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_SEND:
810 case OBD_IOC_CHANGELOG_CLEAR: {
811 struct ioc_changelog *icc = karg;
813 if (icc->icc_mdtindex >= count)
816 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
817 sizeof(*icc), icc, NULL);
822 for (i = 0; i < count; i++) {
825 if (lmv->tgts[i].ltd_exp == NULL)
828 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
830 if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
833 if (lmv->tgts[i].ltd_active) {
834 CERROR("error: iocontrol MDC %s on MDT"
835 "idx %d cmd %x: err = %d\n",
836 lmv->tgts[i].ltd_uuid.uuid,
851 static int lmv_all_chars_policy(int count, const char *name,
862 static int lmv_nid_policy(struct lmv_obd *lmv)
864 struct obd_import *imp;
868 * XXX: To get nid we assume that underlying obd device is mdc.
870 imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
871 id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
872 return id % lmv->desc.ld_tgt_count;
875 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
876 placement_policy_t placement)
879 case PLACEMENT_CHAR_POLICY:
880 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
882 op_data->op_namelen);
883 case PLACEMENT_NID_POLICY:
884 return lmv_nid_policy(lmv);
890 CERROR("Unsupported placement policy %x\n", placement);
895 * This is _inode_ placement policy function (not name).
897 static int lmv_placement_policy(struct obd_device *obd,
898 struct md_op_data *op_data,
901 struct lmv_obd *lmv = &obd->u.lmv;
902 struct lmv_object *obj;
906 LASSERT(mds != NULL);
908 if (lmv->desc.ld_tgt_count == 1) {
914 * Allocate new fid on target according to operation type and parent
917 obj = lmv_object_find(obd, &op_data->op_fid1);
918 if (obj != NULL || op_data->op_name == NULL ||
919 op_data->op_opc != LUSTRE_OPC_MKDIR) {
921 * Allocate fid for non-dir or for null name or for case parent
928 * If we have this flag turned on, and we see that
929 * parent dir is split, this means, that caller did not
930 * notice split yet. This is race and we would like to
931 * let caller know that.
933 if (op_data->op_bias & MDS_CHECK_SPLIT)
938 * Allocate new fid on same mds where parent fid is located and
939 * where operation will be sent. In case of split dir, ->op_fid1
940 * and ->op_mds here will contain fid and mds of slave directory
941 * object (assigned by caller).
943 *mds = op_data->op_mds;
947 * Parent directory is not split and we want to create a
948 * directory in it. Let's calculate where to place it according
949 * to operation data @op_data.
951 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
956 CERROR("Can't choose MDS, err = %d\n", rc);
958 LASSERT(*mds < lmv->desc.ld_tgt_count);
964 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
967 struct lmv_tgt_desc *tgt;
971 tgt = lmv_get_target(lmv, mds);
974 * New seq alloc and FLD setup should be atomic. Otherwise we may find
975 * on server that seq in new allocated fid is not yet known.
977 cfs_down(&tgt->ltd_fid_sem);
979 if (!tgt->ltd_active)
980 GOTO(out, rc = -ENODEV);
983 * Asking underlaying tgt layer to allocate new fid.
985 rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
987 LASSERT(fid_is_sane(fid));
993 cfs_up(&tgt->ltd_fid_sem);
997 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
998 struct md_op_data *op_data)
1000 struct obd_device *obd = class_exp2obd(exp);
1001 struct lmv_obd *lmv = &obd->u.lmv;
1006 LASSERT(op_data != NULL);
1007 LASSERT(fid != NULL);
1009 rc = lmv_placement_policy(obd, op_data, &mds);
1011 CERROR("Can't get target for allocating fid, "
1016 rc = __lmv_fid_alloc(lmv, fid, mds);
1018 CERROR("Can't alloc new fid, rc %d\n", rc);
1025 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1028 LASSERT(exp != NULL && fid != NULL);
1029 if (lmv_object_delete(exp, fid)) {
1030 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1036 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1038 struct lmv_obd *lmv = &obd->u.lmv;
1039 struct lprocfs_static_vars lvars;
1040 struct lmv_desc *desc;
1045 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1046 CERROR("LMV setup requires a descriptor\n");
1050 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1051 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1052 CERROR("Lmv descriptor size wrong: %d > %d\n",
1053 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1057 lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1059 OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1060 if (lmv->tgts == NULL)
1063 for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1064 cfs_sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1065 lmv->tgts[i].ltd_idx = i;
1068 lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1070 OBD_ALLOC(lmv->datas, lmv->datas_size);
1071 if (lmv->datas == NULL)
1072 GOTO(out_free_tgts, rc = -ENOMEM);
1074 obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1075 lmv->desc.ld_tgt_count = 0;
1076 lmv->desc.ld_active_tgt_count = 0;
1077 lmv->max_cookiesize = 0;
1078 lmv->max_def_easize = 0;
1079 lmv->max_easize = 0;
1080 lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1082 cfs_spin_lock_init(&lmv->lmv_lock);
1083 cfs_sema_init(&lmv->init_sem, 1);
1085 rc = lmv_object_setup(obd);
1087 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1088 GOTO(out_free_datas, rc);
1091 lprocfs_lmv_init_vars(&lvars);
1092 lprocfs_obd_setup(obd, lvars.obd_vars);
1095 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1096 0444, &lmv_proc_target_fops, obd);
1098 CWARN("Error adding target_obd_stats file (%d)\n", rc);
1101 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1102 LUSTRE_CLI_FLD_HASH_DHT);
1104 CERROR("Can't init FLD, err %d\n", rc);
1105 GOTO(out_free_datas, rc);
1111 OBD_FREE(lmv->datas, lmv->datas_size);
1114 OBD_FREE(lmv->tgts, lmv->tgts_size);
1119 static int lmv_cleanup(struct obd_device *obd)
1121 struct lmv_obd *lmv = &obd->u.lmv;
1124 fld_client_fini(&lmv->lmv_fld);
1125 lprocfs_obd_cleanup(obd);
1126 lmv_object_cleanup(obd);
1127 OBD_FREE(lmv->datas, lmv->datas_size);
1128 OBD_FREE(lmv->tgts, lmv->tgts_size);
1133 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1135 struct lustre_cfg *lcfg = buf;
1136 struct obd_uuid tgt_uuid;
1140 switch(lcfg->lcfg_command) {
1142 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1143 GOTO(out, rc = -EINVAL);
1145 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1146 rc = lmv_add_target(obd, &tgt_uuid);
1149 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1150 GOTO(out, rc = -EINVAL);
1157 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1158 __u64 max_age, __u32 flags)
1160 struct lmv_obd *lmv = &obd->u.lmv;
1161 struct obd_statfs *temp;
1166 rc = lmv_check_connect(obd);
1170 OBD_ALLOC(temp, sizeof(*temp));
1174 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1175 if (lmv->tgts[i].ltd_exp == NULL)
1178 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1181 CERROR("can't stat MDS #%d (%s), error %d\n", i,
1182 lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1184 GOTO(out_free_temp, rc);
1189 osfs->os_bavail += temp->os_bavail;
1190 osfs->os_blocks += temp->os_blocks;
1191 osfs->os_ffree += temp->os_ffree;
1192 osfs->os_files += temp->os_files;
1198 OBD_FREE(temp, sizeof(*temp));
1202 static int lmv_getstatus(struct obd_export *exp,
1204 struct obd_capa **pc)
1206 struct obd_device *obd = exp->exp_obd;
1207 struct lmv_obd *lmv = &obd->u.lmv;
1211 rc = lmv_check_connect(obd);
1215 rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1219 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1220 struct obd_capa *oc, obd_valid valid, const char *name,
1221 const char *input, int input_size, int output_size,
1222 int flags, struct ptlrpc_request **request)
1224 struct obd_device *obd = exp->exp_obd;
1225 struct lmv_obd *lmv = &obd->u.lmv;
1226 struct lmv_tgt_desc *tgt;
1230 rc = lmv_check_connect(obd);
1234 tgt = lmv_find_target(lmv, fid);
1236 RETURN(PTR_ERR(tgt));
1238 rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1239 input_size, output_size, flags, request);
1244 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1245 struct obd_capa *oc, obd_valid valid, const char *name,
1246 const char *input, int input_size, int output_size,
1247 int flags, __u32 suppgid,
1248 struct ptlrpc_request **request)
1250 struct obd_device *obd = exp->exp_obd;
1251 struct lmv_obd *lmv = &obd->u.lmv;
1252 struct lmv_tgt_desc *tgt;
1256 rc = lmv_check_connect(obd);
1260 tgt = lmv_find_target(lmv, fid);
1262 RETURN(PTR_ERR(tgt));
1264 rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1265 input_size, output_size, flags, suppgid,
1271 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1272 struct ptlrpc_request **request)
1274 struct obd_device *obd = exp->exp_obd;
1275 struct lmv_obd *lmv = &obd->u.lmv;
1276 struct lmv_tgt_desc *tgt;
1277 struct lmv_object *obj;
1282 rc = lmv_check_connect(obd);
1286 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1288 RETURN(PTR_ERR(tgt));
1290 if (op_data->op_valid & OBD_MD_MDTIDX) {
1291 op_data->op_mds = tgt->ltd_idx;
1295 rc = md_getattr(tgt->ltd_exp, op_data, request);
1299 obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1301 CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1302 obj ? "(split)" : "");
1305 * If object is split, then we loop over all the slaves and gather size
1306 * attribute. In ideal world we would have to gather also mds field from
1307 * all slaves, as object is spread over the cluster and this is
1308 * definitely interesting information and it is not good to loss it,
1312 struct mdt_body *body;
1314 if (*request == NULL) {
1315 lmv_object_put(obj);
1319 body = req_capsule_server_get(&(*request)->rq_pill,
1321 LASSERT(body != NULL);
1323 for (i = 0; i < obj->lo_objcount; i++) {
1324 if (lmv->tgts[i].ltd_exp == NULL) {
1325 CWARN("%s: NULL export for %d\n",
1331 * Skip master object.
1333 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1336 body->size += obj->lo_stripes[i].ls_size;
1339 lmv_object_put_unlock(obj);
1345 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1346 ldlm_iterator_t it, void *data)
1348 struct obd_device *obd = exp->exp_obd;
1349 struct lmv_obd *lmv = &obd->u.lmv;
1354 rc = lmv_check_connect(obd);
1358 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1361 * With CMD every object can have two locks in different namespaces:
1362 * lookup lock in space of mds storing direntry and update/open lock in
1363 * space of mds storing inode.
1365 for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1366 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1371 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1372 struct md_open_data *mod, struct ptlrpc_request **request)
1374 struct obd_device *obd = exp->exp_obd;
1375 struct lmv_obd *lmv = &obd->u.lmv;
1376 struct lmv_tgt_desc *tgt;
1380 rc = lmv_check_connect(obd);
1384 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1386 RETURN(PTR_ERR(tgt));
1388 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1389 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1394 * Called in the case MDS returns -ERESTART on create on open, what means that
1395 * directory is split and its LMV presentation object has to be updated.
1397 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1399 struct obd_device *obd = exp->exp_obd;
1400 struct lmv_obd *lmv = &obd->u.lmv;
1401 struct ptlrpc_request *req = NULL;
1402 struct lmv_tgt_desc *tgt;
1403 struct lmv_object *obj;
1404 struct lustre_md md;
1405 struct md_op_data *op_data;
1412 mealen = lmv_get_easize(lmv);
1414 valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1416 tgt = lmv_find_target(lmv, fid);
1418 RETURN(PTR_ERR(tgt));
1421 * Time to update mea of parent fid.
1424 OBD_ALLOC_PTR(op_data);
1425 if (op_data == NULL)
1428 op_data->op_fid1 = *fid;
1429 op_data->op_mode = mealen;
1430 op_data->op_valid = valid;
1432 rc = md_getattr(tgt->ltd_exp, op_data, &req);
1433 OBD_FREE_PTR(op_data);
1435 CERROR("md_getattr() failed, error %d\n", rc);
1439 rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1441 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1446 GOTO(cleanup, rc = -ENODATA);
1448 obj = lmv_object_create(exp, fid, md.mea);
1452 lmv_object_put(obj);
1454 obd_free_memmd(exp, (void *)&md.mea);
1458 ptlrpc_req_finished(req);
1462 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1463 const void *data, int datalen, int mode, __u32 uid,
1464 __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1465 struct ptlrpc_request **request)
1467 struct obd_device *obd = exp->exp_obd;
1468 struct lmv_obd *lmv = &obd->u.lmv;
1469 struct lmv_tgt_desc *tgt;
1470 struct lmv_object *obj;
1476 rc = lmv_check_connect(obd);
1480 if (!lmv->desc.ld_active_tgt_count)
1486 obj = lmv_object_find(obd, &op_data->op_fid1);
1488 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1489 op_data->op_name, op_data->op_namelen);
1490 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1491 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1492 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1493 tgt = lmv_get_target(lmv, op_data->op_mds);
1494 lmv_object_put(obj);
1496 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1497 op_data->op_bias |= MDS_CHECK_SPLIT;
1498 op_data->op_mds = tgt->ltd_idx;
1502 RETURN(PTR_ERR(tgt));
1504 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1505 if (rc == -ERESTART)
1510 CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1511 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1514 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1515 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1516 cap_effective, rdev, request);
1518 if (*request == NULL)
1520 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1521 } else if (rc == -ERESTART) {
1522 LASSERT(*request != NULL);
1523 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1524 "Got -ERESTART during create!\n");
1525 ptlrpc_req_finished(*request);
1529 * Directory got split. Time to update local object and repeat
1530 * the request with proper MDS.
1532 rc = lmv_handle_split(exp, &op_data->op_fid1);
1534 rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1535 op_data, &op_data->op_fid2);
1544 static int lmv_done_writing(struct obd_export *exp,
1545 struct md_op_data *op_data,
1546 struct md_open_data *mod)
1548 struct obd_device *obd = exp->exp_obd;
1549 struct lmv_obd *lmv = &obd->u.lmv;
1550 struct lmv_tgt_desc *tgt;
1554 rc = lmv_check_connect(obd);
1558 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1560 RETURN(PTR_ERR(tgt));
1562 rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1567 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1568 struct lookup_intent *it, struct md_op_data *op_data,
1569 struct lustre_handle *lockh, void *lmm, int lmmsize)
1571 struct obd_device *obd = exp->exp_obd;
1572 struct lmv_obd *lmv = &obd->u.lmv;
1573 struct lmv_stripe_md *mea = op_data->op_mea1;
1574 struct md_op_data *op_data2;
1575 struct lmv_tgt_desc *tgt;
1580 OBD_ALLOC_PTR(op_data2);
1581 if (op_data2 == NULL)
1584 LASSERT(mea != NULL);
1585 for (i = 0; i < mea->mea_count; i++) {
1586 memset(op_data2, 0, sizeof(*op_data2));
1587 op_data2->op_fid1 = mea->mea_ids[i];
1588 op_data2->op_bias = 0;
1590 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1592 GOTO(cleanup, rc = PTR_ERR(tgt));
1594 if (tgt->ltd_exp == NULL)
1597 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1598 lockh + i, lmm, lmmsize, NULL, 0);
1600 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1601 PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1606 if (it->d.lustre.it_data) {
1607 struct ptlrpc_request *req;
1608 req = (struct ptlrpc_request *)it->d.lustre.it_data;
1609 ptlrpc_req_finished(req);
1612 if (it->d.lustre.it_status)
1613 GOTO(cleanup, rc = it->d.lustre.it_status);
1618 OBD_FREE_PTR(op_data2);
1622 * Drop all taken locks.
1625 if (lockh[i].cookie)
1626 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1627 lockh[i].cookie = 0;
1634 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1635 struct lookup_intent *it, struct md_op_data *op_data,
1636 struct lustre_handle *lockh, void *lmm, int lmmsize,
1637 int extra_lock_flags)
1639 struct ptlrpc_request *req = it->d.lustre.it_data;
1640 struct obd_device *obd = exp->exp_obd;
1641 struct lmv_obd *lmv = &obd->u.lmv;
1642 struct lustre_handle plock;
1643 struct lmv_tgt_desc *tgt;
1644 struct md_op_data *rdata;
1646 struct mdt_body *body;
1651 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1652 LASSERT(body != NULL);
1654 if (!(body->valid & OBD_MD_MDS))
1657 CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1658 LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1661 * We got LOOKUP lock, but we really need attrs.
1663 pmode = it->d.lustre.it_lock_mode;
1664 LASSERT(pmode != 0);
1665 memcpy(&plock, lockh, sizeof(plock));
1666 it->d.lustre.it_lock_mode = 0;
1667 it->d.lustre.it_data = NULL;
1670 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1671 ptlrpc_req_finished(req);
1673 tgt = lmv_find_target(lmv, &fid1);
1675 GOTO(out, rc = PTR_ERR(tgt));
1677 OBD_ALLOC_PTR(rdata);
1679 GOTO(out, rc = -ENOMEM);
1681 rdata->op_fid1 = fid1;
1682 rdata->op_bias = MDS_CROSS_REF;
1684 rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1685 lmm, lmmsize, NULL, extra_lock_flags);
1686 OBD_FREE_PTR(rdata);
1689 ldlm_lock_decref(&plock, pmode);
1694 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1695 struct lookup_intent *it, struct md_op_data *op_data,
1696 struct lustre_handle *lockh, void *lmm, int lmmsize,
1697 struct ptlrpc_request **req, int extra_lock_flags)
1699 struct obd_device *obd = exp->exp_obd;
1700 struct lmv_obd *lmv = &obd->u.lmv;
1701 struct lmv_tgt_desc *tgt;
1702 struct lmv_object *obj;
1707 rc = lmv_check_connect(obd);
1711 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1712 LL_IT2STR(it), PFID(&op_data->op_fid1));
1714 if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1715 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1716 lockh, lmm, lmmsize);
1720 obj = lmv_object_find(obd, &op_data->op_fid1);
1721 if (obj && op_data->op_namelen) {
1722 sidx = raw_name2idx(obj->lo_hashtype,
1724 (char *)op_data->op_name,
1725 op_data->op_namelen);
1726 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1727 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1729 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1732 lmv_object_put(obj);
1735 RETURN(PTR_ERR(tgt));
1737 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1738 LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1740 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1741 lmm, lmmsize, req, extra_lock_flags);
1743 if (rc == 0 && it && it->it_op == IT_OPEN) {
1744 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1745 lmm, lmmsize, extra_lock_flags);
1751 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1752 struct ptlrpc_request **request)
1754 struct ptlrpc_request *req = NULL;
1755 struct obd_device *obd = exp->exp_obd;
1756 struct lmv_obd *lmv = &obd->u.lmv;
1757 struct lu_fid rid = op_data->op_fid1;
1758 struct lmv_tgt_desc *tgt;
1759 struct mdt_body *body;
1760 struct lmv_object *obj;
1761 obd_valid valid = op_data->op_valid;
1767 rc = lmv_check_connect(obd);
1774 obj = lmv_object_find(obd, &rid);
1776 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1777 op_data->op_name, op_data->op_namelen);
1778 rid = obj->lo_stripes[sidx].ls_fid;
1779 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1780 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1781 valid &= ~OBD_MD_FLCKSPLIT;
1782 lmv_object_put(obj);
1784 tgt = lmv_find_target(lmv, &rid);
1785 valid |= OBD_MD_FLCKSPLIT;
1786 op_data->op_mds = tgt->ltd_idx;
1789 RETURN(PTR_ERR(tgt));
1791 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1792 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1793 PFID(&rid), tgt->ltd_idx);
1795 op_data->op_valid = valid;
1796 op_data->op_fid1 = rid;
1797 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1799 body = req_capsule_server_get(&(*request)->rq_pill,
1801 LASSERT(body != NULL);
1803 if (body->valid & OBD_MD_MDS) {
1805 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1808 tgt = lmv_find_target(lmv, &rid);
1810 ptlrpc_req_finished(*request);
1811 RETURN(PTR_ERR(tgt));
1814 op_data->op_fid1 = rid;
1815 op_data->op_valid |= OBD_MD_FLCROSSREF;
1816 op_data->op_namelen = 0;
1817 op_data->op_name = NULL;
1818 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1819 ptlrpc_req_finished(*request);
1822 } else if (rc == -ERESTART) {
1823 LASSERT(*request != NULL);
1824 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1825 "Got -ERESTART during getattr!\n");
1826 ptlrpc_req_finished(*request);
1830 * Directory got split. Time to update local object and repeat
1831 * the request with proper MDS.
1833 rc = lmv_handle_split(exp, &rid);
1840 #define md_op_data_fid(op_data, fl) \
1841 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1842 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1843 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1844 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1847 static int lmv_early_cancel_slaves(struct obd_export *exp,
1848 struct md_op_data *op_data, int op_tgt,
1849 ldlm_mode_t mode, int bits, int flag)
1851 struct obd_device *obd = exp->exp_obd;
1852 struct lmv_obd *lmv = &obd->u.lmv;
1853 ldlm_policy_data_t policy = {{0}};
1854 struct lu_fid *op_fid;
1855 struct lu_fid *st_fid;
1856 struct lmv_tgt_desc *tgt;
1857 struct lmv_object *obj;
1862 op_fid = md_op_data_fid(op_data, flag);
1863 if (!fid_is_sane(op_fid))
1866 obj = lmv_object_find(obd, op_fid);
1870 policy.l_inodebits.bits = bits;
1871 for (i = 0; i < obj->lo_objcount; i++) {
1872 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1873 st_fid = &obj->lo_stripes[i].ls_fid;
1874 if (op_tgt != tgt->ltd_idx) {
1875 CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1876 PFID(st_fid), tgt->ltd_idx);
1877 rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1878 mode, LDLM_FL_ASYNC, NULL);
1880 GOTO(out_put_obj, rc);
1883 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1884 op_tgt, PFID(st_fid));
1886 * Do not cancel locks for operation target, they will
1887 * be handled later in underlaying layer when calling
1888 * function we run on behalf of.
1891 op_data->op_flags |= flag;
1896 lmv_object_put(obj);
1900 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1901 int op_tgt, ldlm_mode_t mode, int bits, int flag)
1903 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1904 struct obd_device *obd = exp->exp_obd;
1905 struct lmv_obd *lmv = &obd->u.lmv;
1906 struct lmv_tgt_desc *tgt;
1907 ldlm_policy_data_t policy = {{0}};
1908 struct lmv_object *obj;
1912 if (!fid_is_sane(fid))
1915 obj = lmv_object_find(obd, fid);
1917 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1919 lmv_object_put(obj);
1921 tgt = lmv_find_target(lmv, fid);
1923 RETURN(PTR_ERR(tgt));
1925 if (tgt->ltd_idx != op_tgt) {
1926 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1927 policy.l_inodebits.bits = bits;
1928 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1929 mode, LDLM_FL_ASYNC, NULL);
1932 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1934 op_data->op_flags |= flag;
1943 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1946 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1947 struct ptlrpc_request **request)
1949 struct obd_device *obd = exp->exp_obd;
1950 struct lmv_obd *lmv = &obd->u.lmv;
1951 struct lmv_tgt_desc *tgt;
1952 struct lmv_object *obj;
1959 rc = lmv_check_connect(obd);
1966 LASSERT(op_data->op_namelen != 0);
1968 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1969 PFID(&op_data->op_fid2), op_data->op_namelen,
1970 op_data->op_name, PFID(&op_data->op_fid1));
1972 obj = lmv_object_find(obd, &op_data->op_fid2);
1974 sidx = raw_name2idx(obj->lo_hashtype,
1977 op_data->op_namelen);
1978 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
1979 mds = obj->lo_stripes[sidx].ls_mds;
1980 lmv_object_put(obj);
1982 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1987 CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
1988 mds, PFID(&op_data->op_fid1));
1990 op_data->op_fsuid = cfs_curproc_fsuid();
1991 op_data->op_fsgid = cfs_curproc_fsgid();
1992 op_data->op_cap = cfs_curproc_cap_pack();
1993 tgt = lmv_get_target(lmv, mds);
1996 * Cancel UPDATE lock on child (fid1).
1998 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1999 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2000 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2002 rc = md_link(tgt->ltd_exp, op_data, request);
2003 if (rc == -ERESTART) {
2004 LASSERT(*request != NULL);
2005 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2006 "Got -ERESTART during link!\n");
2007 ptlrpc_req_finished(*request);
2011 * Directory got split. Time to update local object and repeat
2012 * the request with proper MDS.
2014 rc = lmv_handle_split(exp, &op_data->op_fid2);
2022 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2023 const char *old, int oldlen, const char *new, int newlen,
2024 struct ptlrpc_request **request)
2026 struct obd_device *obd = exp->exp_obd;
2027 struct lmv_obd *lmv = &obd->u.lmv;
2028 struct lmv_tgt_desc *src_tgt;
2029 struct lmv_tgt_desc *tgt_tgt;
2033 struct lmv_object *obj;
2038 LASSERT(oldlen != 0);
2040 CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2041 oldlen, old, PFID(&op_data->op_fid1),
2042 newlen, new, PFID(&op_data->op_fid2));
2044 rc = lmv_check_connect(obd);
2051 obj = lmv_object_find(obd, &op_data->op_fid1);
2053 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2054 (char *)old, oldlen);
2055 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2056 mds1 = obj->lo_stripes[sidx].ls_mds;
2057 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2058 lmv_object_put(obj);
2060 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2065 obj = lmv_object_find(obd, &op_data->op_fid2);
2068 * Directory is already split, so we have to forward request to
2071 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2072 (char *)new, newlen);
2074 mds2 = obj->lo_stripes[sidx].ls_mds;
2075 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2076 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2077 lmv_object_put(obj);
2079 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2084 op_data->op_fsuid = cfs_curproc_fsuid();
2085 op_data->op_fsgid = cfs_curproc_fsgid();
2086 op_data->op_cap = cfs_curproc_cap_pack();
2088 src_tgt = lmv_get_target(lmv, mds1);
2089 tgt_tgt = lmv_get_target(lmv, mds2);
2092 * LOOKUP lock on src child (fid3) should also be cancelled for
2093 * src_tgt in mdc_rename.
2095 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2098 * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2101 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2102 LCK_EX, MDS_INODELOCK_UPDATE,
2103 MF_MDC_CANCEL_FID2);
2106 * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2109 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2110 LCK_EX, MDS_INODELOCK_LOOKUP,
2111 MF_MDC_CANCEL_FID4);
2115 * Cancel all the locks on tgt child (fid4).
2118 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2119 LCK_EX, MDS_INODELOCK_FULL,
2120 MF_MDC_CANCEL_FID4);
2123 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2124 new, newlen, request);
2126 if (rc == -ERESTART) {
2127 LASSERT(*request != NULL);
2128 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2129 "Got -ERESTART during rename!\n");
2130 ptlrpc_req_finished(*request);
2134 * Directory got split. Time to update local object and repeat
2135 * the request with proper MDS.
2137 rc = lmv_handle_split(exp, &op_data->op_fid1);
2144 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2145 void *ea, int ealen, void *ea2, int ea2len,
2146 struct ptlrpc_request **request,
2147 struct md_open_data **mod)
2149 struct obd_device *obd = exp->exp_obd;
2150 struct lmv_obd *lmv = &obd->u.lmv;
2151 struct ptlrpc_request *req;
2152 struct lmv_tgt_desc *tgt;
2153 struct lmv_object *obj;
2158 rc = lmv_check_connect(obd);
2162 obj = lmv_object_find(obd, &op_data->op_fid1);
2164 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2165 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2166 obj ? ", split" : "");
2168 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2170 for (i = 0; i < obj->lo_objcount; i++) {
2171 op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2173 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2179 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2180 ea2, ea2len, &req, mod);
2182 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2184 * This is master object and this request should
2185 * be returned back to llite.
2189 ptlrpc_req_finished(req);
2195 lmv_object_put(obj);
2197 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2199 RETURN(PTR_ERR(tgt));
2201 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2202 ea2len, request, mod);
2207 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2208 struct obd_capa *oc, struct ptlrpc_request **request)
2210 struct obd_device *obd = exp->exp_obd;
2211 struct lmv_obd *lmv = &obd->u.lmv;
2212 struct lmv_tgt_desc *tgt;
2216 rc = lmv_check_connect(obd);
2220 tgt = lmv_find_target(lmv, fid);
2222 RETURN(PTR_ERR(tgt));
2224 rc = md_sync(tgt->ltd_exp, fid, oc, request);
2229 * Main purpose of LMV blocking ast is to remove split directory LMV
2230 * presentation object (struct lmv_object) attached to the lock being revoked.
2232 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2233 void *data, int flag)
2235 struct lustre_handle lockh;
2236 struct lmv_object *obj;
2241 case LDLM_CB_BLOCKING:
2242 ldlm_lock2handle(lock, &lockh);
2243 rc = ldlm_cli_cancel(&lockh);
2245 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2249 case LDLM_CB_CANCELING:
2251 * Time to drop cached attrs for split directory object
2253 obj = lock->l_ast_data;
2255 CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2256 ", master "DFID"\n",
2257 lock->l_resource->lr_name.name[3] == 1 ?
2258 "LOOKUP" : "UPDATE",
2259 lock->l_resource->lr_name.name[0],
2260 lock->l_resource->lr_name.name[1],
2261 PFID(&obj->lo_fid));
2262 lmv_object_put(obj);
2271 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2275 val = le64_to_cpu(*hash);
2277 val += MAX_HASH_SIZE;
2278 if (val != DIR_END_OFF)
2279 *hash = cpu_to_le64(val - hash_adj);
2282 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2285 struct obd_import *imp;
2288 * XXX: to get nid we assume that underlying obd device is mdc.
2290 imp = class_exp2cliimp(exp);
2291 id = imp->imp_connection->c_self + fid_flatten(fid);
2293 CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2294 imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2296 return id ^ (id >> 32);
2299 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2300 struct obd_capa *oc, __u64 offset64, struct page *page,
2301 struct ptlrpc_request **request)
2303 struct obd_device *obd = exp->exp_obd;
2304 struct lmv_obd *lmv = &obd->u.lmv;
2305 struct lu_fid rid = *fid;
2306 struct lmv_object *obj;
2316 struct lmv_stripe *los;
2317 struct lmv_tgt_desc *tgt;
2318 struct lu_dirpage *dp;
2319 struct lu_dirent *ent;
2324 rc = lmv_check_connect(obd);
2328 CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2331 * This case handle directory lookup in clustered metadata case (i.e.
2332 * split directory is located on multiple md servers.)
2333 * each server keeps directory entries for certain range of hashes.
2334 * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2335 * first server will keep records with hashes [ 0 ... MAX_HASH / N - 1],
2336 * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2338 * readdir can simply start reading entries from 0 - N server in
2339 * order but that will not scale well as all client will request dir in
2340 * to server in same order.
2341 * Following algorithm does optimization:
2342 * Instead of doing readdir in 1, 2, ...., N order, client with a
2343 * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2344 * (every client has rank R)
2345 * But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2346 * since client ask dir from MDS{R} client has pages with offsets
2347 * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2348 * on hash values that we get.
2350 obj = lmv_object_find_lock(obd, fid);
2352 nr = obj->lo_objcount;
2354 seg_size = MAX_HASH_SIZE;
2355 do_div(seg_size, nr);
2356 los = obj->lo_stripes;
2357 tgt = lmv_get_target(lmv, los[0].ls_mds);
2358 rank = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2360 do_div(tgt_tmp, seg_size);
2361 tgt0_idx = do_div(tgt_tmp, nr);
2362 tgt_idx = (tgt0_idx + rank) % nr;
2364 if (tgt_idx < tgt0_idx)
2368 * Last segment has unusual length due to division
2371 hash_adj = MAX_HASH_SIZE - seg_size * nr;
2375 hash_adj += rank * seg_size;
2377 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2378 LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2379 offset, tgt0_idx, offset + hash_adj, tgt_idx);
2381 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2382 rid = obj->lo_stripes[tgt_idx].ls_fid;
2383 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2385 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2386 PFID(&rid), (unsigned long)offset, tgt_idx);
2388 tgt = lmv_find_target(lmv, &rid);
2391 GOTO(cleanup, rc = PTR_ERR(tgt));
2393 rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2397 dp = cfs_kmap(page);
2399 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2400 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2401 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2403 for (ent = lu_dirent_start(dp); ent != NULL;
2404 ent = lu_dirent_next(ent))
2405 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2407 if (tgt0_idx != nr - 1) {
2410 end = le64_to_cpu(dp->ldp_hash_end);
2411 if (end == DIR_END_OFF) {
2412 dp->ldp_hash_end = cpu_to_le32(seg_size *
2415 ""DFID" reset end "LPX64" tgt %d\n",
2417 le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2425 lmv_object_put_unlock(obj);
2429 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2430 struct ptlrpc_request **request)
2432 struct obd_device *obd = exp->exp_obd;
2433 struct lmv_obd *lmv = &obd->u.lmv;
2434 struct lmv_tgt_desc *tgt = NULL;
2435 struct lmv_object *obj;
2441 rc = lmv_check_connect(obd);
2448 LASSERT(op_data->op_namelen != 0);
2450 obj = lmv_object_find(obd, &op_data->op_fid1);
2452 sidx = raw_name2idx(obj->lo_hashtype,
2455 op_data->op_namelen);
2456 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2457 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2458 tgt = lmv_get_target(lmv,
2459 obj->lo_stripes[sidx].ls_mds);
2460 lmv_object_put(obj);
2461 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2462 op_data->op_namelen, op_data->op_name,
2463 PFID(&op_data->op_fid1), sidx);
2467 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2469 RETURN(PTR_ERR(tgt));
2470 op_data->op_bias |= MDS_CHECK_SPLIT;
2473 op_data->op_fsuid = cfs_curproc_fsuid();
2474 op_data->op_fsgid = cfs_curproc_fsgid();
2475 op_data->op_cap = cfs_curproc_cap_pack();
2478 * If child's fid is given, cancel unused locks for it if it is from
2479 * another export than parent.
2481 * LOOKUP lock for child (fid3) should also be cancelled on parent
2482 * tgt_tgt in mdc_unlink().
2484 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2487 * Cancel FULL locks on child (fid3).
2489 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2490 MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2493 rc = md_unlink(tgt->ltd_exp, op_data, request);
2495 if (rc == -ERESTART) {
2496 LASSERT(*request != NULL);
2497 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2498 "Got -ERESTART during unlink!\n");
2499 ptlrpc_req_finished(*request);
2503 * Directory got split. Time to update local object and repeat
2504 * the request with proper MDS.
2506 rc = lmv_handle_split(exp, &op_data->op_fid1);
2513 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2518 case OBD_CLEANUP_EARLY:
2519 /* XXX: here should be calling obd_precleanup() down to
2522 case OBD_CLEANUP_EXPORTS:
2523 rc = obd_llog_finish(obd, 0);
2525 CERROR("failed to cleanup llogging subsystems\n");
2533 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2534 void *key, __u32 *vallen, void *val,
2535 struct lov_stripe_md *lsm)
2537 struct obd_device *obd;
2538 struct lmv_obd *lmv;
2542 obd = class_exp2obd(exp);
2544 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2545 exp->exp_handle.h_cookie);
2550 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2551 struct lmv_tgt_desc *tgts;
2554 rc = lmv_check_connect(obd);
2558 LASSERT(*vallen == sizeof(__u32));
2559 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2563 * All tgts should be connected when this gets called.
2565 if (!tgts || !tgts->ltd_exp) {
2566 CERROR("target not setup?\n");
2570 if (!obd_get_info(tgts->ltd_exp, keylen, key,
2575 } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2576 rc = lmv_check_connect(obd);
2581 * Forwarding this request to first MDS, it should know LOV
2584 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2586 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2587 exp->exp_connect_flags =
2588 ((struct obd_connect_data *)val)->ocd_connect_flags;
2593 CDEBUG(D_IOCTL, "Invalid key\n");
2597 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2598 void *key, obd_count vallen, void *val,
2599 struct ptlrpc_request_set *set)
2601 struct lmv_tgt_desc *tgt;
2602 struct obd_device *obd;
2603 struct lmv_obd *lmv;
2607 obd = class_exp2obd(exp);
2609 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2610 exp->exp_handle.h_cookie);
2615 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2616 KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2619 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2620 tgt = &lmv->tgts[i];
2625 err = obd_set_info_async(tgt->ltd_exp,
2626 keylen, key, vallen, val, set);
2637 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2638 struct lov_stripe_md *lsm)
2640 struct obd_device *obd = class_exp2obd(exp);
2641 struct lmv_obd *lmv = &obd->u.lmv;
2642 struct lmv_stripe_md *meap;
2643 struct lmv_stripe_md *lsmp;
2648 mea_size = lmv_get_easize(lmv);
2652 if (*lmmp && !lsm) {
2653 OBD_FREE(*lmmp, mea_size);
2658 if (*lmmp == NULL) {
2659 OBD_ALLOC(*lmmp, mea_size);
2667 lsmp = (struct lmv_stripe_md *)lsm;
2668 meap = (struct lmv_stripe_md *)*lmmp;
2670 if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2671 lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2674 meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2675 meap->mea_count = cpu_to_le32(lsmp->mea_count);
2676 meap->mea_master = cpu_to_le32(lsmp->mea_master);
2678 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2679 meap->mea_ids[i] = meap->mea_ids[i];
2680 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2686 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2687 struct lov_mds_md *lmm, int lmm_size)
2689 struct obd_device *obd = class_exp2obd(exp);
2690 struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2691 struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2692 struct lmv_obd *lmv = &obd->u.lmv;
2698 mea_size = lmv_get_easize(lmv);
2702 if (*lsmp != NULL && lmm == NULL) {
2703 OBD_FREE(*tmea, mea_size);
2708 LASSERT(mea_size == lmm_size);
2710 OBD_ALLOC(*tmea, mea_size);
2717 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2718 mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2719 mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2721 magic = le32_to_cpu(mea->mea_magic);
2724 * Old mea is not handled here.
2726 CERROR("Old not supportable EA is found\n");
2730 (*tmea)->mea_magic = magic;
2731 (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2732 (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2734 for (i = 0; i < (*tmea)->mea_count; i++) {
2735 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2736 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2741 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2742 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2743 int flags, void *opaque)
2745 struct obd_device *obd = exp->exp_obd;
2746 struct lmv_obd *lmv = &obd->u.lmv;
2752 LASSERT(fid != NULL);
2754 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2755 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2758 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2759 policy, mode, flags, opaque);
2766 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2769 struct obd_device *obd = exp->exp_obd;
2770 struct lmv_obd *lmv = &obd->u.lmv;
2774 rc = md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2778 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2779 const struct lu_fid *fid, ldlm_type_t type,
2780 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2781 struct lustre_handle *lockh)
2783 struct obd_device *obd = exp->exp_obd;
2784 struct lmv_obd *lmv = &obd->u.lmv;
2789 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2792 * With CMD every object can have two locks in different namespaces:
2793 * lookup lock in space of mds storing direntry and update/open lock in
2794 * space of mds storing inode. Thus we check all targets, not only that
2795 * one fid was created in.
2797 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2798 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2799 type, policy, mode, lockh);
2807 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2808 struct obd_export *dt_exp, struct obd_export *md_exp,
2809 struct lustre_md *md)
2811 struct obd_device *obd = exp->exp_obd;
2812 struct lmv_obd *lmv = &obd->u.lmv;
2815 rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2819 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2821 struct obd_device *obd = exp->exp_obd;
2822 struct lmv_obd *lmv = &obd->u.lmv;
2826 obd_free_memmd(exp, (void *)&md->mea);
2827 RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2830 int lmv_set_open_replay_data(struct obd_export *exp,
2831 struct obd_client_handle *och,
2832 struct ptlrpc_request *open_req)
2834 struct obd_device *obd = exp->exp_obd;
2835 struct lmv_obd *lmv = &obd->u.lmv;
2836 struct lmv_tgt_desc *tgt;
2839 tgt = lmv_find_target(lmv, &och->och_fid);
2841 RETURN(PTR_ERR(tgt));
2843 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2846 int lmv_clear_open_replay_data(struct obd_export *exp,
2847 struct obd_client_handle *och)
2849 struct obd_device *obd = exp->exp_obd;
2850 struct lmv_obd *lmv = &obd->u.lmv;
2851 struct lmv_tgt_desc *tgt;
2854 tgt = lmv_find_target(lmv, &och->och_fid);
2856 RETURN(PTR_ERR(tgt));
2858 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2861 static int lmv_get_remote_perm(struct obd_export *exp,
2862 const struct lu_fid *fid,
2863 struct obd_capa *oc, __u32 suppgid,
2864 struct ptlrpc_request **request)
2866 struct obd_device *obd = exp->exp_obd;
2867 struct lmv_obd *lmv = &obd->u.lmv;
2868 struct lmv_tgt_desc *tgt;
2872 rc = lmv_check_connect(obd);
2876 tgt = lmv_find_target(lmv, fid);
2878 RETURN(PTR_ERR(tgt));
2880 rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2884 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2887 struct obd_device *obd = exp->exp_obd;
2888 struct lmv_obd *lmv = &obd->u.lmv;
2889 struct lmv_tgt_desc *tgt;
2893 rc = lmv_check_connect(obd);
2897 tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2899 RETURN(PTR_ERR(tgt));
2901 rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2905 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2906 const struct req_msg_field *field, struct obd_capa **oc)
2908 struct obd_device *obd = exp->exp_obd;
2909 struct lmv_obd *lmv = &obd->u.lmv;
2913 rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2917 int lmv_intent_getattr_async(struct obd_export *exp,
2918 struct md_enqueue_info *minfo,
2919 struct ldlm_enqueue_info *einfo)
2921 struct md_op_data *op_data = &minfo->mi_data;
2922 struct obd_device *obd = exp->exp_obd;
2923 struct lmv_obd *lmv = &obd->u.lmv;
2924 struct lmv_object *obj;
2925 struct lmv_tgt_desc *tgt;
2930 rc = lmv_check_connect(obd);
2934 if (!fid_is_sane(&op_data->op_fid2)) {
2935 obj = lmv_object_find(obd, &op_data->op_fid1);
2936 if (obj && op_data->op_namelen) {
2937 sidx = raw_name2idx(obj->lo_hashtype,
2939 (char *)op_data->op_name,
2940 op_data->op_namelen);
2941 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2942 tgt = lmv_get_target(lmv,
2943 obj->lo_stripes[sidx].ls_mds);
2945 "Choose slave dir ("DFID") -> mds #%d\n",
2946 PFID(&op_data->op_fid1), tgt->ltd_idx);
2948 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2951 lmv_object_put(obj);
2953 op_data->op_fid1 = op_data->op_fid2;
2954 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2955 op_data->op_bias = MDS_CROSS_REF;
2957 * Unfortunately, we have to lie to MDC/MDS to retrieve
2958 * attributes llite needs.
2960 if (minfo->mi_it.it_op & IT_LOOKUP)
2961 minfo->mi_it.it_op = IT_GETATTR;
2965 RETURN(PTR_ERR(tgt));
2967 rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2971 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
2972 struct lu_fid *fid, __u32 *bits)
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, bits);
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);