4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdt/mdt_handler.c
34 * Lustre Metadata Target (mdt) request handler
36 * Author: Peter Braam <braam@clusterfs.com>
37 * Author: Andreas Dilger <adilger@clusterfs.com>
38 * Author: Phil Schwan <phil@clusterfs.com>
39 * Author: Mike Shaver <shaver@clusterfs.com>
40 * Author: Nikita Danilov <nikita@clusterfs.com>
41 * Author: Huang Hua <huanghua@clusterfs.com>
42 * Author: Yury Umanets <umka@clusterfs.com>
45 #define DEBUG_SUBSYSTEM S_MDS
47 #include <linux/module.h>
48 #include <linux/pagemap.h>
50 #include <dt_object.h>
51 #include <lustre_acl.h>
52 #include <lustre_export.h>
53 #include <lustre_ioctl.h>
54 #include <lustre_lfsck.h>
55 #include <lustre_log.h>
56 #include <lustre_net.h>
57 #include <lustre_nodemap.h>
58 #include <lustre_mds.h>
59 #include <lustre_param.h>
60 #include <lustre_quota.h>
61 #include <lustre_swab.h>
63 #include <obd_support.h>
65 #include <llog_swab.h>
67 #include "mdt_internal.h"
70 static unsigned int max_mod_rpcs_per_client = 8;
71 module_param(max_mod_rpcs_per_client, uint, 0644);
72 MODULE_PARM_DESC(max_mod_rpcs_per_client, "maximum number of modify RPCs in flight allowed per client");
74 mdl_mode_t mdt_mdl_lock_modes[] = {
75 [LCK_MINMODE] = MDL_MINMODE,
82 [LCK_GROUP] = MDL_GROUP
85 enum ldlm_mode mdt_dlm_lock_modes[] = {
86 [MDL_MINMODE] = LCK_MINMODE,
93 [MDL_GROUP] = LCK_GROUP
96 static struct mdt_device *mdt_dev(struct lu_device *d);
97 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
99 static const struct lu_object_operations mdt_obj_ops;
101 /* Slab for MDT object allocation */
102 static struct kmem_cache *mdt_object_kmem;
104 /* For HSM restore handles */
105 struct kmem_cache *mdt_hsm_cdt_kmem;
107 /* For HSM request handles */
108 struct kmem_cache *mdt_hsm_car_kmem;
110 static struct lu_kmem_descr mdt_caches[] = {
112 .ckd_cache = &mdt_object_kmem,
113 .ckd_name = "mdt_obj",
114 .ckd_size = sizeof(struct mdt_object)
117 .ckd_cache = &mdt_hsm_cdt_kmem,
118 .ckd_name = "mdt_cdt_restore_handle",
119 .ckd_size = sizeof(struct cdt_restore_handle)
122 .ckd_cache = &mdt_hsm_car_kmem,
123 .ckd_name = "mdt_cdt_agent_req",
124 .ckd_size = sizeof(struct cdt_agent_req)
131 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
135 return rep->lock_policy_res1 & op_flag;
138 void mdt_clear_disposition(struct mdt_thread_info *info,
139 struct ldlm_reply *rep, __u64 op_flag)
142 info->mti_opdata &= ~op_flag;
143 tgt_opdata_clear(info->mti_env, op_flag);
146 rep->lock_policy_res1 &= ~op_flag;
149 void mdt_set_disposition(struct mdt_thread_info *info,
150 struct ldlm_reply *rep, __u64 op_flag)
153 info->mti_opdata |= op_flag;
154 tgt_opdata_set(info->mti_env, op_flag);
157 rep->lock_policy_res1 |= op_flag;
160 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
162 lh->mlh_pdo_hash = 0;
163 lh->mlh_reg_mode = lm;
164 lh->mlh_rreg_mode = lm;
165 lh->mlh_type = MDT_REG_LOCK;
168 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
169 const struct lu_name *lname)
171 lh->mlh_reg_mode = lock_mode;
172 lh->mlh_pdo_mode = LCK_MINMODE;
173 lh->mlh_rreg_mode = lock_mode;
174 lh->mlh_type = MDT_PDO_LOCK;
176 if (lu_name_is_valid(lname)) {
177 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
179 /* XXX Workaround for LU-2856
181 * Zero is a valid return value of full_name_hash, but
182 * several users of mlh_pdo_hash assume a non-zero
183 * hash value. We therefore map zero onto an
184 * arbitrary, but consistent value (1) to avoid
185 * problems further down the road. */
186 if (unlikely(lh->mlh_pdo_hash == 0))
187 lh->mlh_pdo_hash = 1;
189 lh->mlh_pdo_hash = 0;
193 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
194 struct mdt_lock_handle *lh)
200 * Any dir access needs couple of locks:
202 * 1) on part of dir we gonna take lookup/modify;
204 * 2) on whole dir to protect it from concurrent splitting and/or to
205 * flush client's cache for readdir().
207 * so, for a given mode and object this routine decides what lock mode
208 * to use for lock #2:
210 * 1) if caller's gonna lookup in dir then we need to protect dir from
211 * being splitted only - LCK_CR
213 * 2) if caller's gonna modify dir then we need to protect dir from
214 * being splitted and to flush cache - LCK_CW
216 * 3) if caller's gonna modify dir and that dir seems ready for
217 * splitting then we need to protect it from any type of access
218 * (lookup/modify/split) - LCK_EX --bzzz
221 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
222 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
225 * Ask underlaying level its opinion about preferable PDO lock mode
226 * having access type passed as regular lock mode:
228 * - MDL_MINMODE means that lower layer does not want to specify lock
231 * - MDL_NL means that no PDO lock should be taken. This is used in some
232 * cases. Say, for non-splittable directories no need to use PDO locks
235 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
236 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
238 if (mode != MDL_MINMODE) {
239 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
242 * Lower layer does not want to specify locking mode. We do it
243 * our selves. No special protection is needed, just flush
244 * client's cache on modification and allow concurrent
247 switch (lh->mlh_reg_mode) {
249 lh->mlh_pdo_mode = LCK_EX;
252 lh->mlh_pdo_mode = LCK_CR;
255 lh->mlh_pdo_mode = LCK_CW;
258 CERROR("Not expected lock type (0x%x)\n",
259 (int)lh->mlh_reg_mode);
264 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
268 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
271 struct mdt_device *mdt = info->mti_mdt;
272 struct lu_name *lname = &info->mti_name;
274 struct mdt_object *parent;
278 LASSERT(!info->mti_cross_ref);
280 OBD_ALLOC(name, NAME_MAX + 1);
283 lname->ln_name = name;
286 * We may want to allow this to mount a completely separate
287 * fileset from the MDT in the future, but keeping it to
288 * ROOT/ only for now avoid potential security issues.
290 *fid = mdt->mdt_md_root_fid;
292 while (rc == 0 && fileset != NULL && *fileset != '\0') {
293 const char *s1 = fileset;
299 while (*s2 != '/' && *s2 != '\0')
307 lname->ln_namelen = s2 - s1;
308 if (lname->ln_namelen > NAME_MAX) {
313 /* reject .. as a path component */
314 if (lname->ln_namelen == 2 &&
315 strncmp(s1, "..", 2) == 0) {
320 strncpy(name, s1, lname->ln_namelen);
321 name[lname->ln_namelen] = '\0';
323 parent = mdt_object_find(info->mti_env, mdt, fid);
324 if (IS_ERR(parent)) {
325 rc = PTR_ERR(parent);
328 /* Only got the fid of this obj by name */
330 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
331 fid, &info->mti_spec);
332 mdt_object_put(info->mti_env, parent);
335 parent = mdt_object_find(info->mti_env, mdt, fid);
337 rc = PTR_ERR(parent);
339 mode = lu_object_attr(&parent->mot_obj);
340 mdt_object_put(info->mti_env, parent);
346 OBD_FREE(name, NAME_MAX + 1);
351 static int mdt_get_root(struct tgt_session_info *tsi)
353 struct mdt_thread_info *info = tsi2mdt_info(tsi);
354 struct mdt_device *mdt = info->mti_mdt;
355 struct mdt_body *repbody;
356 char *fileset = NULL, *buffer = NULL;
358 struct obd_export *exp = info->mti_exp;
359 char *nodemap_fileset;
363 rc = mdt_check_ucred(info);
365 GOTO(out, rc = err_serious(rc));
367 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
368 GOTO(out, rc = err_serious(-ENOMEM));
370 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
371 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
372 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
374 GOTO(out, rc = err_serious(-EFAULT));
377 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
378 if (nodemap_fileset && nodemap_fileset[0]) {
379 CDEBUG(D_INFO, "nodemap fileset is %s\n", nodemap_fileset);
381 /* consider fileset from client as a sub-fileset
382 * of the nodemap one */
383 OBD_ALLOC(buffer, PATH_MAX + 1);
385 GOTO(out, rc = err_serious(-ENOMEM));
386 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
387 nodemap_fileset, fileset) >= PATH_MAX + 1)
388 GOTO(out, rc = err_serious(-EINVAL));
391 /* enforce fileset as specified in the nodemap */
392 fileset = nodemap_fileset;
397 CDEBUG(D_INFO, "Getting fileset %s\n", fileset);
398 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
400 GOTO(out, rc = err_serious(rc));
402 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
404 repbody->mbo_valid |= OBD_MD_FLID;
408 mdt_thread_info_fini(info);
410 OBD_FREE(buffer, PATH_MAX+1);
414 static int mdt_statfs(struct tgt_session_info *tsi)
416 struct ptlrpc_request *req = tgt_ses_req(tsi);
417 struct mdt_thread_info *info = tsi2mdt_info(tsi);
418 struct md_device *next = info->mti_mdt->mdt_child;
419 struct ptlrpc_service_part *svcpt;
420 struct obd_statfs *osfs;
425 svcpt = req->rq_rqbd->rqbd_svcpt;
427 /* This will trigger a watchdog timeout */
428 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
429 (MDT_SERVICE_WATCHDOG_FACTOR *
430 at_get(&svcpt->scp_at_estimate)) + 1);
432 rc = mdt_check_ucred(info);
434 GOTO(out, rc = err_serious(rc));
436 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
437 GOTO(out, rc = err_serious(-ENOMEM));
439 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
441 GOTO(out, rc = -EPROTO);
443 /** statfs information are cached in the mdt_device */
444 if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
445 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
446 /** statfs data is too old, get up-to-date one */
447 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
450 spin_lock(&info->mti_mdt->mdt_lock);
451 info->mti_mdt->mdt_osfs = *osfs;
452 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
453 spin_unlock(&info->mti_mdt->mdt_lock);
455 /** use cached statfs data */
456 spin_lock(&info->mti_mdt->mdt_lock);
457 *osfs = info->mti_mdt->mdt_osfs;
458 spin_unlock(&info->mti_mdt->mdt_lock);
462 mdt_counter_incr(req, LPROC_MDT_STATFS);
464 mdt_thread_info_fini(info);
468 #ifdef CONFIG_FS_POSIX_ACL
470 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
472 * \param info thread info object
473 * \param repbody reply to pack ACLs into
474 * \param o mdt object of file to examine
475 * \param nodemap nodemap of client to reply to
477 * \retval -errno error getting or parsing ACL from disk
479 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
480 struct mdt_object *o, struct lu_nodemap *nodemap)
482 const struct lu_env *env = info->mti_env;
483 struct md_object *next = mdt_object_child(o);
484 struct lu_buf *buf = &info->mti_buf;
487 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
488 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
490 if (buf->lb_len == 0)
493 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
495 if (rc == -ENODATA) {
496 repbody->mbo_aclsize = 0;
497 repbody->mbo_valid |= OBD_MD_FLACL;
499 } else if (rc == -EOPNOTSUPP) {
502 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
503 mdt_obd_name(info->mti_mdt),
504 PFID(mdt_object_fid(o)), rc);
507 rc = nodemap_map_acl(nodemap, buf->lb_buf,
508 rc, NODEMAP_FS_TO_CLIENT);
509 /* if all ACLs mapped out, rc is still >= 0 */
511 CERROR("%s: nodemap_map_acl unable to parse "DFID
512 " ACL: rc = %d\n", mdt_obd_name(info->mti_mdt),
513 PFID(mdt_object_fid(o)), rc);
515 repbody->mbo_aclsize = rc;
516 repbody->mbo_valid |= OBD_MD_FLACL;
524 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
525 const struct lu_attr *attr, const struct lu_fid *fid)
527 struct md_attr *ma = &info->mti_attr;
528 struct obd_export *exp = info->mti_exp;
529 struct lu_nodemap *nodemap = NULL;
531 LASSERT(ma->ma_valid & MA_INODE);
533 if (attr->la_valid & LA_ATIME) {
534 b->mbo_atime = attr->la_atime;
535 b->mbo_valid |= OBD_MD_FLATIME;
537 if (attr->la_valid & LA_MTIME) {
538 b->mbo_mtime = attr->la_mtime;
539 b->mbo_valid |= OBD_MD_FLMTIME;
541 if (attr->la_valid & LA_CTIME) {
542 b->mbo_ctime = attr->la_ctime;
543 b->mbo_valid |= OBD_MD_FLCTIME;
545 if (attr->la_valid & LA_FLAGS) {
546 b->mbo_flags = attr->la_flags;
547 b->mbo_valid |= OBD_MD_FLFLAGS;
549 if (attr->la_valid & LA_NLINK) {
550 b->mbo_nlink = attr->la_nlink;
551 b->mbo_valid |= OBD_MD_FLNLINK;
553 if (attr->la_valid & (LA_UID|LA_GID)) {
554 nodemap = nodemap_get_from_exp(exp);
558 if (attr->la_valid & LA_UID) {
559 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
560 NODEMAP_FS_TO_CLIENT,
562 b->mbo_valid |= OBD_MD_FLUID;
564 if (attr->la_valid & LA_GID) {
565 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
566 NODEMAP_FS_TO_CLIENT,
568 b->mbo_valid |= OBD_MD_FLGID;
571 b->mbo_mode = attr->la_mode;
572 if (attr->la_valid & LA_MODE)
573 b->mbo_valid |= OBD_MD_FLMODE;
574 if (attr->la_valid & LA_TYPE)
575 b->mbo_valid |= OBD_MD_FLTYPE;
579 b->mbo_valid |= OBD_MD_FLID;
580 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
581 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
584 if (!(attr->la_valid & LA_TYPE))
587 b->mbo_rdev = attr->la_rdev;
588 b->mbo_size = attr->la_size;
589 b->mbo_blocks = attr->la_blocks;
591 if (!S_ISREG(attr->la_mode)) {
592 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
593 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
594 /* means no objects are allocated on osts. */
595 LASSERT(!(ma->ma_valid & MA_LOV));
596 /* just ignore blocks occupied by extend attributes on MDS */
598 /* if no object is allocated on osts, the size on mds is valid.
600 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
601 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
602 ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
603 /* A released file stores its size on MDS. */
604 /* But return 1 block for released file, unless tools like tar
605 * will consider it fully sparse. (LU-3864)
607 if (unlikely(b->mbo_size == 0))
611 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
614 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
615 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
616 PFID(fid), (unsigned long long)b->mbo_size);
619 if (!IS_ERR_OR_NULL(nodemap))
620 nodemap_putref(nodemap);
623 static inline int mdt_body_has_lov(const struct lu_attr *la,
624 const struct mdt_body *body)
626 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
627 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
630 void mdt_client_compatibility(struct mdt_thread_info *info)
632 struct mdt_body *body;
633 struct ptlrpc_request *req = mdt_info_req(info);
634 struct obd_export *exp = req->rq_export;
635 struct md_attr *ma = &info->mti_attr;
636 struct lu_attr *la = &ma->ma_attr;
639 if (exp_connect_layout(exp))
640 /* the client can deal with 16-bit lmm_stripe_count */
643 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
645 if (!mdt_body_has_lov(la, body))
648 /* now we have a reply with a lov for a client not compatible with the
649 * layout lock so we have to clean the layout generation number */
650 if (S_ISREG(la->la_mode))
651 ma->ma_lmm->lmm_layout_gen = 0;
655 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
656 struct mdt_object *o)
658 const struct lu_env *env = info->mti_env;
661 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
670 /* Is it a directory? Let's check for the LMV as well */
671 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
672 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
676 rc2 = mo_xattr_get(env, mdt_object_child(o),
678 XATTR_NAME_DEFAULT_LMV);
680 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
688 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
691 const struct lu_env *env = info->mti_env;
695 LASSERT(info->mti_big_lmm_used == 0);
696 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
700 /* big_lmm may need to be grown */
701 if (info->mti_big_lmmsize < rc) {
702 int size = size_roundup_power2(rc);
704 if (info->mti_big_lmmsize > 0) {
705 /* free old buffer */
706 LASSERT(info->mti_big_lmm);
707 OBD_FREE_LARGE(info->mti_big_lmm,
708 info->mti_big_lmmsize);
709 info->mti_big_lmm = NULL;
710 info->mti_big_lmmsize = 0;
713 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
714 if (info->mti_big_lmm == NULL)
716 info->mti_big_lmmsize = size;
718 LASSERT(info->mti_big_lmmsize >= rc);
720 info->mti_buf.lb_buf = info->mti_big_lmm;
721 info->mti_buf.lb_len = info->mti_big_lmmsize;
722 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
727 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
728 struct md_attr *ma, const char *name)
730 struct md_object *next = mdt_object_child(o);
731 struct lu_buf *buf = &info->mti_buf;
734 if (strcmp(name, XATTR_NAME_LOV) == 0) {
735 buf->lb_buf = ma->ma_lmm;
736 buf->lb_len = ma->ma_lmm_size;
737 LASSERT(!(ma->ma_valid & MA_LOV));
738 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
739 buf->lb_buf = ma->ma_lmv;
740 buf->lb_len = ma->ma_lmv_size;
741 LASSERT(!(ma->ma_valid & MA_LMV));
742 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
743 buf->lb_buf = ma->ma_lmv;
744 buf->lb_len = ma->ma_lmv_size;
745 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
750 rc = mo_xattr_get(info->mti_env, next, buf, name);
754 if (strcmp(name, XATTR_NAME_LOV) == 0) {
755 if (info->mti_big_lmm_used)
756 ma->ma_lmm = info->mti_big_lmm;
758 /* NOT return LOV EA with hole to old client. */
759 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
760 LOV_PATTERN_F_HOLE) &&
761 !(exp_connect_flags(info->mti_exp) &
762 OBD_CONNECT_LFSCK)) {
765 ma->ma_lmm_size = rc;
766 ma->ma_valid |= MA_LOV;
768 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
769 if (info->mti_big_lmm_used)
770 ma->ma_lmv = info->mti_big_lmm;
772 ma->ma_lmv_size = rc;
773 ma->ma_valid |= MA_LMV;
774 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
775 ma->ma_lmv_size = rc;
776 ma->ma_valid |= MA_LMV_DEF;
779 /* Update mdt_max_mdsize so all clients will be aware that */
780 if (info->mti_mdt->mdt_max_mdsize < rc)
781 info->mti_mdt->mdt_max_mdsize = rc;
784 } else if (rc == -ENODATA) {
787 } else if (rc == -ERANGE) {
788 /* Default LMV has fixed size, so it must be able to fit
789 * in the original buffer */
790 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
792 rc = mdt_big_xattr_get(info, o, name);
794 info->mti_big_lmm_used = 1;
802 static int mdt_attr_get_pfid(struct mdt_thread_info *info,
803 struct mdt_object *o, struct lu_fid *pfid)
805 struct lu_buf *buf = &info->mti_buf;
806 struct link_ea_header *leh;
807 struct link_ea_entry *lee;
811 buf->lb_buf = info->mti_big_lmm;
812 buf->lb_len = info->mti_big_lmmsize;
813 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
814 buf, XATTR_NAME_LINK);
815 /* ignore errors, MA_PFID won't be set and it is
816 * up to the caller to treat this as an error */
817 if (rc == -ERANGE || buf->lb_len == 0) {
818 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
819 buf->lb_buf = info->mti_big_lmm;
820 buf->lb_len = info->mti_big_lmmsize;
825 if (rc < sizeof(*leh)) {
826 CERROR("short LinkEA on "DFID": rc = %d\n",
827 PFID(mdt_object_fid(o)), rc);
831 leh = (struct link_ea_header *) buf->lb_buf;
832 lee = (struct link_ea_entry *)(leh + 1);
833 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
834 leh->leh_magic = LINK_EA_MAGIC;
835 leh->leh_reccount = __swab32(leh->leh_reccount);
836 leh->leh_len = __swab64(leh->leh_len);
838 if (leh->leh_magic != LINK_EA_MAGIC)
840 if (leh->leh_reccount == 0)
843 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
844 fid_be_to_cpu(pfid, pfid);
849 int mdt_attr_get_complex(struct mdt_thread_info *info,
850 struct mdt_object *o, struct md_attr *ma)
852 const struct lu_env *env = info->mti_env;
853 struct md_object *next = mdt_object_child(o);
854 struct lu_buf *buf = &info->mti_buf;
855 int need = ma->ma_need;
862 if (mdt_object_exists(o) == 0)
863 GOTO(out, rc = -ENOENT);
864 mode = lu_object_attr(&next->mo_lu);
866 if (need & MA_INODE) {
867 ma->ma_need = MA_INODE;
868 rc = mo_attr_get(env, next, ma);
871 ma->ma_valid |= MA_INODE;
874 if (need & MA_PFID) {
875 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
877 ma->ma_valid |= MA_PFID;
878 /* ignore this error, parent fid is not mandatory */
882 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
883 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
888 if (need & MA_LMV && S_ISDIR(mode)) {
889 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
894 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
895 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
900 if (need & MA_HSM && S_ISREG(mode)) {
901 buf->lb_buf = info->mti_xattr_buf;
902 buf->lb_len = sizeof(info->mti_xattr_buf);
903 CLASSERT(sizeof(struct hsm_attrs) <=
904 sizeof(info->mti_xattr_buf));
905 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
906 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
908 ma->ma_valid |= MA_HSM;
909 else if (rc2 < 0 && rc2 != -ENODATA)
913 #ifdef CONFIG_FS_POSIX_ACL
914 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
915 buf->lb_buf = ma->ma_acl;
916 buf->lb_len = ma->ma_acl_size;
917 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
919 ma->ma_acl_size = rc2;
920 ma->ma_valid |= MA_ACL_DEF;
921 } else if (rc2 == -ENODATA) {
930 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
931 rc, ma->ma_valid, ma->ma_lmm);
935 static int mdt_getattr_internal(struct mdt_thread_info *info,
936 struct mdt_object *o, int ma_need)
938 struct md_object *next = mdt_object_child(o);
939 const struct mdt_body *reqbody = info->mti_body;
940 struct ptlrpc_request *req = mdt_info_req(info);
941 struct md_attr *ma = &info->mti_attr;
942 struct lu_attr *la = &ma->ma_attr;
943 struct req_capsule *pill = info->mti_pill;
944 const struct lu_env *env = info->mti_env;
945 struct mdt_body *repbody;
946 struct lu_buf *buffer = &info->mti_buf;
947 struct obd_export *exp = info->mti_exp;
952 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
953 RETURN(err_serious(-ENOMEM));
955 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
959 if (mdt_object_remote(o)) {
960 /* This object is located on remote node.*/
961 /* Return -ENOTSUPP for old client */
962 if (!mdt_is_dne_client(req->rq_export))
963 GOTO(out, rc = -ENOTSUPP);
965 repbody->mbo_fid1 = *mdt_object_fid(o);
966 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
970 if (reqbody->mbo_eadatasize > 0) {
971 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
972 if (buffer->lb_buf == NULL)
973 GOTO(out, rc = -EPROTO);
974 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
977 buffer->lb_buf = NULL;
979 ma_need &= ~(MA_LOV | MA_LMV);
980 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
981 mdt_obd_name(info->mti_mdt),
982 req->rq_export->exp_client_uuid.uuid);
985 /* If it is dir object and client require MEA, then we got MEA */
986 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
987 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
988 /* Assumption: MDT_MD size is enough for lmv size. */
989 ma->ma_lmv = buffer->lb_buf;
990 ma->ma_lmv_size = buffer->lb_len;
991 ma->ma_need = MA_INODE;
992 if (ma->ma_lmv_size > 0) {
993 if (reqbody->mbo_valid & OBD_MD_MEA)
994 ma->ma_need |= MA_LMV;
995 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
996 ma->ma_need |= MA_LMV_DEF;
999 ma->ma_lmm = buffer->lb_buf;
1000 ma->ma_lmm_size = buffer->lb_len;
1001 ma->ma_need = MA_INODE | MA_HSM;
1002 if (ma->ma_lmm_size > 0)
1003 ma->ma_need |= MA_LOV;
1006 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1007 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1008 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1009 /* get default stripe info for this dir. */
1010 ma->ma_need |= MA_LOV_DEF;
1012 ma->ma_need |= ma_need;
1014 rc = mdt_attr_get_complex(info, o, ma);
1016 CDEBUG(rc == -ENOENT ? D_OTHER : D_ERROR,
1017 "%s: getattr error for "DFID": rc = %d\n",
1018 mdt_obd_name(info->mti_mdt),
1019 PFID(mdt_object_fid(o)), rc);
1023 /* if file is released, check if a restore is running */
1024 if (ma->ma_valid & MA_HSM) {
1025 repbody->mbo_valid |= OBD_MD_TSTATE;
1026 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1027 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1028 repbody->mbo_t_state = MS_RESTORE;
1031 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
1033 /* the Lustre protocol supposes to return default striping
1034 * on the user-visible root if explicitly requested */
1035 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
1036 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
1037 struct lu_fid rootfid;
1038 struct mdt_object *root;
1039 struct mdt_device *mdt = info->mti_mdt;
1041 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
1044 root = mdt_object_find(env, mdt, &rootfid);
1046 RETURN(PTR_ERR(root));
1047 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
1048 mdt_object_put(info->mti_env, root);
1050 CERROR("%s: getattr error for "DFID": rc = %d\n",
1051 mdt_obd_name(info->mti_mdt),
1052 PFID(mdt_object_fid(o)), rc);
1057 if (likely(ma->ma_valid & MA_INODE))
1058 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1062 if (mdt_body_has_lov(la, reqbody)) {
1063 if (ma->ma_valid & MA_LOV) {
1064 LASSERT(ma->ma_lmm_size);
1065 repbody->mbo_eadatasize = ma->ma_lmm_size;
1066 if (S_ISDIR(la->la_mode))
1067 repbody->mbo_valid |= OBD_MD_FLDIREA;
1069 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1070 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1072 if (ma->ma_valid & MA_LMV) {
1073 /* Return -ENOTSUPP for old client */
1074 if (!mdt_is_striped_client(req->rq_export))
1077 LASSERT(S_ISDIR(la->la_mode));
1078 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1079 repbody->mbo_eadatasize = ma->ma_lmv_size;
1080 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1082 if (ma->ma_valid & MA_LMV_DEF) {
1083 /* Return -ENOTSUPP for old client */
1084 if (!mdt_is_striped_client(req->rq_export))
1086 LASSERT(S_ISDIR(la->la_mode));
1087 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1088 repbody->mbo_eadatasize = ma->ma_lmv_size;
1089 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1090 OBD_MD_DEFAULT_MEA);
1092 } else if (S_ISLNK(la->la_mode) &&
1093 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1094 buffer->lb_buf = ma->ma_lmm;
1095 /* eadatasize from client includes NULL-terminator, so
1096 * there is no need to read it */
1097 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1098 rc = mo_readlink(env, next, buffer);
1099 if (unlikely(rc <= 0)) {
1100 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1101 mdt_obd_name(info->mti_mdt),
1102 PFID(mdt_object_fid(o)), rc);
1105 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1107 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1109 repbody->mbo_valid |= OBD_MD_LINKNAME;
1110 /* we need to report back size with NULL-terminator
1111 * because client expects that */
1112 repbody->mbo_eadatasize = rc + 1;
1113 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1114 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1115 "on "DFID ", expected %d\n",
1116 mdt_obd_name(info->mti_mdt),
1117 rc, PFID(mdt_object_fid(o)),
1118 reqbody->mbo_eadatasize - 1);
1119 /* NULL terminate */
1120 ((char *)ma->ma_lmm)[rc] = 0;
1122 /* If the total CDEBUG() size is larger than a page, it
1123 * will print a warning to the console, avoid this by
1124 * printing just the last part of the symlink. */
1125 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1126 print_limit < rc ? "..." : "", print_limit,
1127 (char *)ma->ma_lmm + rc - print_limit, rc);
1132 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1133 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1134 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1135 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1136 repbody->mbo_max_mdsize);
1139 #ifdef CONFIG_FS_POSIX_ACL
1140 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1141 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1142 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1143 if (IS_ERR(nodemap))
1144 RETURN(PTR_ERR(nodemap));
1146 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1147 nodemap_putref(nodemap);
1153 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1158 static int mdt_getattr(struct tgt_session_info *tsi)
1160 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1161 struct mdt_object *obj = info->mti_object;
1162 struct req_capsule *pill = info->mti_pill;
1163 struct mdt_body *reqbody;
1164 struct mdt_body *repbody;
1168 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1170 LASSERT(obj != NULL);
1171 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1173 /* Unlike intent case where we need to pre-fill out buffers early on
1174 * in intent policy for ldlm reasons, here we can have a much better
1175 * guess at EA size by just reading it from disk.
1176 * Exceptions are readdir and (missing) directory striping */
1178 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1179 /* No easy way to know how long is the symlink, but it cannot
1180 * be more than PATH_MAX, so we allocate +1 */
1183 /* A special case for fs ROOT: getattr there might fetch
1184 * default EA for entire fs, not just for this dir!
1186 } else if (lu_fid_eq(mdt_object_fid(obj),
1187 &info->mti_mdt->mdt_md_root_fid) &&
1188 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1189 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1191 /* Should the default strping be bigger, mdt_fix_reply
1192 * will reallocate */
1193 rc = DEF_REP_MD_SIZE;
1195 /* Read the actual EA size from disk */
1196 rc = mdt_attr_get_eabuf_size(info, obj);
1200 GOTO(out, rc = err_serious(rc));
1202 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1204 rc = req_capsule_server_pack(pill);
1205 if (unlikely(rc != 0))
1206 GOTO(out, rc = err_serious(rc));
1208 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1209 LASSERT(repbody != NULL);
1210 repbody->mbo_eadatasize = 0;
1211 repbody->mbo_aclsize = 0;
1213 rc = mdt_check_ucred(info);
1215 GOTO(out_shrink, rc);
1217 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1219 rc = mdt_getattr_internal(info, obj, 0);
1222 mdt_client_compatibility(info);
1223 rc2 = mdt_fix_reply(info);
1227 mdt_thread_info_fini(info);
1232 * Exchange MOF_LOV_CREATED flags between two objects after a
1233 * layout swap. No assumption is made on whether o1 or o2 have
1234 * created objects or not.
1236 * \param[in,out] o1 First swap layout object
1237 * \param[in,out] o2 Second swap layout object
1239 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1241 unsigned int o1_lov_created = o1->mot_lov_created;
1243 mutex_lock(&o1->mot_lov_mutex);
1244 mutex_lock(&o2->mot_lov_mutex);
1246 o1->mot_lov_created = o2->mot_lov_created;
1247 o2->mot_lov_created = o1_lov_created;
1249 mutex_unlock(&o2->mot_lov_mutex);
1250 mutex_unlock(&o1->mot_lov_mutex);
1253 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1255 struct mdt_thread_info *info;
1256 struct ptlrpc_request *req = tgt_ses_req(tsi);
1257 struct obd_export *exp = req->rq_export;
1258 struct mdt_object *o1, *o2, *o;
1259 struct mdt_lock_handle *lh1, *lh2;
1260 struct mdc_swap_layouts *msl;
1264 /* client does not support layout lock, so layout swaping
1266 * FIXME: there is a problem for old clients which don't support
1267 * layout lock yet. If those clients have already opened the file
1268 * they won't be notified at all so that old layout may still be
1269 * used to do IO. This can be fixed after file release is landed by
1270 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1271 if (!exp_connect_layout(exp))
1272 RETURN(-EOPNOTSUPP);
1274 info = tsi2mdt_info(tsi);
1276 if (info->mti_dlm_req != NULL)
1277 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1279 o1 = info->mti_object;
1280 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1281 &info->mti_body->mbo_fid2);
1283 GOTO(out, rc = PTR_ERR(o));
1285 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1286 GOTO(put, rc = -ENOENT);
1288 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1289 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1295 /* permission check. Make sure the calling process having permission
1296 * to write both files. */
1297 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1302 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1307 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1309 GOTO(put, rc = -EPROTO);
1311 lh1 = &info->mti_lh[MDT_LH_NEW];
1312 mdt_lock_reg_init(lh1, LCK_EX);
1313 lh2 = &info->mti_lh[MDT_LH_OLD];
1314 mdt_lock_reg_init(lh2, LCK_EX);
1316 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1317 MDS_INODELOCK_XATTR);
1321 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1322 MDS_INODELOCK_XATTR);
1326 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1327 mdt_object_child(o2), msl->msl_flags);
1331 mdt_swap_lov_flag(o1, o2);
1334 mdt_object_unlock(info, o2, lh2, rc);
1336 mdt_object_unlock(info, o1, lh1, rc);
1338 mdt_object_put(info->mti_env, o);
1340 mdt_thread_info_fini(info);
1344 static int mdt_raw_lookup(struct mdt_thread_info *info,
1345 struct mdt_object *parent,
1346 const struct lu_name *lname,
1347 struct ldlm_reply *ldlm_rep)
1349 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1353 LASSERT(!info->mti_cross_ref);
1355 /* Only got the fid of this obj by name */
1356 fid_zero(child_fid);
1357 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1358 lname, child_fid, &info->mti_spec);
1360 struct mdt_body *repbody;
1362 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1363 repbody->mbo_fid1 = *child_fid;
1364 repbody->mbo_valid = OBD_MD_FLID;
1365 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1366 } else if (rc == -ENOENT) {
1367 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1374 * UPDATE lock should be taken against parent, and be released before exit;
1375 * child_bits lock should be taken against child, and be returned back:
1376 * (1)normal request should release the child lock;
1377 * (2)intent request will grant the lock to client.
1379 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1380 struct mdt_lock_handle *lhc,
1382 struct ldlm_reply *ldlm_rep)
1384 struct ptlrpc_request *req = mdt_info_req(info);
1385 struct mdt_body *reqbody = NULL;
1386 struct mdt_object *parent = info->mti_object;
1387 struct mdt_object *child;
1388 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1389 struct lu_name *lname = NULL;
1390 struct mdt_lock_handle *lhp = NULL;
1391 struct ldlm_lock *lock;
1398 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1399 LASSERT(ergo(is_resent,
1400 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1405 if (info->mti_cross_ref) {
1406 /* Only getattr on the child. Parent is on another node. */
1407 mdt_set_disposition(info, ldlm_rep,
1408 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1410 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1412 PFID(mdt_object_fid(child)), ldlm_rep);
1414 rc = mdt_check_resent_lock(info, child, lhc);
1417 } else if (rc > 0) {
1418 mdt_lock_handle_init(lhc);
1419 mdt_lock_reg_init(lhc, LCK_PR);
1422 * Object's name is on another MDS, no lookup or layout
1423 * lock is needed here but update lock is.
1425 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1426 MDS_INODELOCK_LAYOUT);
1427 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1429 rc = mdt_object_lock(info, child, lhc, child_bits);
1434 /* Finally, we can get attr for child. */
1435 if (!mdt_object_exists(child)) {
1436 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1438 "remote object doesn't exist.");
1439 mdt_object_unlock(info, child, lhc, 1);
1443 rc = mdt_getattr_internal(info, child, 0);
1444 if (unlikely(rc != 0))
1445 mdt_object_unlock(info, child, lhc, 1);
1450 lname = &info->mti_name;
1451 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1453 if (lu_name_is_valid(lname)) {
1454 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1455 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1456 PNAME(lname), ldlm_rep);
1458 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1459 if (unlikely(reqbody == NULL))
1460 RETURN(err_serious(-EPROTO));
1462 *child_fid = reqbody->mbo_fid2;
1464 if (unlikely(!fid_is_sane(child_fid)))
1465 RETURN(err_serious(-EINVAL));
1467 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1469 PFID(mdt_object_fid(parent)),
1470 PFID(&reqbody->mbo_fid2), ldlm_rep);
1473 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1475 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1476 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1478 "Parent doesn't exist!");
1482 if (mdt_object_remote(parent)) {
1483 CERROR("%s: parent "DFID" is on remote target\n",
1484 mdt_obd_name(info->mti_mdt),
1485 PFID(mdt_object_fid(parent)));
1489 if (lu_name_is_valid(lname)) {
1490 /* Always allow to lookup ".." */
1491 if (unlikely(lname->ln_namelen == 2 &&
1492 lname->ln_name[0] == '.' &&
1493 lname->ln_name[1] == '.'))
1494 info->mti_spec.sp_permitted = 1;
1496 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1497 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1502 /* step 1: lock parent only if parent is a directory */
1503 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1504 lhp = &info->mti_lh[MDT_LH_PARENT];
1505 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1506 rc = mdt_object_lock(info, parent, lhp,
1507 MDS_INODELOCK_UPDATE);
1508 if (unlikely(rc != 0))
1512 /* step 2: lookup child's fid by name */
1513 fid_zero(child_fid);
1514 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1515 child_fid, &info->mti_spec);
1517 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1520 GOTO(out_parent, rc);
1523 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1526 *step 3: find the child object by fid & lock it.
1527 * regardless if it is local or remote.
1529 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1530 * set parent dir fid the same as child fid in getattr by fid case
1531 * we should not lu_object_find() the object again, could lead
1532 * to hung if there is a concurrent unlink destroyed the object.
1534 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1535 mdt_object_get(info->mti_env, parent);
1538 child = mdt_object_find(info->mti_env, info->mti_mdt,
1542 if (unlikely(IS_ERR(child)))
1543 GOTO(out_parent, rc = PTR_ERR(child));
1545 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1546 if (!mdt_object_exists(child)) {
1547 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1549 "Object doesn't exist!");
1550 GOTO(out_child, rc = -ENOENT);
1553 rc = mdt_check_resent_lock(info, child, lhc);
1555 GOTO(out_child, rc);
1556 } else if (rc > 0) {
1557 mdt_lock_handle_init(lhc);
1558 mdt_lock_reg_init(lhc, LCK_PR);
1561 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1562 mdt_object_exists(child) && !mdt_object_remote(child)) {
1563 struct md_attr *ma = &info->mti_attr;
1566 ma->ma_need = MA_INODE;
1567 rc = mdt_attr_get_complex(info, child, ma);
1568 if (unlikely(rc != 0))
1569 GOTO(out_child, rc);
1571 /* If the file has not been changed for some time, we
1572 * return not only a LOOKUP lock, but also an UPDATE
1573 * lock and this might save us RPC on later STAT. For
1574 * directories, it also let negative dentry cache start
1575 * working for this dir. */
1576 if (ma->ma_valid & MA_INODE &&
1577 ma->ma_attr.la_valid & LA_CTIME &&
1578 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1579 ma->ma_attr.la_ctime < cfs_time_current_sec())
1580 child_bits |= MDS_INODELOCK_UPDATE;
1583 /* layout lock must be granted in a best-effort way
1584 * for IT operations */
1585 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1586 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1587 exp_connect_layout(info->mti_exp) &&
1588 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1589 !mdt_object_remote(child) && ldlm_rep != NULL) {
1590 /* try to grant layout lock for regular file. */
1596 child_bits |= MDS_INODELOCK_LAYOUT;
1597 /* try layout lock, it may fail to be granted due to
1598 * contention at LOOKUP or UPDATE */
1599 if (!mdt_object_lock_try(info, child, lhc,
1601 child_bits &= ~MDS_INODELOCK_LAYOUT;
1602 LASSERT(child_bits != 0);
1603 rc = mdt_object_lock(info, child, lhc,
1609 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1610 * client will enqueue the lock to the remote MDT */
1611 if (mdt_object_remote(child))
1612 child_bits &= ~MDS_INODELOCK_UPDATE;
1613 rc = mdt_object_lock(info, child, lhc, child_bits);
1615 if (unlikely(rc != 0))
1616 GOTO(out_child, rc);
1619 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1621 /* finally, we can get attr for child. */
1622 rc = mdt_getattr_internal(info, child, ma_need);
1623 if (unlikely(rc != 0)) {
1624 mdt_object_unlock(info, child, lhc, 1);
1626 /* Debugging code. */
1627 LDLM_DEBUG(lock, "Returning lock to client");
1628 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1629 &lock->l_resource->lr_name),
1630 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1631 PLDLMRES(lock->l_resource),
1632 PFID(mdt_object_fid(child)));
1635 LDLM_LOCK_PUT(lock);
1639 mdt_object_put(info->mti_env, child);
1642 mdt_object_unlock(info, parent, lhp, 1);
1646 /* normal handler: should release the child lock */
1647 static int mdt_getattr_name(struct tgt_session_info *tsi)
1649 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1650 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1651 struct mdt_body *reqbody;
1652 struct mdt_body *repbody;
1656 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1657 LASSERT(reqbody != NULL);
1658 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1659 LASSERT(repbody != NULL);
1661 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1662 repbody->mbo_eadatasize = 0;
1663 repbody->mbo_aclsize = 0;
1665 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1667 GOTO(out_shrink, rc);
1669 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1670 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1671 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1672 lhc->mlh_reg_lh.cookie = 0;
1674 mdt_exit_ucred(info);
1677 mdt_client_compatibility(info);
1678 rc2 = mdt_fix_reply(info);
1681 mdt_thread_info_fini(info);
1685 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1686 void *karg, void __user *uarg);
1688 static int mdt_set_info(struct tgt_session_info *tsi)
1690 struct ptlrpc_request *req = tgt_ses_req(tsi);
1693 int keylen, vallen, rc = 0;
1697 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1699 DEBUG_REQ(D_HA, req, "no set_info key");
1700 RETURN(err_serious(-EFAULT));
1703 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1706 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1708 DEBUG_REQ(D_HA, req, "no set_info val");
1709 RETURN(err_serious(-EFAULT));
1712 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1715 /* Swab any part of val you need to here */
1716 if (KEY_IS(KEY_READ_ONLY)) {
1717 spin_lock(&req->rq_export->exp_lock);
1719 *exp_connect_flags_ptr(req->rq_export) |=
1722 *exp_connect_flags_ptr(req->rq_export) &=
1723 ~OBD_CONNECT_RDONLY;
1724 spin_unlock(&req->rq_export->exp_lock);
1725 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1726 struct changelog_setinfo *cs = val;
1728 if (vallen != sizeof(*cs)) {
1729 CERROR("%s: bad changelog_clear setinfo size %d\n",
1730 tgt_name(tsi->tsi_tgt), vallen);
1733 if (ptlrpc_req_need_swab(req)) {
1734 __swab64s(&cs->cs_recno);
1735 __swab32s(&cs->cs_id);
1738 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1740 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
1742 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
1749 static int mdt_readpage(struct tgt_session_info *tsi)
1751 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1752 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1753 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1754 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1755 struct mdt_body *repbody;
1761 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1762 RETURN(err_serious(-ENOMEM));
1764 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1765 if (repbody == NULL || reqbody == NULL)
1766 RETURN(err_serious(-EFAULT));
1769 * prepare @rdpg before calling lower layers and transfer itself. Here
1770 * reqbody->size contains offset of where to start to read and
1771 * reqbody->nlink contains number bytes to read.
1773 rdpg->rp_hash = reqbody->mbo_size;
1774 if (rdpg->rp_hash != reqbody->mbo_size) {
1775 CERROR("Invalid hash: %#llx != %#llx\n",
1776 rdpg->rp_hash, reqbody->mbo_size);
1780 rdpg->rp_attrs = reqbody->mbo_mode;
1781 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1782 rdpg->rp_attrs |= LUDA_64BITHASH;
1783 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1784 exp_max_brw_size(tsi->tsi_exp));
1785 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
1787 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1788 if (rdpg->rp_pages == NULL)
1791 for (i = 0; i < rdpg->rp_npages; ++i) {
1792 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1793 if (rdpg->rp_pages[i] == NULL)
1794 GOTO(free_rdpg, rc = -ENOMEM);
1797 /* call lower layers to fill allocated pages with directory data */
1798 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1800 GOTO(free_rdpg, rc);
1802 /* send pages to client */
1803 rc = tgt_sendpage(tsi, rdpg, rc);
1808 for (i = 0; i < rdpg->rp_npages; i++)
1809 if (rdpg->rp_pages[i] != NULL)
1810 __free_page(rdpg->rp_pages[i]);
1811 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1813 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1819 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
1821 struct lu_ucred *uc = mdt_ucred_check(info);
1822 struct lu_attr *attr = &info->mti_attr.ma_attr;
1827 if (op != REINT_SETATTR) {
1828 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
1829 attr->la_uid = uc->uc_fsuid;
1830 /* for S_ISGID, inherit gid from his parent, such work will be
1831 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
1832 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
1833 attr->la_gid = uc->uc_fsgid;
1839 static int mdt_reint_internal(struct mdt_thread_info *info,
1840 struct mdt_lock_handle *lhc,
1843 struct req_capsule *pill = info->mti_pill;
1844 struct mdt_body *repbody;
1849 rc = mdt_reint_unpack(info, op);
1851 CERROR("Can't unpack reint, rc %d\n", rc);
1852 RETURN(err_serious(rc));
1855 /* for replay (no_create) lmm is not needed, client has it already */
1856 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1857 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1860 /* llog cookies are always 0, the field is kept for compatibility */
1861 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1862 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1864 rc = req_capsule_server_pack(pill);
1866 CERROR("Can't pack response, rc %d\n", rc);
1867 RETURN(err_serious(rc));
1870 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1871 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1873 repbody->mbo_eadatasize = 0;
1874 repbody->mbo_aclsize = 0;
1877 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1879 /* for replay no cookkie / lmm need, because client have this already */
1880 if (info->mti_spec.no_create)
1881 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1882 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1884 rc = mdt_init_ucred_reint(info);
1886 GOTO(out_shrink, rc);
1888 rc = mdt_fix_attr_ucred(info, op);
1890 GOTO(out_ucred, rc = err_serious(rc));
1892 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
1894 GOTO(out_ucred, rc);
1895 } else if (rc == 1) {
1896 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1897 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1898 GOTO(out_ucred, rc);
1900 rc = mdt_reint_rec(info, lhc);
1903 mdt_exit_ucred(info);
1905 mdt_client_compatibility(info);
1906 rc2 = mdt_fix_reply(info);
1912 static long mdt_reint_opcode(struct ptlrpc_request *req,
1913 const struct req_format **fmt)
1915 struct mdt_device *mdt;
1916 struct mdt_rec_reint *rec;
1919 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1921 opc = rec->rr_opcode;
1922 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1923 if (opc < REINT_MAX && fmt[opc] != NULL)
1924 req_capsule_extend(&req->rq_pill, fmt[opc]);
1926 mdt = mdt_exp2dev(req->rq_export);
1927 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1928 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1929 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1930 opc = err_serious(-EFAULT);
1933 opc = err_serious(-EFAULT);
1938 static int mdt_reint(struct tgt_session_info *tsi)
1942 static const struct req_format *reint_fmts[REINT_MAX] = {
1943 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1944 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1945 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1946 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1947 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1948 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1949 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1950 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
1951 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME
1956 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1958 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1960 * No lock possible here from client to pass it to reint code
1963 rc = mdt_reint_internal(info, NULL, opc);
1964 mdt_thread_info_fini(info);
1969 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1973 /* this should sync the whole device */
1974 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1976 struct dt_device *dt = mdt->mdt_bottom;
1980 rc = dt->dd_ops->dt_sync(env, dt);
1984 /* this should sync this object */
1985 static int mdt_object_sync(struct mdt_thread_info *info)
1987 struct md_object *next;
1991 if (!mdt_object_exists(info->mti_object)) {
1992 CWARN("%s: non existing object "DFID": rc = %d\n",
1993 mdt_obd_name(info->mti_mdt),
1994 PFID(mdt_object_fid(info->mti_object)), -ESTALE);
1997 next = mdt_object_child(info->mti_object);
1998 rc = mo_object_sync(info->mti_env, next);
2003 static int mdt_sync(struct tgt_session_info *tsi)
2005 struct ptlrpc_request *req = tgt_ses_req(tsi);
2006 struct req_capsule *pill = tsi->tsi_pill;
2007 struct mdt_body *body;
2012 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2013 RETURN(err_serious(-ENOMEM));
2015 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2016 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2018 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2020 /* sync an object */
2021 rc = mdt_object_sync(info);
2023 const struct lu_fid *fid;
2024 struct lu_attr *la = &info->mti_attr.ma_attr;
2026 info->mti_attr.ma_need = MA_INODE;
2027 info->mti_attr.ma_valid = 0;
2028 rc = mdt_attr_get_complex(info, info->mti_object,
2031 body = req_capsule_server_get(pill,
2033 fid = mdt_object_fid(info->mti_object);
2034 mdt_pack_attr2body(info, body, la, fid);
2037 mdt_thread_info_fini(info);
2040 mdt_counter_incr(req, LPROC_MDT_SYNC);
2046 * Handle quota control requests to consult current usage/limit, but also
2047 * to configure quota enforcement
2049 static int mdt_quotactl(struct tgt_session_info *tsi)
2051 struct obd_export *exp = tsi->tsi_exp;
2052 struct req_capsule *pill = tsi->tsi_pill;
2053 struct obd_quotactl *oqctl, *repoqc;
2055 struct mdt_device *mdt = mdt_exp2dev(exp);
2056 struct lu_device *qmt = mdt->mdt_qmt_dev;
2057 struct lu_nodemap *nodemap;
2060 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2062 RETURN(err_serious(-EPROTO));
2064 rc = req_capsule_server_pack(pill);
2066 RETURN(err_serious(rc));
2068 nodemap = nodemap_get_from_exp(exp);
2069 if (IS_ERR(nodemap))
2070 RETURN(PTR_ERR(nodemap));
2072 switch (oqctl->qc_cmd) {
2073 /* master quotactl */
2076 if (!nodemap_can_setquota(nodemap))
2077 GOTO(out_nodemap, rc = -EPERM);
2081 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2082 /* slave quotactl */
2087 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2088 GOTO(out_nodemap, rc = -EFAULT);
2092 if (oqctl->qc_type == USRQUOTA)
2093 id = nodemap_map_id(nodemap, NODEMAP_UID,
2094 NODEMAP_CLIENT_TO_FS, id);
2095 else if (oqctl->qc_type == GRPQUOTA)
2096 id = nodemap_map_id(nodemap, NODEMAP_UID,
2097 NODEMAP_CLIENT_TO_FS, id);
2099 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2101 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2103 if (oqctl->qc_id != id)
2104 swap(oqctl->qc_id, id);
2106 switch (oqctl->qc_cmd) {
2112 /* forward quotactl request to QMT */
2113 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2118 /* slave quotactl */
2119 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2124 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2125 GOTO(out_nodemap, rc = -EFAULT);
2128 if (oqctl->qc_id != id)
2129 swap(oqctl->qc_id, id);
2136 nodemap_putref(nodemap);
2141 /** clone llog ctxt from child (mdd)
2142 * This allows remote llog (replicator) access.
2143 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2144 * context was originally set up, or we can handle them directly.
2145 * I choose the latter, but that means I need any llog
2146 * contexts set up by child to be accessable by the mdt. So we clone the
2147 * context into our context list here.
2149 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2152 struct md_device *next = mdt->mdt_child;
2153 struct llog_ctxt *ctxt;
2156 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2159 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2160 if (rc || ctxt == NULL) {
2164 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2166 CERROR("Can't set mdt ctxt %d\n", rc);
2171 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2172 struct mdt_device *mdt, int idx)
2174 struct llog_ctxt *ctxt;
2176 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2179 /* Put once for the get we just did, and once for the clone */
2180 llog_ctxt_put(ctxt);
2181 llog_ctxt_put(ctxt);
2186 * sec context handlers
2188 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2190 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2196 * quota request handlers
2198 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2200 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2201 struct lu_device *qmt = mdt->mdt_qmt_dev;
2206 RETURN(err_serious(-EOPNOTSUPP));
2208 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2212 struct mdt_object *mdt_object_new(const struct lu_env *env,
2213 struct mdt_device *d,
2214 const struct lu_fid *f)
2216 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2217 struct lu_object *o;
2218 struct mdt_object *m;
2221 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2222 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2223 if (unlikely(IS_ERR(o)))
2224 m = (struct mdt_object *)o;
2230 struct mdt_object *mdt_object_find(const struct lu_env *env,
2231 struct mdt_device *d,
2232 const struct lu_fid *f)
2234 struct lu_object *o;
2235 struct mdt_object *m;
2238 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2239 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2240 if (unlikely(IS_ERR(o)))
2241 m = (struct mdt_object *)o;
2249 * Asyncronous commit for mdt device.
2251 * Pass asynchonous commit call down the MDS stack.
2253 * \param env environment
2254 * \param mdt the mdt device
2256 static void mdt_device_commit_async(const struct lu_env *env,
2257 struct mdt_device *mdt)
2259 struct dt_device *dt = mdt->mdt_bottom;
2263 rc = dt->dd_ops->dt_commit_async(env, dt);
2264 if (unlikely(rc != 0))
2265 CWARN("%s: async commit start failed: rc = %d\n",
2266 mdt_obd_name(mdt), rc);
2267 atomic_inc(&mdt->mdt_async_commit_count);
2272 * Mark the lock as "synchonous".
2274 * Mark the lock to deffer transaction commit to the unlock time.
2276 * \param lock the lock to mark as "synchonous"
2278 * \see mdt_is_lock_sync
2279 * \see mdt_save_lock
2281 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2283 lock->l_ast_data = (void*)1;
2287 * Check whehter the lock "synchonous" or not.
2289 * \param lock the lock to check
2290 * \retval 1 the lock is "synchonous"
2291 * \retval 0 the lock isn't "synchronous"
2293 * \see mdt_set_lock_sync
2294 * \see mdt_save_lock
2296 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2298 return lock->l_ast_data != NULL;
2302 * Blocking AST for mdt locks.
2304 * Starts transaction commit if in case of COS lock conflict or
2305 * deffers such a commit to the mdt_save_lock.
2307 * \param lock the lock which blocks a request or cancelling lock
2308 * \param desc unused
2309 * \param data unused
2310 * \param flag indicates whether this cancelling or blocking callback
2312 * \see ldlm_blocking_ast_nocheck
2314 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2315 void *data, int flag)
2317 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2318 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2322 if (flag == LDLM_CB_CANCELING)
2325 lock_res_and_lock(lock);
2326 if (lock->l_blocking_ast != mdt_blocking_ast) {
2327 unlock_res_and_lock(lock);
2330 if (lock->l_req_mode & (LCK_PW | LCK_EX) &&
2331 lock->l_blocking_lock != NULL) {
2332 if (mdt_cos_is_enabled(mdt) &&
2333 lock->l_client_cookie !=
2334 lock->l_blocking_lock->l_client_cookie)
2335 mdt_set_lock_sync(lock);
2336 else if (mdt_slc_is_enabled(mdt) &&
2337 ldlm_is_cos_incompat(lock->l_blocking_lock))
2338 mdt_set_lock_sync(lock);
2340 rc = ldlm_blocking_ast_nocheck(lock);
2342 /* There is no lock conflict if l_blocking_lock == NULL,
2343 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2344 * when the last reference to a local lock was released */
2345 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2348 rc = lu_env_init(&env, LCT_LOCAL);
2349 if (unlikely(rc != 0))
2350 CWARN("%s: lu_env initialization failed, cannot "
2351 "start asynchronous commit: rc = %d\n",
2354 mdt_device_commit_async(&env, mdt);
2361 * Blocking AST for cross-MDT lock
2363 * Discard lock from uncommitted_slc_locks and cancel it.
2365 * \param lock the lock which blocks a request or cancelling lock
2366 * \param desc unused
2367 * \param data unused
2368 * \param flag indicates whether this cancelling or blocking callback
2369 * \retval 0 on success
2370 * \retval negative number on error
2372 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2373 void *data, int flag)
2379 case LDLM_CB_BLOCKING: {
2380 struct lustre_handle lockh;
2382 ldlm_lock2handle(lock, &lockh);
2383 rc = ldlm_cli_cancel(&lockh,
2384 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
2386 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2391 case LDLM_CB_CANCELING: {
2392 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2393 struct mdt_device *mdt =
2394 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
2396 LDLM_DEBUG(lock, "Revoke remote lock\n");
2398 /* discard slc lock here so that it can be cleaned anytime,
2399 * especially for cleanup_resource() */
2400 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
2402 /* once we cache lock, l_ast_data is set to mdt_object */
2403 if (lock->l_ast_data != NULL) {
2404 struct mdt_object *mo = lock->l_ast_data;
2407 rc = lu_env_init(&env, LCT_MD_THREAD);
2408 if (unlikely(rc != 0)) {
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 %#llx",
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 %#llx\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 = %lld\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(&info->mti_mdt->mdt_lut, lock,
2810 ldlm_lock_decref(h, mode);
2819 * Unlock mdt object.
2821 * Immeditely release the regular lock and the PDO lock or save the
2822 * lock in request and keep them referenced until client ACK or
2823 * transaction commit.
2825 * \param info thread info object
2826 * \param o mdt object
2827 * \param lh mdt lock handle referencing regular and PDO locks
2828 * \param decref force immediate lock releasing
2830 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
2832 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2833 struct mdt_lock_handle *lh, int decref)
2837 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2838 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2839 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
2845 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2846 const struct lu_fid *f,
2847 struct mdt_lock_handle *lh,
2850 struct mdt_object *o;
2852 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2856 rc = mdt_object_lock(info, o, lh, ibits);
2858 mdt_object_put(info->mti_env, o);
2865 void mdt_object_unlock_put(struct mdt_thread_info * info,
2866 struct mdt_object * o,
2867 struct mdt_lock_handle *lh,
2870 mdt_object_unlock(info, o, lh, decref);
2871 mdt_object_put(info->mti_env, o);
2875 * Generic code handling requests that have struct mdt_body passed in:
2877 * - extract mdt_body from request and save it in @info, if present;
2879 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2882 * - if HABEO_CORPUS flag is set for this request type check whether object
2883 * actually exists on storage (lu_object_exists()).
2886 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2888 const struct mdt_body *body;
2889 struct mdt_object *obj;
2890 const struct lu_env *env;
2891 struct req_capsule *pill;
2895 env = info->mti_env;
2896 pill = info->mti_pill;
2898 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2902 if (!(body->mbo_valid & OBD_MD_FLID))
2905 if (!fid_is_sane(&body->mbo_fid1)) {
2906 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2910 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2912 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2913 mdt_object_put(env, obj);
2916 info->mti_object = obj;
2925 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2927 struct req_capsule *pill = info->mti_pill;
2931 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2932 rc = mdt_body_unpack(info, flags);
2936 if (rc == 0 && (flags & HABEO_REFERO)) {
2938 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2939 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2941 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2942 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2945 rc = req_capsule_server_pack(pill);
2950 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2952 lh->mlh_type = MDT_NUL_LOCK;
2953 lh->mlh_reg_lh.cookie = 0ull;
2954 lh->mlh_reg_mode = LCK_MINMODE;
2955 lh->mlh_pdo_lh.cookie = 0ull;
2956 lh->mlh_pdo_mode = LCK_MINMODE;
2957 lh->mlh_rreg_lh.cookie = 0ull;
2958 lh->mlh_rreg_mode = LCK_MINMODE;
2961 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2963 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2964 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2968 * Initialize fields of struct mdt_thread_info. Other fields are left in
2969 * uninitialized state, because it's too expensive to zero out whole
2970 * mdt_thread_info (> 1K) on each request arrival.
2972 void mdt_thread_info_init(struct ptlrpc_request *req,
2973 struct mdt_thread_info *info)
2977 info->mti_pill = &req->rq_pill;
2980 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2981 mdt_lock_handle_init(&info->mti_lh[i]);
2983 /* mdt device: it can be NULL while CONNECT */
2984 if (req->rq_export) {
2985 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2986 info->mti_exp = req->rq_export;
2988 info->mti_mdt = NULL;
2989 info->mti_env = req->rq_svc_thread->t_env;
2990 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2992 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2993 info->mti_big_buf = LU_BUF_NULL;
2994 info->mti_body = NULL;
2995 info->mti_object = NULL;
2996 info->mti_dlm_req = NULL;
2997 info->mti_has_trans = 0;
2998 info->mti_cross_ref = 0;
2999 info->mti_opdata = 0;
3000 info->mti_big_lmm_used = 0;
3002 info->mti_spec.no_create = 0;
3003 info->mti_spec.sp_rm_entry = 0;
3004 info->mti_spec.sp_permitted = 0;
3005 info->mti_spec.sp_migrate_close = 0;
3007 info->mti_spec.u.sp_ea.eadata = NULL;
3008 info->mti_spec.u.sp_ea.eadatalen = 0;
3011 void mdt_thread_info_fini(struct mdt_thread_info *info)
3015 if (info->mti_object != NULL) {
3016 mdt_object_put(info->mti_env, info->mti_object);
3017 info->mti_object = NULL;
3020 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3021 mdt_lock_handle_fini(&info->mti_lh[i]);
3022 info->mti_env = NULL;
3023 info->mti_pill = NULL;
3024 info->mti_exp = NULL;
3026 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3027 lu_buf_free(&info->mti_big_buf);
3030 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3032 struct mdt_thread_info *mti;
3034 mti = mdt_th_info(tsi->tsi_env);
3035 LASSERT(mti != NULL);
3037 mdt_thread_info_init(tgt_ses_req(tsi), mti);
3038 if (tsi->tsi_corpus != NULL) {
3039 mti->mti_object = mdt_obj(tsi->tsi_corpus);
3040 lu_object_get(tsi->tsi_corpus);
3042 mti->mti_body = tsi->tsi_mdt_body;
3043 mti->mti_dlm_req = tsi->tsi_dlm_req;
3048 static int mdt_tgt_connect(struct tgt_session_info *tsi)
3050 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
3052 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
3053 set_current_state(TASK_UNINTERRUPTIBLE);
3054 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
3057 return tgt_connect(tsi);
3075 static int mdt_intent_getattr(enum mdt_it_code opcode,
3076 struct mdt_thread_info *info,
3077 struct ldlm_lock **,
3080 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3081 struct mdt_thread_info *info,
3082 struct ldlm_lock **lockp,
3085 static int mdt_intent_layout(enum mdt_it_code opcode,
3086 struct mdt_thread_info *info,
3087 struct ldlm_lock **,
3089 static int mdt_intent_reint(enum mdt_it_code opcode,
3090 struct mdt_thread_info *info,
3091 struct ldlm_lock **,
3094 static struct mdt_it_flavor {
3095 const struct req_format *it_fmt;
3097 int (*it_act)(enum mdt_it_code ,
3098 struct mdt_thread_info *,
3099 struct ldlm_lock **,
3102 } mdt_it_flavor[] = {
3104 .it_fmt = &RQF_LDLM_INTENT,
3105 /*.it_flags = HABEO_REFERO,*/
3107 .it_act = mdt_intent_reint,
3108 .it_reint = REINT_OPEN
3111 .it_fmt = &RQF_LDLM_INTENT,
3113 * OCREAT is not a MUTABOR request as if the file
3115 * We do the extra check of OBD_CONNECT_RDONLY in
3116 * mdt_reint_open() when we really need to create
3120 .it_act = mdt_intent_reint,
3121 .it_reint = REINT_OPEN
3124 .it_fmt = &RQF_LDLM_INTENT,
3125 .it_flags = MUTABOR,
3126 .it_act = mdt_intent_reint,
3127 .it_reint = REINT_CREATE
3129 [MDT_IT_GETATTR] = {
3130 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
3131 .it_flags = HABEO_REFERO,
3132 .it_act = mdt_intent_getattr
3134 [MDT_IT_READDIR] = {
3140 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
3141 .it_flags = HABEO_REFERO,
3142 .it_act = mdt_intent_getattr
3145 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
3146 .it_flags = MUTABOR,
3148 .it_reint = REINT_UNLINK
3152 .it_flags = MUTABOR,
3155 [MDT_IT_GETXATTR] = {
3156 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
3157 .it_flags = HABEO_CORPUS,
3158 .it_act = mdt_intent_getxattr
3161 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
3163 .it_act = mdt_intent_layout
3168 mdt_intent_lock_replace(struct mdt_thread_info *info,
3169 struct ldlm_lock **lockp,
3170 struct mdt_lock_handle *lh,
3171 __u64 flags, int result)
3173 struct ptlrpc_request *req = mdt_info_req(info);
3174 struct ldlm_lock *lock = *lockp;
3175 struct ldlm_lock *new_lock;
3177 /* If possible resent found a lock, @lh is set to its handle */
3178 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3180 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3181 lh->mlh_reg_lh.cookie = 0;
3185 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
3186 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3187 * a resend case, however, it could be already destroyed
3188 * due to client eviction or a raced cancel RPC. */
3189 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
3190 lh->mlh_reg_lh.cookie);
3191 lh->mlh_reg_lh.cookie = 0;
3195 LASSERTF(new_lock != NULL,
3196 "lockh %#llx flags %#llx : rc = %d\n",
3197 lh->mlh_reg_lh.cookie, flags, result);
3200 * If we've already given this lock to a client once, then we should
3201 * have no readers or writers. Otherwise, we should have one reader
3202 * _or_ writer ref (which will be zeroed below) before returning the
3205 if (new_lock->l_export == req->rq_export) {
3206 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3208 LASSERT(new_lock->l_export == NULL);
3209 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3214 if (new_lock->l_export == req->rq_export) {
3216 * Already gave this to the client, which means that we
3217 * reconstructed a reply.
3219 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3222 LDLM_LOCK_RELEASE(new_lock);
3223 lh->mlh_reg_lh.cookie = 0;
3224 RETURN(ELDLM_LOCK_REPLACED);
3228 * Fixup the lock to be given to the client.
3230 lock_res_and_lock(new_lock);
3231 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3232 * possible blocking AST. */
3233 while (new_lock->l_readers > 0) {
3234 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3235 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3236 new_lock->l_readers--;
3238 while (new_lock->l_writers > 0) {
3239 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3240 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3241 new_lock->l_writers--;
3244 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3245 new_lock->l_blocking_ast = lock->l_blocking_ast;
3246 new_lock->l_completion_ast = lock->l_completion_ast;
3247 new_lock->l_remote_handle = lock->l_remote_handle;
3248 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3250 unlock_res_and_lock(new_lock);
3252 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3253 &new_lock->l_remote_handle,
3254 &new_lock->l_exp_hash);
3256 LDLM_LOCK_RELEASE(new_lock);
3257 lh->mlh_reg_lh.cookie = 0;
3259 RETURN(ELDLM_LOCK_REPLACED);
3262 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3263 struct ldlm_lock *new_lock,
3264 struct mdt_lock_handle *lh,
3267 struct ptlrpc_request *req = mdt_info_req(info);
3268 struct ldlm_request *dlmreq;
3270 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3273 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3275 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3276 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3278 if (flags & LDLM_FL_RESENT) {
3279 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3280 lh->mlh_reg_mode = new_lock->l_granted_mode;
3282 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3283 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
3284 lh->mlh_reg_lh.cookie);
3289 * If the xid matches, then we know this is a resent request, and allow
3290 * it. (It's probably an OPEN, for which we don't send a lock.
3292 if (req_can_reconstruct(req, NULL))
3296 * This remote handle isn't enqueued, so we never received or processed
3297 * this request. Clear MSG_RESENT, because it can be handled like any
3298 * normal request now.
3300 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3302 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
3303 dlmreq->lock_handle[0].cookie);
3306 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3307 struct mdt_thread_info *info,
3308 struct ldlm_lock **lockp,
3311 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3312 struct ldlm_reply *ldlm_rep = NULL;
3317 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3318 * (for the resend case) or a new lock. Below we will use it to
3319 * replace the original lock.
3321 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3322 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3323 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3324 rc = mdt_object_lock(info, info->mti_object, lhc,
3325 MDS_INODELOCK_XATTR);
3330 rc = mdt_getxattr(info);
3332 if (mdt_info_req(info)->rq_repmsg != NULL)
3333 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3335 if (ldlm_rep == NULL ||
3336 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
3337 mdt_object_unlock(info, info->mti_object, lhc, 1);
3338 RETURN(err_serious(-EFAULT));
3341 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3343 /* This is left for interop instead of adding a new interop flag.
3345 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
3346 if (ldlm_rep->lock_policy_res2) {
3347 mdt_object_unlock(info, info->mti_object, lhc, 1);
3348 RETURN(ELDLM_LOCK_ABORTED);
3352 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3356 static int mdt_intent_getattr(enum mdt_it_code opcode,
3357 struct mdt_thread_info *info,
3358 struct ldlm_lock **lockp,
3361 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3363 struct ldlm_reply *ldlm_rep;
3364 struct mdt_body *reqbody;
3365 struct mdt_body *repbody;
3369 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3372 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3375 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3376 repbody->mbo_eadatasize = 0;
3377 repbody->mbo_aclsize = 0;
3381 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3383 case MDT_IT_GETATTR:
3384 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3388 CERROR("Unsupported intent (%d)\n", opcode);
3389 GOTO(out_shrink, rc = -EINVAL);
3392 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3394 GOTO(out_shrink, rc);
3396 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3397 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3399 /* Get lock from request for possible resent case. */
3400 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3402 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3403 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3405 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3406 ldlm_rep->lock_policy_res2 = 0;
3407 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3408 ldlm_rep->lock_policy_res2) {
3409 lhc->mlh_reg_lh.cookie = 0ull;
3410 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3413 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3416 mdt_exit_ucred(info);
3418 mdt_client_compatibility(info);
3419 rc2 = mdt_fix_reply(info);
3425 static int mdt_intent_layout(enum mdt_it_code opcode,
3426 struct mdt_thread_info *info,
3427 struct ldlm_lock **lockp,
3430 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3431 struct layout_intent *layout;
3433 struct mdt_object *obj = NULL;
3434 int layout_size = 0;
3438 if (opcode != MDT_IT_LAYOUT) {
3439 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3444 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3448 if (layout->li_opc != LAYOUT_INTENT_ACCESS) {
3449 CERROR("%s: Unsupported layout intent opc %d\n",
3450 mdt_obd_name(info->mti_mdt), layout->li_opc);
3454 fid = &info->mti_tmp_fid2;
3455 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3457 /* Get lock from request for possible resent case. */
3458 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3460 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3462 GOTO(out, rc = PTR_ERR(obj));
3464 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3465 layout_size = mdt_attr_get_eabuf_size(info, obj);
3466 if (layout_size < 0)
3467 GOTO(out_obj, rc = layout_size);
3469 if (layout_size > info->mti_mdt->mdt_max_mdsize)
3470 info->mti_mdt->mdt_max_mdsize = layout_size;
3473 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3474 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3476 rc = req_capsule_server_pack(info->mti_pill);
3480 mdt_object_put(info->mti_env, obj);
3482 if (rc == 0 && lustre_handle_is_used(&lhc->mlh_reg_lh))
3483 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3486 lhc->mlh_reg_lh.cookie = 0;
3491 static int mdt_intent_reint(enum mdt_it_code opcode,
3492 struct mdt_thread_info *info,
3493 struct ldlm_lock **lockp,
3496 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3497 struct ldlm_reply *rep = NULL;
3501 static const struct req_format *intent_fmts[REINT_MAX] = {
3502 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3503 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3508 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3512 if (mdt_it_flavor[opcode].it_reint != opc) {
3513 CERROR("Reint code %ld doesn't match intent: %d\n",
3515 RETURN(err_serious(-EPROTO));
3518 /* Get lock from request for possible resent case. */
3519 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3521 rc = mdt_reint_internal(info, lhc, opc);
3523 /* Check whether the reply has been packed successfully. */
3524 if (mdt_info_req(info)->rq_repmsg != NULL)
3525 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3527 RETURN(err_serious(-EFAULT));
3529 /* MDC expects this in any case */
3531 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3533 /* the open lock or the lock for cross-ref object should be
3534 * returned to the client */
3535 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3536 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3537 rep->lock_policy_res2 = 0;
3538 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3542 rep->lock_policy_res2 = clear_serious(rc);
3544 if (rep->lock_policy_res2 == -ENOENT &&
3545 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
3546 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
3547 rep->lock_policy_res2 = 0;
3549 lhc->mlh_reg_lh.cookie = 0ull;
3550 if (rc == -ENOTCONN || rc == -ENODEV ||
3551 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3553 * If it is the disconnect error (ENODEV & ENOCONN), the error
3554 * will be returned by rq_status, and client at ptlrpc layer
3555 * will detect this, then disconnect, reconnect the import
3556 * immediately, instead of impacting the following the rpc.
3561 * For other cases, the error will be returned by intent, and client
3562 * will retrieve the result from intent.
3564 RETURN(ELDLM_LOCK_ABORTED);
3567 static int mdt_intent_code(enum ldlm_intent_flags itcode)
3575 case IT_OPEN|IT_CREAT:
3582 rc = MDT_IT_READDIR;
3585 rc = MDT_IT_GETATTR;
3597 rc = MDT_IT_GETXATTR;
3602 case IT_QUOTA_DQACQ:
3607 CERROR("Unknown intent opcode: 0x%08x\n", itcode);
3614 static int mdt_intent_opc(enum ldlm_intent_flags itopc,
3615 struct mdt_thread_info *info,
3616 struct ldlm_lock **lockp, __u64 flags)
3618 struct req_capsule *pill = info->mti_pill;
3619 struct ptlrpc_request *req = mdt_info_req(info);
3620 struct mdt_it_flavor *flv;
3625 opc = mdt_intent_code(itopc);
3629 if (opc == MDT_IT_QUOTA) {
3630 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3633 RETURN(-EOPNOTSUPP);
3635 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3636 /* pass the request to quota master */
3637 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3638 mdt_info_req(info), lockp,
3643 flv = &mdt_it_flavor[opc];
3644 if (flv->it_fmt != NULL)
3645 req_capsule_extend(pill, flv->it_fmt);
3647 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3651 if (flv->it_flags & MUTABOR &&
3652 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3655 if (flv->it_act != NULL) {
3656 struct ldlm_reply *rep;
3658 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
3660 /* execute policy */
3661 rc = flv->it_act(opc, info, lockp, flags);
3663 /* Check whether the reply has been packed successfully. */
3664 if (req->rq_repmsg != NULL) {
3665 rep = req_capsule_server_get(info->mti_pill,
3667 rep->lock_policy_res2 =
3668 ptlrpc_status_hton(rep->lock_policy_res2);
3675 static int mdt_intent_policy(struct ldlm_namespace *ns,
3676 struct ldlm_lock **lockp, void *req_cookie,
3677 enum ldlm_mode mode, __u64 flags, void *data)
3679 struct tgt_session_info *tsi;
3680 struct mdt_thread_info *info;
3681 struct ptlrpc_request *req = req_cookie;
3682 struct ldlm_intent *it;
3683 struct req_capsule *pill;
3688 LASSERT(req != NULL);
3690 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3692 info = tsi2mdt_info(tsi);
3693 LASSERT(info != NULL);
3694 pill = info->mti_pill;
3695 LASSERT(pill->rc_req == req);
3697 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3698 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3699 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3701 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3705 /* Lock without inodebits makes no sense and will oops
3706 * later in ldlm. Let's check it now to see if we have
3707 * ibits corrupted somewhere in mdt_intent_opc().
3708 * The case for client miss to set ibits has been
3709 * processed by others. */
3710 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3711 lr_type == LDLM_IBITS,
3712 info->mti_dlm_req->lock_desc.\
3713 l_policy_data.l_inodebits.bits != 0));
3715 rc = err_serious(-EFAULT);
3717 /* No intent was provided */
3718 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3719 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3720 rc = req_capsule_server_pack(pill);
3722 rc = err_serious(rc);
3724 mdt_thread_info_fini(info);
3728 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3730 struct seq_server_site *ss = mdt_seq_site(mdt);
3732 if (ss->ss_node_id == 0)
3735 if (ss->ss_client_seq != NULL) {
3736 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3737 ss->ss_client_seq->lcs_exp = NULL;
3740 if (ss->ss_server_fld != NULL) {
3741 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3742 ss->ss_server_fld->lsf_control_exp = NULL;
3746 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3748 struct seq_server_site *ss = mdt_seq_site(mdt);
3753 if (ss->ss_server_seq != NULL)
3754 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3757 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3759 mdt_seq_fini_cli(mdt);
3760 mdt_deregister_seq_exp(mdt);
3762 return seq_site_fini(env, mdt_seq_site(mdt));
3766 * It will retrieve its FLDB entries from MDT0, and it only happens
3767 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3768 * and it needs to refresh FLDB from the MDT0.
3770 static int mdt_register_lwp_callback(void *data)
3773 struct mdt_device *mdt = data;
3774 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3778 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3780 rc = lu_env_init(&env, LCT_MD_THREAD);
3782 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3786 /* Allocate new sequence now to avoid creating local transaction
3787 * in the normal transaction process */
3788 rc = seq_server_check_and_alloc_super(&env,
3789 mdt_seq_site(mdt)->ss_server_seq);
3794 rc = fld_update_from_controller(&env, fld);
3796 CERROR("%s: cannot update controller: rc = %d\n",
3797 mdt_obd_name(mdt), rc);
3806 static int mdt_register_seq_exp(struct mdt_device *mdt)
3808 struct seq_server_site *ss = mdt_seq_site(mdt);
3809 char *lwp_name = NULL;
3812 if (ss->ss_node_id == 0)
3815 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3816 if (lwp_name == NULL)
3817 GOTO(out_free, rc = -ENOMEM);
3819 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3823 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3828 rc = lustre_register_lwp_item(lwp_name,
3829 &ss->ss_server_fld->lsf_control_exp,
3830 mdt_register_lwp_callback, mdt);
3832 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3833 ss->ss_client_seq->lcs_exp = NULL;
3837 if (lwp_name != NULL)
3838 OBD_FREE(lwp_name, MAX_OBD_NAME);
3844 * Init client sequence manager which is used by local MDS to talk to sequence
3845 * controller on remote node.
3847 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3849 struct seq_server_site *ss = mdt_seq_site(mdt);
3854 /* check if this is adding the first MDC and controller is not yet
3856 OBD_ALLOC_PTR(ss->ss_client_seq);
3857 if (ss->ss_client_seq == NULL)
3860 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3861 if (prefix == NULL) {
3862 OBD_FREE_PTR(ss->ss_client_seq);
3863 ss->ss_client_seq = NULL;
3867 /* Note: seq_client_fini will be called in seq_site_fini */
3868 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3869 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3870 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3872 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3874 OBD_FREE_PTR(ss->ss_client_seq);
3875 ss->ss_client_seq = NULL;
3879 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3884 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3886 struct seq_server_site *ss;
3890 ss = mdt_seq_site(mdt);
3891 /* init sequence controller server(MDT0) */
3892 if (ss->ss_node_id == 0) {
3893 OBD_ALLOC_PTR(ss->ss_control_seq);
3894 if (ss->ss_control_seq == NULL)
3897 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3898 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3901 GOTO(out_seq_fini, rc);
3904 /* Init normal sequence server */
3905 OBD_ALLOC_PTR(ss->ss_server_seq);
3906 if (ss->ss_server_seq == NULL)
3907 GOTO(out_seq_fini, rc = -ENOMEM);
3909 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3910 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3912 GOTO(out_seq_fini, rc);
3914 /* init seq client for seq server to talk to seq controller(MDT0) */
3915 rc = mdt_seq_init_cli(env, mdt);
3917 GOTO(out_seq_fini, rc);
3919 if (ss->ss_node_id != 0)
3920 /* register controller export through lwp */
3921 rc = mdt_register_seq_exp(mdt);
3926 mdt_seq_fini(env, mdt);
3934 static int mdt_fld_fini(const struct lu_env *env,
3935 struct mdt_device *m)
3937 struct seq_server_site *ss = mdt_seq_site(m);
3940 if (ss && ss->ss_server_fld) {
3941 fld_server_fini(env, ss->ss_server_fld);
3942 OBD_FREE_PTR(ss->ss_server_fld);
3943 ss->ss_server_fld = NULL;
3949 static int mdt_fld_init(const struct lu_env *env,
3951 struct mdt_device *m)
3953 struct seq_server_site *ss;
3957 ss = mdt_seq_site(m);
3959 OBD_ALLOC_PTR(ss->ss_server_fld);
3960 if (ss->ss_server_fld == NULL)
3961 RETURN(rc = -ENOMEM);
3963 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3966 OBD_FREE_PTR(ss->ss_server_fld);
3967 ss->ss_server_fld = NULL;
3974 static void mdt_stack_pre_fini(const struct lu_env *env,
3975 struct mdt_device *m, struct lu_device *top)
3977 struct lustre_cfg_bufs *bufs;
3978 struct lustre_cfg *lcfg;
3979 struct mdt_thread_info *info;
3984 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3985 LASSERT(info != NULL);
3987 bufs = &info->mti_u.bufs;
3989 LASSERT(m->mdt_child_exp);
3990 LASSERT(m->mdt_child_exp->exp_obd);
3992 /* process cleanup, pass mdt obd name to get obd umount flags */
3993 /* XXX: this is needed because all layers are referenced by
3994 * objects (some of them are pinned by osd, for example *
3995 * the proper solution should be a model where object used
3996 * by osd only doesn't have mdt/mdd slices -bzzz */
3997 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3998 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3999 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
4003 top->ld_ops->ldo_process_config(env, top, lcfg);
4004 lustre_cfg_free(lcfg);
4008 static void mdt_stack_fini(const struct lu_env *env,
4009 struct mdt_device *m, struct lu_device *top)
4011 struct obd_device *obd = mdt2obd_dev(m);
4012 struct lustre_cfg_bufs *bufs;
4013 struct lustre_cfg *lcfg;
4014 struct mdt_thread_info *info;
4018 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4019 LASSERT(info != NULL);
4021 lu_dev_del_linkage(top->ld_site, top);
4023 lu_site_purge(env, top->ld_site, -1);
4025 bufs = &info->mti_u.bufs;
4026 /* process cleanup, pass mdt obd name to get obd umount flags */
4027 /* another purpose is to let all layers to release their objects */
4028 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4033 lustre_cfg_bufs_set_string(bufs, 1, flags);
4034 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
4039 top->ld_ops->ldo_process_config(env, top, lcfg);
4040 lustre_cfg_free(lcfg);
4042 lu_site_purge(env, top->ld_site, -1);
4044 m->mdt_child = NULL;
4045 m->mdt_bottom = NULL;
4047 obd_disconnect(m->mdt_child_exp);
4048 m->mdt_child_exp = NULL;
4050 obd_disconnect(m->mdt_bottom_exp);
4051 m->mdt_child_exp = NULL;
4054 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
4055 const char *next, struct obd_export **exp)
4057 struct obd_connect_data *data = NULL;
4058 struct obd_device *obd;
4062 OBD_ALLOC_PTR(data);
4064 GOTO(out, rc = -ENOMEM);
4066 obd = class_name2obd(next);
4068 CERROR("%s: can't locate next device: %s\n",
4069 mdt_obd_name(m), next);
4070 GOTO(out, rc = -ENOTCONN);
4073 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4074 data->ocd_version = LUSTRE_VERSION_CODE;
4076 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
4078 CERROR("%s: cannot connect to next dev %s (%d)\n",
4079 mdt_obd_name(m), next, rc);
4089 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
4090 struct lustre_cfg *cfg)
4092 char *dev = lustre_cfg_string(cfg, 0);
4093 int rc, name_size, uuid_size;
4094 char *name, *uuid, *p;
4095 struct lustre_cfg_bufs *bufs;
4096 struct lustre_cfg *lcfg;
4097 struct obd_device *obd;
4098 struct lustre_profile *lprof;
4099 struct lu_site *site;
4102 /* in 1.8 we had the only device in the stack - MDS.
4103 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
4104 * in 2.3 OSD is instantiated by obd_mount.c, so we need
4105 * to generate names and setup MDT, MDD. MDT will be using
4106 * generated name to connect to MDD. for MDD the next device
4107 * will be LOD with name taken from so called "profile" which
4108 * is generated by mount_option line
4110 * 1.8 MGS generates config. commands like this:
4111 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
4112 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
4113 * 2.0 MGS generates config. commands like this:
4114 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
4115 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4116 * 3:lustre-MDT0000-mdtlov 4:f
4118 * we generate MDD name from MDT one, just replacing T with D
4120 * after all the preparations, the logical equivalent will be
4121 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
4122 * 3:lustre-MDT0000-mdtlov 4:f
4123 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4124 * 3:lustre-MDD0000 4:f
4126 * notice we build the stack from down to top: MDD first, then MDT */
4128 name_size = MAX_OBD_NAME;
4129 uuid_size = MAX_OBD_NAME;
4131 OBD_ALLOC(name, name_size);
4132 OBD_ALLOC(uuid, uuid_size);
4133 if (name == NULL || uuid == NULL)
4134 GOTO(cleanup_mem, rc = -ENOMEM);
4136 OBD_ALLOC_PTR(bufs);
4138 GOTO(cleanup_mem, rc = -ENOMEM);
4141 p = strstr(name, "-MDT");
4143 GOTO(free_bufs, rc = -ENOMEM);
4146 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
4148 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4149 if (lprof == NULL || lprof->lp_dt == NULL) {
4150 CERROR("can't find the profile: %s\n",
4151 lustre_cfg_string(cfg, 0));
4152 GOTO(free_bufs, rc = -EINVAL);
4155 lustre_cfg_bufs_reset(bufs, name);
4156 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
4157 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4158 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4160 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4162 GOTO(put_profile, rc = -ENOMEM);
4164 rc = class_attach(lcfg);
4166 GOTO(lcfg_cleanup, rc);
4168 obd = class_name2obd(name);
4170 CERROR("Can not find obd %s (%s in config)\n",
4171 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
4172 GOTO(lcfg_cleanup, rc = -EINVAL);
4175 lustre_cfg_free(lcfg);
4177 lustre_cfg_bufs_reset(bufs, name);
4178 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4179 lustre_cfg_bufs_set_string(bufs, 2, dev);
4180 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4182 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4184 GOTO(class_detach, rc = -ENOMEM);
4186 rc = class_setup(obd, lcfg);
4188 GOTO(class_detach, rc);
4190 /* connect to MDD we just setup */
4191 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
4193 GOTO(class_detach, rc);
4195 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
4197 LASSERT(mdt_lu_site(mdt) == NULL);
4198 mdt->mdt_lu_dev.ld_site = site;
4199 site->ls_top_dev = &mdt->mdt_lu_dev;
4200 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
4202 /* now connect to bottom OSD */
4203 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
4204 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
4206 GOTO(class_detach, rc);
4208 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
4210 rc = lu_env_refill((struct lu_env *)env);
4212 CERROR("Failure to refill session: '%d'\n", rc);
4214 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
4219 class_detach(obd, lcfg);
4221 lustre_cfg_free(lcfg);
4223 class_put_profile(lprof);
4228 OBD_FREE(name, name_size);
4230 OBD_FREE(uuid, uuid_size);
4234 /* setup quota master target on MDT0 */
4235 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
4236 struct lustre_cfg *cfg)
4238 struct obd_device *obd;
4239 char *dev = lustre_cfg_string(cfg, 0);
4240 char *qmtname, *uuid, *p;
4241 struct lustre_cfg_bufs *bufs;
4242 struct lustre_cfg *lcfg;
4243 struct lustre_profile *lprof;
4244 struct obd_connect_data *data;
4248 LASSERT(mdt->mdt_qmt_exp == NULL);
4249 LASSERT(mdt->mdt_qmt_dev == NULL);
4251 /* quota master is on MDT0 only for now */
4252 if (mdt->mdt_seq_site.ss_node_id != 0)
4255 /* MGS generates config commands which look as follows:
4256 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4257 * 3:lustre-MDT0000-mdtlov 4:f
4259 * We generate the QMT name from the MDT one, just replacing MD with QM
4260 * after all the preparations, the logical equivalent will be:
4261 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4262 * 3:lustre-MDT0000-osd 4:f */
4263 OBD_ALLOC(qmtname, MAX_OBD_NAME);
4264 OBD_ALLOC(uuid, UUID_MAX);
4265 OBD_ALLOC_PTR(bufs);
4266 OBD_ALLOC_PTR(data);
4267 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
4268 GOTO(cleanup_mem, rc = -ENOMEM);
4270 strcpy(qmtname, dev);
4271 p = strstr(qmtname, "-MDT");
4273 GOTO(cleanup_mem, rc = -ENOMEM);
4274 /* replace MD with QM */
4278 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4280 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4281 if (lprof == NULL || lprof->lp_dt == NULL) {
4282 CERROR("can't find profile for %s\n",
4283 lustre_cfg_string(cfg, 0));
4284 GOTO(cleanup_mem, rc = -EINVAL);
4287 lustre_cfg_bufs_reset(bufs, qmtname);
4288 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4289 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4290 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4292 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4294 GOTO(put_profile, rc = -ENOMEM);
4296 rc = class_attach(lcfg);
4298 GOTO(lcfg_cleanup, rc);
4300 obd = class_name2obd(qmtname);
4302 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4303 lustre_cfg_string(cfg, 0));
4304 GOTO(lcfg_cleanup, rc = -EINVAL);
4307 lustre_cfg_free(lcfg);
4309 lustre_cfg_bufs_reset(bufs, qmtname);
4310 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4311 lustre_cfg_bufs_set_string(bufs, 2, dev);
4313 /* for quota, the next device should be the OSD device */
4314 lustre_cfg_bufs_set_string(bufs, 3,
4315 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4317 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4319 GOTO(class_detach, rc = -ENOMEM);
4321 rc = class_setup(obd, lcfg);
4323 GOTO(class_detach, rc);
4325 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4327 /* configure local quota objects */
4328 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4332 GOTO(class_cleanup, rc);
4334 /* connect to quota master target */
4335 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4336 data->ocd_version = LUSTRE_VERSION_CODE;
4337 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4340 CERROR("cannot connect to quota master device %s (%d)\n",
4342 GOTO(class_cleanup, rc);
4348 class_manual_cleanup(obd);
4349 mdt->mdt_qmt_dev = NULL;
4353 class_detach(obd, lcfg);
4355 lustre_cfg_free(lcfg);
4357 class_put_profile(lprof);
4362 OBD_FREE(qmtname, MAX_OBD_NAME);
4364 OBD_FREE(uuid, UUID_MAX);
4370 /* Shutdown quota master target associated with mdt */
4371 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4375 if (mdt->mdt_qmt_exp == NULL)
4377 LASSERT(mdt->mdt_qmt_dev != NULL);
4379 /* the qmt automatically shuts down when the mdt disconnects */
4380 obd_disconnect(mdt->mdt_qmt_exp);
4381 mdt->mdt_qmt_exp = NULL;
4382 mdt->mdt_qmt_dev = NULL;
4386 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4387 * for now. This will be removed along with converting rest of MDT code
4388 * to use tgt_session_info */
4389 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4391 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4394 rc = mdt_getxattr(info);
4396 mdt_thread_info_fini(info);
4400 static struct tgt_handler mdt_tgt_handlers[] = {
4401 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4402 0, MDS_CONNECT, mdt_tgt_connect,
4403 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4404 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4405 0, MDS_DISCONNECT, tgt_disconnect,
4406 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4407 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4408 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4409 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4410 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4411 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GET_ROOT, mdt_get_root),
4412 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4413 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4415 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4416 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4417 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4418 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4419 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4420 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4421 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4422 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4424 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4425 mdt_hsm_ct_register),
4426 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4427 mdt_hsm_ct_unregister),
4428 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4430 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4432 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4433 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4435 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4440 static struct tgt_handler mdt_sec_ctx_ops[] = {
4441 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4442 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4443 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4446 static struct tgt_handler mdt_quota_ops[] = {
4447 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4450 static struct tgt_opc_slice mdt_common_slice[] = {
4452 .tos_opc_start = MDS_FIRST_OPC,
4453 .tos_opc_end = MDS_LAST_OPC,
4454 .tos_hs = mdt_tgt_handlers
4457 .tos_opc_start = OBD_FIRST_OPC,
4458 .tos_opc_end = OBD_LAST_OPC,
4459 .tos_hs = tgt_obd_handlers
4462 .tos_opc_start = LDLM_FIRST_OPC,
4463 .tos_opc_end = LDLM_LAST_OPC,
4464 .tos_hs = tgt_dlm_handlers
4467 .tos_opc_start = SEC_FIRST_OPC,
4468 .tos_opc_end = SEC_LAST_OPC,
4469 .tos_hs = mdt_sec_ctx_ops
4472 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4473 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4474 .tos_hs = tgt_out_handlers
4477 .tos_opc_start = FLD_FIRST_OPC,
4478 .tos_opc_end = FLD_LAST_OPC,
4479 .tos_hs = fld_handlers
4482 .tos_opc_start = SEQ_FIRST_OPC,
4483 .tos_opc_end = SEQ_LAST_OPC,
4484 .tos_hs = seq_handlers
4487 .tos_opc_start = QUOTA_DQACQ,
4488 .tos_opc_end = QUOTA_LAST_OPC,
4489 .tos_hs = mdt_quota_ops
4492 .tos_opc_start = LLOG_FIRST_OPC,
4493 .tos_opc_end = LLOG_LAST_OPC,
4494 .tos_hs = tgt_llog_handlers
4497 .tos_opc_start = LFSCK_FIRST_OPC,
4498 .tos_opc_end = LFSCK_LAST_OPC,
4499 .tos_hs = tgt_lfsck_handlers
4507 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4509 struct md_device *next = m->mdt_child;
4510 struct lu_device *d = &m->mdt_lu_dev;
4511 struct obd_device *obd = mdt2obd_dev(m);
4512 struct lfsck_stop stop;
4515 stop.ls_status = LS_PAUSED;
4517 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4519 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4520 ping_evictor_stop();
4522 if (m->mdt_opts.mo_coordinator)
4523 mdt_hsm_cdt_stop(m);
4525 mdt_hsm_cdt_fini(m);
4527 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4528 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4530 if (m->mdt_namespace != NULL)
4531 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
4532 d->ld_obd->obd_force);
4534 obd_exports_barrier(obd);
4535 obd_zombie_barrier();
4539 tgt_fini(env, &m->mdt_lut);
4540 mdt_fs_cleanup(env, m);
4541 upcall_cache_cleanup(m->mdt_identity_cache);
4542 m->mdt_identity_cache = NULL;
4544 if (m->mdt_namespace != NULL) {
4545 ldlm_namespace_free_post(m->mdt_namespace);
4546 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4549 if (m->mdt_md_root != NULL) {
4550 mdt_object_put(env, m->mdt_md_root);
4551 m->mdt_md_root = NULL;
4554 mdt_quota_fini(env, m);
4556 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4558 mdt_seq_fini(env, m);
4559 mdt_fld_fini(env, m);
4564 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4566 LASSERT(atomic_read(&d->ld_ref) == 0);
4568 server_put_mount(mdt_obd_name(m), true);
4573 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4575 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4576 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4578 struct mdt_thread_info *info;
4579 struct obd_device *obd;
4580 const char *dev = lustre_cfg_string(cfg, 0);
4581 const char *num = lustre_cfg_string(cfg, 2);
4582 struct lustre_mount_info *lmi = NULL;
4583 struct lustre_sb_info *lsi;
4585 struct seq_server_site *ss_site;
4586 const char *identity_upcall = "NONE";
4587 struct md_device *next;
4593 lu_device_init(&m->mdt_lu_dev, ldt);
4595 * Environment (env) might be missing mdt_thread_key values at that
4596 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4599 * Usually device allocation path doesn't use module key values, but
4600 * mdt has to do a lot of work here, so allocate key value.
4602 rc = lu_env_refill((struct lu_env *)env);
4606 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4607 LASSERT(info != NULL);
4609 obd = class_name2obd(dev);
4610 LASSERT(obd != NULL);
4612 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4613 m->mdt_opts.mo_evict_tgt_nids = 1;
4614 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4616 /* default is coordinator off, it is started through conf_param
4618 m->mdt_opts.mo_coordinator = 0;
4620 lmi = server_get_mount(dev);
4622 CERROR("Cannot get mount info for %s!\n", dev);
4625 lsi = s2lsi(lmi->lmi_sb);
4626 /* CMD is supported only in IAM mode */
4628 node_id = simple_strtol(num, NULL, 10);
4629 obd->u.obt.obt_magic = OBT_MAGIC;
4630 if (lsi->lsi_lmd != NULL &&
4631 lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
4632 m->mdt_skip_lfsck = 1;
4635 m->mdt_squash.rsi_uid = 0;
4636 m->mdt_squash.rsi_gid = 0;
4637 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4638 init_rwsem(&m->mdt_squash.rsi_sem);
4639 spin_lock_init(&m->mdt_lock);
4640 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4641 m->mdt_enable_remote_dir = 0;
4642 m->mdt_enable_remote_dir_gid = 0;
4644 atomic_set(&m->mdt_mds_mds_conns, 0);
4645 atomic_set(&m->mdt_async_commit_count, 0);
4647 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4648 m->mdt_lu_dev.ld_obd = obd;
4649 /* Set this lu_device to obd for error handling purposes. */
4650 obd->obd_lu_dev = &m->mdt_lu_dev;
4652 /* init the stack */
4653 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4655 CERROR("%s: Can't init device stack, rc %d\n",
4656 mdt_obd_name(m), rc);
4661 ss_site = mdt_seq_site(m);
4662 s->ld_seq_site = ss_site;
4665 /* set server index */
4666 ss_site->ss_node_id = node_id;
4668 /* failover is the default
4669 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4670 * assumed whose ss_node_id == 0 XXX
4672 obd->obd_replayable = 1;
4673 /* No connection accepted until configurations will finish */
4674 obd->obd_no_conn = 1;
4676 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4677 char *str = lustre_cfg_string(cfg, 4);
4678 if (strchr(str, 'n')) {
4679 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4680 obd->obd_replayable = 0;
4684 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4686 GOTO(err_fini_stack, rc);
4688 rc = mdt_seq_init(env, m);
4690 GOTO(err_fini_fld, rc);
4692 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4693 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4694 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4695 LDLM_NAMESPACE_SERVER,
4696 LDLM_NAMESPACE_GREEDY,
4698 if (m->mdt_namespace == NULL)
4699 GOTO(err_fini_seq, rc = -ENOMEM);
4701 m->mdt_namespace->ns_lvbp = m;
4702 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4704 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4705 /* set obd_namespace for compatibility with old code */
4706 obd->obd_namespace = m->mdt_namespace;
4708 rc = mdt_hsm_cdt_init(m);
4710 CERROR("%s: error initializing coordinator, rc %d\n",
4711 mdt_obd_name(m), rc);
4712 GOTO(err_free_ns, rc);
4715 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4716 OBD_FAIL_MDS_ALL_REQUEST_NET,
4717 OBD_FAIL_MDS_ALL_REPLY_NET);
4719 GOTO(err_free_hsm, rc);
4721 rc = mdt_fs_setup(env, m, obd, lsi);
4725 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4727 next = m->mdt_child;
4728 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4731 GOTO(err_fs_cleanup, rc);
4733 if (mntopts & MNTOPT_USERXATTR)
4734 m->mdt_opts.mo_user_xattr = 1;
4736 m->mdt_opts.mo_user_xattr = 0;
4738 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4740 GOTO(err_fs_cleanup, rc);
4742 if (mntopts & MNTOPT_ACL)
4743 m->mdt_opts.mo_acl = 1;
4745 m->mdt_opts.mo_acl = 0;
4747 /* XXX: to support suppgid for ACL, we enable identity_upcall
4748 * by default, otherwise, maybe got unexpected -EACCESS. */
4749 if (m->mdt_opts.mo_acl)
4750 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4752 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4754 &mdt_identity_upcall_cache_ops);
4755 if (IS_ERR(m->mdt_identity_cache)) {
4756 rc = PTR_ERR(m->mdt_identity_cache);
4757 m->mdt_identity_cache = NULL;
4758 GOTO(err_fs_cleanup, rc);
4761 rc = mdt_procfs_init(m, dev);
4763 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4764 GOTO(err_recovery, rc);
4767 rc = mdt_quota_init(env, m, cfg);
4769 GOTO(err_procfs, rc);
4771 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4772 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4773 "mdt_ldlm_client", m->mdt_ldlm_client);
4775 ping_evictor_start();
4777 /* recovery will be started upon mdt_prepare()
4778 * when the whole stack is complete and ready
4779 * to serve the requests */
4781 /* Reduce the initial timeout on an MDS because it doesn't need such
4782 * a long timeout as an OST does. Adaptive timeouts will adjust this
4783 * value appropriately. */
4784 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4785 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4791 target_recovery_fini(obd);
4792 upcall_cache_cleanup(m->mdt_identity_cache);
4793 m->mdt_identity_cache = NULL;
4795 mdt_fs_cleanup(env, m);
4797 tgt_fini(env, &m->mdt_lut);
4799 mdt_hsm_cdt_fini(m);
4801 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4802 obd->obd_namespace = m->mdt_namespace = NULL;
4804 mdt_seq_fini(env, m);
4806 mdt_fld_fini(env, m);
4808 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4811 server_put_mount(dev, true);
4815 /* For interoperability, the left element is old parameter, the right one
4816 * is the new version of the parameter, if some parameter is deprecated,
4817 * the new version should be set as NULL. */
4818 static struct cfg_interop_param mdt_interop_param[] = {
4819 { "mdt.group_upcall", NULL },
4820 { "mdt.quota_type", NULL },
4821 { "mdd.quota_type", NULL },
4822 { "mdt.som", NULL },
4823 { "mdt.rootsquash", "mdt.root_squash" },
4824 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4828 /* used by MGS to process specific configurations */
4829 static int mdt_process_config(const struct lu_env *env,
4830 struct lu_device *d, struct lustre_cfg *cfg)
4832 struct mdt_device *m = mdt_dev(d);
4833 struct md_device *md_next = m->mdt_child;
4834 struct lu_device *next = md2lu_dev(md_next);
4838 switch (cfg->lcfg_command) {
4840 struct obd_device *obd = d->ld_obd;
4842 /* For interoperability */
4843 struct cfg_interop_param *ptr = NULL;
4844 struct lustre_cfg *old_cfg = NULL;
4847 param = lustre_cfg_string(cfg, 1);
4848 if (param == NULL) {
4849 CERROR("param is empty\n");
4854 ptr = class_find_old_param(param, mdt_interop_param);
4856 if (ptr->new_param == NULL) {
4858 CWARN("For interoperability, skip this %s."
4859 " It is obsolete.\n", ptr->old_param);
4863 CWARN("Found old param %s, changed it to %s.\n",
4864 ptr->old_param, ptr->new_param);
4867 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4874 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4876 if (rc > 0 || rc == -ENOSYS) {
4877 /* is it an HSM var ? */
4878 rc = class_process_proc_param(PARAM_HSM,
4879 hsm_cdt_get_proc_vars(),
4881 if (rc > 0 || rc == -ENOSYS)
4882 /* we don't understand; pass it on */
4883 rc = next->ld_ops->ldo_process_config(env, next,
4887 if (old_cfg != NULL)
4888 lustre_cfg_free(cfg);
4893 /* others are passed further */
4894 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4900 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4901 const struct lu_object_header *hdr,
4902 struct lu_device *d)
4904 struct mdt_object *mo;
4908 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4910 struct lu_object *o;
4911 struct lu_object_header *h;
4914 h = &mo->mot_header;
4915 lu_object_header_init(h);
4916 lu_object_init(o, h, d);
4917 lu_object_add_top(h, o);
4918 o->lo_ops = &mdt_obj_ops;
4919 spin_lock_init(&mo->mot_write_lock);
4920 mutex_init(&mo->mot_lov_mutex);
4921 init_rwsem(&mo->mot_open_sem);
4927 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4928 const struct lu_object_conf *unused)
4930 struct mdt_device *d = mdt_dev(o->lo_dev);
4931 struct lu_device *under;
4932 struct lu_object *below;
4936 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4937 PFID(lu_object_fid(o)));
4939 under = &d->mdt_child->md_lu_dev;
4940 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4941 if (below != NULL) {
4942 lu_object_add(o, below);
4949 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4951 struct mdt_object *mo = mdt_obj(o);
4952 struct lu_object_header *h;
4956 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4957 PFID(lu_object_fid(o)));
4959 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4960 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4963 lu_object_header_fini(h);
4964 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4969 static int mdt_object_print(const struct lu_env *env, void *cookie,
4970 lu_printer_t p, const struct lu_object *o)
4972 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4974 return (*p)(env, cookie,
4975 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
4976 mdto, mdto->mot_lov_created ? "lov_created" : "",
4977 mdto->mot_cache_attr ? "cache_attr" : "",
4978 mdto->mot_write_count);
4981 static int mdt_prepare(const struct lu_env *env,
4982 struct lu_device *pdev,
4983 struct lu_device *cdev)
4985 struct mdt_device *mdt = mdt_dev(cdev);
4986 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4987 struct obd_device *obd = cdev->ld_obd;
4994 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4998 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
5002 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
5006 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
5007 /* The LFSCK instance is registered just now, so it must be there when
5008 * register the namespace to such instance. */
5009 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
5011 if (mdt->mdt_seq_site.ss_node_id == 0) {
5012 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
5013 &mdt->mdt_md_root_fid);
5018 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
5020 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
5021 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
5022 LASSERT(obd->obd_no_conn);
5023 spin_lock(&obd->obd_dev_lock);
5024 obd->obd_no_conn = 0;
5025 spin_unlock(&obd->obd_dev_lock);
5027 if (obd->obd_recovering == 0)
5028 mdt_postrecov(env, mdt);
5033 const struct lu_device_operations mdt_lu_ops = {
5034 .ldo_object_alloc = mdt_object_alloc,
5035 .ldo_process_config = mdt_process_config,
5036 .ldo_prepare = mdt_prepare,
5039 static const struct lu_object_operations mdt_obj_ops = {
5040 .loo_object_init = mdt_object_init,
5041 .loo_object_free = mdt_object_free,
5042 .loo_object_print = mdt_object_print
5045 static int mdt_obd_set_info_async(const struct lu_env *env,
5046 struct obd_export *exp,
5047 __u32 keylen, void *key,
5048 __u32 vallen, void *val,
5049 struct ptlrpc_request_set *set)
5055 if (KEY_IS(KEY_SPTLRPC_CONF)) {
5056 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
5064 * Match client and server connection feature flags.
5066 * Compute the compatibility flags for a connection request based on
5067 * features mutually supported by client and server.
5069 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
5070 * must not be updated here, otherwise a partially initialized value may
5071 * be exposed. After the connection request is successfully processed,
5072 * the top-level MDT connect request handler atomically updates the export
5073 * connect flags from the obd_connect_data::ocd_connect_flags field of the
5074 * reply. \see mdt_connect().
5076 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
5077 * full struct obd_connect_data. So care must be taken when accessing fields
5078 * that are not present in struct obd_connect_data_v1. See LU-16.
5080 * \param exp the obd_export associated with this client/target pair
5081 * \param mdt the target device for the connection
5082 * \param data stores data for this connect request
5085 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
5086 * \retval -EBADE client and server feature requirements are incompatible
5088 static int mdt_connect_internal(struct obd_export *exp,
5089 struct mdt_device *mdt,
5090 struct obd_connect_data *data)
5092 LASSERT(data != NULL);
5094 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
5096 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
5097 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
5099 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
5101 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5102 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
5103 CWARN("%s: client %s does not support ibits lock, either "
5104 "very old or an invalid client: flags %#llx\n",
5105 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
5106 data->ocd_connect_flags);
5110 if (!mdt->mdt_opts.mo_acl)
5111 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
5113 if (!mdt->mdt_opts.mo_user_xattr)
5114 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
5116 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
5117 data->ocd_brw_size = min(data->ocd_brw_size,
5118 (__u32)MD_MAX_BRW_SIZE);
5119 if (data->ocd_brw_size == 0) {
5120 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx"
5121 " ocd_version: %x ocd_grant: %d "
5122 "ocd_index: %u ocd_brw_size is "
5123 "unexpectedly zero, network data "
5124 "corruption? Refusing connection of this"
5127 exp->exp_client_uuid.uuid,
5128 exp, data->ocd_connect_flags, data->ocd_version,
5129 data->ocd_grant, data->ocd_index);
5134 /* NB: Disregard the rule against updating
5135 * exp_connect_data.ocd_connect_flags in this case, since
5136 * tgt_client_new() needs to know if this is a lightweight
5137 * connection, and it is safe to expose this flag before
5138 * connection processing completes. */
5139 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
5140 spin_lock(&exp->exp_lock);
5141 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
5142 spin_unlock(&exp->exp_lock);
5145 data->ocd_version = LUSTRE_VERSION_CODE;
5147 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
5148 CWARN("%s: MDS requires FID support, but client not\n",
5153 if (OCD_HAS_FLAG(data, PINGLESS)) {
5154 if (ptlrpc_pinger_suppress_pings()) {
5155 spin_lock(&exp->exp_obd->obd_dev_lock);
5156 list_del_init(&exp->exp_obd_chain_timed);
5157 spin_unlock(&exp->exp_obd->obd_dev_lock);
5159 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
5163 data->ocd_max_easize = mdt->mdt_max_ea_size;
5165 /* NB: Disregard the rule against updating
5166 * exp_connect_data.ocd_connect_flags in this case, since
5167 * tgt_client_new() needs to know if this is client supports
5168 * multiple modify RPCs, and it is safe to expose this flag before
5169 * connection processing completes. */
5170 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
5171 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
5172 spin_lock(&exp->exp_lock);
5173 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
5174 spin_unlock(&exp->exp_lock);
5180 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
5181 struct mdt_thread_info *info,
5182 struct mdt_file_data *mfd)
5184 struct lu_context ses;
5188 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
5193 lu_context_enter(&ses);
5195 mdt_ucred(info)->uc_valid = UCRED_OLD;
5196 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
5198 lu_context_exit(&ses);
5199 lu_context_fini(&ses);
5205 static int mdt_export_cleanup(struct obd_export *exp)
5207 struct list_head closing_list;
5208 struct mdt_export_data *med = &exp->exp_mdt_data;
5209 struct obd_device *obd = exp->exp_obd;
5210 struct mdt_device *mdt;
5211 struct mdt_thread_info *info;
5213 struct mdt_file_data *mfd, *n;
5217 INIT_LIST_HEAD(&closing_list);
5218 spin_lock(&med->med_open_lock);
5219 while (!list_empty(&med->med_open_head)) {
5220 struct list_head *tmp = med->med_open_head.next;
5221 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
5223 /* Remove mfd handle so it can't be found again.
5224 * We are consuming the mfd_list reference here. */
5225 class_handle_unhash(&mfd->mfd_handle);
5226 list_move_tail(&mfd->mfd_list, &closing_list);
5228 spin_unlock(&med->med_open_lock);
5229 mdt = mdt_dev(obd->obd_lu_dev);
5230 LASSERT(mdt != NULL);
5232 rc = lu_env_init(&env, LCT_MD_THREAD);
5236 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5237 LASSERT(info != NULL);
5238 memset(info, 0, sizeof *info);
5239 info->mti_env = &env;
5240 info->mti_mdt = mdt;
5241 info->mti_exp = exp;
5243 if (!list_empty(&closing_list)) {
5244 struct md_attr *ma = &info->mti_attr;
5246 /* Close any open files (which may also cause orphan
5248 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5249 list_del_init(&mfd->mfd_list);
5250 ma->ma_need = ma->ma_valid = 0;
5252 /* This file is being closed due to an eviction, it
5253 * could have been modified and now dirty regarding to
5254 * HSM archive, check this!
5255 * The logic here is to mark a file dirty if there's a
5256 * chance it was dirtied before the client was evicted,
5257 * so that we don't have to wait for a release attempt
5258 * before finding out the file was actually dirty and
5259 * fail the release. Aggressively marking it dirty here
5260 * will cause the policy engine to attempt to
5261 * re-archive it; when rearchiving, we can compare the
5262 * current version to the HSM data_version and make the
5263 * archive request into a noop if it's not actually
5266 if (mfd->mfd_mode & FMODE_WRITE)
5267 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
5269 /* Don't unlink orphan on failover umount, LU-184 */
5270 if (exp->exp_flags & OBD_OPT_FAILOVER) {
5271 ma->ma_valid = MA_FLAGS;
5272 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
5274 mdt_mfd_close(info, mfd);
5277 info->mti_mdt = NULL;
5278 /* cleanup client slot early */
5279 /* Do not erase record for recoverable client. */
5280 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5281 tgt_client_del(&env, exp);
5287 static inline void mdt_enable_slc(struct mdt_device *mdt)
5289 if (mdt->mdt_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
5290 mdt->mdt_lut.lut_sync_lock_cancel = BLOCKING_SYNC_ON_CANCEL;
5293 static inline void mdt_disable_slc(struct mdt_device *mdt)
5295 if (mdt->mdt_lut.lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL)
5296 mdt->mdt_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
5299 static int mdt_obd_disconnect(struct obd_export *exp)
5305 class_export_get(exp);
5307 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
5308 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
5309 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
5311 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
5312 mdt_disable_slc(mdt);
5315 rc = server_disconnect_export(exp);
5317 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
5319 rc = mdt_export_cleanup(exp);
5320 nodemap_del_member(exp);
5321 class_export_put(exp);
5325 /* mds_connect copy */
5326 static int mdt_obd_connect(const struct lu_env *env,
5327 struct obd_export **exp, struct obd_device *obd,
5328 struct obd_uuid *cluuid,
5329 struct obd_connect_data *data,
5332 struct obd_export *lexp;
5333 struct lustre_handle conn = { 0 };
5334 struct mdt_device *mdt;
5336 lnet_nid_t *client_nid = localdata;
5339 LASSERT(env != NULL);
5340 LASSERT(data != NULL);
5342 if (!exp || !obd || !cluuid)
5345 mdt = mdt_dev(obd->obd_lu_dev);
5347 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5348 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
5349 atomic_inc(&mdt->mdt_mds_mds_conns);
5350 mdt_enable_slc(mdt);
5354 * first, check whether the stack is ready to handle requests
5355 * XXX: probably not very appropriate method is used now
5356 * at some point we should find a better one
5358 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
5359 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5360 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5361 rc = obd_get_info(env, mdt->mdt_child_exp,
5362 sizeof(KEY_OSP_CONNECTED),
5363 KEY_OSP_CONNECTED, NULL, NULL);
5366 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5369 rc = class_connect(&conn, obd, cluuid);
5373 lexp = class_conn2export(&conn);
5374 LASSERT(lexp != NULL);
5376 rc = nodemap_add_member(*client_nid, lexp);
5377 if (rc != 0 && rc != -EEXIST)
5380 rc = mdt_connect_internal(lexp, mdt, data);
5382 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5385 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5386 rc = tgt_client_new(env, lexp);
5388 mdt_export_stats_init(obd, lexp, localdata);
5392 class_disconnect(lexp);
5393 nodemap_del_member(lexp);
5397 /* Because we do not want this export to be evicted by pinger,
5398 * let's not add this export to the timed chain list. */
5399 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
5400 spin_lock(&lexp->exp_obd->obd_dev_lock);
5401 list_del_init(&lexp->exp_obd_chain_timed);
5402 spin_unlock(&lexp->exp_obd->obd_dev_lock);
5409 static int mdt_obd_reconnect(const struct lu_env *env,
5410 struct obd_export *exp, struct obd_device *obd,
5411 struct obd_uuid *cluuid,
5412 struct obd_connect_data *data,
5415 lnet_nid_t *client_nid = localdata;
5419 if (exp == NULL || obd == NULL || cluuid == NULL)
5422 rc = nodemap_add_member(*client_nid, exp);
5423 if (rc != 0 && rc != -EEXIST)
5426 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5428 mdt_export_stats_init(obd, exp, localdata);
5430 nodemap_del_member(exp);
5435 /* FIXME: Can we avoid using these two interfaces? */
5436 static int mdt_init_export(struct obd_export *exp)
5438 struct mdt_export_data *med = &exp->exp_mdt_data;
5442 INIT_LIST_HEAD(&med->med_open_head);
5443 spin_lock_init(&med->med_open_lock);
5444 spin_lock(&exp->exp_lock);
5445 exp->exp_connecting = 1;
5446 spin_unlock(&exp->exp_lock);
5448 /* self-export doesn't need client data and ldlm initialization */
5449 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5450 &exp->exp_client_uuid)))
5453 rc = tgt_client_alloc(exp);
5457 rc = ldlm_init_export(exp);
5464 tgt_client_free(exp);
5466 CERROR("%s: Failed to initialize export: rc = %d\n",
5467 exp->exp_obd->obd_name, rc);
5471 static int mdt_destroy_export(struct obd_export *exp)
5475 target_destroy_export(exp);
5476 /* destroy can be called from failed obd_setup, so
5477 * checking uuid is safer than obd_self_export */
5478 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5479 &exp->exp_client_uuid)))
5482 ldlm_destroy_export(exp);
5483 tgt_client_free(exp);
5485 LASSERT(list_empty(&exp->exp_outstanding_replies));
5486 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5491 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5492 struct linkea_data *ldata)
5496 LASSERT(ldata->ld_buf->lb_buf != NULL);
5498 if (!mdt_object_exists(mdt_obj))
5501 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5502 ldata->ld_buf, XATTR_NAME_LINK);
5503 if (rc == -ERANGE) {
5504 /* Buf was too small, figure out what we need. */
5505 lu_buf_free(ldata->ld_buf);
5506 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5507 ldata->ld_buf, XATTR_NAME_LINK);
5510 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5511 if (ldata->ld_buf->lb_buf == NULL)
5513 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5514 ldata->ld_buf, XATTR_NAME_LINK);
5519 return linkea_init_with_rec(ldata);
5523 * Given an MDT object, try to look up the full path to the object.
5524 * Part of the MDT layer implementation of lfs fid2path.
5526 * \param[in] info Per-thread common data shared by MDT level handlers.
5527 * \param[in] obj Object to do path lookup of
5528 * \param[in,out] fp User-provided struct to store path information
5529 * \param[in] root_fid Root FID of current path should reach
5531 * \retval 0 Lookup successful, path information stored in fp
5532 * \retval -EAGAIN Lookup failed, usually because object is being moved
5533 * \retval negative errno if there was a problem
5535 static int mdt_path_current(struct mdt_thread_info *info,
5536 struct mdt_object *obj,
5537 struct getinfo_fid2path *fp,
5538 struct lu_fid *root_fid)
5540 struct mdt_device *mdt = info->mti_mdt;
5541 struct mdt_object *mdt_obj;
5542 struct link_ea_header *leh;
5543 struct link_ea_entry *lee;
5544 struct lu_name *tmpname = &info->mti_name;
5545 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5546 struct lu_buf *buf = &info->mti_big_buf;
5549 struct linkea_data ldata = { NULL };
5554 /* temp buffer for path element, the buffer will be finally freed
5555 * in mdt_thread_info_fini */
5556 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5557 if (buf->lb_buf == NULL)
5561 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
5564 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5566 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
5567 struct lu_buf lmv_buf;
5569 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
5570 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
5571 GOTO(out, rc = -ENOENT);
5573 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
5575 mdt_object_get(info->mti_env, mdt_obj);
5577 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5578 if (IS_ERR(mdt_obj))
5579 GOTO(out, rc = PTR_ERR(mdt_obj));
5582 if (!mdt_object_exists(mdt_obj)) {
5583 mdt_object_put(info->mti_env, mdt_obj);
5584 GOTO(out, rc = -ENOENT);
5587 if (mdt_object_remote(mdt_obj)) {
5588 mdt_object_put(info->mti_env, mdt_obj);
5589 GOTO(remote_out, rc = -EREMOTE);
5592 rc = mdt_links_read(info, mdt_obj, &ldata);
5594 mdt_object_put(info->mti_env, mdt_obj);
5599 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5600 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5601 /* If set, use link #linkno for path lookup, otherwise use
5602 link #0. Only do this for the final path element. */
5603 if (first && fp->gf_linkno < leh->leh_reccount) {
5605 for (count = 0; count < fp->gf_linkno; count++) {
5606 lee = (struct link_ea_entry *)
5607 ((char *)lee + reclen);
5608 linkea_entry_unpack(lee, &reclen, tmpname,
5611 if (fp->gf_linkno < leh->leh_reccount - 1)
5612 /* indicate to user there are more links */
5616 lmv_buf.lb_buf = info->mti_xattr_buf;
5617 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5618 /* Check if it is slave stripes */
5619 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5620 &lmv_buf, XATTR_NAME_LMV);
5621 mdt_object_put(info->mti_env, mdt_obj);
5623 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5625 /* For slave stripes, get its master */
5626 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5627 fp->gf_fid = *tmpfid;
5630 } else if (rc < 0 && rc != -ENODATA) {
5636 /* Pack the name in the end of the buffer */
5637 ptr -= tmpname->ln_namelen;
5638 if (ptr - 1 <= fp->gf_u.gf_path)
5639 GOTO(out, rc = -EOVERFLOW);
5640 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5643 /* keep the last resolved fid to the client, so the
5644 * client will build the left path on another MDT for
5646 fp->gf_fid = *tmpfid;
5652 ptr++; /* skip leading / */
5653 memmove(fp->gf_u.gf_path, ptr,
5654 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
5661 * Given an MDT object, use mdt_path_current to get the path.
5662 * Essentially a wrapper to retry mdt_path_current a set number of times
5663 * if -EAGAIN is returned (usually because an object is being moved).
5665 * Part of the MDT layer implementation of lfs fid2path.
5667 * \param[in] info Per-thread common data shared by mdt level handlers.
5668 * \param[in] obj Object to do path lookup of
5669 * \param[in,out] fp User-provided struct for arguments and to store path
5672 * \retval 0 Lookup successful, path information stored in fp
5673 * \retval negative errno if there was a problem
5675 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5676 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
5678 struct mdt_device *mdt = info->mti_mdt;
5683 if (fp->gf_pathlen < 3)
5686 if (root_fid == NULL)
5687 root_fid = &mdt->mdt_md_root_fid;
5689 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
5690 fp->gf_u.gf_path[0] = '\0';
5694 /* Retry multiple times in case file is being moved */
5695 while (tries-- && rc == -EAGAIN)
5696 rc = mdt_path_current(info, obj, fp, root_fid);
5702 * Get the full path of the provided FID, as of changelog record recno.
5704 * This checks sanity and looks up object for user provided FID
5705 * before calling the actual path lookup code.
5707 * Part of the MDT layer implementation of lfs fid2path.
5709 * \param[in] info Per-thread common data shared by mdt level handlers.
5710 * \param[in,out] fp User-provided struct for arguments and to store path
5713 * \retval 0 Lookup successful, path information and recno stored in fp
5714 * \retval -ENOENT, object does not exist
5715 * \retval negative errno if there was a problem
5717 static int mdt_fid2path(struct mdt_thread_info *info,
5718 struct lu_fid *root_fid,
5719 struct getinfo_fid2path *fp)
5721 struct mdt_device *mdt = info->mti_mdt;
5722 struct mdt_object *obj;
5726 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
5727 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5729 if (!fid_is_sane(&fp->gf_fid))
5732 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5733 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
5734 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
5735 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5739 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5742 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5743 PFID(&fp->gf_fid), rc);
5747 if (mdt_object_remote(obj))
5749 else if (!mdt_object_exists(obj))
5755 mdt_object_put(info->mti_env, obj);
5756 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5757 PFID(&fp->gf_fid), rc);
5761 rc = mdt_path(info, obj, fp, root_fid);
5763 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
5764 PFID(&fp->gf_fid), fp->gf_u.gf_path,
5765 fp->gf_recno, fp->gf_linkno);
5767 mdt_object_put(info->mti_env, obj);
5772 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
5773 void *val, int vallen)
5775 struct getinfo_fid2path *fpout, *fpin;
5776 struct lu_fid *root_fid = NULL;
5779 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5782 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5783 lustre_swab_fid2path(fpin);
5785 memcpy(fpout, fpin, sizeof(*fpin));
5786 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5789 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
5790 sizeof(struct lu_fid)) {
5791 /* client sent its root FID, which is normally fileset FID */
5792 root_fid = fpin->gf_u.gf_root_fid;
5793 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5794 lustre_swab_lu_fid(root_fid);
5796 if (root_fid != NULL && !fid_is_sane(root_fid))
5800 rc = mdt_fid2path(info, root_fid, fpout);
5804 int mdt_get_info(struct tgt_session_info *tsi)
5814 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5816 CDEBUG(D_IOCTL, "No GETINFO key\n");
5817 RETURN(err_serious(-EFAULT));
5819 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5822 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5823 if (vallen == NULL) {
5824 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5825 tgt_name(tsi->tsi_tgt));
5826 RETURN(err_serious(-EFAULT));
5829 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5831 rc = req_capsule_server_pack(tsi->tsi_pill);
5833 RETURN(err_serious(rc));
5835 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5836 if (valout == NULL) {
5837 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5838 tgt_name(tsi->tsi_tgt));
5839 RETURN(err_serious(-EFAULT));
5842 if (KEY_IS(KEY_FID2PATH)) {
5843 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5845 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
5846 mdt_thread_info_fini(info);
5853 /* Pass the ioc down */
5854 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5855 unsigned int cmd, int len, void *data)
5857 struct lu_context ioctl_session;
5858 struct md_device *next = mdt->mdt_child;
5862 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5865 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5866 lu_context_enter(&ioctl_session);
5867 env->le_ses = &ioctl_session;
5869 LASSERT(next->md_ops->mdo_iocontrol);
5870 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5872 lu_context_exit(&ioctl_session);
5873 lu_context_fini(&ioctl_session);
5877 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5879 struct obd_ioctl_data *data = karg;
5882 struct mdt_object *obj;
5883 struct mdt_lock_handle *lh;
5887 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5888 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5891 fid = (struct lu_fid *)data->ioc_inlbuf1;
5893 if (!fid_is_sane(fid))
5896 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5898 lh = &mti->mti_lh[MDT_LH_PARENT];
5899 mdt_lock_reg_init(lh, LCK_CR);
5901 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5903 RETURN(PTR_ERR(obj));
5905 if (mdt_object_remote(obj)) {
5908 * before calling version get the correct MDS should be
5909 * fid, this is error to find remote object here
5911 CERROR("nonlocal object "DFID"\n", PFID(fid));
5912 } else if (!mdt_object_exists(obj)) {
5913 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5916 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5917 *(__u64 *)data->ioc_inlbuf2 = version;
5920 mdt_object_unlock_put(mti, obj, lh, 1);
5924 /* ioctls on obd dev */
5925 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5926 void *karg, void __user *uarg)
5929 struct obd_device *obd = exp->exp_obd;
5930 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5931 struct dt_device *dt = mdt->mdt_bottom;
5935 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5936 rc = lu_env_init(&env, LCT_MD_THREAD);
5942 rc = mdt_device_sync(&env, mdt);
5944 case OBD_IOC_SET_READONLY:
5945 rc = dt_sync(&env, dt);
5947 rc = dt_ro(&env, dt);
5949 case OBD_IOC_ABORT_RECOVERY:
5950 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5951 obd->obd_abort_recovery = 1;
5952 target_stop_recovery_thread(obd);
5955 case OBD_IOC_CHANGELOG_REG:
5956 case OBD_IOC_CHANGELOG_DEREG:
5957 case OBD_IOC_CHANGELOG_CLEAR:
5958 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5960 case OBD_IOC_START_LFSCK: {
5961 struct md_device *next = mdt->mdt_child;
5962 struct obd_ioctl_data *data = karg;
5963 struct lfsck_start_param lsp;
5965 if (unlikely(data == NULL)) {
5970 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5971 lsp.lsp_index_valid = 0;
5972 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5975 case OBD_IOC_STOP_LFSCK: {
5976 struct md_device *next = mdt->mdt_child;
5977 struct obd_ioctl_data *data = karg;
5978 struct lfsck_stop stop;
5980 stop.ls_status = LS_STOPPED;
5981 /* Old lfsck utils may pass NULL @stop. */
5982 if (data->ioc_inlbuf1 == NULL)
5986 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5988 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5991 case OBD_IOC_QUERY_LFSCK: {
5992 struct md_device *next = mdt->mdt_child;
5993 struct obd_ioctl_data *data = karg;
5995 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
5999 case OBD_IOC_GET_OBJ_VERSION: {
6000 struct mdt_thread_info *mti;
6001 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6002 memset(mti, 0, sizeof *mti);
6003 mti->mti_env = &env;
6007 rc = mdt_ioc_version_get(mti, karg);
6010 case OBD_IOC_CATLOGLIST: {
6011 struct mdt_thread_info *mti;
6013 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6014 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
6015 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
6016 &mti->mti_tmp_fid1);
6021 CERROR("%s: Not supported cmd = %d, rc = %d\n",
6022 mdt_obd_name(mdt), cmd, rc);
6029 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
6031 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
6035 if (!mdt->mdt_skip_lfsck) {
6036 struct lfsck_start_param lsp;
6038 lsp.lsp_start = NULL;
6039 lsp.lsp_index_valid = 0;
6040 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
6041 OBD_IOC_START_LFSCK,
6043 if (rc != 0 && rc != -EALREADY)
6044 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
6045 mdt_obd_name(mdt), rc);
6048 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
6052 static int mdt_obd_postrecov(struct obd_device *obd)
6057 rc = lu_env_init(&env, LCT_MD_THREAD);
6060 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
6065 static struct obd_ops mdt_obd_device_ops = {
6066 .o_owner = THIS_MODULE,
6067 .o_set_info_async = mdt_obd_set_info_async,
6068 .o_connect = mdt_obd_connect,
6069 .o_reconnect = mdt_obd_reconnect,
6070 .o_disconnect = mdt_obd_disconnect,
6071 .o_init_export = mdt_init_export,
6072 .o_destroy_export = mdt_destroy_export,
6073 .o_iocontrol = mdt_iocontrol,
6074 .o_postrecov = mdt_obd_postrecov,
6077 static struct lu_device* mdt_device_fini(const struct lu_env *env,
6078 struct lu_device *d)
6080 struct mdt_device *m = mdt_dev(d);
6087 static struct lu_device *mdt_device_free(const struct lu_env *env,
6088 struct lu_device *d)
6090 struct mdt_device *m = mdt_dev(d);
6093 lu_device_fini(&m->mdt_lu_dev);
6099 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
6100 struct lu_device_type *t,
6101 struct lustre_cfg *cfg)
6103 struct lu_device *l;
6104 struct mdt_device *m;
6111 rc = mdt_init0(env, m, t, cfg);
6113 mdt_device_free(env, l);
6118 l = ERR_PTR(-ENOMEM);
6122 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
6123 LU_KEY_INIT(mdt, struct mdt_thread_info);
6125 static void mdt_key_fini(const struct lu_context *ctx,
6126 struct lu_context_key *key, void* data)
6128 struct mdt_thread_info *info = data;
6130 if (info->mti_big_lmm) {
6131 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
6132 info->mti_big_lmm = NULL;
6133 info->mti_big_lmmsize = 0;
6138 /* context key: mdt_thread_key */
6139 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
6141 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
6143 return lu_ucred(info->mti_env);
6146 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
6148 return lu_ucred_check(info->mti_env);
6152 * Enable/disable COS (Commit On Sharing).
6154 * Set/Clear the COS flag in mdt options.
6156 * \param mdt mdt device
6157 * \param val 0 disables COS, other values enable COS
6159 void mdt_enable_cos(struct mdt_device *mdt, int val)
6164 mdt->mdt_opts.mo_cos = !!val;
6165 rc = lu_env_init(&env, LCT_LOCAL);
6166 if (unlikely(rc != 0)) {
6167 CWARN("%s: lu_env initialization failed, cannot "
6168 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
6171 mdt_device_sync(&env, mdt);
6176 * Check COS (Commit On Sharing) status.
6178 * Return COS flag status.
6180 * \param mdt mdt device
6182 int mdt_cos_is_enabled(struct mdt_device *mdt)
6184 return mdt->mdt_opts.mo_cos != 0;
6187 static struct lu_device_type_operations mdt_device_type_ops = {
6188 .ldto_device_alloc = mdt_device_alloc,
6189 .ldto_device_free = mdt_device_free,
6190 .ldto_device_fini = mdt_device_fini
6193 static struct lu_device_type mdt_device_type = {
6194 .ldt_tags = LU_DEVICE_MD,
6195 .ldt_name = LUSTRE_MDT_NAME,
6196 .ldt_ops = &mdt_device_type_ops,
6197 .ldt_ctx_tags = LCT_MD_THREAD
6200 static int __init mdt_init(void)
6204 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
6205 FID_NOBRACE_LEN + 1);
6206 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
6208 rc = lu_kmem_init(mdt_caches);
6212 rc = mds_mod_init();
6216 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
6217 LUSTRE_MDT_NAME, &mdt_device_type);
6222 lu_kmem_fini(mdt_caches);
6229 static void __exit mdt_exit(void)
6231 class_unregister_type(LUSTRE_MDT_NAME);
6233 lu_kmem_fini(mdt_caches);
6236 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
6237 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
6238 MODULE_VERSION(LUSTRE_VERSION_STRING);
6239 MODULE_LICENSE("GPL");
6241 module_init(mdt_init);
6242 module_exit(mdt_exit);