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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
32 #define DEBUG_SUBSYSTEM S_LMV
34 #include <linux/file.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/user_namespace.h>
38 #include <linux/uidgid.h>
39 #include <linux/slab.h>
40 #include <linux/pagemap.h>
42 #include <linux/math64.h>
43 #include <linux/seq_file.h>
44 #include <linux/namei.h>
46 #include <obd_support.h>
47 #include <lustre_lib.h>
48 #include <lustre_net.h>
49 #include <obd_class.h>
50 #include <lustre_lmv.h>
51 #include <lprocfs_status.h>
52 #include <cl_object.h>
53 #include <lustre_fid.h>
54 #include <uapi/linux/lustre/lustre_ioctl.h>
55 #include <lustre_kernelcomm.h>
56 #include "lmv_internal.h"
58 static int lmv_check_connect(struct obd_device *obd);
60 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
63 if (tgt->ltd_active == activate)
66 tgt->ltd_active = activate;
67 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
70 tgt->ltd_exp->exp_obd->obd_inactive = !activate;
76 * -EINVAL : UUID can't be found in the LMV's target list
77 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
78 * -EBADF : The UUID is found, but the OBD of the wrong type (!)
80 static int lmv_set_mdc_active(struct lmv_obd *lmv,
81 const struct obd_uuid *uuid,
84 struct lu_tgt_desc *tgt = NULL;
85 struct obd_device *obd;
90 CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
91 lmv, uuid->uuid, activate);
93 spin_lock(&lmv->lmv_lock);
94 lmv_foreach_connected_tgt(lmv, tgt) {
95 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
96 tgt->ltd_index, tgt->ltd_uuid.uuid,
97 tgt->ltd_exp->exp_handle.h_cookie);
99 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
104 GOTO(out_lmv_lock, rc = -EINVAL);
106 obd = class_exp2obd(tgt->ltd_exp);
108 GOTO(out_lmv_lock, rc = -ENOTCONN);
110 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
111 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
112 obd->obd_type->typ_name, tgt->ltd_index);
113 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
115 if (tgt->ltd_active == activate) {
116 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
117 activate ? "" : "in");
118 GOTO(out_lmv_lock, rc);
121 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
122 activate ? "" : "in");
123 lmv_activate_target(lmv, tgt, activate);
127 spin_unlock(&lmv->lmv_lock);
131 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
133 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
134 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
136 return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
139 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
140 enum obd_notify_event ev)
142 struct obd_connect_data *conn_data;
143 struct lmv_obd *lmv = &obd->u.lmv;
144 struct obd_uuid *uuid;
148 if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
149 CERROR("unexpected notification of %s %s!\n",
150 watched->obd_type->typ_name,
155 uuid = &watched->u.cli.cl_target_uuid;
156 if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
158 * Set MDC as active before notifying the observer, so the
159 * observer can use the MDC normally.
161 rc = lmv_set_mdc_active(lmv, uuid,
162 ev == OBD_NOTIFY_ACTIVE);
164 CERROR("%sactivation of %s failed: %d\n",
165 ev == OBD_NOTIFY_ACTIVE ? "" : "de",
169 } else if (ev == OBD_NOTIFY_OCD) {
170 conn_data = &watched->u.cli.cl_import->imp_connect_data;
172 * XXX: Make sure that ocd_connect_flags from all targets are
173 * the same. Otherwise one of MDTs runs wrong version or
174 * something like this. --umka
176 obd->obd_self_export->exp_connect_data = *conn_data;
180 * Pass the notification up the chain.
182 if (obd->obd_observer)
183 rc = obd_notify(obd->obd_observer, watched, ev);
188 static int lmv_connect(const struct lu_env *env,
189 struct obd_export **pexp, struct obd_device *obd,
190 struct obd_uuid *cluuid, struct obd_connect_data *data,
193 struct lmv_obd *lmv = &obd->u.lmv;
194 struct lustre_handle conn = { 0 };
195 struct obd_export *exp;
199 rc = class_connect(&conn, obd, cluuid);
201 CERROR("class_connection() returned %d\n", rc);
205 exp = class_conn2export(&conn);
208 lmv->conn_data = *data;
209 lmv->lmv_cache = localdata;
211 lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
212 &obd->obd_kset.kobj);
213 if (!lmv->lmv_tgts_kobj) {
214 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
215 obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
218 rc = lmv_check_connect(obd);
227 if (lmv->lmv_tgts_kobj)
228 kobject_put(lmv->lmv_tgts_kobj);
230 class_disconnect(exp);
235 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
238 struct obd_device *obd = exp->exp_obd;
239 struct lmv_obd *lmv = &obd->u.lmv;
240 struct lmv_tgt_desc *tgt;
246 if (lmv->max_easize < easize) {
247 lmv->max_easize = easize;
250 if (lmv->max_def_easize < def_easize) {
251 lmv->max_def_easize = def_easize;
258 if (lmv->connected == 0)
261 lmv_foreach_connected_tgt(lmv, tgt) {
262 if (!tgt->ltd_active)
265 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
267 CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
268 " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
275 #define MAX_STRING_SIZE 128
277 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
279 struct lmv_obd *lmv = &obd->u.lmv;
280 struct obd_device *mdc_obd;
281 struct obd_export *mdc_exp;
282 struct lu_fld_target target;
286 mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
289 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
293 CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
294 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
295 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
297 if (!mdc_obd->obd_set_up) {
298 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
302 rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
303 &lmv->conn_data, lmv->lmv_cache);
305 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
310 * Init fid sequence client for this mdc and add new fld target.
312 rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
316 target.ft_srv = NULL;
317 target.ft_exp = mdc_exp;
318 target.ft_idx = tgt->ltd_index;
320 fld_client_add_target(&lmv->lmv_fld, &target);
322 rc = obd_register_observer(mdc_obd, obd);
324 obd_disconnect(mdc_exp);
325 CERROR("target %s register_observer error %d\n",
326 tgt->ltd_uuid.uuid, rc);
330 if (obd->obd_observer) {
332 * Tell the observer about the new target.
334 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
337 obd_disconnect(mdc_exp);
343 tgt->ltd_exp = mdc_exp;
344 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
346 md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
348 rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
350 obd_disconnect(mdc_exp);
354 CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
355 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
356 atomic_read(&obd->obd_refcount));
358 lmv_statfs_check_update(obd, tgt);
360 if (lmv->lmv_tgts_kobj)
361 /* Even if we failed to create the link, that's fine */
362 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
363 &mdc_obd->obd_kset.kobj,
368 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
371 ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
375 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
376 __u32 index, int gen)
378 struct obd_device *mdc_obd;
379 struct lmv_obd *lmv = &obd->u.lmv;
380 struct lmv_tgt_desc *tgt;
381 struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
386 CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
387 mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
390 CERROR("%s: Target %s not attached: rc = %d\n",
391 obd->obd_name, uuidp->uuid, -EINVAL);
399 mutex_init(&tgt->ltd_fid_mutex);
400 tgt->ltd_index = index;
401 tgt->ltd_uuid = *uuidp;
404 mutex_lock(<d->ltd_mutex);
405 rc = ltd_add_tgt(ltd, tgt);
406 mutex_unlock(<d->ltd_mutex);
412 /* lmv_check_connect() will connect this target. */
415 rc = lmv_connect_mdc(obd, tgt);
417 int easize = sizeof(struct lmv_stripe_md) +
418 lmv->lmv_mdt_count * sizeof(struct lu_fid);
420 lmv_init_ea_size(obd->obd_self_export, easize, 0);
430 static int lmv_check_connect(struct obd_device *obd)
432 struct lmv_obd *lmv = &obd->u.lmv;
433 struct lmv_tgt_desc *tgt;
442 mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
444 GOTO(unlock, rc = 0);
446 if (!lmv->lmv_mdt_count) {
447 CERROR("%s: no targets configured: rc = -EINVAL\n",
449 GOTO(unlock, rc = -EINVAL);
452 if (!lmv_mdt0_inited(lmv)) {
453 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
455 GOTO(unlock, rc = -EINVAL);
458 CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
459 obd->obd_uuid.uuid, obd->obd_name);
461 lmv_foreach_tgt(lmv, tgt) {
462 rc = lmv_connect_mdc(obd, tgt);
468 easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
469 lmv_init_ea_size(obd->obd_self_export, easize, 0);
472 mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
477 lmv_foreach_tgt(lmv, tgt) {
482 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
483 obd_disconnect(tgt->ltd_exp);
489 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
491 struct lmv_obd *lmv = &obd->u.lmv;
492 struct obd_device *mdc_obd;
496 LASSERT(tgt != NULL);
497 LASSERT(obd != NULL);
499 mdc_obd = class_exp2obd(tgt->ltd_exp);
502 mdc_obd->obd_force = obd->obd_force;
503 mdc_obd->obd_fail = obd->obd_fail;
504 mdc_obd->obd_no_recov = obd->obd_no_recov;
506 if (lmv->lmv_tgts_kobj)
507 sysfs_remove_link(lmv->lmv_tgts_kobj,
511 rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
513 CERROR("Can't finalize fids factory\n");
515 CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
516 tgt->ltd_exp->exp_obd->obd_name,
517 tgt->ltd_exp->exp_obd->obd_uuid.uuid);
519 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
520 rc = obd_disconnect(tgt->ltd_exp);
522 if (tgt->ltd_active) {
523 CERROR("Target %s disconnect error %d\n",
524 tgt->ltd_uuid.uuid, rc);
528 lmv_activate_target(lmv, tgt, 0);
533 static int lmv_disconnect(struct obd_export *exp)
535 struct obd_device *obd = class_exp2obd(exp);
536 struct lmv_obd *lmv = &obd->u.lmv;
537 struct lmv_tgt_desc *tgt;
542 lmv_foreach_connected_tgt(lmv, tgt)
543 lmv_disconnect_mdc(obd, tgt);
545 if (lmv->lmv_tgts_kobj)
546 kobject_put(lmv->lmv_tgts_kobj);
549 class_export_put(exp);
550 rc = class_disconnect(exp);
556 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
559 struct obd_device *obd = class_exp2obd(exp);
560 struct lmv_obd *lmv = &obd->u.lmv;
561 struct getinfo_fid2path *gf;
562 struct lmv_tgt_desc *tgt;
563 struct getinfo_fid2path *remote_gf = NULL;
564 struct lu_fid root_fid;
565 int remote_gf_size = 0;
569 tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
571 RETURN(PTR_ERR(tgt));
573 root_fid = *gf->gf_u.gf_root_fid;
574 LASSERT(fid_is_sane(&root_fid));
577 rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
578 if (rc != 0 && rc != -EREMOTE)
579 GOTO(out_fid2path, rc);
581 /* If remote_gf != NULL, it means just building the
582 * path on the remote MDT, copy this path segement to gf */
583 if (remote_gf != NULL) {
584 struct getinfo_fid2path *ori_gf;
588 ori_gf = (struct getinfo_fid2path *)karg;
589 if (strlen(ori_gf->gf_u.gf_path) + 1 +
590 strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
591 GOTO(out_fid2path, rc = -EOVERFLOW);
593 ptr = ori_gf->gf_u.gf_path;
595 len = strlen(gf->gf_u.gf_path);
596 /* move the current path to the right to release space
597 * for closer-to-root part */
598 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
599 memcpy(ptr, gf->gf_u.gf_path, len);
603 CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
604 tgt->ltd_exp->exp_obd->obd_name,
605 gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
609 GOTO(out_fid2path, rc);
611 /* sigh, has to go to another MDT to do path building further */
612 if (remote_gf == NULL) {
613 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
614 OBD_ALLOC(remote_gf, remote_gf_size);
615 if (remote_gf == NULL)
616 GOTO(out_fid2path, rc = -ENOMEM);
617 remote_gf->gf_pathlen = PATH_MAX;
620 if (!fid_is_sane(&gf->gf_fid)) {
621 CERROR("%s: invalid FID "DFID": rc = %d\n",
622 tgt->ltd_exp->exp_obd->obd_name,
623 PFID(&gf->gf_fid), -EINVAL);
624 GOTO(out_fid2path, rc = -EINVAL);
627 tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
629 GOTO(out_fid2path, rc = -EINVAL);
631 remote_gf->gf_fid = gf->gf_fid;
632 remote_gf->gf_recno = -1;
633 remote_gf->gf_linkno = -1;
634 memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
635 *remote_gf->gf_u.gf_root_fid = root_fid;
637 goto repeat_fid2path;
640 if (remote_gf != NULL)
641 OBD_FREE(remote_gf, remote_gf_size);
645 static int lmv_hsm_req_count(struct lmv_obd *lmv,
646 const struct hsm_user_request *hur,
647 const struct lmv_tgt_desc *tgt_mds)
649 struct lmv_tgt_desc *curr_tgt;
653 /* count how many requests must be sent to the given target */
654 for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
655 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
656 if (IS_ERR(curr_tgt))
657 RETURN(PTR_ERR(curr_tgt));
658 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
664 static int lmv_hsm_req_build(struct lmv_obd *lmv,
665 struct hsm_user_request *hur_in,
666 const struct lmv_tgt_desc *tgt_mds,
667 struct hsm_user_request *hur_out)
670 struct lmv_tgt_desc *curr_tgt;
672 /* build the hsm_user_request for the given target */
673 hur_out->hur_request = hur_in->hur_request;
675 for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
676 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
677 if (IS_ERR(curr_tgt))
678 RETURN(PTR_ERR(curr_tgt));
679 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
680 hur_out->hur_user_item[nr_out] =
681 hur_in->hur_user_item[i];
685 hur_out->hur_request.hr_itemcount = nr_out;
686 memcpy(hur_data(hur_out), hur_data(hur_in),
687 hur_in->hur_request.hr_data_len);
692 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
693 int len, struct lustre_kernelcomm *lk,
696 struct lmv_obd *lmv = &obd->u.lmv;
697 struct lu_tgt_desc *tgt;
702 /* unregister request (call from llapi_hsm_copytool_fini) */
703 lmv_foreach_connected_tgt(lmv, tgt)
704 /* best effort: try to clean as much as possible
705 * (continue on error) */
706 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
708 /* Whatever the result, remove copytool from kuc groups.
709 * Unreached coordinators will get EPIPE on next requests
710 * and will unregister automatically.
712 rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
717 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
718 int len, struct lustre_kernelcomm *lk,
721 struct lmv_obd *lmv = &obd->u.lmv;
723 bool any_set = false;
724 struct kkuc_ct_data *kcd;
726 struct lu_tgt_desc *tgt;
733 filp = fget(lk->lk_wfd);
737 if (lk->lk_flags & LK_FLG_DATANR)
738 kcd_size = offsetof(struct kkuc_ct_data,
739 kcd_archives[lk->lk_data_count]);
741 kcd_size = sizeof(*kcd);
743 OBD_ALLOC(kcd, kcd_size);
745 GOTO(err_fput, rc = -ENOMEM);
747 kcd->kcd_nr_archives = lk->lk_data_count;
748 if (lk->lk_flags & LK_FLG_DATANR) {
749 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
750 if (lk->lk_data_count > 0)
751 memcpy(kcd->kcd_archives, lk->lk_data,
752 sizeof(*kcd->kcd_archives) * lk->lk_data_count);
754 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
757 rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
758 lk->lk_group, kcd, kcd_size);
759 OBD_FREE(kcd, kcd_size);
763 /* All or nothing: try to register to all MDS.
764 * In case of failure, unregister from previous MDS,
765 * except if it because of inactive target. */
766 lmv_foreach_connected_tgt(lmv, tgt) {
767 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
769 if (tgt->ltd_active) {
770 /* permanent error */
771 CERROR("%s: iocontrol MDC %s on MDT"
772 " idx %d cmd %x: err = %d\n",
773 lmv2obd_dev(lmv)->obd_name,
774 tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
777 lk->lk_flags |= LK_FLG_STOP;
779 /* unregister from previous MDS */
780 lmv_foreach_connected_tgt(lmv, tgt) {
781 if (tgt->ltd_index >= i)
784 obd_iocontrol(cmd, tgt->ltd_exp, len,
787 GOTO(err_kkuc_rem, rc);
789 /* else: transient error.
790 * kuc will register to the missing MDT
798 /* no registration done: return error */
799 GOTO(err_kkuc_rem, rc = -ENOTCONN);
804 libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
811 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
812 int len, void *karg, void __user *uarg)
814 struct obd_device *obd = class_exp2obd(exp);
815 struct lmv_obd *lmv = &obd->u.lmv;
816 struct lu_tgt_desc *tgt = NULL;
818 __u32 count = lmv->lmv_mdt_count;
827 case IOC_OBD_STATFS: {
828 struct obd_ioctl_data *data = karg;
829 struct obd_device *mdc_obd;
830 struct obd_statfs stat_buf = {0};
833 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
835 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
838 tgt = lmv_tgt(lmv, index);
842 if (!tgt->ltd_active)
845 mdc_obd = class_exp2obd(tgt->ltd_exp);
850 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
851 min((int) data->ioc_plen2,
852 (int) sizeof(struct obd_uuid))))
855 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
856 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
860 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
861 min((int) data->ioc_plen1,
862 (int) sizeof(stat_buf))))
866 case OBD_IOC_QUOTACTL: {
867 struct if_quotactl *qctl = karg;
868 struct obd_quotactl *oqctl;
870 if (qctl->qc_valid == QC_MDTIDX) {
871 tgt = lmv_tgt(lmv, qctl->qc_idx);
872 } else if (qctl->qc_valid == QC_UUID) {
873 lmv_foreach_tgt(lmv, tgt) {
874 if (!obd_uuid_equals(&tgt->ltd_uuid,
887 if (!tgt || !tgt->ltd_exp)
890 OBD_ALLOC_PTR(oqctl);
894 QCTL_COPY(oqctl, qctl);
895 rc = obd_quotactl(tgt->ltd_exp, oqctl);
897 QCTL_COPY(qctl, oqctl);
898 qctl->qc_valid = QC_MDTIDX;
899 qctl->obd_uuid = tgt->ltd_uuid;
904 case LL_IOC_GET_CONNECT_FLAGS: {
905 tgt = lmv_tgt(lmv, 0);
907 if (tgt && tgt->ltd_exp)
908 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
911 case LL_IOC_FID2MDTIDX: {
912 struct lu_fid *fid = karg;
915 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
919 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
920 * point to user space memory for FID2MDTIDX. */
921 *(__u32 *)uarg = mdt_index;
924 case OBD_IOC_FID2PATH: {
925 rc = lmv_fid2path(exp, len, karg, uarg);
928 case LL_IOC_HSM_STATE_GET:
929 case LL_IOC_HSM_STATE_SET:
930 case LL_IOC_HSM_ACTION: {
931 struct md_op_data *op_data = karg;
933 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
935 RETURN(PTR_ERR(tgt));
937 if (tgt->ltd_exp == NULL)
940 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
943 case LL_IOC_HSM_PROGRESS: {
944 const struct hsm_progress_kernel *hpk = karg;
946 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
948 RETURN(PTR_ERR(tgt));
949 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
952 case LL_IOC_HSM_REQUEST: {
953 struct hsm_user_request *hur = karg;
954 unsigned int reqcount = hur->hur_request.hr_itemcount;
959 /* if the request is about a single fid
960 * or if there is a single MDS, no need to split
962 if (reqcount == 1 || count == 1) {
963 tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
965 RETURN(PTR_ERR(tgt));
966 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
968 /* split fid list to their respective MDS */
969 lmv_foreach_connected_tgt(lmv, tgt) {
972 struct hsm_user_request *req;
974 nr = lmv_hsm_req_count(lmv, hur, tgt);
977 if (nr == 0) /* nothing for this MDS */
980 /* build a request with fids for this MDS */
981 reqlen = offsetof(typeof(*hur),
983 + hur->hur_request.hr_data_len;
984 OBD_ALLOC_LARGE(req, reqlen);
987 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
989 GOTO(hsm_req_err, rc1);
990 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
993 if (rc1 != 0 && rc == 0)
995 OBD_FREE_LARGE(req, reqlen);
1000 case LL_IOC_LOV_SWAP_LAYOUTS: {
1001 struct md_op_data *op_data = karg;
1002 struct lmv_tgt_desc *tgt1, *tgt2;
1004 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1006 RETURN(PTR_ERR(tgt1));
1008 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1010 RETURN(PTR_ERR(tgt2));
1012 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1015 /* only files on same MDT can have their layouts swapped */
1016 if (tgt1->ltd_index != tgt2->ltd_index)
1019 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1022 case LL_IOC_HSM_CT_START: {
1023 struct lustre_kernelcomm *lk = karg;
1024 if (lk->lk_flags & LK_FLG_STOP)
1025 rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1027 rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1031 lmv_foreach_connected_tgt(lmv, tgt) {
1032 struct obd_device *mdc_obd;
1035 /* ll_umount_begin() sets force flag but for lmv, not
1036 * mdc. Let's pass it through */
1037 mdc_obd = class_exp2obd(tgt->ltd_exp);
1038 mdc_obd->obd_force = obd->obd_force;
1039 err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1041 if (tgt->ltd_active) {
1042 CERROR("error: iocontrol MDC %s on MDT"
1043 " idx %d cmd %x: err = %d\n",
1045 tgt->ltd_index, cmd, err);
1058 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1059 struct lu_fid *fid, struct md_op_data *op_data)
1061 struct obd_device *obd = class_exp2obd(exp);
1062 struct lmv_obd *lmv = &obd->u.lmv;
1063 struct lmv_tgt_desc *tgt;
1071 tgt = lmv_tgt(lmv, op_data->op_mds);
1075 if (!tgt->ltd_active || !tgt->ltd_exp)
1079 * New seq alloc and FLD setup should be atomic. Otherwise we may find
1080 * on server that seq in new allocated fid is not yet known.
1082 mutex_lock(&tgt->ltd_fid_mutex);
1083 rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1084 mutex_unlock(&tgt->ltd_fid_mutex);
1086 LASSERT(fid_is_sane(fid));
1093 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1095 struct lmv_obd *lmv = &obd->u.lmv;
1096 struct lmv_desc *desc;
1097 struct lnet_process_id lnet_id;
1103 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1104 CERROR("LMV setup requires a descriptor\n");
1108 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1109 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1110 CERROR("Lmv descriptor size wrong: %d > %d\n",
1111 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1115 obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1116 desc->ld_uuid.uuid);
1117 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1118 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1119 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1120 LMV_DESC_QOS_MAXAGE_DEFAULT;
1121 lmv->max_def_easize = 0;
1122 lmv->max_easize = 0;
1124 spin_lock_init(&lmv->lmv_lock);
1127 * initialize rr_index to lower 32bit of netid, so that client
1128 * can distribute subdirs evenly from the beginning.
1130 while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1131 if (lnet_id.nid != LNET_NID_LO_0) {
1132 lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1137 rc = lmv_tunables_init(obd);
1139 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1142 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1143 LUSTRE_CLI_FLD_HASH_DHT);
1145 CERROR("Can't init FLD, err %d\n", rc);
1147 rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1149 CWARN("%s: error initialize target table: rc = %d\n",
1155 static int lmv_cleanup(struct obd_device *obd)
1157 struct lmv_obd *lmv = &obd->u.lmv;
1158 struct lu_tgt_desc *tgt;
1159 struct lu_tgt_desc *tmp;
1163 fld_client_fini(&lmv->lmv_fld);
1164 lmv_foreach_tgt_safe(lmv, tgt, tmp)
1165 lmv_del_target(lmv, tgt);
1166 lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1171 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1173 struct lustre_cfg *lcfg = buf;
1174 struct obd_uuid obd_uuid;
1180 switch (lcfg->lcfg_command) {
1182 /* modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDT0000_UUID
1183 * 2:0 3:1 4:lustre-MDT0000-mdc_UUID */
1184 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1185 GOTO(out, rc = -EINVAL);
1187 obd_str2uuid(&obd_uuid, lustre_cfg_buf(lcfg, 1));
1189 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1190 GOTO(out, rc = -EINVAL);
1191 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1192 GOTO(out, rc = -EINVAL);
1193 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1196 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1197 GOTO(out, rc = -EINVAL);
1203 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1207 if (flags & OBD_STATFS_FOR_MDT0)
1210 if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1211 return lmv->lmv_statfs_start;
1213 /* choose initial MDT for this client */
1215 struct lnet_process_id lnet_id;
1216 if (LNetGetId(i, &lnet_id) == -ENOENT)
1219 if (lnet_id.nid != LNET_NID_LO_0) {
1220 /* We dont need a full 64-bit modulus, just enough
1221 * to distribute the requests across MDTs evenly.
1223 lmv->lmv_statfs_start = (u32)lnet_id.nid %
1229 return lmv->lmv_statfs_start;
1232 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1233 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1235 struct obd_device *obd = class_exp2obd(exp);
1236 struct lmv_obd *lmv = &obd->u.lmv;
1237 struct obd_statfs *temp;
1238 struct lu_tgt_desc *tgt;
1245 OBD_ALLOC(temp, sizeof(*temp));
1249 /* distribute statfs among MDTs */
1250 idx = lmv_select_statfs_mdt(lmv, flags);
1252 for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1253 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1254 tgt = lmv_tgt(lmv, idx);
1255 if (!tgt || !tgt->ltd_exp)
1258 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1259 flags | OBD_STATFS_NESTED);
1261 CERROR("%s: can't stat MDS #%d: rc = %d\n",
1262 tgt->ltd_exp->exp_obd->obd_name, i, rc);
1263 GOTO(out_free_temp, rc);
1266 if (temp->os_state & OS_STATFS_SUM ||
1267 flags == OBD_STATFS_FOR_MDT0) {
1268 /* reset to the last aggregated values
1269 * and don't sum with non-aggrated data */
1270 /* If the statfs is from mount, it needs to retrieve
1271 * necessary information from MDT0. i.e. mount does
1272 * not need the merged osfs from all of MDT. Also
1273 * clients can be mounted as long as MDT0 is in
1282 osfs->os_bavail += temp->os_bavail;
1283 osfs->os_blocks += temp->os_blocks;
1284 osfs->os_ffree += temp->os_ffree;
1285 osfs->os_files += temp->os_files;
1286 osfs->os_granted += temp->os_granted;
1292 OBD_FREE(temp, sizeof(*temp));
1296 static int lmv_statfs_update(void *cookie, int rc)
1298 struct obd_info *oinfo = cookie;
1299 struct obd_device *obd = oinfo->oi_obd;
1300 struct lmv_obd *lmv = &obd->u.lmv;
1301 struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1302 struct obd_statfs *osfs = oinfo->oi_osfs;
1305 * NB: don't deactivate TGT upon error, because we may not trigger async
1306 * statfs any longer, then there is no chance to activate TGT.
1309 spin_lock(&lmv->lmv_lock);
1310 tgt->ltd_statfs = *osfs;
1311 tgt->ltd_statfs_age = ktime_get_seconds();
1312 spin_unlock(&lmv->lmv_lock);
1313 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1319 /* update tgt statfs async if it's ld_qos_maxage old */
1320 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1322 struct obd_info oinfo = {
1325 .oi_cb_up = lmv_statfs_update,
1329 if (ktime_get_seconds() - tgt->ltd_statfs_age <
1330 obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1333 rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1338 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1341 struct obd_device *obd = exp->exp_obd;
1342 struct lmv_obd *lmv = &obd->u.lmv;
1343 struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1351 rc = md_get_root(tgt->ltd_exp, fileset, fid);
1355 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1356 u64 obd_md_valid, const char *name, size_t buf_size,
1357 struct ptlrpc_request **req)
1359 struct obd_device *obd = exp->exp_obd;
1360 struct lmv_obd *lmv = &obd->u.lmv;
1361 struct lmv_tgt_desc *tgt;
1366 tgt = lmv_fid2tgt(lmv, fid);
1368 RETURN(PTR_ERR(tgt));
1370 rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1375 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1376 u64 obd_md_valid, const char *name,
1377 const void *value, size_t value_size,
1378 unsigned int xattr_flags, u32 suppgid,
1379 struct ptlrpc_request **req)
1381 struct obd_device *obd = exp->exp_obd;
1382 struct lmv_obd *lmv = &obd->u.lmv;
1383 struct lmv_tgt_desc *tgt;
1388 tgt = lmv_fid2tgt(lmv, fid);
1390 RETURN(PTR_ERR(tgt));
1392 rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1393 value, value_size, xattr_flags, suppgid, req);
1398 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1399 struct ptlrpc_request **request)
1401 struct obd_device *obd = exp->exp_obd;
1402 struct lmv_obd *lmv = &obd->u.lmv;
1403 struct lmv_tgt_desc *tgt;
1408 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1410 RETURN(PTR_ERR(tgt));
1412 if (op_data->op_flags & MF_GET_MDT_IDX) {
1413 op_data->op_mds = tgt->ltd_index;
1417 rc = md_getattr(tgt->ltd_exp, op_data, request);
1422 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1424 struct obd_device *obd = exp->exp_obd;
1425 struct lmv_obd *lmv = &obd->u.lmv;
1426 struct lu_tgt_desc *tgt;
1430 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1433 * With DNE every object can have two locks in different namespaces:
1434 * lookup lock in space of MDT storing direntry and update/open lock in
1435 * space of MDT storing inode.
1437 lmv_foreach_connected_tgt(lmv, tgt)
1438 md_null_inode(tgt->ltd_exp, fid);
1443 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1444 struct md_open_data *mod, struct ptlrpc_request **request)
1446 struct obd_device *obd = exp->exp_obd;
1447 struct lmv_obd *lmv = &obd->u.lmv;
1448 struct lmv_tgt_desc *tgt;
1453 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1455 RETURN(PTR_ERR(tgt));
1457 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1458 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1462 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt,
1463 unsigned short dir_depth)
1465 struct lu_tgt_desc *tgt, *cur = NULL;
1466 __u64 total_avail = 0;
1467 __u64 total_weight = 0;
1468 __u64 cur_weight = 0;
1469 int total_usable = 0;
1475 if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1476 RETURN(ERR_PTR(-EAGAIN));
1478 down_write(&lmv->lmv_qos.lq_rw_sem);
1480 if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1481 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1483 rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1485 GOTO(unlock, tgt = ERR_PTR(rc));
1487 lmv_foreach_tgt(lmv, tgt) {
1488 if (!tgt->ltd_exp || !tgt->ltd_active) {
1489 tgt->ltd_qos.ltq_usable = 0;
1493 tgt->ltd_qos.ltq_usable = 1;
1494 lu_tgt_qos_weight_calc(tgt);
1495 if (tgt->ltd_index == *mdt)
1497 total_avail += tgt->ltd_qos.ltq_avail;
1498 total_weight += tgt->ltd_qos.ltq_weight;
1502 /* if current MDT has above-average space, within range of the QOS
1503 * threshold, stay on the same MDT to avoid creating needless remote
1504 * MDT directories. It's more likely for low level directories.
1506 rand = total_avail * (256 - lmv->lmv_qos.lq_threshold_rr) /
1507 (total_usable * 256 * (1 + dir_depth / 4));
1508 if (cur && cur->ltd_qos.ltq_avail >= rand) {
1510 GOTO(unlock, rc = 0);
1513 rand = lu_prandom_u64_max(total_weight);
1515 lmv_foreach_connected_tgt(lmv, tgt) {
1516 if (!tgt->ltd_qos.ltq_usable)
1519 cur_weight += tgt->ltd_qos.ltq_weight;
1520 if (cur_weight < rand)
1523 *mdt = tgt->ltd_index;
1524 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1525 GOTO(unlock, rc = 0);
1528 /* no proper target found */
1529 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1531 up_write(&lmv->lmv_qos.lq_rw_sem);
1536 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1538 struct lu_tgt_desc *tgt;
1544 spin_lock(&lmv->lmv_lock);
1545 for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1546 index = (i + lmv->lmv_qos_rr_index) %
1547 lmv->lmv_mdt_descs.ltd_tgts_size;
1548 tgt = lmv_tgt(lmv, index);
1549 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1552 *mdt = tgt->ltd_index;
1553 lmv->lmv_qos_rr_index = (*mdt + 1) %
1554 lmv->lmv_mdt_descs.ltd_tgts_size;
1555 spin_unlock(&lmv->lmv_lock);
1559 spin_unlock(&lmv->lmv_lock);
1561 RETURN(ERR_PTR(-ENODEV));
1564 /* locate MDT by file name, for striped directory, the file name hash decides
1565 * which stripe its dirent is stored.
1567 static struct lmv_tgt_desc *
1568 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1569 const char *name, int namelen, struct lu_fid *fid,
1570 __u32 *mds, bool new_layout)
1572 struct lmv_tgt_desc *tgt;
1573 const struct lmv_oinfo *oinfo;
1575 if (!lmv_dir_striped(lsm) || !namelen) {
1576 tgt = lmv_fid2tgt(lmv, fid);
1580 *mds = tgt->ltd_index;
1584 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1585 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1586 return ERR_PTR(-EBADF);
1587 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1589 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1591 return ERR_CAST(oinfo);
1594 /* check stripe FID is sane */
1595 if (!fid_is_sane(&oinfo->lmo_fid))
1596 return ERR_PTR(-ENODEV);
1598 *fid = oinfo->lmo_fid;
1599 *mds = oinfo->lmo_mds;
1600 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1602 CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1604 return tgt ? tgt : ERR_PTR(-ENODEV);
1608 * Locate MDT of op_data->op_fid1
1610 * For striped directory, it will locate the stripe by name hash, if hash_type
1611 * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1612 * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1613 * indicates whether old or new layout is used to locate.
1615 * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1617 * \param[in] lmv LMV device
1618 * \param[in/out] op_data client MD stack parameters, name, namelen etc,
1619 * op_mds and op_fid1 will be updated if op_mea1
1620 * indicates fid1 represents a striped directory.
1622 * retval pointer to the lmv_tgt_desc if succeed.
1623 * ERR_PTR(errno) if failed.
1625 struct lmv_tgt_desc *
1626 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1628 struct lmv_stripe_md *lsm = op_data->op_mea1;
1629 struct lmv_oinfo *oinfo;
1630 struct lmv_tgt_desc *tgt;
1632 if (lmv_dir_foreign(lsm))
1633 return ERR_PTR(-ENODATA);
1635 /* During creating VOLATILE file, it should honor the mdt
1636 * index if the file under striped dir is being restored, see
1638 if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1639 op_data->op_mds != LMV_OFFSET_DEFAULT) {
1640 tgt = lmv_tgt(lmv, op_data->op_mds);
1642 return ERR_PTR(-ENODEV);
1644 if (lmv_dir_striped(lsm)) {
1647 /* refill the right parent fid */
1648 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1649 oinfo = &lsm->lsm_md_oinfo[i];
1650 if (oinfo->lmo_mds == op_data->op_mds) {
1651 op_data->op_fid1 = oinfo->lmo_fid;
1656 if (i == lsm->lsm_md_stripe_count)
1657 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1659 } else if (lmv_dir_bad_hash(lsm)) {
1660 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1661 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1663 op_data->op_fid1 = oinfo->lmo_fid;
1664 op_data->op_mds = oinfo->lmo_mds;
1665 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1667 return ERR_PTR(-ENODEV);
1669 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1670 op_data->op_name, op_data->op_namelen,
1671 &op_data->op_fid1, &op_data->op_mds,
1672 op_data->op_new_layout);
1678 /* Locate MDT of op_data->op_fid2 for link/rename */
1679 static struct lmv_tgt_desc *
1680 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1682 struct lmv_tgt_desc *tgt;
1685 LASSERT(op_data->op_name);
1686 if (lmv_dir_layout_changing(op_data->op_mea2)) {
1687 struct lu_fid fid1 = op_data->op_fid1;
1688 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1689 struct ptlrpc_request *request = NULL;
1692 * avoid creating new file under old layout of migrating
1693 * directory, check it here.
1695 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1696 op_data->op_name, op_data->op_namelen,
1697 &op_data->op_fid2, &op_data->op_mds, false);
1701 op_data->op_fid1 = op_data->op_fid2;
1702 op_data->op_mea1 = op_data->op_mea2;
1703 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1704 op_data->op_fid1 = fid1;
1705 op_data->op_mea1 = lsm1;
1707 ptlrpc_req_finished(request);
1708 RETURN(ERR_PTR(-EEXIST));
1712 RETURN(ERR_PTR(rc));
1715 return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1716 op_data->op_namelen, &op_data->op_fid2,
1717 &op_data->op_mds, true);
1720 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1722 struct lu_tgt_desc *tgt;
1723 struct ptlrpc_request *request;
1726 LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1727 LASSERT(!op_data->op_new_layout);
1729 tgt = lmv_locate_tgt(lmv, op_data);
1731 return PTR_ERR(tgt);
1733 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1735 ptlrpc_req_finished(request);
1742 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1744 const struct lmv_user_md *lum = op_data->op_data;
1746 return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1747 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1748 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1751 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1753 const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1755 return (op_data->op_flags & MF_QOS_MKDIR) ||
1756 (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT);
1759 /* mkdir by QoS in three cases:
1760 * 1. ROOT default LMV is space balanced.
1761 * 2. 'lfs mkdir -i -1'
1762 * 3. parent default LMV master_mdt_index is -1
1764 * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1765 * directories under striped directory.
1767 static inline bool lmv_op_qos_mkdir(const struct md_op_data *op_data)
1769 if (op_data->op_code != LUSTRE_OPC_MKDIR)
1772 if (lmv_dir_striped(op_data->op_mea1))
1775 if (lmv_op_user_qos_mkdir(op_data))
1778 if (lmv_op_default_qos_mkdir(op_data))
1784 /* if parent default LMV is space balanced, and
1785 * 1. max_inherit_rr is set
1786 * 2. or parent is ROOT
1787 * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1788 * LMV requests roundrobin mkdir, do the same.
1789 * NB, this needs to check server is balanced, which is done by caller.
1791 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1793 const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1795 if (!lmv_op_default_qos_mkdir(op_data))
1798 return (op_data->op_flags & MF_RR_MKDIR) ||
1799 (lsm && lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE) ||
1800 fid_is_root(&op_data->op_fid1);
1803 /* 'lfs mkdir -i <specific_MDT>' */
1804 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1806 const struct lmv_user_md *lum = op_data->op_data;
1808 return op_data->op_code == LUSTRE_OPC_MKDIR &&
1809 op_data->op_cli_flags & CLI_SET_MEA && lum &&
1810 (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1811 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1812 le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1815 /* parent default LMV master_mdt_index is not -1. */
1817 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1819 return op_data->op_code == LUSTRE_OPC_MKDIR &&
1820 op_data->op_default_mea1 &&
1821 op_data->op_default_mea1->lsm_md_master_mdt_index !=
1825 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1826 const void *data, size_t datalen, umode_t mode, uid_t uid,
1827 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
1828 struct ptlrpc_request **request)
1830 struct obd_device *obd = exp->exp_obd;
1831 struct lmv_obd *lmv = &obd->u.lmv;
1832 struct lmv_tgt_desc *tgt;
1833 struct mdt_body *repbody;
1838 if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1841 if (lmv_dir_bad_hash(op_data->op_mea1))
1844 if (lmv_dir_layout_changing(op_data->op_mea1)) {
1846 * if parent is migrating, create() needs to lookup existing
1847 * name in both old and new layout, check old layout on client.
1849 rc = lmv_old_layout_lookup(lmv, op_data);
1853 op_data->op_new_layout = true;
1856 tgt = lmv_locate_tgt(lmv, op_data);
1858 RETURN(PTR_ERR(tgt));
1860 if (lmv_op_user_specific_mkdir(op_data)) {
1861 struct lmv_user_md *lum = op_data->op_data;
1863 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1864 tgt = lmv_tgt(lmv, op_data->op_mds);
1867 } else if (lmv_op_default_specific_mkdir(op_data)) {
1869 op_data->op_default_mea1->lsm_md_master_mdt_index;
1870 tgt = lmv_tgt(lmv, op_data->op_mds);
1873 } else if (lmv_op_qos_mkdir(op_data)) {
1874 struct lmv_tgt_desc *tmp = tgt;
1876 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds,
1877 op_data->op_dir_depth);
1878 if (tgt == ERR_PTR(-EAGAIN)) {
1879 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1880 !lmv_op_default_rr_mkdir(op_data) &&
1881 !lmv_op_user_qos_mkdir(op_data))
1882 /* if it's not necessary, don't create remote
1887 tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1890 RETURN(PTR_ERR(tgt));
1893 * only update statfs after QoS mkdir, this means the cached
1894 * statfs may be stale, and current mkdir may not follow QoS
1895 * accurately, but it's not serious, and avoids periodic statfs
1896 * when client doesn't mkdir by QoS.
1898 lmv_statfs_check_update(obd, tgt);
1902 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1906 CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1907 (int)op_data->op_namelen, op_data->op_name,
1908 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1911 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1912 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1913 cap_effective, rdev, request);
1915 if (*request == NULL)
1917 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1920 /* dir restripe needs to send to MDT where dir is located */
1921 if (rc != -EREMOTE ||
1922 !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
1925 repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
1926 if (repbody == NULL)
1929 /* Not cross-ref case, just get out of here. */
1930 if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
1933 op_data->op_fid2 = repbody->mbo_fid1;
1934 ptlrpc_req_finished(*request);
1937 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1939 RETURN(PTR_ERR(tgt));
1941 op_data->op_mds = tgt->ltd_index;
1946 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1947 const union ldlm_policy_data *policy, struct md_op_data *op_data,
1948 struct lustre_handle *lockh, __u64 extra_lock_flags)
1950 struct obd_device *obd = exp->exp_obd;
1951 struct lmv_obd *lmv = &obd->u.lmv;
1952 struct lmv_tgt_desc *tgt;
1957 CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1959 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1961 RETURN(PTR_ERR(tgt));
1963 CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1964 PFID(&op_data->op_fid1), tgt->ltd_index);
1966 rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1973 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1974 struct ptlrpc_request **preq)
1976 struct obd_device *obd = exp->exp_obd;
1977 struct lmv_obd *lmv = &obd->u.lmv;
1978 struct lmv_tgt_desc *tgt;
1979 struct mdt_body *body;
1985 if (op_data->op_namelen == 2 &&
1986 op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
1987 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1989 tgt = lmv_locate_tgt(lmv, op_data);
1991 RETURN(PTR_ERR(tgt));
1993 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1994 (int)op_data->op_namelen, op_data->op_name,
1995 PFID(&op_data->op_fid1), tgt->ltd_index);
1997 rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1998 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1999 ptlrpc_req_finished(*preq);
2007 body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2008 LASSERT(body != NULL);
2010 if (body->mbo_valid & OBD_MD_MDS) {
2011 op_data->op_fid1 = body->mbo_fid1;
2012 op_data->op_valid |= OBD_MD_FLCROSSREF;
2013 op_data->op_namelen = 0;
2014 op_data->op_name = NULL;
2016 ptlrpc_req_finished(*preq);
2025 #define md_op_data_fid(op_data, fl) \
2026 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2027 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2028 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2029 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2032 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2033 struct md_op_data *op_data, __u32 op_tgt,
2034 enum ldlm_mode mode, int bits, int flag)
2036 struct lu_fid *fid = md_op_data_fid(op_data, flag);
2037 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2038 union ldlm_policy_data policy = { { 0 } };
2042 if (!fid_is_sane(fid))
2046 tgt = lmv_fid2tgt(lmv, fid);
2048 RETURN(PTR_ERR(tgt));
2051 if (tgt->ltd_index != op_tgt) {
2052 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2053 policy.l_inodebits.bits = bits;
2054 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2055 mode, LCF_ASYNC, NULL);
2058 "EARLY_CANCEL skip operation target %d on "DFID"\n",
2060 op_data->op_flags |= flag;
2068 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2071 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2072 struct ptlrpc_request **request)
2074 struct obd_device *obd = exp->exp_obd;
2075 struct lmv_obd *lmv = &obd->u.lmv;
2076 struct lmv_tgt_desc *tgt;
2080 LASSERT(op_data->op_namelen != 0);
2082 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2083 PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2084 op_data->op_name, PFID(&op_data->op_fid1));
2086 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2087 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2088 op_data->op_cap = current_cap();
2090 tgt = lmv_locate_tgt2(lmv, op_data);
2092 RETURN(PTR_ERR(tgt));
2095 * Cancel UPDATE lock on child (fid1).
2097 op_data->op_flags |= MF_MDC_CANCEL_FID2;
2098 rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2099 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2103 rc = md_link(tgt->ltd_exp, op_data, request);
2108 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2109 const char *name, size_t namelen,
2110 struct ptlrpc_request **request)
2112 struct obd_device *obd = exp->exp_obd;
2113 struct lmv_obd *lmv = &obd->u.lmv;
2114 struct lmv_stripe_md *lsm = op_data->op_mea1;
2115 struct lmv_tgt_desc *parent_tgt;
2116 struct lmv_tgt_desc *sp_tgt;
2117 struct lmv_tgt_desc *tp_tgt = NULL;
2118 struct lmv_tgt_desc *child_tgt;
2119 struct lmv_tgt_desc *tgt;
2120 struct lu_fid target_fid = { 0 };
2125 LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2127 CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2128 PFID(&op_data->op_fid1), (int)namelen, name);
2130 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2131 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2132 op_data->op_cap = current_cap();
2134 parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2135 if (IS_ERR(parent_tgt))
2136 RETURN(PTR_ERR(parent_tgt));
2138 if (lmv_dir_striped(lsm)) {
2139 const struct lmv_oinfo *oinfo;
2141 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2143 RETURN(PTR_ERR(oinfo));
2145 /* save source stripe FID in fid4 temporarily for ELC */
2146 op_data->op_fid4 = oinfo->lmo_fid;
2147 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2152 * if parent is being migrated too, fill op_fid2 with target
2153 * stripe fid, otherwise the target stripe is not created yet.
2155 if (lmv_dir_layout_changing(lsm)) {
2156 oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2159 RETURN(PTR_ERR(oinfo));
2161 op_data->op_fid2 = oinfo->lmo_fid;
2162 tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2167 sp_tgt = parent_tgt;
2170 child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2171 if (IS_ERR(child_tgt))
2172 RETURN(PTR_ERR(child_tgt));
2174 /* for directory, migrate to MDT specified by lum_stripe_offset;
2175 * otherwise migrate to the target stripe of parent, but parent
2176 * directory may have finished migration (normally current file too),
2177 * allocate FID on MDT lum_stripe_offset, and server will check
2178 * whether file was migrated already.
2180 if (S_ISDIR(op_data->op_mode) || !tp_tgt) {
2181 struct lmv_user_md *lum = op_data->op_data;
2183 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2185 op_data->op_mds = tp_tgt->ltd_index;
2187 rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2192 * for directory, send migrate request to the MDT where the object will
2193 * be migrated to, because we can't create a striped directory remotely.
2195 * otherwise, send to the MDT where source is located because regular
2196 * file may open lease.
2198 * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2199 * backward compatibility.
2201 if (S_ISDIR(op_data->op_mode) &&
2202 (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2203 tgt = lmv_fid2tgt(lmv, &target_fid);
2205 RETURN(PTR_ERR(tgt));
2210 /* cancel UPDATE lock of parent master object */
2211 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2212 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2216 /* cancel UPDATE lock of source parent */
2217 if (sp_tgt != parent_tgt) {
2219 * migrate RPC packs master object FID, because we can only pack
2220 * two FIDs in reint RPC, but MDS needs to know both source
2221 * parent and target parent, and it will obtain them from master
2222 * FID and LMV, the other FID in RPC is kept for target.
2224 * since this FID is not passed to MDC, cancel it anyway.
2226 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2227 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2231 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2233 op_data->op_fid4 = target_fid;
2235 /* cancel UPDATE locks of target parent */
2236 rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2237 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2241 /* cancel LOOKUP lock of source if source is remote object */
2242 if (child_tgt != sp_tgt) {
2243 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2244 LCK_EX, MDS_INODELOCK_LOOKUP,
2245 MF_MDC_CANCEL_FID3);
2250 /* cancel ELC locks of source */
2251 rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2252 MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2256 rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2261 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2262 const char *old, size_t oldlen,
2263 const char *new, size_t newlen,
2264 struct ptlrpc_request **request)
2266 struct obd_device *obd = exp->exp_obd;
2267 struct lmv_obd *lmv = &obd->u.lmv;
2268 struct lmv_tgt_desc *sp_tgt;
2269 struct lmv_tgt_desc *tp_tgt = NULL;
2270 struct lmv_tgt_desc *src_tgt = NULL;
2271 struct lmv_tgt_desc *tgt;
2272 struct mdt_body *body;
2277 LASSERT(oldlen != 0);
2279 if (op_data->op_cli_flags & CLI_MIGRATE) {
2280 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2284 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2285 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2286 op_data->op_cap = current_cap();
2288 op_data->op_name = new;
2289 op_data->op_namelen = newlen;
2291 tp_tgt = lmv_locate_tgt2(lmv, op_data);
2293 RETURN(PTR_ERR(tp_tgt));
2295 /* Since the target child might be destroyed, and it might become
2296 * orphan, and we can only check orphan on the local MDT right now, so
2297 * we send rename request to the MDT where target child is located. If
2298 * target child does not exist, then it will send the request to the
2300 if (fid_is_sane(&op_data->op_fid4)) {
2301 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2303 RETURN(PTR_ERR(tgt));
2308 op_data->op_flags |= MF_MDC_CANCEL_FID4;
2310 /* cancel UPDATE locks of target parent */
2311 rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2312 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2316 if (fid_is_sane(&op_data->op_fid4)) {
2317 /* cancel LOOKUP lock of target on target parent */
2318 if (tgt != tp_tgt) {
2319 rc = lmv_early_cancel(exp, tp_tgt, op_data,
2320 tgt->ltd_index, LCK_EX,
2321 MDS_INODELOCK_LOOKUP,
2322 MF_MDC_CANCEL_FID4);
2328 if (fid_is_sane(&op_data->op_fid3)) {
2329 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2330 if (IS_ERR(src_tgt))
2331 RETURN(PTR_ERR(src_tgt));
2333 /* cancel ELC locks of source */
2334 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2335 LCK_EX, MDS_INODELOCK_ELC,
2336 MF_MDC_CANCEL_FID3);
2341 op_data->op_name = old;
2342 op_data->op_namelen = oldlen;
2344 sp_tgt = lmv_locate_tgt(lmv, op_data);
2346 RETURN(PTR_ERR(sp_tgt));
2348 /* cancel UPDATE locks of source parent */
2349 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2350 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2354 if (fid_is_sane(&op_data->op_fid3)) {
2355 /* cancel LOOKUP lock of source on source parent */
2356 if (src_tgt != sp_tgt) {
2357 rc = lmv_early_cancel(exp, sp_tgt, op_data,
2358 tgt->ltd_index, LCK_EX,
2359 MDS_INODELOCK_LOOKUP,
2360 MF_MDC_CANCEL_FID3);
2367 CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2368 PFID(&op_data->op_fid1), (int)oldlen, old,
2369 PFID(&op_data->op_fid2), (int)newlen, new);
2371 rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2373 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2374 ptlrpc_req_finished(*request);
2379 if (rc && rc != -EXDEV)
2382 body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2386 /* Not cross-ref case, just get out of here. */
2387 if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2390 op_data->op_fid4 = body->mbo_fid1;
2392 ptlrpc_req_finished(*request);
2395 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2397 RETURN(PTR_ERR(tgt));
2399 if (fid_is_sane(&op_data->op_fid4)) {
2400 /* cancel LOOKUP lock of target on target parent */
2401 if (tgt != tp_tgt) {
2402 rc = lmv_early_cancel(exp, tp_tgt, op_data,
2403 tgt->ltd_index, LCK_EX,
2404 MDS_INODELOCK_LOOKUP,
2405 MF_MDC_CANCEL_FID4);
2414 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2415 void *ea, size_t ealen, struct ptlrpc_request **request)
2417 struct obd_device *obd = exp->exp_obd;
2418 struct lmv_obd *lmv = &obd->u.lmv;
2419 struct lmv_tgt_desc *tgt;
2424 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2425 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2426 op_data->op_xvalid);
2428 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2429 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2431 RETURN(PTR_ERR(tgt));
2433 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2438 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2439 struct ptlrpc_request **request)
2441 struct obd_device *obd = exp->exp_obd;
2442 struct lmv_obd *lmv = &obd->u.lmv;
2443 struct lmv_tgt_desc *tgt;
2448 tgt = lmv_fid2tgt(lmv, fid);
2450 RETURN(PTR_ERR(tgt));
2452 rc = md_fsync(tgt->ltd_exp, fid, request);
2456 struct stripe_dirent {
2457 struct page *sd_page;
2458 struct lu_dirpage *sd_dp;
2459 struct lu_dirent *sd_ent;
2463 struct lmv_dir_ctxt {
2464 struct lmv_obd *ldc_lmv;
2465 struct md_op_data *ldc_op_data;
2466 struct md_callback *ldc_cb_op;
2469 struct stripe_dirent ldc_stripes[0];
2472 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2474 if (stripe->sd_page) {
2475 kunmap(stripe->sd_page);
2476 put_page(stripe->sd_page);
2477 stripe->sd_page = NULL;
2478 stripe->sd_ent = NULL;
2482 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2486 for (i = 0; i < ctxt->ldc_count; i++)
2487 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2490 /* if @ent is dummy, or . .., get next */
2491 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2492 struct lu_dirent *ent,
2495 for (; ent; ent = lu_dirent_next(ent)) {
2496 /* Skip dummy entry */
2497 if (le16_to_cpu(ent->lde_namelen) == 0)
2500 /* skip . and .. for other stripes */
2502 (strncmp(ent->lde_name, ".",
2503 le16_to_cpu(ent->lde_namelen)) == 0 ||
2504 strncmp(ent->lde_name, "..",
2505 le16_to_cpu(ent->lde_namelen)) == 0))
2508 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2515 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2516 struct stripe_dirent *stripe,
2519 struct md_op_data *op_data = ctxt->ldc_op_data;
2520 struct lmv_oinfo *oinfo;
2521 struct lu_fid fid = op_data->op_fid1;
2522 struct inode *inode = op_data->op_data;
2523 struct lmv_tgt_desc *tgt;
2524 struct lu_dirent *ent = stripe->sd_ent;
2525 __u64 hash = ctxt->ldc_hash;
2530 LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2534 if (stripe->sd_page) {
2535 __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2537 /* @hash should be the last dirent hash */
2538 LASSERTF(hash <= end,
2539 "ctxt@%p stripe@%p hash %llx end %llx\n",
2540 ctxt, stripe, hash, end);
2541 /* unload last page */
2542 stripe_dirent_unload(stripe);
2544 if (end == MDS_DIR_END_OFF) {
2545 stripe->sd_eof = true;
2551 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2552 if (!oinfo->lmo_root) {
2557 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2563 /* op_data is shared by stripes, reset after use */
2564 op_data->op_fid1 = oinfo->lmo_fid;
2565 op_data->op_fid2 = oinfo->lmo_fid;
2566 op_data->op_data = oinfo->lmo_root;
2568 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2571 op_data->op_fid1 = fid;
2572 op_data->op_fid2 = fid;
2573 op_data->op_data = inode;
2578 stripe->sd_dp = page_address(stripe->sd_page);
2579 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2581 /* in case a page filled with ., .. and dummy, read next */
2584 stripe->sd_ent = ent;
2587 /* treat error as eof, so dir can be partially accessed */
2588 stripe->sd_eof = true;
2589 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2590 "directory is partially accessed!\n",
2591 PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2598 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2600 struct obd_device *obd = exp->exp_obd;
2601 struct lmv_obd *lmv = &obd->u.lmv;
2602 struct lmv_tgt_desc *tgt;
2607 rc = lmv_check_connect(obd);
2611 tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2613 RETURN(PTR_ERR(tgt));
2615 data->op_flags |= MF_MDC_CANCEL_FID1;
2616 rc = md_file_resync(tgt->ltd_exp, data);
2621 * Get dirent with the closest hash for striped directory
2623 * This function will search the dir entry, whose hash value is the
2624 * closest(>=) to hash from all of sub-stripes, and it is only being called
2625 * for striped directory.
2627 * \param[in] ctxt dir read context
2629 * \retval dirent get the entry successfully
2630 * NULL does not get the entry, normally it means
2631 * it reaches the end of the directory, while read
2632 * stripe dirent error is ignored to allow partial
2635 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2637 struct stripe_dirent *stripe;
2638 struct lu_dirent *ent = NULL;
2642 /* TODO: optimize with k-way merge sort */
2643 for (i = 0; i < ctxt->ldc_count; i++) {
2644 stripe = &ctxt->ldc_stripes[i];
2648 if (!stripe->sd_ent) {
2649 stripe_dirent_load(ctxt, stripe, i);
2650 if (!stripe->sd_ent) {
2651 LASSERT(stripe->sd_eof);
2657 le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2658 le64_to_cpu(stripe->sd_ent->lde_hash)) {
2660 if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2667 stripe = &ctxt->ldc_stripes[min];
2668 ent = stripe->sd_ent;
2669 /* pop found dirent */
2670 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2678 * Build dir entry page for striped directory
2680 * This function gets one entry by @offset from a striped directory. It will
2681 * read entries from all of stripes, and choose one closest to the required
2682 * offset(&offset). A few notes
2683 * 1. skip . and .. for non-zero stripes, because there can only have one .
2684 * and .. in a directory.
2685 * 2. op_data will be shared by all of stripes, instead of allocating new
2686 * one, so need to restore before reusing.
2688 * \param[in] exp obd export refer to LMV
2689 * \param[in] op_data hold those MD parameters of read_entry
2690 * \param[in] cb_op ldlm callback being used in enqueue in mdc_read_entry
2691 * \param[in] offset starting hash offset
2692 * \param[out] ppage the page holding the entry. Note: because the entry
2693 * will be accessed in upper layer, so we need hold the
2694 * page until the usages of entry is finished, see
2695 * ll_dir_entry_next.
2697 * retval =0 if get entry successfully
2698 * <0 cannot get entry
2700 static int lmv_striped_read_page(struct obd_export *exp,
2701 struct md_op_data *op_data,
2702 struct md_callback *cb_op,
2703 __u64 offset, struct page **ppage)
2705 struct page *page = NULL;
2706 struct lu_dirpage *dp;
2708 struct lu_dirent *ent;
2709 struct lu_dirent *last_ent;
2711 struct lmv_dir_ctxt *ctxt;
2712 struct lu_dirent *next = NULL;
2718 /* Allocate a page and read entries from all of stripes and fill
2719 * the page by hash order */
2720 page = alloc_page(GFP_KERNEL);
2724 /* Initialize the entry page */
2726 memset(dp, 0, sizeof(*dp));
2727 dp->ldp_hash_start = cpu_to_le64(offset);
2730 left_bytes = PAGE_SIZE - sizeof(*dp);
2734 /* initalize dir read context */
2735 stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2736 OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2738 GOTO(free_page, rc = -ENOMEM);
2739 ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2740 ctxt->ldc_op_data = op_data;
2741 ctxt->ldc_cb_op = cb_op;
2742 ctxt->ldc_hash = offset;
2743 ctxt->ldc_count = stripe_count;
2746 next = lmv_dirent_next(ctxt);
2748 /* end of directory */
2750 ctxt->ldc_hash = MDS_DIR_END_OFF;
2753 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2755 ent_size = le16_to_cpu(next->lde_reclen);
2757 /* the last entry lde_reclen is 0, but it might not be the last
2758 * one of this temporay dir page */
2760 ent_size = lu_dirent_calc_size(
2761 le16_to_cpu(next->lde_namelen),
2762 le32_to_cpu(next->lde_attrs));
2764 if (ent_size > left_bytes)
2767 memcpy(ent, next, ent_size);
2769 /* Replace . with master FID and Replace .. with the parent FID
2770 * of master object */
2771 if (strncmp(ent->lde_name, ".",
2772 le16_to_cpu(ent->lde_namelen)) == 0 &&
2773 le16_to_cpu(ent->lde_namelen) == 1)
2774 fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2775 else if (strncmp(ent->lde_name, "..",
2776 le16_to_cpu(ent->lde_namelen)) == 0 &&
2777 le16_to_cpu(ent->lde_namelen) == 2)
2778 fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2780 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2781 le16_to_cpu(ent->lde_namelen), ent->lde_name,
2782 le64_to_cpu(ent->lde_hash));
2784 left_bytes -= ent_size;
2785 ent->lde_reclen = cpu_to_le16(ent_size);
2787 ent = (void *)ent + ent_size;
2790 last_ent->lde_reclen = 0;
2793 dp->ldp_flags |= LDF_EMPTY;
2794 else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2795 dp->ldp_flags |= LDF_COLLIDE;
2796 dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2797 dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2799 put_lmv_dir_ctxt(ctxt);
2800 OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2813 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2814 struct md_callback *cb_op, __u64 offset,
2815 struct page **ppage)
2817 struct obd_device *obd = exp->exp_obd;
2818 struct lmv_obd *lmv = &obd->u.lmv;
2819 struct lmv_tgt_desc *tgt;
2824 if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2827 if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2828 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2832 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2834 RETURN(PTR_ERR(tgt));
2836 rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2842 * Unlink a file/directory
2844 * Unlink a file or directory under the parent dir. The unlink request
2845 * usually will be sent to the MDT where the child is located, but if
2846 * the client does not have the child FID then request will be sent to the
2847 * MDT where the parent is located.
2849 * If the parent is a striped directory then it also needs to locate which
2850 * stripe the name of the child is located, and replace the parent FID
2851 * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2852 * it will walk through all of sub-stripes until the child is being
2855 * \param[in] exp export refer to LMV
2856 * \param[in] op_data different parameters transferred beween client
2857 * MD stacks, name, namelen, FIDs etc.
2858 * op_fid1 is the parent FID, op_fid2 is the child
2860 * \param[out] request point to the request of unlink.
2862 * retval 0 if succeed
2863 * negative errno if failed.
2865 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2866 struct ptlrpc_request **request)
2868 struct obd_device *obd = exp->exp_obd;
2869 struct lmv_obd *lmv = &obd->u.lmv;
2870 struct lmv_tgt_desc *tgt;
2871 struct lmv_tgt_desc *parent_tgt;
2872 struct mdt_body *body;
2877 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2878 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2879 op_data->op_cap = current_cap();
2882 parent_tgt = lmv_locate_tgt(lmv, op_data);
2883 if (IS_ERR(parent_tgt))
2884 RETURN(PTR_ERR(parent_tgt));
2886 if (likely(!fid_is_zero(&op_data->op_fid2))) {
2887 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2889 RETURN(PTR_ERR(tgt));
2895 * If child's fid is given, cancel unused locks for it if it is from
2896 * another export than parent.
2898 * LOOKUP lock for child (fid3) should also be cancelled on parent
2899 * tgt_tgt in mdc_unlink().
2901 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2903 if (parent_tgt != tgt)
2904 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2905 LCK_EX, MDS_INODELOCK_LOOKUP,
2906 MF_MDC_CANCEL_FID3);
2908 rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2909 MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2913 CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2914 PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2917 rc = md_unlink(tgt->ltd_exp, op_data, request);
2918 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2919 ptlrpc_req_finished(*request);
2927 body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2931 /* Not cross-ref case, just get out of here. */
2932 if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2935 /* This is a remote object, try remote MDT. */
2936 op_data->op_fid2 = body->mbo_fid1;
2937 ptlrpc_req_finished(*request);
2940 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2942 RETURN(PTR_ERR(tgt));
2947 static int lmv_precleanup(struct obd_device *obd)
2950 libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2951 fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2952 lprocfs_obd_cleanup(obd);
2953 lprocfs_free_md_stats(obd);
2958 * Get by key a value associated with a LMV device.
2960 * Dispatch request to lower-layer devices as needed.
2962 * \param[in] env execution environment for this thread
2963 * \param[in] exp export for the LMV device
2964 * \param[in] keylen length of key identifier
2965 * \param[in] key identifier of key to get value for
2966 * \param[in] vallen size of \a val
2967 * \param[out] val pointer to storage location for value
2968 * \param[in] lsm optional striping metadata of object
2970 * \retval 0 on success
2971 * \retval negative negated errno on failure
2973 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2974 __u32 keylen, void *key, __u32 *vallen, void *val)
2976 struct obd_device *obd;
2977 struct lmv_obd *lmv;
2978 struct lu_tgt_desc *tgt;
2983 obd = class_exp2obd(exp);
2985 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2986 exp->exp_handle.h_cookie);
2991 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2992 LASSERT(*vallen == sizeof(__u32));
2993 lmv_foreach_connected_tgt(lmv, tgt) {
2994 if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2999 } else if (KEY_IS(KEY_MAX_EASIZE) ||
3000 KEY_IS(KEY_DEFAULT_EASIZE) ||
3001 KEY_IS(KEY_CONN_DATA)) {
3003 * Forwarding this request to first MDS, it should know LOV
3006 tgt = lmv_tgt(lmv, 0);
3010 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3011 if (!rc && KEY_IS(KEY_CONN_DATA))
3012 exp->exp_connect_data = *(struct obd_connect_data *)val;
3014 } else if (KEY_IS(KEY_TGT_COUNT)) {
3015 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
3019 CDEBUG(D_IOCTL, "Invalid key\n");
3023 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3024 int *__rcs, struct ptlrpc_request_set *_set)
3026 struct obd_device *obd = class_exp2obd(exp);
3027 struct ptlrpc_request_set *set = _set;
3028 struct lmv_obd *lmv = &obd->u.lmv;
3029 int tgt_count = lmv->lmv_mdt_count;
3030 struct lu_tgt_desc *tgt;
3031 struct fid_array *fat, **fas = NULL;
3032 int i, rc, **rcs = NULL;
3035 set = ptlrpc_prep_set();
3040 /* split FIDs by targets */
3041 OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3043 GOTO(out, rc = -ENOMEM);
3044 OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3046 GOTO(out_fas, rc = -ENOMEM);
3048 for (i = 0; i < fa->fa_nr; i++) {
3051 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3053 CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3054 PFID(&fa->fa_fids[i]), rc);
3057 LASSERT(idx < tgt_count);
3059 OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3060 fa_fids[fa->fa_nr]));
3062 GOTO(out, rc = -ENOMEM);
3064 OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3066 GOTO(out, rc = -ENOMEM);
3069 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3072 lmv_foreach_connected_tgt(lmv, tgt) {
3073 fat = fas[tgt->ltd_index];
3074 if (!fat || fat->fa_nr == 0)
3076 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3079 rc = ptlrpc_set_wait(NULL, set);
3082 for (i = 0; i < tgt_count; i++) {
3084 if (!fat || fat->fa_nr == 0)
3086 /* copy FIDs back */
3087 memcpy(fa->fa_fids + j, fat->fa_fids,
3088 fat->fa_nr * sizeof(struct lu_fid));
3090 memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3095 ptlrpc_set_destroy(set);
3098 for (i = 0; i < tgt_count; i++) {
3100 OBD_FREE(fas[i], offsetof(struct fid_array,
3101 fa_fids[fa->fa_nr]));
3103 OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3106 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3109 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3115 * Asynchronously set by key a value associated with a LMV device.
3117 * Dispatch request to lower-layer devices as needed.
3119 * \param[in] env execution environment for this thread
3120 * \param[in] exp export for the LMV device
3121 * \param[in] keylen length of key identifier
3122 * \param[in] key identifier of key to store value for
3123 * \param[in] vallen size of value to store
3124 * \param[in] val pointer to data to be stored
3125 * \param[in] set optional list of related ptlrpc requests
3127 * \retval 0 on success
3128 * \retval negative negated errno on failure
3130 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3131 __u32 keylen, void *key, __u32 vallen, void *val,
3132 struct ptlrpc_request_set *set)
3134 struct lmv_tgt_desc *tgt;
3135 struct obd_device *obd;
3136 struct lmv_obd *lmv;
3140 obd = class_exp2obd(exp);
3142 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3143 exp->exp_handle.h_cookie);
3148 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3149 KEY_IS(KEY_DEFAULT_EASIZE)) {
3152 lmv_foreach_connected_tgt(lmv, tgt) {
3153 err = obd_set_info_async(env, tgt->ltd_exp,
3154 keylen, key, vallen, val, set);
3165 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3166 const struct lmv_mds_md_v1 *lmm1)
3168 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3175 lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3176 lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3177 lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3178 if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3179 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3181 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3182 lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3183 lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3184 lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3185 cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3186 sizeof(lsm->lsm_md_pool_name));
3188 if (cplen >= sizeof(lsm->lsm_md_pool_name))
3191 CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3192 "layout_version %d\n", lsm->lsm_md_stripe_count,
3193 lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3194 lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3195 lsm->lsm_md_layout_version);
3197 stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3198 for (i = 0; i < stripe_count; i++) {
3199 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3200 &lmm1->lmv_stripe_fids[i]);
3202 * set default value -1, so lmv_locate_tgt() knows this stripe
3203 * target is not initialized.
3205 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3206 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3209 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3210 &lsm->lsm_md_oinfo[i].lmo_mds);
3217 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3218 PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3224 static inline int lmv_unpack_user_md(struct obd_export *exp,
3225 struct lmv_stripe_md *lsm,
3226 const struct lmv_user_md *lmu)
3228 lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3229 lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3230 lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3231 lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3232 lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3233 lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3234 lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3239 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3240 const union lmv_mds_md *lmm, size_t lmm_size)
3242 struct lmv_stripe_md *lsm;
3245 bool allocated = false;
3248 LASSERT(lsmp != NULL);
3252 if (lsm != NULL && lmm == NULL) {
3254 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3256 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3259 lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3261 OBD_FREE_LARGE(lfm, lfm_size);
3265 if (lmv_dir_striped(lsm)) {
3266 for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3267 iput(lsm->lsm_md_oinfo[i].lmo_root);
3268 lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3270 lsm_size = lmv_stripe_md_size(0);
3272 OBD_FREE(lsm, lsm_size);
3277 /* foreign lmv case */
3278 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3279 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3282 OBD_ALLOC_LARGE(lfm, lmm_size);
3285 *lsmp = (struct lmv_stripe_md *)lfm;
3287 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3288 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3289 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3290 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3291 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3296 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3300 if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3301 le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3302 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3303 exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3308 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3309 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3312 * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3313 * stripecount should be 0 then.
3315 lsm_size = lmv_stripe_md_size(0);
3318 OBD_ALLOC(lsm, lsm_size);
3325 switch (le32_to_cpu(lmm->lmv_magic)) {
3327 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3329 case LMV_USER_MAGIC:
3330 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3333 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3334 le32_to_cpu(lmm->lmv_magic));
3339 if (rc != 0 && allocated) {
3340 OBD_FREE(lsm, lsm_size);
3347 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3349 lmv_unpackmd(NULL, &lsm, NULL, 0);
3351 EXPORT_SYMBOL(lmv_free_memmd);
3353 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3354 union ldlm_policy_data *policy,
3355 enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3358 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3359 struct lu_tgt_desc *tgt;
3365 LASSERT(fid != NULL);
3367 lmv_foreach_connected_tgt(lmv, tgt) {
3368 if (!tgt->ltd_active)
3371 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3379 static int lmv_set_lock_data(struct obd_export *exp,
3380 const struct lustre_handle *lockh,
3381 void *data, __u64 *bits)
3383 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3384 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3389 if (tgt == NULL || tgt->ltd_exp == NULL)
3391 rc = md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3395 static enum ldlm_mode
3396 lmv_lock_match(struct obd_export *exp, __u64 flags,
3397 const struct lu_fid *fid, enum ldlm_type type,
3398 union ldlm_policy_data *policy,
3399 enum ldlm_mode mode, struct lustre_handle *lockh)
3401 struct obd_device *obd = exp->exp_obd;
3402 struct lmv_obd *lmv = &obd->u.lmv;
3404 struct lu_tgt_desc *tgt;
3410 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3413 * With DNE every object can have two locks in different namespaces:
3414 * lookup lock in space of MDT storing direntry and update/open lock in
3415 * space of MDT storing inode. Try the MDT that the FID maps to first,
3416 * since this can be easily found, and only try others if that fails.
3418 for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3419 i < lmv->lmv_mdt_descs.ltd_tgts_size;
3420 i++, index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3422 CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3423 obd->obd_name, PFID(fid), index);
3427 tgt = lmv_tgt(lmv, index);
3428 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3431 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3441 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3442 struct obd_export *dt_exp, struct obd_export *md_exp,
3443 struct lustre_md *md)
3445 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3446 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3448 if (!tgt || !tgt->ltd_exp)
3451 return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3454 static int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3456 struct obd_device *obd = exp->exp_obd;
3457 struct lmv_obd *lmv = &obd->u.lmv;
3458 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3462 if (md->default_lmv) {
3463 lmv_free_memmd(md->default_lmv);
3464 md->default_lmv = NULL;
3466 if (md->lmv != NULL) {
3467 lmv_free_memmd(md->lmv);
3470 if (!tgt || !tgt->ltd_exp)
3472 RETURN(md_free_lustre_md(tgt->ltd_exp, md));
3475 static int lmv_set_open_replay_data(struct obd_export *exp,
3476 struct obd_client_handle *och,
3477 struct lookup_intent *it)
3479 struct obd_device *obd = exp->exp_obd;
3480 struct lmv_obd *lmv = &obd->u.lmv;
3481 struct lmv_tgt_desc *tgt;
3485 tgt = lmv_fid2tgt(lmv, &och->och_fid);
3487 RETURN(PTR_ERR(tgt));
3489 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3492 static int lmv_clear_open_replay_data(struct obd_export *exp,
3493 struct obd_client_handle *och)
3495 struct obd_device *obd = exp->exp_obd;
3496 struct lmv_obd *lmv = &obd->u.lmv;
3497 struct lmv_tgt_desc *tgt;
3501 tgt = lmv_fid2tgt(lmv, &och->och_fid);
3503 RETURN(PTR_ERR(tgt));
3505 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3508 static int lmv_intent_getattr_async(struct obd_export *exp,
3509 struct md_enqueue_info *minfo)
3511 struct md_op_data *op_data = &minfo->mi_data;
3512 struct obd_device *obd = exp->exp_obd;
3513 struct lmv_obd *lmv = &obd->u.lmv;
3514 struct lmv_tgt_desc *ptgt;
3515 struct lmv_tgt_desc *ctgt;
3520 if (!fid_is_sane(&op_data->op_fid2))
3523 ptgt = lmv_locate_tgt(lmv, op_data);
3525 RETURN(PTR_ERR(ptgt));
3527 ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3529 RETURN(PTR_ERR(ctgt));
3532 * remote object needs two RPCs to lookup and getattr, considering the
3533 * complexity don't support statahead for now.
3538 rc = md_intent_getattr_async(ptgt->ltd_exp, minfo);
3543 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3544 struct lu_fid *fid, __u64 *bits)
3546 struct obd_device *obd = exp->exp_obd;
3547 struct lmv_obd *lmv = &obd->u.lmv;
3548 struct lmv_tgt_desc *tgt;
3553 tgt = lmv_fid2tgt(lmv, fid);
3555 RETURN(PTR_ERR(tgt));
3557 rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3561 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3562 const struct lmv_stripe_md *lsm,
3563 const char *name, int namelen,
3566 const struct lmv_oinfo *oinfo;
3568 LASSERT(lmv_dir_striped(lsm));
3570 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3572 return PTR_ERR(oinfo);
3574 *fid = oinfo->lmo_fid;
3580 * For lmv, only need to send request to master MDT, and the master MDT will
3581 * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3582 * we directly fetch data from the slave MDTs.
3584 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3585 struct obd_quotactl *oqctl)
3587 struct obd_device *obd = class_exp2obd(exp);
3588 struct lmv_obd *lmv = &obd->u.lmv;
3589 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3590 __u64 curspace, curinodes;
3595 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3596 CERROR("master lmv inactive\n");
3600 if (oqctl->qc_cmd != Q_GETOQUOTA) {
3601 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3605 curspace = curinodes = 0;
3606 lmv_foreach_connected_tgt(lmv, tgt) {
3609 if (!tgt->ltd_active)
3612 err = obd_quotactl(tgt->ltd_exp, oqctl);
3614 CERROR("getquota on mdt %d failed. %d\n",
3615 tgt->ltd_index, err);
3619 curspace += oqctl->qc_dqblk.dqb_curspace;
3620 curinodes += oqctl->qc_dqblk.dqb_curinodes;
3623 oqctl->qc_dqblk.dqb_curspace = curspace;
3624 oqctl->qc_dqblk.dqb_curinodes = curinodes;
3629 static int lmv_merge_attr(struct obd_export *exp,
3630 const struct lmv_stripe_md *lsm,
3631 struct cl_attr *attr,
3632 ldlm_blocking_callback cb_blocking)
3637 if (!lmv_dir_striped(lsm))
3640 rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3644 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3645 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3651 "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3652 PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3653 i_size_read(inode), (unsigned long long)inode->i_blocks,
3654 inode->i_nlink, (s64)inode->i_atime.tv_sec,
3655 (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3657 /* for slave stripe, it needs to subtract nlink for . and .. */
3659 attr->cat_nlink += inode->i_nlink - 2;
3661 attr->cat_nlink = inode->i_nlink;
3663 attr->cat_size += i_size_read(inode);
3664 attr->cat_blocks += inode->i_blocks;
3666 if (attr->cat_atime < inode->i_atime.tv_sec)
3667 attr->cat_atime = inode->i_atime.tv_sec;
3669 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3670 attr->cat_ctime = inode->i_ctime.tv_sec;
3672 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3673 attr->cat_mtime = inode->i_mtime.tv_sec;
3678 static const struct obd_ops lmv_obd_ops = {
3679 .o_owner = THIS_MODULE,
3680 .o_setup = lmv_setup,
3681 .o_cleanup = lmv_cleanup,
3682 .o_precleanup = lmv_precleanup,
3683 .o_process_config = lmv_process_config,
3684 .o_connect = lmv_connect,
3685 .o_disconnect = lmv_disconnect,
3686 .o_statfs = lmv_statfs,
3687 .o_get_info = lmv_get_info,
3688 .o_set_info_async = lmv_set_info_async,
3689 .o_notify = lmv_notify,
3690 .o_get_uuid = lmv_get_uuid,
3691 .o_fid_alloc = lmv_fid_alloc,
3692 .o_iocontrol = lmv_iocontrol,
3693 .o_quotactl = lmv_quotactl
3696 static const struct md_ops lmv_md_ops = {
3697 .m_get_root = lmv_get_root,
3698 .m_null_inode = lmv_null_inode,
3699 .m_close = lmv_close,
3700 .m_create = lmv_create,
3701 .m_enqueue = lmv_enqueue,
3702 .m_getattr = lmv_getattr,
3703 .m_getxattr = lmv_getxattr,
3704 .m_getattr_name = lmv_getattr_name,
3705 .m_intent_lock = lmv_intent_lock,
3707 .m_rename = lmv_rename,
3708 .m_setattr = lmv_setattr,
3709 .m_setxattr = lmv_setxattr,
3710 .m_fsync = lmv_fsync,
3711 .m_file_resync = lmv_file_resync,
3712 .m_read_page = lmv_read_page,
3713 .m_unlink = lmv_unlink,
3714 .m_init_ea_size = lmv_init_ea_size,
3715 .m_cancel_unused = lmv_cancel_unused,
3716 .m_set_lock_data = lmv_set_lock_data,
3717 .m_lock_match = lmv_lock_match,
3718 .m_get_lustre_md = lmv_get_lustre_md,
3719 .m_free_lustre_md = lmv_free_lustre_md,
3720 .m_merge_attr = lmv_merge_attr,
3721 .m_set_open_replay_data = lmv_set_open_replay_data,
3722 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3723 .m_intent_getattr_async = lmv_intent_getattr_async,
3724 .m_revalidate_lock = lmv_revalidate_lock,
3725 .m_get_fid_from_lsm = lmv_get_fid_from_lsm,
3726 .m_unpackmd = lmv_unpackmd,
3727 .m_rmfid = lmv_rmfid,
3730 static int __init lmv_init(void)
3732 return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
3733 LUSTRE_LMV_NAME, NULL);
3736 static void __exit lmv_exit(void)
3738 class_unregister_type(LUSTRE_LMV_NAME);
3741 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3742 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3743 MODULE_VERSION(LUSTRE_VERSION_STRING);
3744 MODULE_LICENSE("GPL");
3746 module_init(lmv_init);
3747 module_exit(lmv_exit);