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 atomic_t lmv_object_count = 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 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 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 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 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 lustre_handle *conn, 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 obd_export *exp;
254 rc = class_connect(conn, obd, cluuid);
256 CERROR("class_connection() returned %d\n", rc);
260 exp = class_conn2export(conn);
263 * We don't want to actually do the underlying connections more than
264 * once, so keep track.
267 if (lmv->refcount > 1) {
268 class_export_put(exp);
274 lmv->cluuid = *cluuid;
277 lmv->conn_data = *data;
280 lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
282 if (IS_ERR(lmv_proc_dir)) {
283 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
284 obd->obd_type->typ_name, obd->obd_name);
290 * All real clients should perform actual connection right away, because
291 * it is possible, that LMV will not have opportunity to connect targets
292 * and MDC stuff will be called directly, for instance while reading
293 * ../mdc/../kbytesfree procfs file, etc.
295 if (data->ocd_connect_flags & OBD_CONNECT_REAL)
296 rc = lmv_check_connect(obd);
301 lprocfs_remove(&lmv_proc_dir);
308 static void lmv_set_timeouts(struct obd_device *obd)
310 struct lmv_tgt_desc *tgts;
315 if (lmv->server_timeout == 0)
318 if (lmv->connected == 0)
321 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
322 if (tgts->ltd_exp == NULL)
325 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
326 KEY_INTERMDS, 0, NULL, NULL);
330 static int lmv_init_ea_size(struct obd_export *exp, int easize,
331 int def_easize, int cookiesize)
333 struct obd_device *obd = exp->exp_obd;
334 struct lmv_obd *lmv = &obd->u.lmv;
340 if (lmv->max_easize < easize) {
341 lmv->max_easize = easize;
344 if (lmv->max_def_easize < def_easize) {
345 lmv->max_def_easize = def_easize;
348 if (lmv->max_cookiesize < cookiesize) {
349 lmv->max_cookiesize = cookiesize;
355 if (lmv->connected == 0)
358 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
359 if (lmv->tgts[i].ltd_exp == NULL) {
360 CWARN("%s: NULL export for %d\n", obd->obd_name, i);
364 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
367 CERROR("obd_init_ea_size() failed on MDT target %d, "
368 "error %d.\n", i, rc);
375 #define MAX_STRING_SIZE 128
377 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
380 struct proc_dir_entry *lmv_proc_dir;
382 struct lmv_obd *lmv = &obd->u.lmv;
383 struct obd_uuid *cluuid = &lmv->cluuid;
384 struct obd_connect_data *mdc_data = NULL;
385 struct obd_uuid lmv_mdc_uuid = { "LMV_MDC_UUID" };
386 struct lustre_handle conn = {0, };
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, &conn, mdc_obd, &lmv_mdc_uuid,
411 &lmv->conn_data, NULL);
413 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
417 mdc_exp = class_conn2export(&conn);
420 * Init fid sequence client for this mdc and add new fld target.
422 rc = obd_fid_init(mdc_exp);
426 target.ft_srv = NULL;
427 target.ft_exp = mdc_exp;
428 target.ft_idx = tgt->ltd_idx;
430 fld_client_add_target(&lmv->lmv_fld, &target);
432 mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
434 rc = obd_register_observer(mdc_obd, obd);
436 obd_disconnect(mdc_exp);
437 CERROR("target %s register_observer error %d\n",
438 tgt->ltd_uuid.uuid, rc);
442 if (obd->obd_observer) {
444 * Tell the observer about the new target.
446 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
447 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
449 obd_disconnect(mdc_exp);
455 tgt->ltd_exp = mdc_exp;
456 lmv->desc.ld_active_tgt_count++;
459 * Copy connect data, it may be used later.
461 lmv->datas[tgt->ltd_idx] = *mdc_data;
463 md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
464 lmv->max_def_easize, lmv->max_cookiesize);
466 CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
467 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
468 atomic_read(&obd->obd_refcount));
471 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
473 struct proc_dir_entry *mdc_symlink;
474 char name[MAX_STRING_SIZE + 1];
476 LASSERT(mdc_obd->obd_type != NULL);
477 LASSERT(mdc_obd->obd_type->typ_name != NULL);
478 name[MAX_STRING_SIZE] = '\0';
479 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
480 mdc_obd->obd_type->typ_name,
482 mdc_symlink = proc_symlink(mdc_obd->obd_name,
484 if (mdc_symlink == NULL) {
485 CERROR("Could not register LMV target "
486 "/proc/fs/lustre/%s/%s/target_obds/%s.",
487 obd->obd_type->typ_name, obd->obd_name,
489 lprocfs_remove(&lmv_proc_dir);
497 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
499 struct lmv_obd *lmv = &obd->u.lmv;
500 struct lmv_tgt_desc *tgt;
504 CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
508 if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
509 lmv_init_unlock(lmv);
510 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
511 "That many MDCs already configured.\n",
512 tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
515 if (lmv->desc.ld_tgt_count == 0) {
516 struct obd_device *mdc_obd;
518 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
521 lmv_init_unlock(lmv);
522 CERROR("Target %s not attached\n", tgt_uuid->uuid);
526 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, 0, NULL, tgt_uuid);
528 lmv_init_unlock(lmv);
529 CERROR("lmv failed to setup llogging subsystems\n");
532 spin_lock(&lmv->lmv_lock);
533 tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
534 tgt->ltd_uuid = *tgt_uuid;
535 spin_unlock(&lmv->lmv_lock);
537 if (lmv->connected) {
538 rc = lmv_connect_mdc(obd, tgt);
540 spin_lock(&lmv->lmv_lock);
541 lmv->desc.ld_tgt_count--;
542 memset(tgt, 0, sizeof(*tgt));
543 spin_unlock(&lmv->lmv_lock);
545 int easize = sizeof(struct lmv_stripe_md) +
546 lmv->desc.ld_tgt_count *
547 sizeof(struct lu_fid);
548 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
552 lmv_init_unlock(lmv);
556 int lmv_check_connect(struct obd_device *obd)
558 struct lmv_obd *lmv = &obd->u.lmv;
559 struct lmv_tgt_desc *tgt;
569 if (lmv->connected) {
570 lmv_init_unlock(lmv);
574 if (lmv->desc.ld_tgt_count == 0) {
575 CERROR("%s: no targets configured.\n", obd->obd_name);
579 CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
580 lmv->cluuid.uuid, obd->obd_name);
582 LASSERT(lmv->tgts != NULL);
584 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
585 rc = lmv_connect_mdc(obd, tgt);
590 lmv_set_timeouts(obd);
591 class_export_put(lmv->exp);
593 easize = lmv_get_easize(lmv);
594 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
595 lmv_init_unlock(lmv);
604 --lmv->desc.ld_active_tgt_count;
605 rc2 = obd_disconnect(tgt->ltd_exp);
607 CERROR("LMV target %s disconnect on "
608 "MDC idx %d: error %d\n",
609 tgt->ltd_uuid.uuid, i, rc2);
613 class_disconnect(lmv->exp);
614 lmv_init_unlock(lmv);
618 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
621 struct proc_dir_entry *lmv_proc_dir;
623 struct lmv_obd *lmv = &obd->u.lmv;
624 struct obd_device *mdc_obd;
628 LASSERT(tgt != NULL);
629 LASSERT(obd != NULL);
631 mdc_obd = class_exp2obd(tgt->ltd_exp);
634 mdc_obd->obd_no_recov = obd->obd_no_recov;
637 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
639 struct proc_dir_entry *mdc_symlink;
641 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
643 lprocfs_remove(&mdc_symlink);
645 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
646 obd->obd_type->typ_name, obd->obd_name,
651 rc = obd_fid_fini(tgt->ltd_exp);
653 CERROR("Can't finanize fids factory\n");
655 CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
656 tgt->ltd_exp->exp_obd->obd_name,
657 tgt->ltd_exp->exp_obd->obd_uuid.uuid);
659 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
660 rc = obd_disconnect(tgt->ltd_exp);
662 if (tgt->ltd_active) {
663 CERROR("Target %s disconnect error %d\n",
664 tgt->ltd_uuid.uuid, rc);
668 lmv_activate_target(lmv, tgt, 0);
673 static int lmv_disconnect(struct obd_export *exp)
675 struct obd_device *obd = class_exp2obd(exp);
677 struct proc_dir_entry *lmv_proc_dir;
679 struct lmv_obd *lmv = &obd->u.lmv;
688 * Only disconnect the underlying layers on the final disconnect.
691 if (lmv->refcount != 0)
694 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
695 if (lmv->tgts[i].ltd_exp == NULL)
697 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
701 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
703 lprocfs_remove(&lmv_proc_dir);
705 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
706 obd->obd_type->typ_name, obd->obd_name);
712 * This is the case when no real connection is established by
713 * lmv_check_connect().
716 class_export_put(exp);
717 rc = class_disconnect(exp);
718 if (lmv->refcount == 0)
723 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
724 int len, void *karg, void *uarg)
726 struct obd_device *obddev = class_exp2obd(exp);
727 struct lmv_obd *lmv = &obddev->u.lmv;
733 if (lmv->desc.ld_tgt_count == 0)
737 case IOC_OBD_STATFS: {
738 struct obd_ioctl_data *data = karg;
739 struct obd_device *mdc_obd;
740 struct obd_statfs stat_buf = {0};
743 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
744 LASSERT(data->ioc_plen1 == sizeof(struct obd_statfs));
746 if ((index >= lmv->desc.ld_tgt_count))
749 if (!lmv->tgts[index].ltd_active)
752 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
756 rc = obd_statfs(mdc_obd, &stat_buf,
757 cfs_time_current_64() - HZ, 0);
760 if (copy_to_user(data->ioc_pbuf1, &stat_buf, data->ioc_plen1))
762 rc = copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
767 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
770 if (lmv->tgts[i].ltd_exp == NULL)
773 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
776 if (lmv->tgts[i].ltd_active) {
777 CERROR("error: iocontrol MDC %s on MDT"
778 "idx %d cmd %x: err = %d\n",
779 lmv->tgts[i].ltd_uuid.uuid,
794 static int lmv_all_chars_policy(int count, const char *name,
805 static int lmv_nid_policy(struct lmv_obd *lmv)
807 struct obd_import *imp;
811 * XXX: To get nid we assume that underlying obd device is mdc.
813 imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
814 id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
815 return id % lmv->desc.ld_tgt_count;
818 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
819 placement_policy_t placement)
822 case PLACEMENT_CHAR_POLICY:
823 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
825 op_data->op_namelen);
826 case PLACEMENT_NID_POLICY:
827 return lmv_nid_policy(lmv);
833 CERROR("Unsupported placement policy %x\n", placement);
838 * This is _inode_ placement policy function (not name).
840 static int lmv_placement_policy(struct obd_device *obd,
841 struct md_op_data *op_data,
844 struct lmv_obd *lmv = &obd->u.lmv;
845 struct lmv_object *obj;
849 LASSERT(mds != NULL);
851 if (lmv->desc.ld_tgt_count == 1) {
857 * Allocate new fid on target according to operation type and parent
860 obj = lmv_object_find(obd, &op_data->op_fid1);
861 if (obj != NULL || op_data->op_name == NULL ||
862 op_data->op_opc != LUSTRE_OPC_MKDIR) {
864 * Allocate fid for non-dir or for null name or for case parent
871 * If we have this flag turned on, and we see that
872 * parent dir is split, this means, that caller did not
873 * notice split yet. This is race and we would like to
874 * let caller know that.
876 if (op_data->op_bias & MDS_CHECK_SPLIT)
881 * Allocate new fid on same mds where parent fid is located and
882 * where operation will be sent. In case of split dir, ->op_fid1
883 * and ->op_mds here will contain fid and mds of slave directory
884 * object (assigned by caller).
886 *mds = op_data->op_mds;
890 * Parent directory is not split and we want to create a
891 * directory in it. Let's calculate where to place it according
892 * to operation data @op_data.
894 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
899 CERROR("Can't choose MDS, err = %d\n", rc);
901 LASSERT(*mds < lmv->desc.ld_tgt_count);
907 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
910 struct lmv_tgt_desc *tgt;
914 tgt = lmv_get_target(lmv, mds);
917 * New seq alloc and FLD setup should be atomic. Otherwise we may find
918 * on server that seq in new allocated fid is not yet known.
920 down(&tgt->ltd_fid_sem);
922 if (!tgt->ltd_active)
923 GOTO(out, rc = -ENODEV);
926 * Asking underlaying tgt layer to allocate new fid.
928 rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
930 LASSERT(fid_is_sane(fid));
933 * Client switches to new sequence, setup FLD.
935 rc = fld_client_create(&lmv->lmv_fld, fid_seq(fid),
939 * Delete just allocated fid sequence in case
942 CERROR("Can't create fld entry, rc %d\n", rc);
943 obd_fid_delete(tgt->ltd_exp, NULL);
949 up(&tgt->ltd_fid_sem);
953 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
954 struct md_op_data *op_data)
956 struct obd_device *obd = class_exp2obd(exp);
957 struct lmv_obd *lmv = &obd->u.lmv;
962 LASSERT(op_data != NULL);
963 LASSERT(fid != NULL);
965 rc = lmv_placement_policy(obd, op_data, &mds);
967 CERROR("Can't get target for allocating fid, "
972 rc = __lmv_fid_alloc(lmv, fid, mds);
974 CERROR("Can't alloc new fid, rc %d\n", rc);
981 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
984 LASSERT(exp != NULL && fid != NULL);
985 if (lmv_object_delete(exp, fid)) {
986 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
992 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
994 struct lmv_obd *lmv = &obd->u.lmv;
995 struct lprocfs_static_vars lvars;
996 struct lmv_desc *desc;
1001 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1002 CERROR("LMV setup requires a descriptor\n");
1006 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1007 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1008 CERROR("Lmv descriptor size wrong: %d > %d\n",
1009 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1013 lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1015 OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1016 if (lmv->tgts == NULL)
1019 for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1020 sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1021 lmv->tgts[i].ltd_idx = i;
1024 lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1026 OBD_ALLOC(lmv->datas, lmv->datas_size);
1027 if (lmv->datas == NULL)
1028 GOTO(out_free_tgts, rc = -ENOMEM);
1030 obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1031 lmv->desc.ld_tgt_count = 0;
1032 lmv->desc.ld_active_tgt_count = 0;
1033 lmv->max_cookiesize = 0;
1034 lmv->max_def_easize = 0;
1035 lmv->max_easize = 0;
1036 lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1038 spin_lock_init(&lmv->lmv_lock);
1039 sema_init(&lmv->init_sem, 1);
1041 rc = lmv_object_setup(obd);
1043 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1044 GOTO(out_free_datas, rc);
1047 lprocfs_lmv_init_vars(&lvars);
1048 lprocfs_obd_setup(obd, lvars.obd_vars);
1051 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1052 0444, &lmv_proc_target_fops, obd);
1054 CWARN("Error adding target_obd_stats file (%d)\n", rc);
1057 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1058 LUSTRE_CLI_FLD_HASH_DHT);
1060 CERROR("Can't init FLD, err %d\n", rc);
1061 GOTO(out_free_datas, rc);
1067 OBD_FREE(lmv->datas, lmv->datas_size);
1070 OBD_FREE(lmv->tgts, lmv->tgts_size);
1075 static int lmv_cleanup(struct obd_device *obd)
1077 struct lmv_obd *lmv = &obd->u.lmv;
1080 fld_client_fini(&lmv->lmv_fld);
1081 lprocfs_obd_cleanup(obd);
1082 lmv_object_cleanup(obd);
1083 OBD_FREE(lmv->datas, lmv->datas_size);
1084 OBD_FREE(lmv->tgts, lmv->tgts_size);
1089 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1091 struct lustre_cfg *lcfg = buf;
1092 struct obd_uuid tgt_uuid;
1096 switch(lcfg->lcfg_command) {
1098 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1099 GOTO(out, rc = -EINVAL);
1101 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1102 rc = lmv_add_target(obd, &tgt_uuid);
1105 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1106 GOTO(out, rc = -EINVAL);
1113 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1114 __u64 max_age, __u32 flags)
1116 struct lmv_obd *lmv = &obd->u.lmv;
1117 struct obd_statfs *temp;
1122 rc = lmv_check_connect(obd);
1126 OBD_ALLOC(temp, sizeof(*temp));
1130 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1131 if (lmv->tgts[i].ltd_exp == NULL)
1134 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1137 CERROR("can't stat MDS #%d (%s), error %d\n", i,
1138 lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1140 GOTO(out_free_temp, rc);
1145 osfs->os_bavail += temp->os_bavail;
1146 osfs->os_blocks += temp->os_blocks;
1147 osfs->os_ffree += temp->os_ffree;
1148 osfs->os_files += temp->os_files;
1154 OBD_FREE(temp, sizeof(*temp));
1158 static int lmv_getstatus(struct obd_export *exp,
1160 struct obd_capa **pc)
1162 struct obd_device *obd = exp->exp_obd;
1163 struct lmv_obd *lmv = &obd->u.lmv;
1167 rc = lmv_check_connect(obd);
1171 rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1175 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1176 struct obd_capa *oc, obd_valid valid, const char *name,
1177 const char *input, int input_size, int output_size,
1178 int flags, struct ptlrpc_request **request)
1180 struct obd_device *obd = exp->exp_obd;
1181 struct lmv_obd *lmv = &obd->u.lmv;
1182 struct lmv_tgt_desc *tgt;
1186 rc = lmv_check_connect(obd);
1190 tgt = lmv_find_target(lmv, fid);
1192 RETURN(PTR_ERR(tgt));
1194 rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1195 input_size, output_size, flags, request);
1200 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1201 struct obd_capa *oc, obd_valid valid, const char *name,
1202 const char *input, int input_size, int output_size,
1203 int flags, __u32 suppgid,
1204 struct ptlrpc_request **request)
1206 struct obd_device *obd = exp->exp_obd;
1207 struct lmv_obd *lmv = &obd->u.lmv;
1208 struct lmv_tgt_desc *tgt;
1212 rc = lmv_check_connect(obd);
1216 tgt = lmv_find_target(lmv, fid);
1218 RETURN(PTR_ERR(tgt));
1220 rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1221 input_size, output_size, flags, suppgid,
1227 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1228 struct obd_capa *oc, obd_valid valid, int ea_size,
1229 struct ptlrpc_request **request)
1231 struct obd_device *obd = exp->exp_obd;
1232 struct lmv_obd *lmv = &obd->u.lmv;
1233 struct lmv_tgt_desc *tgt;
1234 struct lmv_object *obj;
1239 rc = lmv_check_connect(obd);
1243 tgt = lmv_find_target(lmv, fid);
1245 RETURN(PTR_ERR(tgt));
1247 rc = md_getattr(tgt->ltd_exp, fid, oc, valid, ea_size, request);
1251 obj = lmv_object_find_lock(obd, fid);
1253 CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(fid),
1254 obj ? "(split)" : "");
1257 * If object is split, then we loop over all the slaves and gather size
1258 * attribute. In ideal world we would have to gather also mds field from
1259 * all slaves, as object is spread over the cluster and this is
1260 * definitely interesting information and it is not good to loss it,
1264 struct mdt_body *body;
1266 if (*request == NULL) {
1267 lmv_object_put(obj);
1271 body = req_capsule_server_get(&(*request)->rq_pill,
1273 LASSERT(body != NULL);
1275 for (i = 0; i < obj->lo_objcount; i++) {
1276 if (lmv->tgts[i].ltd_exp == NULL) {
1277 CWARN("%s: NULL export for %d\n",
1283 * Skip master object.
1285 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1288 body->size += obj->lo_stripes[i].ls_size;
1291 lmv_object_put_unlock(obj);
1297 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1298 ldlm_iterator_t it, void *data)
1300 struct obd_device *obd = exp->exp_obd;
1301 struct lmv_obd *lmv = &obd->u.lmv;
1306 rc = lmv_check_connect(obd);
1310 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1313 * With CMD every object can have two locks in different namespaces:
1314 * lookup lock in space of mds storing direntry and update/open lock in
1315 * space of mds storing inode.
1317 for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1318 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1323 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1324 struct md_open_data *mod, struct ptlrpc_request **request)
1326 struct obd_device *obd = exp->exp_obd;
1327 struct lmv_obd *lmv = &obd->u.lmv;
1328 struct lmv_tgt_desc *tgt;
1332 rc = lmv_check_connect(obd);
1336 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1338 RETURN(PTR_ERR(tgt));
1340 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1341 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1346 * Called in the case MDS returns -ERESTART on create on open, what means that
1347 * directory is split and its LMV presentation object has to be updated.
1349 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1351 struct obd_device *obd = exp->exp_obd;
1352 struct lmv_obd *lmv = &obd->u.lmv;
1353 struct ptlrpc_request *req = NULL;
1354 struct lmv_tgt_desc *tgt;
1355 struct lmv_object *obj;
1356 struct lustre_md md;
1363 mealen = lmv_get_easize(lmv);
1365 valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1367 tgt = lmv_find_target(lmv, fid);
1369 RETURN(PTR_ERR(tgt));
1372 * Time to update mea of parent fid.
1374 rc = md_getattr(tgt->ltd_exp, fid, NULL, valid, mealen, &req);
1376 CERROR("md_getattr() failed, error %d\n", rc);
1380 rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1382 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1387 GOTO(cleanup, rc = -ENODATA);
1389 obj = lmv_object_create(exp, fid, md.mea);
1393 lmv_object_put(obj);
1395 obd_free_memmd(exp, (void *)&md.mea);
1399 ptlrpc_req_finished(req);
1403 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1404 const void *data, int datalen, int mode, __u32 uid,
1405 __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1406 struct ptlrpc_request **request)
1408 struct obd_device *obd = exp->exp_obd;
1409 struct lmv_obd *lmv = &obd->u.lmv;
1410 struct lmv_tgt_desc *tgt;
1411 struct lmv_object *obj;
1417 rc = lmv_check_connect(obd);
1421 if (!lmv->desc.ld_active_tgt_count)
1427 obj = lmv_object_find(obd, &op_data->op_fid1);
1429 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1430 op_data->op_name, op_data->op_namelen);
1431 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1432 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1433 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1434 tgt = lmv_get_target(lmv, op_data->op_mds);
1435 lmv_object_put(obj);
1437 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1438 op_data->op_bias |= MDS_CHECK_SPLIT;
1439 op_data->op_mds = tgt->ltd_idx;
1443 RETURN(PTR_ERR(tgt));
1445 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1446 if (rc == -ERESTART)
1451 CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #"LPU64"\n",
1452 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1455 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1456 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1457 cap_effective, rdev, request);
1459 if (*request == NULL)
1461 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1462 } else if (rc == -ERESTART) {
1463 LASSERT(*request != NULL);
1464 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1465 "Got -ERESTART during create!\n");
1466 ptlrpc_req_finished(*request);
1470 * Directory got split. Time to update local object and repeat
1471 * the request with proper MDS.
1473 rc = lmv_handle_split(exp, &op_data->op_fid1);
1475 rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1476 op_data, &op_data->op_fid2);
1485 static int lmv_done_writing(struct obd_export *exp,
1486 struct md_op_data *op_data,
1487 struct md_open_data *mod)
1489 struct obd_device *obd = exp->exp_obd;
1490 struct lmv_obd *lmv = &obd->u.lmv;
1491 struct lmv_tgt_desc *tgt;
1495 rc = lmv_check_connect(obd);
1499 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1501 RETURN(PTR_ERR(tgt));
1503 rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1508 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1509 struct lookup_intent *it, struct md_op_data *op_data,
1510 struct lustre_handle *lockh, void *lmm, int lmmsize)
1512 struct obd_device *obd = exp->exp_obd;
1513 struct lmv_obd *lmv = &obd->u.lmv;
1514 struct lmv_stripe_md *mea = op_data->op_mea1;
1515 struct md_op_data *op_data2;
1516 struct lmv_tgt_desc *tgt;
1521 OBD_ALLOC_PTR(op_data2);
1522 if (op_data2 == NULL)
1525 LASSERT(mea != NULL);
1526 for (i = 0; i < mea->mea_count; i++) {
1527 memset(op_data2, 0, sizeof(*op_data2));
1528 op_data2->op_fid1 = mea->mea_ids[i];
1529 op_data2->op_bias = 0;
1531 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1533 GOTO(cleanup, rc = PTR_ERR(tgt));
1535 if (tgt->ltd_exp == NULL)
1538 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1539 lockh + i, lmm, lmmsize, NULL, 0);
1541 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1542 PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1547 if (it->d.lustre.it_data) {
1548 struct ptlrpc_request *req;
1549 req = (struct ptlrpc_request *)it->d.lustre.it_data;
1550 ptlrpc_req_finished(req);
1553 if (it->d.lustre.it_status)
1554 GOTO(cleanup, rc = it->d.lustre.it_status);
1559 OBD_FREE_PTR(op_data2);
1563 * Drop all taken locks.
1566 if (lockh[i].cookie)
1567 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1568 lockh[i].cookie = 0;
1575 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1576 struct lookup_intent *it, struct md_op_data *op_data,
1577 struct lustre_handle *lockh, void *lmm, int lmmsize,
1578 int extra_lock_flags)
1580 struct ptlrpc_request *req = it->d.lustre.it_data;
1581 struct obd_device *obd = exp->exp_obd;
1582 struct lmv_obd *lmv = &obd->u.lmv;
1583 struct lustre_handle plock;
1584 struct lmv_tgt_desc *tgt;
1585 struct md_op_data *rdata;
1587 struct mdt_body *body;
1592 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1593 LASSERT(body != NULL);
1595 if (!(body->valid & OBD_MD_MDS))
1598 CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1599 LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1602 * We got LOOKUP lock, but we really need attrs.
1604 pmode = it->d.lustre.it_lock_mode;
1605 LASSERT(pmode != 0);
1606 memcpy(&plock, lockh, sizeof(plock));
1607 it->d.lustre.it_lock_mode = 0;
1608 it->d.lustre.it_data = NULL;
1611 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1612 ptlrpc_req_finished(req);
1614 tgt = lmv_find_target(lmv, &fid1);
1616 GOTO(out, rc = PTR_ERR(tgt));
1618 OBD_ALLOC_PTR(rdata);
1620 GOTO(out, rc = -ENOMEM);
1622 rdata->op_fid1 = fid1;
1623 rdata->op_bias = MDS_CROSS_REF;
1625 rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1626 lmm, lmmsize, NULL, extra_lock_flags);
1627 OBD_FREE_PTR(rdata);
1630 ldlm_lock_decref(&plock, pmode);
1635 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1636 struct lookup_intent *it, struct md_op_data *op_data,
1637 struct lustre_handle *lockh, void *lmm, int lmmsize,
1638 struct ptlrpc_request **req, int extra_lock_flags)
1640 struct obd_device *obd = exp->exp_obd;
1641 struct lmv_obd *lmv = &obd->u.lmv;
1642 struct lmv_tgt_desc *tgt;
1643 struct lmv_object *obj;
1648 rc = lmv_check_connect(obd);
1652 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1653 LL_IT2STR(it), PFID(&op_data->op_fid1));
1655 if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1656 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1657 lockh, lmm, lmmsize);
1661 obj = lmv_object_find(obd, &op_data->op_fid1);
1662 if (obj && op_data->op_namelen) {
1663 sidx = raw_name2idx(obj->lo_hashtype,
1665 (char *)op_data->op_name,
1666 op_data->op_namelen);
1667 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1668 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1670 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1673 lmv_object_put(obj);
1676 RETURN(PTR_ERR(tgt));
1678 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1679 LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1681 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1682 lmm, lmmsize, req, extra_lock_flags);
1684 if (rc == 0 && it && it->it_op == IT_OPEN) {
1685 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1686 lmm, lmmsize, extra_lock_flags);
1692 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1693 struct obd_capa *oc, const char *name, int namelen,
1694 obd_valid valid, int ea_size, __u32 suppgid,
1695 struct ptlrpc_request **request)
1697 struct ptlrpc_request *req = NULL;
1698 struct obd_device *obd = exp->exp_obd;
1699 struct lmv_obd *lmv = &obd->u.lmv;
1700 struct lu_fid rid = *fid;
1701 struct lmv_tgt_desc *tgt;
1702 struct mdt_body *body;
1703 struct lmv_object *obj;
1709 rc = lmv_check_connect(obd);
1716 obj = lmv_object_find(obd, &rid);
1718 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1720 rid = obj->lo_stripes[sidx].ls_fid;
1721 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1722 valid &= ~OBD_MD_FLCKSPLIT;
1723 lmv_object_put(obj);
1725 tgt = lmv_find_target(lmv, &rid);
1726 valid |= OBD_MD_FLCKSPLIT;
1729 RETURN(PTR_ERR(tgt));
1731 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1732 namelen, name, PFID(fid), PFID(&rid), tgt->ltd_idx);
1734 rc = md_getattr_name(tgt->ltd_exp, &rid, oc, name, namelen, valid,
1735 ea_size, suppgid, request);
1737 body = req_capsule_server_get(&(*request)->rq_pill,
1739 LASSERT(body != NULL);
1741 if (body->valid & OBD_MD_MDS) {
1743 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1746 tgt = lmv_find_target(lmv, &rid);
1748 ptlrpc_req_finished(*request);
1749 RETURN(PTR_ERR(tgt));
1752 rc = md_getattr_name(tgt->ltd_exp, &rid, NULL, NULL,
1753 1, valid | OBD_MD_FLCROSSREF,
1754 ea_size, suppgid, &req);
1755 ptlrpc_req_finished(*request);
1758 } else if (rc == -ERESTART) {
1759 LASSERT(*request != NULL);
1760 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1761 "Got -ERESTART during getattr!\n");
1762 ptlrpc_req_finished(*request);
1766 * Directory got split. Time to update local object and repeat
1767 * the request with proper MDS.
1769 rc = lmv_handle_split(exp, &rid);
1776 #define md_op_data_fid(op_data, fl) \
1777 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1778 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1779 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1780 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1783 static int lmv_early_cancel_slaves(struct obd_export *exp,
1784 struct md_op_data *op_data, int op_tgt,
1785 ldlm_mode_t mode, int bits, int flag)
1787 struct obd_device *obd = exp->exp_obd;
1788 struct lmv_obd *lmv = &obd->u.lmv;
1789 ldlm_policy_data_t policy = {{0}};
1790 struct lu_fid *op_fid;
1791 struct lu_fid *st_fid;
1792 struct lmv_tgt_desc *tgt;
1793 struct lmv_object *obj;
1798 op_fid = md_op_data_fid(op_data, flag);
1799 if (!fid_is_sane(op_fid))
1802 obj = lmv_object_find(obd, op_fid);
1806 policy.l_inodebits.bits = bits;
1807 for (i = 0; i < obj->lo_objcount; i++) {
1808 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1809 st_fid = &obj->lo_stripes[i].ls_fid;
1810 if (op_tgt != tgt->ltd_idx) {
1811 CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1812 PFID(st_fid), tgt->ltd_idx);
1813 rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1814 mode, LDLM_FL_ASYNC, NULL);
1816 GOTO(out_put_obj, rc);
1819 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1820 op_tgt, PFID(st_fid));
1822 * Do not cancel locks for operation target, they will
1823 * be handled later in underlaying layer when calling
1824 * function we run on behalf of.
1827 op_data->op_flags |= flag;
1832 lmv_object_put(obj);
1836 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1837 int op_tgt, ldlm_mode_t mode, int bits, int flag)
1839 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1840 struct obd_device *obd = exp->exp_obd;
1841 struct lmv_obd *lmv = &obd->u.lmv;
1842 struct lmv_tgt_desc *tgt;
1843 ldlm_policy_data_t policy = {{0}};
1844 struct lmv_object *obj;
1848 if (!fid_is_sane(fid))
1851 obj = lmv_object_find(obd, fid);
1853 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1855 lmv_object_put(obj);
1857 tgt = lmv_find_target(lmv, fid);
1859 RETURN(PTR_ERR(tgt));
1861 if (tgt->ltd_idx != op_tgt) {
1862 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1863 policy.l_inodebits.bits = bits;
1864 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1865 mode, LDLM_FL_ASYNC, NULL);
1868 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1870 op_data->op_flags |= flag;
1879 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1882 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1883 struct ptlrpc_request **request)
1885 struct obd_device *obd = exp->exp_obd;
1886 struct lmv_obd *lmv = &obd->u.lmv;
1887 struct lmv_tgt_desc *tgt;
1888 struct lmv_object *obj;
1895 rc = lmv_check_connect(obd);
1902 if (op_data->op_namelen != 0) {
1903 obj = lmv_object_find(obd, &op_data->op_fid2);
1905 sidx = raw_name2idx(obj->lo_hashtype,
1908 op_data->op_namelen);
1909 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
1910 mds = obj->lo_stripes[sidx].ls_mds;
1911 lmv_object_put(obj);
1913 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1918 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1919 PFID(&op_data->op_fid2), op_data->op_namelen,
1920 op_data->op_name, PFID(&op_data->op_fid1));
1922 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds);
1927 * Request from MDS to acquire i_links for inode by fid1.
1929 CDEBUG(D_INODE, "Inc i_nlinks for "DFID"\n",
1930 PFID(&op_data->op_fid1));
1933 CDEBUG(D_INODE, "Forward to mds #"LPU64" ("DFID")\n",
1934 mds, PFID(&op_data->op_fid1));
1936 op_data->op_fsuid = current->fsuid;
1937 op_data->op_fsgid = current->fsgid;
1938 op_data->op_cap = cfs_curproc_cap_pack();
1939 tgt = lmv_get_target(lmv, mds);
1941 if (op_data->op_namelen) {
1943 * Cancel UPDATE lock on child (fid1).
1945 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1946 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
1947 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1950 rc = md_link(tgt->ltd_exp, op_data, request);
1951 if (rc == -ERESTART) {
1952 LASSERT(*request != NULL);
1953 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1954 "Got -ERESTART during link!\n");
1955 ptlrpc_req_finished(*request);
1959 * Directory got split. Time to update local object and repeat
1960 * the request with proper MDS.
1962 rc = lmv_handle_split(exp, &op_data->op_fid2);
1970 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
1971 const char *old, int oldlen, const char *new, int newlen,
1972 struct ptlrpc_request **request)
1974 struct obd_device *obd = exp->exp_obd;
1975 struct lmv_obd *lmv = &obd->u.lmv;
1976 struct lmv_tgt_desc *src_tgt;
1977 struct lmv_tgt_desc *tgt_tgt;
1981 struct lmv_object *obj;
1986 CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
1987 oldlen, old, PFID(&op_data->op_fid1),
1988 newlen, new, PFID(&op_data->op_fid2));
1990 rc = lmv_check_connect(obd);
1996 * MDS with old dir entry is asking another MDS to create name
2000 "Create %*s(%d/%d) in "DFID" pointing "
2001 "to "DFID"\n", newlen, new, oldlen, newlen,
2002 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1));
2004 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds1);
2009 * Target directory can be split, sowe should forward request to
2012 obj = lmv_object_find(obd, &op_data->op_fid2);
2014 sidx = raw_name2idx(obj->lo_hashtype,
2016 (char *)new, newlen);
2017 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2018 CDEBUG(D_INODE, "Parent obj "DFID"\n",
2019 PFID(&op_data->op_fid2));
2020 lmv_object_put(obj);
2028 obj = lmv_object_find(obd, &op_data->op_fid1);
2030 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2031 (char *)old, oldlen);
2032 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2033 mds1 = obj->lo_stripes[sidx].ls_mds;
2034 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2035 lmv_object_put(obj);
2037 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2042 obj = lmv_object_find(obd, &op_data->op_fid2);
2045 * Directory is already split, so we have to forward request to
2048 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2049 (char *)new, newlen);
2051 mds2 = obj->lo_stripes[sidx].ls_mds;
2052 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2053 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2054 lmv_object_put(obj);
2056 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2062 op_data->op_fsuid = current->fsuid;
2063 op_data->op_fsgid = current->fsgid;
2064 op_data->op_cap = cfs_curproc_cap_pack();
2066 src_tgt = lmv_get_target(lmv, mds1);
2067 tgt_tgt = lmv_get_target(lmv, mds2);
2070 * LOOKUP lock on src child (fid3) should also be cancelled for
2071 * src_tgt in mdc_rename.
2073 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2076 * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2079 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2080 LCK_EX, MDS_INODELOCK_UPDATE,
2081 MF_MDC_CANCEL_FID2);
2084 * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2087 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2088 LCK_EX, MDS_INODELOCK_LOOKUP,
2089 MF_MDC_CANCEL_FID4);
2093 * Cancel all the locks on tgt child (fid4).
2096 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2097 LCK_EX, MDS_INODELOCK_FULL,
2098 MF_MDC_CANCEL_FID4);
2102 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2103 new, newlen, request);
2104 if (rc == -ERESTART) {
2105 LASSERT(*request != NULL);
2106 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2107 "Got -ERESTART during rename!\n");
2108 ptlrpc_req_finished(*request);
2112 * Directory got split. Time to update local object and repeat
2113 * the request with proper MDS.
2115 rc = lmv_handle_split(exp, &op_data->op_fid1);
2122 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2123 void *ea, int ealen, void *ea2, int ea2len,
2124 struct ptlrpc_request **request,
2125 struct md_open_data **mod)
2127 struct obd_device *obd = exp->exp_obd;
2128 struct lmv_obd *lmv = &obd->u.lmv;
2129 struct ptlrpc_request *req;
2130 struct lmv_tgt_desc *tgt;
2131 struct lmv_object *obj;
2136 rc = lmv_check_connect(obd);
2140 obj = lmv_object_find(obd, &op_data->op_fid1);
2142 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2143 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2144 obj ? ", split" : "");
2146 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2148 for (i = 0; i < obj->lo_objcount; i++) {
2149 op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2151 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2157 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2158 ea2, ea2len, &req, mod);
2160 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2162 * This is master object and this request should
2163 * be returned back to llite.
2167 ptlrpc_req_finished(req);
2173 lmv_object_put(obj);
2175 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2177 RETURN(PTR_ERR(tgt));
2179 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2180 ea2len, request, mod);
2185 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2186 struct obd_capa *oc, struct ptlrpc_request **request)
2188 struct obd_device *obd = exp->exp_obd;
2189 struct lmv_obd *lmv = &obd->u.lmv;
2190 struct lmv_tgt_desc *tgt;
2194 rc = lmv_check_connect(obd);
2198 tgt = lmv_find_target(lmv, fid);
2200 RETURN(PTR_ERR(tgt));
2202 rc = md_sync(tgt->ltd_exp, fid, oc, request);
2207 * Main purpose of LMV blocking ast is to remove split directory LMV
2208 * presentation object (struct lmv_object) attached to the lock being revoked.
2210 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2211 void *data, int flag)
2213 struct lustre_handle lockh;
2214 struct lmv_object *obj;
2219 case LDLM_CB_BLOCKING:
2220 ldlm_lock2handle(lock, &lockh);
2221 rc = ldlm_cli_cancel(&lockh);
2223 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2227 case LDLM_CB_CANCELING:
2229 * Time to drop cached attrs for split directory object
2231 obj = lock->l_ast_data;
2233 CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2234 ", master "DFID"\n",
2235 lock->l_resource->lr_name.name[3] == 1 ?
2236 "LOOKUP" : "UPDATE",
2237 lock->l_resource->lr_name.name[0],
2238 lock->l_resource->lr_name.name[1],
2239 PFID(&obj->lo_fid));
2240 lmv_object_put(obj);
2249 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2253 val = le64_to_cpu(*hash);
2255 val += MAX_HASH_SIZE;
2256 if (val != DIR_END_OFF)
2257 *hash = cpu_to_le64(val - hash_adj);
2260 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2263 struct obd_import *imp;
2266 * XXX: to get nid we assume that underlying obd device is mdc.
2268 imp = class_exp2cliimp(exp);
2269 id = imp->imp_connection->c_self + fid_flatten(fid);
2271 CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2272 imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2274 return id ^ (id >> 32);
2277 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2278 struct obd_capa *oc, __u64 offset64, struct page *page,
2279 struct ptlrpc_request **request)
2281 struct obd_device *obd = exp->exp_obd;
2282 struct lmv_obd *lmv = &obd->u.lmv;
2283 struct lu_fid rid = *fid;
2284 struct lmv_object *obj;
2294 struct lmv_stripe *los;
2295 struct lmv_tgt_desc *tgt;
2296 struct lu_dirpage *dp;
2297 struct lu_dirent *ent;
2302 rc = lmv_check_connect(obd);
2306 CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2309 * This case handle directory lookup in clustered metadata case (i.e.
2310 * split directory is located on multiple md servers.)
2311 * each server keeps directory entries for certain range of hashes.
2312 * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2313 * first server will keep records with hashes [ 0 ... MAX_HASH / N - 1],
2314 * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2316 * readdir can simply start reading entries from 0 - N server in
2317 * order but that will not scale well as all client will request dir in
2318 * to server in same order.
2319 * Following algorithm does optimization:
2320 * Instead of doing readdir in 1, 2, ...., N order, client with a
2321 * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2322 * (every client has rank R)
2323 * But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2324 * since client ask dir from MDS{R} client has pages with offsets
2325 * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2326 * on hash values that we get.
2328 obj = lmv_object_find_lock(obd, fid);
2330 nr = obj->lo_objcount;
2332 seg_size = MAX_HASH_SIZE;
2333 do_div(seg_size, nr);
2334 los = obj->lo_stripes;
2335 tgt = lmv_get_target(lmv, los[0].ls_mds);
2336 rank = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2338 do_div(tgt_tmp, seg_size);
2339 tgt0_idx = do_div(tgt_tmp, nr);
2340 tgt_idx = (tgt0_idx + rank) % nr;
2342 if (tgt_idx < tgt0_idx)
2346 * Last segment has unusual length due to division
2349 hash_adj = MAX_HASH_SIZE - seg_size * nr;
2353 hash_adj += rank * seg_size;
2355 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2356 LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2357 offset, tgt0_idx, offset + hash_adj, tgt_idx);
2359 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2360 rid = obj->lo_stripes[tgt_idx].ls_fid;
2361 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2363 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2364 PFID(&rid), (unsigned long)offset, tgt_idx);
2366 tgt = lmv_find_target(lmv, &rid);
2369 GOTO(cleanup, rc = PTR_ERR(tgt));
2371 rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2375 dp = cfs_kmap(page);
2377 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2378 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2379 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2381 for (ent = lu_dirent_start(dp); ent != NULL;
2382 ent = lu_dirent_next(ent))
2383 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2385 if (tgt0_idx != nr - 1) {
2388 end = le64_to_cpu(dp->ldp_hash_end);
2389 if (end == DIR_END_OFF) {
2390 dp->ldp_hash_end = cpu_to_le32(seg_size *
2393 ""DFID" reset end "LPX64" tgt %d\n",
2395 le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2403 lmv_object_put_unlock(obj);
2407 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2408 struct ptlrpc_request **request)
2410 struct obd_device *obd = exp->exp_obd;
2411 struct lmv_obd *lmv = &obd->u.lmv;
2412 struct lmv_tgt_desc *tgt = NULL;
2413 struct lmv_object *obj;
2418 rc = lmv_check_connect(obd);
2425 if (op_data->op_namelen != 0) {
2428 obj = lmv_object_find(obd, &op_data->op_fid1);
2430 sidx = raw_name2idx(obj->lo_hashtype,
2433 op_data->op_namelen);
2434 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2435 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2436 tgt = lmv_get_target(lmv,
2437 obj->lo_stripes[sidx].ls_mds);
2438 lmv_object_put(obj);
2439 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2440 op_data->op_namelen, op_data->op_name,
2441 PFID(&op_data->op_fid1), sidx);
2444 CDEBUG(D_INODE, "Drop i_nlink on "DFID"\n",
2445 PFID(&op_data->op_fid1));
2448 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2450 RETURN(PTR_ERR(tgt));
2451 op_data->op_bias |= MDS_CHECK_SPLIT;
2454 op_data->op_fsuid = current->fsuid;
2455 op_data->op_fsgid = current->fsgid;
2456 op_data->op_cap = cfs_curproc_cap_pack();
2459 * If child's fid is given, cancel unused locks for it if it is from
2460 * another export than parent.
2462 if (op_data->op_namelen) {
2464 * LOOKUP lock for child (fid3) should also be cancelled on
2465 * parent tgt_tgt in mdc_unlink().
2467 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2470 * Cancel FULL locks on child (fid3).
2472 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2473 MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2476 rc = md_unlink(tgt->ltd_exp, op_data, request);
2477 if (rc == -ERESTART) {
2478 LASSERT(*request != NULL);
2479 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2480 "Got -ERESTART during unlink!\n");
2481 ptlrpc_req_finished(*request);
2485 * Directory got split. Time to update local object and repeat
2486 * the request with proper MDS.
2488 rc = lmv_handle_split(exp, &op_data->op_fid1);
2495 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2500 case OBD_CLEANUP_EARLY:
2501 /* XXX: here should be calling obd_precleanup() down to
2504 case OBD_CLEANUP_EXPORTS:
2505 rc = obd_llog_finish(obd, 0);
2507 CERROR("failed to cleanup llogging subsystems\n");
2515 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2516 void *key, __u32 *vallen, void *val,
2517 struct lov_stripe_md *lsm)
2519 struct obd_device *obd;
2520 struct lmv_obd *lmv;
2524 obd = class_exp2obd(exp);
2526 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2527 exp->exp_handle.h_cookie);
2532 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2533 struct lmv_tgt_desc *tgts;
2536 rc = lmv_check_connect(obd);
2540 LASSERT(*vallen == sizeof(__u32));
2541 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2545 * All tgts should be connected when this gets called.
2547 if (!tgts || !tgts->ltd_exp) {
2548 CERROR("target not setup?\n");
2552 if (!obd_get_info(tgts->ltd_exp, keylen, key,
2557 } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2558 rc = lmv_check_connect(obd);
2563 * Forwarding this request to first MDS, it should know LOV
2566 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2568 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2569 exp->exp_connect_flags =
2570 ((struct obd_connect_data *)val)->ocd_connect_flags;
2575 CDEBUG(D_IOCTL, "Invalid key\n");
2579 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2580 void *key, obd_count vallen, void *val,
2581 struct ptlrpc_request_set *set)
2583 struct lmv_tgt_desc *tgt;
2584 struct obd_device *obd;
2585 struct lmv_obd *lmv;
2589 obd = class_exp2obd(exp);
2591 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2592 exp->exp_handle.h_cookie);
2597 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2598 KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2601 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2602 tgt = &lmv->tgts[i];
2607 err = obd_set_info_async(tgt->ltd_exp,
2608 keylen, key, vallen, val, set);
2619 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2620 struct lov_stripe_md *lsm)
2622 struct obd_device *obd = class_exp2obd(exp);
2623 struct lmv_obd *lmv = &obd->u.lmv;
2624 struct lmv_stripe_md *meap;
2625 struct lmv_stripe_md *lsmp;
2630 mea_size = lmv_get_easize(lmv);
2634 if (*lmmp && !lsm) {
2635 OBD_FREE(*lmmp, mea_size);
2640 if (*lmmp == NULL) {
2641 OBD_ALLOC(*lmmp, mea_size);
2649 lsmp = (struct lmv_stripe_md *)lsm;
2650 meap = (struct lmv_stripe_md *)*lmmp;
2652 if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2653 lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2656 meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2657 meap->mea_count = cpu_to_le32(lsmp->mea_count);
2658 meap->mea_master = cpu_to_le32(lsmp->mea_master);
2660 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2661 meap->mea_ids[i] = meap->mea_ids[i];
2662 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2668 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2669 struct lov_mds_md *lmm, int lmm_size)
2671 struct obd_device *obd = class_exp2obd(exp);
2672 struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2673 struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2674 struct lmv_obd *lmv = &obd->u.lmv;
2680 mea_size = lmv_get_easize(lmv);
2684 if (*lsmp != NULL && lmm == NULL) {
2685 OBD_FREE(*tmea, mea_size);
2690 LASSERT(mea_size == lmm_size);
2692 OBD_ALLOC(*tmea, mea_size);
2699 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2700 mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2701 mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2703 magic = le32_to_cpu(mea->mea_magic);
2706 * Old mea is not handled here.
2708 CERROR("Old not supportable EA is found\n");
2712 (*tmea)->mea_magic = magic;
2713 (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2714 (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2716 for (i = 0; i < (*tmea)->mea_count; i++) {
2717 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2718 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2723 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2724 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2725 int flags, void *opaque)
2727 struct obd_device *obd = exp->exp_obd;
2728 struct lmv_obd *lmv = &obd->u.lmv;
2734 LASSERT(fid != NULL);
2736 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2737 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2740 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2741 policy, mode, flags, opaque);
2748 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2750 struct obd_device *obd = exp->exp_obd;
2751 struct lmv_obd *lmv = &obd->u.lmv;
2754 rc = md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data);
2758 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2759 const struct lu_fid *fid, ldlm_type_t type,
2760 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2761 struct lustre_handle *lockh)
2763 struct obd_device *obd = exp->exp_obd;
2764 struct lmv_obd *lmv = &obd->u.lmv;
2769 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2772 * With CMD every object can have two locks in different namespaces:
2773 * lookup lock in space of mds storing direntry and update/open lock in
2774 * space of mds storing inode. Thus we check all targets, not only that
2775 * one fid was created in.
2777 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2778 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2779 type, policy, mode, lockh);
2787 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2788 struct obd_export *dt_exp, struct obd_export *md_exp,
2789 struct lustre_md *md)
2791 struct obd_device *obd = exp->exp_obd;
2792 struct lmv_obd *lmv = &obd->u.lmv;
2795 rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2799 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2801 struct obd_device *obd = exp->exp_obd;
2802 struct lmv_obd *lmv = &obd->u.lmv;
2806 obd_free_memmd(exp, (void *)&md->mea);
2807 RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2810 int lmv_set_open_replay_data(struct obd_export *exp,
2811 struct obd_client_handle *och,
2812 struct ptlrpc_request *open_req)
2814 struct obd_device *obd = exp->exp_obd;
2815 struct lmv_obd *lmv = &obd->u.lmv;
2816 struct lmv_tgt_desc *tgt;
2819 tgt = lmv_find_target(lmv, &och->och_fid);
2821 RETURN(PTR_ERR(tgt));
2823 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2826 int lmv_clear_open_replay_data(struct obd_export *exp,
2827 struct obd_client_handle *och)
2829 struct obd_device *obd = exp->exp_obd;
2830 struct lmv_obd *lmv = &obd->u.lmv;
2831 struct lmv_tgt_desc *tgt;
2834 tgt = lmv_find_target(lmv, &och->och_fid);
2836 RETURN(PTR_ERR(tgt));
2838 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2841 static int lmv_get_remote_perm(struct obd_export *exp,
2842 const struct lu_fid *fid,
2843 struct obd_capa *oc, __u32 suppgid,
2844 struct ptlrpc_request **request)
2846 struct obd_device *obd = exp->exp_obd;
2847 struct lmv_obd *lmv = &obd->u.lmv;
2848 struct lmv_tgt_desc *tgt;
2852 rc = lmv_check_connect(obd);
2856 tgt = lmv_find_target(lmv, fid);
2858 RETURN(PTR_ERR(tgt));
2860 rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2864 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2867 struct obd_device *obd = exp->exp_obd;
2868 struct lmv_obd *lmv = &obd->u.lmv;
2869 struct lmv_tgt_desc *tgt;
2873 rc = lmv_check_connect(obd);
2877 tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2879 RETURN(PTR_ERR(tgt));
2881 rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2885 int lmv_intent_getattr_async(struct obd_export *exp,
2886 struct md_enqueue_info *minfo,
2887 struct ldlm_enqueue_info *einfo)
2889 struct md_op_data *op_data = &minfo->mi_data;
2890 struct obd_device *obd = exp->exp_obd;
2891 struct lmv_obd *lmv = &obd->u.lmv;
2892 struct lmv_object *obj;
2893 struct lmv_tgt_desc *tgt;
2898 rc = lmv_check_connect(obd);
2902 if (!fid_is_sane(&op_data->op_fid2)) {
2903 obj = lmv_object_find(obd, &op_data->op_fid1);
2904 if (obj && op_data->op_namelen) {
2905 sidx = raw_name2idx(obj->lo_hashtype,
2907 (char *)op_data->op_name,
2908 op_data->op_namelen);
2909 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2910 tgt = lmv_get_target(lmv,
2911 obj->lo_stripes[sidx].ls_mds);
2913 "Choose slave dir ("DFID") -> mds #%d\n",
2914 PFID(&op_data->op_fid1), tgt->ltd_idx);
2916 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2919 lmv_object_put(obj);
2921 op_data->op_fid1 = op_data->op_fid2;
2922 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2923 op_data->op_bias = MDS_CROSS_REF;
2925 * Unfortunately, we have to lie to MDC/MDS to retrieve
2926 * attributes llite needs.
2928 if (minfo->mi_it.it_op & IT_LOOKUP)
2929 minfo->mi_it.it_op = IT_GETATTR;
2933 RETURN(PTR_ERR(tgt));
2935 rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2939 int lmv_revalidate_lock(struct obd_export *exp,
2940 struct lookup_intent *it,
2943 struct obd_device *obd = exp->exp_obd;
2944 struct lmv_obd *lmv = &obd->u.lmv;
2945 struct lmv_tgt_desc *tgt;
2949 rc = lmv_check_connect(obd);
2953 tgt = lmv_find_target(lmv, fid);
2955 RETURN(PTR_ERR(tgt));
2957 rc = md_revalidate_lock(tgt->ltd_exp, it, fid);
2962 struct obd_ops lmv_obd_ops = {
2963 .o_owner = THIS_MODULE,
2964 .o_setup = lmv_setup,
2965 .o_cleanup = lmv_cleanup,
2966 .o_precleanup = lmv_precleanup,
2967 .o_process_config = lmv_process_config,
2968 .o_connect = lmv_connect,
2969 .o_disconnect = lmv_disconnect,
2970 .o_statfs = lmv_statfs,
2971 .o_get_info = lmv_get_info,
2972 .o_set_info_async = lmv_set_info_async,
2973 .o_packmd = lmv_packmd,
2974 .o_unpackmd = lmv_unpackmd,
2975 .o_notify = lmv_notify,
2976 .o_get_uuid = lmv_get_uuid,
2977 .o_iocontrol = lmv_iocontrol,
2978 .o_fid_delete = lmv_fid_delete
2981 struct md_ops lmv_md_ops = {
2982 .m_getstatus = lmv_getstatus,
2983 .m_change_cbdata = lmv_change_cbdata,
2984 .m_close = lmv_close,
2985 .m_create = lmv_create,
2986 .m_done_writing = lmv_done_writing,
2987 .m_enqueue = lmv_enqueue,
2988 .m_getattr = lmv_getattr,
2989 .m_getxattr = lmv_getxattr,
2990 .m_getattr_name = lmv_getattr_name,
2991 .m_intent_lock = lmv_intent_lock,
2993 .m_rename = lmv_rename,
2994 .m_setattr = lmv_setattr,
2995 .m_setxattr = lmv_setxattr,
2997 .m_readpage = lmv_readpage,
2998 .m_unlink = lmv_unlink,
2999 .m_init_ea_size = lmv_init_ea_size,
3000 .m_cancel_unused = lmv_cancel_unused,
3001 .m_set_lock_data = lmv_set_lock_data,
3002 .m_lock_match = lmv_lock_match,
3003 .m_get_lustre_md = lmv_get_lustre_md,
3004 .m_free_lustre_md = lmv_free_lustre_md,
3005 .m_set_open_replay_data = lmv_set_open_replay_data,
3006 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3007 .m_renew_capa = lmv_renew_capa,
3008 .m_get_remote_perm = lmv_get_remote_perm,
3009 .m_intent_getattr_async = lmv_intent_getattr_async,
3010 .m_revalidate_lock = lmv_revalidate_lock
3013 int __init lmv_init(void)
3015 struct lprocfs_static_vars lvars;
3018 lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3019 sizeof(struct lmv_object),
3021 if (!lmv_object_cache) {
3022 CERROR("Error allocating lmv objects cache\n");
3026 lprocfs_lmv_init_vars(&lvars);
3027 rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3028 lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3030 cfs_mem_cache_destroy(lmv_object_cache);
3036 static void lmv_exit(void)
3038 class_unregister_type(LUSTRE_LMV_NAME);
3040 LASSERTF(atomic_read(&lmv_object_count) == 0,
3041 "Can't free lmv objects cache, %d object(s) busy\n",
3042 atomic_read(&lmv_object_count));
3043 cfs_mem_cache_destroy(lmv_object_cache);
3046 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3047 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3048 MODULE_LICENSE("GPL");
3050 module_init(lmv_init);
3051 module_exit(lmv_exit);