4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LMV
39 #include <linux/slab.h>
40 #include <linux/module.h>
41 #include <linux/init.h>
42 #include <linux/slab.h>
43 #include <linux/pagemap.h>
45 #include <asm/div64.h>
46 #include <linux/seq_file.h>
47 #include <linux/namei.h>
49 #include <liblustre.h>
52 #include <lustre_log.h>
53 #include <obd_support.h>
54 #include <lustre_lib.h>
55 #include <lustre_net.h>
56 #include <obd_class.h>
57 #include <lprocfs_status.h>
58 #include <lustre_lite.h>
59 #include <lustre_fid.h>
60 #include "lmv_internal.h"
63 cfs_mem_cache_t *lmv_object_cache;
64 cfs_atomic_t lmv_object_count = CFS_ATOMIC_INIT(0);
66 static void lmv_activate_target(struct lmv_obd *lmv,
67 struct lmv_tgt_desc *tgt,
70 if (tgt->ltd_active == activate)
73 tgt->ltd_active = activate;
74 lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
80 * -EINVAL : UUID can't be found in the LMV's target list
81 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
82 * -EBADF : The UUID is found, but the OBD of the wrong type (!)
84 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
87 struct lmv_tgt_desc *tgt;
88 struct obd_device *obd;
93 CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
94 lmv, uuid->uuid, activate);
96 cfs_spin_lock(&lmv->lmv_lock);
97 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
98 if (tgt->ltd_exp == NULL)
101 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n",
102 i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
104 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
108 if (i == lmv->desc.ld_tgt_count)
109 GOTO(out_lmv_lock, rc = -EINVAL);
111 obd = class_exp2obd(tgt->ltd_exp);
113 GOTO(out_lmv_lock, rc = -ENOTCONN);
115 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
116 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
117 obd->obd_type->typ_name, i);
118 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
120 if (tgt->ltd_active == activate) {
121 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
122 activate ? "" : "in");
123 GOTO(out_lmv_lock, rc);
126 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
127 activate ? "" : "in");
128 lmv_activate_target(lmv, tgt, activate);
132 cfs_spin_unlock(&lmv->lmv_lock);
136 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
137 struct obd_connect_data *data)
139 struct lmv_tgt_desc *tgt;
143 LASSERT(data != NULL);
145 cfs_spin_lock(&lmv->lmv_lock);
146 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
147 if (tgt->ltd_exp == NULL)
150 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
151 lmv->datas[tgt->ltd_idx] = *data;
155 cfs_spin_unlock(&lmv->lmv_lock);
159 struct obd_uuid *lmv_get_uuid(struct obd_export *exp) {
160 struct obd_device *obd = exp->exp_obd;
161 struct lmv_obd *lmv = &obd->u.lmv;
162 return obd_get_uuid(lmv->tgts[0].ltd_exp);
165 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
166 enum obd_notify_event ev, void *data)
168 struct obd_connect_data *conn_data;
169 struct lmv_obd *lmv = &obd->u.lmv;
170 struct obd_uuid *uuid;
174 if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
175 CERROR("unexpected notification of %s %s!\n",
176 watched->obd_type->typ_name,
181 uuid = &watched->u.cli.cl_target_uuid;
182 if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
184 * Set MDC as active before notifying the observer, so the
185 * observer can use the MDC normally.
187 rc = lmv_set_mdc_active(lmv, uuid,
188 ev == OBD_NOTIFY_ACTIVE);
190 CERROR("%sactivation of %s failed: %d\n",
191 ev == OBD_NOTIFY_ACTIVE ? "" : "de",
195 } else if (ev == OBD_NOTIFY_OCD) {
196 conn_data = &watched->u.cli.cl_import->imp_connect_data;
199 * Set connect data to desired target, update exp_connect_flags.
201 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
203 CERROR("can't set connect data to target %s, rc %d\n",
209 * XXX: Make sure that ocd_connect_flags from all targets are
210 * the same. Otherwise one of MDTs runs wrong version or
211 * something like this. --umka
213 obd->obd_self_export->exp_connect_flags =
214 conn_data->ocd_connect_flags;
217 else if (ev == OBD_NOTIFY_DISCON) {
219 * For disconnect event, flush fld cache for failout MDS case.
221 fld_client_flush(&lmv->lmv_fld);
225 * Pass the notification up the chain.
227 if (obd->obd_observer)
228 rc = obd_notify(obd->obd_observer, watched, ev, data);
234 * This is fake connect function. Its purpose is to initialize lmv and say
235 * caller that everything is okay. Real connection will be performed later.
237 static int lmv_connect(const struct lu_env *env,
238 struct obd_export **exp, struct obd_device *obd,
239 struct obd_uuid *cluuid, struct obd_connect_data *data,
243 struct proc_dir_entry *lmv_proc_dir;
245 struct lmv_obd *lmv = &obd->u.lmv;
246 struct lustre_handle conn = { 0 };
251 * We don't want to actually do the underlying connections more than
252 * once, so keep track.
255 if (lmv->refcount > 1) {
260 rc = class_connect(&conn, obd, cluuid);
262 CERROR("class_connection() returned %d\n", rc);
266 *exp = class_conn2export(&conn);
267 class_export_get(*exp);
271 lmv->cluuid = *cluuid;
274 lmv->conn_data = *data;
277 lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
279 if (IS_ERR(lmv_proc_dir)) {
280 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
281 obd->obd_type->typ_name, obd->obd_name);
287 * All real clients should perform actual connection right away, because
288 * it is possible, that LMV will not have opportunity to connect targets
289 * and MDC stuff will be called directly, for instance while reading
290 * ../mdc/../kbytesfree procfs file, etc.
292 if (data->ocd_connect_flags & OBD_CONNECT_REAL)
293 rc = lmv_check_connect(obd);
298 lprocfs_remove(&lmv_proc_dir);
305 static void lmv_set_timeouts(struct obd_device *obd)
307 struct lmv_tgt_desc *tgts;
312 if (lmv->server_timeout == 0)
315 if (lmv->connected == 0)
318 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
319 if (tgts->ltd_exp == NULL)
322 obd_set_info_async(NULL, tgts->ltd_exp, sizeof(KEY_INTERMDS),
323 KEY_INTERMDS, 0, NULL, NULL);
327 static int lmv_init_ea_size(struct obd_export *exp, int easize,
328 int def_easize, int cookiesize)
330 struct obd_device *obd = exp->exp_obd;
331 struct lmv_obd *lmv = &obd->u.lmv;
337 if (lmv->max_easize < easize) {
338 lmv->max_easize = easize;
341 if (lmv->max_def_easize < def_easize) {
342 lmv->max_def_easize = def_easize;
345 if (lmv->max_cookiesize < cookiesize) {
346 lmv->max_cookiesize = cookiesize;
352 if (lmv->connected == 0)
355 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
356 if (lmv->tgts[i].ltd_exp == NULL) {
357 CWARN("%s: NULL export for %d\n", obd->obd_name, i);
361 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
364 CERROR("obd_init_ea_size() failed on MDT target %d, "
365 "error %d.\n", i, rc);
372 #define MAX_STRING_SIZE 128
374 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
377 struct proc_dir_entry *lmv_proc_dir;
379 struct lmv_obd *lmv = &obd->u.lmv;
380 struct obd_uuid *cluuid = &lmv->cluuid;
381 struct obd_connect_data *mdc_data = NULL;
382 struct obd_uuid lmv_mdc_uuid = { "LMV_MDC_UUID" };
383 struct obd_device *mdc_obd;
384 struct obd_export *mdc_exp;
385 struct lu_fld_target target;
389 mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
392 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
396 CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
397 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
398 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
401 if (!mdc_obd->obd_set_up) {
402 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
406 rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
407 &lmv->conn_data, NULL);
409 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
414 * Init fid sequence client for this mdc and add new fld target.
416 rc = obd_fid_init(mdc_exp);
420 target.ft_srv = NULL;
421 target.ft_exp = mdc_exp;
422 target.ft_idx = tgt->ltd_idx;
424 fld_client_add_target(&lmv->lmv_fld, &target);
426 mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
428 rc = obd_register_observer(mdc_obd, obd);
430 obd_disconnect(mdc_exp);
431 CERROR("target %s register_observer error %d\n",
432 tgt->ltd_uuid.uuid, rc);
436 if (obd->obd_observer) {
438 * Tell the observer about the new target.
440 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
441 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
443 obd_disconnect(mdc_exp);
449 tgt->ltd_exp = mdc_exp;
450 lmv->desc.ld_active_tgt_count++;
453 * Copy connect data, it may be used later.
455 lmv->datas[tgt->ltd_idx] = *mdc_data;
457 md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
458 lmv->max_def_easize, lmv->max_cookiesize);
460 CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
461 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
462 cfs_atomic_read(&obd->obd_refcount));
465 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
467 struct proc_dir_entry *mdc_symlink;
469 LASSERT(mdc_obd->obd_type != NULL);
470 LASSERT(mdc_obd->obd_type->typ_name != NULL);
471 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
474 mdc_obd->obd_type->typ_name,
476 if (mdc_symlink == NULL) {
477 CERROR("Could not register LMV target "
478 "/proc/fs/lustre/%s/%s/target_obds/%s.",
479 obd->obd_type->typ_name, obd->obd_name,
481 lprocfs_remove(&lmv_proc_dir);
489 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
491 struct lmv_obd *lmv = &obd->u.lmv;
492 struct lmv_tgt_desc *tgt;
496 CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
500 if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
501 lmv_init_unlock(lmv);
502 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
503 "That many MDCs already configured.\n",
504 tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
507 if (lmv->desc.ld_tgt_count == 0) {
508 struct obd_device *mdc_obd;
510 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
513 lmv_init_unlock(lmv);
514 CERROR("Target %s not attached\n", tgt_uuid->uuid);
518 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, NULL);
520 lmv_init_unlock(lmv);
521 CERROR("lmv failed to setup llogging subsystems\n");
524 cfs_spin_lock(&lmv->lmv_lock);
525 tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
526 tgt->ltd_uuid = *tgt_uuid;
527 cfs_spin_unlock(&lmv->lmv_lock);
529 if (lmv->connected) {
530 rc = lmv_connect_mdc(obd, tgt);
532 cfs_spin_lock(&lmv->lmv_lock);
533 lmv->desc.ld_tgt_count--;
534 memset(tgt, 0, sizeof(*tgt));
535 cfs_spin_unlock(&lmv->lmv_lock);
537 int easize = sizeof(struct lmv_stripe_md) +
538 lmv->desc.ld_tgt_count *
539 sizeof(struct lu_fid);
540 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
544 lmv_init_unlock(lmv);
548 int lmv_check_connect(struct obd_device *obd)
550 struct lmv_obd *lmv = &obd->u.lmv;
551 struct lmv_tgt_desc *tgt;
561 if (lmv->connected) {
562 lmv_init_unlock(lmv);
566 if (lmv->desc.ld_tgt_count == 0) {
567 lmv_init_unlock(lmv);
568 CERROR("%s: no targets configured.\n", obd->obd_name);
572 CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
573 lmv->cluuid.uuid, obd->obd_name);
575 LASSERT(lmv->tgts != NULL);
577 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
578 rc = lmv_connect_mdc(obd, tgt);
583 lmv_set_timeouts(obd);
584 class_export_put(lmv->exp);
586 easize = lmv_get_easize(lmv);
587 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
588 lmv_init_unlock(lmv);
597 --lmv->desc.ld_active_tgt_count;
598 rc2 = obd_disconnect(tgt->ltd_exp);
600 CERROR("LMV target %s disconnect on "
601 "MDC idx %d: error %d\n",
602 tgt->ltd_uuid.uuid, i, rc2);
606 class_disconnect(lmv->exp);
607 lmv_init_unlock(lmv);
611 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
614 struct proc_dir_entry *lmv_proc_dir;
616 struct lmv_obd *lmv = &obd->u.lmv;
617 struct obd_device *mdc_obd;
621 LASSERT(tgt != NULL);
622 LASSERT(obd != NULL);
624 mdc_obd = class_exp2obd(tgt->ltd_exp);
627 mdc_obd->obd_force = obd->obd_force;
628 mdc_obd->obd_fail = obd->obd_fail;
629 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);
752 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
753 min((int) data->ioc_plen2,
754 (int) sizeof(struct obd_uuid))))
757 rc = obd_statfs(NULL, lmv->tgts[index].ltd_exp, &stat_buf,
758 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
762 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
763 min((int) data->ioc_plen1,
764 (int) sizeof(stat_buf))))
768 case OBD_IOC_QUOTACTL: {
769 struct if_quotactl *qctl = karg;
770 struct lmv_tgt_desc *tgt = NULL;
771 struct obd_quotactl *oqctl;
773 if (qctl->qc_valid == QC_MDTIDX) {
774 if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
777 tgt = &lmv->tgts[qctl->qc_idx];
780 } else if (qctl->qc_valid == QC_UUID) {
781 for (i = 0; i < count; i++) {
783 if (!obd_uuid_equals(&tgt->ltd_uuid,
787 if (tgt->ltd_exp == NULL)
799 LASSERT(tgt && tgt->ltd_exp);
800 OBD_ALLOC_PTR(oqctl);
804 QCTL_COPY(oqctl, qctl);
805 rc = obd_quotactl(tgt->ltd_exp, oqctl);
807 QCTL_COPY(qctl, oqctl);
808 qctl->qc_valid = QC_MDTIDX;
809 qctl->obd_uuid = tgt->ltd_uuid;
814 case OBD_IOC_CHANGELOG_SEND:
815 case OBD_IOC_CHANGELOG_CLEAR: {
816 struct ioc_changelog *icc = karg;
818 if (icc->icc_mdtindex >= count)
821 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
822 sizeof(*icc), icc, NULL);
825 case LL_IOC_GET_CONNECT_FLAGS: {
826 rc = obd_iocontrol(cmd, lmv->tgts[0].ltd_exp, len, karg, uarg);
831 for (i = 0; i < count; i++) {
833 struct obd_device *mdc_obd;
835 if (lmv->tgts[i].ltd_exp == NULL)
837 /* ll_umount_begin() sets force flag but for lmv, not
838 * mdc. Let's pass it through */
839 mdc_obd = class_exp2obd(lmv->tgts[i].ltd_exp);
840 mdc_obd->obd_force = obddev->obd_force;
841 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
843 if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
846 if (lmv->tgts[i].ltd_active) {
847 CERROR("error: iocontrol MDC %s on MDT"
848 "idx %d cmd %x: err = %d\n",
849 lmv->tgts[i].ltd_uuid.uuid,
864 static int lmv_all_chars_policy(int count, const char *name,
875 static int lmv_nid_policy(struct lmv_obd *lmv)
877 struct obd_import *imp;
881 * XXX: To get nid we assume that underlying obd device is mdc.
883 imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
884 id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
885 return id % lmv->desc.ld_tgt_count;
888 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
889 placement_policy_t placement)
892 case PLACEMENT_CHAR_POLICY:
893 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
895 op_data->op_namelen);
896 case PLACEMENT_NID_POLICY:
897 return lmv_nid_policy(lmv);
903 CERROR("Unsupported placement policy %x\n", placement);
908 * This is _inode_ placement policy function (not name).
910 static int lmv_placement_policy(struct obd_device *obd,
911 struct md_op_data *op_data,
914 struct lmv_obd *lmv = &obd->u.lmv;
915 struct lmv_object *obj;
919 LASSERT(mds != NULL);
921 if (lmv->desc.ld_tgt_count == 1) {
927 * Allocate new fid on target according to operation type and parent
930 obj = lmv_object_find(obd, &op_data->op_fid1);
931 if (obj != NULL || op_data->op_name == NULL ||
932 op_data->op_opc != LUSTRE_OPC_MKDIR) {
934 * Allocate fid for non-dir or for null name or for case parent
941 * If we have this flag turned on, and we see that
942 * parent dir is split, this means, that caller did not
943 * notice split yet. This is race and we would like to
944 * let caller know that.
946 if (op_data->op_bias & MDS_CHECK_SPLIT)
951 * Allocate new fid on same mds where parent fid is located and
952 * where operation will be sent. In case of split dir, ->op_fid1
953 * and ->op_mds here will contain fid and mds of slave directory
954 * object (assigned by caller).
956 *mds = op_data->op_mds;
960 * Parent directory is not split and we want to create a
961 * directory in it. Let's calculate where to place it according
962 * to operation data @op_data.
964 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
969 CERROR("Can't choose MDS, err = %d\n", rc);
971 LASSERT(*mds < lmv->desc.ld_tgt_count);
977 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
980 struct lmv_tgt_desc *tgt;
984 tgt = lmv_get_target(lmv, mds);
987 * New seq alloc and FLD setup should be atomic. Otherwise we may find
988 * on server that seq in new allocated fid is not yet known.
990 cfs_mutex_lock(&tgt->ltd_fid_mutex);
992 if (!tgt->ltd_active)
993 GOTO(out, rc = -ENODEV);
996 * Asking underlaying tgt layer to allocate new fid.
998 rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
1000 LASSERT(fid_is_sane(fid));
1006 cfs_mutex_unlock(&tgt->ltd_fid_mutex);
1010 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1011 struct md_op_data *op_data)
1013 struct obd_device *obd = class_exp2obd(exp);
1014 struct lmv_obd *lmv = &obd->u.lmv;
1019 LASSERT(op_data != NULL);
1020 LASSERT(fid != NULL);
1022 rc = lmv_placement_policy(obd, op_data, &mds);
1024 CERROR("Can't get target for allocating fid, "
1029 rc = __lmv_fid_alloc(lmv, fid, mds);
1031 CERROR("Can't alloc new fid, rc %d\n", rc);
1038 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1041 LASSERT(exp != NULL && fid != NULL);
1042 if (lmv_object_delete(exp, fid)) {
1043 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1049 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1051 struct lmv_obd *lmv = &obd->u.lmv;
1052 struct lprocfs_static_vars lvars;
1053 struct lmv_desc *desc;
1058 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1059 CERROR("LMV setup requires a descriptor\n");
1063 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1064 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1065 CERROR("Lmv descriptor size wrong: %d > %d\n",
1066 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1070 lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1072 OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1073 if (lmv->tgts == NULL)
1076 for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1077 cfs_mutex_init(&lmv->tgts[i].ltd_fid_mutex);
1078 lmv->tgts[i].ltd_idx = i;
1081 lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1083 OBD_ALLOC(lmv->datas, lmv->datas_size);
1084 if (lmv->datas == NULL)
1085 GOTO(out_free_tgts, rc = -ENOMEM);
1087 obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1088 lmv->desc.ld_tgt_count = 0;
1089 lmv->desc.ld_active_tgt_count = 0;
1090 lmv->max_cookiesize = 0;
1091 lmv->max_def_easize = 0;
1092 lmv->max_easize = 0;
1093 lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1095 cfs_spin_lock_init(&lmv->lmv_lock);
1096 cfs_mutex_init(&lmv->init_mutex);
1098 rc = lmv_object_setup(obd);
1100 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1101 GOTO(out_free_datas, rc);
1104 lprocfs_lmv_init_vars(&lvars);
1105 lprocfs_obd_setup(obd, lvars.obd_vars);
1108 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
1109 0444, &lmv_proc_target_fops, obd);
1111 CWARN("%s: error adding LMV target_obd file: rc = %d\n",
1115 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1116 LUSTRE_CLI_FLD_HASH_DHT);
1118 CERROR("Can't init FLD, err %d\n", rc);
1119 GOTO(out_free_datas, rc);
1125 OBD_FREE(lmv->datas, lmv->datas_size);
1128 OBD_FREE(lmv->tgts, lmv->tgts_size);
1133 static int lmv_cleanup(struct obd_device *obd)
1135 struct lmv_obd *lmv = &obd->u.lmv;
1138 fld_client_fini(&lmv->lmv_fld);
1139 lmv_object_cleanup(obd);
1140 OBD_FREE(lmv->datas, lmv->datas_size);
1141 OBD_FREE(lmv->tgts, lmv->tgts_size);
1146 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1148 struct lustre_cfg *lcfg = buf;
1149 struct obd_uuid tgt_uuid;
1153 switch(lcfg->lcfg_command) {
1155 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1156 GOTO(out, rc = -EINVAL);
1158 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1159 rc = lmv_add_target(obd, &tgt_uuid);
1162 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1163 GOTO(out, rc = -EINVAL);
1170 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1171 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1173 struct obd_device *obd = class_exp2obd(exp);
1174 struct lmv_obd *lmv = &obd->u.lmv;
1175 struct obd_statfs *temp;
1180 rc = lmv_check_connect(obd);
1184 OBD_ALLOC(temp, sizeof(*temp));
1188 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1189 if (lmv->tgts[i].ltd_exp == NULL)
1192 rc = obd_statfs(env, lmv->tgts[i].ltd_exp, temp,
1195 CERROR("can't stat MDS #%d (%s), error %d\n", i,
1196 lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1198 GOTO(out_free_temp, rc);
1203 osfs->os_bavail += temp->os_bavail;
1204 osfs->os_blocks += temp->os_blocks;
1205 osfs->os_ffree += temp->os_ffree;
1206 osfs->os_files += temp->os_files;
1212 OBD_FREE(temp, sizeof(*temp));
1216 static int lmv_getstatus(struct obd_export *exp,
1218 struct obd_capa **pc)
1220 struct obd_device *obd = exp->exp_obd;
1221 struct lmv_obd *lmv = &obd->u.lmv;
1225 rc = lmv_check_connect(obd);
1229 rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1233 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1234 struct obd_capa *oc, obd_valid valid, const char *name,
1235 const char *input, int input_size, int output_size,
1236 int flags, struct ptlrpc_request **request)
1238 struct obd_device *obd = exp->exp_obd;
1239 struct lmv_obd *lmv = &obd->u.lmv;
1240 struct lmv_tgt_desc *tgt;
1244 rc = lmv_check_connect(obd);
1248 tgt = lmv_find_target(lmv, fid);
1250 RETURN(PTR_ERR(tgt));
1252 rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1253 input_size, output_size, flags, request);
1258 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1259 struct obd_capa *oc, obd_valid valid, const char *name,
1260 const char *input, int input_size, int output_size,
1261 int flags, __u32 suppgid,
1262 struct ptlrpc_request **request)
1264 struct obd_device *obd = exp->exp_obd;
1265 struct lmv_obd *lmv = &obd->u.lmv;
1266 struct lmv_tgt_desc *tgt;
1270 rc = lmv_check_connect(obd);
1274 tgt = lmv_find_target(lmv, fid);
1276 RETURN(PTR_ERR(tgt));
1278 rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1279 input_size, output_size, flags, suppgid,
1285 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1286 struct ptlrpc_request **request)
1288 struct obd_device *obd = exp->exp_obd;
1289 struct lmv_obd *lmv = &obd->u.lmv;
1290 struct lmv_tgt_desc *tgt;
1291 struct lmv_object *obj;
1296 rc = lmv_check_connect(obd);
1300 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1302 RETURN(PTR_ERR(tgt));
1304 if (op_data->op_valid & OBD_MD_MDTIDX) {
1305 op_data->op_mds = tgt->ltd_idx;
1309 rc = md_getattr(tgt->ltd_exp, op_data, request);
1313 obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1315 CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1316 obj ? "(split)" : "");
1319 * If object is split, then we loop over all the slaves and gather size
1320 * attribute. In ideal world we would have to gather also mds field from
1321 * all slaves, as object is spread over the cluster and this is
1322 * definitely interesting information and it is not good to loss it,
1326 struct mdt_body *body;
1328 if (*request == NULL) {
1329 lmv_object_put(obj);
1333 body = req_capsule_server_get(&(*request)->rq_pill,
1335 LASSERT(body != NULL);
1337 for (i = 0; i < obj->lo_objcount; i++) {
1338 if (lmv->tgts[i].ltd_exp == NULL) {
1339 CWARN("%s: NULL export for %d\n",
1345 * Skip master object.
1347 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1350 body->size += obj->lo_stripes[i].ls_size;
1353 lmv_object_put_unlock(obj);
1359 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1360 ldlm_iterator_t it, void *data)
1362 struct obd_device *obd = exp->exp_obd;
1363 struct lmv_obd *lmv = &obd->u.lmv;
1368 rc = lmv_check_connect(obd);
1372 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1375 * With CMD every object can have two locks in different namespaces:
1376 * lookup lock in space of mds storing direntry and update/open lock in
1377 * space of mds storing inode.
1379 for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1380 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1385 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1386 ldlm_iterator_t it, void *data)
1388 struct obd_device *obd = exp->exp_obd;
1389 struct lmv_obd *lmv = &obd->u.lmv;
1394 rc = lmv_check_connect(obd);
1398 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1401 * With CMD every object can have two locks in different namespaces:
1402 * lookup lock in space of mds storing direntry and update/open lock in
1403 * space of mds storing inode.
1405 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1406 rc = md_find_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1415 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1416 struct md_open_data *mod, struct ptlrpc_request **request)
1418 struct obd_device *obd = exp->exp_obd;
1419 struct lmv_obd *lmv = &obd->u.lmv;
1420 struct lmv_tgt_desc *tgt;
1424 rc = lmv_check_connect(obd);
1428 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1430 RETURN(PTR_ERR(tgt));
1432 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1433 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1438 * Called in the case MDS returns -ERESTART on create on open, what means that
1439 * directory is split and its LMV presentation object has to be updated.
1441 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1443 struct obd_device *obd = exp->exp_obd;
1444 struct lmv_obd *lmv = &obd->u.lmv;
1445 struct ptlrpc_request *req = NULL;
1446 struct lmv_tgt_desc *tgt;
1447 struct lmv_object *obj;
1448 struct lustre_md md;
1449 struct md_op_data *op_data;
1456 mealen = lmv_get_easize(lmv);
1458 valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1460 tgt = lmv_find_target(lmv, fid);
1462 RETURN(PTR_ERR(tgt));
1465 * Time to update mea of parent fid.
1468 OBD_ALLOC_PTR(op_data);
1469 if (op_data == NULL)
1472 op_data->op_fid1 = *fid;
1473 op_data->op_mode = mealen;
1474 op_data->op_valid = valid;
1476 rc = md_getattr(tgt->ltd_exp, op_data, &req);
1477 OBD_FREE_PTR(op_data);
1479 CERROR("md_getattr() failed, error %d\n", rc);
1483 rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1485 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1490 GOTO(cleanup, rc = -ENODATA);
1492 obj = lmv_object_create(exp, fid, md.mea);
1496 lmv_object_put(obj);
1498 obd_free_memmd(exp, (void *)&md.mea);
1502 ptlrpc_req_finished(req);
1506 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1507 const void *data, int datalen, int mode, __u32 uid,
1508 __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1509 struct ptlrpc_request **request)
1511 struct obd_device *obd = exp->exp_obd;
1512 struct lmv_obd *lmv = &obd->u.lmv;
1513 struct lmv_tgt_desc *tgt;
1514 struct lmv_object *obj;
1520 rc = lmv_check_connect(obd);
1524 if (!lmv->desc.ld_active_tgt_count)
1530 obj = lmv_object_find(obd, &op_data->op_fid1);
1532 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1533 op_data->op_name, op_data->op_namelen);
1534 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1535 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1536 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1537 tgt = lmv_get_target(lmv, op_data->op_mds);
1538 lmv_object_put(obj);
1540 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1541 op_data->op_bias |= MDS_CHECK_SPLIT;
1542 op_data->op_mds = tgt->ltd_idx;
1546 RETURN(PTR_ERR(tgt));
1548 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1549 if (rc == -ERESTART)
1554 CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1555 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1558 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1559 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1560 cap_effective, rdev, request);
1562 if (*request == NULL)
1564 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1565 } else if (rc == -ERESTART) {
1566 LASSERT(*request != NULL);
1567 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1568 "Got -ERESTART during create!\n");
1569 ptlrpc_req_finished(*request);
1573 * Directory got split. Time to update local object and repeat
1574 * the request with proper MDS.
1576 rc = lmv_handle_split(exp, &op_data->op_fid1);
1578 rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1579 op_data, &op_data->op_fid2);
1588 static int lmv_done_writing(struct obd_export *exp,
1589 struct md_op_data *op_data,
1590 struct md_open_data *mod)
1592 struct obd_device *obd = exp->exp_obd;
1593 struct lmv_obd *lmv = &obd->u.lmv;
1594 struct lmv_tgt_desc *tgt;
1598 rc = lmv_check_connect(obd);
1602 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1604 RETURN(PTR_ERR(tgt));
1606 rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1611 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1612 struct lookup_intent *it, struct md_op_data *op_data,
1613 struct lustre_handle *lockh, void *lmm, int lmmsize)
1615 struct obd_device *obd = exp->exp_obd;
1616 struct lmv_obd *lmv = &obd->u.lmv;
1617 struct lmv_stripe_md *mea = op_data->op_mea1;
1618 struct md_op_data *op_data2;
1619 struct lmv_tgt_desc *tgt;
1624 OBD_ALLOC_PTR(op_data2);
1625 if (op_data2 == NULL)
1628 LASSERT(mea != NULL);
1629 for (i = 0; i < mea->mea_count; i++) {
1630 memset(op_data2, 0, sizeof(*op_data2));
1631 op_data2->op_fid1 = mea->mea_ids[i];
1632 op_data2->op_bias = 0;
1634 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1636 GOTO(cleanup, rc = PTR_ERR(tgt));
1638 if (tgt->ltd_exp == NULL)
1641 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1642 lockh + i, lmm, lmmsize, NULL, 0);
1644 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1645 PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1650 if (it->d.lustre.it_data) {
1651 struct ptlrpc_request *req;
1652 req = (struct ptlrpc_request *)it->d.lustre.it_data;
1653 ptlrpc_req_finished(req);
1656 if (it->d.lustre.it_status)
1657 GOTO(cleanup, rc = it->d.lustre.it_status);
1662 OBD_FREE_PTR(op_data2);
1666 * Drop all taken locks.
1669 if (lockh[i].cookie)
1670 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1671 lockh[i].cookie = 0;
1678 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1679 struct lookup_intent *it, struct md_op_data *op_data,
1680 struct lustre_handle *lockh, void *lmm, int lmmsize,
1681 int extra_lock_flags)
1683 struct ptlrpc_request *req = it->d.lustre.it_data;
1684 struct obd_device *obd = exp->exp_obd;
1685 struct lmv_obd *lmv = &obd->u.lmv;
1686 struct lustre_handle plock;
1687 struct lmv_tgt_desc *tgt;
1688 struct md_op_data *rdata;
1690 struct mdt_body *body;
1695 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1696 LASSERT(body != NULL);
1698 if (!(body->valid & OBD_MD_MDS))
1701 CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1702 LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1705 * We got LOOKUP lock, but we really need attrs.
1707 pmode = it->d.lustre.it_lock_mode;
1708 LASSERT(pmode != 0);
1709 memcpy(&plock, lockh, sizeof(plock));
1710 it->d.lustre.it_lock_mode = 0;
1711 it->d.lustre.it_data = NULL;
1714 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1715 ptlrpc_req_finished(req);
1717 tgt = lmv_find_target(lmv, &fid1);
1719 GOTO(out, rc = PTR_ERR(tgt));
1721 OBD_ALLOC_PTR(rdata);
1723 GOTO(out, rc = -ENOMEM);
1725 rdata->op_fid1 = fid1;
1726 rdata->op_bias = MDS_CROSS_REF;
1728 rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1729 lmm, lmmsize, NULL, extra_lock_flags);
1730 OBD_FREE_PTR(rdata);
1733 ldlm_lock_decref(&plock, pmode);
1738 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1739 struct lookup_intent *it, struct md_op_data *op_data,
1740 struct lustre_handle *lockh, void *lmm, int lmmsize,
1741 struct ptlrpc_request **req, int extra_lock_flags)
1743 struct obd_device *obd = exp->exp_obd;
1744 struct lmv_obd *lmv = &obd->u.lmv;
1745 struct lmv_tgt_desc *tgt;
1746 struct lmv_object *obj;
1751 rc = lmv_check_connect(obd);
1755 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1756 LL_IT2STR(it), PFID(&op_data->op_fid1));
1758 if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1759 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1760 lockh, lmm, lmmsize);
1764 obj = lmv_object_find(obd, &op_data->op_fid1);
1765 if (obj && op_data->op_namelen) {
1766 sidx = raw_name2idx(obj->lo_hashtype,
1768 (char *)op_data->op_name,
1769 op_data->op_namelen);
1770 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1771 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1773 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1776 lmv_object_put(obj);
1779 RETURN(PTR_ERR(tgt));
1781 CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1782 LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1784 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1785 lmm, lmmsize, req, extra_lock_flags);
1787 if (rc == 0 && it && it->it_op == IT_OPEN) {
1788 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1789 lmm, lmmsize, extra_lock_flags);
1795 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1796 struct ptlrpc_request **request)
1798 struct ptlrpc_request *req = NULL;
1799 struct obd_device *obd = exp->exp_obd;
1800 struct lmv_obd *lmv = &obd->u.lmv;
1801 struct lu_fid rid = op_data->op_fid1;
1802 struct lmv_tgt_desc *tgt;
1803 struct mdt_body *body;
1804 struct lmv_object *obj;
1805 obd_valid valid = op_data->op_valid;
1811 rc = lmv_check_connect(obd);
1818 obj = lmv_object_find(obd, &rid);
1820 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1821 op_data->op_name, op_data->op_namelen);
1822 rid = obj->lo_stripes[sidx].ls_fid;
1823 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1824 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1825 valid &= ~OBD_MD_FLCKSPLIT;
1826 lmv_object_put(obj);
1828 tgt = lmv_find_target(lmv, &rid);
1829 valid |= OBD_MD_FLCKSPLIT;
1830 op_data->op_mds = tgt->ltd_idx;
1833 RETURN(PTR_ERR(tgt));
1835 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1836 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1837 PFID(&rid), tgt->ltd_idx);
1839 op_data->op_valid = valid;
1840 op_data->op_fid1 = rid;
1841 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1843 body = req_capsule_server_get(&(*request)->rq_pill,
1845 LASSERT(body != NULL);
1847 if (body->valid & OBD_MD_MDS) {
1849 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1852 tgt = lmv_find_target(lmv, &rid);
1854 ptlrpc_req_finished(*request);
1855 RETURN(PTR_ERR(tgt));
1858 op_data->op_fid1 = rid;
1859 op_data->op_valid |= OBD_MD_FLCROSSREF;
1860 op_data->op_namelen = 0;
1861 op_data->op_name = NULL;
1862 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1863 ptlrpc_req_finished(*request);
1866 } else if (rc == -ERESTART) {
1867 LASSERT(*request != NULL);
1868 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1869 "Got -ERESTART during getattr!\n");
1870 ptlrpc_req_finished(*request);
1874 * Directory got split. Time to update local object and repeat
1875 * the request with proper MDS.
1877 rc = lmv_handle_split(exp, &rid);
1884 #define md_op_data_fid(op_data, fl) \
1885 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1886 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1887 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1888 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1891 static int lmv_early_cancel_slaves(struct obd_export *exp,
1892 struct md_op_data *op_data, int op_tgt,
1893 ldlm_mode_t mode, int bits, int flag)
1895 struct obd_device *obd = exp->exp_obd;
1896 struct lmv_obd *lmv = &obd->u.lmv;
1897 ldlm_policy_data_t policy = {{0}};
1898 struct lu_fid *op_fid;
1899 struct lu_fid *st_fid;
1900 struct lmv_tgt_desc *tgt;
1901 struct lmv_object *obj;
1906 op_fid = md_op_data_fid(op_data, flag);
1907 if (!fid_is_sane(op_fid))
1910 obj = lmv_object_find(obd, op_fid);
1914 policy.l_inodebits.bits = bits;
1915 for (i = 0; i < obj->lo_objcount; i++) {
1916 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1917 st_fid = &obj->lo_stripes[i].ls_fid;
1918 if (op_tgt != tgt->ltd_idx) {
1919 CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1920 PFID(st_fid), tgt->ltd_idx);
1921 rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1922 mode, LCF_ASYNC, NULL);
1924 GOTO(out_put_obj, rc);
1927 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1928 op_tgt, PFID(st_fid));
1930 * Do not cancel locks for operation target, they will
1931 * be handled later in underlaying layer when calling
1932 * function we run on behalf of.
1935 op_data->op_flags |= flag;
1940 lmv_object_put(obj);
1944 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1945 int op_tgt, ldlm_mode_t mode, int bits, int flag)
1947 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1948 struct obd_device *obd = exp->exp_obd;
1949 struct lmv_obd *lmv = &obd->u.lmv;
1950 struct lmv_tgt_desc *tgt;
1951 ldlm_policy_data_t policy = {{0}};
1952 struct lmv_object *obj;
1956 if (!fid_is_sane(fid))
1959 obj = lmv_object_find(obd, fid);
1961 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1963 lmv_object_put(obj);
1965 tgt = lmv_find_target(lmv, fid);
1967 RETURN(PTR_ERR(tgt));
1969 if (tgt->ltd_idx != op_tgt) {
1970 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1971 policy.l_inodebits.bits = bits;
1972 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1973 mode, LCF_ASYNC, NULL);
1976 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1978 op_data->op_flags |= flag;
1987 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1990 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1991 struct ptlrpc_request **request)
1993 struct obd_device *obd = exp->exp_obd;
1994 struct lmv_obd *lmv = &obd->u.lmv;
1995 struct lmv_tgt_desc *tgt;
1996 struct lmv_object *obj;
2003 rc = lmv_check_connect(obd);
2010 LASSERT(op_data->op_namelen != 0);
2012 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2013 PFID(&op_data->op_fid2), op_data->op_namelen,
2014 op_data->op_name, PFID(&op_data->op_fid1));
2016 obj = lmv_object_find(obd, &op_data->op_fid2);
2018 sidx = raw_name2idx(obj->lo_hashtype,
2021 op_data->op_namelen);
2022 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2023 mds = obj->lo_stripes[sidx].ls_mds;
2024 lmv_object_put(obj);
2026 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
2031 CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
2032 mds, PFID(&op_data->op_fid1));
2034 op_data->op_fsuid = cfs_curproc_fsuid();
2035 op_data->op_fsgid = cfs_curproc_fsgid();
2036 op_data->op_cap = cfs_curproc_cap_pack();
2037 tgt = lmv_get_target(lmv, mds);
2040 * Cancel UPDATE lock on child (fid1).
2042 op_data->op_flags |= MF_MDC_CANCEL_FID2;
2043 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2044 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2046 rc = md_link(tgt->ltd_exp, op_data, request);
2047 if (rc == -ERESTART) {
2048 LASSERT(*request != NULL);
2049 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2050 "Got -ERESTART during link!\n");
2051 ptlrpc_req_finished(*request);
2055 * Directory got split. Time to update local object and repeat
2056 * the request with proper MDS.
2058 rc = lmv_handle_split(exp, &op_data->op_fid2);
2066 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2067 const char *old, int oldlen, const char *new, int newlen,
2068 struct ptlrpc_request **request)
2070 struct obd_device *obd = exp->exp_obd;
2071 struct lmv_obd *lmv = &obd->u.lmv;
2072 struct lmv_tgt_desc *src_tgt;
2076 struct lmv_object *obj;
2081 LASSERT(oldlen != 0);
2083 CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2084 oldlen, old, PFID(&op_data->op_fid1),
2085 newlen, new, PFID(&op_data->op_fid2));
2087 rc = lmv_check_connect(obd);
2094 obj = lmv_object_find(obd, &op_data->op_fid1);
2096 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2097 (char *)old, oldlen);
2098 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2099 mds1 = obj->lo_stripes[sidx].ls_mds;
2100 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2101 lmv_object_put(obj);
2103 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2108 obj = lmv_object_find(obd, &op_data->op_fid2);
2111 * Directory is already split, so we have to forward request to
2114 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2115 (char *)new, newlen);
2117 mds2 = obj->lo_stripes[sidx].ls_mds;
2118 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2119 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2120 lmv_object_put(obj);
2122 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2127 op_data->op_fsuid = cfs_curproc_fsuid();
2128 op_data->op_fsgid = cfs_curproc_fsgid();
2129 op_data->op_cap = cfs_curproc_cap_pack();
2131 src_tgt = lmv_get_target(lmv, mds1);
2134 * LOOKUP lock on src child (fid3) should also be cancelled for
2135 * src_tgt in mdc_rename.
2137 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2140 * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2143 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2144 LCK_EX, MDS_INODELOCK_UPDATE,
2145 MF_MDC_CANCEL_FID2);
2148 * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2151 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2152 LCK_EX, MDS_INODELOCK_LOOKUP,
2153 MF_MDC_CANCEL_FID4);
2157 * Cancel all the locks on tgt child (fid4).
2160 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2161 LCK_EX, MDS_INODELOCK_FULL,
2162 MF_MDC_CANCEL_FID4);
2165 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2166 new, newlen, request);
2168 if (rc == -ERESTART) {
2169 LASSERT(*request != NULL);
2170 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2171 "Got -ERESTART during rename!\n");
2172 ptlrpc_req_finished(*request);
2176 * Directory got split. Time to update local object and repeat
2177 * the request with proper MDS.
2179 rc = lmv_handle_split(exp, &op_data->op_fid1);
2186 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2187 void *ea, int ealen, void *ea2, int ea2len,
2188 struct ptlrpc_request **request,
2189 struct md_open_data **mod)
2191 struct obd_device *obd = exp->exp_obd;
2192 struct lmv_obd *lmv = &obd->u.lmv;
2193 struct ptlrpc_request *req;
2194 struct lmv_tgt_desc *tgt;
2195 struct lmv_object *obj;
2200 rc = lmv_check_connect(obd);
2204 obj = lmv_object_find(obd, &op_data->op_fid1);
2206 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2207 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2208 obj ? ", split" : "");
2210 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2212 for (i = 0; i < obj->lo_objcount; i++) {
2213 op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2215 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2221 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2222 ea2, ea2len, &req, mod);
2224 if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2226 * This is master object and this request should
2227 * be returned back to llite.
2231 ptlrpc_req_finished(req);
2237 lmv_object_put(obj);
2239 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2241 RETURN(PTR_ERR(tgt));
2243 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2244 ea2len, request, mod);
2249 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2250 struct obd_capa *oc, struct ptlrpc_request **request)
2252 struct obd_device *obd = exp->exp_obd;
2253 struct lmv_obd *lmv = &obd->u.lmv;
2254 struct lmv_tgt_desc *tgt;
2258 rc = lmv_check_connect(obd);
2262 tgt = lmv_find_target(lmv, fid);
2264 RETURN(PTR_ERR(tgt));
2266 rc = md_sync(tgt->ltd_exp, fid, oc, request);
2271 * Main purpose of LMV blocking ast is to remove split directory LMV
2272 * presentation object (struct lmv_object) attached to the lock being revoked.
2274 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2275 void *data, int flag)
2277 struct lustre_handle lockh;
2278 struct lmv_object *obj;
2283 case LDLM_CB_BLOCKING:
2284 ldlm_lock2handle(lock, &lockh);
2285 rc = ldlm_cli_cancel(&lockh);
2287 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2291 case LDLM_CB_CANCELING:
2293 * Time to drop cached attrs for split directory object
2295 obj = lock->l_ast_data;
2297 CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2298 ", master "DFID"\n",
2299 lock->l_resource->lr_name.name[3] == 1 ?
2300 "LOOKUP" : "UPDATE",
2301 lock->l_resource->lr_name.name[0],
2302 lock->l_resource->lr_name.name[1],
2303 PFID(&obj->lo_fid));
2304 lmv_object_put(obj);
2313 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2317 val = le64_to_cpu(*hash);
2319 val += MAX_HASH_SIZE;
2320 if (val != MDS_DIR_END_OFF)
2321 *hash = cpu_to_le64(val - hash_adj);
2324 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2327 struct obd_import *imp;
2330 * XXX: to get nid we assume that underlying obd device is mdc.
2332 imp = class_exp2cliimp(exp);
2333 id = imp->imp_connection->c_self + fid_flatten(fid);
2335 CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2336 imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2338 return id ^ (id >> 32);
2341 static int lmv_readpage(struct obd_export *exp, struct md_op_data *op_data,
2342 struct page **pages, struct ptlrpc_request **request)
2344 struct obd_device *obd = exp->exp_obd;
2345 struct lmv_obd *lmv = &obd->u.lmv;
2346 struct lmv_object *obj;
2347 struct lu_fid rid = op_data->op_fid1;
2348 __u64 offset = op_data->op_offset;
2358 /* number of pages read, in CFS_PAGE_SIZE */
2360 /* number of pages transferred in LU_PAGE_SIZE */
2362 struct lmv_stripe *los;
2363 struct lmv_tgt_desc *tgt;
2364 struct lu_dirpage *dp;
2365 struct lu_dirent *ent;
2368 rc = lmv_check_connect(obd);
2372 CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2375 * This case handle directory lookup in clustered metadata case (i.e.
2376 * split directory is located on multiple md servers.)
2377 * each server keeps directory entries for certain range of hashes.
2378 * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2379 * first server will keep records with hashes [ 0 ... MAX_HASH / N - 1],
2380 * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2382 * readdir can simply start reading entries from 0 - N server in
2383 * order but that will not scale well as all client will request dir in
2384 * to server in same order.
2385 * Following algorithm does optimization:
2386 * Instead of doing readdir in 1, 2, ...., N order, client with a
2387 * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2388 * (every client has rank R)
2389 * But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2390 * since client ask dir from MDS{R} client has pages with offsets
2391 * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2392 * on hash values that we get.
2394 obj = lmv_object_find_lock(obd, &rid);
2396 nr = obj->lo_objcount;
2398 seg_size = MAX_HASH_SIZE;
2399 do_div(seg_size, nr);
2400 los = obj->lo_stripes;
2401 tgt = lmv_get_target(lmv, los[0].ls_mds);
2402 rank = lmv_node_rank(tgt->ltd_exp, &rid) % nr;
2404 do_div(tgt_tmp, seg_size);
2405 tgt0_idx = do_div(tgt_tmp, nr);
2406 tgt_idx = (tgt0_idx + rank) % nr;
2408 if (tgt_idx < tgt0_idx)
2412 * Last segment has unusual length due to division
2415 hash_adj = MAX_HASH_SIZE - seg_size * nr;
2419 hash_adj += rank * seg_size;
2421 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2422 LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2423 offset, tgt0_idx, offset + hash_adj, tgt_idx);
2425 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2426 rid = obj->lo_stripes[tgt_idx].ls_fid;
2427 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2429 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2430 PFID(&rid), (unsigned long)offset, tgt_idx);
2432 tgt = lmv_find_target(lmv, &rid);
2435 GOTO(cleanup, rc = PTR_ERR(tgt));
2437 op_data->op_fid1 = rid;
2438 rc = md_readpage(tgt->ltd_exp, op_data, pages, request);
2442 nrdpgs = ((*request)->rq_bulk->bd_nob_transferred + CFS_PAGE_SIZE - 1)
2444 nlupgs = (*request)->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
2445 LASSERT(!((*request)->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
2446 LASSERT(nrdpgs > 0 && nrdpgs <= op_data->op_npages);
2448 CDEBUG(D_INODE, "read %d(%d)/%d pages\n", nrdpgs, nlupgs,
2449 op_data->op_npages);
2451 for (i = 0; i < nrdpgs; i++) {
2452 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2453 struct lu_dirpage *first;
2457 struct lu_dirent *tmp = NULL;
2459 dp = cfs_kmap(pages[i]);
2461 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2462 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2463 LASSERT(le64_to_cpu(dp->ldp_hash_start) <=
2464 op_data->op_offset);
2466 if ((tgt0_idx != nr - 1) &&
2467 (le64_to_cpu(dp->ldp_hash_end) == MDS_DIR_END_OFF))
2469 dp->ldp_hash_end = cpu_to_le32(seg_size *
2472 ""DFID" reset end "LPX64" tgt %d\n",
2474 (__u64)le64_to_cpu(dp->ldp_hash_end),
2479 ent = lu_dirent_start(dp);
2480 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2482 hash_end = dp->ldp_hash_end;
2486 for (tmp = ent; ent != NULL;
2487 tmp = ent, ent = lu_dirent_next(ent)) {
2489 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2492 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2493 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
2494 if (((unsigned long)dp & ~CFS_PAGE_MASK) && nlupgs > 0) {
2495 ent = lu_dirent_start(dp);
2498 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2499 if ((tgt0_idx != nr - 1) &&
2500 (le64_to_cpu(dp->ldp_hash_end) ==
2502 hash_end = cpu_to_le32(seg_size *
2505 ""DFID" reset end "LPX64" tgt %d\n",
2507 (__u64)le64_to_cpu(hash_end),
2511 hash_end = dp->ldp_hash_end;
2512 flags = dp->ldp_flags;
2515 /* enlarge the end entry lde_reclen from 0 to
2516 * first entry of next lu_dirpage, in this way
2517 * several lu_dirpages can be stored into one
2518 * client page on client. */
2519 tmp = ((void *)tmp) +
2520 le16_to_cpu(tmp->lde_reclen);
2522 cpu_to_le16((char *)(dp->ldp_entries) -
2527 first->ldp_hash_end = hash_end;
2528 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
2529 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
2531 SET_BUT_UNUSED(tmp);
2533 cfs_kunmap(pages[i]);
2538 lmv_object_put_unlock(obj);
2542 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2543 struct ptlrpc_request **request)
2545 struct obd_device *obd = exp->exp_obd;
2546 struct lmv_obd *lmv = &obd->u.lmv;
2547 struct lmv_tgt_desc *tgt = NULL;
2548 struct lmv_object *obj;
2554 rc = lmv_check_connect(obd);
2561 LASSERT(op_data->op_namelen != 0);
2563 obj = lmv_object_find(obd, &op_data->op_fid1);
2565 sidx = raw_name2idx(obj->lo_hashtype,
2568 op_data->op_namelen);
2569 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2570 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2571 tgt = lmv_get_target(lmv,
2572 obj->lo_stripes[sidx].ls_mds);
2573 lmv_object_put(obj);
2574 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2575 op_data->op_namelen, op_data->op_name,
2576 PFID(&op_data->op_fid1), sidx);
2580 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2582 RETURN(PTR_ERR(tgt));
2583 op_data->op_bias |= MDS_CHECK_SPLIT;
2586 op_data->op_fsuid = cfs_curproc_fsuid();
2587 op_data->op_fsgid = cfs_curproc_fsgid();
2588 op_data->op_cap = cfs_curproc_cap_pack();
2591 * If child's fid is given, cancel unused locks for it if it is from
2592 * another export than parent.
2594 * LOOKUP lock for child (fid3) should also be cancelled on parent
2595 * tgt_tgt in mdc_unlink().
2597 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2600 * Cancel FULL locks on child (fid3).
2602 rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2603 MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2606 rc = md_unlink(tgt->ltd_exp, op_data, request);
2608 if (rc == -ERESTART) {
2609 LASSERT(*request != NULL);
2610 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2611 "Got -ERESTART during unlink!\n");
2612 ptlrpc_req_finished(*request);
2616 * Directory got split. Time to update local object and repeat
2617 * the request with proper MDS.
2619 rc = lmv_handle_split(exp, &op_data->op_fid1);
2626 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2628 struct lmv_obd *lmv = &obd->u.lmv;
2632 case OBD_CLEANUP_EARLY:
2633 /* XXX: here should be calling obd_precleanup() down to
2636 case OBD_CLEANUP_EXPORTS:
2637 fld_client_proc_fini(&lmv->lmv_fld);
2638 lprocfs_obd_cleanup(obd);
2639 rc = obd_llog_finish(obd, 0);
2641 CERROR("failed to cleanup llogging subsystems\n");
2649 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2650 __u32 keylen, void *key, __u32 *vallen, void *val,
2651 struct lov_stripe_md *lsm)
2653 struct obd_device *obd;
2654 struct lmv_obd *lmv;
2658 obd = class_exp2obd(exp);
2660 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2661 exp->exp_handle.h_cookie);
2666 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2667 struct lmv_tgt_desc *tgts;
2670 rc = lmv_check_connect(obd);
2674 LASSERT(*vallen == sizeof(__u32));
2675 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2679 * All tgts should be connected when this gets called.
2681 if (!tgts || !tgts->ltd_exp) {
2682 CERROR("target not setup?\n");
2686 if (!obd_get_info(env, tgts->ltd_exp, keylen, key,
2691 } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2692 rc = lmv_check_connect(obd);
2697 * Forwarding this request to first MDS, it should know LOV
2700 rc = obd_get_info(env, lmv->tgts[0].ltd_exp, keylen, key,
2702 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2703 exp->exp_connect_flags =
2704 ((struct obd_connect_data *)val)->ocd_connect_flags;
2707 } else if (KEY_IS(KEY_TGT_COUNT)) {
2708 *((int *)val) = lmv->desc.ld_tgt_count;
2712 CDEBUG(D_IOCTL, "Invalid key\n");
2716 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
2717 obd_count keylen, void *key, obd_count vallen,
2718 void *val, struct ptlrpc_request_set *set)
2720 struct lmv_tgt_desc *tgt;
2721 struct obd_device *obd;
2722 struct lmv_obd *lmv;
2726 obd = class_exp2obd(exp);
2728 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2729 exp->exp_handle.h_cookie);
2734 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX)) {
2737 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2738 tgt = &lmv->tgts[i];
2743 err = obd_set_info_async(env, tgt->ltd_exp,
2744 keylen, key, vallen, val, set);
2755 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2756 struct lov_stripe_md *lsm)
2758 struct obd_device *obd = class_exp2obd(exp);
2759 struct lmv_obd *lmv = &obd->u.lmv;
2760 struct lmv_stripe_md *meap;
2761 struct lmv_stripe_md *lsmp;
2766 mea_size = lmv_get_easize(lmv);
2770 if (*lmmp && !lsm) {
2771 OBD_FREE_LARGE(*lmmp, mea_size);
2776 if (*lmmp == NULL) {
2777 OBD_ALLOC_LARGE(*lmmp, mea_size);
2785 lsmp = (struct lmv_stripe_md *)lsm;
2786 meap = (struct lmv_stripe_md *)*lmmp;
2788 if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2789 lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2792 meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2793 meap->mea_count = cpu_to_le32(lsmp->mea_count);
2794 meap->mea_master = cpu_to_le32(lsmp->mea_master);
2796 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2797 meap->mea_ids[i] = meap->mea_ids[i];
2798 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2804 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2805 struct lov_mds_md *lmm, int lmm_size)
2807 struct obd_device *obd = class_exp2obd(exp);
2808 struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2809 struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2810 struct lmv_obd *lmv = &obd->u.lmv;
2816 mea_size = lmv_get_easize(lmv);
2820 if (*lsmp != NULL && lmm == NULL) {
2821 OBD_FREE_LARGE(*tmea, mea_size);
2826 LASSERT(mea_size == lmm_size);
2828 OBD_ALLOC_LARGE(*tmea, mea_size);
2835 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2836 mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2837 mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2839 magic = le32_to_cpu(mea->mea_magic);
2842 * Old mea is not handled here.
2844 CERROR("Old not supportable EA is found\n");
2848 (*tmea)->mea_magic = magic;
2849 (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2850 (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2852 for (i = 0; i < (*tmea)->mea_count; i++) {
2853 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2854 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2859 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2860 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2861 ldlm_cancel_flags_t flags, void *opaque)
2863 struct obd_device *obd = exp->exp_obd;
2864 struct lmv_obd *lmv = &obd->u.lmv;
2870 LASSERT(fid != NULL);
2872 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2873 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2876 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2877 policy, mode, flags, opaque);
2884 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2887 struct obd_device *obd = exp->exp_obd;
2888 struct lmv_obd *lmv = &obd->u.lmv;
2892 rc = md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2896 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2897 const struct lu_fid *fid, ldlm_type_t type,
2898 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2899 struct lustre_handle *lockh)
2901 struct obd_device *obd = exp->exp_obd;
2902 struct lmv_obd *lmv = &obd->u.lmv;
2907 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2910 * With CMD every object can have two locks in different namespaces:
2911 * lookup lock in space of mds storing direntry and update/open lock in
2912 * space of mds storing inode. Thus we check all targets, not only that
2913 * one fid was created in.
2915 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2916 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2917 type, policy, mode, lockh);
2925 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2926 struct obd_export *dt_exp, struct obd_export *md_exp,
2927 struct lustre_md *md)
2929 struct obd_device *obd = exp->exp_obd;
2930 struct lmv_obd *lmv = &obd->u.lmv;
2933 rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2937 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2939 struct obd_device *obd = exp->exp_obd;
2940 struct lmv_obd *lmv = &obd->u.lmv;
2944 obd_free_memmd(exp, (void *)&md->mea);
2945 RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2948 int lmv_set_open_replay_data(struct obd_export *exp,
2949 struct obd_client_handle *och,
2950 struct ptlrpc_request *open_req)
2952 struct obd_device *obd = exp->exp_obd;
2953 struct lmv_obd *lmv = &obd->u.lmv;
2954 struct lmv_tgt_desc *tgt;
2957 tgt = lmv_find_target(lmv, &och->och_fid);
2959 RETURN(PTR_ERR(tgt));
2961 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2964 int lmv_clear_open_replay_data(struct obd_export *exp,
2965 struct obd_client_handle *och)
2967 struct obd_device *obd = exp->exp_obd;
2968 struct lmv_obd *lmv = &obd->u.lmv;
2969 struct lmv_tgt_desc *tgt;
2972 tgt = lmv_find_target(lmv, &och->och_fid);
2974 RETURN(PTR_ERR(tgt));
2976 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2979 static int lmv_get_remote_perm(struct obd_export *exp,
2980 const struct lu_fid *fid,
2981 struct obd_capa *oc, __u32 suppgid,
2982 struct ptlrpc_request **request)
2984 struct obd_device *obd = exp->exp_obd;
2985 struct lmv_obd *lmv = &obd->u.lmv;
2986 struct lmv_tgt_desc *tgt;
2990 rc = lmv_check_connect(obd);
2994 tgt = lmv_find_target(lmv, fid);
2996 RETURN(PTR_ERR(tgt));
2998 rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
3002 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
3005 struct obd_device *obd = exp->exp_obd;
3006 struct lmv_obd *lmv = &obd->u.lmv;
3007 struct lmv_tgt_desc *tgt;
3011 rc = lmv_check_connect(obd);
3015 tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
3017 RETURN(PTR_ERR(tgt));
3019 rc = md_renew_capa(tgt->ltd_exp, oc, cb);
3023 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
3024 const struct req_msg_field *field, struct obd_capa **oc)
3026 struct obd_device *obd = exp->exp_obd;
3027 struct lmv_obd *lmv = &obd->u.lmv;
3031 rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
3035 int lmv_intent_getattr_async(struct obd_export *exp,
3036 struct md_enqueue_info *minfo,
3037 struct ldlm_enqueue_info *einfo)
3039 struct md_op_data *op_data = &minfo->mi_data;
3040 struct obd_device *obd = exp->exp_obd;
3041 struct lmv_obd *lmv = &obd->u.lmv;
3042 struct lmv_object *obj;
3043 struct lmv_tgt_desc *tgt = NULL;
3048 rc = lmv_check_connect(obd);
3052 if (op_data->op_namelen) {
3053 obj = lmv_object_find(obd, &op_data->op_fid1);
3055 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
3056 (char *)op_data->op_name,
3057 op_data->op_namelen);
3058 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
3059 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
3060 lmv_object_put(obj);
3065 tgt = lmv_find_target(lmv, &op_data->op_fid1);
3068 RETURN(PTR_ERR(tgt));
3070 rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
3074 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3075 struct lu_fid *fid, __u64 *bits)
3077 struct obd_device *obd = exp->exp_obd;
3078 struct lmv_obd *lmv = &obd->u.lmv;
3079 struct lmv_tgt_desc *tgt;
3083 rc = lmv_check_connect(obd);
3087 tgt = lmv_find_target(lmv, fid);
3089 RETURN(PTR_ERR(tgt));
3091 rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3096 * For lmv, only need to send request to master MDT, and the master MDT will
3097 * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3098 * we directly fetch data from the slave MDTs.
3100 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3101 struct obd_quotactl *oqctl)
3103 struct obd_device *obd = class_exp2obd(exp);
3104 struct lmv_obd *lmv = &obd->u.lmv;
3105 struct lmv_tgt_desc *tgt = &lmv->tgts[0];
3107 __u64 curspace, curinodes;
3110 if (!lmv->desc.ld_tgt_count || !tgt->ltd_active) {
3111 CERROR("master lmv inactive\n");
3115 if (oqctl->qc_cmd != Q_GETOQUOTA) {
3116 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3120 curspace = curinodes = 0;
3121 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3123 tgt = &lmv->tgts[i];
3125 if (tgt->ltd_exp == NULL)
3127 if (!tgt->ltd_active) {
3128 CDEBUG(D_HA, "mdt %d is inactive.\n", i);
3132 err = obd_quotactl(tgt->ltd_exp, oqctl);
3134 CERROR("getquota on mdt %d failed. %d\n", i, err);
3138 curspace += oqctl->qc_dqblk.dqb_curspace;
3139 curinodes += oqctl->qc_dqblk.dqb_curinodes;
3142 oqctl->qc_dqblk.dqb_curspace = curspace;
3143 oqctl->qc_dqblk.dqb_curinodes = curinodes;
3148 int lmv_quotacheck(struct obd_device *unused, struct obd_export *exp,
3149 struct obd_quotactl *oqctl)
3151 struct obd_device *obd = class_exp2obd(exp);
3152 struct lmv_obd *lmv = &obd->u.lmv;
3153 struct lmv_tgt_desc *tgt;
3157 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
3160 if (!tgt->ltd_active) {
3161 CERROR("lmv idx %d inactive\n", i);
3165 err = obd_quotacheck(tgt->ltd_exp, oqctl);
3173 struct obd_ops lmv_obd_ops = {
3174 .o_owner = THIS_MODULE,
3175 .o_setup = lmv_setup,
3176 .o_cleanup = lmv_cleanup,
3177 .o_precleanup = lmv_precleanup,
3178 .o_process_config = lmv_process_config,
3179 .o_connect = lmv_connect,
3180 .o_disconnect = lmv_disconnect,
3181 .o_statfs = lmv_statfs,
3182 .o_get_info = lmv_get_info,
3183 .o_set_info_async = lmv_set_info_async,
3184 .o_packmd = lmv_packmd,
3185 .o_unpackmd = lmv_unpackmd,
3186 .o_notify = lmv_notify,
3187 .o_get_uuid = lmv_get_uuid,
3188 .o_iocontrol = lmv_iocontrol,
3189 .o_fid_delete = lmv_fid_delete,
3190 .o_quotacheck = lmv_quotacheck,
3191 .o_quotactl = lmv_quotactl
3194 struct md_ops lmv_md_ops = {
3195 .m_getstatus = lmv_getstatus,
3196 .m_change_cbdata = lmv_change_cbdata,
3197 .m_find_cbdata = lmv_find_cbdata,
3198 .m_close = lmv_close,
3199 .m_create = lmv_create,
3200 .m_done_writing = lmv_done_writing,
3201 .m_enqueue = lmv_enqueue,
3202 .m_getattr = lmv_getattr,
3203 .m_getxattr = lmv_getxattr,
3204 .m_getattr_name = lmv_getattr_name,
3205 .m_intent_lock = lmv_intent_lock,
3207 .m_rename = lmv_rename,
3208 .m_setattr = lmv_setattr,
3209 .m_setxattr = lmv_setxattr,
3211 .m_readpage = lmv_readpage,
3212 .m_unlink = lmv_unlink,
3213 .m_init_ea_size = lmv_init_ea_size,
3214 .m_cancel_unused = lmv_cancel_unused,
3215 .m_set_lock_data = lmv_set_lock_data,
3216 .m_lock_match = lmv_lock_match,
3217 .m_get_lustre_md = lmv_get_lustre_md,
3218 .m_free_lustre_md = lmv_free_lustre_md,
3219 .m_set_open_replay_data = lmv_set_open_replay_data,
3220 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3221 .m_renew_capa = lmv_renew_capa,
3222 .m_unpack_capa = lmv_unpack_capa,
3223 .m_get_remote_perm = lmv_get_remote_perm,
3224 .m_intent_getattr_async = lmv_intent_getattr_async,
3225 .m_revalidate_lock = lmv_revalidate_lock
3228 int __init lmv_init(void)
3230 struct lprocfs_static_vars lvars;
3233 lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3234 sizeof(struct lmv_object),
3236 if (!lmv_object_cache) {
3237 CERROR("Error allocating lmv objects cache\n");
3241 lprocfs_lmv_init_vars(&lvars);
3243 rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3244 lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3246 cfs_mem_cache_destroy(lmv_object_cache);
3252 static void lmv_exit(void)
3254 class_unregister_type(LUSTRE_LMV_NAME);
3256 LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3257 "Can't free lmv objects cache, %d object(s) busy\n",
3258 cfs_atomic_read(&lmv_object_count));
3259 cfs_mem_cache_destroy(lmv_object_cache);
3262 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3263 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3264 MODULE_LICENSE("GPL");
3266 module_init(lmv_init);
3267 module_exit(lmv_exit);