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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdt/mdt_handler.c
34 * Lustre Metadata Target (mdt) request handler
36 * Author: Peter Braam <braam@clusterfs.com>
37 * Author: Andreas Dilger <adilger@clusterfs.com>
38 * Author: Phil Schwan <phil@clusterfs.com>
39 * Author: Mike Shaver <shaver@clusterfs.com>
40 * Author: Nikita Danilov <nikita@clusterfs.com>
41 * Author: Huang Hua <huanghua@clusterfs.com>
42 * Author: Yury Umanets <umka@clusterfs.com>
45 #define DEBUG_SUBSYSTEM S_MDS
47 #include <linux/module.h>
48 #include <linux/pagemap.h>
50 #include <dt_object.h>
51 #include <lustre_acl.h>
52 #include <lustre_export.h>
53 #include <uapi/linux/lustre/lustre_ioctl.h>
54 #include <lustre_lfsck.h>
55 #include <lustre_log.h>
56 #include <lustre_nodemap.h>
57 #include <lustre_mds.h>
58 #include <uapi/linux/lustre/lustre_param.h>
59 #include <lustre_quota.h>
60 #include <lustre_swab.h>
62 #include <obd_support.h>
63 #include <lustre_barrier.h>
64 #include <obd_cksum.h>
65 #include <llog_swab.h>
67 #include "mdt_internal.h"
70 static unsigned int max_mod_rpcs_per_client = 8;
71 module_param(max_mod_rpcs_per_client, uint, 0644);
72 MODULE_PARM_DESC(max_mod_rpcs_per_client, "maximum number of modify RPCs in flight allowed per client");
74 mdl_mode_t mdt_mdl_lock_modes[] = {
75 [LCK_MINMODE] = MDL_MINMODE,
82 [LCK_GROUP] = MDL_GROUP
85 enum ldlm_mode mdt_dlm_lock_modes[] = {
86 [MDL_MINMODE] = LCK_MINMODE,
93 [MDL_GROUP] = LCK_GROUP
96 static struct mdt_device *mdt_dev(struct lu_device *d);
97 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
99 static const struct lu_object_operations mdt_obj_ops;
101 /* Slab for MDT object allocation */
102 static struct kmem_cache *mdt_object_kmem;
104 /* For HSM restore handles */
105 struct kmem_cache *mdt_hsm_cdt_kmem;
107 /* For HSM request handles */
108 struct kmem_cache *mdt_hsm_car_kmem;
110 static struct lu_kmem_descr mdt_caches[] = {
112 .ckd_cache = &mdt_object_kmem,
113 .ckd_name = "mdt_obj",
114 .ckd_size = sizeof(struct mdt_object)
117 .ckd_cache = &mdt_hsm_cdt_kmem,
118 .ckd_name = "mdt_cdt_restore_handle",
119 .ckd_size = sizeof(struct cdt_restore_handle)
122 .ckd_cache = &mdt_hsm_car_kmem,
123 .ckd_name = "mdt_cdt_agent_req",
124 .ckd_size = sizeof(struct cdt_agent_req)
131 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
135 return rep->lock_policy_res1 & op_flag;
138 void mdt_clear_disposition(struct mdt_thread_info *info,
139 struct ldlm_reply *rep, __u64 op_flag)
142 info->mti_opdata &= ~op_flag;
143 tgt_opdata_clear(info->mti_env, op_flag);
146 rep->lock_policy_res1 &= ~op_flag;
149 void mdt_set_disposition(struct mdt_thread_info *info,
150 struct ldlm_reply *rep, __u64 op_flag)
153 info->mti_opdata |= op_flag;
154 tgt_opdata_set(info->mti_env, op_flag);
157 rep->lock_policy_res1 |= op_flag;
160 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
162 lh->mlh_pdo_hash = 0;
163 lh->mlh_reg_mode = lm;
164 lh->mlh_rreg_mode = lm;
165 lh->mlh_type = MDT_REG_LOCK;
168 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
169 const struct lu_name *lname)
171 lh->mlh_reg_mode = lock_mode;
172 lh->mlh_pdo_mode = LCK_MINMODE;
173 lh->mlh_rreg_mode = lock_mode;
174 lh->mlh_type = MDT_PDO_LOCK;
176 if (lu_name_is_valid(lname)) {
177 lh->mlh_pdo_hash = ll_full_name_hash(NULL, lname->ln_name,
179 /* XXX Workaround for LU-2856
181 * Zero is a valid return value of full_name_hash, but
182 * several users of mlh_pdo_hash assume a non-zero
183 * hash value. We therefore map zero onto an
184 * arbitrary, but consistent value (1) to avoid
185 * problems further down the road. */
186 if (unlikely(lh->mlh_pdo_hash == 0))
187 lh->mlh_pdo_hash = 1;
189 lh->mlh_pdo_hash = 0;
193 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
194 struct mdt_lock_handle *lh)
200 * Any dir access needs couple of locks:
202 * 1) on part of dir we gonna take lookup/modify;
204 * 2) on whole dir to protect it from concurrent splitting and/or to
205 * flush client's cache for readdir().
207 * so, for a given mode and object this routine decides what lock mode
208 * to use for lock #2:
210 * 1) if caller's gonna lookup in dir then we need to protect dir from
211 * being splitted only - LCK_CR
213 * 2) if caller's gonna modify dir then we need to protect dir from
214 * being splitted and to flush cache - LCK_CW
216 * 3) if caller's gonna modify dir and that dir seems ready for
217 * splitting then we need to protect it from any type of access
218 * (lookup/modify/split) - LCK_EX --bzzz
221 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
222 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
225 * Ask underlaying level its opinion about preferable PDO lock mode
226 * having access type passed as regular lock mode:
228 * - MDL_MINMODE means that lower layer does not want to specify lock
231 * - MDL_NL means that no PDO lock should be taken. This is used in some
232 * cases. Say, for non-splittable directories no need to use PDO locks
235 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
236 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
238 if (mode != MDL_MINMODE) {
239 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
242 * Lower layer does not want to specify locking mode. We do it
243 * our selves. No special protection is needed, just flush
244 * client's cache on modification and allow concurrent
247 switch (lh->mlh_reg_mode) {
249 lh->mlh_pdo_mode = LCK_EX;
252 lh->mlh_pdo_mode = LCK_CR;
255 lh->mlh_pdo_mode = LCK_CW;
258 CERROR("Not expected lock type (0x%x)\n",
259 (int)lh->mlh_reg_mode);
264 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
268 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
271 struct mdt_device *mdt = info->mti_mdt;
272 struct lu_name *lname = &info->mti_name;
274 struct mdt_object *parent;
278 LASSERT(!info->mti_cross_ref);
280 OBD_ALLOC(name, NAME_MAX + 1);
283 lname->ln_name = name;
286 * We may want to allow this to mount a completely separate
287 * fileset from the MDT in the future, but keeping it to
288 * ROOT/ only for now avoid potential security issues.
290 *fid = mdt->mdt_md_root_fid;
292 while (rc == 0 && fileset != NULL && *fileset != '\0') {
293 const char *s1 = fileset;
299 while (*s2 != '/' && *s2 != '\0')
307 lname->ln_namelen = s2 - s1;
308 if (lname->ln_namelen > NAME_MAX) {
313 /* reject .. as a path component */
314 if (lname->ln_namelen == 2 &&
315 strncmp(s1, "..", 2) == 0) {
320 strncpy(name, s1, lname->ln_namelen);
321 name[lname->ln_namelen] = '\0';
323 parent = mdt_object_find(info->mti_env, mdt, fid);
324 if (IS_ERR(parent)) {
325 rc = PTR_ERR(parent);
328 /* Only got the fid of this obj by name */
330 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
331 fid, &info->mti_spec);
332 mdt_object_put(info->mti_env, parent);
335 parent = mdt_object_find(info->mti_env, mdt, fid);
337 rc = PTR_ERR(parent);
339 mode = lu_object_attr(&parent->mot_obj);
340 mdt_object_put(info->mti_env, parent);
346 OBD_FREE(name, NAME_MAX + 1);
351 static int mdt_get_root(struct tgt_session_info *tsi)
353 struct mdt_thread_info *info = tsi2mdt_info(tsi);
354 struct mdt_device *mdt = info->mti_mdt;
355 struct mdt_body *repbody;
356 char *fileset = NULL, *buffer = NULL;
358 struct obd_export *exp = info->mti_exp;
359 char *nodemap_fileset;
363 rc = mdt_check_ucred(info);
365 GOTO(out, rc = err_serious(rc));
367 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
368 GOTO(out, rc = err_serious(-ENOMEM));
370 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
371 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
372 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
374 GOTO(out, rc = err_serious(-EFAULT));
377 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
378 if (nodemap_fileset && nodemap_fileset[0]) {
379 CDEBUG(D_INFO, "nodemap fileset is %s\n", nodemap_fileset);
381 /* consider fileset from client as a sub-fileset
382 * of the nodemap one */
383 OBD_ALLOC(buffer, PATH_MAX + 1);
385 GOTO(out, rc = err_serious(-ENOMEM));
386 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
387 nodemap_fileset, fileset) >= PATH_MAX + 1)
388 GOTO(out, rc = err_serious(-EINVAL));
391 /* enforce fileset as specified in the nodemap */
392 fileset = nodemap_fileset;
397 CDEBUG(D_INFO, "Getting fileset %s\n", fileset);
398 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
400 GOTO(out, rc = err_serious(rc));
402 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
404 repbody->mbo_valid |= OBD_MD_FLID;
408 mdt_thread_info_fini(info);
410 OBD_FREE(buffer, PATH_MAX+1);
414 static int mdt_statfs(struct tgt_session_info *tsi)
416 struct ptlrpc_request *req = tgt_ses_req(tsi);
417 struct mdt_thread_info *info = tsi2mdt_info(tsi);
418 struct mdt_device *mdt = info->mti_mdt;
419 struct tg_grants_data *tgd = &mdt->mdt_lut.lut_tgd;
420 struct ptlrpc_service_part *svcpt;
421 struct obd_statfs *osfs;
426 svcpt = req->rq_rqbd->rqbd_svcpt;
428 /* This will trigger a watchdog timeout */
429 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
430 (MDT_SERVICE_WATCHDOG_FACTOR *
431 at_get(&svcpt->scp_at_estimate)) + 1);
433 rc = mdt_check_ucred(info);
435 GOTO(out, rc = err_serious(rc));
437 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
438 GOTO(out, rc = err_serious(-ENOMEM));
440 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
442 GOTO(out, rc = -EPROTO);
444 rc = tgt_statfs_internal(tsi->tsi_env, &mdt->mdt_lut, osfs,
445 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
450 /* at least try to account for cached pages. its still racy and
451 * might be under-reporting if clients haven't announced their
452 * caches with brw recently */
453 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
454 " pending %llu free %llu avail %llu\n",
455 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
456 tgd->tgd_tot_pending,
457 osfs->os_bfree << tgd->tgd_blockbits,
458 osfs->os_bavail << tgd->tgd_blockbits);
460 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
461 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
462 osfs->os_bsize - 1) >> tgd->tgd_blockbits));
464 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
465 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
466 "%llu objects: %llu free; state %x\n",
467 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
468 osfs->os_files, osfs->os_ffree, osfs->os_state);
470 if (!exp_grant_param_supp(tsi->tsi_exp) &&
471 tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT) {
472 /* clients which don't support OBD_CONNECT_GRANT_PARAM
473 * should not see a block size > page size, otherwise
474 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
475 * block size which is the biggest block size known to work
476 * with all client's page size. */
477 osfs->os_blocks <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
478 osfs->os_bfree <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
479 osfs->os_bavail <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
480 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
483 mdt_counter_incr(req, LPROC_MDT_STATFS);
485 mdt_thread_info_fini(info);
490 * Pack size attributes into the reply.
492 int mdt_pack_size2body(struct mdt_thread_info *info,
493 const struct lu_fid *fid, bool dom_lock)
496 struct md_attr *ma = &info->mti_attr;
501 LASSERT(ma->ma_attr.la_valid & LA_MODE);
503 if (!S_ISREG(ma->ma_attr.la_mode) ||
504 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
507 dom_stripe = mdt_lmm_dom_entry(ma->ma_lmm);
508 /* no DoM stripe, no size in reply */
509 if (dom_stripe == LMM_NO_DOM)
512 /* no DoM lock, no size in reply */
516 /* Either DoM lock exists or LMM has only DoM stripe then
517 * return size on body. */
518 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
520 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
524 #ifdef CONFIG_FS_POSIX_ACL
526 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
528 * \param info thread info object
529 * \param repbody reply to pack ACLs into
530 * \param o mdt object of file to examine
531 * \param nodemap nodemap of client to reply to
533 * \retval -errno error getting or parsing ACL from disk
535 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
536 struct mdt_object *o, struct lu_nodemap *nodemap)
538 const struct lu_env *env = info->mti_env;
539 struct md_object *next = mdt_object_child(o);
540 struct lu_buf *buf = &info->mti_buf;
541 struct mdt_device *mdt = info->mti_mdt;
546 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
547 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
549 if (buf->lb_len == 0)
553 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
555 if (rc == -ENODATA) {
556 repbody->mbo_aclsize = 0;
557 repbody->mbo_valid |= OBD_MD_FLACL;
559 } else if (rc == -EOPNOTSUPP) {
563 exp_connect_large_acl(info->mti_exp) &&
564 buf->lb_buf != info->mti_big_acl) {
565 if (info->mti_big_acl == NULL) {
566 OBD_ALLOC_LARGE(info->mti_big_acl,
567 mdt->mdt_max_ea_size);
568 if (info->mti_big_acl == NULL) {
569 CERROR("%s: unable to grow "
572 PFID(mdt_object_fid(o)));
576 info->mti_big_aclsize =
577 mdt->mdt_max_ea_size;
580 CDEBUG(D_INODE, "%s: grow the "DFID
581 " ACL buffer to size %d\n",
583 PFID(mdt_object_fid(o)),
584 mdt->mdt_max_ea_size);
586 buf->lb_buf = info->mti_big_acl;
587 buf->lb_len = info->mti_big_aclsize;
592 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
593 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
596 if (buf->lb_buf == info->mti_big_acl)
597 info->mti_big_acl_used = 1;
599 rc = nodemap_map_acl(nodemap, buf->lb_buf,
600 rc, NODEMAP_FS_TO_CLIENT);
601 /* if all ACLs mapped out, rc is still >= 0 */
603 CERROR("%s: nodemap_map_acl unable to parse "DFID
604 " ACL: rc = %d\n", mdt_obd_name(mdt),
605 PFID(mdt_object_fid(o)), rc);
607 repbody->mbo_aclsize = rc;
608 repbody->mbo_valid |= OBD_MD_FLACL;
617 /* XXX Look into layout in MDT layer. */
618 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
620 struct lov_comp_md_v1 *comp_v1;
621 struct lov_mds_md *v1;
624 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
625 comp_v1 = (struct lov_comp_md_v1 *)lmm;
627 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
628 v1 = (struct lov_mds_md *)((char *)comp_v1 +
629 comp_v1->lcm_entries[i].lcme_offset);
630 /* We don't support partial release for now */
631 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
636 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
641 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
642 const struct lu_attr *attr, const struct lu_fid *fid)
644 struct md_attr *ma = &info->mti_attr;
645 struct obd_export *exp = info->mti_exp;
646 struct lu_nodemap *nodemap = NULL;
648 LASSERT(ma->ma_valid & MA_INODE);
650 if (attr->la_valid & LA_ATIME) {
651 b->mbo_atime = attr->la_atime;
652 b->mbo_valid |= OBD_MD_FLATIME;
654 if (attr->la_valid & LA_MTIME) {
655 b->mbo_mtime = attr->la_mtime;
656 b->mbo_valid |= OBD_MD_FLMTIME;
658 if (attr->la_valid & LA_CTIME) {
659 b->mbo_ctime = attr->la_ctime;
660 b->mbo_valid |= OBD_MD_FLCTIME;
662 if (attr->la_valid & LA_FLAGS) {
663 b->mbo_flags = attr->la_flags;
664 b->mbo_valid |= OBD_MD_FLFLAGS;
666 if (attr->la_valid & LA_NLINK) {
667 b->mbo_nlink = attr->la_nlink;
668 b->mbo_valid |= OBD_MD_FLNLINK;
670 if (attr->la_valid & (LA_UID|LA_GID)) {
671 nodemap = nodemap_get_from_exp(exp);
675 if (attr->la_valid & LA_UID) {
676 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
677 NODEMAP_FS_TO_CLIENT,
679 b->mbo_valid |= OBD_MD_FLUID;
681 if (attr->la_valid & LA_GID) {
682 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
683 NODEMAP_FS_TO_CLIENT,
685 b->mbo_valid |= OBD_MD_FLGID;
688 if (attr->la_valid & LA_PROJID) {
689 /* TODO, nodemap for project id */
690 b->mbo_projid = attr->la_projid;
691 b->mbo_valid |= OBD_MD_FLPROJID;
694 b->mbo_mode = attr->la_mode;
695 if (attr->la_valid & LA_MODE)
696 b->mbo_valid |= OBD_MD_FLMODE;
697 if (attr->la_valid & LA_TYPE)
698 b->mbo_valid |= OBD_MD_FLTYPE;
702 b->mbo_valid |= OBD_MD_FLID;
703 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
704 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
707 if (!(attr->la_valid & LA_TYPE))
710 b->mbo_rdev = attr->la_rdev;
711 b->mbo_size = attr->la_size;
712 b->mbo_blocks = attr->la_blocks;
714 if (!S_ISREG(attr->la_mode)) {
715 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
716 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
717 /* means no objects are allocated on osts. */
718 LASSERT(!(ma->ma_valid & MA_LOV));
719 /* just ignore blocks occupied by extend attributes on MDS */
721 /* if no object is allocated on osts, the size on mds is valid.
723 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
724 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
725 if (mdt_hsm_is_released(ma->ma_lmm)) {
726 /* A released file stores its size on MDS. */
727 /* But return 1 block for released file, unless tools
728 * like tar will consider it fully sparse. (LU-3864)
730 if (unlikely(b->mbo_size == 0))
734 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
735 } else if (info->mti_som_valid) { /* som is valid */
736 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
740 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
741 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
742 PFID(fid), (unsigned long long)b->mbo_size);
745 if (!IS_ERR_OR_NULL(nodemap))
746 nodemap_putref(nodemap);
749 static inline int mdt_body_has_lov(const struct lu_attr *la,
750 const struct mdt_body *body)
752 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
753 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
756 void mdt_client_compatibility(struct mdt_thread_info *info)
758 struct mdt_body *body;
759 struct ptlrpc_request *req = mdt_info_req(info);
760 struct obd_export *exp = req->rq_export;
761 struct md_attr *ma = &info->mti_attr;
762 struct lu_attr *la = &ma->ma_attr;
765 if (exp_connect_layout(exp))
766 /* the client can deal with 16-bit lmm_stripe_count */
769 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
771 if (!mdt_body_has_lov(la, body))
774 /* now we have a reply with a lov for a client not compatible with the
775 * layout lock so we have to clean the layout generation number */
776 if (S_ISREG(la->la_mode))
777 ma->ma_lmm->lmm_layout_gen = 0;
781 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
782 struct mdt_object *o)
784 const struct lu_env *env = info->mti_env;
787 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
796 /* Is it a directory? Let's check for the LMV as well */
797 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
798 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
802 rc2 = mo_xattr_get(env, mdt_object_child(o),
804 XATTR_NAME_DEFAULT_LMV);
806 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
814 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
817 const struct lu_env *env = info->mti_env;
821 LASSERT(info->mti_big_lmm_used == 0);
822 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
826 /* big_lmm may need to be grown */
827 if (info->mti_big_lmmsize < rc) {
828 int size = size_roundup_power2(rc);
830 if (info->mti_big_lmmsize > 0) {
831 /* free old buffer */
832 LASSERT(info->mti_big_lmm);
833 OBD_FREE_LARGE(info->mti_big_lmm,
834 info->mti_big_lmmsize);
835 info->mti_big_lmm = NULL;
836 info->mti_big_lmmsize = 0;
839 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
840 if (info->mti_big_lmm == NULL)
842 info->mti_big_lmmsize = size;
844 LASSERT(info->mti_big_lmmsize >= rc);
846 info->mti_buf.lb_buf = info->mti_big_lmm;
847 info->mti_buf.lb_len = info->mti_big_lmmsize;
848 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
853 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
854 struct md_attr *ma, const char *name)
856 struct md_object *next = mdt_object_child(o);
857 struct lu_buf *buf = &info->mti_buf;
860 if (strcmp(name, XATTR_NAME_LOV) == 0) {
861 buf->lb_buf = ma->ma_lmm;
862 buf->lb_len = ma->ma_lmm_size;
863 LASSERT(!(ma->ma_valid & MA_LOV));
864 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
865 buf->lb_buf = ma->ma_lmv;
866 buf->lb_len = ma->ma_lmv_size;
867 LASSERT(!(ma->ma_valid & MA_LMV));
868 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
869 buf->lb_buf = ma->ma_lmv;
870 buf->lb_len = ma->ma_lmv_size;
871 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
876 rc = mo_xattr_get(info->mti_env, next, buf, name);
880 if (strcmp(name, XATTR_NAME_LOV) == 0) {
881 if (info->mti_big_lmm_used)
882 ma->ma_lmm = info->mti_big_lmm;
884 /* NOT return LOV EA with hole to old client. */
885 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
886 LOV_PATTERN_F_HOLE) &&
887 !(exp_connect_flags(info->mti_exp) &
888 OBD_CONNECT_LFSCK)) {
891 ma->ma_lmm_size = rc;
892 ma->ma_valid |= MA_LOV;
894 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
895 if (info->mti_big_lmm_used)
896 ma->ma_lmv = info->mti_big_lmm;
898 ma->ma_lmv_size = rc;
899 ma->ma_valid |= MA_LMV;
900 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
901 ma->ma_lmv_size = rc;
902 ma->ma_valid |= MA_LMV_DEF;
905 /* Update mdt_max_mdsize so all clients will be aware that */
906 if (info->mti_mdt->mdt_max_mdsize < rc)
907 info->mti_mdt->mdt_max_mdsize = rc;
910 } else if (rc == -ENODATA) {
913 } else if (rc == -ERANGE) {
914 /* Default LMV has fixed size, so it must be able to fit
915 * in the original buffer */
916 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
918 rc = mdt_big_xattr_get(info, o, name);
920 info->mti_big_lmm_used = 1;
928 static int mdt_attr_get_pfid(struct mdt_thread_info *info,
929 struct mdt_object *o, struct lu_fid *pfid)
931 struct lu_buf *buf = &info->mti_buf;
932 struct link_ea_header *leh;
933 struct link_ea_entry *lee;
937 buf->lb_buf = info->mti_big_lmm;
938 buf->lb_len = info->mti_big_lmmsize;
939 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
940 buf, XATTR_NAME_LINK);
941 /* ignore errors, MA_PFID won't be set and it is
942 * up to the caller to treat this as an error */
943 if (rc == -ERANGE || buf->lb_len == 0) {
944 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
945 buf->lb_buf = info->mti_big_lmm;
946 buf->lb_len = info->mti_big_lmmsize;
951 if (rc < sizeof(*leh)) {
952 CERROR("short LinkEA on "DFID": rc = %d\n",
953 PFID(mdt_object_fid(o)), rc);
957 leh = (struct link_ea_header *) buf->lb_buf;
958 lee = (struct link_ea_entry *)(leh + 1);
959 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
960 leh->leh_magic = LINK_EA_MAGIC;
961 leh->leh_reccount = __swab32(leh->leh_reccount);
962 leh->leh_len = __swab64(leh->leh_len);
964 if (leh->leh_magic != LINK_EA_MAGIC)
966 if (leh->leh_reccount == 0)
969 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
970 fid_be_to_cpu(pfid, pfid);
975 int mdt_attr_get_complex(struct mdt_thread_info *info,
976 struct mdt_object *o, struct md_attr *ma)
978 const struct lu_env *env = info->mti_env;
979 struct md_object *next = mdt_object_child(o);
980 struct lu_buf *buf = &info->mti_buf;
981 int need = ma->ma_need;
988 if (mdt_object_exists(o) == 0)
989 GOTO(out, rc = -ENOENT);
990 mode = lu_object_attr(&next->mo_lu);
992 if (need & MA_INODE) {
993 ma->ma_need = MA_INODE;
994 rc = mo_attr_get(env, next, ma);
999 (void) mdt_get_som(info, o, &ma->ma_attr);
1000 ma->ma_valid |= MA_INODE;
1003 if (need & MA_PFID) {
1004 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1006 ma->ma_valid |= MA_PFID;
1007 /* ignore this error, parent fid is not mandatory */
1011 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1012 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1017 if (need & MA_LMV && S_ISDIR(mode)) {
1018 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1023 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1024 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1029 if (need & MA_HSM && S_ISREG(mode)) {
1030 buf->lb_buf = info->mti_xattr_buf;
1031 buf->lb_len = sizeof(info->mti_xattr_buf);
1032 CLASSERT(sizeof(struct hsm_attrs) <=
1033 sizeof(info->mti_xattr_buf));
1034 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1035 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1037 ma->ma_valid |= MA_HSM;
1038 else if (rc2 < 0 && rc2 != -ENODATA)
1039 GOTO(out, rc = rc2);
1042 #ifdef CONFIG_FS_POSIX_ACL
1043 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1044 buf->lb_buf = ma->ma_acl;
1045 buf->lb_len = ma->ma_acl_size;
1046 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1048 ma->ma_acl_size = rc2;
1049 ma->ma_valid |= MA_ACL_DEF;
1050 } else if (rc2 == -ENODATA) {
1052 ma->ma_acl_size = 0;
1054 GOTO(out, rc = rc2);
1059 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1060 rc, ma->ma_valid, ma->ma_lmm);
1064 static int mdt_getattr_internal(struct mdt_thread_info *info,
1065 struct mdt_object *o, int ma_need)
1067 struct md_object *next = mdt_object_child(o);
1068 const struct mdt_body *reqbody = info->mti_body;
1069 struct ptlrpc_request *req = mdt_info_req(info);
1070 struct md_attr *ma = &info->mti_attr;
1071 struct lu_attr *la = &ma->ma_attr;
1072 struct req_capsule *pill = info->mti_pill;
1073 const struct lu_env *env = info->mti_env;
1074 struct mdt_body *repbody;
1075 struct lu_buf *buffer = &info->mti_buf;
1076 struct obd_export *exp = info->mti_exp;
1080 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1081 RETURN(err_serious(-ENOMEM));
1083 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1087 if (mdt_object_remote(o)) {
1088 /* This object is located on remote node.*/
1089 /* Return -ENOTSUPP for old client */
1090 if (!mdt_is_dne_client(req->rq_export))
1091 GOTO(out, rc = -ENOTSUPP);
1093 repbody->mbo_fid1 = *mdt_object_fid(o);
1094 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1098 if (reqbody->mbo_eadatasize > 0) {
1099 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1100 if (buffer->lb_buf == NULL)
1101 GOTO(out, rc = -EPROTO);
1102 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1105 buffer->lb_buf = NULL;
1107 ma_need &= ~(MA_LOV | MA_LMV);
1108 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1109 mdt_obd_name(info->mti_mdt),
1110 req->rq_export->exp_client_uuid.uuid);
1113 /* If it is dir object and client require MEA, then we got MEA */
1114 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1115 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1116 /* Assumption: MDT_MD size is enough for lmv size. */
1117 ma->ma_lmv = buffer->lb_buf;
1118 ma->ma_lmv_size = buffer->lb_len;
1119 ma->ma_need = MA_INODE;
1120 if (ma->ma_lmv_size > 0) {
1121 if (reqbody->mbo_valid & OBD_MD_MEA)
1122 ma->ma_need |= MA_LMV;
1123 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
1124 ma->ma_need |= MA_LMV_DEF;
1127 ma->ma_lmm = buffer->lb_buf;
1128 ma->ma_lmm_size = buffer->lb_len;
1129 ma->ma_need = MA_INODE | MA_HSM;
1130 if (ma->ma_lmm_size > 0)
1131 ma->ma_need |= MA_LOV;
1134 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1135 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1136 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1137 /* get default stripe info for this dir. */
1138 ma->ma_need |= MA_LOV_DEF;
1140 ma->ma_need |= ma_need;
1142 rc = mdt_attr_get_complex(info, o, ma);
1144 CDEBUG(rc == -ENOENT ? D_OTHER : D_ERROR,
1145 "%s: getattr error for "DFID": rc = %d\n",
1146 mdt_obd_name(info->mti_mdt),
1147 PFID(mdt_object_fid(o)), rc);
1151 /* if file is released, check if a restore is running */
1152 if (ma->ma_valid & MA_HSM) {
1153 repbody->mbo_valid |= OBD_MD_TSTATE;
1154 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1155 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1156 repbody->mbo_t_state = MS_RESTORE;
1159 if (likely(ma->ma_valid & MA_INODE))
1160 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1164 if (mdt_body_has_lov(la, reqbody)) {
1165 if (ma->ma_valid & MA_LOV) {
1166 LASSERT(ma->ma_lmm_size);
1167 repbody->mbo_eadatasize = ma->ma_lmm_size;
1168 if (S_ISDIR(la->la_mode))
1169 repbody->mbo_valid |= OBD_MD_FLDIREA;
1171 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1172 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1174 if (ma->ma_valid & MA_LMV) {
1175 /* Return -ENOTSUPP for old client */
1176 if (!mdt_is_striped_client(req->rq_export))
1179 LASSERT(S_ISDIR(la->la_mode));
1180 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1181 repbody->mbo_eadatasize = ma->ma_lmv_size;
1182 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1184 if (ma->ma_valid & MA_LMV_DEF) {
1185 /* Return -ENOTSUPP for old client */
1186 if (!mdt_is_striped_client(req->rq_export))
1188 LASSERT(S_ISDIR(la->la_mode));
1189 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1190 repbody->mbo_eadatasize = ma->ma_lmv_size;
1191 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1192 OBD_MD_DEFAULT_MEA);
1194 } else if (S_ISLNK(la->la_mode) &&
1195 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1196 buffer->lb_buf = ma->ma_lmm;
1197 /* eadatasize from client includes NULL-terminator, so
1198 * there is no need to read it */
1199 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1200 rc = mo_readlink(env, next, buffer);
1201 if (unlikely(rc <= 0)) {
1202 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1203 mdt_obd_name(info->mti_mdt),
1204 PFID(mdt_object_fid(o)), rc);
1207 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1209 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1211 repbody->mbo_valid |= OBD_MD_LINKNAME;
1212 /* we need to report back size with NULL-terminator
1213 * because client expects that */
1214 repbody->mbo_eadatasize = rc + 1;
1215 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1216 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1217 "on "DFID ", expected %d\n",
1218 mdt_obd_name(info->mti_mdt),
1219 rc, PFID(mdt_object_fid(o)),
1220 reqbody->mbo_eadatasize - 1);
1221 /* NULL terminate */
1222 ((char *)ma->ma_lmm)[rc] = 0;
1224 /* If the total CDEBUG() size is larger than a page, it
1225 * will print a warning to the console, avoid this by
1226 * printing just the last part of the symlink. */
1227 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1228 print_limit < rc ? "..." : "", print_limit,
1229 (char *)ma->ma_lmm + rc - print_limit, rc);
1234 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1235 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1236 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1237 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1238 repbody->mbo_max_mdsize);
1241 #ifdef CONFIG_FS_POSIX_ACL
1242 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1243 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1244 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1245 if (IS_ERR(nodemap))
1246 RETURN(PTR_ERR(nodemap));
1248 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1249 nodemap_putref(nodemap);
1255 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1260 static int mdt_getattr(struct tgt_session_info *tsi)
1262 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1263 struct mdt_object *obj = info->mti_object;
1264 struct req_capsule *pill = info->mti_pill;
1265 struct mdt_body *reqbody;
1266 struct mdt_body *repbody;
1270 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1272 LASSERT(obj != NULL);
1273 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1275 /* Unlike intent case where we need to pre-fill out buffers early on
1276 * in intent policy for ldlm reasons, here we can have a much better
1277 * guess at EA size by just reading it from disk.
1278 * Exceptions are readdir and (missing) directory striping */
1280 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1281 /* No easy way to know how long is the symlink, but it cannot
1282 * be more than PATH_MAX, so we allocate +1 */
1285 /* A special case for fs ROOT: getattr there might fetch
1286 * default EA for entire fs, not just for this dir!
1288 } else if (lu_fid_eq(mdt_object_fid(obj),
1289 &info->mti_mdt->mdt_md_root_fid) &&
1290 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1291 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1293 /* Should the default strping be bigger, mdt_fix_reply
1294 * will reallocate */
1295 rc = DEF_REP_MD_SIZE;
1297 /* Read the actual EA size from disk */
1298 rc = mdt_attr_get_eabuf_size(info, obj);
1302 GOTO(out, rc = err_serious(rc));
1304 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1306 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1307 * by default. If the target object has more ACL entries, then
1308 * enlarge the buffer when necessary. */
1309 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1310 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1312 rc = req_capsule_server_pack(pill);
1313 if (unlikely(rc != 0))
1314 GOTO(out, rc = err_serious(rc));
1316 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1317 LASSERT(repbody != NULL);
1318 repbody->mbo_eadatasize = 0;
1319 repbody->mbo_aclsize = 0;
1321 rc = mdt_check_ucred(info);
1323 GOTO(out_shrink, rc);
1325 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1327 rc = mdt_getattr_internal(info, obj, 0);
1330 mdt_client_compatibility(info);
1331 rc2 = mdt_fix_reply(info);
1335 mdt_thread_info_fini(info);
1340 * Handler of layout intent RPC requiring the layout modification
1342 * \param[in] info thread environment
1343 * \param[in] obj object
1344 * \param[in] layout layout change descriptor
1346 * \retval 0 on success
1347 * \retval < 0 error code
1349 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1350 struct md_layout_change *layout)
1352 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1356 if (!mdt_object_exists(obj))
1357 GOTO(out, rc = -ENOENT);
1359 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1360 GOTO(out, rc = -EINVAL);
1362 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1367 /* take layout lock to prepare layout change */
1368 mdt_lock_reg_init(lh, LCK_EX);
1369 rc = mdt_object_lock(info, obj, lh, MDS_INODELOCK_LAYOUT);
1373 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1375 mdt_object_unlock(info, obj, lh, 1);
1381 * Exchange MOF_LOV_CREATED flags between two objects after a
1382 * layout swap. No assumption is made on whether o1 or o2 have
1383 * created objects or not.
1385 * \param[in,out] o1 First swap layout object
1386 * \param[in,out] o2 Second swap layout object
1388 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1390 unsigned int o1_lov_created = o1->mot_lov_created;
1392 mutex_lock(&o1->mot_lov_mutex);
1393 mutex_lock(&o2->mot_lov_mutex);
1395 o1->mot_lov_created = o2->mot_lov_created;
1396 o2->mot_lov_created = o1_lov_created;
1398 mutex_unlock(&o2->mot_lov_mutex);
1399 mutex_unlock(&o1->mot_lov_mutex);
1402 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1404 struct mdt_thread_info *info;
1405 struct ptlrpc_request *req = tgt_ses_req(tsi);
1406 struct obd_export *exp = req->rq_export;
1407 struct mdt_object *o1, *o2, *o;
1408 struct mdt_lock_handle *lh1, *lh2;
1409 struct mdc_swap_layouts *msl;
1413 /* client does not support layout lock, so layout swaping
1415 * FIXME: there is a problem for old clients which don't support
1416 * layout lock yet. If those clients have already opened the file
1417 * they won't be notified at all so that old layout may still be
1418 * used to do IO. This can be fixed after file release is landed by
1419 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1420 if (!exp_connect_layout(exp))
1421 RETURN(-EOPNOTSUPP);
1423 info = tsi2mdt_info(tsi);
1425 if (info->mti_dlm_req != NULL)
1426 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1428 o1 = info->mti_object;
1429 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1430 &info->mti_body->mbo_fid2);
1432 GOTO(out, rc = PTR_ERR(o));
1434 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1435 GOTO(put, rc = -ENOENT);
1437 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1438 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1444 /* permission check. Make sure the calling process having permission
1445 * to write both files. */
1446 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1451 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1456 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1458 GOTO(put, rc = -EPROTO);
1460 lh1 = &info->mti_lh[MDT_LH_NEW];
1461 mdt_lock_reg_init(lh1, LCK_EX);
1462 lh2 = &info->mti_lh[MDT_LH_OLD];
1463 mdt_lock_reg_init(lh2, LCK_EX);
1465 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1466 MDS_INODELOCK_XATTR);
1470 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1471 MDS_INODELOCK_XATTR);
1475 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1476 mdt_object_child(o2), msl->msl_flags);
1480 mdt_swap_lov_flag(o1, o2);
1483 mdt_object_unlock(info, o2, lh2, rc);
1485 mdt_object_unlock(info, o1, lh1, rc);
1487 mdt_object_put(info->mti_env, o);
1489 mdt_thread_info_fini(info);
1493 static int mdt_raw_lookup(struct mdt_thread_info *info,
1494 struct mdt_object *parent,
1495 const struct lu_name *lname,
1496 struct ldlm_reply *ldlm_rep)
1498 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1502 LASSERT(!info->mti_cross_ref);
1504 /* Only got the fid of this obj by name */
1505 fid_zero(child_fid);
1506 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1507 lname, child_fid, &info->mti_spec);
1509 struct mdt_body *repbody;
1511 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1512 repbody->mbo_fid1 = *child_fid;
1513 repbody->mbo_valid = OBD_MD_FLID;
1514 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1515 } else if (rc == -ENOENT) {
1516 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1523 * UPDATE lock should be taken against parent, and be released before exit;
1524 * child_bits lock should be taken against child, and be returned back:
1525 * (1)normal request should release the child lock;
1526 * (2)intent request will grant the lock to client.
1528 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1529 struct mdt_lock_handle *lhc,
1531 struct ldlm_reply *ldlm_rep)
1533 struct ptlrpc_request *req = mdt_info_req(info);
1534 struct mdt_body *reqbody = NULL;
1535 struct mdt_object *parent = info->mti_object;
1536 struct mdt_object *child;
1537 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1538 struct lu_name *lname = NULL;
1539 struct mdt_lock_handle *lhp = NULL;
1540 struct ldlm_lock *lock;
1548 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1549 LASSERT(ergo(is_resent,
1550 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1555 if (info->mti_cross_ref) {
1556 /* Only getattr on the child. Parent is on another node. */
1557 mdt_set_disposition(info, ldlm_rep,
1558 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1560 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1562 PFID(mdt_object_fid(child)), ldlm_rep);
1564 rc = mdt_check_resent_lock(info, child, lhc);
1567 } else if (rc > 0) {
1568 mdt_lock_handle_init(lhc);
1569 mdt_lock_reg_init(lhc, LCK_PR);
1572 * Object's name is on another MDS, no lookup or layout
1573 * lock is needed here but update lock is.
1575 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1576 MDS_INODELOCK_LAYOUT);
1577 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1579 rc = mdt_object_lock(info, child, lhc, child_bits);
1584 /* Finally, we can get attr for child. */
1585 if (!mdt_object_exists(child)) {
1586 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1588 "remote object doesn't exist.");
1589 mdt_object_unlock(info, child, lhc, 1);
1593 rc = mdt_getattr_internal(info, child, 0);
1594 if (unlikely(rc != 0))
1595 mdt_object_unlock(info, child, lhc, 1);
1600 lname = &info->mti_name;
1601 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1603 if (lu_name_is_valid(lname)) {
1604 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1605 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1606 PNAME(lname), ldlm_rep);
1608 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1609 if (unlikely(reqbody == NULL))
1610 RETURN(err_serious(-EPROTO));
1612 *child_fid = reqbody->mbo_fid2;
1614 if (unlikely(!fid_is_sane(child_fid)))
1615 RETURN(err_serious(-EINVAL));
1617 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1619 PFID(mdt_object_fid(parent)),
1620 PFID(&reqbody->mbo_fid2), ldlm_rep);
1623 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1625 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1626 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1628 "Parent doesn't exist!");
1632 if (mdt_object_remote(parent)) {
1633 CERROR("%s: parent "DFID" is on remote target\n",
1634 mdt_obd_name(info->mti_mdt),
1635 PFID(mdt_object_fid(parent)));
1639 if (lu_name_is_valid(lname)) {
1640 /* Always allow to lookup ".." */
1641 if (unlikely(lname->ln_namelen == 2 &&
1642 lname->ln_name[0] == '.' &&
1643 lname->ln_name[1] == '.'))
1644 info->mti_spec.sp_permitted = 1;
1646 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1647 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1652 /* step 1: lock parent only if parent is a directory */
1653 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1654 lhp = &info->mti_lh[MDT_LH_PARENT];
1655 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1656 rc = mdt_object_lock(info, parent, lhp,
1657 MDS_INODELOCK_UPDATE);
1658 if (unlikely(rc != 0))
1662 /* step 2: lookup child's fid by name */
1663 fid_zero(child_fid);
1664 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1665 child_fid, &info->mti_spec);
1667 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1670 GOTO(out_parent, rc);
1673 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1676 *step 3: find the child object by fid & lock it.
1677 * regardless if it is local or remote.
1679 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1680 * set parent dir fid the same as child fid in getattr by fid case
1681 * we should not lu_object_find() the object again, could lead
1682 * to hung if there is a concurrent unlink destroyed the object.
1684 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1685 mdt_object_get(info->mti_env, parent);
1688 child = mdt_object_find(info->mti_env, info->mti_mdt,
1692 if (unlikely(IS_ERR(child)))
1693 GOTO(out_parent, rc = PTR_ERR(child));
1695 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1696 if (!mdt_object_exists(child)) {
1697 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1699 "Object doesn't exist!");
1700 GOTO(out_child, rc = -ENOENT);
1703 rc = mdt_check_resent_lock(info, child, lhc);
1705 GOTO(out_child, rc);
1706 } else if (rc > 0) {
1707 mdt_lock_handle_init(lhc);
1708 mdt_lock_reg_init(lhc, LCK_PR);
1710 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1711 mdt_object_exists(child) && !mdt_object_remote(child)) {
1712 struct md_attr *ma = &info->mti_attr;
1715 ma->ma_need = MA_INODE;
1716 rc = mdt_attr_get_complex(info, child, ma);
1717 if (unlikely(rc != 0))
1718 GOTO(out_child, rc);
1720 /* If the file has not been changed for some time, we
1721 * return not only a LOOKUP lock, but also an UPDATE
1722 * lock and this might save us RPC on later STAT. For
1723 * directories, it also let negative dentry cache start
1724 * working for this dir. */
1725 if (ma->ma_valid & MA_INODE &&
1726 ma->ma_attr.la_valid & LA_CTIME &&
1727 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1728 ma->ma_attr.la_ctime < ktime_get_real_seconds())
1729 child_bits |= MDS_INODELOCK_UPDATE;
1732 /* layout lock must be granted in a best-effort way
1733 * for IT operations */
1734 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1735 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1736 !mdt_object_remote(child) && ldlm_rep != NULL) {
1737 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1738 exp_connect_layout(info->mti_exp)) {
1739 /* try to grant layout lock for regular file. */
1740 try_bits = MDS_INODELOCK_LAYOUT;
1742 /* Acquire DOM lock in advance for data-on-mdt file */
1743 if (child != parent)
1744 try_bits |= MDS_INODELOCK_DOM;
1747 if (try_bits != 0) {
1748 /* try layout lock, it may fail to be granted due to
1749 * contention at LOOKUP or UPDATE */
1750 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
1752 if (child_bits & MDS_INODELOCK_LAYOUT)
1755 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1756 * client will enqueue the lock to the remote MDT */
1757 if (mdt_object_remote(child))
1758 child_bits &= ~MDS_INODELOCK_UPDATE;
1759 rc = mdt_object_lock(info, child, lhc, child_bits);
1761 if (unlikely(rc != 0))
1762 GOTO(out_child, rc);
1765 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1767 /* finally, we can get attr for child. */
1768 rc = mdt_getattr_internal(info, child, ma_need);
1769 if (unlikely(rc != 0)) {
1770 mdt_object_unlock(info, child, lhc, 1);
1772 /* Debugging code. */
1773 LDLM_DEBUG(lock, "Returning lock to client");
1774 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1775 &lock->l_resource->lr_name),
1776 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1777 PLDLMRES(lock->l_resource),
1778 PFID(mdt_object_fid(child)));
1780 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1781 mdt_object_exists(child) && !mdt_object_remote(child) &&
1783 LDLM_LOCK_PUT(lock);
1784 mdt_object_put(info->mti_env, child);
1785 /* NB: call the mdt_pack_size2body always after
1786 * mdt_object_put(), that is why this speacial
1787 * exit path is used. */
1788 rc = mdt_pack_size2body(info, child_fid,
1789 child_bits & MDS_INODELOCK_DOM);
1790 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
1791 /* DOM lock was taken in advance but this is
1792 * not DoM file. Drop the lock. */
1793 lock_res_and_lock(lock);
1794 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
1795 unlock_res_and_lock(lock);
1798 GOTO(out_parent, rc = 0);
1802 LDLM_LOCK_PUT(lock);
1806 mdt_object_put(info->mti_env, child);
1809 mdt_object_unlock(info, parent, lhp, 1);
1813 /* normal handler: should release the child lock */
1814 static int mdt_getattr_name(struct tgt_session_info *tsi)
1816 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1817 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1818 struct mdt_body *reqbody;
1819 struct mdt_body *repbody;
1823 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1824 LASSERT(reqbody != NULL);
1825 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1826 LASSERT(repbody != NULL);
1828 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1829 repbody->mbo_eadatasize = 0;
1830 repbody->mbo_aclsize = 0;
1832 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1834 GOTO(out_shrink, rc);
1836 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1837 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1838 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1839 lhc->mlh_reg_lh.cookie = 0;
1841 mdt_exit_ucred(info);
1844 mdt_client_compatibility(info);
1845 rc2 = mdt_fix_reply(info);
1848 mdt_thread_info_fini(info);
1852 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1853 void *karg, void __user *uarg);
1855 static int mdt_set_info(struct tgt_session_info *tsi)
1857 struct ptlrpc_request *req = tgt_ses_req(tsi);
1860 int keylen, vallen, rc = 0;
1864 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1866 DEBUG_REQ(D_HA, req, "no set_info key");
1867 RETURN(err_serious(-EFAULT));
1870 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1873 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1875 DEBUG_REQ(D_HA, req, "no set_info val");
1876 RETURN(err_serious(-EFAULT));
1879 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1882 /* Swab any part of val you need to here */
1883 if (KEY_IS(KEY_READ_ONLY)) {
1884 spin_lock(&req->rq_export->exp_lock);
1886 *exp_connect_flags_ptr(req->rq_export) |=
1889 *exp_connect_flags_ptr(req->rq_export) &=
1890 ~OBD_CONNECT_RDONLY;
1891 spin_unlock(&req->rq_export->exp_lock);
1892 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1893 struct changelog_setinfo *cs = val;
1895 if (vallen != sizeof(*cs)) {
1896 CERROR("%s: bad changelog_clear setinfo size %d\n",
1897 tgt_name(tsi->tsi_tgt), vallen);
1900 if (ptlrpc_req_need_swab(req)) {
1901 __swab64s(&cs->cs_recno);
1902 __swab32s(&cs->cs_id);
1905 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1907 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
1909 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
1916 static int mdt_readpage(struct tgt_session_info *tsi)
1918 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1919 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1920 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1921 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1922 struct mdt_body *repbody;
1928 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1929 RETURN(err_serious(-ENOMEM));
1931 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1932 if (repbody == NULL || reqbody == NULL)
1933 RETURN(err_serious(-EFAULT));
1936 * prepare @rdpg before calling lower layers and transfer itself. Here
1937 * reqbody->size contains offset of where to start to read and
1938 * reqbody->nlink contains number bytes to read.
1940 rdpg->rp_hash = reqbody->mbo_size;
1941 if (rdpg->rp_hash != reqbody->mbo_size) {
1942 CERROR("Invalid hash: %#llx != %#llx\n",
1943 rdpg->rp_hash, reqbody->mbo_size);
1947 rdpg->rp_attrs = reqbody->mbo_mode;
1948 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1949 rdpg->rp_attrs |= LUDA_64BITHASH;
1950 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1951 exp_max_brw_size(tsi->tsi_exp));
1952 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
1954 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1955 if (rdpg->rp_pages == NULL)
1958 for (i = 0; i < rdpg->rp_npages; ++i) {
1959 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1960 if (rdpg->rp_pages[i] == NULL)
1961 GOTO(free_rdpg, rc = -ENOMEM);
1964 /* call lower layers to fill allocated pages with directory data */
1965 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1967 GOTO(free_rdpg, rc);
1969 /* send pages to client */
1970 rc = tgt_sendpage(tsi, rdpg, rc);
1975 for (i = 0; i < rdpg->rp_npages; i++)
1976 if (rdpg->rp_pages[i] != NULL)
1977 __free_page(rdpg->rp_pages[i]);
1978 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1980 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1986 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
1988 struct lu_ucred *uc = mdt_ucred_check(info);
1989 struct lu_attr *attr = &info->mti_attr.ma_attr;
1994 if (op != REINT_SETATTR) {
1995 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
1996 attr->la_uid = uc->uc_fsuid;
1997 /* for S_ISGID, inherit gid from his parent, such work will be
1998 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
1999 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2000 attr->la_gid = uc->uc_fsgid;
2006 static int mdt_reint_internal(struct mdt_thread_info *info,
2007 struct mdt_lock_handle *lhc,
2010 struct req_capsule *pill = info->mti_pill;
2011 struct mdt_body *repbody;
2016 rc = mdt_reint_unpack(info, op);
2018 CERROR("Can't unpack reint, rc %d\n", rc);
2019 RETURN(err_serious(rc));
2022 /* for replay (no_create) lmm is not needed, client has it already */
2023 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2024 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2027 /* llog cookies are always 0, the field is kept for compatibility */
2028 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2029 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
2031 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
2032 * by default. If the target object has more ACL entries, then
2033 * enlarge the buffer when necessary. */
2034 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
2035 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
2036 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
2038 rc = req_capsule_server_pack(pill);
2040 CERROR("Can't pack response, rc %d\n", rc);
2041 RETURN(err_serious(rc));
2044 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
2045 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
2047 repbody->mbo_eadatasize = 0;
2048 repbody->mbo_aclsize = 0;
2051 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
2053 /* for replay no cookkie / lmm need, because client have this already */
2054 if (info->mti_spec.no_create)
2055 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2056 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
2058 rc = mdt_init_ucred_reint(info);
2060 GOTO(out_shrink, rc);
2062 rc = mdt_fix_attr_ucred(info, op);
2064 GOTO(out_ucred, rc = err_serious(rc));
2066 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
2068 GOTO(out_ucred, rc);
2069 } else if (rc == 1) {
2070 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
2071 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
2072 GOTO(out_ucred, rc);
2074 rc = mdt_reint_rec(info, lhc);
2077 mdt_exit_ucred(info);
2079 mdt_client_compatibility(info);
2080 rc2 = mdt_fix_reply(info);
2086 static long mdt_reint_opcode(struct ptlrpc_request *req,
2087 const struct req_format **fmt)
2089 struct mdt_device *mdt;
2090 struct mdt_rec_reint *rec;
2093 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
2095 opc = rec->rr_opcode;
2096 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
2097 if (opc < REINT_MAX && fmt[opc] != NULL)
2098 req_capsule_extend(&req->rq_pill, fmt[opc]);
2100 mdt = mdt_exp2dev(req->rq_export);
2101 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
2102 " rc = %d\n", req->rq_export->exp_obd->obd_name,
2103 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
2104 opc = err_serious(-EFAULT);
2107 opc = err_serious(-EFAULT);
2112 static int mdt_reint(struct tgt_session_info *tsi)
2116 static const struct req_format *reint_fmts[REINT_MAX] = {
2117 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
2118 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
2119 [REINT_LINK] = &RQF_MDS_REINT_LINK,
2120 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
2121 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
2122 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
2123 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
2124 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
2125 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME,
2126 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
2131 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
2133 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2135 * No lock possible here from client to pass it to reint code
2138 rc = mdt_reint_internal(info, NULL, opc);
2139 mdt_thread_info_fini(info);
2144 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
2148 /* this should sync the whole device */
2149 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
2151 struct dt_device *dt = mdt->mdt_bottom;
2155 rc = dt->dd_ops->dt_sync(env, dt);
2159 /* this should sync this object */
2160 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
2161 struct mdt_object *mo)
2167 if (!mdt_object_exists(mo)) {
2168 CWARN("%s: non existing object "DFID": rc = %d\n",
2169 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
2174 rc = mo_object_sync(env, mdt_object_child(mo));
2179 static int mdt_sync(struct tgt_session_info *tsi)
2181 struct ptlrpc_request *req = tgt_ses_req(tsi);
2182 struct req_capsule *pill = tsi->tsi_pill;
2183 struct mdt_body *body;
2188 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2189 RETURN(err_serious(-ENOMEM));
2191 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2192 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2194 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2196 /* sync an object */
2197 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
2200 const struct lu_fid *fid;
2201 struct lu_attr *la = &info->mti_attr.ma_attr;
2203 info->mti_attr.ma_need = MA_INODE;
2204 info->mti_attr.ma_valid = 0;
2205 rc = mdt_attr_get_complex(info, info->mti_object,
2208 body = req_capsule_server_get(pill,
2210 fid = mdt_object_fid(info->mti_object);
2211 mdt_pack_attr2body(info, body, la, fid);
2214 mdt_thread_info_fini(info);
2217 mdt_counter_incr(req, LPROC_MDT_SYNC);
2222 static int mdt_data_sync(struct tgt_session_info *tsi)
2224 struct mdt_thread_info *info;
2225 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2226 struct ost_body *body = tsi->tsi_ost_body;
2227 struct ost_body *repbody;
2228 struct mdt_object *mo = NULL;
2234 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2236 /* if no fid is specified then do nothing,
2237 * device sync is done via MDS_SYNC */
2238 if (fid_is_zero(&tsi->tsi_fid))
2241 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
2243 RETURN(PTR_ERR(mo));
2245 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
2249 repbody->oa.o_oi = body->oa.o_oi;
2250 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2252 info = tsi2mdt_info(tsi);
2253 ma = &info->mti_attr;
2254 ma->ma_need = MA_INODE;
2256 rc = mdt_attr_get_complex(info, mo, ma);
2258 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
2261 mdt_thread_info_fini(info);
2266 mdt_object_put(tsi->tsi_env, mo);
2271 * Handle quota control requests to consult current usage/limit, but also
2272 * to configure quota enforcement
2274 static int mdt_quotactl(struct tgt_session_info *tsi)
2276 struct obd_export *exp = tsi->tsi_exp;
2277 struct req_capsule *pill = tsi->tsi_pill;
2278 struct obd_quotactl *oqctl, *repoqc;
2280 struct mdt_device *mdt = mdt_exp2dev(exp);
2281 struct lu_device *qmt = mdt->mdt_qmt_dev;
2282 struct lu_nodemap *nodemap;
2285 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2287 RETURN(err_serious(-EPROTO));
2289 rc = req_capsule_server_pack(pill);
2291 RETURN(err_serious(rc));
2293 nodemap = nodemap_get_from_exp(exp);
2294 if (IS_ERR(nodemap))
2295 RETURN(PTR_ERR(nodemap));
2297 switch (oqctl->qc_cmd) {
2298 /* master quotactl */
2301 if (!nodemap_can_setquota(nodemap))
2302 GOTO(out_nodemap, rc = -EPERM);
2306 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2307 /* slave quotactl */
2312 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2313 GOTO(out_nodemap, rc = -EFAULT);
2317 switch (oqctl->qc_type) {
2319 id = nodemap_map_id(nodemap, NODEMAP_UID,
2320 NODEMAP_CLIENT_TO_FS, id);
2323 id = nodemap_map_id(nodemap, NODEMAP_GID,
2324 NODEMAP_CLIENT_TO_FS, id);
2327 /* todo: check/map project id */
2331 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2333 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2335 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2337 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
2338 barrier_exit(tsi->tsi_tgt->lut_bottom);
2340 if (oqctl->qc_id != id)
2341 swap(oqctl->qc_id, id);
2343 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
2344 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
2345 RETURN(-EINPROGRESS);
2348 switch (oqctl->qc_cmd) {
2354 /* forward quotactl request to QMT */
2355 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2360 /* slave quotactl */
2361 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2366 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2367 GOTO(out_nodemap, rc = -EFAULT);
2370 if (oqctl->qc_id != id)
2371 swap(oqctl->qc_id, id);
2378 nodemap_putref(nodemap);
2383 /** clone llog ctxt from child (mdd)
2384 * This allows remote llog (replicator) access.
2385 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2386 * context was originally set up, or we can handle them directly.
2387 * I choose the latter, but that means I need any llog
2388 * contexts set up by child to be accessable by the mdt. So we clone the
2389 * context into our context list here.
2391 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2394 struct md_device *next = mdt->mdt_child;
2395 struct llog_ctxt *ctxt;
2398 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2401 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2402 if (rc || ctxt == NULL) {
2406 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2408 CERROR("Can't set mdt ctxt %d\n", rc);
2413 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2414 struct mdt_device *mdt, int idx)
2416 struct llog_ctxt *ctxt;
2418 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2421 /* Put once for the get we just did, and once for the clone */
2422 llog_ctxt_put(ctxt);
2423 llog_ctxt_put(ctxt);
2428 * sec context handlers
2430 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2432 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2438 * quota request handlers
2440 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2442 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2443 struct lu_device *qmt = mdt->mdt_qmt_dev;
2448 RETURN(err_serious(-EOPNOTSUPP));
2450 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2454 struct mdt_object *mdt_object_new(const struct lu_env *env,
2455 struct mdt_device *d,
2456 const struct lu_fid *f)
2458 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2459 struct lu_object *o;
2460 struct mdt_object *m;
2463 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2464 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2465 if (unlikely(IS_ERR(o)))
2466 m = (struct mdt_object *)o;
2472 struct mdt_object *mdt_object_find(const struct lu_env *env,
2473 struct mdt_device *d,
2474 const struct lu_fid *f)
2476 struct lu_object *o;
2477 struct mdt_object *m;
2480 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2481 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2482 if (unlikely(IS_ERR(o)))
2483 m = (struct mdt_object *)o;
2491 * Asyncronous commit for mdt device.
2493 * Pass asynchonous commit call down the MDS stack.
2495 * \param env environment
2496 * \param mdt the mdt device
2498 static void mdt_device_commit_async(const struct lu_env *env,
2499 struct mdt_device *mdt)
2501 struct dt_device *dt = mdt->mdt_bottom;
2505 rc = dt->dd_ops->dt_commit_async(env, dt);
2506 if (unlikely(rc != 0))
2507 CWARN("%s: async commit start failed: rc = %d\n",
2508 mdt_obd_name(mdt), rc);
2509 atomic_inc(&mdt->mdt_async_commit_count);
2514 * Mark the lock as "synchonous".
2516 * Mark the lock to deffer transaction commit to the unlock time.
2518 * \param lock the lock to mark as "synchonous"
2520 * \see mdt_is_lock_sync
2521 * \see mdt_save_lock
2523 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2525 lock->l_ast_data = (void*)1;
2529 * Check whehter the lock "synchonous" or not.
2531 * \param lock the lock to check
2532 * \retval 1 the lock is "synchonous"
2533 * \retval 0 the lock isn't "synchronous"
2535 * \see mdt_set_lock_sync
2536 * \see mdt_save_lock
2538 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2540 return lock->l_ast_data != NULL;
2544 * Blocking AST for mdt locks.
2546 * Starts transaction commit if in case of COS lock conflict or
2547 * deffers such a commit to the mdt_save_lock.
2549 * \param lock the lock which blocks a request or cancelling lock
2550 * \param desc unused
2551 * \param data unused
2552 * \param flag indicates whether this cancelling or blocking callback
2554 * \see ldlm_blocking_ast_nocheck
2556 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2557 void *data, int flag)
2559 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2560 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2561 bool commit_async = false;
2565 if (flag == LDLM_CB_CANCELING)
2568 lock_res_and_lock(lock);
2569 if (lock->l_blocking_ast != mdt_blocking_ast) {
2570 unlock_res_and_lock(lock);
2573 /* There is no lock conflict if l_blocking_lock == NULL,
2574 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2575 * when the last reference to a local lock was released */
2576 if (lock->l_req_mode & (LCK_PW | LCK_EX) &&
2577 lock->l_blocking_lock != NULL) {
2578 if (mdt_cos_is_enabled(mdt)) {
2579 if (lock->l_client_cookie !=
2580 lock->l_blocking_lock->l_client_cookie)
2581 mdt_set_lock_sync(lock);
2582 } else if (mdt_slc_is_enabled(mdt) &&
2583 ldlm_is_cos_incompat(lock->l_blocking_lock)) {
2584 mdt_set_lock_sync(lock);
2586 * we may do extra commit here, but there is a small
2587 * window to miss a commit: lock was unlocked (saved),
2588 * then a conflict lock queued and we come here, but
2589 * REP-ACK not received, so lock was not converted to
2591 * Fortunately this window is quite small, so the
2592 * extra commit should be rare (not to say distributed
2593 * operation is rare too).
2595 commit_async = true;
2597 } else if (lock->l_req_mode == LCK_COS &&
2598 lock->l_blocking_lock != NULL) {
2599 commit_async = true;
2602 rc = ldlm_blocking_ast_nocheck(lock);
2607 rc = lu_env_init(&env, LCT_LOCAL);
2608 if (unlikely(rc != 0))
2609 CWARN("%s: lu_env initialization failed, cannot "
2610 "start asynchronous commit: rc = %d\n",
2613 mdt_device_commit_async(&env, mdt);
2620 * Blocking AST for cross-MDT lock
2622 * Discard lock from uncommitted_slc_locks and cancel it.
2624 * \param lock the lock which blocks a request or cancelling lock
2625 * \param desc unused
2626 * \param data unused
2627 * \param flag indicates whether this cancelling or blocking callback
2628 * \retval 0 on success
2629 * \retval negative number on error
2631 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2632 void *data, int flag)
2638 case LDLM_CB_BLOCKING: {
2639 struct lustre_handle lockh;
2641 ldlm_lock2handle(lock, &lockh);
2642 rc = ldlm_cli_cancel(&lockh,
2643 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
2645 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2650 case LDLM_CB_CANCELING: {
2651 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2652 struct mdt_device *mdt =
2653 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
2655 LDLM_DEBUG(lock, "Revoke remote lock\n");
2657 /* discard slc lock here so that it can be cleaned anytime,
2658 * especially for cleanup_resource() */
2659 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
2661 /* once we cache lock, l_ast_data is set to mdt_object */
2662 if (lock->l_ast_data != NULL) {
2663 struct mdt_object *mo = lock->l_ast_data;
2666 rc = lu_env_init(&env, LCT_MD_THREAD);
2667 if (unlikely(rc != 0)) {
2668 CWARN("%s: lu_env initialization failed, object"
2669 "%p "DFID" is leaked!\n",
2671 PFID(mdt_object_fid(mo)));
2675 if (lock->l_policy_data.l_inodebits.bits &
2676 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
2677 rc = mo_invalidate(&env, mdt_object_child(mo));
2678 mo->mot_cache_attr = 0;
2680 mdt_object_put(&env, mo);
2692 int mdt_check_resent_lock(struct mdt_thread_info *info,
2693 struct mdt_object *mo,
2694 struct mdt_lock_handle *lhc)
2696 /* the lock might already be gotten in ldlm_handle_enqueue() */
2697 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2698 struct ptlrpc_request *req = mdt_info_req(info);
2699 struct ldlm_lock *lock;
2701 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2702 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2704 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2705 * a resend case, however, it could be already destroyed
2706 * due to client eviction or a raced cancel RPC. */
2707 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
2708 lhc->mlh_reg_lh.cookie);
2712 if (!fid_res_name_eq(mdt_object_fid(mo),
2713 &lock->l_resource->lr_name)) {
2714 CWARN("%s: Although resent, but still not "
2715 "get child lock:"DFID"\n",
2716 info->mti_exp->exp_obd->obd_name,
2717 PFID(mdt_object_fid(mo)));
2718 LDLM_LOCK_PUT(lock);
2721 LDLM_LOCK_PUT(lock);
2727 int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
2728 struct mdt_object *o, const struct lu_fid *fid,
2729 struct lustre_handle *lh, enum ldlm_mode mode,
2730 __u64 *ibits, __u64 trybits, bool cache)
2732 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2733 union ldlm_policy_data *policy = &mti->mti_policy;
2734 struct ldlm_res_id *res_id = &mti->mti_res_id;
2738 LASSERT(mdt_object_remote(o));
2740 fid_build_reg_res_name(fid, res_id);
2742 memset(einfo, 0, sizeof(*einfo));
2743 einfo->ei_type = LDLM_IBITS;
2744 einfo->ei_mode = mode;
2745 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2746 einfo->ei_cb_cp = ldlm_completion_ast;
2747 einfo->ei_enq_slave = 0;
2748 einfo->ei_res_id = res_id;
2752 * if we cache lock, couple lock with mdt_object, so that object
2753 * can be easily found in lock ASTs.
2755 mdt_object_get(mti->mti_env, o);
2756 einfo->ei_cbdata = o;
2760 memset(policy, 0, sizeof(*policy));
2761 policy->l_inodebits.bits = *ibits;
2762 policy->l_inodebits.try_bits = trybits;
2764 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2766 if (rc < 0 && cache) {
2767 mdt_object_put(mti->mti_env, o);
2768 einfo->ei_cbdata = NULL;
2771 /* Return successfully acquired bits to a caller */
2773 struct ldlm_lock *lock = ldlm_handle2lock(lh);
2776 *ibits = lock->l_policy_data.l_inodebits.bits;
2777 LDLM_LOCK_PUT(lock);
2782 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
2783 const struct lu_fid *fid, struct lustre_handle *lh,
2784 enum ldlm_mode mode, __u64 ibits, bool cache)
2786 return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0,
2790 static int mdt_object_local_lock(struct mdt_thread_info *info,
2791 struct mdt_object *o,
2792 struct mdt_lock_handle *lh, __u64 *ibits,
2793 __u64 trybits, bool cos_incompat)
2795 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2796 union ldlm_policy_data *policy = &info->mti_policy;
2797 struct ldlm_res_id *res_id = &info->mti_res_id;
2802 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2803 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2804 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2805 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2808 LASSERT(lh->mlh_reg_mode == LCK_PW ||
2809 lh->mlh_reg_mode == LCK_EX);
2810 dlmflags |= LDLM_FL_COS_INCOMPAT;
2811 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
2812 dlmflags |= LDLM_FL_COS_ENABLED;
2815 /* Only enqueue LOOKUP lock for remote object */
2816 LASSERT(ergo(mdt_object_remote(o), *ibits == MDS_INODELOCK_LOOKUP));
2818 if (lh->mlh_type == MDT_PDO_LOCK) {
2819 /* check for exists after object is locked */
2820 if (mdt_object_exists(o) == 0) {
2821 /* Non-existent object shouldn't have PDO lock */
2824 /* Non-dir object shouldn't have PDO lock */
2825 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2831 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2832 dlmflags |= LDLM_FL_ATOMIC_CB;
2835 * Take PDO lock on whole directory and build correct @res_id for lock
2836 * on part of directory.
2838 if (lh->mlh_pdo_hash != 0) {
2839 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2840 mdt_lock_pdo_mode(info, o, lh);
2841 if (lh->mlh_pdo_mode != LCK_NL) {
2843 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2844 * is never going to be sent to client and we do not
2845 * want it slowed down due to possible cancels.
2847 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2848 policy->l_inodebits.try_bits = 0;
2849 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2850 policy, res_id, dlmflags,
2851 info->mti_exp == NULL ? NULL :
2852 &info->mti_exp->exp_handle.h_cookie);
2853 if (unlikely(rc != 0))
2854 GOTO(out_unlock, rc);
2858 * Finish res_id initializing by name hash marking part of
2859 * directory which is taking modification.
2861 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2864 policy->l_inodebits.bits = *ibits;
2865 policy->l_inodebits.try_bits = trybits;
2868 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2869 * going to be sent to client. If it is - mdt_intent_policy() path will
2870 * fix it up and turn FL_LOCAL flag off.
2872 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2873 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2874 info->mti_exp == NULL ? NULL :
2875 &info->mti_exp->exp_handle.h_cookie);
2878 mdt_object_unlock(info, o, lh, 1);
2879 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2880 lh->mlh_pdo_hash != 0 &&
2881 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2882 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2884 /* Return successfully acquired bits to a caller */
2886 struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh);
2889 *ibits = lock->l_policy_data.l_inodebits.bits;
2890 LDLM_LOCK_PUT(lock);
2896 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2897 struct mdt_lock_handle *lh, __u64 *ibits,
2898 __u64 trybits, bool cos_incompat)
2900 struct mdt_lock_handle *local_lh = NULL;
2904 if (!mdt_object_remote(o)) {
2905 rc = mdt_object_local_lock(info, o, lh, ibits, trybits,
2910 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2911 *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2912 MDS_INODELOCK_XATTR);
2914 /* Only enqueue LOOKUP lock for remote object */
2915 if (*ibits & MDS_INODELOCK_LOOKUP) {
2916 __u64 local = MDS_INODELOCK_LOOKUP;
2918 rc = mdt_object_local_lock(info, o, lh, &local, 0,
2926 if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) {
2927 /* Sigh, PDO needs to enqueue 2 locks right now, but
2928 * enqueue RPC can only request 1 lock, to avoid extra
2929 * RPC, so it will instead enqueue EX lock for remote
2930 * object anyway XXX*/
2931 if (lh->mlh_type == MDT_PDO_LOCK &&
2932 lh->mlh_pdo_hash != 0) {
2933 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2934 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2935 PFID(mdt_object_fid(o)));
2936 lh->mlh_pdo_hash = 0;
2937 lh->mlh_rreg_mode = LCK_EX;
2938 lh->mlh_type = MDT_REG_LOCK;
2941 rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o),
2944 ibits, trybits, false);
2945 if (rc != ELDLM_OK) {
2946 if (local_lh != NULL)
2947 mdt_object_unlock(info, o, local_lh, rc);
2955 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2956 struct mdt_lock_handle *lh, __u64 ibits)
2958 return mdt_object_lock_internal(info, o, lh, &ibits, 0, false);
2961 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2962 struct mdt_lock_handle *lh, __u64 ibits,
2965 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
2966 return mdt_object_lock_internal(info, o, lh, &ibits, 0,
2970 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2971 struct mdt_lock_handle *lh, __u64 *ibits,
2972 __u64 trybits, bool cos_incompat)
2974 bool trylock_only = *ibits == 0;
2977 LASSERT(!(*ibits & trybits));
2978 rc = mdt_object_lock_internal(info, o, lh, ibits, trybits,
2980 if (rc && trylock_only) { /* clear error for try ibits lock only */
2981 LASSERT(*ibits == 0);
2988 * Save a lock within request object.
2990 * Keep the lock referenced until whether client ACK or transaction
2991 * commit happens or release the lock immediately depending on input
2992 * parameters. If COS is ON, a write lock is converted to COS lock
2995 * \param info thead info object
2996 * \param h lock handle
2997 * \param mode lock mode
2998 * \param decref force immediate lock releasing
3000 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
3001 enum ldlm_mode mode, int decref)
3005 if (lustre_handle_is_used(h)) {
3006 if (decref || !info->mti_has_trans ||
3007 !(mode & (LCK_PW | LCK_EX))) {
3008 mdt_fid_unlock(h, mode);
3010 struct mdt_device *mdt = info->mti_mdt;
3011 struct ldlm_lock *lock = ldlm_handle2lock(h);
3012 struct ptlrpc_request *req = mdt_info_req(info);
3013 bool cos = mdt_cos_is_enabled(mdt);
3014 bool convert_lock = !cos && mdt_slc_is_enabled(mdt);
3016 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
3019 /* there is no request if mdt_object_unlock() is called
3020 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
3021 if (likely(req != NULL)) {
3022 LDLM_DEBUG(lock, "save lock request %p reply "
3023 "state %p transno %lld\n", req,
3024 req->rq_reply_state, req->rq_transno);
3026 ldlm_lock_downgrade(lock, LCK_COS);
3029 if (req->rq_export->exp_disconnected)
3030 mdt_fid_unlock(h, mode);
3032 ptlrpc_save_lock(req, h, mode, cos,
3035 mdt_fid_unlock(h, mode);
3037 if (mdt_is_lock_sync(lock)) {
3038 CDEBUG(D_HA, "found sync-lock,"
3039 " async commit started\n");
3040 mdt_device_commit_async(info->mti_env,
3043 LDLM_LOCK_PUT(lock);
3052 * Save cross-MDT lock in uncommitted_slc_locks
3054 * Keep the lock referenced until transaction commit happens or release the lock
3055 * immediately depending on input parameters.
3057 * \param info thead info object
3058 * \param h lock handle
3059 * \param mode lock mode
3060 * \param decref force immediate lock releasing
3062 static void mdt_save_remote_lock(struct mdt_thread_info *info,
3063 struct mdt_object *o, struct lustre_handle *h,
3064 enum ldlm_mode mode, int decref)
3068 if (lustre_handle_is_used(h)) {
3069 struct ldlm_lock *lock = ldlm_handle2lock(h);
3072 (lock->l_policy_data.l_inodebits.bits &
3073 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
3074 mo_invalidate(info->mti_env, mdt_object_child(o));
3076 if (decref || !info->mti_has_trans ||
3077 !(mode & (LCK_PW | LCK_EX))) {
3078 ldlm_lock_decref_and_cancel(h, mode);
3079 LDLM_LOCK_PUT(lock);
3081 struct ptlrpc_request *req = mdt_info_req(info);
3083 LASSERT(req != NULL);
3084 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
3086 ldlm_lock_decref(h, mode);
3095 * Unlock mdt object.
3097 * Immeditely release the regular lock and the PDO lock or save the
3098 * lock in request and keep them referenced until client ACK or
3099 * transaction commit.
3101 * \param info thread info object
3102 * \param o mdt object
3103 * \param lh mdt lock handle referencing regular and PDO locks
3104 * \param decref force immediate lock releasing
3106 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
3108 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
3109 struct mdt_lock_handle *lh, int decref)
3113 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
3114 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
3115 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
3121 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
3122 const struct lu_fid *f,
3123 struct mdt_lock_handle *lh,
3126 struct mdt_object *o;
3128 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
3132 rc = mdt_object_lock(info, o, lh, ibits);
3134 mdt_object_put(info->mti_env, o);
3141 void mdt_object_unlock_put(struct mdt_thread_info * info,
3142 struct mdt_object * o,
3143 struct mdt_lock_handle *lh,
3146 mdt_object_unlock(info, o, lh, decref);
3147 mdt_object_put(info->mti_env, o);
3151 * Generic code handling requests that have struct mdt_body passed in:
3153 * - extract mdt_body from request and save it in @info, if present;
3155 * - create lu_object, corresponding to the fid in mdt_body, and save it in
3158 * - if HABEO_CORPUS flag is set for this request type check whether object
3159 * actually exists on storage (lu_object_exists()).
3162 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
3164 const struct mdt_body *body;
3165 struct mdt_object *obj;
3166 const struct lu_env *env;
3167 struct req_capsule *pill;
3171 env = info->mti_env;
3172 pill = info->mti_pill;
3174 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
3178 if (!(body->mbo_valid & OBD_MD_FLID))
3181 if (!fid_is_sane(&body->mbo_fid1)) {
3182 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
3186 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
3188 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
3189 mdt_object_put(env, obj);
3192 info->mti_object = obj;
3201 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
3203 struct req_capsule *pill = info->mti_pill;
3207 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
3208 rc = mdt_body_unpack(info, flags);
3212 if (rc == 0 && (flags & HABEO_REFERO)) {
3214 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3215 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
3217 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3218 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
3221 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3222 * by default. If the target object has more ACL entries, then
3223 * enlarge the buffer when necessary. */
3224 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3225 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3226 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3228 rc = req_capsule_server_pack(pill);
3233 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
3235 lh->mlh_type = MDT_NUL_LOCK;
3236 lh->mlh_reg_lh.cookie = 0ull;
3237 lh->mlh_reg_mode = LCK_MINMODE;
3238 lh->mlh_pdo_lh.cookie = 0ull;
3239 lh->mlh_pdo_mode = LCK_MINMODE;
3240 lh->mlh_rreg_lh.cookie = 0ull;
3241 lh->mlh_rreg_mode = LCK_MINMODE;
3244 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
3246 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3247 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3251 * Initialize fields of struct mdt_thread_info. Other fields are left in
3252 * uninitialized state, because it's too expensive to zero out whole
3253 * mdt_thread_info (> 1K) on each request arrival.
3255 void mdt_thread_info_init(struct ptlrpc_request *req,
3256 struct mdt_thread_info *info)
3260 info->mti_pill = &req->rq_pill;
3263 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3264 mdt_lock_handle_init(&info->mti_lh[i]);
3266 /* mdt device: it can be NULL while CONNECT */
3267 if (req->rq_export) {
3268 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
3269 info->mti_exp = req->rq_export;
3271 info->mti_mdt = NULL;
3272 info->mti_env = req->rq_svc_thread->t_env;
3273 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
3275 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
3276 info->mti_big_buf = LU_BUF_NULL;
3277 info->mti_body = NULL;
3278 info->mti_object = NULL;
3279 info->mti_dlm_req = NULL;
3280 info->mti_has_trans = 0;
3281 info->mti_cross_ref = 0;
3282 info->mti_opdata = 0;
3283 info->mti_big_lmm_used = 0;
3284 info->mti_big_acl_used = 0;
3285 info->mti_som_valid = 0;
3287 info->mti_spec.no_create = 0;
3288 info->mti_spec.sp_rm_entry = 0;
3289 info->mti_spec.sp_permitted = 0;
3290 info->mti_spec.sp_migrate_close = 0;
3292 info->mti_spec.u.sp_ea.eadata = NULL;
3293 info->mti_spec.u.sp_ea.eadatalen = 0;
3296 void mdt_thread_info_fini(struct mdt_thread_info *info)
3300 if (info->mti_object != NULL) {
3301 mdt_object_put(info->mti_env, info->mti_object);
3302 info->mti_object = NULL;
3305 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3306 mdt_lock_handle_fini(&info->mti_lh[i]);
3307 info->mti_env = NULL;
3308 info->mti_pill = NULL;
3309 info->mti_exp = NULL;
3311 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3312 lu_buf_free(&info->mti_big_buf);
3315 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3317 struct mdt_thread_info *mti;
3319 mti = mdt_th_info(tsi->tsi_env);
3320 LASSERT(mti != NULL);
3322 mdt_thread_info_init(tgt_ses_req(tsi), mti);
3323 if (tsi->tsi_corpus != NULL) {
3324 mti->mti_object = mdt_obj(tsi->tsi_corpus);
3325 lu_object_get(tsi->tsi_corpus);
3327 mti->mti_body = tsi->tsi_mdt_body;
3328 mti->mti_dlm_req = tsi->tsi_dlm_req;
3333 static int mdt_tgt_connect(struct tgt_session_info *tsi)
3335 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
3337 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
3338 set_current_state(TASK_UNINTERRUPTIBLE);
3339 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
3342 return tgt_connect(tsi);
3362 static int mdt_intent_getattr(enum mdt_it_code opcode,
3363 struct mdt_thread_info *info,
3364 struct ldlm_lock **, __u64);
3366 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3367 struct mdt_thread_info *info,
3368 struct ldlm_lock **lockp,
3371 static int mdt_intent_layout(enum mdt_it_code opcode,
3372 struct mdt_thread_info *info,
3373 struct ldlm_lock **,
3375 static int mdt_intent_reint(enum mdt_it_code opcode,
3376 struct mdt_thread_info *info,
3377 struct ldlm_lock **,
3379 static int mdt_intent_glimpse(enum mdt_it_code opcode,
3380 struct mdt_thread_info *info,
3381 struct ldlm_lock **lockp, __u64 flags)
3383 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
3386 static int mdt_intent_brw(enum mdt_it_code opcode,
3387 struct mdt_thread_info *info,
3388 struct ldlm_lock **lockp, __u64 flags)
3390 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
3394 static struct mdt_it_flavor {
3395 const struct req_format *it_fmt;
3397 int (*it_act)(enum mdt_it_code ,
3398 struct mdt_thread_info *,
3399 struct ldlm_lock **,
3402 } mdt_it_flavor[] = {
3404 .it_fmt = &RQF_LDLM_INTENT,
3405 /*.it_flags = HABEO_REFERO,*/
3407 .it_act = mdt_intent_reint,
3408 .it_reint = REINT_OPEN
3411 .it_fmt = &RQF_LDLM_INTENT,
3413 * OCREAT is not a MUTABOR request as if the file
3415 * We do the extra check of OBD_CONNECT_RDONLY in
3416 * mdt_reint_open() when we really need to create
3420 .it_act = mdt_intent_reint,
3421 .it_reint = REINT_OPEN
3424 .it_fmt = &RQF_LDLM_INTENT,
3425 .it_flags = MUTABOR,
3426 .it_act = mdt_intent_reint,
3427 .it_reint = REINT_CREATE
3429 [MDT_IT_GETATTR] = {
3430 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
3431 .it_flags = HABEO_REFERO,
3432 .it_act = mdt_intent_getattr
3434 [MDT_IT_READDIR] = {
3440 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
3441 .it_flags = HABEO_REFERO,
3442 .it_act = mdt_intent_getattr
3445 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
3446 .it_flags = MUTABOR,
3448 .it_reint = REINT_UNLINK
3452 .it_flags = MUTABOR,
3455 [MDT_IT_GETXATTR] = {
3456 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
3457 .it_flags = HABEO_CORPUS,
3458 .it_act = mdt_intent_getxattr
3461 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
3463 .it_act = mdt_intent_layout
3465 [MDT_IT_GLIMPSE] = {
3466 .it_fmt = &RQF_LDLM_INTENT,
3468 .it_act = mdt_intent_glimpse,
3471 .it_fmt = &RQF_LDLM_INTENT,
3473 .it_act = mdt_intent_brw,
3478 int mdt_intent_lock_replace(struct mdt_thread_info *info,
3479 struct ldlm_lock **lockp,
3480 struct mdt_lock_handle *lh,
3481 __u64 flags, int result)
3483 struct ptlrpc_request *req = mdt_info_req(info);
3484 struct ldlm_lock *lock = *lockp;
3485 struct ldlm_lock *new_lock;
3487 /* If possible resent found a lock, @lh is set to its handle */
3488 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3490 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3491 lh->mlh_reg_lh.cookie = 0;
3495 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
3496 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3497 * a resend case, however, it could be already destroyed
3498 * due to client eviction or a raced cancel RPC. */
3499 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
3500 lh->mlh_reg_lh.cookie);
3501 lh->mlh_reg_lh.cookie = 0;
3505 LASSERTF(new_lock != NULL,
3506 "lockh %#llx flags %#llx : rc = %d\n",
3507 lh->mlh_reg_lh.cookie, flags, result);
3510 * If we've already given this lock to a client once, then we should
3511 * have no readers or writers. Otherwise, we should have one reader
3512 * _or_ writer ref (which will be zeroed below) before returning the
3515 if (new_lock->l_export == req->rq_export) {
3516 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3518 LASSERT(new_lock->l_export == NULL);
3519 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3524 if (new_lock->l_export == req->rq_export) {
3526 * Already gave this to the client, which means that we
3527 * reconstructed a reply.
3529 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3532 LDLM_LOCK_RELEASE(new_lock);
3533 lh->mlh_reg_lh.cookie = 0;
3534 RETURN(ELDLM_LOCK_REPLACED);
3538 * Fixup the lock to be given to the client.
3540 lock_res_and_lock(new_lock);
3541 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3542 * possible blocking AST. */
3543 while (new_lock->l_readers > 0) {
3544 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3545 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3546 new_lock->l_readers--;
3548 while (new_lock->l_writers > 0) {
3549 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3550 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3551 new_lock->l_writers--;
3554 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3555 new_lock->l_blocking_ast = lock->l_blocking_ast;
3556 new_lock->l_completion_ast = lock->l_completion_ast;
3557 if (ldlm_has_dom(new_lock))
3558 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
3559 new_lock->l_remote_handle = lock->l_remote_handle;
3560 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3562 unlock_res_and_lock(new_lock);
3564 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3565 &new_lock->l_remote_handle,
3566 &new_lock->l_exp_hash);
3568 LDLM_LOCK_RELEASE(new_lock);
3569 lh->mlh_reg_lh.cookie = 0;
3571 RETURN(ELDLM_LOCK_REPLACED);
3574 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3575 struct ldlm_lock *new_lock,
3576 struct mdt_lock_handle *lh, __u64 flags)
3578 struct ptlrpc_request *req = mdt_info_req(info);
3579 struct ldlm_request *dlmreq;
3581 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3584 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3586 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3587 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3589 if (flags & LDLM_FL_RESENT) {
3590 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3591 lh->mlh_reg_mode = new_lock->l_granted_mode;
3593 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3594 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
3595 lh->mlh_reg_lh.cookie);
3600 * If the xid matches, then we know this is a resent request, and allow
3601 * it. (It's probably an OPEN, for which we don't send a lock.
3603 if (req_can_reconstruct(req, NULL))
3607 * This remote handle isn't enqueued, so we never received or processed
3608 * this request. Clear MSG_RESENT, because it can be handled like any
3609 * normal request now.
3611 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3613 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
3614 dlmreq->lock_handle[0].cookie);
3617 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3618 struct mdt_thread_info *info,
3619 struct ldlm_lock **lockp,
3622 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3623 struct ldlm_reply *ldlm_rep = NULL;
3628 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3629 * (for the resend case) or a new lock. Below we will use it to
3630 * replace the original lock.
3632 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3633 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3634 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3635 rc = mdt_object_lock(info, info->mti_object, lhc,
3636 MDS_INODELOCK_XATTR);
3641 rc = mdt_getxattr(info);
3643 if (mdt_info_req(info)->rq_repmsg != NULL)
3644 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3646 if (ldlm_rep == NULL ||
3647 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
3648 mdt_object_unlock(info, info->mti_object, lhc, 1);
3649 RETURN(err_serious(-EFAULT));
3652 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3654 /* This is left for interop instead of adding a new interop flag.
3656 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
3657 if (ldlm_rep->lock_policy_res2) {
3658 mdt_object_unlock(info, info->mti_object, lhc, 1);
3659 RETURN(ELDLM_LOCK_ABORTED);
3663 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3667 static int mdt_intent_getattr(enum mdt_it_code opcode,
3668 struct mdt_thread_info *info,
3669 struct ldlm_lock **lockp,
3672 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3674 struct ldlm_reply *ldlm_rep;
3675 struct mdt_body *reqbody;
3676 struct mdt_body *repbody;
3680 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3683 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3686 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3687 repbody->mbo_eadatasize = 0;
3688 repbody->mbo_aclsize = 0;
3692 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3694 case MDT_IT_GETATTR:
3695 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3699 CERROR("Unsupported intent (%d)\n", opcode);
3700 GOTO(out_shrink, rc = -EINVAL);
3703 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3705 GOTO(out_shrink, rc);
3707 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3708 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3710 /* Get lock from request for possible resent case. */
3711 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3713 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3714 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3716 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3717 ldlm_rep->lock_policy_res2 = 0;
3718 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3719 ldlm_rep->lock_policy_res2) {
3720 lhc->mlh_reg_lh.cookie = 0ull;
3721 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3724 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3727 mdt_exit_ucred(info);
3729 mdt_client_compatibility(info);
3730 rc2 = mdt_fix_reply(info);
3736 static int mdt_intent_layout(enum mdt_it_code opcode,
3737 struct mdt_thread_info *info,
3738 struct ldlm_lock **lockp,
3741 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3742 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
3743 struct layout_intent *intent;
3744 struct lu_fid *fid = &info->mti_tmp_fid2;
3745 struct mdt_object *obj = NULL;
3746 int layout_size = 0;
3750 if (opcode != MDT_IT_LAYOUT) {
3751 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3756 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3758 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3762 CDEBUG(D_INFO, DFID "got layout change request from client: "
3763 "opc:%u flags:%#x extent "DEXT"\n",
3764 PFID(fid), intent->li_opc, intent->li_flags,
3765 PEXT(&intent->li_extent));
3767 switch (intent->li_opc) {
3768 case LAYOUT_INTENT_TRUNC:
3769 case LAYOUT_INTENT_WRITE:
3770 layout.mlc_opc = MD_LAYOUT_WRITE;
3771 layout.mlc_intent = intent;
3773 case LAYOUT_INTENT_ACCESS:
3775 case LAYOUT_INTENT_READ:
3776 case LAYOUT_INTENT_GLIMPSE:
3777 case LAYOUT_INTENT_RELEASE:
3778 case LAYOUT_INTENT_RESTORE:
3779 CERROR("%s: Unsupported layout intent opc %d\n",
3780 mdt_obd_name(info->mti_mdt), intent->li_opc);
3784 CERROR("%s: Unknown layout intent opc %d\n",
3785 mdt_obd_name(info->mti_mdt), intent->li_opc);
3792 /* Get lock from request for possible resent case. */
3793 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3795 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3797 GOTO(out, rc = PTR_ERR(obj));
3799 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3800 layout_size = mdt_attr_get_eabuf_size(info, obj);
3801 if (layout_size < 0)
3802 GOTO(out_obj, rc = layout_size);
3804 if (layout_size > info->mti_mdt->mdt_max_mdsize)
3805 info->mti_mdt->mdt_max_mdsize = layout_size;
3809 * set reply buffer size, so that ldlm_handle_enqueue0()->
3810 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
3812 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3813 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3815 rc = req_capsule_server_pack(info->mti_pill);
3820 if (layout.mlc_opc != MD_LAYOUT_NOP) {
3821 struct lu_buf *buf = &layout.mlc_buf;
3824 * mdt_layout_change is a reint operation, when the request
3825 * is resent, layout write shouldn't reprocess it again.
3827 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
3829 GOTO(out_obj, rc = rc < 0 ? rc : 0);
3832 * There is another resent case: the client's job has been
3833 * done by another client, referring lod_declare_layout_change
3834 * -EALREADY case, and it became a operation w/o transaction,
3835 * so we should not do the layout change, otherwise
3836 * mdt_layout_change() will try to cancel the granted server
3837 * CR lock whose remote counterpart is still in hold on the
3838 * client, and a deadlock ensues.
3840 rc = mdt_check_resent_lock(info, obj, lhc);
3846 if (unlikely(req_is_replay(mdt_info_req(info)))) {
3847 buf->lb_buf = req_capsule_client_get(info->mti_pill,
3849 buf->lb_len = req_capsule_get_size(info->mti_pill,
3850 &RMF_EADATA, RCL_CLIENT);
3852 * If it's a replay of layout write intent RPC, the
3853 * client has saved the extended lovea when
3854 * it get reply then.
3856 if (buf->lb_len > 0)
3857 mdt_fix_lov_magic(info, buf->lb_buf);
3861 * Instantiate some layout components, if @buf contains
3862 * lovea, then it's a replay of the layout intent write
3865 rc = mdt_layout_change(info, obj, &layout);
3870 mdt_object_put(info->mti_env, obj);
3872 if (rc == 0 && lustre_handle_is_used(&lhc->mlh_reg_lh))
3873 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3876 lhc->mlh_reg_lh.cookie = 0;
3881 static int mdt_intent_reint(enum mdt_it_code opcode,
3882 struct mdt_thread_info *info,
3883 struct ldlm_lock **lockp,
3886 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3887 struct ldlm_reply *rep = NULL;
3891 static const struct req_format *intent_fmts[REINT_MAX] = {
3892 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3893 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3898 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3902 if (mdt_it_flavor[opcode].it_reint != opc) {
3903 CERROR("Reint code %ld doesn't match intent: %d\n",
3905 RETURN(err_serious(-EPROTO));
3908 /* Get lock from request for possible resent case. */
3909 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3911 rc = mdt_reint_internal(info, lhc, opc);
3913 /* Check whether the reply has been packed successfully. */
3914 if (mdt_info_req(info)->rq_repmsg != NULL)
3915 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3917 RETURN(err_serious(-EFAULT));
3919 /* MDC expects this in any case */
3921 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3923 /* the open lock or the lock for cross-ref object should be
3924 * returned to the client */
3925 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3926 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3927 rep->lock_policy_res2 = 0;
3928 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3932 rep->lock_policy_res2 = clear_serious(rc);
3934 if (rep->lock_policy_res2 == -ENOENT &&
3935 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
3936 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
3937 rep->lock_policy_res2 = 0;
3939 lhc->mlh_reg_lh.cookie = 0ull;
3940 if (rc == -ENOTCONN || rc == -ENODEV ||
3941 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3943 * If it is the disconnect error (ENODEV & ENOCONN), the error
3944 * will be returned by rq_status, and client at ptlrpc layer
3945 * will detect this, then disconnect, reconnect the import
3946 * immediately, instead of impacting the following the rpc.
3951 * For other cases, the error will be returned by intent, and client
3952 * will retrieve the result from intent.
3954 RETURN(ELDLM_LOCK_ABORTED);
3957 static int mdt_intent_code(enum ldlm_intent_flags itcode)
3965 case IT_OPEN|IT_CREAT:
3972 rc = MDT_IT_READDIR;
3975 rc = MDT_IT_GETATTR;
3987 rc = MDT_IT_GETXATTR;
3992 case IT_QUOTA_DQACQ:
3997 rc = MDT_IT_GLIMPSE;
4003 CERROR("Unknown intent opcode: 0x%08x\n", itcode);
4010 static int mdt_intent_opc(enum ldlm_intent_flags itopc,
4011 struct mdt_thread_info *info,
4012 struct ldlm_lock **lockp, __u64 flags)
4014 struct req_capsule *pill = info->mti_pill;
4015 struct ptlrpc_request *req = mdt_info_req(info);
4016 struct mdt_it_flavor *flv;
4021 opc = mdt_intent_code(itopc);
4025 if (opc == MDT_IT_QUOTA) {
4026 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
4029 RETURN(-EOPNOTSUPP);
4031 if (mdt_rdonly(req->rq_export))
4034 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
4035 /* pass the request to quota master */
4036 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
4037 mdt_info_req(info), lockp,
4042 flv = &mdt_it_flavor[opc];
4043 if (flv->it_fmt != NULL)
4044 req_capsule_extend(pill, flv->it_fmt);
4046 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
4050 if (flv->it_flags & MUTABOR && mdt_rdonly(req->rq_export))
4053 if (flv->it_act != NULL) {
4054 struct ldlm_reply *rep;
4056 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
4058 /* execute policy */
4059 rc = flv->it_act(opc, info, lockp, flags);
4061 /* Check whether the reply has been packed successfully. */
4062 if (req->rq_repmsg != NULL) {
4063 rep = req_capsule_server_get(info->mti_pill,
4065 rep->lock_policy_res2 =
4066 ptlrpc_status_hton(rep->lock_policy_res2);
4073 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
4074 enum ldlm_intent_flags it_opc)
4076 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
4078 /* update stats when IT code is known */
4079 if (srv_stats != NULL)
4080 lprocfs_counter_incr(srv_stats,
4081 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
4082 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
4085 static int mdt_intent_policy(struct ldlm_namespace *ns,
4086 struct ldlm_lock **lockp, void *req_cookie,
4087 enum ldlm_mode mode, __u64 flags, void *data)
4089 struct tgt_session_info *tsi;
4090 struct mdt_thread_info *info;
4091 struct ptlrpc_request *req = req_cookie;
4092 struct ldlm_intent *it;
4093 struct req_capsule *pill;
4094 const struct ldlm_lock_desc *ldesc;
4099 LASSERT(req != NULL);
4101 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
4103 info = tsi2mdt_info(tsi);
4104 LASSERT(info != NULL);
4105 pill = info->mti_pill;
4106 LASSERT(pill->rc_req == req);
4107 ldesc = &info->mti_dlm_req->lock_desc;
4109 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
4110 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
4111 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
4113 mdt_ptlrpc_stats_update(req, it->opc);
4114 rc = mdt_intent_opc(it->opc, info, lockp, flags);
4118 /* Lock without inodebits makes no sense and will oops
4119 * later in ldlm. Let's check it now to see if we have
4120 * ibits corrupted somewhere in mdt_intent_opc().
4121 * The case for client miss to set ibits has been
4122 * processed by others. */
4123 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
4124 ldesc->l_policy_data.l_inodebits.bits != 0));
4126 rc = err_serious(-EFAULT);
4129 /* No intent was provided */
4130 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
4131 rc = req_capsule_server_pack(pill);
4133 rc = err_serious(rc);
4135 mdt_thread_info_fini(info);
4139 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
4141 struct seq_server_site *ss = mdt_seq_site(mdt);
4143 if (ss->ss_node_id == 0)
4146 if (ss->ss_client_seq != NULL) {
4147 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4148 ss->ss_client_seq->lcs_exp = NULL;
4151 if (ss->ss_server_fld != NULL) {
4152 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
4153 ss->ss_server_fld->lsf_control_exp = NULL;
4157 static void mdt_seq_fini_cli(struct mdt_device *mdt)
4159 struct seq_server_site *ss = mdt_seq_site(mdt);
4164 if (ss->ss_server_seq != NULL)
4165 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
4168 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
4170 mdt_seq_fini_cli(mdt);
4171 mdt_deregister_seq_exp(mdt);
4173 return seq_site_fini(env, mdt_seq_site(mdt));
4177 * It will retrieve its FLDB entries from MDT0, and it only happens
4178 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
4179 * and it needs to refresh FLDB from the MDT0.
4181 static int mdt_register_lwp_callback(void *data)
4184 struct mdt_device *mdt = data;
4185 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
4189 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
4191 rc = lu_env_init(&env, LCT_MD_THREAD);
4193 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
4197 /* Allocate new sequence now to avoid creating local transaction
4198 * in the normal transaction process */
4199 rc = seq_server_check_and_alloc_super(&env,
4200 mdt_seq_site(mdt)->ss_server_seq);
4205 rc = fld_update_from_controller(&env, fld);
4207 CERROR("%s: cannot update controller: rc = %d\n",
4208 mdt_obd_name(mdt), rc);
4217 static int mdt_register_seq_exp(struct mdt_device *mdt)
4219 struct seq_server_site *ss = mdt_seq_site(mdt);
4220 char *lwp_name = NULL;
4223 if (ss->ss_node_id == 0)
4226 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
4227 if (lwp_name == NULL)
4228 GOTO(out_free, rc = -ENOMEM);
4230 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
4234 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
4239 rc = lustre_register_lwp_item(lwp_name,
4240 &ss->ss_server_fld->lsf_control_exp,
4241 mdt_register_lwp_callback, mdt);
4243 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4244 ss->ss_client_seq->lcs_exp = NULL;
4248 if (lwp_name != NULL)
4249 OBD_FREE(lwp_name, MAX_OBD_NAME);
4255 * Init client sequence manager which is used by local MDS to talk to sequence
4256 * controller on remote node.
4258 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
4260 struct seq_server_site *ss = mdt_seq_site(mdt);
4265 /* check if this is adding the first MDC and controller is not yet
4267 OBD_ALLOC_PTR(ss->ss_client_seq);
4268 if (ss->ss_client_seq == NULL)
4271 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
4272 if (prefix == NULL) {
4273 OBD_FREE_PTR(ss->ss_client_seq);
4274 ss->ss_client_seq = NULL;
4278 /* Note: seq_client_fini will be called in seq_site_fini */
4279 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
4280 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
4281 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
4283 OBD_FREE(prefix, MAX_OBD_NAME + 5);
4285 OBD_FREE_PTR(ss->ss_client_seq);
4286 ss->ss_client_seq = NULL;
4290 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
4295 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
4297 struct seq_server_site *ss;
4301 ss = mdt_seq_site(mdt);
4302 /* init sequence controller server(MDT0) */
4303 if (ss->ss_node_id == 0) {
4304 OBD_ALLOC_PTR(ss->ss_control_seq);
4305 if (ss->ss_control_seq == NULL)
4308 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
4309 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
4312 GOTO(out_seq_fini, rc);
4315 /* Init normal sequence server */
4316 OBD_ALLOC_PTR(ss->ss_server_seq);
4317 if (ss->ss_server_seq == NULL)
4318 GOTO(out_seq_fini, rc = -ENOMEM);
4320 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
4321 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
4323 GOTO(out_seq_fini, rc);
4325 /* init seq client for seq server to talk to seq controller(MDT0) */
4326 rc = mdt_seq_init_cli(env, mdt);
4328 GOTO(out_seq_fini, rc);
4330 if (ss->ss_node_id != 0)
4331 /* register controller export through lwp */
4332 rc = mdt_register_seq_exp(mdt);
4337 mdt_seq_fini(env, mdt);
4345 static int mdt_fld_fini(const struct lu_env *env,
4346 struct mdt_device *m)
4348 struct seq_server_site *ss = mdt_seq_site(m);
4351 if (ss && ss->ss_server_fld) {
4352 fld_server_fini(env, ss->ss_server_fld);
4353 OBD_FREE_PTR(ss->ss_server_fld);
4354 ss->ss_server_fld = NULL;
4360 static int mdt_fld_init(const struct lu_env *env,
4362 struct mdt_device *m)
4364 struct seq_server_site *ss;
4368 ss = mdt_seq_site(m);
4370 OBD_ALLOC_PTR(ss->ss_server_fld);
4371 if (ss->ss_server_fld == NULL)
4372 RETURN(rc = -ENOMEM);
4374 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
4377 OBD_FREE_PTR(ss->ss_server_fld);
4378 ss->ss_server_fld = NULL;
4385 static void mdt_stack_pre_fini(const struct lu_env *env,
4386 struct mdt_device *m, struct lu_device *top)
4388 struct lustre_cfg_bufs *bufs;
4389 struct lustre_cfg *lcfg;
4390 struct mdt_thread_info *info;
4395 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4396 LASSERT(info != NULL);
4398 bufs = &info->mti_u.bufs;
4400 LASSERT(m->mdt_child_exp);
4401 LASSERT(m->mdt_child_exp->exp_obd);
4403 /* process cleanup, pass mdt obd name to get obd umount flags */
4404 /* XXX: this is needed because all layers are referenced by
4405 * objects (some of them are pinned by osd, for example *
4406 * the proper solution should be a model where object used
4407 * by osd only doesn't have mdt/mdd slices -bzzz */
4408 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4409 lustre_cfg_bufs_set_string(bufs, 1, NULL);
4410 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4413 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
4415 top->ld_ops->ldo_process_config(env, top, lcfg);
4416 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4420 static void mdt_stack_fini(const struct lu_env *env,
4421 struct mdt_device *m, struct lu_device *top)
4423 struct obd_device *obd = mdt2obd_dev(m);
4424 struct lustre_cfg_bufs *bufs;
4425 struct lustre_cfg *lcfg;
4426 struct mdt_thread_info *info;
4430 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4431 LASSERT(info != NULL);
4433 lu_dev_del_linkage(top->ld_site, top);
4435 lu_site_purge(env, top->ld_site, -1);
4437 bufs = &info->mti_u.bufs;
4438 /* process cleanup, pass mdt obd name to get obd umount flags */
4439 /* another purpose is to let all layers to release their objects */
4440 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4445 lustre_cfg_bufs_set_string(bufs, 1, flags);
4446 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4449 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
4452 top->ld_ops->ldo_process_config(env, top, lcfg);
4453 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4455 lu_site_purge(env, top->ld_site, -1);
4457 m->mdt_child = NULL;
4458 m->mdt_bottom = NULL;
4460 obd_disconnect(m->mdt_child_exp);
4461 m->mdt_child_exp = NULL;
4463 obd_disconnect(m->mdt_bottom_exp);
4464 m->mdt_child_exp = NULL;
4467 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
4468 const char *next, struct obd_export **exp)
4470 struct obd_connect_data *data = NULL;
4471 struct obd_device *obd;
4475 OBD_ALLOC_PTR(data);
4477 GOTO(out, rc = -ENOMEM);
4479 obd = class_name2obd(next);
4481 CERROR("%s: can't locate next device: %s\n",
4482 mdt_obd_name(m), next);
4483 GOTO(out, rc = -ENOTCONN);
4486 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4487 data->ocd_version = LUSTRE_VERSION_CODE;
4489 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
4491 CERROR("%s: cannot connect to next dev %s (%d)\n",
4492 mdt_obd_name(m), next, rc);
4502 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
4503 struct lustre_cfg *cfg)
4505 char *dev = lustre_cfg_string(cfg, 0);
4506 int rc, name_size, uuid_size;
4507 char *name, *uuid, *p;
4508 struct lustre_cfg_bufs *bufs;
4509 struct lustre_cfg *lcfg;
4510 struct obd_device *obd;
4511 struct lustre_profile *lprof;
4512 struct lu_site *site;
4515 /* in 1.8 we had the only device in the stack - MDS.
4516 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
4517 * in 2.3 OSD is instantiated by obd_mount.c, so we need
4518 * to generate names and setup MDT, MDD. MDT will be using
4519 * generated name to connect to MDD. for MDD the next device
4520 * will be LOD with name taken from so called "profile" which
4521 * is generated by mount_option line
4523 * 1.8 MGS generates config. commands like this:
4524 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
4525 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
4526 * 2.0 MGS generates config. commands like this:
4527 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
4528 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4529 * 3:lustre-MDT0000-mdtlov 4:f
4531 * we generate MDD name from MDT one, just replacing T with D
4533 * after all the preparations, the logical equivalent will be
4534 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
4535 * 3:lustre-MDT0000-mdtlov 4:f
4536 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4537 * 3:lustre-MDD0000 4:f
4539 * notice we build the stack from down to top: MDD first, then MDT */
4541 name_size = MAX_OBD_NAME;
4542 uuid_size = MAX_OBD_NAME;
4544 OBD_ALLOC(name, name_size);
4545 OBD_ALLOC(uuid, uuid_size);
4546 if (name == NULL || uuid == NULL)
4547 GOTO(cleanup_mem, rc = -ENOMEM);
4549 OBD_ALLOC_PTR(bufs);
4551 GOTO(cleanup_mem, rc = -ENOMEM);
4554 p = strstr(name, "-MDT");
4556 GOTO(free_bufs, rc = -ENOMEM);
4559 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
4561 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4562 if (lprof == NULL || lprof->lp_dt == NULL) {
4563 CERROR("can't find the profile: %s\n",
4564 lustre_cfg_string(cfg, 0));
4565 GOTO(free_bufs, rc = -EINVAL);
4568 lustre_cfg_bufs_reset(bufs, name);
4569 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
4570 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4571 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4573 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4575 GOTO(put_profile, rc = -ENOMEM);
4576 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4578 rc = class_attach(lcfg);
4580 GOTO(lcfg_cleanup, rc);
4582 obd = class_name2obd(name);
4584 CERROR("Can not find obd %s (%s in config)\n",
4585 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
4586 GOTO(lcfg_cleanup, rc = -EINVAL);
4589 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4591 lustre_cfg_bufs_reset(bufs, name);
4592 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4593 lustre_cfg_bufs_set_string(bufs, 2, dev);
4594 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4596 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4598 GOTO(class_detach, rc = -ENOMEM);
4599 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4601 rc = class_setup(obd, lcfg);
4603 GOTO(class_detach, rc);
4605 /* connect to MDD we just setup */
4606 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
4608 GOTO(class_detach, rc);
4610 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
4612 LASSERT(mdt_lu_site(mdt) == NULL);
4613 mdt->mdt_lu_dev.ld_site = site;
4614 site->ls_top_dev = &mdt->mdt_lu_dev;
4615 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
4617 /* now connect to bottom OSD */
4618 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
4619 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
4621 GOTO(class_detach, rc);
4623 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
4625 rc = lu_env_refill((struct lu_env *)env);
4627 CERROR("Failure to refill session: '%d'\n", rc);
4629 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
4634 class_detach(obd, lcfg);
4636 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4638 class_put_profile(lprof);
4643 OBD_FREE(name, name_size);
4645 OBD_FREE(uuid, uuid_size);
4649 /* setup quota master target on MDT0 */
4650 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
4651 struct lustre_cfg *cfg)
4653 struct obd_device *obd;
4654 char *dev = lustre_cfg_string(cfg, 0);
4655 char *qmtname, *uuid, *p;
4656 struct lustre_cfg_bufs *bufs;
4657 struct lustre_cfg *lcfg;
4658 struct lustre_profile *lprof;
4659 struct obd_connect_data *data;
4663 LASSERT(mdt->mdt_qmt_exp == NULL);
4664 LASSERT(mdt->mdt_qmt_dev == NULL);
4666 /* quota master is on MDT0 only for now */
4667 if (mdt->mdt_seq_site.ss_node_id != 0)
4670 /* MGS generates config commands which look as follows:
4671 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4672 * 3:lustre-MDT0000-mdtlov 4:f
4674 * We generate the QMT name from the MDT one, just replacing MD with QM
4675 * after all the preparations, the logical equivalent will be:
4676 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4677 * 3:lustre-MDT0000-osd 4:f */
4678 OBD_ALLOC(qmtname, MAX_OBD_NAME);
4679 OBD_ALLOC(uuid, UUID_MAX);
4680 OBD_ALLOC_PTR(bufs);
4681 OBD_ALLOC_PTR(data);
4682 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
4683 GOTO(cleanup_mem, rc = -ENOMEM);
4685 strcpy(qmtname, dev);
4686 p = strstr(qmtname, "-MDT");
4688 GOTO(cleanup_mem, rc = -ENOMEM);
4689 /* replace MD with QM */
4693 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4695 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4696 if (lprof == NULL || lprof->lp_dt == NULL) {
4697 CERROR("can't find profile for %s\n",
4698 lustre_cfg_string(cfg, 0));
4699 GOTO(cleanup_mem, rc = -EINVAL);
4702 lustre_cfg_bufs_reset(bufs, qmtname);
4703 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4704 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4705 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4707 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4709 GOTO(put_profile, rc = -ENOMEM);
4710 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4712 rc = class_attach(lcfg);
4714 GOTO(lcfg_cleanup, rc);
4716 obd = class_name2obd(qmtname);
4718 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4719 lustre_cfg_string(cfg, 0));
4720 GOTO(lcfg_cleanup, rc = -EINVAL);
4723 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4725 lustre_cfg_bufs_reset(bufs, qmtname);
4726 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4727 lustre_cfg_bufs_set_string(bufs, 2, dev);
4729 /* for quota, the next device should be the OSD device */
4730 lustre_cfg_bufs_set_string(bufs, 3,
4731 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4733 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4735 GOTO(class_detach, rc = -ENOMEM);
4736 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4738 rc = class_setup(obd, lcfg);
4740 GOTO(class_detach, rc);
4742 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4744 /* configure local quota objects */
4745 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4749 GOTO(class_cleanup, rc);
4751 /* connect to quota master target */
4752 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4753 data->ocd_version = LUSTRE_VERSION_CODE;
4754 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4757 CERROR("cannot connect to quota master device %s (%d)\n",
4759 GOTO(class_cleanup, rc);
4765 class_manual_cleanup(obd);
4766 mdt->mdt_qmt_dev = NULL;
4770 class_detach(obd, lcfg);
4772 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4774 class_put_profile(lprof);
4779 OBD_FREE(qmtname, MAX_OBD_NAME);
4781 OBD_FREE(uuid, UUID_MAX);
4787 /* Shutdown quota master target associated with mdt */
4788 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4792 if (mdt->mdt_qmt_exp == NULL)
4794 LASSERT(mdt->mdt_qmt_dev != NULL);
4796 /* the qmt automatically shuts down when the mdt disconnects */
4797 obd_disconnect(mdt->mdt_qmt_exp);
4798 mdt->mdt_qmt_exp = NULL;
4799 mdt->mdt_qmt_dev = NULL;
4803 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4804 * for now. This will be removed along with converting rest of MDT code
4805 * to use tgt_session_info */
4806 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4808 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4811 rc = mdt_getxattr(info);
4813 mdt_thread_info_fini(info);
4817 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
4818 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
4819 #define OST_BRW_READ OST_READ
4820 #define OST_BRW_WRITE OST_WRITE
4822 static struct tgt_handler mdt_tgt_handlers[] = {
4823 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4824 0, MDS_CONNECT, mdt_tgt_connect,
4825 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4826 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4827 0, MDS_DISCONNECT, tgt_disconnect,
4828 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4829 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4830 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4831 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4832 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4833 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GET_ROOT, mdt_get_root),
4834 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4835 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4837 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4838 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4839 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4840 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4841 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4842 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4843 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4844 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4846 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4847 mdt_hsm_ct_register),
4848 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4849 mdt_hsm_ct_unregister),
4850 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4852 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4854 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4855 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4857 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4862 static struct tgt_handler mdt_io_ops[] = {
4863 TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO, OST_BRW_READ, tgt_brw_read),
4864 TGT_OST_HDL(HABEO_CORPUS | MUTABOR, OST_BRW_WRITE, tgt_brw_write),
4865 TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4866 OST_PUNCH, mdt_punch_hdl),
4867 TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO, OST_SYNC, mdt_data_sync),
4870 static struct tgt_handler mdt_sec_ctx_ops[] = {
4871 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4872 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4873 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4876 static struct tgt_handler mdt_quota_ops[] = {
4877 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4880 static struct tgt_opc_slice mdt_common_slice[] = {
4882 .tos_opc_start = MDS_FIRST_OPC,
4883 .tos_opc_end = MDS_LAST_OPC,
4884 .tos_hs = mdt_tgt_handlers
4887 .tos_opc_start = OBD_FIRST_OPC,
4888 .tos_opc_end = OBD_LAST_OPC,
4889 .tos_hs = tgt_obd_handlers
4892 .tos_opc_start = LDLM_FIRST_OPC,
4893 .tos_opc_end = LDLM_LAST_OPC,
4894 .tos_hs = tgt_dlm_handlers
4897 .tos_opc_start = SEC_FIRST_OPC,
4898 .tos_opc_end = SEC_LAST_OPC,
4899 .tos_hs = mdt_sec_ctx_ops
4902 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4903 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4904 .tos_hs = tgt_out_handlers
4907 .tos_opc_start = FLD_FIRST_OPC,
4908 .tos_opc_end = FLD_LAST_OPC,
4909 .tos_hs = fld_handlers
4912 .tos_opc_start = SEQ_FIRST_OPC,
4913 .tos_opc_end = SEQ_LAST_OPC,
4914 .tos_hs = seq_handlers
4917 .tos_opc_start = QUOTA_DQACQ,
4918 .tos_opc_end = QUOTA_LAST_OPC,
4919 .tos_hs = mdt_quota_ops
4922 .tos_opc_start = LLOG_FIRST_OPC,
4923 .tos_opc_end = LLOG_LAST_OPC,
4924 .tos_hs = tgt_llog_handlers
4927 .tos_opc_start = LFSCK_FIRST_OPC,
4928 .tos_opc_end = LFSCK_LAST_OPC,
4929 .tos_hs = tgt_lfsck_handlers
4932 .tos_opc_start = OST_FIRST_OPC,
4933 .tos_opc_end = OST_LAST_OPC,
4934 .tos_hs = mdt_io_ops
4941 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4943 struct md_device *next = m->mdt_child;
4944 struct lu_device *d = &m->mdt_lu_dev;
4945 struct obd_device *obd = mdt2obd_dev(m);
4946 struct lfsck_stop stop;
4949 stop.ls_status = LS_PAUSED;
4951 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4953 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4954 ping_evictor_stop();
4956 /* Remove the HSM /proc entry so the coordinator cannot be
4957 * restarted by a user while it's shutting down. */
4958 hsm_cdt_procfs_fini(m);
4959 mdt_hsm_cdt_stop(m);
4961 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4962 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4964 if (m->mdt_namespace != NULL)
4965 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
4966 d->ld_obd->obd_force);
4968 obd_exports_barrier(obd);
4969 obd_zombie_barrier();
4971 mdt_quota_fini(env, m);
4973 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4975 /* Calling the cleanup functions in the same order as in the mdt_init0
4980 target_recovery_fini(obd);
4981 upcall_cache_cleanup(m->mdt_identity_cache);
4982 m->mdt_identity_cache = NULL;
4984 mdt_fs_cleanup(env, m);
4986 tgt_fini(env, &m->mdt_lut);
4988 mdt_hsm_cdt_fini(m);
4990 if (m->mdt_namespace != NULL) {
4991 ldlm_namespace_free_post(m->mdt_namespace);
4992 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4995 if (m->mdt_md_root != NULL) {
4996 mdt_object_put(env, m->mdt_md_root);
4997 m->mdt_md_root = NULL;
5000 mdt_seq_fini(env, m);
5002 mdt_fld_fini(env, m);
5007 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5009 LASSERT(atomic_read(&d->ld_ref) == 0);
5011 server_put_mount(mdt_obd_name(m), true);
5016 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
5018 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
5019 struct lu_device_type *ldt, struct lustre_cfg *cfg)
5021 const struct dt_device_param *dt_conf;
5022 struct mdt_thread_info *info;
5023 struct obd_device *obd;
5024 const char *dev = lustre_cfg_string(cfg, 0);
5025 const char *num = lustre_cfg_string(cfg, 2);
5026 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
5027 struct lustre_mount_info *lmi = NULL;
5028 struct lustre_sb_info *lsi;
5030 struct seq_server_site *ss_site;
5031 const char *identity_upcall = "NONE";
5032 struct md_device *next;
5038 lu_device_init(&m->mdt_lu_dev, ldt);
5040 * Environment (env) might be missing mdt_thread_key values at that
5041 * point, if device is allocated when mdt_thread_key is in QUIESCENT
5044 * Usually device allocation path doesn't use module key values, but
5045 * mdt has to do a lot of work here, so allocate key value.
5047 rc = lu_env_refill((struct lu_env *)env);
5051 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5052 LASSERT(info != NULL);
5054 obd = class_name2obd(dev);
5055 LASSERT(obd != NULL);
5057 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
5058 m->mdt_opts.mo_evict_tgt_nids = 1;
5059 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
5061 lmi = server_get_mount(dev);
5063 CERROR("Cannot get mount info for %s!\n", dev);
5066 lsi = s2lsi(lmi->lmi_sb);
5067 /* CMD is supported only in IAM mode */
5069 node_id = simple_strtol(num, NULL, 10);
5070 obd->u.obt.obt_magic = OBT_MAGIC;
5071 if (lsi->lsi_lmd != NULL &&
5072 lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
5073 m->mdt_skip_lfsck = 1;
5076 /* DoM files get IO lock at open by default */
5077 m->mdt_opts.mo_dom_lock = 1;
5079 m->mdt_squash.rsi_uid = 0;
5080 m->mdt_squash.rsi_gid = 0;
5081 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
5082 init_rwsem(&m->mdt_squash.rsi_sem);
5083 spin_lock_init(&m->mdt_lock);
5084 m->mdt_enable_remote_dir = 0;
5085 m->mdt_enable_remote_dir_gid = 0;
5087 atomic_set(&m->mdt_mds_mds_conns, 0);
5088 atomic_set(&m->mdt_async_commit_count, 0);
5090 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
5091 m->mdt_lu_dev.ld_obd = obd;
5092 /* Set this lu_device to obd for error handling purposes. */
5093 obd->obd_lu_dev = &m->mdt_lu_dev;
5095 /* init the stack */
5096 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
5098 CERROR("%s: Can't init device stack, rc %d\n",
5099 mdt_obd_name(m), rc);
5104 ss_site = mdt_seq_site(m);
5105 s->ld_seq_site = ss_site;
5108 /* set server index */
5109 ss_site->ss_node_id = node_id;
5111 /* failover is the default
5112 * FIXME: we do not failout mds0/mgs, which may cause some problems.
5113 * assumed whose ss_node_id == 0 XXX
5115 obd->obd_replayable = 1;
5116 /* No connection accepted until configurations will finish */
5117 obd->obd_no_conn = 1;
5119 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
5120 char *str = lustre_cfg_string(cfg, 4);
5121 if (strchr(str, 'n')) {
5122 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
5123 obd->obd_replayable = 0;
5127 rc = mdt_fld_init(env, mdt_obd_name(m), m);
5129 GOTO(err_fini_stack, rc);
5131 rc = mdt_seq_init(env, m);
5133 GOTO(err_fini_fld, rc);
5135 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
5136 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
5137 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
5138 LDLM_NAMESPACE_SERVER,
5139 LDLM_NAMESPACE_GREEDY,
5141 if (m->mdt_namespace == NULL)
5142 GOTO(err_fini_seq, rc = -ENOMEM);
5144 m->mdt_namespace->ns_lvbp = m;
5145 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
5147 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
5148 /* set obd_namespace for compatibility with old code */
5149 obd->obd_namespace = m->mdt_namespace;
5151 rc = mdt_hsm_cdt_init(m);
5153 CERROR("%s: error initializing coordinator, rc %d\n",
5154 mdt_obd_name(m), rc);
5155 GOTO(err_free_ns, rc);
5158 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
5159 OBD_FAIL_MDS_ALL_REQUEST_NET,
5160 OBD_FAIL_MDS_ALL_REPLY_NET);
5162 GOTO(err_free_hsm, rc);
5164 /* Amount of available space excluded from granting and reserved
5165 * for metadata. It is in percentage and 50% is default value. */
5166 tgd->tgd_reserved_pcnt = 50;
5168 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
5169 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
5171 m->mdt_brw_size = ONE_MB_BRW_SIZE;
5173 rc = mdt_fs_setup(env, m, obd, lsi);
5177 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
5179 next = m->mdt_child;
5180 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
5182 mntopts = dt_conf->ddp_mntopts;
5184 if (mntopts & MNTOPT_USERXATTR)
5185 m->mdt_opts.mo_user_xattr = 1;
5187 m->mdt_opts.mo_user_xattr = 0;
5189 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
5191 if (mntopts & MNTOPT_ACL)
5192 m->mdt_opts.mo_acl = 1;
5194 m->mdt_opts.mo_acl = 0;
5196 /* XXX: to support suppgid for ACL, we enable identity_upcall
5197 * by default, otherwise, maybe got unexpected -EACCESS. */
5198 if (m->mdt_opts.mo_acl)
5199 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
5201 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
5203 &mdt_identity_upcall_cache_ops);
5204 if (IS_ERR(m->mdt_identity_cache)) {
5205 rc = PTR_ERR(m->mdt_identity_cache);
5206 m->mdt_identity_cache = NULL;
5207 GOTO(err_fs_cleanup, rc);
5210 rc = mdt_procfs_init(m, dev);
5212 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
5213 GOTO(err_recovery, rc);
5216 rc = mdt_quota_init(env, m, cfg);
5218 GOTO(err_procfs, rc);
5220 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
5221 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
5222 "mdt_ldlm_client", m->mdt_ldlm_client);
5224 ping_evictor_start();
5226 /* recovery will be started upon mdt_prepare()
5227 * when the whole stack is complete and ready
5228 * to serve the requests */
5230 /* Reduce the initial timeout on an MDS because it doesn't need such
5231 * a long timeout as an OST does. Adaptive timeouts will adjust this
5232 * value appropriately. */
5233 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
5234 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
5240 target_recovery_fini(obd);
5241 upcall_cache_cleanup(m->mdt_identity_cache);
5242 m->mdt_identity_cache = NULL;
5244 mdt_fs_cleanup(env, m);
5246 tgt_fini(env, &m->mdt_lut);
5248 mdt_hsm_cdt_fini(m);
5250 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
5251 obd->obd_namespace = m->mdt_namespace = NULL;
5253 mdt_seq_fini(env, m);
5255 mdt_fld_fini(env, m);
5257 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5260 server_put_mount(dev, true);
5264 /* For interoperability, the left element is old parameter, the right one
5265 * is the new version of the parameter, if some parameter is deprecated,
5266 * the new version should be set as NULL. */
5267 static struct cfg_interop_param mdt_interop_param[] = {
5268 { "mdt.group_upcall", NULL },
5269 { "mdt.quota_type", NULL },
5270 { "mdd.quota_type", NULL },
5271 { "mdt.som", NULL },
5272 { "mdt.rootsquash", "mdt.root_squash" },
5273 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
5277 /* used by MGS to process specific configurations */
5278 static int mdt_process_config(const struct lu_env *env,
5279 struct lu_device *d, struct lustre_cfg *cfg)
5281 struct mdt_device *m = mdt_dev(d);
5282 struct md_device *md_next = m->mdt_child;
5283 struct lu_device *next = md2lu_dev(md_next);
5287 switch (cfg->lcfg_command) {
5289 struct obd_device *obd = d->ld_obd;
5291 /* For interoperability */
5292 struct cfg_interop_param *ptr = NULL;
5293 struct lustre_cfg *old_cfg = NULL;
5296 param = lustre_cfg_string(cfg, 1);
5297 if (param == NULL) {
5298 CERROR("param is empty\n");
5303 ptr = class_find_old_param(param, mdt_interop_param);
5305 if (ptr->new_param == NULL) {
5307 CWARN("For interoperability, skip this %s."
5308 " It is obsolete.\n", ptr->old_param);
5312 CWARN("Found old param %s, changed it to %s.\n",
5313 ptr->old_param, ptr->new_param);
5316 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
5323 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
5325 if (rc > 0 || rc == -ENOSYS) {
5326 /* is it an HSM var ? */
5327 rc = class_process_proc_param(PARAM_HSM,
5328 hsm_cdt_get_proc_vars(),
5330 if (rc > 0 || rc == -ENOSYS)
5331 /* we don't understand; pass it on */
5332 rc = next->ld_ops->ldo_process_config(env, next,
5337 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
5338 cfg->lcfg_buflens));
5342 /* others are passed further */
5343 rc = next->ld_ops->ldo_process_config(env, next, cfg);
5349 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
5350 const struct lu_object_header *hdr,
5351 struct lu_device *d)
5353 struct mdt_object *mo;
5357 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
5359 struct lu_object *o;
5360 struct lu_object_header *h;
5363 h = &mo->mot_header;
5364 lu_object_header_init(h);
5365 lu_object_init(o, h, d);
5366 lu_object_add_top(h, o);
5367 o->lo_ops = &mdt_obj_ops;
5368 spin_lock_init(&mo->mot_write_lock);
5369 mutex_init(&mo->mot_lov_mutex);
5370 init_rwsem(&mo->mot_dom_sem);
5371 init_rwsem(&mo->mot_open_sem);
5372 atomic_set(&mo->mot_open_count, 0);
5378 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
5379 const struct lu_object_conf *unused)
5381 struct mdt_device *d = mdt_dev(o->lo_dev);
5382 struct lu_device *under;
5383 struct lu_object *below;
5387 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
5388 PFID(lu_object_fid(o)));
5390 under = &d->mdt_child->md_lu_dev;
5391 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
5392 if (below != NULL) {
5393 lu_object_add(o, below);
5400 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
5402 struct mdt_object *mo = mdt_obj(o);
5403 struct lu_object_header *h;
5407 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
5408 PFID(lu_object_fid(o)));
5410 LASSERT(atomic_read(&mo->mot_open_count) == 0);
5411 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
5414 lu_object_header_fini(h);
5415 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
5420 static int mdt_object_print(const struct lu_env *env, void *cookie,
5421 lu_printer_t p, const struct lu_object *o)
5423 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
5425 return (*p)(env, cookie,
5426 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
5427 mdto, mdto->mot_lov_created ? "lov_created" : "",
5428 mdto->mot_cache_attr ? "cache_attr" : "",
5429 mdto->mot_write_count);
5432 static int mdt_prepare(const struct lu_env *env,
5433 struct lu_device *pdev,
5434 struct lu_device *cdev)
5436 struct mdt_device *mdt = mdt_dev(cdev);
5437 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
5438 struct obd_device *obd = cdev->ld_obd;
5445 rc = next->ld_ops->ldo_prepare(env, cdev, next);
5449 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
5453 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
5457 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
5458 /* The LFSCK instance is registered just now, so it must be there when
5459 * register the namespace to such instance. */
5460 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
5462 if (mdt->mdt_seq_site.ss_node_id == 0) {
5463 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
5464 &mdt->mdt_md_root_fid);
5469 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
5471 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
5472 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
5473 LASSERT(obd->obd_no_conn);
5474 spin_lock(&obd->obd_dev_lock);
5475 obd->obd_no_conn = 0;
5476 spin_unlock(&obd->obd_dev_lock);
5478 if (obd->obd_recovering == 0)
5479 mdt_postrecov(env, mdt);
5484 const struct lu_device_operations mdt_lu_ops = {
5485 .ldo_object_alloc = mdt_object_alloc,
5486 .ldo_process_config = mdt_process_config,
5487 .ldo_prepare = mdt_prepare,
5490 static const struct lu_object_operations mdt_obj_ops = {
5491 .loo_object_init = mdt_object_init,
5492 .loo_object_free = mdt_object_free,
5493 .loo_object_print = mdt_object_print
5496 static int mdt_obd_set_info_async(const struct lu_env *env,
5497 struct obd_export *exp,
5498 __u32 keylen, void *key,
5499 __u32 vallen, void *val,
5500 struct ptlrpc_request_set *set)
5506 if (KEY_IS(KEY_SPTLRPC_CONF)) {
5507 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
5515 * Match client and server connection feature flags.
5517 * Compute the compatibility flags for a connection request based on
5518 * features mutually supported by client and server.
5520 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
5521 * must not be updated here, otherwise a partially initialized value may
5522 * be exposed. After the connection request is successfully processed,
5523 * the top-level MDT connect request handler atomically updates the export
5524 * connect flags from the obd_connect_data::ocd_connect_flags field of the
5525 * reply. \see mdt_connect().
5527 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
5528 * full struct obd_connect_data. So care must be taken when accessing fields
5529 * that are not present in struct obd_connect_data_v1. See LU-16.
5531 * \param exp the obd_export associated with this client/target pair
5532 * \param mdt the target device for the connection
5533 * \param data stores data for this connect request
5536 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
5537 * \retval -EBADE client and server feature requirements are incompatible
5539 static int mdt_connect_internal(const struct lu_env *env,
5540 struct obd_export *exp,
5541 struct mdt_device *mdt,
5542 struct obd_connect_data *data, bool reconnect)
5544 LASSERT(data != NULL);
5546 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
5548 if (mdt->mdt_bottom->dd_rdonly &&
5549 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5550 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
5553 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
5554 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
5556 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
5558 if (!mdt->mdt_opts.mo_acl)
5559 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
5561 if (!mdt->mdt_opts.mo_user_xattr)
5562 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
5564 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
5565 data->ocd_brw_size = min(data->ocd_brw_size,
5567 if (data->ocd_brw_size == 0) {
5568 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
5569 "ocd_version: %x ocd_grant: %d ocd_index: %u "
5570 "ocd_brw_size unexpectedly zero, network data "
5571 "corruption? Refusing to connect this client\n",
5573 exp->exp_client_uuid.uuid,
5574 exp, data->ocd_connect_flags, data->ocd_version,
5575 data->ocd_grant, data->ocd_index);
5580 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
5581 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
5583 /* client is reporting its page size, for future use */
5584 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
5585 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
5586 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
5587 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
5588 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
5589 /* ocd_grant_tax_kb is in 1K byte blocks */
5590 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
5591 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
5594 if (OCD_HAS_FLAG(data, GRANT)) {
5595 /* Save connect_data we have so far because tgt_grant_connect()
5596 * uses it to calculate grant. */
5597 exp->exp_connect_data = *data;
5598 tgt_grant_connect(env, exp, data, !reconnect);
5601 if (OCD_HAS_FLAG(data, MAXBYTES))
5602 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
5604 /* NB: Disregard the rule against updating
5605 * exp_connect_data.ocd_connect_flags in this case, since
5606 * tgt_client_new() needs to know if this is a lightweight
5607 * connection, and it is safe to expose this flag before
5608 * connection processing completes. */
5609 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
5610 spin_lock(&exp->exp_lock);
5611 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
5612 spin_unlock(&exp->exp_lock);
5615 data->ocd_version = LUSTRE_VERSION_CODE;
5617 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
5618 CWARN("%s: MDS requires FID support, but client not\n",
5623 if (OCD_HAS_FLAG(data, PINGLESS)) {
5624 if (ptlrpc_pinger_suppress_pings()) {
5625 spin_lock(&exp->exp_obd->obd_dev_lock);
5626 list_del_init(&exp->exp_obd_chain_timed);
5627 spin_unlock(&exp->exp_obd->obd_dev_lock);
5629 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
5633 data->ocd_max_easize = mdt->mdt_max_ea_size;
5635 /* NB: Disregard the rule against updating
5636 * exp_connect_data.ocd_connect_flags in this case, since
5637 * tgt_client_new() needs to know if this is client supports
5638 * multiple modify RPCs, and it is safe to expose this flag before
5639 * connection processing completes. */
5640 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
5641 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
5642 spin_lock(&exp->exp_lock);
5643 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
5644 spin_unlock(&exp->exp_lock);
5647 if (OCD_HAS_FLAG(data, CKSUM)) {
5648 __u32 cksum_types = data->ocd_cksum_types;
5650 /* The client set in ocd_cksum_types the checksum types it
5651 * supports. We have to mask off the algorithms that we don't
5653 data->ocd_cksum_types &= cksum_types_supported_server();
5655 if (unlikely(data->ocd_cksum_types == 0)) {
5656 CERROR("%s: Connect with checksum support but no "
5657 "ocd_cksum_types is set\n",
5658 exp->exp_obd->obd_name);
5662 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
5663 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
5664 cksum_types, data->ocd_cksum_types);
5666 /* This client does not support OBD_CONNECT_CKSUM
5667 * fall back to CRC32 */
5668 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
5669 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
5670 exp->exp_obd->obd_name, obd_export_nid2str(exp));
5676 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
5677 struct mdt_thread_info *info,
5678 struct mdt_file_data *mfd)
5680 struct lu_context ses;
5684 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
5689 lu_context_enter(&ses);
5691 mdt_ucred(info)->uc_valid = UCRED_OLD;
5692 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
5694 lu_context_exit(&ses);
5695 lu_context_fini(&ses);
5701 static int mdt_export_cleanup(struct obd_export *exp)
5703 struct list_head closing_list;
5704 struct mdt_export_data *med = &exp->exp_mdt_data;
5705 struct obd_device *obd = exp->exp_obd;
5706 struct mdt_device *mdt;
5707 struct mdt_thread_info *info;
5709 struct mdt_file_data *mfd, *n;
5713 INIT_LIST_HEAD(&closing_list);
5714 spin_lock(&med->med_open_lock);
5715 while (!list_empty(&med->med_open_head)) {
5716 struct list_head *tmp = med->med_open_head.next;
5717 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
5719 /* Remove mfd handle so it can't be found again.
5720 * We are consuming the mfd_list reference here. */
5721 class_handle_unhash(&mfd->mfd_handle);
5722 list_move_tail(&mfd->mfd_list, &closing_list);
5724 spin_unlock(&med->med_open_lock);
5725 mdt = mdt_dev(obd->obd_lu_dev);
5726 LASSERT(mdt != NULL);
5728 rc = lu_env_init(&env, LCT_MD_THREAD);
5732 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5733 LASSERT(info != NULL);
5734 memset(info, 0, sizeof *info);
5735 info->mti_env = &env;
5736 info->mti_mdt = mdt;
5737 info->mti_exp = exp;
5739 if (!list_empty(&closing_list)) {
5740 struct md_attr *ma = &info->mti_attr;
5742 /* Close any open files (which may also cause orphan
5744 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5745 list_del_init(&mfd->mfd_list);
5746 ma->ma_need = ma->ma_valid = 0;
5748 /* This file is being closed due to an eviction, it
5749 * could have been modified and now dirty regarding to
5750 * HSM archive, check this!
5751 * The logic here is to mark a file dirty if there's a
5752 * chance it was dirtied before the client was evicted,
5753 * so that we don't have to wait for a release attempt
5754 * before finding out the file was actually dirty and
5755 * fail the release. Aggressively marking it dirty here
5756 * will cause the policy engine to attempt to
5757 * re-archive it; when rearchiving, we can compare the
5758 * current version to the HSM data_version and make the
5759 * archive request into a noop if it's not actually
5762 if (mfd->mfd_mode & FMODE_WRITE)
5763 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
5765 /* Don't unlink orphan on failover umount, LU-184 */
5766 if (exp->exp_flags & OBD_OPT_FAILOVER) {
5767 ma->ma_valid = MA_FLAGS;
5768 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
5770 mdt_mfd_close(info, mfd);
5773 info->mti_mdt = NULL;
5774 /* cleanup client slot early */
5775 /* Do not erase record for recoverable client. */
5776 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5777 tgt_client_del(&env, exp);
5783 static inline void mdt_enable_slc(struct mdt_device *mdt)
5785 if (mdt->mdt_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
5786 mdt->mdt_lut.lut_sync_lock_cancel = BLOCKING_SYNC_ON_CANCEL;
5789 static inline void mdt_disable_slc(struct mdt_device *mdt)
5791 if (mdt->mdt_lut.lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL)
5792 mdt->mdt_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
5795 static int mdt_obd_disconnect(struct obd_export *exp)
5802 class_export_get(exp);
5804 if (!(exp->exp_flags & OBD_OPT_FORCE))
5805 tgt_grant_sanity_check(exp->exp_obd, __func__);
5807 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
5808 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
5809 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
5811 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
5812 mdt_disable_slc(mdt);
5815 rc = server_disconnect_export(exp);
5817 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
5819 tgt_grant_discard(exp);
5821 rc = mdt_export_cleanup(exp);
5822 nodemap_del_member(exp);
5823 class_export_put(exp);
5827 /* mds_connect copy */
5828 static int mdt_obd_connect(const struct lu_env *env,
5829 struct obd_export **exp, struct obd_device *obd,
5830 struct obd_uuid *cluuid,
5831 struct obd_connect_data *data,
5834 struct obd_export *lexp;
5835 struct lustre_handle conn = { 0 };
5836 struct mdt_device *mdt;
5838 lnet_nid_t *client_nid = localdata;
5841 LASSERT(env != NULL);
5842 LASSERT(data != NULL);
5844 if (!exp || !obd || !cluuid)
5847 mdt = mdt_dev(obd->obd_lu_dev);
5849 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5850 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
5851 atomic_inc(&mdt->mdt_mds_mds_conns);
5852 mdt_enable_slc(mdt);
5856 * first, check whether the stack is ready to handle requests
5857 * XXX: probably not very appropriate method is used now
5858 * at some point we should find a better one
5860 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
5861 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5862 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5863 rc = obd_get_info(env, mdt->mdt_child_exp,
5864 sizeof(KEY_OSP_CONNECTED),
5865 KEY_OSP_CONNECTED, NULL, NULL);
5868 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5871 rc = class_connect(&conn, obd, cluuid);
5875 lexp = class_conn2export(&conn);
5876 LASSERT(lexp != NULL);
5878 rc = nodemap_add_member(*client_nid, lexp);
5879 if (rc != 0 && rc != -EEXIST)
5882 rc = mdt_connect_internal(env, lexp, mdt, data, false);
5884 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5887 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5888 rc = tgt_client_new(env, lexp);
5890 mdt_export_stats_init(obd, lexp, localdata);
5894 class_disconnect(lexp);
5895 nodemap_del_member(lexp);
5899 /* Because we do not want this export to be evicted by pinger,
5900 * let's not add this export to the timed chain list. */
5901 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
5902 spin_lock(&lexp->exp_obd->obd_dev_lock);
5903 list_del_init(&lexp->exp_obd_chain_timed);
5904 spin_unlock(&lexp->exp_obd->obd_dev_lock);
5911 static int mdt_obd_reconnect(const struct lu_env *env,
5912 struct obd_export *exp, struct obd_device *obd,
5913 struct obd_uuid *cluuid,
5914 struct obd_connect_data *data,
5917 lnet_nid_t *client_nid = localdata;
5921 if (exp == NULL || obd == NULL || cluuid == NULL)
5924 rc = nodemap_add_member(*client_nid, exp);
5925 if (rc != 0 && rc != -EEXIST)
5928 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
5931 mdt_export_stats_init(obd, exp, localdata);
5933 nodemap_del_member(exp);
5938 /* FIXME: Can we avoid using these two interfaces? */
5939 static int mdt_init_export(struct obd_export *exp)
5941 struct mdt_export_data *med = &exp->exp_mdt_data;
5945 INIT_LIST_HEAD(&med->med_open_head);
5946 spin_lock_init(&med->med_open_lock);
5947 spin_lock(&exp->exp_lock);
5948 exp->exp_connecting = 1;
5949 spin_unlock(&exp->exp_lock);
5951 /* self-export doesn't need client data and ldlm initialization */
5952 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5953 &exp->exp_client_uuid)))
5956 rc = tgt_client_alloc(exp);
5960 rc = ldlm_init_export(exp);
5967 tgt_client_free(exp);
5969 CERROR("%s: Failed to initialize export: rc = %d\n",
5970 exp->exp_obd->obd_name, rc);
5974 static int mdt_destroy_export(struct obd_export *exp)
5978 target_destroy_export(exp);
5979 /* destroy can be called from failed obd_setup, so
5980 * checking uuid is safer than obd_self_export */
5981 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5982 &exp->exp_client_uuid)))
5985 ldlm_destroy_export(exp);
5986 tgt_client_free(exp);
5988 LASSERT(list_empty(&exp->exp_outstanding_replies));
5989 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5992 * discard grants once we're sure no more
5993 * interaction with the client is possible
5995 tgt_grant_discard(exp);
5996 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
5997 exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--;
5999 if (!(exp->exp_flags & OBD_OPT_FORCE))
6000 tgt_grant_sanity_check(exp->exp_obd, __func__);
6005 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
6006 struct linkea_data *ldata)
6010 LASSERT(ldata->ld_buf->lb_buf != NULL);
6012 if (!mdt_object_exists(mdt_obj))
6015 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6016 ldata->ld_buf, XATTR_NAME_LINK);
6017 if (rc == -ERANGE) {
6018 /* Buf was too small, figure out what we need. */
6019 lu_buf_free(ldata->ld_buf);
6020 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6021 ldata->ld_buf, XATTR_NAME_LINK);
6024 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
6025 if (ldata->ld_buf->lb_buf == NULL)
6027 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6028 ldata->ld_buf, XATTR_NAME_LINK);
6033 return linkea_init_with_rec(ldata);
6037 * Given an MDT object, try to look up the full path to the object.
6038 * Part of the MDT layer implementation of lfs fid2path.
6040 * \param[in] info Per-thread common data shared by MDT level handlers.
6041 * \param[in] obj Object to do path lookup of
6042 * \param[in,out] fp User-provided struct to store path information
6043 * \param[in] root_fid Root FID of current path should reach
6045 * \retval 0 Lookup successful, path information stored in fp
6046 * \retval -EAGAIN Lookup failed, usually because object is being moved
6047 * \retval negative errno if there was a problem
6049 static int mdt_path_current(struct mdt_thread_info *info,
6050 struct mdt_object *obj,
6051 struct getinfo_fid2path *fp,
6052 struct lu_fid *root_fid)
6054 struct mdt_device *mdt = info->mti_mdt;
6055 struct mdt_object *mdt_obj;
6056 struct link_ea_header *leh;
6057 struct link_ea_entry *lee;
6058 struct lu_name *tmpname = &info->mti_name;
6059 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
6060 struct lu_buf *buf = &info->mti_big_buf;
6063 struct linkea_data ldata = { NULL };
6068 /* temp buffer for path element, the buffer will be finally freed
6069 * in mdt_thread_info_fini */
6070 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
6071 if (buf->lb_buf == NULL)
6075 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
6078 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
6080 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
6081 struct lu_buf lmv_buf;
6083 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
6084 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
6085 GOTO(out, rc = -ENOENT);
6087 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
6089 mdt_object_get(info->mti_env, mdt_obj);
6091 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
6092 if (IS_ERR(mdt_obj))
6093 GOTO(out, rc = PTR_ERR(mdt_obj));
6096 if (!mdt_object_exists(mdt_obj)) {
6097 mdt_object_put(info->mti_env, mdt_obj);
6098 GOTO(out, rc = -ENOENT);
6101 if (mdt_object_remote(mdt_obj)) {
6102 mdt_object_put(info->mti_env, mdt_obj);
6103 GOTO(remote_out, rc = -EREMOTE);
6106 rc = mdt_links_read(info, mdt_obj, &ldata);
6108 mdt_object_put(info->mti_env, mdt_obj);
6113 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
6114 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
6115 /* If set, use link #linkno for path lookup, otherwise use
6116 link #0. Only do this for the final path element. */
6117 if (first && fp->gf_linkno < leh->leh_reccount) {
6119 for (count = 0; count < fp->gf_linkno; count++) {
6120 lee = (struct link_ea_entry *)
6121 ((char *)lee + reclen);
6122 linkea_entry_unpack(lee, &reclen, tmpname,
6125 if (fp->gf_linkno < leh->leh_reccount - 1)
6126 /* indicate to user there are more links */
6130 lmv_buf.lb_buf = info->mti_xattr_buf;
6131 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
6132 /* Check if it is slave stripes */
6133 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6134 &lmv_buf, XATTR_NAME_LMV);
6135 mdt_object_put(info->mti_env, mdt_obj);
6137 union lmv_mds_md *lmm = lmv_buf.lb_buf;
6139 /* For slave stripes, get its master */
6140 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
6141 fp->gf_fid = *tmpfid;
6144 } else if (rc < 0 && rc != -ENODATA) {
6150 /* Pack the name in the end of the buffer */
6151 ptr -= tmpname->ln_namelen;
6152 if (ptr - 1 <= fp->gf_u.gf_path)
6153 GOTO(out, rc = -EOVERFLOW);
6154 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
6157 /* keep the last resolved fid to the client, so the
6158 * client will build the left path on another MDT for
6160 fp->gf_fid = *tmpfid;
6166 ptr++; /* skip leading / */
6167 memmove(fp->gf_u.gf_path, ptr,
6168 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
6175 * Given an MDT object, use mdt_path_current to get the path.
6176 * Essentially a wrapper to retry mdt_path_current a set number of times
6177 * if -EAGAIN is returned (usually because an object is being moved).
6179 * Part of the MDT layer implementation of lfs fid2path.
6181 * \param[in] info Per-thread common data shared by mdt level handlers.
6182 * \param[in] obj Object to do path lookup of
6183 * \param[in,out] fp User-provided struct for arguments and to store path
6186 * \retval 0 Lookup successful, path information stored in fp
6187 * \retval negative errno if there was a problem
6189 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
6190 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
6192 struct mdt_device *mdt = info->mti_mdt;
6197 if (fp->gf_pathlen < 3)
6200 if (root_fid == NULL)
6201 root_fid = &mdt->mdt_md_root_fid;
6203 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
6204 fp->gf_u.gf_path[0] = '\0';
6208 /* Retry multiple times in case file is being moved */
6209 while (tries-- && rc == -EAGAIN)
6210 rc = mdt_path_current(info, obj, fp, root_fid);
6216 * Get the full path of the provided FID, as of changelog record recno.
6218 * This checks sanity and looks up object for user provided FID
6219 * before calling the actual path lookup code.
6221 * Part of the MDT layer implementation of lfs fid2path.
6223 * \param[in] info Per-thread common data shared by mdt level handlers.
6224 * \param[in,out] fp User-provided struct for arguments and to store path
6227 * \retval 0 Lookup successful, path information and recno stored in fp
6228 * \retval -ENOENT, object does not exist
6229 * \retval negative errno if there was a problem
6231 static int mdt_fid2path(struct mdt_thread_info *info,
6232 struct lu_fid *root_fid,
6233 struct getinfo_fid2path *fp)
6235 struct mdt_device *mdt = info->mti_mdt;
6236 struct mdt_object *obj;
6240 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
6241 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
6243 if (!fid_is_sane(&fp->gf_fid))
6246 if (!fid_is_namespace_visible(&fp->gf_fid)) {
6247 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
6248 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
6249 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
6253 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
6256 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
6257 PFID(&fp->gf_fid), rc);
6261 if (mdt_object_remote(obj))
6263 else if (!mdt_object_exists(obj))
6269 mdt_object_put(info->mti_env, obj);
6270 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
6271 PFID(&fp->gf_fid), rc);
6275 rc = mdt_path(info, obj, fp, root_fid);
6277 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
6278 PFID(&fp->gf_fid), fp->gf_u.gf_path,
6279 fp->gf_recno, fp->gf_linkno);
6281 mdt_object_put(info->mti_env, obj);
6286 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
6287 void *val, int vallen)
6289 struct getinfo_fid2path *fpout, *fpin;
6290 struct lu_fid *root_fid = NULL;
6293 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
6296 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6297 lustre_swab_fid2path(fpin);
6299 memcpy(fpout, fpin, sizeof(*fpin));
6300 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
6303 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
6304 sizeof(struct lu_fid)) {
6305 /* client sent its root FID, which is normally fileset FID */
6306 root_fid = fpin->gf_u.gf_root_fid;
6307 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6308 lustre_swab_lu_fid(root_fid);
6310 if (root_fid != NULL && !fid_is_sane(root_fid))
6314 rc = mdt_fid2path(info, root_fid, fpout);
6318 int mdt_get_info(struct tgt_session_info *tsi)
6328 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
6330 CDEBUG(D_IOCTL, "No GETINFO key\n");
6331 RETURN(err_serious(-EFAULT));
6333 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
6336 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
6337 if (vallen == NULL) {
6338 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
6339 tgt_name(tsi->tsi_tgt));
6340 RETURN(err_serious(-EFAULT));
6343 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
6345 rc = req_capsule_server_pack(tsi->tsi_pill);
6347 RETURN(err_serious(rc));
6349 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
6350 if (valout == NULL) {
6351 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
6352 tgt_name(tsi->tsi_tgt));
6353 RETURN(err_serious(-EFAULT));
6356 if (KEY_IS(KEY_FID2PATH)) {
6357 struct mdt_thread_info *info = tsi2mdt_info(tsi);
6359 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
6360 mdt_thread_info_fini(info);
6367 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
6369 struct obd_ioctl_data *data = karg;
6372 struct mdt_object *obj;
6373 struct mdt_lock_handle *lh;
6377 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
6378 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
6381 fid = (struct lu_fid *)data->ioc_inlbuf1;
6383 if (!fid_is_sane(fid))
6386 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
6388 lh = &mti->mti_lh[MDT_LH_PARENT];
6389 mdt_lock_reg_init(lh, LCK_CR);
6391 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
6393 RETURN(PTR_ERR(obj));
6395 if (mdt_object_remote(obj)) {
6398 * before calling version get the correct MDS should be
6399 * fid, this is error to find remote object here
6401 CERROR("nonlocal object "DFID"\n", PFID(fid));
6402 } else if (!mdt_object_exists(obj)) {
6403 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
6406 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
6407 *(__u64 *)data->ioc_inlbuf2 = version;
6410 mdt_object_unlock_put(mti, obj, lh, 1);
6414 /* ioctls on obd dev */
6415 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
6416 void *karg, void __user *uarg)
6419 struct obd_device *obd = exp->exp_obd;
6420 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
6421 struct dt_device *dt = mdt->mdt_bottom;
6425 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
6426 rc = lu_env_init(&env, LCT_MD_THREAD);
6432 rc = mdt_device_sync(&env, mdt);
6434 case OBD_IOC_SET_READONLY:
6435 rc = dt_sync(&env, dt);
6437 rc = dt_ro(&env, dt);
6439 case OBD_IOC_ABORT_RECOVERY:
6440 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
6441 obd->obd_abort_recovery = 1;
6442 target_stop_recovery_thread(obd);
6445 case OBD_IOC_CHANGELOG_REG:
6446 case OBD_IOC_CHANGELOG_DEREG:
6447 case OBD_IOC_CHANGELOG_CLEAR:
6448 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env,
6452 case OBD_IOC_START_LFSCK: {
6453 struct md_device *next = mdt->mdt_child;
6454 struct obd_ioctl_data *data = karg;
6455 struct lfsck_start_param lsp;
6457 if (unlikely(data == NULL)) {
6462 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
6463 lsp.lsp_index_valid = 0;
6464 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
6467 case OBD_IOC_STOP_LFSCK: {
6468 struct md_device *next = mdt->mdt_child;
6469 struct obd_ioctl_data *data = karg;
6470 struct lfsck_stop stop;
6472 stop.ls_status = LS_STOPPED;
6473 /* Old lfsck utils may pass NULL @stop. */
6474 if (data->ioc_inlbuf1 == NULL)
6478 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
6480 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
6483 case OBD_IOC_QUERY_LFSCK: {
6484 struct md_device *next = mdt->mdt_child;
6485 struct obd_ioctl_data *data = karg;
6487 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
6491 case OBD_IOC_GET_OBJ_VERSION: {
6492 struct mdt_thread_info *mti;
6493 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6494 memset(mti, 0, sizeof *mti);
6495 mti->mti_env = &env;
6499 rc = mdt_ioc_version_get(mti, karg);
6502 case OBD_IOC_CATLOGLIST: {
6503 struct mdt_thread_info *mti;
6505 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6506 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
6507 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
6508 &mti->mti_tmp_fid1);
6513 CERROR("%s: Not supported cmd = %d, rc = %d\n",
6514 mdt_obd_name(mdt), cmd, rc);
6521 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
6523 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
6527 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
6528 struct lfsck_start_param lsp;
6530 lsp.lsp_start = NULL;
6531 lsp.lsp_index_valid = 0;
6532 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
6533 OBD_IOC_START_LFSCK,
6535 if (rc != 0 && rc != -EALREADY)
6536 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
6537 mdt_obd_name(mdt), rc);
6540 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
6544 static int mdt_obd_postrecov(struct obd_device *obd)
6549 rc = lu_env_init(&env, LCT_MD_THREAD);
6552 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
6557 static struct obd_ops mdt_obd_device_ops = {
6558 .o_owner = THIS_MODULE,
6559 .o_set_info_async = mdt_obd_set_info_async,
6560 .o_connect = mdt_obd_connect,
6561 .o_reconnect = mdt_obd_reconnect,
6562 .o_disconnect = mdt_obd_disconnect,
6563 .o_init_export = mdt_init_export,
6564 .o_destroy_export = mdt_destroy_export,
6565 .o_iocontrol = mdt_iocontrol,
6566 .o_postrecov = mdt_obd_postrecov,
6567 /* Data-on-MDT IO methods */
6568 .o_preprw = mdt_obd_preprw,
6569 .o_commitrw = mdt_obd_commitrw,
6572 static struct lu_device* mdt_device_fini(const struct lu_env *env,
6573 struct lu_device *d)
6575 struct mdt_device *m = mdt_dev(d);
6582 static struct lu_device *mdt_device_free(const struct lu_env *env,
6583 struct lu_device *d)
6585 struct mdt_device *m = mdt_dev(d);
6588 lu_device_fini(&m->mdt_lu_dev);
6594 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
6595 struct lu_device_type *t,
6596 struct lustre_cfg *cfg)
6598 struct lu_device *l;
6599 struct mdt_device *m;
6606 rc = mdt_init0(env, m, t, cfg);
6608 mdt_device_free(env, l);
6613 l = ERR_PTR(-ENOMEM);
6617 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
6618 LU_KEY_INIT(mdt, struct mdt_thread_info);
6620 static void mdt_key_fini(const struct lu_context *ctx,
6621 struct lu_context_key *key, void* data)
6623 struct mdt_thread_info *info = data;
6625 if (info->mti_big_lmm) {
6626 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
6627 info->mti_big_lmm = NULL;
6628 info->mti_big_lmmsize = 0;
6631 if (info->mti_big_acl) {
6632 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
6633 info->mti_big_acl = NULL;
6634 info->mti_big_aclsize = 0;
6640 /* context key: mdt_thread_key */
6641 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
6643 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
6645 return lu_ucred(info->mti_env);
6648 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
6650 return lu_ucred_check(info->mti_env);
6654 * Enable/disable COS (Commit On Sharing).
6656 * Set/Clear the COS flag in mdt options.
6658 * \param mdt mdt device
6659 * \param val 0 disables COS, other values enable COS
6661 void mdt_enable_cos(struct mdt_device *mdt, int val)
6666 mdt->mdt_opts.mo_cos = !!val;
6667 rc = lu_env_init(&env, LCT_LOCAL);
6668 if (unlikely(rc != 0)) {
6669 CWARN("%s: lu_env initialization failed, cannot "
6670 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
6673 mdt_device_sync(&env, mdt);
6678 * Check COS (Commit On Sharing) status.
6680 * Return COS flag status.
6682 * \param mdt mdt device
6684 int mdt_cos_is_enabled(struct mdt_device *mdt)
6686 return mdt->mdt_opts.mo_cos != 0;
6689 static struct lu_device_type_operations mdt_device_type_ops = {
6690 .ldto_device_alloc = mdt_device_alloc,
6691 .ldto_device_free = mdt_device_free,
6692 .ldto_device_fini = mdt_device_fini
6695 static struct lu_device_type mdt_device_type = {
6696 .ldt_tags = LU_DEVICE_MD,
6697 .ldt_name = LUSTRE_MDT_NAME,
6698 .ldt_ops = &mdt_device_type_ops,
6699 .ldt_ctx_tags = LCT_MD_THREAD
6702 static int __init mdt_init(void)
6706 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
6707 FID_NOBRACE_LEN + 1);
6708 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
6710 rc = lu_kmem_init(mdt_caches);
6714 rc = mds_mod_init();
6718 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
6719 LUSTRE_MDT_NAME, &mdt_device_type);
6724 lu_kmem_fini(mdt_caches);
6731 static void __exit mdt_exit(void)
6733 class_unregister_type(LUSTRE_MDT_NAME);
6735 lu_kmem_fini(mdt_caches);
6738 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
6739 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
6740 MODULE_VERSION(LUSTRE_VERSION_STRING);
6741 MODULE_LICENSE("GPL");
6743 module_init(mdt_init);
6744 module_exit(mdt_exit);