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 #ifdef HAVE_UIDGID_HEADER
40 # include <linux/uidgid.h>
42 #include <linux/slab.h>
43 #include <linux/pagemap.h>
45 #include <linux/math64.h>
46 #include <linux/seq_file.h>
47 #include <linux/namei.h>
49 #include <obd_support.h>
50 #include <lustre_lib.h>
51 #include <lustre_net.h>
52 #include <obd_class.h>
53 #include <lustre_lmv.h>
54 #include <lprocfs_status.h>
55 #include <cl_object.h>
56 #include <lustre_fid.h>
57 #include <uapi/linux/lustre/lustre_ioctl.h>
58 #include <lustre_kernelcomm.h>
59 #include "lmv_internal.h"
61 static int lmv_check_connect(struct obd_device *obd);
63 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
66 if (tgt->ltd_active == activate)
69 tgt->ltd_active = activate;
70 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
73 tgt->ltd_exp->exp_obd->obd_inactive = !activate;
79 * -EINVAL : UUID can't be found in the LMV's target list
80 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
81 * -EBADF : The UUID is found, but the OBD of the wrong type (!)
83 static int lmv_set_mdc_active(struct lmv_obd *lmv,
84 const struct obd_uuid *uuid,
87 struct lu_tgt_desc *tgt = NULL;
88 struct obd_device *obd;
93 CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
94 lmv, uuid->uuid, activate);
96 spin_lock(&lmv->lmv_lock);
97 lmv_foreach_connected_tgt(lmv, tgt) {
98 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
99 tgt->ltd_index, tgt->ltd_uuid.uuid,
100 tgt->ltd_exp->exp_handle.h_cookie);
102 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
107 GOTO(out_lmv_lock, rc = -EINVAL);
109 obd = class_exp2obd(tgt->ltd_exp);
111 GOTO(out_lmv_lock, rc = -ENOTCONN);
113 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
114 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
115 obd->obd_type->typ_name, tgt->ltd_index);
116 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
118 if (tgt->ltd_active == activate) {
119 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
120 activate ? "" : "in");
121 GOTO(out_lmv_lock, rc);
124 CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
125 activate ? "" : "in");
126 lmv_activate_target(lmv, tgt, activate);
130 spin_unlock(&lmv->lmv_lock);
134 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
136 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
137 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
139 return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
142 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
143 enum obd_notify_event ev)
145 struct obd_connect_data *conn_data;
146 struct lmv_obd *lmv = &obd->u.lmv;
147 struct obd_uuid *uuid;
151 if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
152 CERROR("unexpected notification of %s %s!\n",
153 watched->obd_type->typ_name,
158 uuid = &watched->u.cli.cl_target_uuid;
159 if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
161 * Set MDC as active before notifying the observer, so the
162 * observer can use the MDC normally.
164 rc = lmv_set_mdc_active(lmv, uuid,
165 ev == OBD_NOTIFY_ACTIVE);
167 CERROR("%sactivation of %s failed: %d\n",
168 ev == OBD_NOTIFY_ACTIVE ? "" : "de",
172 } else if (ev == OBD_NOTIFY_OCD) {
173 conn_data = &watched->u.cli.cl_import->imp_connect_data;
175 * XXX: Make sure that ocd_connect_flags from all targets are
176 * the same. Otherwise one of MDTs runs wrong version or
177 * something like this. --umka
179 obd->obd_self_export->exp_connect_data = *conn_data;
183 * Pass the notification up the chain.
185 if (obd->obd_observer)
186 rc = obd_notify(obd->obd_observer, watched, ev);
191 static int lmv_connect(const struct lu_env *env,
192 struct obd_export **pexp, struct obd_device *obd,
193 struct obd_uuid *cluuid, struct obd_connect_data *data,
196 struct lmv_obd *lmv = &obd->u.lmv;
197 struct lustre_handle conn = { 0 };
198 struct obd_export *exp;
202 rc = class_connect(&conn, obd, cluuid);
204 CERROR("class_connection() returned %d\n", rc);
208 exp = class_conn2export(&conn);
211 lmv->conn_data = *data;
212 lmv->lmv_cache = localdata;
214 lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
215 &obd->obd_kset.kobj);
216 if (!lmv->lmv_tgts_kobj) {
217 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
218 obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
221 rc = lmv_check_connect(obd);
230 if (lmv->lmv_tgts_kobj)
231 kobject_put(lmv->lmv_tgts_kobj);
233 class_disconnect(exp);
238 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
241 struct obd_device *obd = exp->exp_obd;
242 struct lmv_obd *lmv = &obd->u.lmv;
243 struct lmv_tgt_desc *tgt;
249 if (lmv->max_easize < easize) {
250 lmv->max_easize = easize;
253 if (lmv->max_def_easize < def_easize) {
254 lmv->max_def_easize = def_easize;
261 if (lmv->connected == 0)
264 lmv_foreach_connected_tgt(lmv, tgt) {
265 if (!tgt->ltd_active)
268 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
270 CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
271 " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
278 #define MAX_STRING_SIZE 128
280 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
282 struct lmv_obd *lmv = &obd->u.lmv;
283 struct obd_device *mdc_obd;
284 struct obd_export *mdc_exp;
285 struct lu_fld_target target;
289 mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
292 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
296 CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
297 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
298 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
300 if (!mdc_obd->obd_set_up) {
301 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
305 rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
306 &lmv->conn_data, lmv->lmv_cache);
308 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
313 * Init fid sequence client for this mdc and add new fld target.
315 rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
319 target.ft_srv = NULL;
320 target.ft_exp = mdc_exp;
321 target.ft_idx = tgt->ltd_index;
323 fld_client_add_target(&lmv->lmv_fld, &target);
325 rc = obd_register_observer(mdc_obd, obd);
327 obd_disconnect(mdc_exp);
328 CERROR("target %s register_observer error %d\n",
329 tgt->ltd_uuid.uuid, rc);
333 if (obd->obd_observer) {
335 * Tell the observer about the new target.
337 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
340 obd_disconnect(mdc_exp);
346 tgt->ltd_exp = mdc_exp;
347 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
349 md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
351 rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
353 obd_disconnect(mdc_exp);
357 CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
358 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
359 atomic_read(&obd->obd_refcount));
361 lmv_statfs_check_update(obd, tgt);
363 if (lmv->lmv_tgts_kobj)
364 /* Even if we failed to create the link, that's fine */
365 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
366 &mdc_obd->obd_kset.kobj,
371 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
374 ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
378 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
379 __u32 index, int gen)
381 struct obd_device *mdc_obd;
382 struct lmv_obd *lmv = &obd->u.lmv;
383 struct lmv_tgt_desc *tgt;
384 struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
389 CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
390 mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
393 CERROR("%s: Target %s not attached: rc = %d\n",
394 obd->obd_name, uuidp->uuid, -EINVAL);
402 mutex_init(&tgt->ltd_fid_mutex);
403 tgt->ltd_index = index;
404 tgt->ltd_uuid = *uuidp;
407 mutex_lock(<d->ltd_mutex);
408 rc = ltd_add_tgt(ltd, tgt);
409 mutex_unlock(<d->ltd_mutex);
415 /* lmv_check_connect() will connect this target. */
418 rc = lmv_connect_mdc(obd, tgt);
420 int easize = sizeof(struct lmv_stripe_md) +
421 lmv->lmv_mdt_count * sizeof(struct lu_fid);
423 lmv_init_ea_size(obd->obd_self_export, easize, 0);
433 static int lmv_check_connect(struct obd_device *obd)
435 struct lmv_obd *lmv = &obd->u.lmv;
436 struct lmv_tgt_desc *tgt;
445 mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
447 GOTO(unlock, rc = 0);
449 if (!lmv->lmv_mdt_count) {
450 CERROR("%s: no targets configured: rc = -EINVAL\n",
452 GOTO(unlock, rc = -EINVAL);
455 if (!lmv_mdt0_inited(lmv)) {
456 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
458 GOTO(unlock, rc = -EINVAL);
461 CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
462 obd->obd_uuid.uuid, obd->obd_name);
464 lmv_foreach_tgt(lmv, tgt) {
465 rc = lmv_connect_mdc(obd, tgt);
471 easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
472 lmv_init_ea_size(obd->obd_self_export, easize, 0);
475 mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
480 lmv_foreach_tgt(lmv, tgt) {
485 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
486 obd_disconnect(tgt->ltd_exp);
492 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
494 struct lmv_obd *lmv = &obd->u.lmv;
495 struct obd_device *mdc_obd;
499 LASSERT(tgt != NULL);
500 LASSERT(obd != NULL);
502 mdc_obd = class_exp2obd(tgt->ltd_exp);
505 mdc_obd->obd_force = obd->obd_force;
506 mdc_obd->obd_fail = obd->obd_fail;
507 mdc_obd->obd_no_recov = obd->obd_no_recov;
509 if (lmv->lmv_tgts_kobj)
510 sysfs_remove_link(lmv->lmv_tgts_kobj,
514 rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
516 CERROR("Can't finanize fids factory\n");
518 CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
519 tgt->ltd_exp->exp_obd->obd_name,
520 tgt->ltd_exp->exp_obd->obd_uuid.uuid);
522 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
523 rc = obd_disconnect(tgt->ltd_exp);
525 if (tgt->ltd_active) {
526 CERROR("Target %s disconnect error %d\n",
527 tgt->ltd_uuid.uuid, rc);
531 lmv_activate_target(lmv, tgt, 0);
536 static int lmv_disconnect(struct obd_export *exp)
538 struct obd_device *obd = class_exp2obd(exp);
539 struct lmv_obd *lmv = &obd->u.lmv;
540 struct lmv_tgt_desc *tgt;
545 lmv_foreach_connected_tgt(lmv, tgt)
546 lmv_disconnect_mdc(obd, tgt);
548 if (lmv->lmv_tgts_kobj)
549 kobject_put(lmv->lmv_tgts_kobj);
552 class_export_put(exp);
553 rc = class_disconnect(exp);
559 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
562 struct obd_device *obddev = class_exp2obd(exp);
563 struct lmv_obd *lmv = &obddev->u.lmv;
564 struct getinfo_fid2path *gf;
565 struct lmv_tgt_desc *tgt;
566 struct getinfo_fid2path *remote_gf = NULL;
567 struct lu_fid root_fid;
568 int remote_gf_size = 0;
572 tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
574 RETURN(PTR_ERR(tgt));
576 root_fid = *gf->gf_u.gf_root_fid;
577 LASSERT(fid_is_sane(&root_fid));
580 rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
581 if (rc != 0 && rc != -EREMOTE)
582 GOTO(out_fid2path, rc);
584 /* If remote_gf != NULL, it means just building the
585 * path on the remote MDT, copy this path segement to gf */
586 if (remote_gf != NULL) {
587 struct getinfo_fid2path *ori_gf;
591 ori_gf = (struct getinfo_fid2path *)karg;
592 if (strlen(ori_gf->gf_u.gf_path) + 1 +
593 strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
594 GOTO(out_fid2path, rc = -EOVERFLOW);
596 ptr = ori_gf->gf_u.gf_path;
598 len = strlen(gf->gf_u.gf_path);
599 /* move the current path to the right to release space
600 * for closer-to-root part */
601 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
602 memcpy(ptr, gf->gf_u.gf_path, len);
606 CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
607 tgt->ltd_exp->exp_obd->obd_name,
608 gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
612 GOTO(out_fid2path, rc);
614 /* sigh, has to go to another MDT to do path building further */
615 if (remote_gf == NULL) {
616 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
617 OBD_ALLOC(remote_gf, remote_gf_size);
618 if (remote_gf == NULL)
619 GOTO(out_fid2path, rc = -ENOMEM);
620 remote_gf->gf_pathlen = PATH_MAX;
623 if (!fid_is_sane(&gf->gf_fid)) {
624 CERROR("%s: invalid FID "DFID": rc = %d\n",
625 tgt->ltd_exp->exp_obd->obd_name,
626 PFID(&gf->gf_fid), -EINVAL);
627 GOTO(out_fid2path, rc = -EINVAL);
630 tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
632 GOTO(out_fid2path, rc = -EINVAL);
634 remote_gf->gf_fid = gf->gf_fid;
635 remote_gf->gf_recno = -1;
636 remote_gf->gf_linkno = -1;
637 memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
638 *remote_gf->gf_u.gf_root_fid = root_fid;
640 goto repeat_fid2path;
643 if (remote_gf != NULL)
644 OBD_FREE(remote_gf, remote_gf_size);
648 static int lmv_hsm_req_count(struct lmv_obd *lmv,
649 const struct hsm_user_request *hur,
650 const struct lmv_tgt_desc *tgt_mds)
652 struct lmv_tgt_desc *curr_tgt;
656 /* count how many requests must be sent to the given target */
657 for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
658 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
659 if (IS_ERR(curr_tgt))
660 RETURN(PTR_ERR(curr_tgt));
661 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
667 static int lmv_hsm_req_build(struct lmv_obd *lmv,
668 struct hsm_user_request *hur_in,
669 const struct lmv_tgt_desc *tgt_mds,
670 struct hsm_user_request *hur_out)
673 struct lmv_tgt_desc *curr_tgt;
675 /* build the hsm_user_request for the given target */
676 hur_out->hur_request = hur_in->hur_request;
678 for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
679 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
680 if (IS_ERR(curr_tgt))
681 RETURN(PTR_ERR(curr_tgt));
682 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
683 hur_out->hur_user_item[nr_out] =
684 hur_in->hur_user_item[i];
688 hur_out->hur_request.hr_itemcount = nr_out;
689 memcpy(hur_data(hur_out), hur_data(hur_in),
690 hur_in->hur_request.hr_data_len);
695 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
696 int len, struct lustre_kernelcomm *lk,
699 struct lmv_obd *lmv = &obd->u.lmv;
700 struct lu_tgt_desc *tgt;
705 /* unregister request (call from llapi_hsm_copytool_fini) */
706 lmv_foreach_connected_tgt(lmv, tgt)
707 /* best effort: try to clean as much as possible
708 * (continue on error) */
709 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
711 /* Whatever the result, remove copytool from kuc groups.
712 * Unreached coordinators will get EPIPE on next requests
713 * and will unregister automatically.
715 rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
720 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
721 int len, struct lustre_kernelcomm *lk,
724 struct lmv_obd *lmv = &obd->u.lmv;
726 bool any_set = false;
727 struct kkuc_ct_data *kcd;
729 struct lu_tgt_desc *tgt;
736 filp = fget(lk->lk_wfd);
740 if (lk->lk_flags & LK_FLG_DATANR)
741 kcd_size = offsetof(struct kkuc_ct_data,
742 kcd_archives[lk->lk_data_count]);
744 kcd_size = sizeof(*kcd);
746 OBD_ALLOC(kcd, kcd_size);
748 GOTO(err_fput, rc = -ENOMEM);
750 kcd->kcd_nr_archives = lk->lk_data_count;
751 if (lk->lk_flags & LK_FLG_DATANR) {
752 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
753 if (lk->lk_data_count > 0)
754 memcpy(kcd->kcd_archives, lk->lk_data,
755 sizeof(*kcd->kcd_archives) * lk->lk_data_count);
757 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
760 rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
761 lk->lk_group, kcd, kcd_size);
762 OBD_FREE(kcd, kcd_size);
766 /* All or nothing: try to register to all MDS.
767 * In case of failure, unregister from previous MDS,
768 * except if it because of inactive target. */
769 lmv_foreach_connected_tgt(lmv, tgt) {
770 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
772 if (tgt->ltd_active) {
773 /* permanent error */
774 CERROR("%s: iocontrol MDC %s on MDT"
775 " idx %d cmd %x: err = %d\n",
776 lmv2obd_dev(lmv)->obd_name,
777 tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
780 lk->lk_flags |= LK_FLG_STOP;
782 /* unregister from previous MDS */
783 lmv_foreach_connected_tgt(lmv, tgt) {
784 if (tgt->ltd_index >= i)
787 obd_iocontrol(cmd, tgt->ltd_exp, len,
790 GOTO(err_kkuc_rem, rc);
792 /* else: transient error.
793 * kuc will register to the missing MDT
801 /* no registration done: return error */
802 GOTO(err_kkuc_rem, rc = -ENOTCONN);
807 libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
814 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
815 int len, void *karg, void __user *uarg)
817 struct obd_device *obddev = class_exp2obd(exp);
818 struct lmv_obd *lmv = &obddev->u.lmv;
819 struct lu_tgt_desc *tgt = NULL;
821 __u32 count = lmv->lmv_mdt_count;
830 case IOC_OBD_STATFS: {
831 struct obd_ioctl_data *data = karg;
832 struct obd_device *mdc_obd;
833 struct obd_statfs stat_buf = {0};
836 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
838 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
841 tgt = lmv_tgt(lmv, index);
842 if (!tgt || !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(obddev, cmd, len, lk, uarg);
1027 rc = lmv_hsm_ct_register(obddev, 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 = obddev->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);
1059 * This is _inode_ placement policy function (not name).
1061 static u32 lmv_placement_policy(struct obd_device *obd,
1062 struct md_op_data *op_data)
1064 struct lmv_obd *lmv = &obd->u.lmv;
1065 struct lmv_user_md *lum;
1070 if (lmv->lmv_mdt_count == 1)
1073 lum = op_data->op_data;
1076 * 1. See if the stripe offset is specified by lum.
1077 * 2. If parent has default LMV, and its hash type is "space", choose
1078 * MDT with QoS. (see lmv_locate_tgt_qos()).
1079 * 3. Then check if default LMV stripe offset is not -1.
1080 * 4. Finally choose MDS by name hash if the parent
1081 * is striped directory. (see lmv_locate_tgt()).
1083 * presently explicit MDT location is not supported
1084 * for foreign dirs (as it can't be embedded into free
1085 * format LMV, like with lum_stripe_offset), so we only
1086 * rely on default stripe offset or then name hashing.
1088 if (op_data->op_cli_flags & CLI_SET_MEA && lum != NULL &&
1089 le32_to_cpu(lum->lum_magic != LMV_MAGIC_FOREIGN) &&
1090 le32_to_cpu(lum->lum_stripe_offset) != (__u32)-1) {
1091 mdt = le32_to_cpu(lum->lum_stripe_offset);
1092 } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1093 !lmv_dir_striped(op_data->op_mea1) &&
1094 lmv_dir_qos_mkdir(op_data->op_default_mea1)) {
1095 mdt = op_data->op_mds;
1096 } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1097 op_data->op_default_mea1 &&
1098 op_data->op_default_mea1->lsm_md_master_mdt_index !=
1100 mdt = op_data->op_default_mea1->lsm_md_master_mdt_index;
1101 op_data->op_mds = mdt;
1103 mdt = op_data->op_mds;
1109 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid, u32 mds)
1111 struct lmv_tgt_desc *tgt;
1116 tgt = lmv_tgt(lmv, mds);
1121 * New seq alloc and FLD setup should be atomic. Otherwise we may find
1122 * on server that seq in new allocated fid is not yet known.
1124 mutex_lock(&tgt->ltd_fid_mutex);
1126 if (tgt->ltd_active == 0 || tgt->ltd_exp == NULL)
1127 GOTO(out, rc = -ENODEV);
1130 * Asking underlying tgt layer to allocate new fid.
1132 rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1134 LASSERT(fid_is_sane(fid));
1140 mutex_unlock(&tgt->ltd_fid_mutex);
1144 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1145 struct lu_fid *fid, struct md_op_data *op_data)
1147 struct obd_device *obd = class_exp2obd(exp);
1148 struct lmv_obd *lmv = &obd->u.lmv;
1154 LASSERT(op_data != NULL);
1155 LASSERT(fid != NULL);
1157 mds = lmv_placement_policy(obd, op_data);
1159 rc = __lmv_fid_alloc(lmv, fid, mds);
1161 CERROR("Can't alloc new fid, rc %d\n", rc);
1166 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1168 struct lmv_obd *lmv = &obd->u.lmv;
1169 struct lmv_desc *desc;
1170 struct lnet_process_id lnet_id;
1176 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1177 CERROR("LMV setup requires a descriptor\n");
1181 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1182 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1183 CERROR("Lmv descriptor size wrong: %d > %d\n",
1184 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1188 obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1189 desc->ld_uuid.uuid);
1190 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1191 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1192 lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1193 LMV_DESC_QOS_MAXAGE_DEFAULT;
1194 lmv->max_def_easize = 0;
1195 lmv->max_easize = 0;
1197 spin_lock_init(&lmv->lmv_lock);
1200 * initialize rr_index to lower 32bit of netid, so that client
1201 * can distribute subdirs evenly from the beginning.
1203 while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1204 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1205 lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1210 rc = lmv_tunables_init(obd);
1212 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1215 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1216 LUSTRE_CLI_FLD_HASH_DHT);
1218 CERROR("Can't init FLD, err %d\n", rc);
1220 rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1222 CWARN("%s: error initialize target table: rc = %d\n",
1228 static int lmv_cleanup(struct obd_device *obd)
1230 struct lmv_obd *lmv = &obd->u.lmv;
1231 struct lu_tgt_desc *tgt;
1232 struct lu_tgt_desc *tmp;
1236 fld_client_fini(&lmv->lmv_fld);
1237 lmv_foreach_tgt_safe(lmv, tgt, tmp)
1238 lmv_del_target(lmv, tgt);
1239 lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1244 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1246 struct lustre_cfg *lcfg = buf;
1247 struct obd_uuid obd_uuid;
1253 switch (lcfg->lcfg_command) {
1255 /* modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDT0000_UUID
1256 * 2:0 3:1 4:lustre-MDT0000-mdc_UUID */
1257 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1258 GOTO(out, rc = -EINVAL);
1260 obd_str2uuid(&obd_uuid, lustre_cfg_buf(lcfg, 1));
1262 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1263 GOTO(out, rc = -EINVAL);
1264 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1265 GOTO(out, rc = -EINVAL);
1266 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1269 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1270 GOTO(out, rc = -EINVAL);
1276 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1280 if (flags & OBD_STATFS_FOR_MDT0)
1283 if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1284 return lmv->lmv_statfs_start;
1286 /* choose initial MDT for this client */
1288 struct lnet_process_id lnet_id;
1289 if (LNetGetId(i, &lnet_id) == -ENOENT)
1292 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1293 /* We dont need a full 64-bit modulus, just enough
1294 * to distribute the requests across MDTs evenly.
1296 lmv->lmv_statfs_start = (u32)lnet_id.nid %
1302 return lmv->lmv_statfs_start;
1305 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1306 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1308 struct obd_device *obd = class_exp2obd(exp);
1309 struct lmv_obd *lmv = &obd->u.lmv;
1310 struct obd_statfs *temp;
1311 struct lu_tgt_desc *tgt;
1318 OBD_ALLOC(temp, sizeof(*temp));
1322 /* distribute statfs among MDTs */
1323 idx = lmv_select_statfs_mdt(lmv, flags);
1325 for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1326 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1327 tgt = lmv_tgt(lmv, idx);
1328 if (!tgt || !tgt->ltd_exp)
1331 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age, flags);
1333 CERROR("%s: can't stat MDS #%d: rc = %d\n",
1334 tgt->ltd_exp->exp_obd->obd_name, i, rc);
1335 GOTO(out_free_temp, rc);
1338 if (temp->os_state & OS_STATE_SUM ||
1339 flags == OBD_STATFS_FOR_MDT0) {
1340 /* reset to the last aggregated values
1341 * and don't sum with non-aggrated data */
1342 /* If the statfs is from mount, it needs to retrieve
1343 * necessary information from MDT0. i.e. mount does
1344 * not need the merged osfs from all of MDT. Also
1345 * clients can be mounted as long as MDT0 is in
1354 osfs->os_bavail += temp->os_bavail;
1355 osfs->os_blocks += temp->os_blocks;
1356 osfs->os_ffree += temp->os_ffree;
1357 osfs->os_files += temp->os_files;
1358 osfs->os_granted += temp->os_granted;
1364 OBD_FREE(temp, sizeof(*temp));
1368 static int lmv_statfs_update(void *cookie, int rc)
1370 struct obd_info *oinfo = cookie;
1371 struct obd_device *obd = oinfo->oi_obd;
1372 struct lmv_obd *lmv = &obd->u.lmv;
1373 struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1374 struct obd_statfs *osfs = oinfo->oi_osfs;
1377 * NB: don't deactivate TGT upon error, because we may not trigger async
1378 * statfs any longer, then there is no chance to activate TGT.
1381 spin_lock(&lmv->lmv_lock);
1382 tgt->ltd_statfs = *osfs;
1383 tgt->ltd_statfs_age = ktime_get_seconds();
1384 spin_unlock(&lmv->lmv_lock);
1385 lmv->lmv_qos.lq_dirty = 1;
1391 /* update tgt statfs async if it's ld_qos_maxage old */
1392 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1394 struct obd_info oinfo = {
1397 .oi_cb_up = lmv_statfs_update,
1401 if (ktime_get_seconds() - tgt->ltd_statfs_age <
1402 obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1405 rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1410 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1413 struct obd_device *obd = exp->exp_obd;
1414 struct lmv_obd *lmv = &obd->u.lmv;
1415 struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1423 rc = md_get_root(tgt->ltd_exp, fileset, fid);
1427 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1428 u64 obd_md_valid, const char *name, size_t buf_size,
1429 struct ptlrpc_request **req)
1431 struct obd_device *obd = exp->exp_obd;
1432 struct lmv_obd *lmv = &obd->u.lmv;
1433 struct lmv_tgt_desc *tgt;
1438 tgt = lmv_fid2tgt(lmv, fid);
1440 RETURN(PTR_ERR(tgt));
1442 rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1447 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1448 u64 obd_md_valid, const char *name,
1449 const void *value, size_t value_size,
1450 unsigned int xattr_flags, u32 suppgid,
1451 struct ptlrpc_request **req)
1453 struct obd_device *obd = exp->exp_obd;
1454 struct lmv_obd *lmv = &obd->u.lmv;
1455 struct lmv_tgt_desc *tgt;
1460 tgt = lmv_fid2tgt(lmv, fid);
1462 RETURN(PTR_ERR(tgt));
1464 rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1465 value, value_size, xattr_flags, suppgid, req);
1470 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1471 struct ptlrpc_request **request)
1473 struct obd_device *obd = exp->exp_obd;
1474 struct lmv_obd *lmv = &obd->u.lmv;
1475 struct lmv_tgt_desc *tgt;
1480 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1482 RETURN(PTR_ERR(tgt));
1484 if (op_data->op_flags & MF_GET_MDT_IDX) {
1485 op_data->op_mds = tgt->ltd_index;
1489 rc = md_getattr(tgt->ltd_exp, op_data, request);
1494 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1496 struct obd_device *obd = exp->exp_obd;
1497 struct lmv_obd *lmv = &obd->u.lmv;
1498 struct lu_tgt_desc *tgt;
1502 CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1505 * With DNE every object can have two locks in different namespaces:
1506 * lookup lock in space of MDT storing direntry and update/open lock in
1507 * space of MDT storing inode.
1509 lmv_foreach_connected_tgt(lmv, tgt)
1510 md_null_inode(tgt->ltd_exp, fid);
1515 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1516 struct md_open_data *mod, struct ptlrpc_request **request)
1518 struct obd_device *obd = exp->exp_obd;
1519 struct lmv_obd *lmv = &obd->u.lmv;
1520 struct lmv_tgt_desc *tgt;
1525 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1527 RETURN(PTR_ERR(tgt));
1529 CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1530 rc = md_close(tgt->ltd_exp, op_data, mod, request);
1534 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
1536 struct lu_tgt_desc *tgt;
1537 __u64 total_weight = 0;
1538 __u64 cur_weight = 0;
1544 if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1545 RETURN(ERR_PTR(-EAGAIN));
1547 down_write(&lmv->lmv_qos.lq_rw_sem);
1549 if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1550 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1552 rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1554 GOTO(unlock, tgt = ERR_PTR(rc));
1556 lmv_foreach_tgt(lmv, tgt) {
1557 tgt->ltd_qos.ltq_usable = 0;
1558 if (!tgt->ltd_exp || !tgt->ltd_active)
1561 tgt->ltd_qos.ltq_usable = 1;
1562 lu_tgt_qos_weight_calc(tgt);
1563 total_weight += tgt->ltd_qos.ltq_weight;
1566 rand = lu_prandom_u64_max(total_weight);
1568 lmv_foreach_connected_tgt(lmv, tgt) {
1569 if (!tgt->ltd_qos.ltq_usable)
1572 cur_weight += tgt->ltd_qos.ltq_weight;
1573 if (cur_weight < rand)
1576 *mdt = tgt->ltd_index;
1577 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1578 GOTO(unlock, rc = 0);
1581 /* no proper target found */
1582 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1584 up_write(&lmv->lmv_qos.lq_rw_sem);
1589 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1591 struct lu_tgt_desc *tgt;
1597 spin_lock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1598 for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1599 index = (i + lmv->lmv_qos_rr_index) %
1600 lmv->lmv_mdt_descs.ltd_tgts_size;
1601 tgt = lmv_tgt(lmv, index);
1602 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1605 *mdt = tgt->ltd_index;
1606 lmv->lmv_qos_rr_index = (*mdt + 1) %
1607 lmv->lmv_mdt_descs.ltd_tgts_size;
1608 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1612 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
1614 RETURN(ERR_PTR(-ENODEV));
1617 static struct lmv_tgt_desc *
1618 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1619 const char *name, int namelen, struct lu_fid *fid,
1620 __u32 *mds, bool post_migrate)
1622 struct lmv_tgt_desc *tgt;
1623 const struct lmv_oinfo *oinfo;
1625 if (!lmv_dir_striped(lsm) || !namelen) {
1626 tgt = lmv_fid2tgt(lmv, fid);
1630 *mds = tgt->ltd_index;
1634 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1635 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1636 return ERR_PTR(-EBADF);
1637 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1639 oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
1642 return ERR_CAST(oinfo);
1645 *fid = oinfo->lmo_fid;
1646 *mds = oinfo->lmo_mds;
1647 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1649 CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1651 return tgt ? tgt : ERR_PTR(-ENODEV);
1655 * Locate MDT of op_data->op_fid1
1657 * For striped directory, it will locate the stripe by name hash, if hash_type
1658 * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1659 * which is set outside, and if dir is migrating, 'op_data->op_post_migrate'
1660 * indicates whether old or new layout is used to locate.
1662 * For plain direcotry, normally it will locate MDT by FID, but if this
1663 * directory has default LMV, and its hash type is "space", locate MDT with QoS.
1665 * \param[in] lmv LMV device
1666 * \param[in] op_data client MD stack parameters, name, namelen
1669 * retval pointer to the lmv_tgt_desc if succeed.
1670 * ERR_PTR(errno) if failed.
1672 struct lmv_tgt_desc *
1673 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1675 struct lmv_stripe_md *lsm = op_data->op_mea1;
1676 struct lmv_oinfo *oinfo;
1677 struct lmv_tgt_desc *tgt;
1679 if (lmv_dir_foreign(lsm))
1680 return ERR_PTR(-ENODATA);
1682 /* During creating VOLATILE file, it should honor the mdt
1683 * index if the file under striped dir is being restored, see
1685 if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1686 (int)op_data->op_mds != -1) {
1687 tgt = lmv_tgt(lmv, op_data->op_mds);
1689 return ERR_PTR(-ENODEV);
1691 if (lmv_dir_striped(lsm)) {
1694 /* refill the right parent fid */
1695 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1696 oinfo = &lsm->lsm_md_oinfo[i];
1697 if (oinfo->lmo_mds == op_data->op_mds) {
1698 op_data->op_fid1 = oinfo->lmo_fid;
1703 if (i == lsm->lsm_md_stripe_count)
1704 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1706 } else if (lmv_dir_bad_hash(lsm)) {
1707 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1708 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1710 op_data->op_fid1 = oinfo->lmo_fid;
1711 op_data->op_mds = oinfo->lmo_mds;
1712 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1714 tgt = ERR_PTR(-ENODEV);
1715 } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1716 lmv_dir_qos_mkdir(op_data->op_default_mea1) &&
1717 !lmv_dir_striped(lsm)) {
1718 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds);
1719 if (tgt == ERR_PTR(-EAGAIN))
1720 tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1722 * only update statfs when mkdir under dir with "space" hash,
1723 * this means the cached statfs may be stale, and current mkdir
1724 * may not follow QoS accurately, but it's not serious, and it
1725 * avoids periodic statfs when client doesn't mkdir under
1726 * "space" hashed directories.
1728 * TODO: after MDT support QoS object allocation, also update
1729 * statfs for 'lfs mkdir -i -1 ...", currently it's done in user
1733 struct obd_device *obd;
1735 obd = container_of(lmv, struct obd_device, u.lmv);
1736 lmv_statfs_check_update(obd, tgt);
1739 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1740 op_data->op_name, op_data->op_namelen,
1741 &op_data->op_fid1, &op_data->op_mds,
1742 op_data->op_post_migrate);
1748 /* Locate MDT of op_data->op_fid2 for link/rename */
1749 static struct lmv_tgt_desc *
1750 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1752 struct lmv_tgt_desc *tgt;
1755 LASSERT(op_data->op_name);
1756 if (lmv_dir_migrating(op_data->op_mea2)) {
1757 struct lu_fid fid1 = op_data->op_fid1;
1758 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1759 struct ptlrpc_request *request = NULL;
1762 * avoid creating new file under old layout of migrating
1763 * directory, check it here.
1765 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1766 op_data->op_name, op_data->op_namelen,
1767 &op_data->op_fid2, &op_data->op_mds, false);
1771 op_data->op_fid1 = op_data->op_fid2;
1772 op_data->op_mea1 = op_data->op_mea2;
1773 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1774 op_data->op_fid1 = fid1;
1775 op_data->op_mea1 = lsm1;
1777 ptlrpc_req_finished(request);
1778 RETURN(ERR_PTR(-EEXIST));
1782 RETURN(ERR_PTR(rc));
1785 return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1786 op_data->op_namelen, &op_data->op_fid2,
1787 &op_data->op_mds, true);
1790 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1791 const void *data, size_t datalen, umode_t mode, uid_t uid,
1792 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1793 struct ptlrpc_request **request)
1795 struct obd_device *obd = exp->exp_obd;
1796 struct lmv_obd *lmv = &obd->u.lmv;
1797 struct lmv_tgt_desc *tgt;
1802 if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1805 if (lmv_dir_bad_hash(op_data->op_mea1))
1808 if (lmv_dir_migrating(op_data->op_mea1)) {
1810 * if parent is migrating, create() needs to lookup existing
1811 * name, to avoid creating new file under old layout of
1812 * migrating directory, check old layout here.
1814 tgt = lmv_locate_tgt(lmv, op_data);
1816 RETURN(PTR_ERR(tgt));
1818 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1820 ptlrpc_req_finished(*request);
1828 op_data->op_post_migrate = true;
1831 tgt = lmv_locate_tgt(lmv, op_data);
1833 RETURN(PTR_ERR(tgt));
1835 CDEBUG(D_INODE, "CREATE name '%.*s' on "DFID" -> mds #%x\n",
1836 (int)op_data->op_namelen, op_data->op_name,
1837 PFID(&op_data->op_fid1), op_data->op_mds);
1839 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1843 if (exp_connect_flags(exp) & OBD_CONNECT_DIR_STRIPE) {
1844 /* Send the create request to the MDT where the object
1845 * will be located */
1846 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1848 RETURN(PTR_ERR(tgt));
1850 op_data->op_mds = tgt->ltd_index;
1853 CDEBUG(D_INODE, "CREATE obj "DFID" -> mds #%x\n",
1854 PFID(&op_data->op_fid2), op_data->op_mds);
1856 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1857 rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1858 cap_effective, rdev, request);
1860 if (*request == NULL)
1862 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1868 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1869 const union ldlm_policy_data *policy, struct md_op_data *op_data,
1870 struct lustre_handle *lockh, __u64 extra_lock_flags)
1872 struct obd_device *obd = exp->exp_obd;
1873 struct lmv_obd *lmv = &obd->u.lmv;
1874 struct lmv_tgt_desc *tgt;
1879 CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1881 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1883 RETURN(PTR_ERR(tgt));
1885 CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1886 PFID(&op_data->op_fid1), tgt->ltd_index);
1888 rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1895 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1896 struct ptlrpc_request **preq)
1898 struct obd_device *obd = exp->exp_obd;
1899 struct lmv_obd *lmv = &obd->u.lmv;
1900 struct lmv_tgt_desc *tgt;
1901 struct mdt_body *body;
1907 tgt = lmv_locate_tgt(lmv, op_data);
1909 RETURN(PTR_ERR(tgt));
1911 CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1912 (int)op_data->op_namelen, op_data->op_name,
1913 PFID(&op_data->op_fid1), tgt->ltd_index);
1915 rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1916 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1917 ptlrpc_req_finished(*preq);
1925 body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1926 LASSERT(body != NULL);
1928 if (body->mbo_valid & OBD_MD_MDS) {
1929 op_data->op_fid1 = body->mbo_fid1;
1930 op_data->op_valid |= OBD_MD_FLCROSSREF;
1931 op_data->op_namelen = 0;
1932 op_data->op_name = NULL;
1934 ptlrpc_req_finished(*preq);
1943 #define md_op_data_fid(op_data, fl) \
1944 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1945 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1946 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1947 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1950 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1951 struct md_op_data *op_data, __u32 op_tgt,
1952 enum ldlm_mode mode, int bits, int flag)
1954 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1955 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1956 union ldlm_policy_data policy = { { 0 } };
1960 if (!fid_is_sane(fid))
1964 tgt = lmv_fid2tgt(lmv, fid);
1966 RETURN(PTR_ERR(tgt));
1969 if (tgt->ltd_index != op_tgt) {
1970 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1971 policy.l_inodebits.bits = bits;
1972 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1973 mode, LCF_ASYNC, NULL);
1976 "EARLY_CANCEL skip operation target %d on "DFID"\n",
1978 op_data->op_flags |= flag;
1986 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1989 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1990 struct ptlrpc_request **request)
1992 struct obd_device *obd = exp->exp_obd;
1993 struct lmv_obd *lmv = &obd->u.lmv;
1994 struct lmv_tgt_desc *tgt;
1998 LASSERT(op_data->op_namelen != 0);
2000 CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2001 PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2002 op_data->op_name, PFID(&op_data->op_fid1));
2004 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2005 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2006 op_data->op_cap = cfs_curproc_cap_pack();
2008 tgt = lmv_locate_tgt2(lmv, op_data);
2010 RETURN(PTR_ERR(tgt));
2013 * Cancel UPDATE lock on child (fid1).
2015 op_data->op_flags |= MF_MDC_CANCEL_FID2;
2016 rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2017 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2021 rc = md_link(tgt->ltd_exp, op_data, request);
2026 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2027 const char *name, size_t namelen,
2028 struct ptlrpc_request **request)
2030 struct obd_device *obd = exp->exp_obd;
2031 struct lmv_obd *lmv = &obd->u.lmv;
2032 struct lmv_stripe_md *lsm = op_data->op_mea1;
2033 struct lmv_tgt_desc *parent_tgt;
2034 struct lmv_tgt_desc *sp_tgt;
2035 struct lmv_tgt_desc *tp_tgt = NULL;
2036 struct lmv_tgt_desc *child_tgt;
2037 struct lmv_tgt_desc *tgt;
2038 struct lu_fid target_fid;
2043 LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2045 CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2046 PFID(&op_data->op_fid1), (int)namelen, name);
2048 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2049 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2050 op_data->op_cap = cfs_curproc_cap_pack();
2052 parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2053 if (IS_ERR(parent_tgt))
2054 RETURN(PTR_ERR(parent_tgt));
2056 if (lmv_dir_striped(lsm)) {
2057 __u32 hash_type = lsm->lsm_md_hash_type;
2058 __u32 stripe_count = lsm->lsm_md_stripe_count;
2061 * old stripes are appended after new stripes for migrating
2064 if (lmv_dir_migrating(lsm)) {
2065 hash_type = lsm->lsm_md_migrate_hash;
2066 stripe_count -= lsm->lsm_md_migrate_offset;
2069 rc = lmv_name_to_stripe_index(hash_type, stripe_count, name,
2074 if (lmv_dir_migrating(lsm))
2075 rc += lsm->lsm_md_migrate_offset;
2077 /* save it in fid4 temporarily for early cancel */
2078 op_data->op_fid4 = lsm->lsm_md_oinfo[rc].lmo_fid;
2079 sp_tgt = lmv_tgt(lmv, lsm->lsm_md_oinfo[rc].lmo_mds);
2084 * if parent is being migrated too, fill op_fid2 with target
2085 * stripe fid, otherwise the target stripe is not created yet.
2087 if (lmv_dir_migrating(lsm)) {
2088 hash_type = lsm->lsm_md_hash_type &
2089 ~LMV_HASH_FLAG_MIGRATION;
2090 stripe_count = lsm->lsm_md_migrate_offset;
2092 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
2097 op_data->op_fid2 = lsm->lsm_md_oinfo[rc].lmo_fid;
2098 tp_tgt = lmv_tgt(lmv, lsm->lsm_md_oinfo[rc].lmo_mds);
2103 sp_tgt = parent_tgt;
2106 child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2107 if (IS_ERR(child_tgt))
2108 RETURN(PTR_ERR(child_tgt));
2110 if (!S_ISDIR(op_data->op_mode) && tp_tgt)
2111 rc = __lmv_fid_alloc(lmv, &target_fid, tp_tgt->ltd_index);
2113 rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2118 * for directory, send migrate request to the MDT where the object will
2119 * be migrated to, because we can't create a striped directory remotely.
2121 * otherwise, send to the MDT where source is located because regular
2122 * file may open lease.
2124 * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2125 * backward compatibility.
2127 if (S_ISDIR(op_data->op_mode) &&
2128 (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2129 tgt = lmv_fid2tgt(lmv, &target_fid);
2131 RETURN(PTR_ERR(tgt));
2136 /* cancel UPDATE lock of parent master object */
2137 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2138 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2142 /* cancel UPDATE lock of source parent */
2143 if (sp_tgt != parent_tgt) {
2145 * migrate RPC packs master object FID, because we can only pack
2146 * two FIDs in reint RPC, but MDS needs to know both source
2147 * parent and target parent, and it will obtain them from master
2148 * FID and LMV, the other FID in RPC is kept for target.
2150 * since this FID is not passed to MDC, cancel it anyway.
2152 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2153 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2157 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2159 op_data->op_fid4 = target_fid;
2161 /* cancel UPDATE locks of target parent */
2162 rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2163 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2167 /* cancel LOOKUP lock of source if source is remote object */
2168 if (child_tgt != sp_tgt) {
2169 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2170 LCK_EX, MDS_INODELOCK_LOOKUP,
2171 MF_MDC_CANCEL_FID3);
2176 /* cancel ELC locks of source */
2177 rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2178 MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2182 rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2187 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2188 const char *old, size_t oldlen,
2189 const char *new, size_t newlen,
2190 struct ptlrpc_request **request)
2192 struct obd_device *obd = exp->exp_obd;
2193 struct lmv_obd *lmv = &obd->u.lmv;
2194 struct lmv_tgt_desc *sp_tgt;
2195 struct lmv_tgt_desc *tp_tgt = NULL;
2196 struct lmv_tgt_desc *src_tgt = NULL;
2197 struct lmv_tgt_desc *tgt;
2198 struct mdt_body *body;
2203 LASSERT(oldlen != 0);
2205 if (op_data->op_cli_flags & CLI_MIGRATE) {
2206 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2210 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2211 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2212 op_data->op_cap = cfs_curproc_cap_pack();
2214 op_data->op_name = new;
2215 op_data->op_namelen = newlen;
2217 tp_tgt = lmv_locate_tgt2(lmv, op_data);
2219 RETURN(PTR_ERR(tp_tgt));
2221 /* Since the target child might be destroyed, and it might become
2222 * orphan, and we can only check orphan on the local MDT right now, so
2223 * we send rename request to the MDT where target child is located. If
2224 * target child does not exist, then it will send the request to the
2226 if (fid_is_sane(&op_data->op_fid4)) {
2227 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2229 RETURN(PTR_ERR(tgt));
2234 op_data->op_flags |= MF_MDC_CANCEL_FID4;
2236 /* cancel UPDATE locks of target parent */
2237 rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2238 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2242 if (fid_is_sane(&op_data->op_fid4)) {
2243 /* cancel LOOKUP lock of target on target parent */
2244 if (tgt != tp_tgt) {
2245 rc = lmv_early_cancel(exp, tp_tgt, op_data,
2246 tgt->ltd_index, LCK_EX,
2247 MDS_INODELOCK_LOOKUP,
2248 MF_MDC_CANCEL_FID4);
2254 if (fid_is_sane(&op_data->op_fid3)) {
2255 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2256 if (IS_ERR(src_tgt))
2257 RETURN(PTR_ERR(src_tgt));
2259 /* cancel ELC locks of source */
2260 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2261 LCK_EX, MDS_INODELOCK_ELC,
2262 MF_MDC_CANCEL_FID3);
2267 op_data->op_name = old;
2268 op_data->op_namelen = oldlen;
2270 sp_tgt = lmv_locate_tgt(lmv, op_data);
2272 RETURN(PTR_ERR(sp_tgt));
2274 /* cancel UPDATE locks of source parent */
2275 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2276 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2280 if (fid_is_sane(&op_data->op_fid3)) {
2281 /* cancel LOOKUP lock of source on source parent */
2282 if (src_tgt != sp_tgt) {
2283 rc = lmv_early_cancel(exp, sp_tgt, op_data,
2284 tgt->ltd_index, LCK_EX,
2285 MDS_INODELOCK_LOOKUP,
2286 MF_MDC_CANCEL_FID3);
2293 CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2294 PFID(&op_data->op_fid1), (int)oldlen, old,
2295 PFID(&op_data->op_fid2), (int)newlen, new);
2297 rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2299 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2300 ptlrpc_req_finished(*request);
2305 if (rc && rc != -EXDEV)
2308 body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2312 /* Not cross-ref case, just get out of here. */
2313 if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2316 op_data->op_fid4 = body->mbo_fid1;
2318 ptlrpc_req_finished(*request);
2321 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2323 RETURN(PTR_ERR(tgt));
2325 if (fid_is_sane(&op_data->op_fid4)) {
2326 /* cancel LOOKUP lock of target on target parent */
2327 if (tgt != tp_tgt) {
2328 rc = lmv_early_cancel(exp, tp_tgt, op_data,
2329 tgt->ltd_index, LCK_EX,
2330 MDS_INODELOCK_LOOKUP,
2331 MF_MDC_CANCEL_FID4);
2340 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2341 void *ea, size_t ealen, struct ptlrpc_request **request)
2343 struct obd_device *obd = exp->exp_obd;
2344 struct lmv_obd *lmv = &obd->u.lmv;
2345 struct lmv_tgt_desc *tgt;
2350 CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2351 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2352 op_data->op_xvalid);
2354 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2355 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2357 RETURN(PTR_ERR(tgt));
2359 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2364 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2365 struct ptlrpc_request **request)
2367 struct obd_device *obd = exp->exp_obd;
2368 struct lmv_obd *lmv = &obd->u.lmv;
2369 struct lmv_tgt_desc *tgt;
2374 tgt = lmv_fid2tgt(lmv, fid);
2376 RETURN(PTR_ERR(tgt));
2378 rc = md_fsync(tgt->ltd_exp, fid, request);
2382 struct stripe_dirent {
2383 struct page *sd_page;
2384 struct lu_dirpage *sd_dp;
2385 struct lu_dirent *sd_ent;
2389 struct lmv_dir_ctxt {
2390 struct lmv_obd *ldc_lmv;
2391 struct md_op_data *ldc_op_data;
2392 struct md_callback *ldc_cb_op;
2395 struct stripe_dirent ldc_stripes[0];
2398 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2400 if (stripe->sd_page) {
2401 kunmap(stripe->sd_page);
2402 put_page(stripe->sd_page);
2403 stripe->sd_page = NULL;
2404 stripe->sd_ent = NULL;
2408 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2412 for (i = 0; i < ctxt->ldc_count; i++)
2413 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2416 /* if @ent is dummy, or . .., get next */
2417 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2418 struct lu_dirent *ent,
2421 for (; ent; ent = lu_dirent_next(ent)) {
2422 /* Skip dummy entry */
2423 if (le16_to_cpu(ent->lde_namelen) == 0)
2426 /* skip . and .. for other stripes */
2428 (strncmp(ent->lde_name, ".",
2429 le16_to_cpu(ent->lde_namelen)) == 0 ||
2430 strncmp(ent->lde_name, "..",
2431 le16_to_cpu(ent->lde_namelen)) == 0))
2434 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2441 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2442 struct stripe_dirent *stripe,
2445 struct md_op_data *op_data = ctxt->ldc_op_data;
2446 struct lmv_oinfo *oinfo;
2447 struct lu_fid fid = op_data->op_fid1;
2448 struct inode *inode = op_data->op_data;
2449 struct lmv_tgt_desc *tgt;
2450 struct lu_dirent *ent = stripe->sd_ent;
2451 __u64 hash = ctxt->ldc_hash;
2456 LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2460 if (stripe->sd_page) {
2461 __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2463 /* @hash should be the last dirent hash */
2464 LASSERTF(hash <= end,
2465 "ctxt@%p stripe@%p hash %llx end %llx\n",
2466 ctxt, stripe, hash, end);
2467 /* unload last page */
2468 stripe_dirent_unload(stripe);
2470 if (end == MDS_DIR_END_OFF) {
2471 stripe->sd_eof = true;
2477 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2478 if (!oinfo->lmo_root) {
2483 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2489 /* op_data is shared by stripes, reset after use */
2490 op_data->op_fid1 = oinfo->lmo_fid;
2491 op_data->op_fid2 = oinfo->lmo_fid;
2492 op_data->op_data = oinfo->lmo_root;
2494 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2497 op_data->op_fid1 = fid;
2498 op_data->op_fid2 = fid;
2499 op_data->op_data = inode;
2504 stripe->sd_dp = page_address(stripe->sd_page);
2505 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2507 /* in case a page filled with ., .. and dummy, read next */
2510 stripe->sd_ent = ent;
2513 /* treat error as eof, so dir can be partially accessed */
2514 stripe->sd_eof = true;
2515 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2516 "directory is partially accessed!\n",
2517 PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2524 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2526 struct obd_device *obd = exp->exp_obd;
2527 struct lmv_obd *lmv = &obd->u.lmv;
2528 struct lmv_tgt_desc *tgt;
2533 rc = lmv_check_connect(obd);
2537 tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2539 RETURN(PTR_ERR(tgt));
2541 data->op_flags |= MF_MDC_CANCEL_FID1;
2542 rc = md_file_resync(tgt->ltd_exp, data);
2547 * Get dirent with the closest hash for striped directory
2549 * This function will search the dir entry, whose hash value is the
2550 * closest(>=) to hash from all of sub-stripes, and it is only being called
2551 * for striped directory.
2553 * \param[in] ctxt dir read context
2555 * \retval dirent get the entry successfully
2556 * NULL does not get the entry, normally it means
2557 * it reaches the end of the directory, while read
2558 * stripe dirent error is ignored to allow partial
2561 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2563 struct stripe_dirent *stripe;
2564 struct lu_dirent *ent = NULL;
2568 /* TODO: optimize with k-way merge sort */
2569 for (i = 0; i < ctxt->ldc_count; i++) {
2570 stripe = &ctxt->ldc_stripes[i];
2574 if (!stripe->sd_ent) {
2575 stripe_dirent_load(ctxt, stripe, i);
2576 if (!stripe->sd_ent) {
2577 LASSERT(stripe->sd_eof);
2583 le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2584 le64_to_cpu(stripe->sd_ent->lde_hash)) {
2586 if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2593 stripe = &ctxt->ldc_stripes[min];
2594 ent = stripe->sd_ent;
2595 /* pop found dirent */
2596 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2604 * Build dir entry page for striped directory
2606 * This function gets one entry by @offset from a striped directory. It will
2607 * read entries from all of stripes, and choose one closest to the required
2608 * offset(&offset). A few notes
2609 * 1. skip . and .. for non-zero stripes, because there can only have one .
2610 * and .. in a directory.
2611 * 2. op_data will be shared by all of stripes, instead of allocating new
2612 * one, so need to restore before reusing.
2614 * \param[in] exp obd export refer to LMV
2615 * \param[in] op_data hold those MD parameters of read_entry
2616 * \param[in] cb_op ldlm callback being used in enqueue in mdc_read_entry
2617 * \param[in] offset starting hash offset
2618 * \param[out] ppage the page holding the entry. Note: because the entry
2619 * will be accessed in upper layer, so we need hold the
2620 * page until the usages of entry is finished, see
2621 * ll_dir_entry_next.
2623 * retval =0 if get entry successfully
2624 * <0 cannot get entry
2626 static int lmv_striped_read_page(struct obd_export *exp,
2627 struct md_op_data *op_data,
2628 struct md_callback *cb_op,
2629 __u64 offset, struct page **ppage)
2631 struct page *page = NULL;
2632 struct lu_dirpage *dp;
2634 struct lu_dirent *ent;
2635 struct lu_dirent *last_ent;
2637 struct lmv_dir_ctxt *ctxt;
2638 struct lu_dirent *next = NULL;
2644 /* Allocate a page and read entries from all of stripes and fill
2645 * the page by hash order */
2646 page = alloc_page(GFP_KERNEL);
2650 /* Initialize the entry page */
2652 memset(dp, 0, sizeof(*dp));
2653 dp->ldp_hash_start = cpu_to_le64(offset);
2656 left_bytes = PAGE_SIZE - sizeof(*dp);
2660 /* initalize dir read context */
2661 stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2662 OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2664 GOTO(free_page, rc = -ENOMEM);
2665 ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2666 ctxt->ldc_op_data = op_data;
2667 ctxt->ldc_cb_op = cb_op;
2668 ctxt->ldc_hash = offset;
2669 ctxt->ldc_count = stripe_count;
2672 next = lmv_dirent_next(ctxt);
2674 /* end of directory */
2676 ctxt->ldc_hash = MDS_DIR_END_OFF;
2679 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2681 ent_size = le16_to_cpu(next->lde_reclen);
2683 /* the last entry lde_reclen is 0, but it might not be the last
2684 * one of this temporay dir page */
2686 ent_size = lu_dirent_calc_size(
2687 le16_to_cpu(next->lde_namelen),
2688 le32_to_cpu(next->lde_attrs));
2690 if (ent_size > left_bytes)
2693 memcpy(ent, next, ent_size);
2695 /* Replace . with master FID and Replace .. with the parent FID
2696 * of master object */
2697 if (strncmp(ent->lde_name, ".",
2698 le16_to_cpu(ent->lde_namelen)) == 0 &&
2699 le16_to_cpu(ent->lde_namelen) == 1)
2700 fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2701 else if (strncmp(ent->lde_name, "..",
2702 le16_to_cpu(ent->lde_namelen)) == 0 &&
2703 le16_to_cpu(ent->lde_namelen) == 2)
2704 fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2706 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2707 le16_to_cpu(ent->lde_namelen), ent->lde_name,
2708 le64_to_cpu(ent->lde_hash));
2710 left_bytes -= ent_size;
2711 ent->lde_reclen = cpu_to_le16(ent_size);
2713 ent = (void *)ent + ent_size;
2716 last_ent->lde_reclen = 0;
2719 dp->ldp_flags |= LDF_EMPTY;
2720 else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2721 dp->ldp_flags |= LDF_COLLIDE;
2722 dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2723 dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2725 put_lmv_dir_ctxt(ctxt);
2726 OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2739 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2740 struct md_callback *cb_op, __u64 offset,
2741 struct page **ppage)
2743 struct obd_device *obd = exp->exp_obd;
2744 struct lmv_obd *lmv = &obd->u.lmv;
2745 struct lmv_tgt_desc *tgt;
2750 if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2753 if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2754 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2758 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2760 RETURN(PTR_ERR(tgt));
2762 rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2768 * Unlink a file/directory
2770 * Unlink a file or directory under the parent dir. The unlink request
2771 * usually will be sent to the MDT where the child is located, but if
2772 * the client does not have the child FID then request will be sent to the
2773 * MDT where the parent is located.
2775 * If the parent is a striped directory then it also needs to locate which
2776 * stripe the name of the child is located, and replace the parent FID
2777 * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2778 * it will walk through all of sub-stripes until the child is being
2781 * \param[in] exp export refer to LMV
2782 * \param[in] op_data different parameters transferred beween client
2783 * MD stacks, name, namelen, FIDs etc.
2784 * op_fid1 is the parent FID, op_fid2 is the child
2786 * \param[out] request point to the request of unlink.
2788 * retval 0 if succeed
2789 * negative errno if failed.
2791 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2792 struct ptlrpc_request **request)
2794 struct obd_device *obd = exp->exp_obd;
2795 struct lmv_obd *lmv = &obd->u.lmv;
2796 struct lmv_tgt_desc *tgt;
2797 struct lmv_tgt_desc *parent_tgt;
2798 struct mdt_body *body;
2803 op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2804 op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2805 op_data->op_cap = cfs_curproc_cap_pack();
2808 parent_tgt = lmv_locate_tgt(lmv, op_data);
2809 if (IS_ERR(parent_tgt))
2810 RETURN(PTR_ERR(parent_tgt));
2812 if (likely(!fid_is_zero(&op_data->op_fid2))) {
2813 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2815 RETURN(PTR_ERR(tgt));
2821 * If child's fid is given, cancel unused locks for it if it is from
2822 * another export than parent.
2824 * LOOKUP lock for child (fid3) should also be cancelled on parent
2825 * tgt_tgt in mdc_unlink().
2827 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2829 if (parent_tgt != tgt)
2830 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2831 LCK_EX, MDS_INODELOCK_LOOKUP,
2832 MF_MDC_CANCEL_FID3);
2834 rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2835 MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2839 CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2840 PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2843 rc = md_unlink(tgt->ltd_exp, op_data, request);
2844 if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2845 ptlrpc_req_finished(*request);
2853 body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2857 /* Not cross-ref case, just get out of here. */
2858 if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2861 /* This is a remote object, try remote MDT. */
2862 op_data->op_fid2 = body->mbo_fid1;
2863 ptlrpc_req_finished(*request);
2866 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2868 RETURN(PTR_ERR(tgt));
2873 static int lmv_precleanup(struct obd_device *obd)
2876 libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2877 fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2878 lprocfs_obd_cleanup(obd);
2879 lprocfs_free_md_stats(obd);
2884 * Get by key a value associated with a LMV device.
2886 * Dispatch request to lower-layer devices as needed.
2888 * \param[in] env execution environment for this thread
2889 * \param[in] exp export for the LMV device
2890 * \param[in] keylen length of key identifier
2891 * \param[in] key identifier of key to get value for
2892 * \param[in] vallen size of \a val
2893 * \param[out] val pointer to storage location for value
2894 * \param[in] lsm optional striping metadata of object
2896 * \retval 0 on success
2897 * \retval negative negated errno on failure
2899 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2900 __u32 keylen, void *key, __u32 *vallen, void *val)
2902 struct obd_device *obd;
2903 struct lmv_obd *lmv;
2904 struct lu_tgt_desc *tgt;
2909 obd = class_exp2obd(exp);
2911 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2912 exp->exp_handle.h_cookie);
2917 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2918 LASSERT(*vallen == sizeof(__u32));
2919 lmv_foreach_connected_tgt(lmv, tgt) {
2920 if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2925 } else if (KEY_IS(KEY_MAX_EASIZE) ||
2926 KEY_IS(KEY_DEFAULT_EASIZE) ||
2927 KEY_IS(KEY_CONN_DATA)) {
2929 * Forwarding this request to first MDS, it should know LOV
2932 tgt = lmv_tgt(lmv, 0);
2936 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
2937 if (!rc && KEY_IS(KEY_CONN_DATA))
2938 exp->exp_connect_data = *(struct obd_connect_data *)val;
2940 } else if (KEY_IS(KEY_TGT_COUNT)) {
2941 *((int *)val) = lmv->lmv_mdt_descs.ltd_tgts_size;
2945 CDEBUG(D_IOCTL, "Invalid key\n");
2949 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
2950 int *__rcs, struct ptlrpc_request_set *_set)
2952 struct obd_device *obddev = class_exp2obd(exp);
2953 struct ptlrpc_request_set *set = _set;
2954 struct lmv_obd *lmv = &obddev->u.lmv;
2955 int tgt_count = lmv->lmv_mdt_count;
2956 struct lu_tgt_desc *tgt;
2957 struct fid_array *fat, **fas = NULL;
2958 int i, rc, **rcs = NULL;
2961 set = ptlrpc_prep_set();
2966 /* split FIDs by targets */
2967 OBD_ALLOC(fas, sizeof(fas) * tgt_count);
2969 GOTO(out, rc = -ENOMEM);
2970 OBD_ALLOC(rcs, sizeof(int *) * tgt_count);
2972 GOTO(out_fas, rc = -ENOMEM);
2974 for (i = 0; i < fa->fa_nr; i++) {
2977 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
2979 CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
2980 PFID(&fa->fa_fids[i]), rc);
2983 LASSERT(idx < tgt_count);
2985 OBD_ALLOC(fas[idx], offsetof(struct fid_array,
2986 fa_fids[fa->fa_nr]));
2988 GOTO(out, rc = -ENOMEM);
2990 OBD_ALLOC(rcs[idx], sizeof(int) * fa->fa_nr);
2992 GOTO(out, rc = -ENOMEM);
2995 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
2998 lmv_foreach_connected_tgt(lmv, tgt) {
2999 fat = fas[tgt->ltd_index];
3000 if (!fat || fat->fa_nr == 0)
3002 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3005 rc = ptlrpc_set_wait(NULL, set);
3008 for (i = 0; i < tgt_count; i++) {
3010 if (!fat || fat->fa_nr == 0)
3012 /* copy FIDs back */
3013 memcpy(fa->fa_fids + j, fat->fa_fids,
3014 fat->fa_nr * sizeof(struct lu_fid));
3016 memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3021 ptlrpc_set_destroy(set);
3024 for (i = 0; i < tgt_count; i++) {
3026 OBD_FREE(fas[i], offsetof(struct fid_array,
3027 fa_fids[fa->fa_nr]));
3029 OBD_FREE(rcs[i], sizeof(int) * fa->fa_nr);
3032 OBD_FREE(rcs, sizeof(int *) * tgt_count);
3035 OBD_FREE(fas, sizeof(fas) * tgt_count);
3041 * Asynchronously set by key a value associated with a LMV device.
3043 * Dispatch request to lower-layer devices as needed.
3045 * \param[in] env execution environment for this thread
3046 * \param[in] exp export for the LMV device
3047 * \param[in] keylen length of key identifier
3048 * \param[in] key identifier of key to store value for
3049 * \param[in] vallen size of value to store
3050 * \param[in] val pointer to data to be stored
3051 * \param[in] set optional list of related ptlrpc requests
3053 * \retval 0 on success
3054 * \retval negative negated errno on failure
3056 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3057 __u32 keylen, void *key, __u32 vallen, void *val,
3058 struct ptlrpc_request_set *set)
3060 struct lmv_tgt_desc *tgt = NULL;
3061 struct obd_device *obd;
3062 struct lmv_obd *lmv;
3066 obd = class_exp2obd(exp);
3068 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3069 exp->exp_handle.h_cookie);
3074 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3075 KEY_IS(KEY_DEFAULT_EASIZE)) {
3078 lmv_foreach_connected_tgt(lmv, tgt) {
3079 err = obd_set_info_async(env, tgt->ltd_exp,
3080 keylen, key, vallen, val, set);
3091 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3092 const struct lmv_mds_md_v1 *lmm1)
3094 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3101 lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3102 lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3103 lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3104 if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3105 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3107 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3108 lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3109 lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3110 lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3111 cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3112 sizeof(lsm->lsm_md_pool_name));
3114 if (cplen >= sizeof(lsm->lsm_md_pool_name))
3117 CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %#x "
3118 "layout_version %d\n", lsm->lsm_md_stripe_count,
3119 lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
3120 lsm->lsm_md_layout_version);
3122 stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3123 for (i = 0; i < stripe_count; i++) {
3124 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3125 &lmm1->lmv_stripe_fids[i]);
3127 * set default value -1, so lmv_locate_tgt() knows this stripe
3128 * target is not initialized.
3130 lsm->lsm_md_oinfo[i].lmo_mds = (u32)-1;
3131 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3134 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3135 &lsm->lsm_md_oinfo[i].lmo_mds);
3142 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3143 PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3149 static inline int lmv_unpack_user_md(struct obd_export *exp,
3150 struct lmv_stripe_md *lsm,
3151 const struct lmv_user_md *lmu)
3153 lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3154 lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3155 lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3156 lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3161 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3162 const union lmv_mds_md *lmm, size_t lmm_size)
3164 struct lmv_stripe_md *lsm;
3167 bool allocated = false;
3170 LASSERT(lsmp != NULL);
3174 if (lsm != NULL && lmm == NULL) {
3176 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3178 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3181 lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3183 OBD_FREE_LARGE(lfm, lfm_size);
3187 if (lmv_dir_striped(lsm)) {
3188 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3189 if (lsm->lsm_md_oinfo[i].lmo_root)
3190 iput(lsm->lsm_md_oinfo[i].lmo_root);
3192 lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3194 lsm_size = lmv_stripe_md_size(0);
3196 OBD_FREE(lsm, lsm_size);
3201 /* foreign lmv case */
3202 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3203 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3206 OBD_ALLOC_LARGE(lfm, lmm_size);
3209 *lsmp = (struct lmv_stripe_md *)lfm;
3211 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3212 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3213 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3214 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3215 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3220 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3224 if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3225 le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3226 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3227 exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3232 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3233 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3236 * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3237 * stripecount should be 0 then.
3239 lsm_size = lmv_stripe_md_size(0);
3242 OBD_ALLOC(lsm, lsm_size);
3249 switch (le32_to_cpu(lmm->lmv_magic)) {
3251 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3253 case LMV_USER_MAGIC:
3254 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3257 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3258 le32_to_cpu(lmm->lmv_magic));
3263 if (rc != 0 && allocated) {
3264 OBD_FREE(lsm, lsm_size);
3271 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3273 lmv_unpackmd(NULL, &lsm, NULL, 0);
3275 EXPORT_SYMBOL(lmv_free_memmd);
3277 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3278 union ldlm_policy_data *policy,
3279 enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3282 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3283 struct lu_tgt_desc *tgt;
3289 LASSERT(fid != NULL);
3291 lmv_foreach_connected_tgt(lmv, tgt) {
3292 if (!tgt->ltd_active)
3295 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3303 static int lmv_set_lock_data(struct obd_export *exp,
3304 const struct lustre_handle *lockh,
3305 void *data, __u64 *bits)
3307 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3308 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3313 if (tgt == NULL || tgt->ltd_exp == NULL)
3315 rc = md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3319 enum ldlm_mode lmv_lock_match(struct obd_export *exp, __u64 flags,
3320 const struct lu_fid *fid, enum ldlm_type type,
3321 union ldlm_policy_data *policy,
3322 enum ldlm_mode mode, struct lustre_handle *lockh)
3324 struct obd_device *obd = exp->exp_obd;
3325 struct lmv_obd *lmv = &obd->u.lmv;
3327 struct lu_tgt_desc *tgt;
3333 CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3336 * With DNE every object can have two locks in different namespaces:
3337 * lookup lock in space of MDT storing direntry and update/open lock in
3338 * space of MDT storing inode. Try the MDT that the FID maps to first,
3339 * since this can be easily found, and only try others if that fails.
3341 for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3342 i < lmv->lmv_mdt_descs.ltd_tgts_size;
3343 i++, index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3345 CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3346 obd->obd_name, PFID(fid), index);
3350 tgt = lmv_tgt(lmv, index);
3351 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3354 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3363 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
3364 struct obd_export *dt_exp, struct obd_export *md_exp,
3365 struct lustre_md *md)
3367 struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3368 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3370 if (!tgt || !tgt->ltd_exp)
3373 return md_get_lustre_md(tgt->ltd_exp, req, dt_exp, md_exp, md);
3376 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3378 struct obd_device *obd = exp->exp_obd;
3379 struct lmv_obd *lmv = &obd->u.lmv;
3380 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3384 if (md->default_lmv) {
3385 lmv_free_memmd(md->default_lmv);
3386 md->default_lmv = NULL;
3388 if (md->lmv != NULL) {
3389 lmv_free_memmd(md->lmv);
3392 if (!tgt || !tgt->ltd_exp)
3394 RETURN(md_free_lustre_md(tgt->ltd_exp, md));
3397 int lmv_set_open_replay_data(struct obd_export *exp,
3398 struct obd_client_handle *och,
3399 struct lookup_intent *it)
3401 struct obd_device *obd = exp->exp_obd;
3402 struct lmv_obd *lmv = &obd->u.lmv;
3403 struct lmv_tgt_desc *tgt;
3407 tgt = lmv_fid2tgt(lmv, &och->och_fid);
3409 RETURN(PTR_ERR(tgt));
3411 RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3414 int lmv_clear_open_replay_data(struct obd_export *exp,
3415 struct obd_client_handle *och)
3417 struct obd_device *obd = exp->exp_obd;
3418 struct lmv_obd *lmv = &obd->u.lmv;
3419 struct lmv_tgt_desc *tgt;
3423 tgt = lmv_fid2tgt(lmv, &och->och_fid);
3425 RETURN(PTR_ERR(tgt));
3427 RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3430 int lmv_intent_getattr_async(struct obd_export *exp,
3431 struct md_enqueue_info *minfo)
3433 struct md_op_data *op_data = &minfo->mi_data;
3434 struct obd_device *obd = exp->exp_obd;
3435 struct lmv_obd *lmv = &obd->u.lmv;
3436 struct lmv_tgt_desc *ptgt;
3437 struct lmv_tgt_desc *ctgt;
3442 if (!fid_is_sane(&op_data->op_fid2))
3445 ptgt = lmv_locate_tgt(lmv, op_data);
3447 RETURN(PTR_ERR(ptgt));
3449 ctgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
3451 RETURN(PTR_ERR(ctgt));
3454 * remote object needs two RPCs to lookup and getattr, considering the
3455 * complexity don't support statahead for now.
3460 rc = md_intent_getattr_async(ptgt->ltd_exp, minfo);
3465 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3466 struct lu_fid *fid, __u64 *bits)
3468 struct obd_device *obd = exp->exp_obd;
3469 struct lmv_obd *lmv = &obd->u.lmv;
3470 struct lmv_tgt_desc *tgt;
3475 tgt = lmv_fid2tgt(lmv, fid);
3477 RETURN(PTR_ERR(tgt));
3479 rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3483 int lmv_get_fid_from_lsm(struct obd_export *exp,
3484 const struct lmv_stripe_md *lsm,
3485 const char *name, int namelen, struct lu_fid *fid)
3487 const struct lmv_oinfo *oinfo;
3489 LASSERT(lmv_dir_striped(lsm));
3491 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3493 return PTR_ERR(oinfo);
3495 *fid = oinfo->lmo_fid;
3501 * For lmv, only need to send request to master MDT, and the master MDT will
3502 * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3503 * we directly fetch data from the slave MDTs.
3505 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3506 struct obd_quotactl *oqctl)
3508 struct obd_device *obd = class_exp2obd(exp);
3509 struct lmv_obd *lmv = &obd->u.lmv;
3510 struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3511 __u64 curspace, curinodes;
3516 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3517 CERROR("master lmv inactive\n");
3521 if (oqctl->qc_cmd != Q_GETOQUOTA) {
3522 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3526 curspace = curinodes = 0;
3527 lmv_foreach_connected_tgt(lmv, tgt) {
3530 if (!tgt->ltd_active)
3533 err = obd_quotactl(tgt->ltd_exp, oqctl);
3535 CERROR("getquota on mdt %d failed. %d\n",
3536 tgt->ltd_index, err);
3540 curspace += oqctl->qc_dqblk.dqb_curspace;
3541 curinodes += oqctl->qc_dqblk.dqb_curinodes;
3544 oqctl->qc_dqblk.dqb_curspace = curspace;
3545 oqctl->qc_dqblk.dqb_curinodes = curinodes;
3550 static int lmv_merge_attr(struct obd_export *exp,
3551 const struct lmv_stripe_md *lsm,
3552 struct cl_attr *attr,
3553 ldlm_blocking_callback cb_blocking)
3558 if (!lmv_dir_striped(lsm))
3561 rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3565 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3566 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3572 "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3573 PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3574 i_size_read(inode), (unsigned long long)inode->i_blocks,
3575 inode->i_nlink, (s64)inode->i_atime.tv_sec,
3576 (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3578 /* for slave stripe, it needs to subtract nlink for . and .. */
3580 attr->cat_nlink += inode->i_nlink - 2;
3582 attr->cat_nlink = inode->i_nlink;
3584 attr->cat_size += i_size_read(inode);
3585 attr->cat_blocks += inode->i_blocks;
3587 if (attr->cat_atime < inode->i_atime.tv_sec)
3588 attr->cat_atime = inode->i_atime.tv_sec;
3590 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3591 attr->cat_ctime = inode->i_ctime.tv_sec;
3593 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3594 attr->cat_mtime = inode->i_mtime.tv_sec;
3599 struct obd_ops lmv_obd_ops = {
3600 .o_owner = THIS_MODULE,
3601 .o_setup = lmv_setup,
3602 .o_cleanup = lmv_cleanup,
3603 .o_precleanup = lmv_precleanup,
3604 .o_process_config = lmv_process_config,
3605 .o_connect = lmv_connect,
3606 .o_disconnect = lmv_disconnect,
3607 .o_statfs = lmv_statfs,
3608 .o_get_info = lmv_get_info,
3609 .o_set_info_async = lmv_set_info_async,
3610 .o_notify = lmv_notify,
3611 .o_get_uuid = lmv_get_uuid,
3612 .o_fid_alloc = lmv_fid_alloc,
3613 .o_iocontrol = lmv_iocontrol,
3614 .o_quotactl = lmv_quotactl
3617 struct md_ops lmv_md_ops = {
3618 .m_get_root = lmv_get_root,
3619 .m_null_inode = lmv_null_inode,
3620 .m_close = lmv_close,
3621 .m_create = lmv_create,
3622 .m_enqueue = lmv_enqueue,
3623 .m_getattr = lmv_getattr,
3624 .m_getxattr = lmv_getxattr,
3625 .m_getattr_name = lmv_getattr_name,
3626 .m_intent_lock = lmv_intent_lock,
3628 .m_rename = lmv_rename,
3629 .m_setattr = lmv_setattr,
3630 .m_setxattr = lmv_setxattr,
3631 .m_fsync = lmv_fsync,
3632 .m_file_resync = lmv_file_resync,
3633 .m_read_page = lmv_read_page,
3634 .m_unlink = lmv_unlink,
3635 .m_init_ea_size = lmv_init_ea_size,
3636 .m_cancel_unused = lmv_cancel_unused,
3637 .m_set_lock_data = lmv_set_lock_data,
3638 .m_lock_match = lmv_lock_match,
3639 .m_get_lustre_md = lmv_get_lustre_md,
3640 .m_free_lustre_md = lmv_free_lustre_md,
3641 .m_merge_attr = lmv_merge_attr,
3642 .m_set_open_replay_data = lmv_set_open_replay_data,
3643 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3644 .m_intent_getattr_async = lmv_intent_getattr_async,
3645 .m_revalidate_lock = lmv_revalidate_lock,
3646 .m_get_fid_from_lsm = lmv_get_fid_from_lsm,
3647 .m_unpackmd = lmv_unpackmd,
3648 .m_rmfid = lmv_rmfid,
3651 static int __init lmv_init(void)
3653 return class_register_type(&lmv_obd_ops, &lmv_md_ops, true, NULL,
3654 LUSTRE_LMV_NAME, NULL);
3657 static void __exit lmv_exit(void)
3659 class_unregister_type(LUSTRE_LMV_NAME);
3662 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3663 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3664 MODULE_VERSION(LUSTRE_VERSION_STRING);
3665 MODULE_LICENSE("GPL");
3667 module_init(lmv_init);
3668 module_exit(lmv_exit);