4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_handler.c
38 * Lustre Metadata Target (mdt) request handler
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
43 * Author: Mike Shaver <shaver@clusterfs.com>
44 * Author: Nikita Danilov <nikita@clusterfs.com>
45 * Author: Huang Hua <huanghua@clusterfs.com>
46 * Author: Yury Umanets <umka@clusterfs.com>
49 #define DEBUG_SUBSYSTEM S_MDS
51 #include <linux/module.h>
52 #include <linux/pagemap.h>
54 #include <dt_object.h>
55 #include <lustre_acl.h>
56 #include <lustre_export.h>
57 #include <lustre_ioctl.h>
58 #include <lustre_lfsck.h>
59 #include <lustre_log.h>
60 #include <lustre_net.h>
61 #include <lustre_nodemap.h>
62 #include <lustre_mds.h>
63 #include <lustre_param.h>
64 #include <lustre_quota.h>
65 #include <lustre_swab.h>
67 #include <obd_support.h>
69 #include <llog_swab.h>
71 #include "mdt_internal.h"
74 static unsigned int max_mod_rpcs_per_client = 8;
75 module_param(max_mod_rpcs_per_client, uint, 0644);
76 MODULE_PARM_DESC(max_mod_rpcs_per_client, "maximum number of modify RPCs in flight allowed per client");
78 mdl_mode_t mdt_mdl_lock_modes[] = {
79 [LCK_MINMODE] = MDL_MINMODE,
86 [LCK_GROUP] = MDL_GROUP
89 enum ldlm_mode mdt_dlm_lock_modes[] = {
90 [MDL_MINMODE] = LCK_MINMODE,
97 [MDL_GROUP] = LCK_GROUP
100 static struct mdt_device *mdt_dev(struct lu_device *d);
101 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
103 static const struct lu_object_operations mdt_obj_ops;
105 /* Slab for MDT object allocation */
106 static struct kmem_cache *mdt_object_kmem;
108 /* For HSM restore handles */
109 struct kmem_cache *mdt_hsm_cdt_kmem;
111 /* For HSM request handles */
112 struct kmem_cache *mdt_hsm_car_kmem;
114 static struct lu_kmem_descr mdt_caches[] = {
116 .ckd_cache = &mdt_object_kmem,
117 .ckd_name = "mdt_obj",
118 .ckd_size = sizeof(struct mdt_object)
121 .ckd_cache = &mdt_hsm_cdt_kmem,
122 .ckd_name = "mdt_cdt_restore_handle",
123 .ckd_size = sizeof(struct cdt_restore_handle)
126 .ckd_cache = &mdt_hsm_car_kmem,
127 .ckd_name = "mdt_cdt_agent_req",
128 .ckd_size = sizeof(struct cdt_agent_req)
135 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
139 return rep->lock_policy_res1 & op_flag;
142 void mdt_clear_disposition(struct mdt_thread_info *info,
143 struct ldlm_reply *rep, __u64 op_flag)
146 info->mti_opdata &= ~op_flag;
147 tgt_opdata_clear(info->mti_env, op_flag);
150 rep->lock_policy_res1 &= ~op_flag;
153 void mdt_set_disposition(struct mdt_thread_info *info,
154 struct ldlm_reply *rep, __u64 op_flag)
157 info->mti_opdata |= op_flag;
158 tgt_opdata_set(info->mti_env, op_flag);
161 rep->lock_policy_res1 |= op_flag;
164 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
166 lh->mlh_pdo_hash = 0;
167 lh->mlh_reg_mode = lm;
168 lh->mlh_rreg_mode = lm;
169 lh->mlh_type = MDT_REG_LOCK;
172 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
173 const struct lu_name *lname)
175 lh->mlh_reg_mode = lock_mode;
176 lh->mlh_pdo_mode = LCK_MINMODE;
177 lh->mlh_rreg_mode = lock_mode;
178 lh->mlh_type = MDT_PDO_LOCK;
180 if (lu_name_is_valid(lname)) {
181 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
183 /* XXX Workaround for LU-2856
185 * Zero is a valid return value of full_name_hash, but
186 * several users of mlh_pdo_hash assume a non-zero
187 * hash value. We therefore map zero onto an
188 * arbitrary, but consistent value (1) to avoid
189 * problems further down the road. */
190 if (unlikely(lh->mlh_pdo_hash == 0))
191 lh->mlh_pdo_hash = 1;
193 lh->mlh_pdo_hash = 0;
197 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
198 struct mdt_lock_handle *lh)
204 * Any dir access needs couple of locks:
206 * 1) on part of dir we gonna take lookup/modify;
208 * 2) on whole dir to protect it from concurrent splitting and/or to
209 * flush client's cache for readdir().
211 * so, for a given mode and object this routine decides what lock mode
212 * to use for lock #2:
214 * 1) if caller's gonna lookup in dir then we need to protect dir from
215 * being splitted only - LCK_CR
217 * 2) if caller's gonna modify dir then we need to protect dir from
218 * being splitted and to flush cache - LCK_CW
220 * 3) if caller's gonna modify dir and that dir seems ready for
221 * splitting then we need to protect it from any type of access
222 * (lookup/modify/split) - LCK_EX --bzzz
225 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
226 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
229 * Ask underlaying level its opinion about preferable PDO lock mode
230 * having access type passed as regular lock mode:
232 * - MDL_MINMODE means that lower layer does not want to specify lock
235 * - MDL_NL means that no PDO lock should be taken. This is used in some
236 * cases. Say, for non-splittable directories no need to use PDO locks
239 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
240 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
242 if (mode != MDL_MINMODE) {
243 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
246 * Lower layer does not want to specify locking mode. We do it
247 * our selves. No special protection is needed, just flush
248 * client's cache on modification and allow concurrent
251 switch (lh->mlh_reg_mode) {
253 lh->mlh_pdo_mode = LCK_EX;
256 lh->mlh_pdo_mode = LCK_CR;
259 lh->mlh_pdo_mode = LCK_CW;
262 CERROR("Not expected lock type (0x%x)\n",
263 (int)lh->mlh_reg_mode);
268 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
272 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
275 struct mdt_device *mdt = info->mti_mdt;
276 struct lu_name *lname = &info->mti_name;
278 struct mdt_object *parent;
282 LASSERT(!info->mti_cross_ref);
284 OBD_ALLOC(name, NAME_MAX + 1);
287 lname->ln_name = name;
290 * We may want to allow this to mount a completely separate
291 * fileset from the MDT in the future, but keeping it to
292 * ROOT/ only for now avoid potential security issues.
294 *fid = mdt->mdt_md_root_fid;
296 while (rc == 0 && fileset != NULL && *fileset != '\0') {
297 const char *s1 = fileset;
303 while (*s2 != '/' && *s2 != '\0')
311 lname->ln_namelen = s2 - s1;
312 if (lname->ln_namelen > NAME_MAX) {
317 /* reject .. as a path component */
318 if (lname->ln_namelen == 2 &&
319 strncmp(s1, "..", 2) == 0) {
324 strncpy(name, s1, lname->ln_namelen);
325 name[lname->ln_namelen] = '\0';
327 parent = mdt_object_find(info->mti_env, mdt, fid);
328 if (IS_ERR(parent)) {
329 rc = PTR_ERR(parent);
332 /* Only got the fid of this obj by name */
334 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
335 fid, &info->mti_spec);
336 mdt_object_put(info->mti_env, parent);
339 parent = mdt_object_find(info->mti_env, mdt, fid);
341 rc = PTR_ERR(parent);
343 mode = lu_object_attr(&parent->mot_obj);
344 mdt_object_put(info->mti_env, parent);
350 OBD_FREE(name, NAME_MAX + 1);
355 static int mdt_get_root(struct tgt_session_info *tsi)
357 struct mdt_thread_info *info = tsi2mdt_info(tsi);
358 struct mdt_device *mdt = info->mti_mdt;
359 struct mdt_body *repbody;
360 char *fileset = NULL, *buffer = NULL;
362 struct obd_export *exp = info->mti_exp;
363 char *nodemap_fileset;
367 rc = mdt_check_ucred(info);
369 GOTO(out, rc = err_serious(rc));
371 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
372 GOTO(out, rc = err_serious(-ENOMEM));
374 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
375 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
376 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
378 GOTO(out, rc = err_serious(-EFAULT));
381 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
382 if (nodemap_fileset && nodemap_fileset[0]) {
384 /* consider fileset from client as a sub-fileset
385 * of the nodemap one */
386 OBD_ALLOC(buffer, PATH_MAX + 1);
388 GOTO(out, rc = err_serious(-ENOMEM));
389 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
390 nodemap_fileset, fileset) >= PATH_MAX + 1)
391 GOTO(out, rc = err_serious(-EINVAL));
394 /* enforce fileset as specified in the nodemap */
395 fileset = nodemap_fileset;
400 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
402 GOTO(out, rc = err_serious(rc));
404 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
406 repbody->mbo_valid |= OBD_MD_FLID;
410 mdt_thread_info_fini(info);
412 OBD_FREE(buffer, PATH_MAX+1);
416 static int mdt_statfs(struct tgt_session_info *tsi)
418 struct ptlrpc_request *req = tgt_ses_req(tsi);
419 struct mdt_thread_info *info = tsi2mdt_info(tsi);
420 struct md_device *next = info->mti_mdt->mdt_child;
421 struct ptlrpc_service_part *svcpt;
422 struct obd_statfs *osfs;
427 svcpt = req->rq_rqbd->rqbd_svcpt;
429 /* This will trigger a watchdog timeout */
430 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
431 (MDT_SERVICE_WATCHDOG_FACTOR *
432 at_get(&svcpt->scp_at_estimate)) + 1);
434 rc = mdt_check_ucred(info);
436 GOTO(out, rc = err_serious(rc));
438 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
439 GOTO(out, rc = err_serious(-ENOMEM));
441 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
443 GOTO(out, rc = -EPROTO);
445 /** statfs information are cached in the mdt_device */
446 if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
447 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
448 /** statfs data is too old, get up-to-date one */
449 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
452 spin_lock(&info->mti_mdt->mdt_lock);
453 info->mti_mdt->mdt_osfs = *osfs;
454 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
455 spin_unlock(&info->mti_mdt->mdt_lock);
457 /** use cached statfs data */
458 spin_lock(&info->mti_mdt->mdt_lock);
459 *osfs = info->mti_mdt->mdt_osfs;
460 spin_unlock(&info->mti_mdt->mdt_lock);
464 mdt_counter_incr(req, LPROC_MDT_STATFS);
466 mdt_thread_info_fini(info);
470 #ifdef CONFIG_FS_POSIX_ACL
472 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
474 * \param info thread info object
475 * \param repbody reply to pack ACLs into
476 * \param o mdt object of file to examine
477 * \param nodemap nodemap of client to reply to
479 * \retval -errno error getting or parsing ACL from disk
481 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
482 struct mdt_object *o, struct lu_nodemap *nodemap)
484 const struct lu_env *env = info->mti_env;
485 struct md_object *next = mdt_object_child(o);
486 struct lu_buf *buf = &info->mti_buf;
489 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
490 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
492 if (buf->lb_len == 0)
495 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
497 if (rc == -ENODATA) {
498 repbody->mbo_aclsize = 0;
499 repbody->mbo_valid |= OBD_MD_FLACL;
501 } else if (rc == -EOPNOTSUPP) {
504 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
505 mdt_obd_name(info->mti_mdt),
506 PFID(mdt_object_fid(o)), rc);
509 rc = nodemap_map_acl(nodemap, buf->lb_buf,
510 rc, NODEMAP_FS_TO_CLIENT);
511 /* if all ACLs mapped out, rc is still >= 0 */
513 CERROR("%s: nodemap_map_acl unable to parse "DFID
514 " ACL: rc = %d\n", mdt_obd_name(info->mti_mdt),
515 PFID(mdt_object_fid(o)), rc);
517 repbody->mbo_aclsize = rc;
518 repbody->mbo_valid |= OBD_MD_FLACL;
526 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
527 const struct lu_attr *attr, const struct lu_fid *fid)
529 struct md_attr *ma = &info->mti_attr;
530 struct obd_export *exp = info->mti_exp;
531 struct lu_nodemap *nodemap = NULL;
533 LASSERT(ma->ma_valid & MA_INODE);
535 if (attr->la_valid & LA_ATIME) {
536 b->mbo_atime = attr->la_atime;
537 b->mbo_valid |= OBD_MD_FLATIME;
539 if (attr->la_valid & LA_MTIME) {
540 b->mbo_mtime = attr->la_mtime;
541 b->mbo_valid |= OBD_MD_FLMTIME;
543 if (attr->la_valid & LA_CTIME) {
544 b->mbo_ctime = attr->la_ctime;
545 b->mbo_valid |= OBD_MD_FLCTIME;
547 if (attr->la_valid & LA_FLAGS) {
548 b->mbo_flags = attr->la_flags;
549 b->mbo_valid |= OBD_MD_FLFLAGS;
551 if (attr->la_valid & LA_NLINK) {
552 b->mbo_nlink = attr->la_nlink;
553 b->mbo_valid |= OBD_MD_FLNLINK;
555 if (attr->la_valid & (LA_UID|LA_GID)) {
556 nodemap = nodemap_get_from_exp(exp);
560 if (attr->la_valid & LA_UID) {
561 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
562 NODEMAP_FS_TO_CLIENT,
564 b->mbo_valid |= OBD_MD_FLUID;
566 if (attr->la_valid & LA_GID) {
567 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
568 NODEMAP_FS_TO_CLIENT,
570 b->mbo_valid |= OBD_MD_FLGID;
573 b->mbo_mode = attr->la_mode;
574 if (attr->la_valid & LA_MODE)
575 b->mbo_valid |= OBD_MD_FLMODE;
576 if (attr->la_valid & LA_TYPE)
577 b->mbo_valid |= OBD_MD_FLTYPE;
581 b->mbo_valid |= OBD_MD_FLID;
582 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid="LPX64"\n",
583 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
586 if (!(attr->la_valid & LA_TYPE))
589 b->mbo_rdev = attr->la_rdev;
590 b->mbo_size = attr->la_size;
591 b->mbo_blocks = attr->la_blocks;
593 if (!S_ISREG(attr->la_mode)) {
594 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
595 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
596 /* means no objects are allocated on osts. */
597 LASSERT(!(ma->ma_valid & MA_LOV));
598 /* just ignore blocks occupied by extend attributes on MDS */
600 /* if no object is allocated on osts, the size on mds is valid.
602 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
603 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
604 ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
605 /* A released file stores its size on MDS. */
606 /* But return 1 block for released file, unless tools like tar
607 * will consider it fully sparse. (LU-3864)
609 if (unlikely(b->mbo_size == 0))
613 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
616 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
617 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
618 PFID(fid), (unsigned long long)b->mbo_size);
621 if (!IS_ERR_OR_NULL(nodemap))
622 nodemap_putref(nodemap);
625 static inline int mdt_body_has_lov(const struct lu_attr *la,
626 const struct mdt_body *body)
628 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
629 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
632 void mdt_client_compatibility(struct mdt_thread_info *info)
634 struct mdt_body *body;
635 struct ptlrpc_request *req = mdt_info_req(info);
636 struct obd_export *exp = req->rq_export;
637 struct md_attr *ma = &info->mti_attr;
638 struct lu_attr *la = &ma->ma_attr;
641 if (exp_connect_layout(exp))
642 /* the client can deal with 16-bit lmm_stripe_count */
645 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
647 if (!mdt_body_has_lov(la, body))
650 /* now we have a reply with a lov for a client not compatible with the
651 * layout lock so we have to clean the layout generation number */
652 if (S_ISREG(la->la_mode))
653 ma->ma_lmm->lmm_layout_gen = 0;
657 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
658 struct mdt_object *o)
660 const struct lu_env *env = info->mti_env;
663 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
672 /* Is it a directory? Let's check for the LMV as well */
673 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
674 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
678 rc2 = mo_xattr_get(env, mdt_object_child(o),
680 XATTR_NAME_DEFAULT_LMV);
682 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
690 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
693 const struct lu_env *env = info->mti_env;
697 LASSERT(info->mti_big_lmm_used == 0);
698 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
702 /* big_lmm may need to be grown */
703 if (info->mti_big_lmmsize < rc) {
704 int size = size_roundup_power2(rc);
706 if (info->mti_big_lmmsize > 0) {
707 /* free old buffer */
708 LASSERT(info->mti_big_lmm);
709 OBD_FREE_LARGE(info->mti_big_lmm,
710 info->mti_big_lmmsize);
711 info->mti_big_lmm = NULL;
712 info->mti_big_lmmsize = 0;
715 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
716 if (info->mti_big_lmm == NULL)
718 info->mti_big_lmmsize = size;
720 LASSERT(info->mti_big_lmmsize >= rc);
722 info->mti_buf.lb_buf = info->mti_big_lmm;
723 info->mti_buf.lb_len = info->mti_big_lmmsize;
724 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
729 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
730 struct md_attr *ma, const char *name)
732 struct md_object *next = mdt_object_child(o);
733 struct lu_buf *buf = &info->mti_buf;
736 if (strcmp(name, XATTR_NAME_LOV) == 0) {
737 buf->lb_buf = ma->ma_lmm;
738 buf->lb_len = ma->ma_lmm_size;
739 LASSERT(!(ma->ma_valid & MA_LOV));
740 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
741 buf->lb_buf = ma->ma_lmv;
742 buf->lb_len = ma->ma_lmv_size;
743 LASSERT(!(ma->ma_valid & MA_LMV));
744 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
745 buf->lb_buf = ma->ma_lmv;
746 buf->lb_len = ma->ma_lmv_size;
747 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
752 rc = mo_xattr_get(info->mti_env, next, buf, name);
756 if (strcmp(name, XATTR_NAME_LOV) == 0) {
757 if (info->mti_big_lmm_used)
758 ma->ma_lmm = info->mti_big_lmm;
760 /* NOT return LOV EA with hole to old client. */
761 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
762 LOV_PATTERN_F_HOLE) &&
763 !(exp_connect_flags(info->mti_exp) &
764 OBD_CONNECT_LFSCK)) {
767 ma->ma_lmm_size = rc;
768 ma->ma_valid |= MA_LOV;
770 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
771 if (info->mti_big_lmm_used)
772 ma->ma_lmv = info->mti_big_lmm;
774 ma->ma_lmv_size = rc;
775 ma->ma_valid |= MA_LMV;
776 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
777 ma->ma_lmv_size = rc;
778 ma->ma_valid |= MA_LMV_DEF;
781 /* Update mdt_max_mdsize so all clients will be aware that */
782 if (info->mti_mdt->mdt_max_mdsize < rc)
783 info->mti_mdt->mdt_max_mdsize = rc;
786 } else if (rc == -ENODATA) {
789 } else if (rc == -ERANGE) {
790 /* Default LMV has fixed size, so it must be able to fit
791 * in the original buffer */
792 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
794 rc = mdt_big_xattr_get(info, o, name);
796 info->mti_big_lmm_used = 1;
804 static int mdt_attr_get_pfid(struct mdt_thread_info *info,
805 struct mdt_object *o, struct lu_fid *pfid)
807 struct lu_buf *buf = &info->mti_buf;
808 struct link_ea_header *leh;
809 struct link_ea_entry *lee;
813 buf->lb_buf = info->mti_big_lmm;
814 buf->lb_len = info->mti_big_lmmsize;
815 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
816 buf, XATTR_NAME_LINK);
817 /* ignore errors, MA_PFID won't be set and it is
818 * up to the caller to treat this as an error */
819 if (rc == -ERANGE || buf->lb_len == 0) {
820 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
821 buf->lb_buf = info->mti_big_lmm;
822 buf->lb_len = info->mti_big_lmmsize;
827 if (rc < sizeof(*leh)) {
828 CERROR("short LinkEA on "DFID": rc = %d\n",
829 PFID(mdt_object_fid(o)), rc);
833 leh = (struct link_ea_header *) buf->lb_buf;
834 lee = (struct link_ea_entry *)(leh + 1);
835 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
836 leh->leh_magic = LINK_EA_MAGIC;
837 leh->leh_reccount = __swab32(leh->leh_reccount);
838 leh->leh_len = __swab64(leh->leh_len);
840 if (leh->leh_magic != LINK_EA_MAGIC)
842 if (leh->leh_reccount == 0)
845 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
846 fid_be_to_cpu(pfid, pfid);
851 int mdt_attr_get_complex(struct mdt_thread_info *info,
852 struct mdt_object *o, struct md_attr *ma)
854 const struct lu_env *env = info->mti_env;
855 struct md_object *next = mdt_object_child(o);
856 struct lu_buf *buf = &info->mti_buf;
857 int need = ma->ma_need;
864 if (mdt_object_exists(o) == 0)
865 GOTO(out, rc = -ENOENT);
866 mode = lu_object_attr(&next->mo_lu);
868 if (need & MA_INODE) {
869 ma->ma_need = MA_INODE;
870 rc = mo_attr_get(env, next, ma);
873 ma->ma_valid |= MA_INODE;
876 if (need & MA_PFID) {
877 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
879 ma->ma_valid |= MA_PFID;
880 /* ignore this error, parent fid is not mandatory */
884 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
885 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
890 if (need & MA_LMV && S_ISDIR(mode)) {
891 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
896 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
897 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
902 if (need & MA_HSM && S_ISREG(mode)) {
903 buf->lb_buf = info->mti_xattr_buf;
904 buf->lb_len = sizeof(info->mti_xattr_buf);
905 CLASSERT(sizeof(struct hsm_attrs) <=
906 sizeof(info->mti_xattr_buf));
907 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
908 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
910 ma->ma_valid |= MA_HSM;
911 else if (rc2 < 0 && rc2 != -ENODATA)
915 #ifdef CONFIG_FS_POSIX_ACL
916 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
917 buf->lb_buf = ma->ma_acl;
918 buf->lb_len = ma->ma_acl_size;
919 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
921 ma->ma_acl_size = rc2;
922 ma->ma_valid |= MA_ACL_DEF;
923 } else if (rc2 == -ENODATA) {
932 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = "LPX64" ma_lmm=%p\n",
933 rc, ma->ma_valid, ma->ma_lmm);
937 static int mdt_getattr_internal(struct mdt_thread_info *info,
938 struct mdt_object *o, int ma_need)
940 struct md_object *next = mdt_object_child(o);
941 const struct mdt_body *reqbody = info->mti_body;
942 struct ptlrpc_request *req = mdt_info_req(info);
943 struct md_attr *ma = &info->mti_attr;
944 struct lu_attr *la = &ma->ma_attr;
945 struct req_capsule *pill = info->mti_pill;
946 const struct lu_env *env = info->mti_env;
947 struct mdt_body *repbody;
948 struct lu_buf *buffer = &info->mti_buf;
949 struct obd_export *exp = info->mti_exp;
954 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
955 RETURN(err_serious(-ENOMEM));
957 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
961 if (mdt_object_remote(o)) {
962 /* This object is located on remote node.*/
963 /* Return -ENOTSUPP for old client */
964 if (!mdt_is_dne_client(req->rq_export))
965 GOTO(out, rc = -ENOTSUPP);
967 repbody->mbo_fid1 = *mdt_object_fid(o);
968 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
972 if (reqbody->mbo_eadatasize > 0) {
973 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
974 if (buffer->lb_buf == NULL)
975 GOTO(out, rc = -EPROTO);
976 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
979 buffer->lb_buf = NULL;
981 ma_need &= ~(MA_LOV | MA_LMV);
982 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
983 mdt_obd_name(info->mti_mdt),
984 req->rq_export->exp_client_uuid.uuid);
987 /* If it is dir object and client require MEA, then we got MEA */
988 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
989 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
990 /* Assumption: MDT_MD size is enough for lmv size. */
991 ma->ma_lmv = buffer->lb_buf;
992 ma->ma_lmv_size = buffer->lb_len;
993 ma->ma_need = MA_INODE;
994 if (ma->ma_lmv_size > 0) {
995 if (reqbody->mbo_valid & OBD_MD_MEA)
996 ma->ma_need |= MA_LMV;
997 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
998 ma->ma_need |= MA_LMV_DEF;
1001 ma->ma_lmm = buffer->lb_buf;
1002 ma->ma_lmm_size = buffer->lb_len;
1003 ma->ma_need = MA_INODE | MA_HSM;
1004 if (ma->ma_lmm_size > 0)
1005 ma->ma_need |= MA_LOV;
1008 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1009 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1010 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1011 /* get default stripe info for this dir. */
1012 ma->ma_need |= MA_LOV_DEF;
1014 ma->ma_need |= ma_need;
1016 rc = mdt_attr_get_complex(info, o, ma);
1018 CDEBUG(rc == -ENOENT ? D_OTHER : D_ERROR,
1019 "%s: getattr error for "DFID": rc = %d\n",
1020 mdt_obd_name(info->mti_mdt),
1021 PFID(mdt_object_fid(o)), rc);
1025 /* if file is released, check if a restore is running */
1026 if (ma->ma_valid & MA_HSM) {
1027 repbody->mbo_valid |= OBD_MD_TSTATE;
1028 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1029 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1030 repbody->mbo_t_state = MS_RESTORE;
1033 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
1035 /* the Lustre protocol supposes to return default striping
1036 * on the user-visible root if explicitly requested */
1037 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
1038 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
1039 struct lu_fid rootfid;
1040 struct mdt_object *root;
1041 struct mdt_device *mdt = info->mti_mdt;
1043 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
1046 root = mdt_object_find(env, mdt, &rootfid);
1048 RETURN(PTR_ERR(root));
1049 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
1050 mdt_object_put(info->mti_env, root);
1052 CERROR("%s: getattr error for "DFID": rc = %d\n",
1053 mdt_obd_name(info->mti_mdt),
1054 PFID(mdt_object_fid(o)), rc);
1059 if (likely(ma->ma_valid & MA_INODE))
1060 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1064 if (mdt_body_has_lov(la, reqbody)) {
1065 if (ma->ma_valid & MA_LOV) {
1066 LASSERT(ma->ma_lmm_size);
1067 repbody->mbo_eadatasize = ma->ma_lmm_size;
1068 if (S_ISDIR(la->la_mode))
1069 repbody->mbo_valid |= OBD_MD_FLDIREA;
1071 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1072 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1074 if (ma->ma_valid & MA_LMV) {
1075 /* Return -ENOTSUPP for old client */
1076 if (!mdt_is_striped_client(req->rq_export))
1079 LASSERT(S_ISDIR(la->la_mode));
1080 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1081 repbody->mbo_eadatasize = ma->ma_lmv_size;
1082 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1084 if (ma->ma_valid & MA_LMV_DEF) {
1085 /* Return -ENOTSUPP for old client */
1086 if (!mdt_is_striped_client(req->rq_export))
1088 LASSERT(S_ISDIR(la->la_mode));
1089 repbody->mbo_eadatasize = ma->ma_lmv_size;
1090 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1091 OBD_MD_DEFAULT_MEA);
1093 } else if (S_ISLNK(la->la_mode) &&
1094 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1095 buffer->lb_buf = ma->ma_lmm;
1096 /* eadatasize from client includes NULL-terminator, so
1097 * there is no need to read it */
1098 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1099 rc = mo_readlink(env, next, buffer);
1100 if (unlikely(rc <= 0)) {
1101 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1102 mdt_obd_name(info->mti_mdt),
1103 PFID(mdt_object_fid(o)), rc);
1106 int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
1108 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1110 repbody->mbo_valid |= OBD_MD_LINKNAME;
1111 /* we need to report back size with NULL-terminator
1112 * because client expects that */
1113 repbody->mbo_eadatasize = rc + 1;
1114 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1115 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1116 "on "DFID ", expected %d\n",
1117 mdt_obd_name(info->mti_mdt),
1118 rc, PFID(mdt_object_fid(o)),
1119 reqbody->mbo_eadatasize - 1);
1120 /* NULL terminate */
1121 ((char *)ma->ma_lmm)[rc] = 0;
1123 /* If the total CDEBUG() size is larger than a page, it
1124 * will print a warning to the console, avoid this by
1125 * printing just the last part of the symlink. */
1126 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1127 print_limit < rc ? "..." : "", print_limit,
1128 (char *)ma->ma_lmm + rc - print_limit, rc);
1133 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1134 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1135 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1136 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1137 repbody->mbo_max_mdsize);
1140 #ifdef CONFIG_FS_POSIX_ACL
1141 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1142 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1143 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1144 if (IS_ERR(nodemap))
1145 RETURN(PTR_ERR(nodemap));
1147 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1148 nodemap_putref(nodemap);
1154 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1159 static int mdt_getattr(struct tgt_session_info *tsi)
1161 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1162 struct mdt_object *obj = info->mti_object;
1163 struct req_capsule *pill = info->mti_pill;
1164 struct mdt_body *reqbody;
1165 struct mdt_body *repbody;
1169 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1171 LASSERT(obj != NULL);
1172 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1174 /* Unlike intent case where we need to pre-fill out buffers early on
1175 * in intent policy for ldlm reasons, here we can have a much better
1176 * guess at EA size by just reading it from disk.
1177 * Exceptions are readdir and (missing) directory striping */
1179 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1180 /* No easy way to know how long is the symlink, but it cannot
1181 * be more than PATH_MAX, so we allocate +1 */
1184 /* A special case for fs ROOT: getattr there might fetch
1185 * default EA for entire fs, not just for this dir!
1187 } else if (lu_fid_eq(mdt_object_fid(obj),
1188 &info->mti_mdt->mdt_md_root_fid) &&
1189 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1190 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1192 /* Should the default strping be bigger, mdt_fix_reply
1193 * will reallocate */
1194 rc = DEF_REP_MD_SIZE;
1196 /* Read the actual EA size from disk */
1197 rc = mdt_attr_get_eabuf_size(info, obj);
1201 GOTO(out, rc = err_serious(rc));
1203 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1205 rc = req_capsule_server_pack(pill);
1206 if (unlikely(rc != 0))
1207 GOTO(out, rc = err_serious(rc));
1209 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1210 LASSERT(repbody != NULL);
1211 repbody->mbo_eadatasize = 0;
1212 repbody->mbo_aclsize = 0;
1214 rc = mdt_check_ucred(info);
1216 GOTO(out_shrink, rc);
1218 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1220 rc = mdt_getattr_internal(info, obj, 0);
1223 mdt_client_compatibility(info);
1224 rc2 = mdt_fix_reply(info);
1228 mdt_thread_info_fini(info);
1233 * Exchange MOF_LOV_CREATED flags between two objects after a
1234 * layout swap. No assumption is made on whether o1 or o2 have
1235 * created objects or not.
1237 * \param[in,out] o1 First swap layout object
1238 * \param[in,out] o2 Second swap layout object
1240 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1242 unsigned int o1_lov_created = o1->mot_lov_created;
1244 mutex_lock(&o1->mot_lov_mutex);
1245 mutex_lock(&o2->mot_lov_mutex);
1247 o1->mot_lov_created = o2->mot_lov_created;
1248 o2->mot_lov_created = o1_lov_created;
1250 mutex_unlock(&o2->mot_lov_mutex);
1251 mutex_unlock(&o1->mot_lov_mutex);
1254 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1256 struct mdt_thread_info *info;
1257 struct ptlrpc_request *req = tgt_ses_req(tsi);
1258 struct obd_export *exp = req->rq_export;
1259 struct mdt_object *o1, *o2, *o;
1260 struct mdt_lock_handle *lh1, *lh2;
1261 struct mdc_swap_layouts *msl;
1265 /* client does not support layout lock, so layout swaping
1267 * FIXME: there is a problem for old clients which don't support
1268 * layout lock yet. If those clients have already opened the file
1269 * they won't be notified at all so that old layout may still be
1270 * used to do IO. This can be fixed after file release is landed by
1271 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1272 if (!exp_connect_layout(exp))
1273 RETURN(-EOPNOTSUPP);
1275 info = tsi2mdt_info(tsi);
1277 if (info->mti_dlm_req != NULL)
1278 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1280 o1 = info->mti_object;
1281 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1282 &info->mti_body->mbo_fid2);
1284 GOTO(out, rc = PTR_ERR(o));
1286 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1287 GOTO(put, rc = -ENOENT);
1289 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1290 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1296 /* permission check. Make sure the calling process having permission
1297 * to write both files. */
1298 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1303 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1308 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1310 GOTO(put, rc = -EPROTO);
1312 lh1 = &info->mti_lh[MDT_LH_NEW];
1313 mdt_lock_reg_init(lh1, LCK_EX);
1314 lh2 = &info->mti_lh[MDT_LH_OLD];
1315 mdt_lock_reg_init(lh2, LCK_EX);
1317 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1318 MDS_INODELOCK_XATTR);
1322 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1323 MDS_INODELOCK_XATTR);
1327 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1328 mdt_object_child(o2), msl->msl_flags);
1332 mdt_swap_lov_flag(o1, o2);
1335 mdt_object_unlock(info, o2, lh2, rc);
1337 mdt_object_unlock(info, o1, lh1, rc);
1339 mdt_object_put(info->mti_env, o);
1341 mdt_thread_info_fini(info);
1345 static int mdt_raw_lookup(struct mdt_thread_info *info,
1346 struct mdt_object *parent,
1347 const struct lu_name *lname,
1348 struct ldlm_reply *ldlm_rep)
1350 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1354 LASSERT(!info->mti_cross_ref);
1356 /* Only got the fid of this obj by name */
1357 fid_zero(child_fid);
1358 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1359 lname, child_fid, &info->mti_spec);
1361 struct mdt_body *repbody;
1363 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1364 repbody->mbo_fid1 = *child_fid;
1365 repbody->mbo_valid = OBD_MD_FLID;
1366 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1367 } else if (rc == -ENOENT) {
1368 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1375 * UPDATE lock should be taken against parent, and be released before exit;
1376 * child_bits lock should be taken against child, and be returned back:
1377 * (1)normal request should release the child lock;
1378 * (2)intent request will grant the lock to client.
1380 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1381 struct mdt_lock_handle *lhc,
1383 struct ldlm_reply *ldlm_rep)
1385 struct ptlrpc_request *req = mdt_info_req(info);
1386 struct mdt_body *reqbody = NULL;
1387 struct mdt_object *parent = info->mti_object;
1388 struct mdt_object *child;
1389 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1390 struct lu_name *lname = NULL;
1391 struct mdt_lock_handle *lhp = NULL;
1392 struct ldlm_lock *lock;
1399 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1400 LASSERT(ergo(is_resent,
1401 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1406 if (info->mti_cross_ref) {
1407 /* Only getattr on the child. Parent is on another node. */
1408 mdt_set_disposition(info, ldlm_rep,
1409 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1411 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1413 PFID(mdt_object_fid(child)), ldlm_rep);
1415 rc = mdt_check_resent_lock(info, child, lhc);
1418 } else if (rc > 0) {
1419 mdt_lock_handle_init(lhc);
1420 mdt_lock_reg_init(lhc, LCK_PR);
1423 * Object's name is on another MDS, no lookup or layout
1424 * lock is needed here but update lock is.
1426 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1427 MDS_INODELOCK_LAYOUT);
1428 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1430 rc = mdt_object_lock(info, child, lhc, child_bits);
1435 /* Finally, we can get attr for child. */
1436 if (!mdt_object_exists(child)) {
1437 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1439 "remote object doesn't exist.");
1440 mdt_object_unlock(info, child, lhc, 1);
1444 rc = mdt_getattr_internal(info, child, 0);
1445 if (unlikely(rc != 0))
1446 mdt_object_unlock(info, child, lhc, 1);
1451 lname = &info->mti_name;
1452 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1454 if (lu_name_is_valid(lname)) {
1455 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1456 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1457 PNAME(lname), ldlm_rep);
1459 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1460 if (unlikely(reqbody == NULL))
1461 RETURN(err_serious(-EPROTO));
1463 *child_fid = reqbody->mbo_fid2;
1465 if (unlikely(!fid_is_sane(child_fid)))
1466 RETURN(err_serious(-EINVAL));
1468 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1470 PFID(mdt_object_fid(parent)),
1471 PFID(&reqbody->mbo_fid2), ldlm_rep);
1474 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1476 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1477 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1479 "Parent doesn't exist!");
1483 if (mdt_object_remote(parent)) {
1484 CERROR("%s: parent "DFID" is on remote target\n",
1485 mdt_obd_name(info->mti_mdt),
1486 PFID(mdt_object_fid(parent)));
1490 if (lu_name_is_valid(lname)) {
1491 /* Always allow to lookup ".." */
1492 if (unlikely(lname->ln_namelen == 2 &&
1493 lname->ln_name[0] == '.' &&
1494 lname->ln_name[1] == '.'))
1495 info->mti_spec.sp_permitted = 1;
1497 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1498 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1503 /* step 1: lock parent only if parent is a directory */
1504 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1505 lhp = &info->mti_lh[MDT_LH_PARENT];
1506 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1507 rc = mdt_object_lock(info, parent, lhp,
1508 MDS_INODELOCK_UPDATE);
1509 if (unlikely(rc != 0))
1513 /* step 2: lookup child's fid by name */
1514 fid_zero(child_fid);
1515 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1516 child_fid, &info->mti_spec);
1518 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1521 GOTO(out_parent, rc);
1524 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1527 *step 3: find the child object by fid & lock it.
1528 * regardless if it is local or remote.
1530 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1531 * set parent dir fid the same as child fid in getattr by fid case
1532 * we should not lu_object_find() the object again, could lead
1533 * to hung if there is a concurrent unlink destroyed the object.
1535 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1536 mdt_object_get(info->mti_env, parent);
1539 child = mdt_object_find(info->mti_env, info->mti_mdt,
1543 if (unlikely(IS_ERR(child)))
1544 GOTO(out_parent, rc = PTR_ERR(child));
1546 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1547 if (!mdt_object_exists(child)) {
1548 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1550 "Object doesn't exist!");
1551 GOTO(out_child, rc = -ENOENT);
1554 rc = mdt_check_resent_lock(info, child, lhc);
1556 GOTO(out_child, rc);
1557 } else if (rc > 0) {
1558 mdt_lock_handle_init(lhc);
1559 mdt_lock_reg_init(lhc, LCK_PR);
1562 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1563 mdt_object_exists(child) && !mdt_object_remote(child)) {
1564 struct md_attr *ma = &info->mti_attr;
1567 ma->ma_need = MA_INODE;
1568 rc = mdt_attr_get_complex(info, child, ma);
1569 if (unlikely(rc != 0))
1570 GOTO(out_child, rc);
1572 /* If the file has not been changed for some time, we
1573 * return not only a LOOKUP lock, but also an UPDATE
1574 * lock and this might save us RPC on later STAT. For
1575 * directories, it also let negative dentry cache start
1576 * working for this dir. */
1577 if (ma->ma_valid & MA_INODE &&
1578 ma->ma_attr.la_valid & LA_CTIME &&
1579 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1580 ma->ma_attr.la_ctime < cfs_time_current_sec())
1581 child_bits |= MDS_INODELOCK_UPDATE;
1584 /* layout lock must be granted in a best-effort way
1585 * for IT operations */
1586 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1587 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1588 exp_connect_layout(info->mti_exp) &&
1589 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1590 !mdt_object_remote(child) && ldlm_rep != NULL) {
1591 /* try to grant layout lock for regular file. */
1597 child_bits |= MDS_INODELOCK_LAYOUT;
1598 /* try layout lock, it may fail to be granted due to
1599 * contention at LOOKUP or UPDATE */
1600 if (!mdt_object_lock_try(info, child, lhc,
1602 child_bits &= ~MDS_INODELOCK_LAYOUT;
1603 LASSERT(child_bits != 0);
1604 rc = mdt_object_lock(info, child, lhc,
1610 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1611 * client will enqueue the lock to the remote MDT */
1612 if (mdt_object_remote(child))
1613 child_bits &= ~MDS_INODELOCK_UPDATE;
1614 rc = mdt_object_lock(info, child, lhc, child_bits);
1616 if (unlikely(rc != 0))
1617 GOTO(out_child, rc);
1620 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1622 /* finally, we can get attr for child. */
1623 rc = mdt_getattr_internal(info, child, ma_need);
1624 if (unlikely(rc != 0)) {
1625 mdt_object_unlock(info, child, lhc, 1);
1627 /* Debugging code. */
1628 LDLM_DEBUG(lock, "Returning lock to client");
1629 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1630 &lock->l_resource->lr_name),
1631 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1632 PLDLMRES(lock->l_resource),
1633 PFID(mdt_object_fid(child)));
1636 LDLM_LOCK_PUT(lock);
1640 mdt_object_put(info->mti_env, child);
1643 mdt_object_unlock(info, parent, lhp, 1);
1647 /* normal handler: should release the child lock */
1648 static int mdt_getattr_name(struct tgt_session_info *tsi)
1650 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1651 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1652 struct mdt_body *reqbody;
1653 struct mdt_body *repbody;
1657 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1658 LASSERT(reqbody != NULL);
1659 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1660 LASSERT(repbody != NULL);
1662 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1663 repbody->mbo_eadatasize = 0;
1664 repbody->mbo_aclsize = 0;
1666 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1668 GOTO(out_shrink, rc);
1670 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1671 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1672 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1673 lhc->mlh_reg_lh.cookie = 0;
1675 mdt_exit_ucred(info);
1678 mdt_client_compatibility(info);
1679 rc2 = mdt_fix_reply(info);
1682 mdt_thread_info_fini(info);
1686 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1687 void *karg, void __user *uarg);
1689 static int mdt_set_info(struct tgt_session_info *tsi)
1691 struct ptlrpc_request *req = tgt_ses_req(tsi);
1694 int keylen, vallen, rc = 0;
1698 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1700 DEBUG_REQ(D_HA, req, "no set_info key");
1701 RETURN(err_serious(-EFAULT));
1704 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1707 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1709 DEBUG_REQ(D_HA, req, "no set_info val");
1710 RETURN(err_serious(-EFAULT));
1713 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1716 /* Swab any part of val you need to here */
1717 if (KEY_IS(KEY_READ_ONLY)) {
1718 spin_lock(&req->rq_export->exp_lock);
1720 *exp_connect_flags_ptr(req->rq_export) |=
1723 *exp_connect_flags_ptr(req->rq_export) &=
1724 ~OBD_CONNECT_RDONLY;
1725 spin_unlock(&req->rq_export->exp_lock);
1726 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1727 struct changelog_setinfo *cs = val;
1729 if (vallen != sizeof(*cs)) {
1730 CERROR("%s: bad changelog_clear setinfo size %d\n",
1731 tgt_name(tsi->tsi_tgt), vallen);
1734 if (ptlrpc_req_need_swab(req)) {
1735 __swab64s(&cs->cs_recno);
1736 __swab32s(&cs->cs_id);
1739 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1741 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
1743 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
1750 static int mdt_readpage(struct tgt_session_info *tsi)
1752 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1753 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1754 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1755 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1756 struct mdt_body *repbody;
1762 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1763 RETURN(err_serious(-ENOMEM));
1765 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1766 if (repbody == NULL || reqbody == NULL)
1767 RETURN(err_serious(-EFAULT));
1770 * prepare @rdpg before calling lower layers and transfer itself. Here
1771 * reqbody->size contains offset of where to start to read and
1772 * reqbody->nlink contains number bytes to read.
1774 rdpg->rp_hash = reqbody->mbo_size;
1775 if (rdpg->rp_hash != reqbody->mbo_size) {
1776 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1777 rdpg->rp_hash, reqbody->mbo_size);
1781 rdpg->rp_attrs = reqbody->mbo_mode;
1782 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1783 rdpg->rp_attrs |= LUDA_64BITHASH;
1784 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1785 exp_max_brw_size(tsi->tsi_exp));
1786 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1788 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1789 if (rdpg->rp_pages == NULL)
1792 for (i = 0; i < rdpg->rp_npages; ++i) {
1793 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1794 if (rdpg->rp_pages[i] == NULL)
1795 GOTO(free_rdpg, rc = -ENOMEM);
1798 /* call lower layers to fill allocated pages with directory data */
1799 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1801 GOTO(free_rdpg, rc);
1803 /* send pages to client */
1804 rc = tgt_sendpage(tsi, rdpg, rc);
1809 for (i = 0; i < rdpg->rp_npages; i++)
1810 if (rdpg->rp_pages[i] != NULL)
1811 __free_page(rdpg->rp_pages[i]);
1812 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1814 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1820 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
1822 struct lu_ucred *uc = mdt_ucred_check(info);
1823 struct lu_attr *attr = &info->mti_attr.ma_attr;
1828 if (op != REINT_SETATTR) {
1829 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
1830 attr->la_uid = uc->uc_fsuid;
1831 /* for S_ISGID, inherit gid from his parent, such work will be
1832 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
1833 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
1834 attr->la_gid = uc->uc_fsgid;
1840 static int mdt_reint_internal(struct mdt_thread_info *info,
1841 struct mdt_lock_handle *lhc,
1844 struct req_capsule *pill = info->mti_pill;
1845 struct mdt_body *repbody;
1850 rc = mdt_reint_unpack(info, op);
1852 CERROR("Can't unpack reint, rc %d\n", rc);
1853 RETURN(err_serious(rc));
1856 /* for replay (no_create) lmm is not needed, client has it already */
1857 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1858 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1861 /* llog cookies are always 0, the field is kept for compatibility */
1862 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1863 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1865 rc = req_capsule_server_pack(pill);
1867 CERROR("Can't pack response, rc %d\n", rc);
1868 RETURN(err_serious(rc));
1871 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1872 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1874 repbody->mbo_eadatasize = 0;
1875 repbody->mbo_aclsize = 0;
1878 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1880 /* for replay no cookkie / lmm need, because client have this already */
1881 if (info->mti_spec.no_create)
1882 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1883 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1885 rc = mdt_init_ucred_reint(info);
1887 GOTO(out_shrink, rc);
1889 rc = mdt_fix_attr_ucred(info, op);
1891 GOTO(out_ucred, rc = err_serious(rc));
1893 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
1895 GOTO(out_ucred, rc);
1896 } else if (rc == 1) {
1897 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1898 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1899 GOTO(out_ucred, rc);
1901 rc = mdt_reint_rec(info, lhc);
1904 mdt_exit_ucred(info);
1906 mdt_client_compatibility(info);
1907 rc2 = mdt_fix_reply(info);
1913 static long mdt_reint_opcode(struct ptlrpc_request *req,
1914 const struct req_format **fmt)
1916 struct mdt_device *mdt;
1917 struct mdt_rec_reint *rec;
1920 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1922 opc = rec->rr_opcode;
1923 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1924 if (opc < REINT_MAX && fmt[opc] != NULL)
1925 req_capsule_extend(&req->rq_pill, fmt[opc]);
1927 mdt = mdt_exp2dev(req->rq_export);
1928 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1929 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1930 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1931 opc = err_serious(-EFAULT);
1934 opc = err_serious(-EFAULT);
1939 static int mdt_reint(struct tgt_session_info *tsi)
1943 static const struct req_format *reint_fmts[REINT_MAX] = {
1944 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1945 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1946 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1947 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1948 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1949 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1950 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1951 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
1952 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME
1957 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1959 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1961 * No lock possible here from client to pass it to reint code
1964 rc = mdt_reint_internal(info, NULL, opc);
1965 mdt_thread_info_fini(info);
1970 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1974 /* this should sync the whole device */
1975 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1977 struct dt_device *dt = mdt->mdt_bottom;
1981 rc = dt->dd_ops->dt_sync(env, dt);
1985 /* this should sync this object */
1986 static int mdt_object_sync(struct mdt_thread_info *info)
1988 struct md_object *next;
1992 if (!mdt_object_exists(info->mti_object)) {
1993 CWARN("%s: non existing object "DFID": rc = %d\n",
1994 mdt_obd_name(info->mti_mdt),
1995 PFID(mdt_object_fid(info->mti_object)), -ESTALE);
1998 next = mdt_object_child(info->mti_object);
1999 rc = mo_object_sync(info->mti_env, next);
2004 static int mdt_sync(struct tgt_session_info *tsi)
2006 struct ptlrpc_request *req = tgt_ses_req(tsi);
2007 struct req_capsule *pill = tsi->tsi_pill;
2008 struct mdt_body *body;
2013 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2014 RETURN(err_serious(-ENOMEM));
2016 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2017 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2019 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2021 /* sync an object */
2022 rc = mdt_object_sync(info);
2024 const struct lu_fid *fid;
2025 struct lu_attr *la = &info->mti_attr.ma_attr;
2027 info->mti_attr.ma_need = MA_INODE;
2028 info->mti_attr.ma_valid = 0;
2029 rc = mdt_attr_get_complex(info, info->mti_object,
2032 body = req_capsule_server_get(pill,
2034 fid = mdt_object_fid(info->mti_object);
2035 mdt_pack_attr2body(info, body, la, fid);
2038 mdt_thread_info_fini(info);
2041 mdt_counter_incr(req, LPROC_MDT_SYNC);
2047 * Handle quota control requests to consult current usage/limit, but also
2048 * to configure quota enforcement
2050 static int mdt_quotactl(struct tgt_session_info *tsi)
2052 struct obd_export *exp = tsi->tsi_exp;
2053 struct req_capsule *pill = tsi->tsi_pill;
2054 struct obd_quotactl *oqctl, *repoqc;
2056 struct mdt_device *mdt = mdt_exp2dev(exp);
2057 struct lu_device *qmt = mdt->mdt_qmt_dev;
2058 struct lu_nodemap *nodemap;
2061 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2063 RETURN(err_serious(-EPROTO));
2065 rc = req_capsule_server_pack(pill);
2067 RETURN(err_serious(rc));
2069 nodemap = nodemap_get_from_exp(exp);
2070 if (IS_ERR(nodemap))
2071 RETURN(PTR_ERR(nodemap));
2073 switch (oqctl->qc_cmd) {
2074 /* master quotactl */
2077 if (!nodemap_can_setquota(nodemap))
2078 GOTO(out_nodemap, rc = -EPERM);
2082 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2083 /* slave quotactl */
2088 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2089 GOTO(out_nodemap, rc = -EFAULT);
2093 if (oqctl->qc_type == USRQUOTA)
2094 id = nodemap_map_id(nodemap, NODEMAP_UID,
2095 NODEMAP_CLIENT_TO_FS, id);
2096 else if (oqctl->qc_type == GRPQUOTA)
2097 id = nodemap_map_id(nodemap, NODEMAP_UID,
2098 NODEMAP_CLIENT_TO_FS, id);
2100 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2102 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2104 if (oqctl->qc_id != id)
2105 swap(oqctl->qc_id, id);
2107 switch (oqctl->qc_cmd) {
2113 /* forward quotactl request to QMT */
2114 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2119 /* slave quotactl */
2120 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2125 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2126 GOTO(out_nodemap, rc = -EFAULT);
2129 if (oqctl->qc_id != id)
2130 swap(oqctl->qc_id, id);
2137 nodemap_putref(nodemap);
2142 /** clone llog ctxt from child (mdd)
2143 * This allows remote llog (replicator) access.
2144 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2145 * context was originally set up, or we can handle them directly.
2146 * I choose the latter, but that means I need any llog
2147 * contexts set up by child to be accessable by the mdt. So we clone the
2148 * context into our context list here.
2150 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2153 struct md_device *next = mdt->mdt_child;
2154 struct llog_ctxt *ctxt;
2157 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2160 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2161 if (rc || ctxt == NULL) {
2165 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2167 CERROR("Can't set mdt ctxt %d\n", rc);
2172 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2173 struct mdt_device *mdt, int idx)
2175 struct llog_ctxt *ctxt;
2177 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2180 /* Put once for the get we just did, and once for the clone */
2181 llog_ctxt_put(ctxt);
2182 llog_ctxt_put(ctxt);
2187 * sec context handlers
2189 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2191 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2197 * quota request handlers
2199 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2201 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2202 struct lu_device *qmt = mdt->mdt_qmt_dev;
2207 RETURN(err_serious(-EOPNOTSUPP));
2209 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2213 struct mdt_object *mdt_object_new(const struct lu_env *env,
2214 struct mdt_device *d,
2215 const struct lu_fid *f)
2217 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2218 struct lu_object *o;
2219 struct mdt_object *m;
2222 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2223 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2224 if (unlikely(IS_ERR(o)))
2225 m = (struct mdt_object *)o;
2231 struct mdt_object *mdt_object_find(const struct lu_env *env,
2232 struct mdt_device *d,
2233 const struct lu_fid *f)
2235 struct lu_object *o;
2236 struct mdt_object *m;
2239 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2240 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2241 if (unlikely(IS_ERR(o)))
2242 m = (struct mdt_object *)o;
2250 * Asyncronous commit for mdt device.
2252 * Pass asynchonous commit call down the MDS stack.
2254 * \param env environment
2255 * \param mdt the mdt device
2257 static void mdt_device_commit_async(const struct lu_env *env,
2258 struct mdt_device *mdt)
2260 struct dt_device *dt = mdt->mdt_bottom;
2264 rc = dt->dd_ops->dt_commit_async(env, dt);
2265 if (unlikely(rc != 0))
2266 CWARN("%s: async commit start failed: rc = %d\n",
2267 mdt_obd_name(mdt), rc);
2268 atomic_inc(&mdt->mdt_async_commit_count);
2273 * Mark the lock as "synchonous".
2275 * Mark the lock to deffer transaction commit to the unlock time.
2277 * \param lock the lock to mark as "synchonous"
2279 * \see mdt_is_lock_sync
2280 * \see mdt_save_lock
2282 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2284 lock->l_ast_data = (void*)1;
2288 * Check whehter the lock "synchonous" or not.
2290 * \param lock the lock to check
2291 * \retval 1 the lock is "synchonous"
2292 * \retval 0 the lock isn't "synchronous"
2294 * \see mdt_set_lock_sync
2295 * \see mdt_save_lock
2297 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2299 return lock->l_ast_data != NULL;
2303 * Blocking AST for mdt locks.
2305 * Starts transaction commit if in case of COS lock conflict or
2306 * deffers such a commit to the mdt_save_lock.
2308 * \param lock the lock which blocks a request or cancelling lock
2309 * \param desc unused
2310 * \param data unused
2311 * \param flag indicates whether this cancelling or blocking callback
2313 * \see ldlm_blocking_ast_nocheck
2315 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2316 void *data, int flag)
2318 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2319 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2323 if (flag == LDLM_CB_CANCELING)
2326 lock_res_and_lock(lock);
2327 if (lock->l_blocking_ast != mdt_blocking_ast) {
2328 unlock_res_and_lock(lock);
2331 if (lock->l_req_mode & (LCK_PW | LCK_EX) &&
2332 lock->l_blocking_lock != NULL) {
2333 if (mdt_cos_is_enabled(mdt) &&
2334 lock->l_client_cookie !=
2335 lock->l_blocking_lock->l_client_cookie)
2336 mdt_set_lock_sync(lock);
2337 else if (mdt_slc_is_enabled(mdt) &&
2338 ldlm_is_cos_incompat(lock->l_blocking_lock))
2339 mdt_set_lock_sync(lock);
2341 rc = ldlm_blocking_ast_nocheck(lock);
2343 /* There is no lock conflict if l_blocking_lock == NULL,
2344 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2345 * when the last reference to a local lock was released */
2346 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2349 rc = lu_env_init(&env, LCT_LOCAL);
2350 if (unlikely(rc != 0))
2351 CWARN("%s: lu_env initialization failed, cannot "
2352 "start asynchronous commit: rc = %d\n",
2355 mdt_device_commit_async(&env, mdt);
2362 * Blocking AST for cross-MDT lock
2364 * Discard lock from uncommitted_slc_locks and cancel it.
2366 * \param lock the lock which blocks a request or cancelling lock
2367 * \param desc unused
2368 * \param data unused
2369 * \param flag indicates whether this cancelling or blocking callback
2370 * \retval 0 on success
2371 * \retval negative number on error
2373 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2374 void *data, int flag)
2380 case LDLM_CB_BLOCKING: {
2381 struct lustre_handle lockh;
2383 ldlm_lock2handle(lock, &lockh);
2384 rc = ldlm_cli_cancel(&lockh,
2385 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
2387 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2392 case LDLM_CB_CANCELING: {
2393 LDLM_DEBUG(lock, "Revoke remote lock\n");
2395 /* discard slc lock here so that it can be cleaned anytime,
2396 * especially for cleanup_resource() */
2397 tgt_discard_slc_lock(lock);
2399 /* once we cache lock, l_ast_data is set to mdt_object */
2400 if (lock->l_ast_data != NULL) {
2401 struct mdt_object *mo = lock->l_ast_data;
2404 rc = lu_env_init(&env, LCT_MD_THREAD);
2405 if (unlikely(rc != 0)) {
2406 struct obd_device *obd;
2408 obd = ldlm_lock_to_ns(lock)->ns_obd;
2409 CWARN("%s: lu_env initialization failed, object"
2410 "%p "DFID" is leaked!\n",
2412 PFID(mdt_object_fid(mo)));
2416 if (lock->l_policy_data.l_inodebits.bits &
2417 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
2418 rc = mo_invalidate(&env, mdt_object_child(mo));
2419 mo->mot_cache_attr = 0;
2421 mdt_object_put(&env, mo);
2433 int mdt_check_resent_lock(struct mdt_thread_info *info,
2434 struct mdt_object *mo,
2435 struct mdt_lock_handle *lhc)
2437 /* the lock might already be gotten in ldlm_handle_enqueue() */
2438 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2439 struct ptlrpc_request *req = mdt_info_req(info);
2440 struct ldlm_lock *lock;
2442 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2443 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2445 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2446 * a resend case, however, it could be already destroyed
2447 * due to client eviction or a raced cancel RPC. */
2448 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64,
2449 lhc->mlh_reg_lh.cookie);
2453 if (!fid_res_name_eq(mdt_object_fid(mo),
2454 &lock->l_resource->lr_name)) {
2455 CWARN("%s: Although resent, but still not "
2456 "get child lock:"DFID"\n",
2457 info->mti_exp->exp_obd->obd_name,
2458 PFID(mdt_object_fid(mo)));
2459 LDLM_LOCK_PUT(lock);
2462 LDLM_LOCK_PUT(lock);
2468 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
2469 const struct lu_fid *fid, struct lustre_handle *lh,
2470 enum ldlm_mode mode, __u64 ibits, bool nonblock,
2473 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2474 union ldlm_policy_data *policy = &mti->mti_policy;
2475 struct ldlm_res_id *res_id = &mti->mti_res_id;
2479 LASSERT(mdt_object_remote(o));
2481 fid_build_reg_res_name(fid, res_id);
2483 memset(einfo, 0, sizeof(*einfo));
2484 einfo->ei_type = LDLM_IBITS;
2485 einfo->ei_mode = mode;
2486 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2487 einfo->ei_cb_cp = ldlm_completion_ast;
2488 einfo->ei_enq_slave = 0;
2489 einfo->ei_res_id = res_id;
2491 einfo->ei_nonblock = 1;
2494 * if we cache lock, couple lock with mdt_object, so that object
2495 * can be easily found in lock ASTs.
2497 mdt_object_get(mti->mti_env, o);
2498 einfo->ei_cbdata = o;
2501 memset(policy, 0, sizeof(*policy));
2502 policy->l_inodebits.bits = ibits;
2504 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2506 if (rc < 0 && cache) {
2507 mdt_object_put(mti->mti_env, o);
2508 einfo->ei_cbdata = NULL;
2513 static int mdt_object_local_lock(struct mdt_thread_info *info,
2514 struct mdt_object *o,
2515 struct mdt_lock_handle *lh, __u64 ibits,
2516 bool nonblock, bool cos_incompat)
2518 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2519 union ldlm_policy_data *policy = &info->mti_policy;
2520 struct ldlm_res_id *res_id = &info->mti_res_id;
2525 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2526 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2527 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2528 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2531 LASSERT(lh->mlh_reg_mode == LCK_PW ||
2532 lh->mlh_reg_mode == LCK_EX);
2533 dlmflags |= LDLM_FL_COS_INCOMPAT;
2534 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
2535 dlmflags |= LDLM_FL_COS_ENABLED;
2538 /* Only enqueue LOOKUP lock for remote object */
2539 if (mdt_object_remote(o))
2540 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2542 if (lh->mlh_type == MDT_PDO_LOCK) {
2543 /* check for exists after object is locked */
2544 if (mdt_object_exists(o) == 0) {
2545 /* Non-existent object shouldn't have PDO lock */
2548 /* Non-dir object shouldn't have PDO lock */
2549 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2554 memset(policy, 0, sizeof(*policy));
2555 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2557 dlmflags |= LDLM_FL_ATOMIC_CB;
2559 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2562 * Take PDO lock on whole directory and build correct @res_id for lock
2563 * on part of directory.
2565 if (lh->mlh_pdo_hash != 0) {
2566 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2567 mdt_lock_pdo_mode(info, o, lh);
2568 if (lh->mlh_pdo_mode != LCK_NL) {
2570 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2571 * is never going to be sent to client and we do not
2572 * want it slowed down due to possible cancels.
2574 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2575 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2576 policy, res_id, dlmflags,
2577 info->mti_exp == NULL ? NULL :
2578 &info->mti_exp->exp_handle.h_cookie);
2579 if (unlikely(rc != 0))
2580 GOTO(out_unlock, rc);
2584 * Finish res_id initializing by name hash marking part of
2585 * directory which is taking modification.
2587 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2590 policy->l_inodebits.bits = ibits;
2593 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2594 * going to be sent to client. If it is - mdt_intent_policy() path will
2595 * fix it up and turn FL_LOCAL flag off.
2597 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2598 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2599 info->mti_exp == NULL ? NULL :
2600 &info->mti_exp->exp_handle.h_cookie);
2603 mdt_object_unlock(info, o, lh, 1);
2604 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2605 lh->mlh_pdo_hash != 0 &&
2606 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2607 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2613 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2614 struct mdt_lock_handle *lh, __u64 ibits, bool nonblock,
2617 struct mdt_lock_handle *local_lh = NULL;
2621 if (!mdt_object_remote(o)) {
2622 rc = mdt_object_local_lock(info, o, lh, ibits, nonblock,
2627 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2628 ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2629 MDS_INODELOCK_XATTR);
2631 /* Only enqueue LOOKUP lock for remote object */
2632 if (ibits & MDS_INODELOCK_LOOKUP) {
2633 rc = mdt_object_local_lock(info, o, lh, MDS_INODELOCK_LOOKUP,
2634 nonblock, cos_incompat);
2641 if (ibits & MDS_INODELOCK_UPDATE) {
2642 /* Sigh, PDO needs to enqueue 2 locks right now, but
2643 * enqueue RPC can only request 1 lock, to avoid extra
2644 * RPC, so it will instead enqueue EX lock for remote
2645 * object anyway XXX*/
2646 if (lh->mlh_type == MDT_PDO_LOCK &&
2647 lh->mlh_pdo_hash != 0) {
2648 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2649 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2650 PFID(mdt_object_fid(o)));
2651 lh->mlh_pdo_hash = 0;
2652 lh->mlh_rreg_mode = LCK_EX;
2653 lh->mlh_type = MDT_REG_LOCK;
2655 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2658 MDS_INODELOCK_UPDATE, nonblock,
2660 if (rc != ELDLM_OK) {
2661 if (local_lh != NULL)
2662 mdt_object_unlock(info, o, local_lh, rc);
2670 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2671 struct mdt_lock_handle *lh, __u64 ibits)
2673 return mdt_object_lock_internal(info, o, lh, ibits, false, false);
2676 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2677 struct mdt_lock_handle *lh, __u64 ibits,
2680 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
2681 return mdt_object_lock_internal(info, o, lh, ibits, false,
2685 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2686 struct mdt_lock_handle *lh, __u64 ibits)
2688 struct mdt_lock_handle tmp = *lh;
2691 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true, false);
2698 int mdt_reint_object_lock_try(struct mdt_thread_info *info,
2699 struct mdt_object *o, struct mdt_lock_handle *lh,
2700 __u64 ibits, bool cos_incompat)
2702 struct mdt_lock_handle tmp = *lh;
2705 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
2706 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true, cos_incompat);
2714 * Save a lock within request object.
2716 * Keep the lock referenced until whether client ACK or transaction
2717 * commit happens or release the lock immediately depending on input
2718 * parameters. If COS is ON, a write lock is converted to COS lock
2721 * \param info thead info object
2722 * \param h lock handle
2723 * \param mode lock mode
2724 * \param decref force immediate lock releasing
2726 static void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2727 enum ldlm_mode mode, int decref)
2731 if (lustre_handle_is_used(h)) {
2732 if (decref || !info->mti_has_trans ||
2733 !(mode & (LCK_PW | LCK_EX))) {
2734 mdt_fid_unlock(h, mode);
2736 struct mdt_device *mdt = info->mti_mdt;
2737 struct ldlm_lock *lock = ldlm_handle2lock(h);
2738 struct ptlrpc_request *req = mdt_info_req(info);
2741 cos = (mdt_cos_is_enabled(mdt) ||
2742 mdt_slc_is_enabled(mdt));
2744 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2747 /* there is no request if mdt_object_unlock() is called
2748 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2749 if (likely(req != NULL)) {
2750 CDEBUG(D_HA, "request = %p reply state = %p"
2751 " transno = "LPD64"\n", req,
2752 req->rq_reply_state, req->rq_transno);
2754 ldlm_lock_downgrade(lock, LCK_COS);
2757 ptlrpc_save_lock(req, h, mode, cos);
2759 mdt_fid_unlock(h, mode);
2761 if (mdt_is_lock_sync(lock)) {
2762 CDEBUG(D_HA, "found sync-lock,"
2763 " async commit started\n");
2764 mdt_device_commit_async(info->mti_env,
2767 LDLM_LOCK_PUT(lock);
2776 * Save cross-MDT lock in uncommitted_slc_locks
2778 * Keep the lock referenced until transaction commit happens or release the lock
2779 * immediately depending on input parameters.
2781 * \param info thead info object
2782 * \param h lock handle
2783 * \param mode lock mode
2784 * \param decref force immediate lock releasing
2786 static void mdt_save_remote_lock(struct mdt_thread_info *info,
2787 struct mdt_object *o, struct lustre_handle *h,
2788 enum ldlm_mode mode, int decref)
2792 if (lustre_handle_is_used(h)) {
2793 struct ldlm_lock *lock = ldlm_handle2lock(h);
2796 (lock->l_policy_data.l_inodebits.bits &
2797 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
2798 mo_invalidate(info->mti_env, mdt_object_child(o));
2800 if (decref || !info->mti_has_trans ||
2801 !(mode & (LCK_PW | LCK_EX))) {
2802 ldlm_lock_decref_and_cancel(h, mode);
2803 LDLM_LOCK_PUT(lock);
2805 struct ptlrpc_request *req = mdt_info_req(info);
2807 LASSERT(req != NULL);
2808 tgt_save_slc_lock(lock, req->rq_transno);
2809 ldlm_lock_decref(h, mode);
2818 * Unlock mdt object.
2820 * Immeditely release the regular lock and the PDO lock or save the
2821 * lock in request and keep them referenced until client ACK or
2822 * transaction commit.
2824 * \param info thread info object
2825 * \param o mdt object
2826 * \param lh mdt lock handle referencing regular and PDO locks
2827 * \param decref force immediate lock releasing
2829 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2830 struct mdt_lock_handle *lh, int decref)
2834 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2835 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2836 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
2842 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2843 const struct lu_fid *f,
2844 struct mdt_lock_handle *lh,
2847 struct mdt_object *o;
2849 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2853 rc = mdt_object_lock(info, o, lh, ibits);
2855 mdt_object_put(info->mti_env, o);
2862 void mdt_object_unlock_put(struct mdt_thread_info * info,
2863 struct mdt_object * o,
2864 struct mdt_lock_handle *lh,
2867 mdt_object_unlock(info, o, lh, decref);
2868 mdt_object_put(info->mti_env, o);
2872 * Generic code handling requests that have struct mdt_body passed in:
2874 * - extract mdt_body from request and save it in @info, if present;
2876 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2879 * - if HABEO_CORPUS flag is set for this request type check whether object
2880 * actually exists on storage (lu_object_exists()).
2883 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2885 const struct mdt_body *body;
2886 struct mdt_object *obj;
2887 const struct lu_env *env;
2888 struct req_capsule *pill;
2892 env = info->mti_env;
2893 pill = info->mti_pill;
2895 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2899 if (!(body->mbo_valid & OBD_MD_FLID))
2902 if (!fid_is_sane(&body->mbo_fid1)) {
2903 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2907 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2909 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2910 mdt_object_put(env, obj);
2913 info->mti_object = obj;
2922 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2924 struct req_capsule *pill = info->mti_pill;
2928 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2929 rc = mdt_body_unpack(info, flags);
2933 if (rc == 0 && (flags & HABEO_REFERO)) {
2935 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2936 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2938 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2939 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2942 rc = req_capsule_server_pack(pill);
2947 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2949 lh->mlh_type = MDT_NUL_LOCK;
2950 lh->mlh_reg_lh.cookie = 0ull;
2951 lh->mlh_reg_mode = LCK_MINMODE;
2952 lh->mlh_pdo_lh.cookie = 0ull;
2953 lh->mlh_pdo_mode = LCK_MINMODE;
2954 lh->mlh_rreg_lh.cookie = 0ull;
2955 lh->mlh_rreg_mode = LCK_MINMODE;
2958 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2960 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2961 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2965 * Initialize fields of struct mdt_thread_info. Other fields are left in
2966 * uninitialized state, because it's too expensive to zero out whole
2967 * mdt_thread_info (> 1K) on each request arrival.
2969 void mdt_thread_info_init(struct ptlrpc_request *req,
2970 struct mdt_thread_info *info)
2974 info->mti_pill = &req->rq_pill;
2977 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2978 mdt_lock_handle_init(&info->mti_lh[i]);
2980 /* mdt device: it can be NULL while CONNECT */
2981 if (req->rq_export) {
2982 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2983 info->mti_exp = req->rq_export;
2985 info->mti_mdt = NULL;
2986 info->mti_env = req->rq_svc_thread->t_env;
2987 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2989 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2990 info->mti_big_buf = LU_BUF_NULL;
2991 info->mti_body = NULL;
2992 info->mti_object = NULL;
2993 info->mti_dlm_req = NULL;
2994 info->mti_has_trans = 0;
2995 info->mti_cross_ref = 0;
2996 info->mti_opdata = 0;
2997 info->mti_big_lmm_used = 0;
2999 info->mti_spec.no_create = 0;
3000 info->mti_spec.sp_rm_entry = 0;
3001 info->mti_spec.sp_permitted = 0;
3002 info->mti_spec.sp_migrate_close = 0;
3004 info->mti_spec.u.sp_ea.eadata = NULL;
3005 info->mti_spec.u.sp_ea.eadatalen = 0;
3008 void mdt_thread_info_fini(struct mdt_thread_info *info)
3012 if (info->mti_object != NULL) {
3013 mdt_object_put(info->mti_env, info->mti_object);
3014 info->mti_object = NULL;
3017 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3018 mdt_lock_handle_fini(&info->mti_lh[i]);
3019 info->mti_env = NULL;
3020 info->mti_pill = NULL;
3021 info->mti_exp = NULL;
3023 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3024 lu_buf_free(&info->mti_big_buf);
3027 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3029 struct mdt_thread_info *mti;
3031 mti = mdt_th_info(tsi->tsi_env);
3032 LASSERT(mti != NULL);
3034 mdt_thread_info_init(tgt_ses_req(tsi), mti);
3035 if (tsi->tsi_corpus != NULL) {
3036 mti->mti_object = mdt_obj(tsi->tsi_corpus);
3037 lu_object_get(tsi->tsi_corpus);
3039 mti->mti_body = tsi->tsi_mdt_body;
3040 mti->mti_dlm_req = tsi->tsi_dlm_req;
3045 static int mdt_tgt_connect(struct tgt_session_info *tsi)
3047 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
3049 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
3050 set_current_state(TASK_UNINTERRUPTIBLE);
3051 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
3054 return tgt_connect(tsi);
3072 static int mdt_intent_getattr(enum mdt_it_code opcode,
3073 struct mdt_thread_info *info,
3074 struct ldlm_lock **,
3077 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3078 struct mdt_thread_info *info,
3079 struct ldlm_lock **lockp,
3082 static int mdt_intent_layout(enum mdt_it_code opcode,
3083 struct mdt_thread_info *info,
3084 struct ldlm_lock **,
3086 static int mdt_intent_reint(enum mdt_it_code opcode,
3087 struct mdt_thread_info *info,
3088 struct ldlm_lock **,
3091 static struct mdt_it_flavor {
3092 const struct req_format *it_fmt;
3094 int (*it_act)(enum mdt_it_code ,
3095 struct mdt_thread_info *,
3096 struct ldlm_lock **,
3099 } mdt_it_flavor[] = {
3101 .it_fmt = &RQF_LDLM_INTENT,
3102 /*.it_flags = HABEO_REFERO,*/
3104 .it_act = mdt_intent_reint,
3105 .it_reint = REINT_OPEN
3108 .it_fmt = &RQF_LDLM_INTENT,
3110 * OCREAT is not a MUTABOR request as if the file
3112 * We do the extra check of OBD_CONNECT_RDONLY in
3113 * mdt_reint_open() when we really need to create
3117 .it_act = mdt_intent_reint,
3118 .it_reint = REINT_OPEN
3121 .it_fmt = &RQF_LDLM_INTENT,
3122 .it_flags = MUTABOR,
3123 .it_act = mdt_intent_reint,
3124 .it_reint = REINT_CREATE
3126 [MDT_IT_GETATTR] = {
3127 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
3128 .it_flags = HABEO_REFERO,
3129 .it_act = mdt_intent_getattr
3131 [MDT_IT_READDIR] = {
3137 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
3138 .it_flags = HABEO_REFERO,
3139 .it_act = mdt_intent_getattr
3142 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
3143 .it_flags = MUTABOR,
3145 .it_reint = REINT_UNLINK
3149 .it_flags = MUTABOR,
3152 [MDT_IT_GETXATTR] = {
3153 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
3154 .it_flags = HABEO_CORPUS,
3155 .it_act = mdt_intent_getxattr
3158 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
3160 .it_act = mdt_intent_layout
3165 mdt_intent_lock_replace(struct mdt_thread_info *info,
3166 struct ldlm_lock **lockp,
3167 struct mdt_lock_handle *lh,
3168 __u64 flags, int result)
3170 struct ptlrpc_request *req = mdt_info_req(info);
3171 struct ldlm_lock *lock = *lockp;
3172 struct ldlm_lock *new_lock;
3174 /* If possible resent found a lock, @lh is set to its handle */
3175 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3177 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3178 lh->mlh_reg_lh.cookie = 0;
3182 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
3183 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3184 * a resend case, however, it could be already destroyed
3185 * due to client eviction or a raced cancel RPC. */
3186 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
3187 lh->mlh_reg_lh.cookie);
3188 lh->mlh_reg_lh.cookie = 0;
3192 LASSERTF(new_lock != NULL,
3193 "lockh "LPX64" flags "LPX64" rc %d\n",
3194 lh->mlh_reg_lh.cookie, flags, result);
3197 * If we've already given this lock to a client once, then we should
3198 * have no readers or writers. Otherwise, we should have one reader
3199 * _or_ writer ref (which will be zeroed below) before returning the
3202 if (new_lock->l_export == req->rq_export) {
3203 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3205 LASSERT(new_lock->l_export == NULL);
3206 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3211 if (new_lock->l_export == req->rq_export) {
3213 * Already gave this to the client, which means that we
3214 * reconstructed a reply.
3216 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3219 LDLM_LOCK_RELEASE(new_lock);
3220 lh->mlh_reg_lh.cookie = 0;
3221 RETURN(ELDLM_LOCK_REPLACED);
3225 * Fixup the lock to be given to the client.
3227 lock_res_and_lock(new_lock);
3228 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3229 * possible blocking AST. */
3230 while (new_lock->l_readers > 0) {
3231 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3232 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3233 new_lock->l_readers--;
3235 while (new_lock->l_writers > 0) {
3236 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3237 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3238 new_lock->l_writers--;
3241 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3242 new_lock->l_blocking_ast = lock->l_blocking_ast;
3243 new_lock->l_completion_ast = lock->l_completion_ast;
3244 new_lock->l_remote_handle = lock->l_remote_handle;
3245 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3247 unlock_res_and_lock(new_lock);
3249 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3250 &new_lock->l_remote_handle,
3251 &new_lock->l_exp_hash);
3253 LDLM_LOCK_RELEASE(new_lock);
3254 lh->mlh_reg_lh.cookie = 0;
3256 RETURN(ELDLM_LOCK_REPLACED);
3259 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3260 struct ldlm_lock *new_lock,
3261 struct mdt_lock_handle *lh,
3264 struct ptlrpc_request *req = mdt_info_req(info);
3265 struct ldlm_request *dlmreq;
3267 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3270 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3272 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3273 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3275 if (flags & LDLM_FL_RESENT) {
3276 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3277 lh->mlh_reg_mode = new_lock->l_granted_mode;
3279 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3280 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3281 lh->mlh_reg_lh.cookie);
3286 * If the xid matches, then we know this is a resent request, and allow
3287 * it. (It's probably an OPEN, for which we don't send a lock.
3289 if (req_can_reconstruct(req, NULL))
3293 * This remote handle isn't enqueued, so we never received or processed
3294 * this request. Clear MSG_RESENT, because it can be handled like any
3295 * normal request now.
3297 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3299 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3300 dlmreq->lock_handle[0].cookie);
3303 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3304 struct mdt_thread_info *info,
3305 struct ldlm_lock **lockp,
3308 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3309 struct ldlm_reply *ldlm_rep = NULL;
3313 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3314 * (for the resend case) or a new lock. Below we will use it to
3315 * replace the original lock.
3317 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3318 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3319 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3320 rc = mdt_object_lock(info, info->mti_object, lhc,
3321 MDS_INODELOCK_XATTR);
3326 grc = mdt_getxattr(info);
3328 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, 0);
3330 if (mdt_info_req(info)->rq_repmsg != NULL)
3331 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3332 if (ldlm_rep == NULL)
3333 RETURN(err_serious(-EFAULT));
3335 ldlm_rep->lock_policy_res2 = grc;
3340 static int mdt_intent_getattr(enum mdt_it_code opcode,
3341 struct mdt_thread_info *info,
3342 struct ldlm_lock **lockp,
3345 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3347 struct ldlm_reply *ldlm_rep;
3348 struct mdt_body *reqbody;
3349 struct mdt_body *repbody;
3353 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3356 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3359 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3360 repbody->mbo_eadatasize = 0;
3361 repbody->mbo_aclsize = 0;
3365 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3367 case MDT_IT_GETATTR:
3368 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3372 CERROR("Unsupported intent (%d)\n", opcode);
3373 GOTO(out_shrink, rc = -EINVAL);
3376 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3378 GOTO(out_shrink, rc);
3380 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3381 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3383 /* Get lock from request for possible resent case. */
3384 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3386 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3387 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3389 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3390 ldlm_rep->lock_policy_res2 = 0;
3391 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3392 ldlm_rep->lock_policy_res2) {
3393 lhc->mlh_reg_lh.cookie = 0ull;
3394 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3397 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3400 mdt_exit_ucred(info);
3402 mdt_client_compatibility(info);
3403 rc2 = mdt_fix_reply(info);
3409 static int mdt_intent_layout(enum mdt_it_code opcode,
3410 struct mdt_thread_info *info,
3411 struct ldlm_lock **lockp,
3414 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3415 struct layout_intent *layout;
3417 struct mdt_object *obj = NULL;
3418 int layout_size = 0;
3422 if (opcode != MDT_IT_LAYOUT) {
3423 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3428 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3432 if (layout->li_opc != LAYOUT_INTENT_ACCESS) {
3433 CERROR("%s: Unsupported layout intent opc %d\n",
3434 mdt_obd_name(info->mti_mdt), layout->li_opc);
3438 fid = &info->mti_tmp_fid2;
3439 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3441 /* Get lock from request for possible resent case. */
3442 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3444 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3446 GOTO(out, rc = PTR_ERR(obj));
3448 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3449 layout_size = mdt_attr_get_eabuf_size(info, obj);
3450 if (layout_size < 0)
3451 GOTO(out_obj, rc = layout_size);
3453 if (layout_size > info->mti_mdt->mdt_max_mdsize)
3454 info->mti_mdt->mdt_max_mdsize = layout_size;
3457 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3458 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3460 rc = req_capsule_server_pack(info->mti_pill);
3464 mdt_object_put(info->mti_env, obj);
3466 if (rc == 0 && lustre_handle_is_used(&lhc->mlh_reg_lh))
3467 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3470 lhc->mlh_reg_lh.cookie = 0;
3475 static int mdt_intent_reint(enum mdt_it_code opcode,
3476 struct mdt_thread_info *info,
3477 struct ldlm_lock **lockp,
3480 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3481 struct ldlm_reply *rep = NULL;
3485 static const struct req_format *intent_fmts[REINT_MAX] = {
3486 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3487 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3492 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3496 if (mdt_it_flavor[opcode].it_reint != opc) {
3497 CERROR("Reint code %ld doesn't match intent: %d\n",
3499 RETURN(err_serious(-EPROTO));
3502 /* Get lock from request for possible resent case. */
3503 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3505 rc = mdt_reint_internal(info, lhc, opc);
3507 /* Check whether the reply has been packed successfully. */
3508 if (mdt_info_req(info)->rq_repmsg != NULL)
3509 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3511 RETURN(err_serious(-EFAULT));
3513 /* MDC expects this in any case */
3515 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3517 /* the open lock or the lock for cross-ref object should be
3518 * returned to the client */
3519 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3520 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3521 rep->lock_policy_res2 = 0;
3522 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3526 rep->lock_policy_res2 = clear_serious(rc);
3528 if (rep->lock_policy_res2 == -ENOENT &&
3529 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
3530 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
3531 rep->lock_policy_res2 = 0;
3533 lhc->mlh_reg_lh.cookie = 0ull;
3534 if (rc == -ENOTCONN || rc == -ENODEV ||
3535 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3537 * If it is the disconnect error (ENODEV & ENOCONN), the error
3538 * will be returned by rq_status, and client at ptlrpc layer
3539 * will detect this, then disconnect, reconnect the import
3540 * immediately, instead of impacting the following the rpc.
3545 * For other cases, the error will be returned by intent, and client
3546 * will retrieve the result from intent.
3548 RETURN(ELDLM_LOCK_ABORTED);
3551 static int mdt_intent_code(enum ldlm_intent_flags itcode)
3559 case IT_OPEN|IT_CREAT:
3566 rc = MDT_IT_READDIR;
3569 rc = MDT_IT_GETATTR;
3581 rc = MDT_IT_GETXATTR;
3586 case IT_QUOTA_DQACQ:
3591 CERROR("Unknown intent opcode: 0x%08x\n", itcode);
3598 static int mdt_intent_opc(enum ldlm_intent_flags itopc,
3599 struct mdt_thread_info *info,
3600 struct ldlm_lock **lockp, __u64 flags)
3602 struct req_capsule *pill = info->mti_pill;
3603 struct ptlrpc_request *req = mdt_info_req(info);
3604 struct mdt_it_flavor *flv;
3609 opc = mdt_intent_code(itopc);
3613 if (opc == MDT_IT_QUOTA) {
3614 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3617 RETURN(-EOPNOTSUPP);
3619 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3620 /* pass the request to quota master */
3621 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3622 mdt_info_req(info), lockp,
3627 flv = &mdt_it_flavor[opc];
3628 if (flv->it_fmt != NULL)
3629 req_capsule_extend(pill, flv->it_fmt);
3631 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3635 if (flv->it_flags & MUTABOR &&
3636 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3639 if (flv->it_act != NULL) {
3640 struct ldlm_reply *rep;
3642 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
3644 /* execute policy */
3645 rc = flv->it_act(opc, info, lockp, flags);
3647 /* Check whether the reply has been packed successfully. */
3648 if (req->rq_repmsg != NULL) {
3649 rep = req_capsule_server_get(info->mti_pill,
3651 rep->lock_policy_res2 =
3652 ptlrpc_status_hton(rep->lock_policy_res2);
3659 static int mdt_intent_policy(struct ldlm_namespace *ns,
3660 struct ldlm_lock **lockp, void *req_cookie,
3661 enum ldlm_mode mode, __u64 flags, void *data)
3663 struct tgt_session_info *tsi;
3664 struct mdt_thread_info *info;
3665 struct ptlrpc_request *req = req_cookie;
3666 struct ldlm_intent *it;
3667 struct req_capsule *pill;
3672 LASSERT(req != NULL);
3674 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3676 info = tsi2mdt_info(tsi);
3677 LASSERT(info != NULL);
3678 pill = info->mti_pill;
3679 LASSERT(pill->rc_req == req);
3681 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3682 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3683 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3685 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3689 /* Lock without inodebits makes no sense and will oops
3690 * later in ldlm. Let's check it now to see if we have
3691 * ibits corrupted somewhere in mdt_intent_opc().
3692 * The case for client miss to set ibits has been
3693 * processed by others. */
3694 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3695 lr_type == LDLM_IBITS,
3696 info->mti_dlm_req->lock_desc.\
3697 l_policy_data.l_inodebits.bits != 0));
3699 rc = err_serious(-EFAULT);
3701 /* No intent was provided */
3702 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3703 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3704 rc = req_capsule_server_pack(pill);
3706 rc = err_serious(rc);
3708 mdt_thread_info_fini(info);
3712 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3714 struct seq_server_site *ss = mdt_seq_site(mdt);
3716 if (ss->ss_node_id == 0)
3719 if (ss->ss_client_seq != NULL) {
3720 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3721 ss->ss_client_seq->lcs_exp = NULL;
3724 if (ss->ss_server_fld != NULL) {
3725 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3726 ss->ss_server_fld->lsf_control_exp = NULL;
3730 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3732 struct seq_server_site *ss = mdt_seq_site(mdt);
3737 if (ss->ss_server_seq != NULL)
3738 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3741 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3743 mdt_seq_fini_cli(mdt);
3744 mdt_deregister_seq_exp(mdt);
3746 return seq_site_fini(env, mdt_seq_site(mdt));
3750 * It will retrieve its FLDB entries from MDT0, and it only happens
3751 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3752 * and it needs to refresh FLDB from the MDT0.
3754 static int mdt_register_lwp_callback(void *data)
3757 struct mdt_device *mdt = data;
3758 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3762 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3764 rc = lu_env_init(&env, LCT_MD_THREAD);
3766 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3770 /* Allocate new sequence now to avoid creating local transaction
3771 * in the normal transaction process */
3772 rc = seq_server_check_and_alloc_super(&env,
3773 mdt_seq_site(mdt)->ss_server_seq);
3778 rc = fld_update_from_controller(&env, fld);
3780 CERROR("%s: cannot update controller: rc = %d\n",
3781 mdt_obd_name(mdt), rc);
3790 static int mdt_register_seq_exp(struct mdt_device *mdt)
3792 struct seq_server_site *ss = mdt_seq_site(mdt);
3793 char *lwp_name = NULL;
3796 if (ss->ss_node_id == 0)
3799 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3800 if (lwp_name == NULL)
3801 GOTO(out_free, rc = -ENOMEM);
3803 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3807 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3812 rc = lustre_register_lwp_item(lwp_name,
3813 &ss->ss_server_fld->lsf_control_exp,
3814 mdt_register_lwp_callback, mdt);
3816 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3817 ss->ss_client_seq->lcs_exp = NULL;
3821 if (lwp_name != NULL)
3822 OBD_FREE(lwp_name, MAX_OBD_NAME);
3828 * Init client sequence manager which is used by local MDS to talk to sequence
3829 * controller on remote node.
3831 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3833 struct seq_server_site *ss = mdt_seq_site(mdt);
3838 /* check if this is adding the first MDC and controller is not yet
3840 OBD_ALLOC_PTR(ss->ss_client_seq);
3841 if (ss->ss_client_seq == NULL)
3844 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3845 if (prefix == NULL) {
3846 OBD_FREE_PTR(ss->ss_client_seq);
3847 ss->ss_client_seq = NULL;
3851 /* Note: seq_client_fini will be called in seq_site_fini */
3852 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3853 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3854 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3856 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3858 OBD_FREE_PTR(ss->ss_client_seq);
3859 ss->ss_client_seq = NULL;
3863 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3868 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3870 struct seq_server_site *ss;
3874 ss = mdt_seq_site(mdt);
3875 /* init sequence controller server(MDT0) */
3876 if (ss->ss_node_id == 0) {
3877 OBD_ALLOC_PTR(ss->ss_control_seq);
3878 if (ss->ss_control_seq == NULL)
3881 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3882 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3885 GOTO(out_seq_fini, rc);
3888 /* Init normal sequence server */
3889 OBD_ALLOC_PTR(ss->ss_server_seq);
3890 if (ss->ss_server_seq == NULL)
3891 GOTO(out_seq_fini, rc = -ENOMEM);
3893 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3894 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3896 GOTO(out_seq_fini, rc);
3898 /* init seq client for seq server to talk to seq controller(MDT0) */
3899 rc = mdt_seq_init_cli(env, mdt);
3901 GOTO(out_seq_fini, rc);
3903 if (ss->ss_node_id != 0)
3904 /* register controller export through lwp */
3905 rc = mdt_register_seq_exp(mdt);
3910 mdt_seq_fini(env, mdt);
3918 static int mdt_fld_fini(const struct lu_env *env,
3919 struct mdt_device *m)
3921 struct seq_server_site *ss = mdt_seq_site(m);
3924 if (ss && ss->ss_server_fld) {
3925 fld_server_fini(env, ss->ss_server_fld);
3926 OBD_FREE_PTR(ss->ss_server_fld);
3927 ss->ss_server_fld = NULL;
3933 static int mdt_fld_init(const struct lu_env *env,
3935 struct mdt_device *m)
3937 struct seq_server_site *ss;
3941 ss = mdt_seq_site(m);
3943 OBD_ALLOC_PTR(ss->ss_server_fld);
3944 if (ss->ss_server_fld == NULL)
3945 RETURN(rc = -ENOMEM);
3947 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3950 OBD_FREE_PTR(ss->ss_server_fld);
3951 ss->ss_server_fld = NULL;
3958 static void mdt_stack_pre_fini(const struct lu_env *env,
3959 struct mdt_device *m, struct lu_device *top)
3961 struct lustre_cfg_bufs *bufs;
3962 struct lustre_cfg *lcfg;
3963 struct mdt_thread_info *info;
3968 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3969 LASSERT(info != NULL);
3971 bufs = &info->mti_u.bufs;
3973 LASSERT(m->mdt_child_exp);
3974 LASSERT(m->mdt_child_exp->exp_obd);
3976 /* process cleanup, pass mdt obd name to get obd umount flags */
3977 /* XXX: this is needed because all layers are referenced by
3978 * objects (some of them are pinned by osd, for example *
3979 * the proper solution should be a model where object used
3980 * by osd only doesn't have mdt/mdd slices -bzzz */
3981 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3982 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3983 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3987 top->ld_ops->ldo_process_config(env, top, lcfg);
3988 lustre_cfg_free(lcfg);
3992 static void mdt_stack_fini(const struct lu_env *env,
3993 struct mdt_device *m, struct lu_device *top)
3995 struct obd_device *obd = mdt2obd_dev(m);
3996 struct lustre_cfg_bufs *bufs;
3997 struct lustre_cfg *lcfg;
3998 struct mdt_thread_info *info;
4002 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4003 LASSERT(info != NULL);
4005 lu_dev_del_linkage(top->ld_site, top);
4007 lu_site_purge(env, top->ld_site, -1);
4009 bufs = &info->mti_u.bufs;
4010 /* process cleanup, pass mdt obd name to get obd umount flags */
4011 /* another purpose is to let all layers to release their objects */
4012 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4017 lustre_cfg_bufs_set_string(bufs, 1, flags);
4018 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
4023 top->ld_ops->ldo_process_config(env, top, lcfg);
4024 lustre_cfg_free(lcfg);
4026 lu_site_purge(env, top->ld_site, -1);
4028 m->mdt_child = NULL;
4029 m->mdt_bottom = NULL;
4031 obd_disconnect(m->mdt_child_exp);
4032 m->mdt_child_exp = NULL;
4034 obd_disconnect(m->mdt_bottom_exp);
4035 m->mdt_child_exp = NULL;
4038 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
4039 const char *next, struct obd_export **exp)
4041 struct obd_connect_data *data = NULL;
4042 struct obd_device *obd;
4046 OBD_ALLOC_PTR(data);
4048 GOTO(out, rc = -ENOMEM);
4050 obd = class_name2obd(next);
4052 CERROR("%s: can't locate next device: %s\n",
4053 mdt_obd_name(m), next);
4054 GOTO(out, rc = -ENOTCONN);
4057 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4058 data->ocd_version = LUSTRE_VERSION_CODE;
4060 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
4062 CERROR("%s: cannot connect to next dev %s (%d)\n",
4063 mdt_obd_name(m), next, rc);
4073 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
4074 struct lustre_cfg *cfg)
4076 char *dev = lustre_cfg_string(cfg, 0);
4077 int rc, name_size, uuid_size;
4078 char *name, *uuid, *p;
4079 struct lustre_cfg_bufs *bufs;
4080 struct lustre_cfg *lcfg;
4081 struct obd_device *obd;
4082 struct lustre_profile *lprof;
4083 struct lu_site *site;
4086 /* in 1.8 we had the only device in the stack - MDS.
4087 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
4088 * in 2.3 OSD is instantiated by obd_mount.c, so we need
4089 * to generate names and setup MDT, MDD. MDT will be using
4090 * generated name to connect to MDD. for MDD the next device
4091 * will be LOD with name taken from so called "profile" which
4092 * is generated by mount_option line
4094 * 1.8 MGS generates config. commands like this:
4095 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
4096 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
4097 * 2.0 MGS generates config. commands like this:
4098 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
4099 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4100 * 3:lustre-MDT0000-mdtlov 4:f
4102 * we generate MDD name from MDT one, just replacing T with D
4104 * after all the preparations, the logical equivalent will be
4105 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
4106 * 3:lustre-MDT0000-mdtlov 4:f
4107 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4108 * 3:lustre-MDD0000 4:f
4110 * notice we build the stack from down to top: MDD first, then MDT */
4112 name_size = MAX_OBD_NAME;
4113 uuid_size = MAX_OBD_NAME;
4115 OBD_ALLOC(name, name_size);
4116 OBD_ALLOC(uuid, uuid_size);
4117 if (name == NULL || uuid == NULL)
4118 GOTO(cleanup_mem, rc = -ENOMEM);
4120 OBD_ALLOC_PTR(bufs);
4122 GOTO(cleanup_mem, rc = -ENOMEM);
4125 p = strstr(name, "-MDT");
4127 GOTO(free_bufs, rc = -ENOMEM);
4130 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
4132 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4133 if (lprof == NULL || lprof->lp_dt == NULL) {
4134 CERROR("can't find the profile: %s\n",
4135 lustre_cfg_string(cfg, 0));
4136 GOTO(free_bufs, rc = -EINVAL);
4139 lustre_cfg_bufs_reset(bufs, name);
4140 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
4141 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4142 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4144 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4146 GOTO(put_profile, rc = -ENOMEM);
4148 rc = class_attach(lcfg);
4150 GOTO(lcfg_cleanup, rc);
4152 obd = class_name2obd(name);
4154 CERROR("Can not find obd %s (%s in config)\n",
4155 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
4156 GOTO(lcfg_cleanup, rc = -EINVAL);
4159 lustre_cfg_free(lcfg);
4161 lustre_cfg_bufs_reset(bufs, name);
4162 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4163 lustre_cfg_bufs_set_string(bufs, 2, dev);
4164 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4166 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4168 GOTO(class_detach, rc = -ENOMEM);
4170 rc = class_setup(obd, lcfg);
4172 GOTO(class_detach, rc);
4174 /* connect to MDD we just setup */
4175 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
4177 GOTO(class_detach, rc);
4179 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
4181 LASSERT(mdt_lu_site(mdt) == NULL);
4182 mdt->mdt_lu_dev.ld_site = site;
4183 site->ls_top_dev = &mdt->mdt_lu_dev;
4184 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
4186 /* now connect to bottom OSD */
4187 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
4188 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
4190 GOTO(class_detach, rc);
4192 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
4194 rc = lu_env_refill((struct lu_env *)env);
4196 CERROR("Failure to refill session: '%d'\n", rc);
4198 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
4203 class_detach(obd, lcfg);
4205 lustre_cfg_free(lcfg);
4207 class_put_profile(lprof);
4212 OBD_FREE(name, name_size);
4214 OBD_FREE(uuid, uuid_size);
4218 /* setup quota master target on MDT0 */
4219 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
4220 struct lustre_cfg *cfg)
4222 struct obd_device *obd;
4223 char *dev = lustre_cfg_string(cfg, 0);
4224 char *qmtname, *uuid, *p;
4225 struct lustre_cfg_bufs *bufs;
4226 struct lustre_cfg *lcfg;
4227 struct lustre_profile *lprof;
4228 struct obd_connect_data *data;
4232 LASSERT(mdt->mdt_qmt_exp == NULL);
4233 LASSERT(mdt->mdt_qmt_dev == NULL);
4235 /* quota master is on MDT0 only for now */
4236 if (mdt->mdt_seq_site.ss_node_id != 0)
4239 /* MGS generates config commands which look as follows:
4240 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4241 * 3:lustre-MDT0000-mdtlov 4:f
4243 * We generate the QMT name from the MDT one, just replacing MD with QM
4244 * after all the preparations, the logical equivalent will be:
4245 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4246 * 3:lustre-MDT0000-osd 4:f */
4247 OBD_ALLOC(qmtname, MAX_OBD_NAME);
4248 OBD_ALLOC(uuid, UUID_MAX);
4249 OBD_ALLOC_PTR(bufs);
4250 OBD_ALLOC_PTR(data);
4251 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
4252 GOTO(cleanup_mem, rc = -ENOMEM);
4254 strcpy(qmtname, dev);
4255 p = strstr(qmtname, "-MDT");
4257 GOTO(cleanup_mem, rc = -ENOMEM);
4258 /* replace MD with QM */
4262 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4264 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4265 if (lprof == NULL || lprof->lp_dt == NULL) {
4266 CERROR("can't find profile for %s\n",
4267 lustre_cfg_string(cfg, 0));
4268 GOTO(cleanup_mem, rc = -EINVAL);
4271 lustre_cfg_bufs_reset(bufs, qmtname);
4272 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4273 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4274 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4276 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4278 GOTO(put_profile, rc = -ENOMEM);
4280 rc = class_attach(lcfg);
4282 GOTO(lcfg_cleanup, rc);
4284 obd = class_name2obd(qmtname);
4286 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4287 lustre_cfg_string(cfg, 0));
4288 GOTO(lcfg_cleanup, rc = -EINVAL);
4291 lustre_cfg_free(lcfg);
4293 lustre_cfg_bufs_reset(bufs, qmtname);
4294 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4295 lustre_cfg_bufs_set_string(bufs, 2, dev);
4297 /* for quota, the next device should be the OSD device */
4298 lustre_cfg_bufs_set_string(bufs, 3,
4299 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4301 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4303 GOTO(class_detach, rc = -ENOMEM);
4305 rc = class_setup(obd, lcfg);
4307 GOTO(class_detach, rc);
4309 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4311 /* configure local quota objects */
4312 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4316 GOTO(class_cleanup, rc);
4318 /* connect to quota master target */
4319 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4320 data->ocd_version = LUSTRE_VERSION_CODE;
4321 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4324 CERROR("cannot connect to quota master device %s (%d)\n",
4326 GOTO(class_cleanup, rc);
4332 class_manual_cleanup(obd);
4333 mdt->mdt_qmt_dev = NULL;
4337 class_detach(obd, lcfg);
4339 lustre_cfg_free(lcfg);
4341 class_put_profile(lprof);
4346 OBD_FREE(qmtname, MAX_OBD_NAME);
4348 OBD_FREE(uuid, UUID_MAX);
4354 /* Shutdown quota master target associated with mdt */
4355 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4359 if (mdt->mdt_qmt_exp == NULL)
4361 LASSERT(mdt->mdt_qmt_dev != NULL);
4363 /* the qmt automatically shuts down when the mdt disconnects */
4364 obd_disconnect(mdt->mdt_qmt_exp);
4365 mdt->mdt_qmt_exp = NULL;
4366 mdt->mdt_qmt_dev = NULL;
4370 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4371 * for now. This will be removed along with converting rest of MDT code
4372 * to use tgt_session_info */
4373 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4375 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4378 rc = mdt_getxattr(info);
4380 mdt_thread_info_fini(info);
4384 static struct tgt_handler mdt_tgt_handlers[] = {
4385 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4386 0, MDS_CONNECT, mdt_tgt_connect,
4387 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4388 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4389 0, MDS_DISCONNECT, tgt_disconnect,
4390 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4391 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4392 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4393 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4394 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4395 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GET_ROOT, mdt_get_root),
4396 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4397 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4399 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4400 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4401 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4402 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4403 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4404 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4405 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4406 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4408 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4409 mdt_hsm_ct_register),
4410 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4411 mdt_hsm_ct_unregister),
4412 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4414 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4416 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4417 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4419 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4424 static struct tgt_handler mdt_sec_ctx_ops[] = {
4425 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4426 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4427 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4430 static struct tgt_handler mdt_quota_ops[] = {
4431 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4434 static struct tgt_opc_slice mdt_common_slice[] = {
4436 .tos_opc_start = MDS_FIRST_OPC,
4437 .tos_opc_end = MDS_LAST_OPC,
4438 .tos_hs = mdt_tgt_handlers
4441 .tos_opc_start = OBD_FIRST_OPC,
4442 .tos_opc_end = OBD_LAST_OPC,
4443 .tos_hs = tgt_obd_handlers
4446 .tos_opc_start = LDLM_FIRST_OPC,
4447 .tos_opc_end = LDLM_LAST_OPC,
4448 .tos_hs = tgt_dlm_handlers
4451 .tos_opc_start = SEC_FIRST_OPC,
4452 .tos_opc_end = SEC_LAST_OPC,
4453 .tos_hs = mdt_sec_ctx_ops
4456 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4457 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4458 .tos_hs = tgt_out_handlers
4461 .tos_opc_start = FLD_FIRST_OPC,
4462 .tos_opc_end = FLD_LAST_OPC,
4463 .tos_hs = fld_handlers
4466 .tos_opc_start = SEQ_FIRST_OPC,
4467 .tos_opc_end = SEQ_LAST_OPC,
4468 .tos_hs = seq_handlers
4471 .tos_opc_start = QUOTA_DQACQ,
4472 .tos_opc_end = QUOTA_LAST_OPC,
4473 .tos_hs = mdt_quota_ops
4476 .tos_opc_start = LLOG_FIRST_OPC,
4477 .tos_opc_end = LLOG_LAST_OPC,
4478 .tos_hs = tgt_llog_handlers
4481 .tos_opc_start = LFSCK_FIRST_OPC,
4482 .tos_opc_end = LFSCK_LAST_OPC,
4483 .tos_hs = tgt_lfsck_handlers
4491 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4493 struct md_device *next = m->mdt_child;
4494 struct lu_device *d = &m->mdt_lu_dev;
4495 struct obd_device *obd = mdt2obd_dev(m);
4496 struct lfsck_stop stop;
4499 if (m->mdt_md_root != NULL) {
4500 mdt_object_put(env, m->mdt_md_root);
4501 m->mdt_md_root = NULL;
4504 stop.ls_status = LS_PAUSED;
4506 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4508 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4509 target_recovery_fini(obd);
4510 ping_evictor_stop();
4512 if (m->mdt_opts.mo_coordinator)
4513 mdt_hsm_cdt_stop(m);
4515 mdt_hsm_cdt_fini(m);
4517 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4518 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4520 if (m->mdt_namespace != NULL)
4521 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
4522 d->ld_obd->obd_force);
4524 obd_exports_barrier(obd);
4525 obd_zombie_barrier();
4529 tgt_fini(env, &m->mdt_lut);
4530 mdt_fs_cleanup(env, m);
4531 upcall_cache_cleanup(m->mdt_identity_cache);
4532 m->mdt_identity_cache = NULL;
4534 if (m->mdt_namespace != NULL) {
4535 ldlm_namespace_free_post(m->mdt_namespace);
4536 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4539 mdt_quota_fini(env, m);
4541 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4543 mdt_seq_fini(env, m);
4544 mdt_fld_fini(env, m);
4549 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4551 LASSERT(atomic_read(&d->ld_ref) == 0);
4553 server_put_mount(mdt_obd_name(m), true);
4558 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4560 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4561 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4563 struct mdt_thread_info *info;
4564 struct obd_device *obd;
4565 const char *dev = lustre_cfg_string(cfg, 0);
4566 const char *num = lustre_cfg_string(cfg, 2);
4567 struct lustre_mount_info *lmi = NULL;
4568 struct lustre_sb_info *lsi;
4570 struct seq_server_site *ss_site;
4571 const char *identity_upcall = "NONE";
4572 struct md_device *next;
4578 lu_device_init(&m->mdt_lu_dev, ldt);
4580 * Environment (env) might be missing mdt_thread_key values at that
4581 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4584 * Usually device allocation path doesn't use module key values, but
4585 * mdt has to do a lot of work here, so allocate key value.
4587 rc = lu_env_refill((struct lu_env *)env);
4591 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4592 LASSERT(info != NULL);
4594 obd = class_name2obd(dev);
4595 LASSERT(obd != NULL);
4597 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4598 m->mdt_opts.mo_evict_tgt_nids = 1;
4599 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4601 /* default is coordinator off, it is started through conf_param
4603 m->mdt_opts.mo_coordinator = 0;
4605 lmi = server_get_mount(dev);
4607 CERROR("Cannot get mount info for %s!\n", dev);
4610 lsi = s2lsi(lmi->lmi_sb);
4611 /* CMD is supported only in IAM mode */
4613 node_id = simple_strtol(num, NULL, 10);
4614 obd->u.obt.obt_magic = OBT_MAGIC;
4615 if (lsi->lsi_lmd != NULL &&
4616 lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
4617 m->mdt_skip_lfsck = 1;
4620 m->mdt_squash.rsi_uid = 0;
4621 m->mdt_squash.rsi_gid = 0;
4622 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4623 init_rwsem(&m->mdt_squash.rsi_sem);
4624 spin_lock_init(&m->mdt_lock);
4625 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4626 m->mdt_enable_remote_dir = 0;
4627 m->mdt_enable_remote_dir_gid = 0;
4629 atomic_set(&m->mdt_mds_mds_conns, 0);
4630 atomic_set(&m->mdt_async_commit_count, 0);
4632 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4633 m->mdt_lu_dev.ld_obd = obd;
4634 /* Set this lu_device to obd for error handling purposes. */
4635 obd->obd_lu_dev = &m->mdt_lu_dev;
4637 /* init the stack */
4638 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4640 CERROR("%s: Can't init device stack, rc %d\n",
4641 mdt_obd_name(m), rc);
4646 ss_site = mdt_seq_site(m);
4647 s->ld_seq_site = ss_site;
4650 /* set server index */
4651 ss_site->ss_node_id = node_id;
4653 /* failover is the default
4654 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4655 * assumed whose ss_node_id == 0 XXX
4657 obd->obd_replayable = 1;
4658 /* No connection accepted until configurations will finish */
4659 obd->obd_no_conn = 1;
4661 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4662 char *str = lustre_cfg_string(cfg, 4);
4663 if (strchr(str, 'n')) {
4664 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4665 obd->obd_replayable = 0;
4669 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4671 GOTO(err_fini_stack, rc);
4673 rc = mdt_seq_init(env, m);
4675 GOTO(err_fini_fld, rc);
4677 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4678 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4679 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4680 LDLM_NAMESPACE_SERVER,
4681 LDLM_NAMESPACE_GREEDY,
4683 if (m->mdt_namespace == NULL)
4684 GOTO(err_fini_seq, rc = -ENOMEM);
4686 m->mdt_namespace->ns_lvbp = m;
4687 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4689 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4690 /* set obd_namespace for compatibility with old code */
4691 obd->obd_namespace = m->mdt_namespace;
4693 rc = mdt_hsm_cdt_init(m);
4695 CERROR("%s: error initializing coordinator, rc %d\n",
4696 mdt_obd_name(m), rc);
4697 GOTO(err_free_ns, rc);
4700 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4701 OBD_FAIL_MDS_ALL_REQUEST_NET,
4702 OBD_FAIL_MDS_ALL_REPLY_NET);
4704 GOTO(err_free_hsm, rc);
4706 rc = mdt_fs_setup(env, m, obd, lsi);
4710 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4712 next = m->mdt_child;
4713 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4716 GOTO(err_fs_cleanup, rc);
4718 if (mntopts & MNTOPT_USERXATTR)
4719 m->mdt_opts.mo_user_xattr = 1;
4721 m->mdt_opts.mo_user_xattr = 0;
4723 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4725 GOTO(err_fs_cleanup, rc);
4727 if (mntopts & MNTOPT_ACL)
4728 m->mdt_opts.mo_acl = 1;
4730 m->mdt_opts.mo_acl = 0;
4732 /* XXX: to support suppgid for ACL, we enable identity_upcall
4733 * by default, otherwise, maybe got unexpected -EACCESS. */
4734 if (m->mdt_opts.mo_acl)
4735 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4737 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4739 &mdt_identity_upcall_cache_ops);
4740 if (IS_ERR(m->mdt_identity_cache)) {
4741 rc = PTR_ERR(m->mdt_identity_cache);
4742 m->mdt_identity_cache = NULL;
4743 GOTO(err_fs_cleanup, rc);
4746 rc = mdt_procfs_init(m, dev);
4748 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4749 GOTO(err_recovery, rc);
4752 rc = mdt_quota_init(env, m, cfg);
4754 GOTO(err_procfs, rc);
4756 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4757 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4758 "mdt_ldlm_client", m->mdt_ldlm_client);
4760 ping_evictor_start();
4762 /* recovery will be started upon mdt_prepare()
4763 * when the whole stack is complete and ready
4764 * to serve the requests */
4766 /* Reduce the initial timeout on an MDS because it doesn't need such
4767 * a long timeout as an OST does. Adaptive timeouts will adjust this
4768 * value appropriately. */
4769 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4770 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4776 target_recovery_fini(obd);
4777 upcall_cache_cleanup(m->mdt_identity_cache);
4778 m->mdt_identity_cache = NULL;
4780 mdt_fs_cleanup(env, m);
4782 tgt_fini(env, &m->mdt_lut);
4784 mdt_hsm_cdt_fini(m);
4786 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4787 obd->obd_namespace = m->mdt_namespace = NULL;
4789 mdt_seq_fini(env, m);
4791 mdt_fld_fini(env, m);
4793 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4796 server_put_mount(dev, true);
4800 /* For interoperability, the left element is old parameter, the right one
4801 * is the new version of the parameter, if some parameter is deprecated,
4802 * the new version should be set as NULL. */
4803 static struct cfg_interop_param mdt_interop_param[] = {
4804 { "mdt.group_upcall", NULL },
4805 { "mdt.quota_type", NULL },
4806 { "mdd.quota_type", NULL },
4807 { "mdt.som", NULL },
4808 { "mdt.rootsquash", "mdt.root_squash" },
4809 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4813 /* used by MGS to process specific configurations */
4814 static int mdt_process_config(const struct lu_env *env,
4815 struct lu_device *d, struct lustre_cfg *cfg)
4817 struct mdt_device *m = mdt_dev(d);
4818 struct md_device *md_next = m->mdt_child;
4819 struct lu_device *next = md2lu_dev(md_next);
4823 switch (cfg->lcfg_command) {
4825 struct obd_device *obd = d->ld_obd;
4827 /* For interoperability */
4828 struct cfg_interop_param *ptr = NULL;
4829 struct lustre_cfg *old_cfg = NULL;
4832 param = lustre_cfg_string(cfg, 1);
4833 if (param == NULL) {
4834 CERROR("param is empty\n");
4839 ptr = class_find_old_param(param, mdt_interop_param);
4841 if (ptr->new_param == NULL) {
4843 CWARN("For interoperability, skip this %s."
4844 " It is obsolete.\n", ptr->old_param);
4848 CWARN("Found old param %s, changed it to %s.\n",
4849 ptr->old_param, ptr->new_param);
4852 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4859 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4861 if (rc > 0 || rc == -ENOSYS) {
4862 /* is it an HSM var ? */
4863 rc = class_process_proc_param(PARAM_HSM,
4864 hsm_cdt_get_proc_vars(),
4866 if (rc > 0 || rc == -ENOSYS)
4867 /* we don't understand; pass it on */
4868 rc = next->ld_ops->ldo_process_config(env, next,
4872 if (old_cfg != NULL)
4873 lustre_cfg_free(cfg);
4878 /* others are passed further */
4879 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4885 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4886 const struct lu_object_header *hdr,
4887 struct lu_device *d)
4889 struct mdt_object *mo;
4893 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4895 struct lu_object *o;
4896 struct lu_object_header *h;
4899 h = &mo->mot_header;
4900 lu_object_header_init(h);
4901 lu_object_init(o, h, d);
4902 lu_object_add_top(h, o);
4903 o->lo_ops = &mdt_obj_ops;
4904 spin_lock_init(&mo->mot_write_lock);
4905 mutex_init(&mo->mot_lov_mutex);
4906 init_rwsem(&mo->mot_open_sem);
4912 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4913 const struct lu_object_conf *unused)
4915 struct mdt_device *d = mdt_dev(o->lo_dev);
4916 struct lu_device *under;
4917 struct lu_object *below;
4921 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4922 PFID(lu_object_fid(o)));
4924 under = &d->mdt_child->md_lu_dev;
4925 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4926 if (below != NULL) {
4927 lu_object_add(o, below);
4934 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4936 struct mdt_object *mo = mdt_obj(o);
4937 struct lu_object_header *h;
4941 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4942 PFID(lu_object_fid(o)));
4944 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4945 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4948 lu_object_header_fini(h);
4949 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4954 static int mdt_object_print(const struct lu_env *env, void *cookie,
4955 lu_printer_t p, const struct lu_object *o)
4957 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4959 return (*p)(env, cookie,
4960 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
4961 mdto, mdto->mot_lov_created ? "lov_created" : "",
4962 mdto->mot_cache_attr ? "cache_attr" : "",
4963 mdto->mot_write_count);
4966 static int mdt_prepare(const struct lu_env *env,
4967 struct lu_device *pdev,
4968 struct lu_device *cdev)
4970 struct mdt_device *mdt = mdt_dev(cdev);
4971 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4972 struct obd_device *obd = cdev->ld_obd;
4979 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4983 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4987 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4991 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4992 /* The LFSCK instance is registered just now, so it must be there when
4993 * register the namespace to such instance. */
4994 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4996 if (mdt->mdt_seq_site.ss_node_id == 0) {
4997 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4998 &mdt->mdt_md_root_fid);
5003 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
5005 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
5006 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
5007 LASSERT(obd->obd_no_conn);
5008 spin_lock(&obd->obd_dev_lock);
5009 obd->obd_no_conn = 0;
5010 spin_unlock(&obd->obd_dev_lock);
5012 if (obd->obd_recovering == 0)
5013 mdt_postrecov(env, mdt);
5018 const struct lu_device_operations mdt_lu_ops = {
5019 .ldo_object_alloc = mdt_object_alloc,
5020 .ldo_process_config = mdt_process_config,
5021 .ldo_prepare = mdt_prepare,
5024 static const struct lu_object_operations mdt_obj_ops = {
5025 .loo_object_init = mdt_object_init,
5026 .loo_object_free = mdt_object_free,
5027 .loo_object_print = mdt_object_print
5030 static int mdt_obd_set_info_async(const struct lu_env *env,
5031 struct obd_export *exp,
5032 __u32 keylen, void *key,
5033 __u32 vallen, void *val,
5034 struct ptlrpc_request_set *set)
5040 if (KEY_IS(KEY_SPTLRPC_CONF)) {
5041 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
5049 * Match client and server connection feature flags.
5051 * Compute the compatibility flags for a connection request based on
5052 * features mutually supported by client and server.
5054 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
5055 * must not be updated here, otherwise a partially initialized value may
5056 * be exposed. After the connection request is successfully processed,
5057 * the top-level MDT connect request handler atomically updates the export
5058 * connect flags from the obd_connect_data::ocd_connect_flags field of the
5059 * reply. \see mdt_connect().
5061 * \param exp the obd_export associated with this client/target pair
5062 * \param mdt the target device for the connection
5063 * \param data stores data for this connect request
5066 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
5067 * \retval -EBADE client and server feature requirements are incompatible
5069 static int mdt_connect_internal(struct obd_export *exp,
5070 struct mdt_device *mdt,
5071 struct obd_connect_data *data)
5073 LASSERT(data != NULL);
5075 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
5076 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
5077 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
5079 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5080 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
5081 CWARN("%s: client %s does not support ibits lock, either "
5082 "very old or an invalid client: flags "LPX64"\n",
5083 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
5084 data->ocd_connect_flags);
5088 if (!mdt->mdt_opts.mo_acl)
5089 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
5091 if (!mdt->mdt_opts.mo_user_xattr)
5092 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
5094 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
5095 data->ocd_brw_size = min(data->ocd_brw_size,
5096 (__u32)MD_MAX_BRW_SIZE);
5097 if (data->ocd_brw_size == 0) {
5098 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
5099 " ocd_version: %x ocd_grant: %d "
5100 "ocd_index: %u ocd_brw_size is "
5101 "unexpectedly zero, network data "
5102 "corruption? Refusing connection of this"
5105 exp->exp_client_uuid.uuid,
5106 exp, data->ocd_connect_flags, data->ocd_version,
5107 data->ocd_grant, data->ocd_index);
5112 /* NB: Disregard the rule against updating
5113 * exp_connect_data.ocd_connect_flags in this case, since
5114 * tgt_client_new() needs to know if this is a lightweight
5115 * connection, and it is safe to expose this flag before
5116 * connection processing completes. */
5117 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
5118 spin_lock(&exp->exp_lock);
5119 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
5120 spin_unlock(&exp->exp_lock);
5123 data->ocd_version = LUSTRE_VERSION_CODE;
5125 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
5126 CWARN("%s: MDS requires FID support, but client not\n",
5131 if (OCD_HAS_FLAG(data, PINGLESS)) {
5132 if (ptlrpc_pinger_suppress_pings()) {
5133 spin_lock(&exp->exp_obd->obd_dev_lock);
5134 list_del_init(&exp->exp_obd_chain_timed);
5135 spin_unlock(&exp->exp_obd->obd_dev_lock);
5137 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
5141 data->ocd_max_easize = mdt->mdt_max_ea_size;
5143 /* NB: Disregard the rule against updating
5144 * exp_connect_data.ocd_connect_flags in this case, since
5145 * tgt_client_new() needs to know if this is client supports
5146 * multiple modify RPCs, and it is safe to expose this flag before
5147 * connection processing completes. */
5148 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
5149 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
5150 spin_lock(&exp->exp_lock);
5151 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
5152 spin_unlock(&exp->exp_lock);
5158 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
5159 struct mdt_thread_info *info,
5160 struct mdt_file_data *mfd)
5162 struct lu_context ses;
5166 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
5171 lu_context_enter(&ses);
5173 mdt_ucred(info)->uc_valid = UCRED_OLD;
5174 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
5176 lu_context_exit(&ses);
5177 lu_context_fini(&ses);
5183 static int mdt_export_cleanup(struct obd_export *exp)
5185 struct list_head closing_list;
5186 struct mdt_export_data *med = &exp->exp_mdt_data;
5187 struct obd_device *obd = exp->exp_obd;
5188 struct mdt_device *mdt;
5189 struct mdt_thread_info *info;
5191 struct mdt_file_data *mfd, *n;
5195 INIT_LIST_HEAD(&closing_list);
5196 spin_lock(&med->med_open_lock);
5197 while (!list_empty(&med->med_open_head)) {
5198 struct list_head *tmp = med->med_open_head.next;
5199 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
5201 /* Remove mfd handle so it can't be found again.
5202 * We are consuming the mfd_list reference here. */
5203 class_handle_unhash(&mfd->mfd_handle);
5204 list_move_tail(&mfd->mfd_list, &closing_list);
5206 spin_unlock(&med->med_open_lock);
5207 mdt = mdt_dev(obd->obd_lu_dev);
5208 LASSERT(mdt != NULL);
5210 rc = lu_env_init(&env, LCT_MD_THREAD);
5214 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5215 LASSERT(info != NULL);
5216 memset(info, 0, sizeof *info);
5217 info->mti_env = &env;
5218 info->mti_mdt = mdt;
5219 info->mti_exp = exp;
5221 if (!list_empty(&closing_list)) {
5222 struct md_attr *ma = &info->mti_attr;
5224 /* Close any open files (which may also cause orphan
5226 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5227 list_del_init(&mfd->mfd_list);
5228 ma->ma_need = ma->ma_valid = 0;
5230 /* This file is being closed due to an eviction, it
5231 * could have been modified and now dirty regarding to
5232 * HSM archive, check this!
5233 * The logic here is to mark a file dirty if there's a
5234 * chance it was dirtied before the client was evicted,
5235 * so that we don't have to wait for a release attempt
5236 * before finding out the file was actually dirty and
5237 * fail the release. Aggressively marking it dirty here
5238 * will cause the policy engine to attempt to
5239 * re-archive it; when rearchiving, we can compare the
5240 * current version to the HSM data_version and make the
5241 * archive request into a noop if it's not actually
5244 if (mfd->mfd_mode & FMODE_WRITE)
5245 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
5247 /* Don't unlink orphan on failover umount, LU-184 */
5248 if (exp->exp_flags & OBD_OPT_FAILOVER) {
5249 ma->ma_valid = MA_FLAGS;
5250 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
5252 mdt_mfd_close(info, mfd);
5255 info->mti_mdt = NULL;
5256 /* cleanup client slot early */
5257 /* Do not erase record for recoverable client. */
5258 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5259 tgt_client_del(&env, exp);
5265 static inline void mdt_enable_slc(struct mdt_device *mdt)
5267 if (mdt->mdt_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
5268 mdt->mdt_lut.lut_sync_lock_cancel = BLOCKING_SYNC_ON_CANCEL;
5271 static inline void mdt_disable_slc(struct mdt_device *mdt)
5273 if (mdt->mdt_lut.lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL)
5274 mdt->mdt_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
5277 static int mdt_obd_disconnect(struct obd_export *exp)
5283 class_export_get(exp);
5285 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
5286 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
5287 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
5289 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
5290 mdt_disable_slc(mdt);
5293 rc = server_disconnect_export(exp);
5295 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
5297 rc = mdt_export_cleanup(exp);
5298 nodemap_del_member(exp);
5299 class_export_put(exp);
5303 /* mds_connect copy */
5304 static int mdt_obd_connect(const struct lu_env *env,
5305 struct obd_export **exp, struct obd_device *obd,
5306 struct obd_uuid *cluuid,
5307 struct obd_connect_data *data,
5310 struct obd_export *lexp;
5311 struct lustre_handle conn = { 0 };
5312 struct mdt_device *mdt;
5314 lnet_nid_t *client_nid = localdata;
5317 LASSERT(env != NULL);
5318 LASSERT(data != NULL);
5320 if (!exp || !obd || !cluuid)
5323 mdt = mdt_dev(obd->obd_lu_dev);
5325 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5326 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
5327 atomic_inc(&mdt->mdt_mds_mds_conns);
5328 mdt_enable_slc(mdt);
5332 * first, check whether the stack is ready to handle requests
5333 * XXX: probably not very appropriate method is used now
5334 * at some point we should find a better one
5336 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
5337 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5338 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5339 rc = obd_get_info(env, mdt->mdt_child_exp,
5340 sizeof(KEY_OSP_CONNECTED),
5341 KEY_OSP_CONNECTED, NULL, NULL);
5344 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5347 rc = class_connect(&conn, obd, cluuid);
5351 lexp = class_conn2export(&conn);
5352 LASSERT(lexp != NULL);
5354 rc = nodemap_add_member(*client_nid, lexp);
5355 if (rc != 0 && rc != -EEXIST)
5358 rc = mdt_connect_internal(lexp, mdt, data);
5360 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5363 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5364 rc = tgt_client_new(env, lexp);
5366 mdt_export_stats_init(obd, lexp, localdata);
5370 class_disconnect(lexp);
5371 nodemap_del_member(lexp);
5375 /* Because we do not want this export to be evicted by pinger,
5376 * let's not add this export to the timed chain list. */
5377 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
5378 spin_lock(&lexp->exp_obd->obd_dev_lock);
5379 list_del_init(&lexp->exp_obd_chain_timed);
5380 spin_unlock(&lexp->exp_obd->obd_dev_lock);
5387 static int mdt_obd_reconnect(const struct lu_env *env,
5388 struct obd_export *exp, struct obd_device *obd,
5389 struct obd_uuid *cluuid,
5390 struct obd_connect_data *data,
5393 lnet_nid_t *client_nid = localdata;
5397 if (exp == NULL || obd == NULL || cluuid == NULL)
5400 rc = nodemap_add_member(*client_nid, exp);
5401 if (rc != 0 && rc != -EEXIST)
5404 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5406 mdt_export_stats_init(obd, exp, localdata);
5408 nodemap_del_member(exp);
5413 /* FIXME: Can we avoid using these two interfaces? */
5414 static int mdt_init_export(struct obd_export *exp)
5416 struct mdt_export_data *med = &exp->exp_mdt_data;
5420 INIT_LIST_HEAD(&med->med_open_head);
5421 spin_lock_init(&med->med_open_lock);
5422 spin_lock(&exp->exp_lock);
5423 exp->exp_connecting = 1;
5424 spin_unlock(&exp->exp_lock);
5426 /* self-export doesn't need client data and ldlm initialization */
5427 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5428 &exp->exp_client_uuid)))
5431 rc = tgt_client_alloc(exp);
5435 rc = ldlm_init_export(exp);
5442 tgt_client_free(exp);
5444 CERROR("%s: Failed to initialize export: rc = %d\n",
5445 exp->exp_obd->obd_name, rc);
5449 static int mdt_destroy_export(struct obd_export *exp)
5453 target_destroy_export(exp);
5454 /* destroy can be called from failed obd_setup, so
5455 * checking uuid is safer than obd_self_export */
5456 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5457 &exp->exp_client_uuid)))
5460 ldlm_destroy_export(exp);
5461 tgt_client_free(exp);
5463 LASSERT(list_empty(&exp->exp_outstanding_replies));
5464 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5469 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5470 struct linkea_data *ldata)
5474 LASSERT(ldata->ld_buf->lb_buf != NULL);
5476 if (!mdt_object_exists(mdt_obj))
5479 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5480 ldata->ld_buf, XATTR_NAME_LINK);
5481 if (rc == -ERANGE) {
5482 /* Buf was too small, figure out what we need. */
5483 lu_buf_free(ldata->ld_buf);
5484 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5485 ldata->ld_buf, XATTR_NAME_LINK);
5488 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5489 if (ldata->ld_buf->lb_buf == NULL)
5491 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5492 ldata->ld_buf, XATTR_NAME_LINK);
5497 return linkea_init(ldata);
5501 * Given an MDT object, try to look up the full path to the object.
5502 * Part of the MDT layer implementation of lfs fid2path.
5504 * \param[in] info Per-thread common data shared by MDT level handlers.
5505 * \param[in] obj Object to do path lookup of
5506 * \param[in,out] fp User-provided struct to store path information
5507 * \param[in] root_fid Root FID of current path should reach
5509 * \retval 0 Lookup successful, path information stored in fp
5510 * \retval -EAGAIN Lookup failed, usually because object is being moved
5511 * \retval negative errno if there was a problem
5513 static int mdt_path_current(struct mdt_thread_info *info,
5514 struct mdt_object *obj,
5515 struct getinfo_fid2path *fp,
5516 struct lu_fid *root_fid)
5518 struct mdt_device *mdt = info->mti_mdt;
5519 struct mdt_object *mdt_obj;
5520 struct link_ea_header *leh;
5521 struct link_ea_entry *lee;
5522 struct lu_name *tmpname = &info->mti_name;
5523 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5524 struct lu_buf *buf = &info->mti_big_buf;
5527 struct linkea_data ldata = { NULL };
5532 /* temp buffer for path element, the buffer will be finally freed
5533 * in mdt_thread_info_fini */
5534 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5535 if (buf->lb_buf == NULL)
5539 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
5542 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5544 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
5545 struct lu_buf lmv_buf;
5547 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
5548 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
5549 GOTO(out, rc = -ENOENT);
5551 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5552 if (IS_ERR(mdt_obj))
5553 GOTO(out, rc = PTR_ERR(mdt_obj));
5555 if (!mdt_object_exists(mdt_obj)) {
5556 mdt_object_put(info->mti_env, mdt_obj);
5557 GOTO(out, rc = -ENOENT);
5560 if (mdt_object_remote(mdt_obj)) {
5561 mdt_object_put(info->mti_env, mdt_obj);
5562 GOTO(remote_out, rc = -EREMOTE);
5565 rc = mdt_links_read(info, mdt_obj, &ldata);
5567 mdt_object_put(info->mti_env, mdt_obj);
5572 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5573 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5574 /* If set, use link #linkno for path lookup, otherwise use
5575 link #0. Only do this for the final path element. */
5576 if (first && fp->gf_linkno < leh->leh_reccount) {
5578 for (count = 0; count < fp->gf_linkno; count++) {
5579 lee = (struct link_ea_entry *)
5580 ((char *)lee + reclen);
5581 linkea_entry_unpack(lee, &reclen, tmpname,
5584 if (fp->gf_linkno < leh->leh_reccount - 1)
5585 /* indicate to user there are more links */
5589 lmv_buf.lb_buf = info->mti_xattr_buf;
5590 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5591 /* Check if it is slave stripes */
5592 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5593 &lmv_buf, XATTR_NAME_LMV);
5594 mdt_object_put(info->mti_env, mdt_obj);
5596 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5598 /* For slave stripes, get its master */
5599 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5600 fp->gf_fid = *tmpfid;
5603 } else if (rc < 0 && rc != -ENODATA) {
5609 /* Pack the name in the end of the buffer */
5610 ptr -= tmpname->ln_namelen;
5611 if (ptr - 1 <= fp->gf_u.gf_path)
5612 GOTO(out, rc = -EOVERFLOW);
5613 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5616 /* keep the last resolved fid to the client, so the
5617 * client will build the left path on another MDT for
5619 fp->gf_fid = *tmpfid;
5625 ptr++; /* skip leading / */
5626 memmove(fp->gf_u.gf_path, ptr,
5627 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
5634 * Given an MDT object, use mdt_path_current to get the path.
5635 * Essentially a wrapper to retry mdt_path_current a set number of times
5636 * if -EAGAIN is returned (usually because an object is being moved).
5638 * Part of the MDT layer implementation of lfs fid2path.
5640 * \param[in] info Per-thread common data shared by mdt level handlers.
5641 * \param[in] obj Object to do path lookup of
5642 * \param[in,out] fp User-provided struct for arguments and to store path
5645 * \retval 0 Lookup successful, path information stored in fp
5646 * \retval negative errno if there was a problem
5648 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5649 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
5651 struct mdt_device *mdt = info->mti_mdt;
5656 if (fp->gf_pathlen < 3)
5659 if (root_fid == NULL)
5660 root_fid = &mdt->mdt_md_root_fid;
5662 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
5663 fp->gf_u.gf_path[0] = '\0';
5667 /* Retry multiple times in case file is being moved */
5668 while (tries-- && rc == -EAGAIN)
5669 rc = mdt_path_current(info, obj, fp, root_fid);
5675 * Get the full path of the provided FID, as of changelog record recno.
5677 * This checks sanity and looks up object for user provided FID
5678 * before calling the actual path lookup code.
5680 * Part of the MDT layer implementation of lfs fid2path.
5682 * \param[in] info Per-thread common data shared by mdt level handlers.
5683 * \param[in,out] fp User-provided struct for arguments and to store path
5686 * \retval 0 Lookup successful, path information and recno stored in fp
5687 * \retval -ENOENT, object does not exist
5688 * \retval negative errno if there was a problem
5690 static int mdt_fid2path(struct mdt_thread_info *info,
5691 struct lu_fid *root_fid,
5692 struct getinfo_fid2path *fp)
5694 struct mdt_device *mdt = info->mti_mdt;
5695 struct mdt_object *obj;
5699 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5700 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5702 if (!fid_is_sane(&fp->gf_fid))
5705 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5706 CWARN("%s: "DFID" is invalid, sequence should be "
5707 ">= "LPX64"\n", mdt_obd_name(mdt),
5708 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5712 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5715 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5716 PFID(&fp->gf_fid), rc);
5720 if (mdt_object_remote(obj))
5722 else if (!mdt_object_exists(obj))
5728 mdt_object_put(info->mti_env, obj);
5729 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5730 PFID(&fp->gf_fid), rc);
5734 rc = mdt_path(info, obj, fp, root_fid);
5736 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5737 PFID(&fp->gf_fid), fp->gf_u.gf_path,
5738 fp->gf_recno, fp->gf_linkno);
5740 mdt_object_put(info->mti_env, obj);
5745 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
5746 void *val, int vallen)
5748 struct getinfo_fid2path *fpout, *fpin;
5749 struct lu_fid *root_fid = NULL;
5752 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5755 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5756 lustre_swab_fid2path(fpin);
5758 memcpy(fpout, fpin, sizeof(*fpin));
5759 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5762 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
5763 sizeof(struct lu_fid)) {
5764 /* client sent its root FID, which is normally fileset FID */
5765 root_fid = fpin->gf_u.gf_root_fid;
5766 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5767 lustre_swab_lu_fid(root_fid);
5769 if (root_fid != NULL && !fid_is_sane(root_fid))
5773 rc = mdt_fid2path(info, root_fid, fpout);
5777 int mdt_get_info(struct tgt_session_info *tsi)
5787 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5789 CDEBUG(D_IOCTL, "No GETINFO key\n");
5790 RETURN(err_serious(-EFAULT));
5792 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5795 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5796 if (vallen == NULL) {
5797 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5798 tgt_name(tsi->tsi_tgt));
5799 RETURN(err_serious(-EFAULT));
5802 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5804 rc = req_capsule_server_pack(tsi->tsi_pill);
5806 RETURN(err_serious(rc));
5808 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5809 if (valout == NULL) {
5810 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5811 tgt_name(tsi->tsi_tgt));
5812 RETURN(err_serious(-EFAULT));
5815 if (KEY_IS(KEY_FID2PATH)) {
5816 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5818 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
5819 mdt_thread_info_fini(info);
5826 /* Pass the ioc down */
5827 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5828 unsigned int cmd, int len, void *data)
5830 struct lu_context ioctl_session;
5831 struct md_device *next = mdt->mdt_child;
5835 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5838 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5839 lu_context_enter(&ioctl_session);
5840 env->le_ses = &ioctl_session;
5842 LASSERT(next->md_ops->mdo_iocontrol);
5843 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5845 lu_context_exit(&ioctl_session);
5846 lu_context_fini(&ioctl_session);
5850 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5852 struct obd_ioctl_data *data = karg;
5855 struct mdt_object *obj;
5856 struct mdt_lock_handle *lh;
5860 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5861 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5864 fid = (struct lu_fid *)data->ioc_inlbuf1;
5866 if (!fid_is_sane(fid))
5869 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5871 lh = &mti->mti_lh[MDT_LH_PARENT];
5872 mdt_lock_reg_init(lh, LCK_CR);
5874 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5876 RETURN(PTR_ERR(obj));
5878 if (mdt_object_remote(obj)) {
5881 * before calling version get the correct MDS should be
5882 * fid, this is error to find remote object here
5884 CERROR("nonlocal object "DFID"\n", PFID(fid));
5885 } else if (!mdt_object_exists(obj)) {
5886 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5889 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5890 *(__u64 *)data->ioc_inlbuf2 = version;
5893 mdt_object_unlock_put(mti, obj, lh, 1);
5897 /* ioctls on obd dev */
5898 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5899 void *karg, void __user *uarg)
5902 struct obd_device *obd = exp->exp_obd;
5903 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5904 struct dt_device *dt = mdt->mdt_bottom;
5908 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5909 rc = lu_env_init(&env, LCT_MD_THREAD);
5915 rc = mdt_device_sync(&env, mdt);
5917 case OBD_IOC_SET_READONLY:
5918 rc = dt_sync(&env, dt);
5920 rc = dt_ro(&env, dt);
5922 case OBD_IOC_ABORT_RECOVERY:
5923 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5924 obd->obd_abort_recovery = 1;
5925 target_stop_recovery_thread(obd);
5928 case OBD_IOC_CHANGELOG_REG:
5929 case OBD_IOC_CHANGELOG_DEREG:
5930 case OBD_IOC_CHANGELOG_CLEAR:
5931 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5933 case OBD_IOC_START_LFSCK: {
5934 struct md_device *next = mdt->mdt_child;
5935 struct obd_ioctl_data *data = karg;
5936 struct lfsck_start_param lsp;
5938 if (unlikely(data == NULL)) {
5943 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5944 lsp.lsp_index_valid = 0;
5945 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5948 case OBD_IOC_STOP_LFSCK: {
5949 struct md_device *next = mdt->mdt_child;
5950 struct obd_ioctl_data *data = karg;
5951 struct lfsck_stop stop;
5953 stop.ls_status = LS_STOPPED;
5954 /* Old lfsck utils may pass NULL @stop. */
5955 if (data->ioc_inlbuf1 == NULL)
5959 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5961 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5964 case OBD_IOC_QUERY_LFSCK: {
5965 struct md_device *next = mdt->mdt_child;
5966 struct obd_ioctl_data *data = karg;
5968 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
5972 case OBD_IOC_GET_OBJ_VERSION: {
5973 struct mdt_thread_info *mti;
5974 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5975 memset(mti, 0, sizeof *mti);
5976 mti->mti_env = &env;
5980 rc = mdt_ioc_version_get(mti, karg);
5983 case OBD_IOC_CATLOGLIST: {
5984 struct mdt_thread_info *mti;
5986 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5987 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5988 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5989 &mti->mti_tmp_fid1);
5994 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5995 mdt_obd_name(mdt), cmd, rc);
6002 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
6004 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
6008 if (!mdt->mdt_skip_lfsck) {
6009 struct lfsck_start_param lsp;
6011 lsp.lsp_start = NULL;
6012 lsp.lsp_index_valid = 0;
6013 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
6014 OBD_IOC_START_LFSCK,
6016 if (rc != 0 && rc != -EALREADY)
6017 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
6018 mdt_obd_name(mdt), rc);
6021 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
6025 static int mdt_obd_postrecov(struct obd_device *obd)
6030 rc = lu_env_init(&env, LCT_MD_THREAD);
6033 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
6038 static struct obd_ops mdt_obd_device_ops = {
6039 .o_owner = THIS_MODULE,
6040 .o_set_info_async = mdt_obd_set_info_async,
6041 .o_connect = mdt_obd_connect,
6042 .o_reconnect = mdt_obd_reconnect,
6043 .o_disconnect = mdt_obd_disconnect,
6044 .o_init_export = mdt_init_export,
6045 .o_destroy_export = mdt_destroy_export,
6046 .o_iocontrol = mdt_iocontrol,
6047 .o_postrecov = mdt_obd_postrecov,
6050 static struct lu_device* mdt_device_fini(const struct lu_env *env,
6051 struct lu_device *d)
6053 struct mdt_device *m = mdt_dev(d);
6060 static struct lu_device *mdt_device_free(const struct lu_env *env,
6061 struct lu_device *d)
6063 struct mdt_device *m = mdt_dev(d);
6066 lu_device_fini(&m->mdt_lu_dev);
6072 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
6073 struct lu_device_type *t,
6074 struct lustre_cfg *cfg)
6076 struct lu_device *l;
6077 struct mdt_device *m;
6084 rc = mdt_init0(env, m, t, cfg);
6086 mdt_device_free(env, l);
6091 l = ERR_PTR(-ENOMEM);
6095 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
6096 LU_KEY_INIT(mdt, struct mdt_thread_info);
6098 static void mdt_key_fini(const struct lu_context *ctx,
6099 struct lu_context_key *key, void* data)
6101 struct mdt_thread_info *info = data;
6103 if (info->mti_big_lmm) {
6104 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
6105 info->mti_big_lmm = NULL;
6106 info->mti_big_lmmsize = 0;
6111 /* context key: mdt_thread_key */
6112 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
6114 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
6116 return lu_ucred(info->mti_env);
6119 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
6121 return lu_ucred_check(info->mti_env);
6125 * Enable/disable COS (Commit On Sharing).
6127 * Set/Clear the COS flag in mdt options.
6129 * \param mdt mdt device
6130 * \param val 0 disables COS, other values enable COS
6132 void mdt_enable_cos(struct mdt_device *mdt, int val)
6137 mdt->mdt_opts.mo_cos = !!val;
6138 rc = lu_env_init(&env, LCT_LOCAL);
6139 if (unlikely(rc != 0)) {
6140 CWARN("%s: lu_env initialization failed, cannot "
6141 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
6144 mdt_device_sync(&env, mdt);
6149 * Check COS (Commit On Sharing) status.
6151 * Return COS flag status.
6153 * \param mdt mdt device
6155 int mdt_cos_is_enabled(struct mdt_device *mdt)
6157 return mdt->mdt_opts.mo_cos != 0;
6160 static struct lu_device_type_operations mdt_device_type_ops = {
6161 .ldto_device_alloc = mdt_device_alloc,
6162 .ldto_device_free = mdt_device_free,
6163 .ldto_device_fini = mdt_device_fini
6166 static struct lu_device_type mdt_device_type = {
6167 .ldt_tags = LU_DEVICE_MD,
6168 .ldt_name = LUSTRE_MDT_NAME,
6169 .ldt_ops = &mdt_device_type_ops,
6170 .ldt_ctx_tags = LCT_MD_THREAD
6173 static int __init mdt_init(void)
6177 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
6178 FID_NOBRACE_LEN + 1);
6179 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
6181 rc = lu_kmem_init(mdt_caches);
6185 rc = mds_mod_init();
6189 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
6190 LUSTRE_MDT_NAME, &mdt_device_type);
6195 lu_kmem_fini(mdt_caches);
6202 static void __exit mdt_exit(void)
6204 class_unregister_type(LUSTRE_MDT_NAME);
6206 lu_kmem_fini(mdt_caches);
6209 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
6210 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
6211 MODULE_VERSION(LUSTRE_VERSION_STRING);
6212 MODULE_LICENSE("GPL");
6214 module_init(mdt_init);
6215 module_exit(mdt_exit);