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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_LMV
35 #include <linux/file.h>
36 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/user_namespace.h>
39 #include <linux/uidgid.h>
40 #include <linux/slab.h>
41 #include <linux/pagemap.h>
43 #include <linux/math64.h>
44 #include <linux/seq_file.h>
45 #include <linux/namei.h>
47 #include <obd_support.h>
48 #include <lustre_lib.h>
49 #include <lustre_net.h>
50 #include <obd_class.h>
51 #include <lustre_lmv.h>
52 #include <lprocfs_status.h>
53 #include <cl_object.h>
54 #include <lustre_fid.h>
55 #include <uapi/linux/lustre/lustre_ioctl.h>
56 #include <lustre_kernelcomm.h>
57 #include "lmv_internal.h"
59 static int lmv_check_connect(struct obd_device *obd);
61 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
64 if (tgt->ltd_active == activate)
67 tgt->ltd_active = activate;
68 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
71 tgt->ltd_exp->exp_obd->obd_inactive = !activate;
77 * -EINVAL : UUID can't be found in the LMV's target list
78 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
79 * -EBADF : The UUID is found, but the OBD of the wrong type (!)
81 static int lmv_set_mdc_active(struct lmv_obd *lmv,
82 const struct obd_uuid *uuid,
85 struct lu_tgt_desc *tgt = NULL;
86 struct obd_device *obd;
91 CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
92 lmv, uuid->uuid, activate);
94 spin_lock(&lmv->lmv_lock);
95 lmv_foreach_connected_tgt(lmv, tgt) {
96 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
97 tgt->ltd_index, tgt->ltd_uuid.uuid,
98 tgt->ltd_exp->exp_handle.h_cookie);
100 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
105 GOTO(out_lmv_lock, rc = -EINVAL);
107 obd = class_exp2obd(tgt->ltd_exp);
109 GOTO(out_lmv_lock, rc = -ENOTCONN);
111 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
112 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
113 obd->obd_type->typ_name, tgt->ltd_index);
114 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
116 if (tgt->ltd_active == activate) {
117 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
118 activate ? "" : "in");
119 GOTO(out_lmv_lock, rc);
122 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
123 activate ? "" : "in");
124 lmv_activate_target(lmv, tgt, activate);
128 spin_unlock(&lmv->lmv_lock);
132 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
134 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
135 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
137 return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
140 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
141 enum obd_notify_event ev)
143 struct obd_connect_data *conn_data;
144 struct lmv_obd *lmv = &obd->u.lmv;
145 struct obd_uuid *uuid;
149 if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
150 CERROR("unexpected notification of %s %s!\n",
151 watched->obd_type->typ_name,
156 uuid = &watched->u.cli.cl_target_uuid;
157 if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
159 * Set MDC as active before notifying the observer, so the
160 * observer can use the MDC normally.
162 rc = lmv_set_mdc_active(lmv, uuid,
163 ev == OBD_NOTIFY_ACTIVE);
165 CERROR("%sactivation of %s failed: %d\n",
166 ev == OBD_NOTIFY_ACTIVE ? "" : "de",
170 } else if (ev == OBD_NOTIFY_OCD) {
171 conn_data = &watched->u.cli.cl_import->imp_connect_data;
173 * XXX: Make sure that ocd_connect_flags from all targets are
174 * the same. Otherwise one of MDTs runs wrong version or
175 * something like this. --umka
177 obd->obd_self_export->exp_connect_data = *conn_data;
181 * Pass the notification up the chain.
183 if (obd->obd_observer)
184 rc = obd_notify(obd->obd_observer, watched, ev);
189 static int lmv_connect(const struct lu_env *env,
190 struct obd_export **pexp, struct obd_device *obd,
191 struct obd_uuid *cluuid, struct obd_connect_data *data,
194 struct lmv_obd *lmv = &obd->u.lmv;
195 struct lustre_handle conn = { 0 };
196 struct obd_export *exp;
200 rc = class_connect(&conn, obd, cluuid);
202 CERROR("class_connection() returned %d\n", rc);
206 exp = class_conn2export(&conn);
209 lmv->conn_data = *data;
210 lmv->lmv_cache = localdata;
212 lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
213 &obd->obd_kset.kobj);
214 if (!lmv->lmv_tgts_kobj) {
215 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
216 obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
219 rc = lmv_check_connect(obd);
228 if (lmv->lmv_tgts_kobj)
229 kobject_put(lmv->lmv_tgts_kobj);
231 class_disconnect(exp);
236 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
239 struct obd_device *obd = exp->exp_obd;
240 struct lmv_obd *lmv = &obd->u.lmv;
241 struct lmv_tgt_desc *tgt;
247 if (lmv->max_easize < easize) {
248 lmv->max_easize = easize;
251 if (lmv->max_def_easize < def_easize) {
252 lmv->max_def_easize = def_easize;
259 if (lmv->connected == 0)
262 lmv_foreach_connected_tgt(lmv, tgt) {
263 if (!tgt->ltd_active)
266 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
268 CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
269 " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
276 #define MAX_STRING_SIZE 128
278 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
280 struct lmv_obd *lmv = &obd->u.lmv;
281 struct obd_device *mdc_obd;
282 struct obd_export *mdc_exp;
283 struct lu_fld_target target;
287 mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
290 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
294 CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
295 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
296 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
298 if (!mdc_obd->obd_set_up) {
299 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
303 rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
304 &lmv->conn_data, lmv->lmv_cache);
306 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
311 * Init fid sequence client for this mdc and add new fld target.
313 rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
317 target.ft_srv = NULL;
318 target.ft_exp = mdc_exp;
319 target.ft_idx = tgt->ltd_index;
321 fld_client_add_target(&lmv->lmv_fld, &target);
323 rc = obd_register_observer(mdc_obd, obd);
325 obd_disconnect(mdc_exp);
326 CERROR("target %s register_observer error %d\n",
327 tgt->ltd_uuid.uuid, rc);
331 if (obd->obd_observer) {
333 * Tell the observer about the new target.
335 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
338 obd_disconnect(mdc_exp);
344 tgt->ltd_exp = mdc_exp;
345 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
347 md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
349 rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
351 obd_disconnect(mdc_exp);
355 CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
356 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
357 atomic_read(&obd->obd_refcount));
359 lmv_statfs_check_update(obd, tgt);
361 if (lmv->lmv_tgts_kobj)
362 /* Even if we failed to create the link, that's fine */
363 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
364 &mdc_obd->obd_kset.kobj,
369 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
372 ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
376 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
377 __u32 index, int gen)
379 struct obd_device *mdc_obd;
380 struct lmv_obd *lmv = &obd->u.lmv;
381 struct lmv_tgt_desc *tgt;
382 struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
387 CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
388 mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
391 CERROR("%s: Target %s not attached: rc = %d\n",
392 obd->obd_name, uuidp->uuid, -EINVAL);
400 mutex_init(&tgt->ltd_fid_mutex);
401 tgt->ltd_index = index;
402 tgt->ltd_uuid = *uuidp;
405 mutex_lock(<d->ltd_mutex);
406 rc = ltd_add_tgt(ltd, tgt);
407 mutex_unlock(<d->ltd_mutex);
413 /* lmv_check_connect() will connect this target. */
416 rc = lmv_connect_mdc(obd, tgt);
418 int easize = sizeof(struct lmv_stripe_md) +
419 lmv->lmv_mdt_count * sizeof(struct lu_fid);
421 lmv_init_ea_size(obd->obd_self_export, easize, 0);
431 static int lmv_check_connect(struct obd_device *obd)
433 struct lmv_obd *lmv = &obd->u.lmv;
434 struct lmv_tgt_desc *tgt;
443 mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
445 GOTO(unlock, rc = 0);
447 if (!lmv->lmv_mdt_count) {
448 CERROR("%s: no targets configured: rc = -EINVAL\n",
450 GOTO(unlock, rc = -EINVAL);
453 if (!lmv_mdt0_inited(lmv)) {
454 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
456 GOTO(unlock, rc = -EINVAL);
459 CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
460 obd->obd_uuid.uuid, obd->obd_name);
462 lmv_foreach_tgt(lmv, tgt) {
463 rc = lmv_connect_mdc(obd, tgt);
469 easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
470 lmv_init_ea_size(obd->obd_self_export, easize, 0);
473 mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
478 lmv_foreach_tgt(lmv, tgt) {
483 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
484 obd_disconnect(tgt->ltd_exp);
490 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
492 struct lmv_obd *lmv = &obd->u.lmv;
493 struct obd_device *mdc_obd;
497 LASSERT(tgt != NULL);
498 LASSERT(obd != NULL);
500 mdc_obd = class_exp2obd(tgt->ltd_exp);
503 mdc_obd->obd_force = obd->obd_force;
504 mdc_obd->obd_fail = obd->obd_fail;
505 mdc_obd->obd_no_recov = obd->obd_no_recov;
507 if (lmv->lmv_tgts_kobj)
508 sysfs_remove_link(lmv->lmv_tgts_kobj,
512 rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
514 CERROR("Can't finanize fids factory\n");
516 CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
517 tgt->ltd_exp->exp_obd->obd_name,
518 tgt->ltd_exp->exp_obd->obd_uuid.uuid);
520 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
521 rc = obd_disconnect(tgt->ltd_exp);
523 if (tgt->ltd_active) {
524 CERROR("Target %s disconnect error %d\n",
525 tgt->ltd_uuid.uuid, rc);
529 lmv_activate_target(lmv, tgt, 0);
534 static int lmv_disconnect(struct obd_export *exp)
536 struct obd_device *obd = class_exp2obd(exp);
537 struct lmv_obd *lmv = &obd->u.lmv;
538 struct lmv_tgt_desc *tgt;
543 lmv_foreach_connected_tgt(lmv, tgt)
544 lmv_disconnect_mdc(obd, tgt);
546 if (lmv->lmv_tgts_kobj)
547 kobject_put(lmv->lmv_tgts_kobj);
550 class_export_put(exp);
551 rc = class_disconnect(exp);
557 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
560 struct obd_device *obd = class_exp2obd(exp);
561 struct lmv_obd *lmv = &obd->u.lmv;
562 struct getinfo_fid2path *gf;
563 struct lmv_tgt_desc *tgt;
564 struct getinfo_fid2path *remote_gf = NULL;
565 struct lu_fid root_fid;
566 int remote_gf_size = 0;
570 tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
572 RETURN(PTR_ERR(tgt));
574 root_fid = *gf->gf_u.gf_root_fid;
575 LASSERT(fid_is_sane(&root_fid));
578 rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
579 if (rc != 0 && rc != -EREMOTE)
580 GOTO(out_fid2path, rc);
582 /* If remote_gf != NULL, it means just building the
583 * path on the remote MDT, copy this path segement to gf */
584 if (remote_gf != NULL) {
585 struct getinfo_fid2path *ori_gf;
589 ori_gf = (struct getinfo_fid2path *)karg;
590 if (strlen(ori_gf->gf_u.gf_path) + 1 +
591 strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
592 GOTO(out_fid2path, rc = -EOVERFLOW);
594 ptr = ori_gf->gf_u.gf_path;
596 len = strlen(gf->gf_u.gf_path);
597 /* move the current path to the right to release space
598 * for closer-to-root part */
599 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
600 memcpy(ptr, gf->gf_u.gf_path, len);
604 CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
605 tgt->ltd_exp->exp_obd->obd_name,
606 gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
610 GOTO(out_fid2path, rc);
612 /* sigh, has to go to another MDT to do path building further */
613 if (remote_gf == NULL) {
614 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
615 OBD_ALLOC(remote_gf, remote_gf_size);
616 if (remote_gf == NULL)
617 GOTO(out_fid2path, rc = -ENOMEM);
618 remote_gf->gf_pathlen = PATH_MAX;
621 if (!fid_is_sane(&gf->gf_fid)) {
622 CERROR("%s: invalid FID "DFID": rc = %d\n",
623 tgt->ltd_exp->exp_obd->obd_name,
624 PFID(&gf->gf_fid), -EINVAL);
625 GOTO(out_fid2path, rc = -EINVAL);
628 tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
630 GOTO(out_fid2path, rc = -EINVAL);
632 remote_gf->gf_fid = gf->gf_fid;
633 remote_gf->gf_recno = -1;
634 remote_gf->gf_linkno = -1;
635 memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
636 *remote_gf->gf_u.gf_root_fid = root_fid;
638 goto repeat_fid2path;
641 if (remote_gf != NULL)
642 OBD_FREE(remote_gf, remote_gf_size);
646 static int lmv_hsm_req_count(struct lmv_obd *lmv,
647 const struct hsm_user_request *hur,
648 const struct lmv_tgt_desc *tgt_mds)
650 struct lmv_tgt_desc *curr_tgt;
654 /* count how many requests must be sent to the given target */
655 for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
656 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
657 if (IS_ERR(curr_tgt))
658 RETURN(PTR_ERR(curr_tgt));
659 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
665 static int lmv_hsm_req_build(struct lmv_obd *lmv,
666 struct hsm_user_request *hur_in,
667 const struct lmv_tgt_desc *tgt_mds,
668 struct hsm_user_request *hur_out)
671 struct lmv_tgt_desc *curr_tgt;
673 /* build the hsm_user_request for the given target */
674 hur_out->hur_request = hur_in->hur_request;
676 for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
677 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
678 if (IS_ERR(curr_tgt))
679 RETURN(PTR_ERR(curr_tgt));
680 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
681 hur_out->hur_user_item[nr_out] =
682 hur_in->hur_user_item[i];
686 hur_out->hur_request.hr_itemcount = nr_out;
687 memcpy(hur_data(hur_out), hur_data(hur_in),
688 hur_in->hur_request.hr_data_len);
693 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
694 int len, struct lustre_kernelcomm *lk,
697 struct lmv_obd *lmv = &obd->u.lmv;
698 struct lu_tgt_desc *tgt;
703 /* unregister request (call from llapi_hsm_copytool_fini) */
704 lmv_foreach_connected_tgt(lmv, tgt)
705 /* best effort: try to clean as much as possible
706 * (continue on error) */
707 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
709 /* Whatever the result, remove copytool from kuc groups.
710 * Unreached coordinators will get EPIPE on next requests
711 * and will unregister automatically.
713 rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
718 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
719 int len, struct lustre_kernelcomm *lk,
722 struct lmv_obd *lmv = &obd->u.lmv;
724 bool any_set = false;
725 struct kkuc_ct_data *kcd;
727 struct lu_tgt_desc *tgt;
734 filp = fget(lk->lk_wfd);
738 if (lk->lk_flags & LK_FLG_DATANR)
739 kcd_size = offsetof(struct kkuc_ct_data,
740 kcd_archives[lk->lk_data_count]);
742 kcd_size = sizeof(*kcd);
744 OBD_ALLOC(kcd, kcd_size);
746 GOTO(err_fput, rc = -ENOMEM);
748 kcd->kcd_nr_archives = lk->lk_data_count;
749 if (lk->lk_flags & LK_FLG_DATANR) {
750 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
751 if (lk->lk_data_count > 0)
752 memcpy(kcd->kcd_archives, lk->lk_data,
753 sizeof(*kcd->kcd_archives) * lk->lk_data_count);
755 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
758 rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
759 lk->lk_group, kcd, kcd_size);
760 OBD_FREE(kcd, kcd_size);
764 /* All or nothing: try to register to all MDS.
765 * In case of failure, unregister from previous MDS,
766 * except if it because of inactive target. */
767 lmv_foreach_connected_tgt(lmv, tgt) {
768 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
770 if (tgt->ltd_active) {
771 /* permanent error */
772 CERROR("%s: iocontrol MDC %s on MDT"
773 " idx %d cmd %x: err = %d\n",
774 lmv2obd_dev(lmv)->obd_name,
775 tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
778 lk->lk_flags |= LK_FLG_STOP;
780 /* unregister from previous MDS */
781 lmv_foreach_connected_tgt(lmv, tgt) {
782 if (tgt->ltd_index >= i)
785 obd_iocontrol(cmd, tgt->ltd_exp, len,
788 GOTO(err_kkuc_rem, rc);
790 /* else: transient error.
791 * kuc will register to the missing MDT
799 /* no registration done: return error */
800 GOTO(err_kkuc_rem, rc = -ENOTCONN);
805 libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
812 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
813 int len, void *karg, void __user *uarg)
815 struct obd_device *obd = class_exp2obd(exp);
816 struct lmv_obd *lmv = &obd->u.lmv;
817 struct lu_tgt_desc *tgt = NULL;
819 __u32 count = lmv->lmv_mdt_count;
828 case IOC_OBD_STATFS: {
829 struct obd_ioctl_data *data = karg;
830 struct obd_device *mdc_obd;
831 struct obd_statfs stat_buf = {0};
834 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
836 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
839 tgt = lmv_tgt(lmv, index);
840 if (!tgt || !tgt->ltd_active)
843 mdc_obd = class_exp2obd(tgt->ltd_exp);
848 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
849 min((int) data->ioc_plen2,
850 (int) sizeof(struct obd_uuid))))
853 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
854 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
858 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
859 min((int) data->ioc_plen1,
860 (int) sizeof(stat_buf))))
864 case OBD_IOC_QUOTACTL: {
865 struct if_quotactl *qctl = karg;
866 struct obd_quotactl *oqctl;
868 if (qctl->qc_valid == QC_MDTIDX) {
869 tgt = lmv_tgt(lmv, qctl->qc_idx);
870 } else if (qctl->qc_valid == QC_UUID) {
871 lmv_foreach_tgt(lmv, tgt) {
872 if (!obd_uuid_equals(&tgt->ltd_uuid,
885 if (!tgt || !tgt->ltd_exp)
888 OBD_ALLOC_PTR(oqctl);
892 QCTL_COPY(oqctl, qctl);
893 rc = obd_quotactl(tgt->ltd_exp, oqctl);
895 QCTL_COPY(qctl, oqctl);
896 qctl->qc_valid = QC_MDTIDX;
897 qctl->obd_uuid = tgt->ltd_uuid;
902 case LL_IOC_GET_CONNECT_FLAGS: {
903 tgt = lmv_tgt(lmv, 0);
905 if (tgt && tgt->ltd_exp)
906 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
909 case LL_IOC_FID2MDTIDX: {
910 struct lu_fid *fid = karg;
913 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
917 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
918 * point to user space memory for FID2MDTIDX. */
919 *(__u32 *)uarg = mdt_index;
922 case OBD_IOC_FID2PATH: {
923 rc = lmv_fid2path(exp, len, karg, uarg);
926 case LL_IOC_HSM_STATE_GET:
927 case LL_IOC_HSM_STATE_SET:
928 case LL_IOC_HSM_ACTION: {
929 struct md_op_data *op_data = karg;
931 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
933 RETURN(PTR_ERR(tgt));
935 if (tgt->ltd_exp == NULL)
938 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
941 case LL_IOC_HSM_PROGRESS: {
942 const struct hsm_progress_kernel *hpk = karg;
944 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
946 RETURN(PTR_ERR(tgt));
947 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
950 case LL_IOC_HSM_REQUEST: {
951 struct hsm_user_request *hur = karg;
952 unsigned int reqcount = hur->hur_request.hr_itemcount;
957 /* if the request is about a single fid
958 * or if there is a single MDS, no need to split
960 if (reqcount == 1 || count == 1) {
961 tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
963 RETURN(PTR_ERR(tgt));
964 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
966 /* split fid list to their respective MDS */
967 lmv_foreach_connected_tgt(lmv, tgt) {
970 struct hsm_user_request *req;
972 nr = lmv_hsm_req_count(lmv, hur, tgt);
975 if (nr == 0) /* nothing for this MDS */
978 /* build a request with fids for this MDS */
979 reqlen = offsetof(typeof(*hur),
981 + hur->hur_request.hr_data_len;
982 OBD_ALLOC_LARGE(req, reqlen);
985 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
987 GOTO(hsm_req_err, rc1);
988 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
991 if (rc1 != 0 && rc == 0)
993 OBD_FREE_LARGE(req, reqlen);
998 case LL_IOC_LOV_SWAP_LAYOUTS: {
999 struct md_op_data *op_data = karg;
1000 struct lmv_tgt_desc *tgt1, *tgt2;
1002 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1004 RETURN(PTR_ERR(tgt1));
1006 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1008 RETURN(PTR_ERR(tgt2));
1010 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1013 /* only files on same MDT can have their layouts swapped */
1014 if (tgt1->ltd_index != tgt2->ltd_index)
1017 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1020 case LL_IOC_HSM_CT_START: {
1021 struct lustre_kernelcomm *lk = karg;
1022 if (lk->lk_flags & LK_FLG_STOP)
1023 rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1025 rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1029 lmv_foreach_connected_tgt(lmv, tgt) {
1030 struct obd_device *mdc_obd;
1033 /* ll_umount_begin() sets force flag but for lmv, not
1034 * mdc. Let's pass it through */
1035 mdc_obd = class_exp2obd(tgt->ltd_exp);
1036 mdc_obd->obd_force = obd->obd_force;
1037 err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1039 if (tgt->ltd_active) {
1040 CERROR("error: iocontrol MDC %s on MDT"
1041 " idx %d cmd %x: err = %d\n",
1043 tgt->ltd_index, cmd, err);
1056 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1057 struct lu_fid *fid, struct md_op_data *op_data)
1059 struct obd_device *obd = class_exp2obd(exp);
1060 struct lmv_obd *lmv = &obd->u.lmv;
1061 struct lmv_tgt_desc *tgt;
1069 tgt = lmv_tgt(lmv, op_data->op_mds);
1073 if (!tgt->ltd_active || !tgt->ltd_exp)
1077 * New seq alloc and FLD setup should be atomic. Otherwise we may find
1078 * on server that seq in new allocated fid is not yet known.
1080 mutex_lock(&tgt->ltd_fid_mutex);
1081 rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1082 mutex_unlock(&tgt->ltd_fid_mutex);
1084 LASSERT(fid_is_sane(fid));
1091 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1093 struct lmv_obd *lmv = &obd->u.lmv;
1094 struct lmv_desc *desc;
1095 struct lnet_process_id lnet_id;
1101 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1102 CERROR("LMV setup requires a descriptor\n");
1106 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1107 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1108 CERROR("Lmv descriptor size wrong: %d > %d\n",
1109 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1113 obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1114 desc->ld_uuid.uuid);
1115 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1116 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1117 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1118 LMV_DESC_QOS_MAXAGE_DEFAULT;
1119 lmv->max_def_easize = 0;
1120 lmv->max_easize = 0;
1122 spin_lock_init(&lmv->lmv_lock);
1125 * initialize rr_index to lower 32bit of netid, so that client
1126 * can distribute subdirs evenly from the beginning.
1128 while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1129 if (lnet_id.nid != LNET_NID_LO_0) {
1130 lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1135 rc = lmv_tunables_init(obd);
1137 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1140 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1141 LUSTRE_CLI_FLD_HASH_DHT);
1143 CERROR("Can't init FLD, err %d\n", rc);
1145 rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1147 CWARN("%s: error initialize target table: rc = %d\n",
1153 static int lmv_cleanup(struct obd_device *obd)
1155 struct lmv_obd *lmv = &obd->u.lmv;
1156 struct lu_tgt_desc *tgt;
1157 struct lu_tgt_desc *tmp;
1161 fld_client_fini(&lmv->lmv_fld);
1162 lmv_foreach_tgt_safe(lmv, tgt, tmp)
1163 lmv_del_target(lmv, tgt);
1164 lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1169 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1171 struct lustre_cfg *lcfg = buf;
1172 struct obd_uuid obd_uuid;
1178 switch (lcfg->lcfg_command) {
1180 /* modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDT0000_UUID
1181 * 2:0 3:1 4:lustre-MDT0000-mdc_UUID */
1182 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1183 GOTO(out, rc = -EINVAL);
1185 obd_str2uuid(&obd_uuid, lustre_cfg_buf(lcfg, 1));
1187 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1188 GOTO(out, rc = -EINVAL);
1189 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1190 GOTO(out, rc = -EINVAL);
1191 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1194 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1195 GOTO(out, rc = -EINVAL);
1201 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1205 if (flags & OBD_STATFS_FOR_MDT0)
1208 if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1209 return lmv->lmv_statfs_start;
1211 /* choose initial MDT for this client */
1213 struct lnet_process_id lnet_id;
1214 if (LNetGetId(i, &lnet_id) == -ENOENT)
1217 if (lnet_id.nid != LNET_NID_LO_0) {
1218 /* We dont need a full 64-bit modulus, just enough
1219 * to distribute the requests across MDTs evenly.
1221 lmv->lmv_statfs_start = (u32)lnet_id.nid %
1227 return lmv->lmv_statfs_start;
1230 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1231 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1233 struct obd_device *obd = class_exp2obd(exp);
1234 struct lmv_obd *lmv = &obd->u.lmv;
1235 struct obd_statfs *temp;
1236 struct lu_tgt_desc *tgt;
1243 OBD_ALLOC(temp, sizeof(*temp));
1247 /* distribute statfs among MDTs */
1248 idx = lmv_select_statfs_mdt(lmv, flags);
1250 for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1251 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1252 tgt = lmv_tgt(lmv, idx);
1253 if (!tgt || !tgt->ltd_exp)
1256 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age, flags);
1258 CERROR("%s: can't stat MDS #%d: rc = %d\n",
1259 tgt->ltd_exp->exp_obd->obd_name, i, rc);
1260 GOTO(out_free_temp, rc);
1263 if (temp->os_state & OS_STATE_SUM ||
1264 flags == OBD_STATFS_FOR_MDT0) {
1265 /* reset to the last aggregated values
1266 * and don't sum with non-aggrated data */
1267 /* If the statfs is from mount, it needs to retrieve
1268 * necessary information from MDT0. i.e. mount does
1269 * not need the merged osfs from all of MDT. Also
1270 * clients can be mounted as long as MDT0 is in
1279 osfs->os_bavail += temp->os_bavail;
1280 osfs->os_blocks += temp->os_blocks;
1281 osfs->os_ffree += temp->os_ffree;
1282 osfs->os_files += temp->os_files;
1283 osfs->os_granted += temp->os_granted;
1289 OBD_FREE(temp, sizeof(*temp));
1293 static int lmv_statfs_update(void *cookie, int rc)
1295 struct obd_info *oinfo = cookie;
1296 struct obd_device *obd = oinfo->oi_obd;
1297 struct lmv_obd *lmv = &obd->u.lmv;
1298 struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1299 struct obd_statfs *osfs = oinfo->oi_osfs;
1302 * NB: don't deactivate TGT upon error, because we may not trigger async
1303 * statfs any longer, then there is no chance to activate TGT.
1306 spin_lock(&lmv->lmv_lock);
1307 tgt->ltd_statfs = *osfs;
1308 tgt->ltd_statfs_age = ktime_get_seconds();
1309 spin_unlock(&lmv->lmv_lock);
1310 lmv->lmv_qos.lq_dirty = 1;
1316 /* update tgt statfs async if it's ld_qos_maxage old */
1317 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1319 struct obd_info oinfo = {
1322 .oi_cb_up = lmv_statfs_update,
1326 if (ktime_get_seconds() - tgt->ltd_statfs_age <
1327 obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1330 rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1335 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1338 struct obd_device *obd = exp->exp_obd;
1339 struct lmv_obd *lmv = &obd->u.lmv;
1340 struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1348 rc = md_get_root(tgt->ltd_exp, fileset, fid);
1352 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1353 u64 obd_md_valid, const char *name, size_t buf_size,
1354 struct ptlrpc_request **req)
1356 struct obd_device *obd = exp->exp_obd;
1357 struct lmv_obd *lmv = &obd->u.lmv;
1358 struct lmv_tgt_desc *tgt;
1363 tgt = lmv_fid2tgt(lmv, fid);
1365 RETURN(PTR_ERR(tgt));
1367 rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1372 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1373 u64 obd_md_valid, const char *name,
1374 const void *value, size_t value_size,
1375 unsigned int xattr_flags, u32 suppgid,
1376 struct ptlrpc_request **req)
1378 struct obd_device *obd = exp->exp_obd;
1379 struct lmv_obd *lmv = &obd->u.lmv;
1380 struct lmv_tgt_desc *tgt;
1385 tgt = lmv_fid2tgt(lmv, fid);
1387 RETURN(PTR_ERR(tgt));
1389 rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1390 value, value_size, xattr_flags, suppgid, req);
1395 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1396 struct ptlrpc_request **request)
1398 struct obd_device *obd = exp->exp_obd;
1399 struct lmv_obd *lmv = &obd->u.lmv;
1400 struct lmv_tgt_desc *tgt;
1405 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1407 RETURN(PTR_ERR(tgt));
1409 if (op_data->op_flags & MF_GET_MDT_IDX) {
1410 op_data->op_mds = tgt->ltd_index;
1414 rc = md_getattr(tgt->ltd_exp, op_data, request);
1419 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1421 struct obd_device *obd = exp->exp_obd;
1422 struct lmv_obd *lmv = &obd->u.lmv;
1423 struct lu_tgt_desc *tgt;
1427 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1430 * With DNE every object can have two locks in different namespaces:
1431 * lookup lock in space of MDT storing direntry and update/open lock in
1432 * space of MDT storing inode.
1434 lmv_foreach_connected_tgt(lmv, tgt)
1435 md_null_inode(tgt->ltd_exp, fid);
1440 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1441 struct md_open_data *mod, struct ptlrpc_request **request)
1443 struct obd_device *obd = exp->exp_obd;
1444 struct lmv_obd *lmv = &obd->u.lmv;
1445 struct lmv_tgt_desc *tgt;
1450 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1452 RETURN(PTR_ERR(tgt));
1454 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1455 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1459 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
1461 struct lu_tgt_desc *tgt;
1462 __u64 total_weight = 0;
1463 __u64 cur_weight = 0;
1469 if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1470 RETURN(ERR_PTR(-EAGAIN));
1472 down_write(&lmv->lmv_qos.lq_rw_sem);
1474 if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1475 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1477 rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1479 GOTO(unlock, tgt = ERR_PTR(rc));
1481 lmv_foreach_tgt(lmv, tgt) {
1482 tgt->ltd_qos.ltq_usable = 0;
1483 if (!tgt->ltd_exp || !tgt->ltd_active)
1486 tgt->ltd_qos.ltq_usable = 1;
1487 lu_tgt_qos_weight_calc(tgt);
1488 total_weight += tgt->ltd_qos.ltq_weight;
1491 rand = lu_prandom_u64_max(total_weight);
1493 lmv_foreach_connected_tgt(lmv, tgt) {
1494 if (!tgt->ltd_qos.ltq_usable)
1497 cur_weight += tgt->ltd_qos.ltq_weight;
1498 if (cur_weight < rand)
1501 *mdt = tgt->ltd_index;
1502 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1503 GOTO(unlock, rc = 0);
1506 /* no proper target found */
1507 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1509 up_write(&lmv->lmv_qos.lq_rw_sem);
1514 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1516 struct lu_tgt_desc *tgt;
1522 spin_lock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1523 for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1524 index = (i + lmv->lmv_qos_rr_index) %
1525 lmv->lmv_mdt_descs.ltd_tgts_size;
1526 tgt = lmv_tgt(lmv, index);
1527 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1530 *mdt = tgt->ltd_index;
1531 lmv->lmv_qos_rr_index = (*mdt + 1) %
1532 lmv->lmv_mdt_descs.ltd_tgts_size;
1533 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1537 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1539 RETURN(ERR_PTR(-ENODEV));
1542 static struct lmv_tgt_desc *
1543 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1544 const char *name, int namelen, struct lu_fid *fid,
1545 __u32 *mds, bool new_layout)
1547 struct lmv_tgt_desc *tgt;
1548 const struct lmv_oinfo *oinfo;
1550 if (!lmv_dir_striped(lsm) || !namelen) {
1551 tgt = lmv_fid2tgt(lmv, fid);
1555 *mds = tgt->ltd_index;
1559 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1560 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1561 return ERR_PTR(-EBADF);
1562 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1564 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1566 return ERR_CAST(oinfo);
1569 *fid = oinfo->lmo_fid;
1570 *mds = oinfo->lmo_mds;
1571 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1573 CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1575 return tgt ? tgt : ERR_PTR(-ENODEV);
1579 * Locate MDT of op_data->op_fid1
1581 * For striped directory, it will locate the stripe by name hash, if hash_type
1582 * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1583 * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1584 * indicates whether old or new layout is used to locate.
1586 * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1588 * \param[in] lmv LMV device
1589 * \param[in] op_data client MD stack parameters, name, namelen
1592 * retval pointer to the lmv_tgt_desc if succeed.
1593 * ERR_PTR(errno) if failed.
1595 struct lmv_tgt_desc *
1596 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1598 struct lmv_stripe_md *lsm = op_data->op_mea1;
1599 struct lmv_oinfo *oinfo;
1600 struct lmv_tgt_desc *tgt;
1602 if (lmv_dir_foreign(lsm))
1603 return ERR_PTR(-ENODATA);
1605 /* During creating VOLATILE file, it should honor the mdt
1606 * index if the file under striped dir is being restored, see
1608 if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1609 op_data->op_mds != LMV_OFFSET_DEFAULT) {
1610 tgt = lmv_tgt(lmv, op_data->op_mds);
1612 return ERR_PTR(-ENODEV);
1614 if (lmv_dir_striped(lsm)) {
1617 /* refill the right parent fid */
1618 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1619 oinfo = &lsm->lsm_md_oinfo[i];
1620 if (oinfo->lmo_mds == op_data->op_mds) {
1621 op_data->op_fid1 = oinfo->lmo_fid;
1626 if (i == lsm->lsm_md_stripe_count)
1627 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1629 } else if (lmv_dir_bad_hash(lsm)) {
1630 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1631 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1633 op_data->op_fid1 = oinfo->lmo_fid;
1634 op_data->op_mds = oinfo->lmo_mds;
1635 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1637 return ERR_PTR(-ENODEV);
1639 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1640 op_data->op_name, op_data->op_namelen,
1641 &op_data->op_fid1, &op_data->op_mds,
1642 op_data->op_new_layout);
1648 /* Locate MDT of op_data->op_fid2 for link/rename */
1649 static struct lmv_tgt_desc *
1650 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1652 struct lmv_tgt_desc *tgt;
1655 LASSERT(op_data->op_name);
1656 if (lmv_dir_layout_changing(op_data->op_mea2)) {
1657 struct lu_fid fid1 = op_data->op_fid1;
1658 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1659 struct ptlrpc_request *request = NULL;
1662 * avoid creating new file under old layout of migrating
1663 * directory, check it here.
1665 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1666 op_data->op_name, op_data->op_namelen,
1667 &op_data->op_fid2, &op_data->op_mds, false);
1671 op_data->op_fid1 = op_data->op_fid2;
1672 op_data->op_mea1 = op_data->op_mea2;
1673 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1674 op_data->op_fid1 = fid1;
1675 op_data->op_mea1 = lsm1;
1677 ptlrpc_req_finished(request);
1678 RETURN(ERR_PTR(-EEXIST));
1682 RETURN(ERR_PTR(rc));
1685 return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1686 op_data->op_namelen, &op_data->op_fid2,
1687 &op_data->op_mds, true);
1690 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1692 struct lu_tgt_desc *tgt;
1693 struct ptlrpc_request *request;
1696 LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1697 LASSERT(!op_data->op_new_layout);
1699 tgt = lmv_locate_tgt(lmv, op_data);
1701 return PTR_ERR(tgt);
1703 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1705 ptlrpc_req_finished(request);
1712 /* mkdir by QoS in two cases:
1713 * 1. 'lfs mkdir -i -1'
1714 * 2. parent default LMV master_mdt_index is -1
1716 * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1717 * directories under striped directory.
1719 static inline bool lmv_op_qos_mkdir(const struct md_op_data *op_data)
1721 const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1722 const struct lmv_user_md *lum = op_data->op_data;
1724 if (op_data->op_code != LUSTRE_OPC_MKDIR)
1727 if (lmv_dir_striped(op_data->op_mea1))
1730 if (op_data->op_cli_flags & CLI_SET_MEA && lum &&
1731 (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1732 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1733 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
1736 if (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT)
1742 /* 'lfs mkdir -i <specific_MDT>' */
1743 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1745 const struct lmv_user_md *lum = op_data->op_data;
1747 return op_data->op_code == LUSTRE_OPC_MKDIR &&
1748 op_data->op_cli_flags & CLI_SET_MEA && lum &&
1749 (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1750 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1751 le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1754 /* parent default LMV master_mdt_index is not -1. */
1756 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1758 return op_data->op_code == LUSTRE_OPC_MKDIR &&
1759 op_data->op_default_mea1 &&
1760 op_data->op_default_mea1->lsm_md_master_mdt_index !=
1763 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1764 const void *data, size_t datalen, umode_t mode, uid_t uid,
1765 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1766 struct ptlrpc_request **request)
1768 struct obd_device *obd = exp->exp_obd;
1769 struct lmv_obd *lmv = &obd->u.lmv;
1770 struct lmv_tgt_desc *tgt;
1771 struct mdt_body *repbody;
1776 if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1779 if (lmv_dir_bad_hash(op_data->op_mea1))
1782 if (lmv_dir_layout_changing(op_data->op_mea1)) {
1784 * if parent is migrating, create() needs to lookup existing
1785 * name in both old and new layout, check old layout on client.
1787 rc = lmv_old_layout_lookup(lmv, op_data);
1791 op_data->op_new_layout = true;
1794 tgt = lmv_locate_tgt(lmv, op_data);
1796 RETURN(PTR_ERR(tgt));
1798 if (lmv_op_user_specific_mkdir(op_data)) {
1799 struct lmv_user_md *lum = op_data->op_data;
1801 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1802 tgt = lmv_tgt(lmv, op_data->op_mds);
1805 } else if (lmv_op_default_specific_mkdir(op_data)) {
1807 op_data->op_default_mea1->lsm_md_master_mdt_index;
1808 tgt = lmv_tgt(lmv, op_data->op_mds);
1811 } else if (lmv_op_qos_mkdir(op_data)) {
1812 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds);
1813 if (tgt == ERR_PTR(-EAGAIN))
1814 tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1816 RETURN(PTR_ERR(tgt));
1818 * only update statfs after QoS mkdir, this means the cached
1819 * statfs may be stale, and current mkdir may not follow QoS
1820 * accurately, but it's not serious, and avoids periodic statfs
1821 * when client doesn't mkdir by QoS.
1823 lmv_statfs_check_update(obd, tgt);
1827 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1831 CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1832 (int)op_data->op_namelen, op_data->op_name,
1833 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1836 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1837 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1838 cap_effective, rdev, request);
1840 if (*request == NULL)
1842 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1845 /* dir restripe needs to send to MDT where dir is located */
1846 if (rc != -EREMOTE ||
1847 !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
1850 repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
1851 if (repbody == NULL)
1854 /* Not cross-ref case, just get out of here. */
1855 if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
1858 op_data->op_fid2 = repbody->mbo_fid1;
1859 ptlrpc_req_finished(*request);
1862 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1864 RETURN(PTR_ERR(tgt));
1866 op_data->op_mds = tgt->ltd_index;
1871 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1872 const union ldlm_policy_data *policy, struct md_op_data *op_data,
1873 struct lustre_handle *lockh, __u64 extra_lock_flags)
1875 struct obd_device *obd = exp->exp_obd;
1876 struct lmv_obd *lmv = &obd->u.lmv;
1877 struct lmv_tgt_desc *tgt;
1882 CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1884 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1886 RETURN(PTR_ERR(tgt));
1888 CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1889 PFID(&op_data->op_fid1), tgt->ltd_index);
1891 rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1898 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1899 struct ptlrpc_request **preq)
1901 struct obd_device *obd = exp->exp_obd;
1902 struct lmv_obd *lmv = &obd->u.lmv;
1903 struct lmv_tgt_desc *tgt;
1904 struct mdt_body *body;
1910 tgt = lmv_locate_tgt(lmv, op_data);
1912 RETURN(PTR_ERR(tgt));
1914 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1915 (int)op_data->op_namelen, op_data->op_name,
1916 PFID(&op_data->op_fid1), tgt->ltd_index);
1918 rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1919 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1920 ptlrpc_req_finished(*preq);
1928 body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1929 LASSERT(body != NULL);
1931 if (body->mbo_valid & OBD_MD_MDS) {
1932 op_data->op_fid1 = body->mbo_fid1;
1933 op_data->op_valid |= OBD_MD_FLCROSSREF;
1934 op_data->op_namelen = 0;
1935 op_data->op_name = NULL;
1937 ptlrpc_req_finished(*preq);
1946 #define md_op_data_fid(op_data, fl) \
1947 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1948 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1949 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1950 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1953 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1954 struct md_op_data *op_data, __u32 op_tgt,
1955 enum ldlm_mode mode, int bits, int flag)
1957 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1958 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1959 union ldlm_policy_data policy = { { 0 } };
1963 if (!fid_is_sane(fid))
1967 tgt = lmv_fid2tgt(lmv, fid);
1969 RETURN(PTR_ERR(tgt));
1972 if (tgt->ltd_index != op_tgt) {
1973 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1974 policy.l_inodebits.bits = bits;
1975 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1976 mode, LCF_ASYNC, NULL);
1979 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1981 op_data->op_flags |= flag;
1989 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1992 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1993 struct ptlrpc_request **request)
1995 struct obd_device *obd = exp->exp_obd;
1996 struct lmv_obd *lmv = &obd->u.lmv;
1997 struct lmv_tgt_desc *tgt;
2001 LASSERT(op_data->op_namelen != 0);
2003 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2004 PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2005 op_data->op_name, PFID(&op_data->op_fid1));
2007 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2008 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2009 op_data->op_cap = cfs_curproc_cap_pack();
2011 tgt = lmv_locate_tgt2(lmv, op_data);
2013 RETURN(PTR_ERR(tgt));
2016 * Cancel UPDATE lock on child (fid1).
2018 op_data->op_flags |= MF_MDC_CANCEL_FID2;
2019 rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2020 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2024 rc = md_link(tgt->ltd_exp, op_data, request);
2029 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2030 const char *name, size_t namelen,
2031 struct ptlrpc_request **request)
2033 struct obd_device *obd = exp->exp_obd;
2034 struct lmv_obd *lmv = &obd->u.lmv;
2035 struct lmv_stripe_md *lsm = op_data->op_mea1;
2036 struct lmv_tgt_desc *parent_tgt;
2037 struct lmv_tgt_desc *sp_tgt;
2038 struct lmv_tgt_desc *tp_tgt = NULL;
2039 struct lmv_tgt_desc *child_tgt;
2040 struct lmv_tgt_desc *tgt;
2041 struct lu_fid target_fid = { 0 };
2046 LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2048 CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2049 PFID(&op_data->op_fid1), (int)namelen, name);
2051 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2052 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2053 op_data->op_cap = cfs_curproc_cap_pack();
2055 parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2056 if (IS_ERR(parent_tgt))
2057 RETURN(PTR_ERR(parent_tgt));
2059 if (lmv_dir_striped(lsm)) {
2060 const struct lmv_oinfo *oinfo;
2062 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2064 RETURN(PTR_ERR(oinfo));
2066 /* save source stripe FID in fid4 temporarily for ELC */
2067 op_data->op_fid4 = oinfo->lmo_fid;
2068 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2073 * if parent is being migrated too, fill op_fid2 with target
2074 * stripe fid, otherwise the target stripe is not created yet.
2076 if (lmv_dir_layout_changing(lsm)) {
2077 oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2080 RETURN(PTR_ERR(oinfo));
2082 op_data->op_fid2 = oinfo->lmo_fid;
2083 tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2088 sp_tgt = parent_tgt;
2091 child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2092 if (IS_ERR(child_tgt))
2093 RETURN(PTR_ERR(child_tgt));
2095 /* for directory, migrate to MDT specified by lum_stripe_offset;
2096 * otherwise migrate to the target stripe of parent, but parent
2097 * directory may have finished migration (normally current file too),
2098 * allocate FID on MDT lum_stripe_offset, and server will check
2099 * whether file was migrated already.
2101 if (S_ISDIR(op_data->op_mode) || !tp_tgt) {
2102 struct lmv_user_md *lum = op_data->op_data;
2104 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2106 op_data->op_mds = tp_tgt->ltd_index;
2108 rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2113 * for directory, send migrate request to the MDT where the object will
2114 * be migrated to, because we can't create a striped directory remotely.
2116 * otherwise, send to the MDT where source is located because regular
2117 * file may open lease.
2119 * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2120 * backward compatibility.
2122 if (S_ISDIR(op_data->op_mode) &&
2123 (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2124 tgt = lmv_fid2tgt(lmv, &target_fid);
2126 RETURN(PTR_ERR(tgt));
2131 /* cancel UPDATE lock of parent master object */
2132 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2133 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2137 /* cancel UPDATE lock of source parent */
2138 if (sp_tgt != parent_tgt) {
2140 * migrate RPC packs master object FID, because we can only pack
2141 * two FIDs in reint RPC, but MDS needs to know both source
2142 * parent and target parent, and it will obtain them from master
2143 * FID and LMV, the other FID in RPC is kept for target.
2145 * since this FID is not passed to MDC, cancel it anyway.
2147 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2148 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2152 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2154 op_data->op_fid4 = target_fid;
2156 /* cancel UPDATE locks of target parent */
2157 rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2158 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2162 /* cancel LOOKUP lock of source if source is remote object */
2163 if (child_tgt != sp_tgt) {
2164 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2165 LCK_EX, MDS_INODELOCK_LOOKUP,
2166 MF_MDC_CANCEL_FID3);
2171 /* cancel ELC locks of source */
2172 rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2173 MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2177 rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2182 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2183 const char *old, size_t oldlen,
2184 const char *new, size_t newlen,
2185 struct ptlrpc_request **request)
2187 struct obd_device *obd = exp->exp_obd;
2188 struct lmv_obd *lmv = &obd->u.lmv;
2189 struct lmv_tgt_desc *sp_tgt;
2190 struct lmv_tgt_desc *tp_tgt = NULL;
2191 struct lmv_tgt_desc *src_tgt = NULL;
2192 struct lmv_tgt_desc *tgt;
2193 struct mdt_body *body;
2198 LASSERT(oldlen != 0);
2200 if (op_data->op_cli_flags & CLI_MIGRATE) {
2201 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2205 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2206 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2207 op_data->op_cap = cfs_curproc_cap_pack();
2209 op_data->op_name = new;
2210 op_data->op_namelen = newlen;
2212 tp_tgt = lmv_locate_tgt2(lmv, op_data);
2214 RETURN(PTR_ERR(tp_tgt));
2216 /* Since the target child might be destroyed, and it might become
2217 * orphan, and we can only check orphan on the local MDT right now, so
2218 * we send rename request to the MDT where target child is located. If
2219 * target child does not exist, then it will send the request to the
2221 if (fid_is_sane(&op_data->op_fid4)) {
2222 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2224 RETURN(PTR_ERR(tgt));
2229 op_data->op_flags |= MF_MDC_CANCEL_FID4;
2231 /* cancel UPDATE locks of target parent */
2232 rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2233 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2237 if (fid_is_sane(&op_data->op_fid4)) {
2238 /* cancel LOOKUP lock of target on target parent */
2239 if (tgt != tp_tgt) {
2240 rc = lmv_early_cancel(exp, tp_tgt, op_data,
2241 tgt->ltd_index, LCK_EX,
2242 MDS_INODELOCK_LOOKUP,
2243 MF_MDC_CANCEL_FID4);
2249 if (fid_is_sane(&op_data->op_fid3)) {
2250 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2251 if (IS_ERR(src_tgt))
2252 RETURN(PTR_ERR(src_tgt));
2254 /* cancel ELC locks of source */
2255 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2256 LCK_EX, MDS_INODELOCK_ELC,
2257 MF_MDC_CANCEL_FID3);
2262 op_data->op_name = old;
2263 op_data->op_namelen = oldlen;
2265 sp_tgt = lmv_locate_tgt(lmv, op_data);
2267 RETURN(PTR_ERR(sp_tgt));
2269 /* cancel UPDATE locks of source parent */
2270 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2271 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2275 if (fid_is_sane(&op_data->op_fid3)) {
2276 /* cancel LOOKUP lock of source on source parent */
2277 if (src_tgt != sp_tgt) {
2278 rc = lmv_early_cancel(exp, sp_tgt, op_data,
2279 tgt->ltd_index, LCK_EX,
2280 MDS_INODELOCK_LOOKUP,
2281 MF_MDC_CANCEL_FID3);
2288 CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2289 PFID(&op_data->op_fid1), (int)oldlen, old,
2290 PFID(&op_data->op_fid2), (int)newlen, new);
2292 rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2294 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2295 ptlrpc_req_finished(*request);
2300 if (rc && rc != -EXDEV)
2303 body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2307 /* Not cross-ref case, just get out of here. */
2308 if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2311 op_data->op_fid4 = body->mbo_fid1;
2313 ptlrpc_req_finished(*request);
2316 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2318 RETURN(PTR_ERR(tgt));
2320 if (fid_is_sane(&op_data->op_fid4)) {
2321 /* cancel LOOKUP lock of target on target parent */
2322 if (tgt != tp_tgt) {
2323 rc = lmv_early_cancel(exp, tp_tgt, op_data,
2324 tgt->ltd_index, LCK_EX,
2325 MDS_INODELOCK_LOOKUP,
2326 MF_MDC_CANCEL_FID4);
2335 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2336 void *ea, size_t ealen, struct ptlrpc_request **request)
2338 struct obd_device *obd = exp->exp_obd;
2339 struct lmv_obd *lmv = &obd->u.lmv;
2340 struct lmv_tgt_desc *tgt;
2345 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2346 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2347 op_data->op_xvalid);
2349 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2350 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2352 RETURN(PTR_ERR(tgt));
2354 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2359 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2360 struct ptlrpc_request **request)
2362 struct obd_device *obd = exp->exp_obd;
2363 struct lmv_obd *lmv = &obd->u.lmv;
2364 struct lmv_tgt_desc *tgt;
2369 tgt = lmv_fid2tgt(lmv, fid);
2371 RETURN(PTR_ERR(tgt));
2373 rc = md_fsync(tgt->ltd_exp, fid, request);
2377 struct stripe_dirent {
2378 struct page *sd_page;
2379 struct lu_dirpage *sd_dp;
2380 struct lu_dirent *sd_ent;
2384 struct lmv_dir_ctxt {
2385 struct lmv_obd *ldc_lmv;
2386 struct md_op_data *ldc_op_data;
2387 struct md_callback *ldc_cb_op;
2390 struct stripe_dirent ldc_stripes[0];
2393 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2395 if (stripe->sd_page) {
2396 kunmap(stripe->sd_page);
2397 put_page(stripe->sd_page);
2398 stripe->sd_page = NULL;
2399 stripe->sd_ent = NULL;
2403 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2407 for (i = 0; i < ctxt->ldc_count; i++)
2408 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2411 /* if @ent is dummy, or . .., get next */
2412 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2413 struct lu_dirent *ent,
2416 for (; ent; ent = lu_dirent_next(ent)) {
2417 /* Skip dummy entry */
2418 if (le16_to_cpu(ent->lde_namelen) == 0)
2421 /* skip . and .. for other stripes */
2423 (strncmp(ent->lde_name, ".",
2424 le16_to_cpu(ent->lde_namelen)) == 0 ||
2425 strncmp(ent->lde_name, "..",
2426 le16_to_cpu(ent->lde_namelen)) == 0))
2429 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2436 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2437 struct stripe_dirent *stripe,
2440 struct md_op_data *op_data = ctxt->ldc_op_data;
2441 struct lmv_oinfo *oinfo;
2442 struct lu_fid fid = op_data->op_fid1;
2443 struct inode *inode = op_data->op_data;
2444 struct lmv_tgt_desc *tgt;
2445 struct lu_dirent *ent = stripe->sd_ent;
2446 __u64 hash = ctxt->ldc_hash;
2451 LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2455 if (stripe->sd_page) {
2456 __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2458 /* @hash should be the last dirent hash */
2459 LASSERTF(hash <= end,
2460 "ctxt@%p stripe@%p hash %llx end %llx\n",
2461 ctxt, stripe, hash, end);
2462 /* unload last page */
2463 stripe_dirent_unload(stripe);
2465 if (end == MDS_DIR_END_OFF) {
2466 stripe->sd_eof = true;
2472 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2473 if (!oinfo->lmo_root) {
2478 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2484 /* op_data is shared by stripes, reset after use */
2485 op_data->op_fid1 = oinfo->lmo_fid;
2486 op_data->op_fid2 = oinfo->lmo_fid;
2487 op_data->op_data = oinfo->lmo_root;
2489 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2492 op_data->op_fid1 = fid;
2493 op_data->op_fid2 = fid;
2494 op_data->op_data = inode;
2499 stripe->sd_dp = page_address(stripe->sd_page);
2500 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2502 /* in case a page filled with ., .. and dummy, read next */
2505 stripe->sd_ent = ent;
2508 /* treat error as eof, so dir can be partially accessed */
2509 stripe->sd_eof = true;
2510 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2511 "directory is partially accessed!\n",
2512 PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2519 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2521 struct obd_device *obd = exp->exp_obd;
2522 struct lmv_obd *lmv = &obd->u.lmv;
2523 struct lmv_tgt_desc *tgt;
2528 rc = lmv_check_connect(obd);
2532 tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2534 RETURN(PTR_ERR(tgt));
2536 data->op_flags |= MF_MDC_CANCEL_FID1;
2537 rc = md_file_resync(tgt->ltd_exp, data);
2542 * Get dirent with the closest hash for striped directory
2544 * This function will search the dir entry, whose hash value is the
2545 * closest(>=) to hash from all of sub-stripes, and it is only being called
2546 * for striped directory.
2548 * \param[in] ctxt dir read context
2550 * \retval dirent get the entry successfully
2551 * NULL does not get the entry, normally it means
2552 * it reaches the end of the directory, while read
2553 * stripe dirent error is ignored to allow partial
2556 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2558 struct stripe_dirent *stripe;
2559 struct lu_dirent *ent = NULL;
2563 /* TODO: optimize with k-way merge sort */
2564 for (i = 0; i < ctxt->ldc_count; i++) {
2565 stripe = &ctxt->ldc_stripes[i];
2569 if (!stripe->sd_ent) {
2570 stripe_dirent_load(ctxt, stripe, i);
2571 if (!stripe->sd_ent) {
2572 LASSERT(stripe->sd_eof);
2578 le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2579 le64_to_cpu(stripe->sd_ent->lde_hash)) {
2581 if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2588 stripe = &ctxt->ldc_stripes[min];
2589 ent = stripe->sd_ent;
2590 /* pop found dirent */
2591 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2599 * Build dir entry page for striped directory
2601 * This function gets one entry by @offset from a striped directory. It will
2602 * read entries from all of stripes, and choose one closest to the required
2603 * offset(&offset). A few notes
2604 * 1. skip . and .. for non-zero stripes, because there can only have one .
2605 * and .. in a directory.
2606 * 2. op_data will be shared by all of stripes, instead of allocating new
2607 * one, so need to restore before reusing.
2609 * \param[in] exp obd export refer to LMV
2610 * \param[in] op_data hold those MD parameters of read_entry
2611 * \param[in] cb_op ldlm callback being used in enqueue in mdc_read_entry
2612 * \param[in] offset starting hash offset
2613 * \param[out] ppage the page holding the entry. Note: because the entry
2614 * will be accessed in upper layer, so we need hold the
2615 * page until the usages of entry is finished, see
2616 * ll_dir_entry_next.
2618 * retval =0 if get entry successfully
2619 * <0 cannot get entry
2621 static int lmv_striped_read_page(struct obd_export *exp,
2622 struct md_op_data *op_data,
2623 struct md_callback *cb_op,
2624 __u64 offset, struct page **ppage)
2626 struct page *page = NULL;
2627 struct lu_dirpage *dp;
2629 struct lu_dirent *ent;
2630 struct lu_dirent *last_ent;
2632 struct lmv_dir_ctxt *ctxt;
2633 struct lu_dirent *next = NULL;
2639 /* Allocate a page and read entries from all of stripes and fill
2640 * the page by hash order */
2641 page = alloc_page(GFP_KERNEL);
2645 /* Initialize the entry page */
2647 memset(dp, 0, sizeof(*dp));
2648 dp->ldp_hash_start = cpu_to_le64(offset);
2651 left_bytes = PAGE_SIZE - sizeof(*dp);
2655 /* initalize dir read context */
2656 stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2657 OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2659 GOTO(free_page, rc = -ENOMEM);
2660 ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2661 ctxt->ldc_op_data = op_data;
2662 ctxt->ldc_cb_op = cb_op;
2663 ctxt->ldc_hash = offset;
2664 ctxt->ldc_count = stripe_count;
2667 next = lmv_dirent_next(ctxt);
2669 /* end of directory */
2671 ctxt->ldc_hash = MDS_DIR_END_OFF;
2674 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2676 ent_size = le16_to_cpu(next->lde_reclen);
2678 /* the last entry lde_reclen is 0, but it might not be the last
2679 * one of this temporay dir page */
2681 ent_size = lu_dirent_calc_size(
2682 le16_to_cpu(next->lde_namelen),
2683 le32_to_cpu(next->lde_attrs));
2685 if (ent_size > left_bytes)
2688 memcpy(ent, next, ent_size);
2690 /* Replace . with master FID and Replace .. with the parent FID
2691 * of master object */
2692 if (strncmp(ent->lde_name, ".",
2693 le16_to_cpu(ent->lde_namelen)) == 0 &&
2694 le16_to_cpu(ent->lde_namelen) == 1)
2695 fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2696 else if (strncmp(ent->lde_name, "..",
2697 le16_to_cpu(ent->lde_namelen)) == 0 &&
2698 le16_to_cpu(ent->lde_namelen) == 2)
2699 fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2701 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2702 le16_to_cpu(ent->lde_namelen), ent->lde_name,
2703 le64_to_cpu(ent->lde_hash));
2705 left_bytes -= ent_size;
2706 ent->lde_reclen = cpu_to_le16(ent_size);
2708 ent = (void *)ent + ent_size;
2711 last_ent->lde_reclen = 0;
2714 dp->ldp_flags |= LDF_EMPTY;
2715 else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2716 dp->ldp_flags |= LDF_COLLIDE;
2717 dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2718 dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2720 put_lmv_dir_ctxt(ctxt);
2721 OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2734 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2735 struct md_callback *cb_op, __u64 offset,
2736 struct page **ppage)
2738 struct obd_device *obd = exp->exp_obd;
2739 struct lmv_obd *lmv = &obd->u.lmv;
2740 struct lmv_tgt_desc *tgt;
2745 if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2748 if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2749 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2753 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2755 RETURN(PTR_ERR(tgt));
2757 rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2763 * Unlink a file/directory
2765 * Unlink a file or directory under the parent dir. The unlink request
2766 * usually will be sent to the MDT where the child is located, but if
2767 * the client does not have the child FID then request will be sent to the
2768 * MDT where the parent is located.
2770 * If the parent is a striped directory then it also needs to locate which
2771 * stripe the name of the child is located, and replace the parent FID
2772 * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2773 * it will walk through all of sub-stripes until the child is being
2776 * \param[in] exp export refer to LMV
2777 * \param[in] op_data different parameters transferred beween client
2778 * MD stacks, name, namelen, FIDs etc.
2779 * op_fid1 is the parent FID, op_fid2 is the child
2781 * \param[out] request point to the request of unlink.
2783 * retval 0 if succeed
2784 * negative errno if failed.
2786 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2787 struct ptlrpc_request **request)
2789 struct obd_device *obd = exp->exp_obd;
2790 struct lmv_obd *lmv = &obd->u.lmv;
2791 struct lmv_tgt_desc *tgt;
2792 struct lmv_tgt_desc *parent_tgt;
2793 struct mdt_body *body;
2798 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2799 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2800 op_data->op_cap = cfs_curproc_cap_pack();
2803 parent_tgt = lmv_locate_tgt(lmv, op_data);
2804 if (IS_ERR(parent_tgt))
2805 RETURN(PTR_ERR(parent_tgt));
2807 if (likely(!fid_is_zero(&op_data->op_fid2))) {
2808 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2810 RETURN(PTR_ERR(tgt));
2816 * If child's fid is given, cancel unused locks for it if it is from
2817 * another export than parent.
2819 * LOOKUP lock for child (fid3) should also be cancelled on parent
2820 * tgt_tgt in mdc_unlink().
2822 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2824 if (parent_tgt != tgt)
2825 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2826 LCK_EX, MDS_INODELOCK_LOOKUP,
2827 MF_MDC_CANCEL_FID3);
2829 rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2830 MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2834 CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2835 PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2838 rc = md_unlink(tgt->ltd_exp, op_data, request);
2839 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2840 ptlrpc_req_finished(*request);
2848 body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2852 /* Not cross-ref case, just get out of here. */
2853 if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2856 /* This is a remote object, try remote MDT. */
2857 op_data->op_fid2 = body->mbo_fid1;
2858 ptlrpc_req_finished(*request);
2861 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2863 RETURN(PTR_ERR(tgt));
2868 static int lmv_precleanup(struct obd_device *obd)
2871 libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2872 fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2873 lprocfs_obd_cleanup(obd);
2874 lprocfs_free_md_stats(obd);
2879 * Get by key a value associated with a LMV device.
2881 * Dispatch request to lower-layer devices as needed.
2883 * \param[in] env execution environment for this thread
2884 * \param[in] exp export for the LMV device
2885 * \param[in] keylen length of key identifier
2886 * \param[in] key identifier of key to get value for
2887 * \param[in] vallen size of \a val
2888 * \param[out] val pointer to storage location for value
2889 * \param[in] lsm optional striping metadata of object
2891 * \retval 0 on success
2892 * \retval negative negated errno on failure
2894 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2895 __u32 keylen, void *key, __u32 *vallen, void *val)
2897 struct obd_device *obd;
2898 struct lmv_obd *lmv;
2899 struct lu_tgt_desc *tgt;
2904 obd = class_exp2obd(exp);
2906 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2907 exp->exp_handle.h_cookie);
2912 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2913 LASSERT(*vallen == sizeof(__u32));
2914 lmv_foreach_connected_tgt(lmv, tgt) {
2915 if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2920 } else if (KEY_IS(KEY_MAX_EASIZE) ||
2921 KEY_IS(KEY_DEFAULT_EASIZE) ||
2922 KEY_IS(KEY_CONN_DATA)) {
2924 * Forwarding this request to first MDS, it should know LOV
2927 tgt = lmv_tgt(lmv, 0);
2931 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
2932 if (!rc && KEY_IS(KEY_CONN_DATA))
2933 exp->exp_connect_data = *(struct obd_connect_data *)val;
2935 } else if (KEY_IS(KEY_TGT_COUNT)) {
2936 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
2940 CDEBUG(D_IOCTL, "Invalid key\n");
2944 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
2945 int *__rcs, struct ptlrpc_request_set *_set)
2947 struct obd_device *obd = class_exp2obd(exp);
2948 struct ptlrpc_request_set *set = _set;
2949 struct lmv_obd *lmv = &obd->u.lmv;
2950 int tgt_count = lmv->lmv_mdt_count;
2951 struct lu_tgt_desc *tgt;
2952 struct fid_array *fat, **fas = NULL;
2953 int i, rc, **rcs = NULL;
2956 set = ptlrpc_prep_set();
2961 /* split FIDs by targets */
2962 OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
2964 GOTO(out, rc = -ENOMEM);
2965 OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
2967 GOTO(out_fas, rc = -ENOMEM);
2969 for (i = 0; i < fa->fa_nr; i++) {
2972 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
2974 CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
2975 PFID(&fa->fa_fids[i]), rc);
2978 LASSERT(idx < tgt_count);
2980 OBD_ALLOC(fas[idx], offsetof(struct fid_array,
2981 fa_fids[fa->fa_nr]));
2983 GOTO(out, rc = -ENOMEM);
2985 OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
2987 GOTO(out, rc = -ENOMEM);
2990 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
2993 lmv_foreach_connected_tgt(lmv, tgt) {
2994 fat = fas[tgt->ltd_index];
2995 if (!fat || fat->fa_nr == 0)
2997 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3000 rc = ptlrpc_set_wait(NULL, set);
3003 for (i = 0; i < tgt_count; i++) {
3005 if (!fat || fat->fa_nr == 0)
3007 /* copy FIDs back */
3008 memcpy(fa->fa_fids + j, fat->fa_fids,
3009 fat->fa_nr * sizeof(struct lu_fid));
3011 memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3016 ptlrpc_set_destroy(set);
3019 for (i = 0; i < tgt_count; i++) {
3021 OBD_FREE(fas[i], offsetof(struct fid_array,
3022 fa_fids[fa->fa_nr]));
3024 OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3027 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3030 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3036 * Asynchronously set by key a value associated with a LMV device.
3038 * Dispatch request to lower-layer devices as needed.
3040 * \param[in] env execution environment for this thread
3041 * \param[in] exp export for the LMV device
3042 * \param[in] keylen length of key identifier
3043 * \param[in] key identifier of key to store value for
3044 * \param[in] vallen size of value to store
3045 * \param[in] val pointer to data to be stored
3046 * \param[in] set optional list of related ptlrpc requests
3048 * \retval 0 on success
3049 * \retval negative negated errno on failure
3051 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3052 __u32 keylen, void *key, __u32 vallen, void *val,
3053 struct ptlrpc_request_set *set)
3055 struct lmv_tgt_desc *tgt = NULL;
3056 struct obd_device *obd;
3057 struct lmv_obd *lmv;
3061 obd = class_exp2obd(exp);
3063 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3064 exp->exp_handle.h_cookie);
3069 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3070 KEY_IS(KEY_DEFAULT_EASIZE)) {
3073 lmv_foreach_connected_tgt(lmv, tgt) {
3074 err = obd_set_info_async(env, tgt->ltd_exp,
3075 keylen, key, vallen, val, set);
3086 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3087 const struct lmv_mds_md_v1 *lmm1)
3089 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3096 lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3097 lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3098 lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3099 if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3100 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3102 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3103 lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3104 lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3105 lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3106 cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3107 sizeof(lsm->lsm_md_pool_name));
3109 if (cplen >= sizeof(lsm->lsm_md_pool_name))
3112 CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3113 "layout_version %d\n", lsm->lsm_md_stripe_count,
3114 lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3115 lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3116 lsm->lsm_md_layout_version);
3118 stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3119 for (i = 0; i < stripe_count; i++) {
3120 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3121 &lmm1->lmv_stripe_fids[i]);
3123 * set default value -1, so lmv_locate_tgt() knows this stripe
3124 * target is not initialized.
3126 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3127 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3130 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3131 &lsm->lsm_md_oinfo[i].lmo_mds);
3138 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3139 PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3145 static inline int lmv_unpack_user_md(struct obd_export *exp,
3146 struct lmv_stripe_md *lsm,
3147 const struct lmv_user_md *lmu)
3149 lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3150 lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3151 lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3152 lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3153 lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3158 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3159 const union lmv_mds_md *lmm, size_t lmm_size)
3161 struct lmv_stripe_md *lsm;
3164 bool allocated = false;
3167 LASSERT(lsmp != NULL);
3171 if (lsm != NULL && lmm == NULL) {
3173 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3175 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3178 lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3180 OBD_FREE_LARGE(lfm, lfm_size);
3184 if (lmv_dir_striped(lsm)) {
3185 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3186 if (lsm->lsm_md_oinfo[i].lmo_root)
3187 iput(lsm->lsm_md_oinfo[i].lmo_root);
3189 lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3191 lsm_size = lmv_stripe_md_size(0);
3193 OBD_FREE(lsm, lsm_size);
3198 /* foreign lmv case */
3199 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3200 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3203 OBD_ALLOC_LARGE(lfm, lmm_size);
3206 *lsmp = (struct lmv_stripe_md *)lfm;
3208 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3209 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3210 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3211 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3212 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3217 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3221 if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3222 le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3223 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3224 exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3229 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3230 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3233 * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3234 * stripecount should be 0 then.
3236 lsm_size = lmv_stripe_md_size(0);
3239 OBD_ALLOC(lsm, lsm_size);
3246 switch (le32_to_cpu(lmm->lmv_magic)) {
3248 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3250 case LMV_USER_MAGIC:
3251 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3254 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3255 le32_to_cpu(lmm->lmv_magic));
3260 if (rc != 0 && allocated) {
3261 OBD_FREE(lsm, lsm_size);
3268 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3270 lmv_unpackmd(NULL, &lsm, NULL, 0);
3272 EXPORT_SYMBOL(lmv_free_memmd);
3274 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3275 union ldlm_policy_data *policy,
3276 enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3279 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3280 struct lu_tgt_desc *tgt;
3286 LASSERT(fid != NULL);
3288 lmv_foreach_connected_tgt(lmv, tgt) {
3289 if (!tgt->ltd_active)
3292 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3300 static int lmv_set_lock_data(struct obd_export *exp,
3301 const struct lustre_handle *lockh,
3302 void *data, __u64 *bits)
3304 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3305 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3310 if (tgt == NULL || tgt->ltd_exp == NULL)
3312 rc = md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3316 enum ldlm_mode lmv_lock_match(struct obd_export *exp, __u64 flags,
3317 const struct lu_fid *fid, enum ldlm_type type,
3318 union ldlm_policy_data *policy,
3319 enum ldlm_mode mode, struct lustre_handle *lockh)
3321 struct obd_device *obd = exp->exp_obd;
3322 struct lmv_obd *lmv = &obd->u.lmv;
3324 struct lu_tgt_desc *tgt;
3330 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3333 * With DNE every object can have two locks in different namespaces:
3334 * lookup lock in space of MDT storing direntry and update/open lock in
3335 * space of MDT storing inode. Try the MDT that the FID maps to first,
3336 * since this can be easily found, and only try others if that fails.
3338 for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3339 i < lmv->lmv_mdt_descs.ltd_tgts_size;
3340 i++, index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3342 CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3343 obd->obd_name, PFID(fid), index);
3347 tgt = lmv_tgt(lmv, index);
3348 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3351 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3360 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
3361 struct obd_export *dt_exp, struct obd_export *md_exp,
3362 struct lustre_md *md)
3364 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3365 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3367 if (!tgt || !tgt->ltd_exp)
3370 return md_get_lustre_md(tgt->ltd_exp, req, dt_exp, md_exp, md);
3373 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3375 struct obd_device *obd = exp->exp_obd;
3376 struct lmv_obd *lmv = &obd->u.lmv;
3377 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3381 if (md->default_lmv) {
3382 lmv_free_memmd(md->default_lmv);
3383 md->default_lmv = NULL;
3385 if (md->lmv != NULL) {
3386 lmv_free_memmd(md->lmv);
3389 if (!tgt || !tgt->ltd_exp)
3391 RETURN(md_free_lustre_md(tgt->ltd_exp, md));
3394 int lmv_set_open_replay_data(struct obd_export *exp,
3395 struct obd_client_handle *och,
3396 struct lookup_intent *it)
3398 struct obd_device *obd = exp->exp_obd;
3399 struct lmv_obd *lmv = &obd->u.lmv;
3400 struct lmv_tgt_desc *tgt;
3404 tgt = lmv_fid2tgt(lmv, &och->och_fid);
3406 RETURN(PTR_ERR(tgt));
3408 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3411 int lmv_clear_open_replay_data(struct obd_export *exp,
3412 struct obd_client_handle *och)
3414 struct obd_device *obd = exp->exp_obd;
3415 struct lmv_obd *lmv = &obd->u.lmv;
3416 struct lmv_tgt_desc *tgt;
3420 tgt = lmv_fid2tgt(lmv, &och->och_fid);
3422 RETURN(PTR_ERR(tgt));
3424 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3427 int lmv_intent_getattr_async(struct obd_export *exp,
3428 struct md_enqueue_info *minfo)
3430 struct md_op_data *op_data = &minfo->mi_data;
3431 struct obd_device *obd = exp->exp_obd;
3432 struct lmv_obd *lmv = &obd->u.lmv;
3433 struct lmv_tgt_desc *ptgt;
3434 struct lmv_tgt_desc *ctgt;
3439 if (!fid_is_sane(&op_data->op_fid2))
3442 ptgt = lmv_locate_tgt(lmv, op_data);
3444 RETURN(PTR_ERR(ptgt));
3446 ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3448 RETURN(PTR_ERR(ctgt));
3451 * remote object needs two RPCs to lookup and getattr, considering the
3452 * complexity don't support statahead for now.
3457 rc = md_intent_getattr_async(ptgt->ltd_exp, minfo);
3462 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3463 struct lu_fid *fid, __u64 *bits)
3465 struct obd_device *obd = exp->exp_obd;
3466 struct lmv_obd *lmv = &obd->u.lmv;
3467 struct lmv_tgt_desc *tgt;
3472 tgt = lmv_fid2tgt(lmv, fid);
3474 RETURN(PTR_ERR(tgt));
3476 rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3480 int lmv_get_fid_from_lsm(struct obd_export *exp,
3481 const struct lmv_stripe_md *lsm,
3482 const char *name, int namelen, struct lu_fid *fid)
3484 const struct lmv_oinfo *oinfo;
3486 LASSERT(lmv_dir_striped(lsm));
3488 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3490 return PTR_ERR(oinfo);
3492 *fid = oinfo->lmo_fid;
3498 * For lmv, only need to send request to master MDT, and the master MDT will
3499 * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3500 * we directly fetch data from the slave MDTs.
3502 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3503 struct obd_quotactl *oqctl)
3505 struct obd_device *obd = class_exp2obd(exp);
3506 struct lmv_obd *lmv = &obd->u.lmv;
3507 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3508 __u64 curspace, curinodes;
3513 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3514 CERROR("master lmv inactive\n");
3518 if (oqctl->qc_cmd != Q_GETOQUOTA) {
3519 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3523 curspace = curinodes = 0;
3524 lmv_foreach_connected_tgt(lmv, tgt) {
3527 if (!tgt->ltd_active)
3530 err = obd_quotactl(tgt->ltd_exp, oqctl);
3532 CERROR("getquota on mdt %d failed. %d\n",
3533 tgt->ltd_index, err);
3537 curspace += oqctl->qc_dqblk.dqb_curspace;
3538 curinodes += oqctl->qc_dqblk.dqb_curinodes;
3541 oqctl->qc_dqblk.dqb_curspace = curspace;
3542 oqctl->qc_dqblk.dqb_curinodes = curinodes;
3547 static int lmv_merge_attr(struct obd_export *exp,
3548 const struct lmv_stripe_md *lsm,
3549 struct cl_attr *attr,
3550 ldlm_blocking_callback cb_blocking)
3555 if (!lmv_dir_striped(lsm))
3558 rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3562 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3563 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3569 "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3570 PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3571 i_size_read(inode), (unsigned long long)inode->i_blocks,
3572 inode->i_nlink, (s64)inode->i_atime.tv_sec,
3573 (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3575 /* for slave stripe, it needs to subtract nlink for . and .. */
3577 attr->cat_nlink += inode->i_nlink - 2;
3579 attr->cat_nlink = inode->i_nlink;
3581 attr->cat_size += i_size_read(inode);
3582 attr->cat_blocks += inode->i_blocks;
3584 if (attr->cat_atime < inode->i_atime.tv_sec)
3585 attr->cat_atime = inode->i_atime.tv_sec;
3587 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3588 attr->cat_ctime = inode->i_ctime.tv_sec;
3590 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3591 attr->cat_mtime = inode->i_mtime.tv_sec;
3596 static const struct obd_ops lmv_obd_ops = {
3597 .o_owner = THIS_MODULE,
3598 .o_setup = lmv_setup,
3599 .o_cleanup = lmv_cleanup,
3600 .o_precleanup = lmv_precleanup,
3601 .o_process_config = lmv_process_config,
3602 .o_connect = lmv_connect,
3603 .o_disconnect = lmv_disconnect,
3604 .o_statfs = lmv_statfs,
3605 .o_get_info = lmv_get_info,
3606 .o_set_info_async = lmv_set_info_async,
3607 .o_notify = lmv_notify,
3608 .o_get_uuid = lmv_get_uuid,
3609 .o_fid_alloc = lmv_fid_alloc,
3610 .o_iocontrol = lmv_iocontrol,
3611 .o_quotactl = lmv_quotactl
3614 static const struct md_ops lmv_md_ops = {
3615 .m_get_root = lmv_get_root,
3616 .m_null_inode = lmv_null_inode,
3617 .m_close = lmv_close,
3618 .m_create = lmv_create,
3619 .m_enqueue = lmv_enqueue,
3620 .m_getattr = lmv_getattr,
3621 .m_getxattr = lmv_getxattr,
3622 .m_getattr_name = lmv_getattr_name,
3623 .m_intent_lock = lmv_intent_lock,
3625 .m_rename = lmv_rename,
3626 .m_setattr = lmv_setattr,
3627 .m_setxattr = lmv_setxattr,
3628 .m_fsync = lmv_fsync,
3629 .m_file_resync = lmv_file_resync,
3630 .m_read_page = lmv_read_page,
3631 .m_unlink = lmv_unlink,
3632 .m_init_ea_size = lmv_init_ea_size,
3633 .m_cancel_unused = lmv_cancel_unused,
3634 .m_set_lock_data = lmv_set_lock_data,
3635 .m_lock_match = lmv_lock_match,
3636 .m_get_lustre_md = lmv_get_lustre_md,
3637 .m_free_lustre_md = lmv_free_lustre_md,
3638 .m_merge_attr = lmv_merge_attr,
3639 .m_set_open_replay_data = lmv_set_open_replay_data,
3640 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3641 .m_intent_getattr_async = lmv_intent_getattr_async,
3642 .m_revalidate_lock = lmv_revalidate_lock,
3643 .m_get_fid_from_lsm = lmv_get_fid_from_lsm,
3644 .m_unpackmd = lmv_unpackmd,
3645 .m_rmfid = lmv_rmfid,
3648 static int __init lmv_init(void)
3650 return class_register_type(&lmv_obd_ops, &lmv_md_ops, true, NULL,
3651 LUSTRE_LMV_NAME, NULL);
3654 static void __exit lmv_exit(void)
3656 class_unregister_type(LUSTRE_LMV_NAME);
3659 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3660 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3661 MODULE_VERSION(LUSTRE_VERSION_STRING);
3662 MODULE_LICENSE("GPL");
3664 module_init(lmv_init);
3665 module_exit(lmv_exit);