid1);
1533 rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1534 op_data, &op_data->op_fid2);
1543 static int lmv_done_writing(struct obd_export *exp,
1544 struct md_op_data *op_data,
1545 struct md_open_data *mod)
1547 struct obd_device *obd = exp->exp_obd;
1548 struct lmv_obd *lmv = &obd->u.lmv;
1549 struct lmv_tgt_desc *tgt;
1553 rc = lmv_check_connect(obd);
1557 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1559 RETURN(PTR_ERR(tgt));
1561 rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1566 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1567 struct lookup_intent *it, struct md_op_data *op_data,
1568 struct lustre_handle *lockh, void *lmm, int lmmsize)
1570 struct obd_device *obd = exp->exp_obd;
1571 struct lmv_obd *lmv = &obd->u.lmv;
1572 struct lmv_stripe_md *mea = op_data->op_mea1;
1573 struct md_op_data *op_data2;
1574 struct lmv_tgt_desc *tgt;
1579 OBD_ALLOC_PTR(op_data2);
1580 if (op_data2 == NULL)
1583 LASSERT(mea != NULL);
1584 for (i = 0; i < mea->mea_count; i++) {
1585 memset(op_data2, 0, sizeof(*op_data2));
1586 op_data2->op_fid1 = mea->mea_ids[i];
1587 op_data2->op_bias = 0;
1589 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1591 GOTO(cleanup, rc = PTR_ERR(tgt));
1593 if (tgt->ltd_exp == NULL)
1596 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1597 lockh + i, lmm, lmmsize, NULL, 0);
1599 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1600 PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1605 if (it->d.lustre.it_data) {
1606 struct ptlrpc_request *req;
1607 req = (struct ptlrpc_request *)it->d.lustre.it_data;
1608 ptlrpc_req_finished(req);
1611 if (it->d.lustre.it_status)
1612 GOTO(cleanup, rc = it->d.lustre.it_status);
1617 OBD_FREE_PTR(op_data2);
1621 * Drop all taken locks.
1624 if (lockh[i].cookie)
1625 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1626 lockh[i].cookie = 0;
1633 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1634 struct lookup_intent *it, struct md_op_data *op_data,
1635 struct lustre_handle *lockh, void *lmm, int lmmsize,
1636 int extra_lock_flags)
1638 struct ptlrpc_request *req = it->d.lustre.it_data;
1639 struct obd_device *obd = exp->exp_obd;
1640 struct lmv_obd *lmv = &obd->u.lmv;
1641 struct lustre_handle plock;
1642 struct lmv_tgt_desc *tgt;
1643 struct md_op_data *rdata;
1645 struct mdt_body *body;
1650 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1651 LASSERT(body != NULL);
1653 if (!(body->valid & OBD_MD_MDS))
1656 CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1657 LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1660 * We got LOOKUP lock, but we really need attrs.
1662 pmode = it->d.lustre.it_lock_mode;
1663 LASSERT(pmode != 0);
1664 memcpy(&plock, lockh, sizeof(plock));
1665 it->d.lustre.it_lock_mode = 0;
1666 it->d.lustre.it_data = NULL;
1669 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1670 ptlrpc_req_finished(req);
1672 tgt = lmv_find_target(lmv, &fid1);
1674 GOTO(out, rc = PTR_ERR(tgt));
1676 OBD_ALLOC_PTR(rdata);
1678 GOTO(out, rc = -ENOMEM);
1680 rdata->op_fid1 = fid1;
1681 rdata->op_bias = MDS_CROSS_REF;
1683 rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1684 lmm, lmmsize, NULL, extra_lock_flags);
1685 OBD_FREE_PTR(rdata);
1688 ldlm_lock_decref(&plock, pmode);
1693 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1694 struct lookup_intent *it, struct md_op_data *op_data,
1695 struct lustre_handle *lockh, void *lmm, int lmmsize,
1696 struct ptlrpc_request **req, int extra_lock_flags)
1698 struct obd_device *obd = exp->exp_obd;
1699 struct lmv_obd *lmv = &obd->u.lmv;
1700 struct lmv_tgt_desc *tgt;
1701 struct lmv_object *obj;
1706 rc = lmv_check_connect(obd);
1710 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1711 LL_IT2STR(it), PFID(&op_data->op_fid1));
1713 if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1714 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1715 lockh, lmm, lmmsize);
1719 obj = lmv_object_find(obd, &op_data->op_fid1);
1720 if (obj && op_data->op_namelen) {
1721 sidx = raw_name2idx(obj->lo_hashtype,
1723 (char *)op_data->op_name,
1724 op_data->op_namelen);
1725 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1726 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1728 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1731 lmv_object_put(obj);
1734 RETURN(PTR_ERR(tgt));
1736 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1737 LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1739 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1740 lmm, lmmsize, req, extra_lock_flags);
1742 if (rc == 0 && it && it->it_op == IT_OPEN) {
1743 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1744 lmm, lmmsize, extra_lock_flags);
1750 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1751 struct ptlrpc_request **request)
1753 struct ptlrpc_request *req = NULL;
1754 struct obd_device *obd = exp->exp_obd;
1755 struct lmv_obd *lmv = &obd->u.lmv;
1756 struct lu_fid rid = op_data->op_fid1;
1757 struct lmv_tgt_desc *tgt;
1758 struct mdt_body *body;
1759 struct lmv_object *obj;
1760 obd_valid valid = op_data->op_valid;
1766 rc = lmv_check_connect(obd);
1773 obj = lmv_object_find(obd, &rid);
1775 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1776 op_data->op_name, op_data->op_namelen);
1777 rid = obj->lo_stripes[sidx].ls_fid;
1778 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1779 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1780 valid &= ~OBD_MD_FLCKSPLIT;
1781 lmv_object_put(obj);
1783 tgt = lmv_find_target(lmv, &rid);
1784 valid |= OBD_MD_FLCKSPLIT;
1785 op_data->op_mds = tgt->ltd_idx;
1788 RETURN(PTR_ERR(tgt));
1790 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1791 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1792 PFID(&rid), tgt->ltd_idx);
1794 op_data->op_valid = valid;
1795 op_data->op_fid1 = rid;
1796 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1798 body = req_capsule_server_get(&(*request)->rq_pill,
1800 LASSERT(body != NULL);
1802 if (body->valid & OBD_MD_MDS) {
1804 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1807 tgt = lmv_find_target(lmv, &rid);
1809 ptlrpc_req_finished(*request);
1810 RETURN(PTR_ERR(tgt));
1813 op_data->op_fid1 = rid;
1814 op_data->op_valid |= OBD_MD_FLCROSSREF;
1815 op_data->op_namelen = 0;
1816 op_data->op_name = NULL;
1817 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1818 ptlrpc_req_finished(*request);
1821 } else if (rc == -ERESTART) {
1822 LASSERT(*request != NULL);
1823 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1824 "Got -ERESTART during getattr!\n");
1825 ptlrpc_req_finished(*request);
1829 * Directory got split. Time to update local object and repeat
1830 * the request with proper MDS.
1832 rc = lmv_handle_split(exp, &rid);
1839 #define md_op_data_fid(op_data, fl) \
1840 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1841 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1842 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1843 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1846 static int lmv_early_cancel_slaves(struct obd_export *exp,
1847 struct md_op_data *op_data, int op_tgt,
1848 ldlm_mode_t mode, int bits, int flag)
1850 struct obd_device *obd = exp->exp_obd;
1851 struct lmv_obd *lmv = &obd->u.lmv;
1852 ldlm_policy_data_t policy = {{0}};
1853 struct lu_fid *op_fid;
1854 struct lu_fid *st_fid;
1855 struct lmv_tgt_desc *tgt;
1856 struct lmv_object *obj;
1861 op_fid = md_op_data_fid(op_data, flag);
1862 if (!fid_is_sane(op_fid))
1865 obj = lmv_object_find(obd, op_fid);
1869 policy.l_inodebits.bits = bits;
1870 for (i = 0; i < obj->lo_objcount; i++) {
1871 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1872 st_fid = &obj->lo_stripes[i].ls_fid;
1873 if (op_tgt != tgt->ltd_idx) {
1874 CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1875 PFID(st_fid), tgt->ltd_idx);
1876 rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1877 mode, LDLM_FL_ASYNC, NULL);
1879 GOTO(out_put_obj, rc);
1882 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1883 op_tgt, PFID(st_fid));
1885 * Do not cancel locks for operation target, they will
1886 * be handled later in underlaying layer when calling
1887 * function we run on behalf of.
1890 op_data->op_flags |= flag;
1895 lmv_object_put(obj);
1899 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1900 int op_tgt, ldlm_mode_t mode, int bits, int flag)
1902 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1903 struct obd_device *obd = exp->exp_obd;
1904 struct lmv_obd *lmv = &obd->u.lmv;
1905 struct lmv_tgt_desc *tgt;
1906 ldlm_policy_data_t policy = {{0}};
1907 struct lmv_object *obj;
1911 if (!fid_is_sane(fid))
1914 obj = lmv_object_find(obd, fid);
1916 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1918 lmv_object_put(obj);
1920 tgt = lmv_find_target(lmv, fid);
1922 RETURN(PTR_ERR(tgt));
1924 if (tgt->ltd_idx != op_tgt) {
1925 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1926 policy.l_inodebits.bits = bits;
1927 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1928 mode, LDLM_FL_ASYNC, NULL);
1931 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1933 op_data->op_flags |= flag;
1942 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1945 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1946 struct ptlrpc_request **request)
1948 struct obd_device *obd = exp->exp_obd;
1949 struct lmv_obd *lmv = &obd->u.lmv;
1950 struct lmv_tgt_desc *tgt;
1951 struct lmv_object *obj;
1958 rc = lmv_check_connect(obd);
1965 LASSERT(op_data->op_namelen != 0);
1967 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1968 PFID(&op_data->op_fid2), op_data->op_namelen,
1969 op_data->op_name, PFID(&op_data->op_fid1));
1971 obj = lmv_object_find(obd, &op_data->op_fid2);
1973 sidx = raw_name2idx(obj->lo_hashtype,
1976 op_data->op_namelen);
1977 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
1978 mds = obj->lo_stripes[sidx].ls_mds;
1979 lmv_object_put(obj);
1981 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1986 CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
1987 mds, PFID(&op_data->op_fid1));
1989 op_data->op_fsuid = cfs_curproc_fsuid();
1990 op_data->op_fsgid = cfs_curproc_fsgid();
1991 op_data->op_cap = cfs_curproc_cap_pack();
1992 tgt = lmv_get_target(lmv, mds);
1995 * Cancel UPDATE lock on child (fid1).
1997 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1998 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
1999 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2001 rc = md_link(tgt->ltd_exp, op_data, request);
2002 if (rc == -ERESTART) {
2003 LASSERT(*request != NULL);
2004 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2005 "Got -ERESTART during link!\n");
2006 ptlrpc_req_finished(*request);
2010 * Directory got split. Time to update local object and repeat
2011 * the request with proper MDS.
2013 rc = lmv_handle_split(exp, &op_data->op_fid2);
2021 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2022 const char *old, int oldlen, const char *new, int newlen,
2023 struct ptlrpc_request **request)
2025 struct obd_device *obd = exp->exp_obd;
2026 struct lmv_obd *lmv = &obd->u.lmv;
2027 struct lmv_tgt_desc *src_tgt;
2028 struct lmv_tgt_desc *tgt_tgt;
2032 struct lmv_object *obj;
2037 LASSERT(oldlen != 0);
2039 CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2040 oldlen, old, PFID(&op_data->op_fid1),
2041 newlen, new, PFID(&op_data->op_fid2));
2043 rc = lmv_check_connect(obd);
2050 obj = lmv_object_find(obd, &op_data->op_fid1);
2052 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2053 (char *)old, oldlen);
2054 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2055 mds1 = obj->lo_stripes[sidx].ls_mds;
2056 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2057 lmv_object_put(obj);
2059 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2064 obj = lmv_object_find(obd, &op_data->op_fid2);
2067 * Directory is already split, so we have to forward request to
2070 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2071 (char *)new, newlen);
2073 mds2 = obj->lo_stripes[sidx].ls_mds;
2074 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2075 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2076 lmv_object_put(obj);
2078 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2083 op_data->op_fsuid = cfs_curproc_fsuid();
2084 op_data->op_fsgid = cfs_curproc_fsgid();
2085 op_data->op_cap = cfs_curproc_cap_pack();
2087 src_tgt = lmv_get_target(lmv, mds1);
2088 tgt_tgt = lmv_get_target(lmv, mds2);
2091 * LOOKUP lock on src child (fid3) should also be cancelled for
2092 * src_tgt in mdc_rename.
2094 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2097 * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2100 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2101 LCK_EX, MDS_INODELOCK_UPDATE,
2102 MF_MDC_CANCEL_FID2);
2105 * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2108 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2109 LCK_EX, MDS_INODELOCK_LOOKUP,
2110 MF_MDC_CANCEL_FID4);
2114 * Cancel all the locks on tgt child (fid4).
2117 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2118 LCK_EX, MDS_INODELOCK_FULL,
2119 MF_MDC_CANCEL_FID4);
2122 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2123 new, newlen, request);
2125 if (rc == -ERESTART) {
2126 LASSERT(*request != NULL);
2127 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2128 "Got -ERESTART during rename!\n");
2129 ptlrpc_req_finished(*request);
2133 * Directory got split. Time to update local object and repeat
2134 * the request with proper MDS.
2136 rc = lmv_handle_split(exp, &op_data->op_fid1);
2143 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2144 void *ea, int ealen, void *ea2, int ea2len,
2145 struct ptlrpc_request **request,
2146 struct md_open_data **mod)
2148 struct obd_device *obd = exp->exp_obd;
2149 struct lmv_obd *lmv = &obd->u.lmv;
2150 struct ptlrpc_request *req;
2151 struct lmv_tgt_desc *tgt;
2152 struct lmv_object *obj;
2157 rc = lmv_check_connect(obd);
2161 obj = lmv_object_find(obd, &op_data->op_fid1);
2163 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2164 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2165 obj ? ", split" : "");
2167 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2169 for (i = 0; i < obj->lo_objcount; i++) {
2170 op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2172 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2178 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2179 ea2, ea2len, &req, mod);
2181 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2183 * This is master object and this request should
2184 * be returned back to llite.
2188 ptlrpc_req_finished(req);
2194 lmv_object_put(obj);
2196 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2198 RETURN(PTR_ERR(tgt));
2200 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2201 ea2len, request, mod);
2206 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2207 struct obd_capa *oc, struct ptlrpc_request **request)
2209 struct obd_device *obd = exp->exp_obd;
2210 struct lmv_obd *lmv = &obd->u.lmv;
2211 struct lmv_tgt_desc *tgt;
2215 rc = lmv_check_connect(obd);
2219 tgt = lmv_find_target(lmv, fid);
2221 RETURN(PTR_ERR(tgt));
2223 rc = md_sync(tgt->ltd_exp, fid, oc, request);
2228 * Main purpose of LMV blocking ast is to remove split directory LMV
2229 * presentation object (struct lmv_object) attached to the lock being revoked.
2231 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2232 void *data, int flag)
2234 struct lustre_handle lockh;
2235 struct lmv_object *obj;
2240 case LDLM_CB_BLOCKING:
2241 ldlm_lock2handle(lock, &lockh);
2242 rc = ldlm_cli_cancel(&lockh);
2244 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2248 case LDLM_CB_CANCELING:
2250 * Time to drop cached attrs for split directory object
2252 obj = lock->l_ast_data;
2254 CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2255 ", master "DFID"\n",
2256 lock->l_resource->lr_name.name[3] == 1 ?
2257 "LOOKUP" : "UPDATE",
2258 lock->l_resource->lr_name.name[0],
2259 lock->l_resource->lr_name.name[1],
2260 PFID(&obj->lo_fid));
2261 lmv_object_put(obj);
2270 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2274 val = le64_to_cpu(*hash);
2276 val += MAX_HASH_SIZE;
2277 if (val != DIR_END_OFF)
2278 *hash = cpu_to_le64(val - hash_adj);
2281 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2284 struct obd_import *imp;
2287 * XXX: to get nid we assume that underlying obd device is mdc.
2289 imp = class_exp2cliimp(exp);
2290 id = imp->imp_connection->c_self + fid_flatten(fid);
2292 CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2293 imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2295 return id ^ (id >> 32);
2298 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2299 struct obd_capa *oc, __u64 offset64, struct page *page,
2300 struct ptlrpc_request **request)
2302 struct obd_device *obd = exp->exp_obd;
2303 struct lmv_obd *lmv = &obd->u.lmv;
2304 struct lu_fid rid = *fid;
2305 struct lmv_object *obj;
2315 struct lmv_stripe *los;
2316 struct lmv_tgt_desc *tgt;
2317 struct lu_dirpage *dp;
2318 struct lu_dirent *ent;
2323 rc = lmv_check_connect(obd);
2327 CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2330 * This case handle directory lookup in clustered metadata case (i.e.
2331 * split directory is located on multiple md servers.)
2332 * each server keeps directory entries for certain range of hashes.
2333 * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2334 * first server will keep records with hashes [ 0 ... MAX_HASH / N - 1],
2335 * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2337 * readdir can simply start reading entries from 0 - N server in
2338 * order but that will not scale well as all client will request dir in
2339 * to server in same order.
2340 * Following algorithm does optimization:
2341 * Instead of doing readdir in 1, 2, ...., N order, client with a
2342 * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2343 * (every client has rank R)
2344 * But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2345 * since client ask dir from MDS{R} client has pages with offsets
2346 * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2347 * on hash values that we get.
2349 obj = lmv_object_find_lock(obd, fid);
2351 nr = obj->lo_objcount;
2353 seg_size = MAX_HASH_SIZE;
2354 do_div(seg_size, nr);
2355 los = obj->lo_stripes;
2356 tgt = lmv_get_target(lmv, los[0].ls_mds);
2357 rank = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2359 do_div(tgt_tmp, seg_size);
2360 tgt0_idx = do_div(tgt_tmp, nr);
2361 tgt_idx = (tgt0_idx + rank) % nr;
2363 if (tgt_idx < tgt0_idx)
2367 * Last segment has unusual length due to division
2370 hash_adj = MAX_HASH_SIZE - seg_size * nr;
2374 hash_adj += rank * seg_size;
2376 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2377 LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2378 offset, tgt0_idx, offset + hash_adj, tgt_idx);
2380 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2381 rid = obj->lo_stripes[tgt_idx].ls_fid;
2382 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2384 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2385 PFID(&rid), (unsigned long)offset, tgt_idx);
2387 tgt = lmv_find_target(lmv, &rid);
2390 GOTO(cleanup, rc = PTR_ERR(tgt));
2392 rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2396 dp = cfs_kmap(page);
2398 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2399 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2400 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2402 for (ent = lu_dirent_start(dp); ent != NULL;
2403 ent = lu_dirent_next(ent))
2404 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2406 if (tgt0_idx != nr - 1) {
2409 end = le64_to_cpu(dp->ldp_hash_end);
2410 if (end == DIR_END_OFF) {
2411 dp->ldp_hash_end = cpu_to_le32(seg_size *
2414 ""DFID" reset end "LPX64" tgt %d\n",
2416 le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2424 lmv_object_put_unlock(obj);
2428 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2429 struct ptlrpc_request **request)
2431 struct obd_device *obd = exp->exp_obd;
2432 struct lmv_obd *lmv = &obd->u.lmv;
2433 struct lmv_tgt_desc *tgt = NULL;
2434 struct lmv_object *obj;
2440 rc = lmv_check_connect(obd);
2447 LASSERT(op_data->op_namelen != 0);
2449 obj = lmv_object_find(obd, &op_data->op_fid1);
2451 sidx = raw_name2idx(obj->lo_hashtype,
2454 op_data->op_namelen);
2455 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2456 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2457 tgt = lmv_get_target(lmv,
2458 obj->lo_stripes[sidx].ls_mds);
2459 lmv_object_put(obj);
2460 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2461 op_data->op_namelen, op_data->op_name,
2462 PFID(&op_data->op_fid1), sidx);
2466 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2468 RETURN(PTR_ERR(tgt));
2469 op_data->op_bias |= MDS_CHECK_SPLIT;
2472 op_data->op_fsuid = cfs_curproc_fsuid();
2473 op_data->op_fsgid = cfs_curproc_fsgid();
2474 op_data->op_cap = cfs_curproc_cap_pack();
2477 * If child's fid is given, cancel unused locks for it if it is from
2478 * another export than parent.
2480 * LOOKUP lock for child (fid3) should also be cancelled on parent
2481 * tgt_tgt in mdc_unlink().
2483 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2486 * Cancel FULL locks on child (fid3).
2488 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2489 MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2492 rc = md_unlink(tgt->ltd_exp, op_data, request);
2494 if (rc == -ERESTART) {
2495 LASSERT(*request != NULL);
2496 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2497 "Got -ERESTART during unlink!\n");
2498 ptlrpc_req_finished(*request);
2502 * Directory got split. Time to update local object and repeat
2503 * the request with proper MDS.
2505 rc = lmv_handle_split(exp, &op_data->op_fid1);
2512 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2517 case OBD_CLEANUP_EARLY:
2518 /* XXX: here should be calling obd_precleanup() down to
2521 case OBD_CLEANUP_EXPORTS:
2522 rc = obd_llog_finish(obd, 0);
2524 CERROR("failed to cleanup llogging subsystems\n");
2532 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2533 void *key, __u32 *vallen, void *val,
2534 struct lov_stripe_md *lsm)
2536 struct obd_device *obd;
2537 struct lmv_obd *lmv;
2541 obd = class_exp2obd(exp);
2543 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2544 exp->exp_handle.h_cookie);
2549 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2550 struct lmv_tgt_desc *tgts;
2553 rc = lmv_check_connect(obd);
2557 LASSERT(*vallen == sizeof(__u32));
2558 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2562 * All tgts should be connected when this gets called.
2564 if (!tgts || !tgts->ltd_exp) {
2565 CERROR("target not setup?\n");
2569 if (!obd_get_info(tgts->ltd_exp, keylen, key,
2574 } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2575 rc = lmv_check_connect(obd);
2580 * Forwarding this request to first MDS, it should know LOV
2583 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2585 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2586 exp->exp_connect_flags =
2587 ((struct obd_connect_data *)val)->ocd_connect_flags;
2592 CDEBUG(D_IOCTL, "Invalid key\n");
2596 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2597 void *key, obd_count vallen, void *val,
2598 struct ptlrpc_request_set *set)
2600 struct lmv_tgt_desc *tgt;
2601 struct obd_device *obd;
2602 struct lmv_obd *lmv;
2606 obd = class_exp2obd(exp);
2608 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2609 exp->exp_handle.h_cookie);
2614 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2615 KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2618 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2619 tgt = &lmv->tgts[i];
2624 err = obd_set_info_async(tgt->ltd_exp,
2625 keylen, key, vallen, val, set);
2636 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2637 struct lov_stripe_md *lsm)
2639 struct obd_device *obd = class_exp2obd(exp);
2640 struct lmv_obd *lmv = &obd->u.lmv;
2641 struct lmv_stripe_md *meap;
2642 struct lmv_stripe_md *lsmp;
2647 mea_size = lmv_get_easize(lmv);
2651 if (*lmmp && !lsm) {
2652 OBD_FREE(*lmmp, mea_size);
2657 if (*lmmp == NULL) {
2658 OBD_ALLOC(*lmmp, mea_size);
2666 lsmp = (struct lmv_stripe_md *)lsm;
2667 meap = (struct lmv_stripe_md *)*lmmp;
2669 if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2670 lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2673 meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2674 meap->mea_count = cpu_to_le32(lsmp->mea_count);
2675 meap->mea_master = cpu_to_le32(lsmp->mea_master);
2677 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2678 meap->mea_ids[i] = meap->mea_ids[i];
2679 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2685 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2686 struct lov_mds_md *lmm, int lmm_size)
2688 struct obd_device *obd = class_exp2obd(exp);
2689 struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2690 struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2691 struct lmv_obd *lmv = &obd->u.lmv;
2697 mea_size = lmv_get_easize(lmv);
2701 if (*lsmp != NULL && lmm == NULL) {
2702 OBD_FREE(*tmea, mea_size);
2707 LASSERT(mea_size == lmm_size);
2709 OBD_ALLOC(*tmea, mea_size);
2716 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2717 mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2718 mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2720 magic = le32_to_cpu(mea->mea_magic);
2723 * Old mea is not handled here.
2725 CERROR("Old not supportable EA is found\n");
2729 (*tmea)->mea_magic = magic;
2730 (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2731 (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2733 for (i = 0; i < (*tmea)->mea_count; i++) {
2734 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2735 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2740 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2741 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2742 int flags, void *opaque)
2744 struct obd_device *obd = exp->exp_obd;
2745 struct lmv_obd *lmv = &obd->u.lmv;
2751 LASSERT(fid != NULL);
2753 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2754 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2757 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2758 policy, mode, flags, opaque);
2765 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2768 struct obd_device *obd = exp->exp_obd;
2769 struct lmv_obd *lmv = &obd->u.lmv;
2773 rc = md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2777 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2778 const struct lu_fid *fid, ldlm_type_t type,
2779 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2780 struct lustre_handle *lockh)
2782 struct obd_device *obd = exp->exp_obd;
2783 struct lmv_obd *lmv = &obd->u.lmv;
2788 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2791 * With CMD every object can have two locks in different namespaces:
2792 * lookup lock in space of mds storing direntry and update/open lock in
2793 * space of mds storing inode. Thus we check all targets, not only that
2794 * one fid was created in.
2796 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2797 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2798 type, policy, mode, lockh);
2806 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2807 struct obd_export *dt_exp, struct obd_export *md_exp,
2808 struct lustre_md *md)
2810 struct obd_device *obd = exp->exp_obd;
2811 struct lmv_obd *lmv = &obd->u.lmv;
2814 rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2818 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2820 struct obd_device *obd = exp->exp_obd;
2821 struct lmv_obd *lmv = &obd->u.lmv;
2825 obd_free_memmd(exp, (void *)&md->mea);
2826 RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2829 int lmv_set_open_replay_data(struct obd_export *exp,
2830 struct obd_client_handle *och,
2831 struct ptlrpc_request *open_req)
2833 struct obd_device *obd = exp->exp_obd;
2834 struct lmv_obd *lmv = &obd->u.lmv;
2835 struct lmv_tgt_desc *tgt;
2838 tgt = lmv_find_target(lmv, &och->och_fid);
2840 RETURN(PTR_ERR(tgt));
2842 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2845 int lmv_clear_open_replay_data(struct obd_export *exp,
2846 struct obd_client_handle *och)
2848 struct obd_device *obd = exp->exp_obd;
2849 struct lmv_obd *lmv = &obd->u.lmv;
2850 struct lmv_tgt_desc *tgt;
2853 tgt = lmv_find_target(lmv, &och->och_fid);
2855 RETURN(PTR_ERR(tgt));
2857 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2860 static int lmv_get_remote_perm(struct obd_export *exp,
2861 const struct lu_fid *fid,
2862 struct obd_capa *oc, __u32 suppgid,
2863 struct ptlrpc_request **request)
2865 struct obd_device *obd = exp->exp_obd;
2866 struct lmv_obd *lmv = &obd->u.lmv;
2867 struct lmv_tgt_desc *tgt;
2871 rc = lmv_check_connect(obd);
2875 tgt = lmv_find_target(lmv, fid);
2877 RETURN(PTR_ERR(tgt));
2879 rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2883 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2886 struct obd_device *obd = exp->exp_obd;
2887 struct lmv_obd *lmv = &obd->u.lmv;
2888 struct lmv_tgt_desc *tgt;
2892 rc = lmv_check_connect(obd);
2896 tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2898 RETURN(PTR_ERR(tgt));
2900 rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2904 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2905 const struct req_msg_field *field, struct obd_capa **oc)
2907 struct obd_device *obd = exp->exp_obd;
2908 struct lmv_obd *lmv = &obd->u.lmv;
2912 rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2916 int lmv_intent_getattr_async(struct obd_export *exp,
2917 struct md_enqueue_info *minfo,
2918 struct ldlm_enqueue_info *einfo)
2920 struct md_op_data *op_data = &minfo->mi_data;
2921 struct obd_device *obd = exp->exp_obd;
2922 struct lmv_obd *lmv = &obd->u.lmv;
2923 struct lmv_object *obj;
2924 struct lmv_tgt_desc *tgt;
2929 rc = lmv_check_connect(obd);
2933 if (!fid_is_sane(&op_data->op_fid2)) {
2934 obj = lmv_object_find(obd, &op_data->op_fid1);
2935 if (obj && op_data->op_namelen) {
2936 sidx = raw_name2idx(obj->lo_hashtype,
2938 (char *)op_data->op_name,
2939 op_data->op_namelen);
2940 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2941 tgt = lmv_get_target(lmv,
2942 obj->lo_stripes[sidx].ls_mds);
2944 "Choose slave dir ("DFID") -> mds #%d\n",
2945 PFID(&op_data->op_fid1), tgt->ltd_idx);
2947 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2950 lmv_object_put(obj);
2952 op_data->op_fid1 = op_data->op_fid2;
2953 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2954 op_data->op_bias = MDS_CROSS_REF;
2956 * Unfortunately, we have to lie to MDC/MDS to retrieve
2957 * attributes llite needs.
2959 if (minfo->mi_it.it_op & IT_LOOKUP)
2960 minfo->mi_it.it_op = IT_GETATTR;
2964 RETURN(PTR_ERR(tgt));
2966 rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2970 int lmv_revalidate_lock(struct obd_export *exp,
2971 struct lookup_intent *it,
2974 struct obd_device *obd = exp->exp_obd;
2975 struct lmv_obd *lmv = &obd->u.lmv;
2976 struct lmv_tgt_desc *tgt;
2980 rc = lmv_check_connect(obd);
2984 tgt = lmv_find_target(lmv, fid);
2986 RETURN(PTR_ERR(tgt));
2988 rc = md_revalidate_lock(tgt->ltd_exp, it, fid);
2993 struct obd_ops lmv_obd_ops = {
2994 .o_owner = THIS_MODULE,
2995 .o_setup = lmv_setup,
2996 .o_cleanup = lmv_cleanup,
2997 .o_precleanup = lmv_precleanup,
2998 .o_process_config = lmv_process_config,
2999 .o_connect = lmv_connect,
3000 .o_disconnect = lmv_disconnect,
3001 .o_statfs = lmv_statfs,
3002 .o_get_info = lmv_get_info,
3003 .o_set_info_async = lmv_set_info_async,
3004 .o_packmd = lmv_packmd,
3005 .o_unpackmd = lmv_unpackmd,
3006 .o_notify = lmv_notify,
3007 .o_get_uuid = lmv_get_uuid,
3008 .o_iocontrol = lmv_iocontrol,
3009 .o_fid_delete = lmv_fid_delete
3012 struct md_ops lmv_md_ops = {
3013 .m_getstatus = lmv_getstatus,
3014 .m_change_cbdata = lmv_change_cbdata,
3015 .m_close = lmv_close,
3016 .m_create = lmv_create,
3017 .m_done_writing = lmv_done_writing,
3018 .m_enqueue = lmv_enqueue,
3019 .m_getattr = lmv_getattr,
3020 .m_getxattr = lmv_getxattr,
3021 .m_getattr_name = lmv_getattr_name,
3022 .m_intent_lock = lmv_intent_lock,
3024 .m_rename = lmv_rename,
3025 .m_setattr = lmv_setattr,
3026 .m_setxattr = lmv_setxattr,
3028 .m_readpage = lmv_readpage,
3029 .m_unlink = lmv_unlink,
3030 .m_init_ea_size = lmv_init_ea_size,
3031 .m_cancel_unused = lmv_cancel_unused,
3032 .m_set_lock_data = lmv_set_lock_data,
3033 .m_lock_match = lmv_lock_match,
3034 .m_get_lustre_md = lmv_get_lustre_md,
3035 .m_free_lustre_md = lmv_free_lustre_md,
3036 .m_set_open_replay_data = lmv_set_open_replay_data,
3037 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3038 .m_renew_capa = lmv_renew_capa,
3039 .m_unpack_capa = lmv_unpack_capa,
3040 .m_get_remote_perm = lmv_get_remote_perm,
3041 .m_intent_getattr_async = lmv_intent_getattr_async,
3042 .m_revalidate_lock = lmv_revalidate_lock
3045 static quota_interface_t *quota_interface;
3046 extern quota_interface_t lmv_quota_interface;
3048 int __init lmv_init(void)
3050 struct lprocfs_static_vars lvars;
3053 lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3054 sizeof(struct lmv_object),
3056 if (!lmv_object_cache) {
3057 CERROR("Error allocating lmv objects cache\n");
3061 lprocfs_lmv_init_vars(&lvars);
3063 cfs_request_module("lquota");
3064 quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3065 init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3067 rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3068 lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3070 if (quota_interface)
3071 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3072 cfs_mem_cache_destroy(lmv_object_cache);
3079 static void lmv_exit(void)
3081 if (quota_interface)
3082 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3084 class_unregister_type(LUSTRE_LMV_NAME);
3086 LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3087 "Can't free lmv objects cache, %d object(s) busy\n",
3088 cfs_atomic_read(&lmv_object_count));
3089 cfs_mem_cache_destroy(lmv_object_cache);
3092 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3093 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3094 MODULE_LICENSE("GPL");
3096 module_init(lmv_init);
3097 module_exit(lmv_exit);