4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_handler.c
38 * Lustre Metadata Target (mdt) request handler
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
43 * Author: Mike Shaver <shaver@clusterfs.com>
44 * Author: Nikita Danilov <nikita@clusterfs.com>
45 * Author: Huang Hua <huanghua@clusterfs.com>
46 * Author: Yury Umanets <umka@clusterfs.com>
49 #define DEBUG_SUBSYSTEM S_MDS
51 #include <linux/module.h>
53 * struct OBD_{ALLOC,FREE}*()
55 #include <obd_support.h>
56 #include <lustre_ioctl.h>
57 /* struct ptlrpc_request */
58 #include <lustre_net.h>
59 /* struct obd_export */
60 #include <lustre_export.h>
61 /* struct obd_device */
64 #include <dt_object.h>
65 #include <lustre_mds.h>
66 #include <lustre_log.h>
67 #include "mdt_internal.h"
68 #include <lustre_acl.h>
69 #include <lustre_param.h>
70 #include <lustre_quota.h>
71 #include <lustre_lfsck.h>
72 #include <lustre_nodemap.h>
74 mdl_mode_t mdt_mdl_lock_modes[] = {
75 [LCK_MINMODE] = MDL_MINMODE,
82 [LCK_GROUP] = MDL_GROUP
85 ldlm_mode_t 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, ldlm_mode_t 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, ldlm_mode_t lock_mode,
169 const struct lu_name *lname)
171 lh->mlh_reg_mode = lock_mode;
172 lh->mlh_rreg_mode = lock_mode;
173 lh->mlh_type = MDT_PDO_LOCK;
175 if (lu_name_is_valid(lname)) {
176 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
178 /* XXX Workaround for LU-2856
180 * Zero is a valid return value of full_name_hash, but
181 * several users of mlh_pdo_hash assume a non-zero
182 * hash value. We therefore map zero onto an
183 * arbitrary, but consistent value (1) to avoid
184 * problems further down the road. */
185 if (unlikely(lh->mlh_pdo_hash == 0))
186 lh->mlh_pdo_hash = 1;
188 lh->mlh_pdo_hash = 0;
192 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
193 struct mdt_lock_handle *lh)
199 * Any dir access needs couple of locks:
201 * 1) on part of dir we gonna take lookup/modify;
203 * 2) on whole dir to protect it from concurrent splitting and/or to
204 * flush client's cache for readdir().
206 * so, for a given mode and object this routine decides what lock mode
207 * to use for lock #2:
209 * 1) if caller's gonna lookup in dir then we need to protect dir from
210 * being splitted only - LCK_CR
212 * 2) if caller's gonna modify dir then we need to protect dir from
213 * being splitted and to flush cache - LCK_CW
215 * 3) if caller's gonna modify dir and that dir seems ready for
216 * splitting then we need to protect it from any type of access
217 * (lookup/modify/split) - LCK_EX --bzzz
220 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
221 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
224 * Ask underlaying level its opinion about preferable PDO lock mode
225 * having access type passed as regular lock mode:
227 * - MDL_MINMODE means that lower layer does not want to specify lock
230 * - MDL_NL means that no PDO lock should be taken. This is used in some
231 * cases. Say, for non-splittable directories no need to use PDO locks
234 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
235 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
237 if (mode != MDL_MINMODE) {
238 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
241 * Lower layer does not want to specify locking mode. We do it
242 * our selves. No special protection is needed, just flush
243 * client's cache on modification and allow concurrent
246 switch (lh->mlh_reg_mode) {
248 lh->mlh_pdo_mode = LCK_EX;
251 lh->mlh_pdo_mode = LCK_CR;
254 lh->mlh_pdo_mode = LCK_CW;
257 CERROR("Not expected lock type (0x%x)\n",
258 (int)lh->mlh_reg_mode);
263 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
267 static int mdt_getstatus(struct tgt_session_info *tsi)
269 struct mdt_thread_info *info = tsi2mdt_info(tsi);
270 struct mdt_device *mdt = info->mti_mdt;
271 struct mdt_body *repbody;
276 rc = mdt_check_ucred(info);
278 GOTO(out, rc = err_serious(rc));
280 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK))
281 GOTO(out, rc = err_serious(-ENOMEM));
283 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
284 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
285 repbody->mbo_valid |= OBD_MD_FLID;
289 mdt_thread_info_fini(info);
293 static int mdt_statfs(struct tgt_session_info *tsi)
295 struct ptlrpc_request *req = tgt_ses_req(tsi);
296 struct mdt_thread_info *info = tsi2mdt_info(tsi);
297 struct md_device *next = info->mti_mdt->mdt_child;
298 struct ptlrpc_service_part *svcpt;
299 struct obd_statfs *osfs;
304 svcpt = req->rq_rqbd->rqbd_svcpt;
306 /* This will trigger a watchdog timeout */
307 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
308 (MDT_SERVICE_WATCHDOG_FACTOR *
309 at_get(&svcpt->scp_at_estimate)) + 1);
311 rc = mdt_check_ucred(info);
313 GOTO(out, rc = err_serious(rc));
315 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
316 GOTO(out, rc = err_serious(-ENOMEM));
318 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
320 GOTO(out, rc = -EPROTO);
322 /** statfs information are cached in the mdt_device */
323 if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
324 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
325 /** statfs data is too old, get up-to-date one */
326 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
329 spin_lock(&info->mti_mdt->mdt_osfs_lock);
330 info->mti_mdt->mdt_osfs = *osfs;
331 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
332 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
334 /** use cached statfs data */
335 spin_lock(&info->mti_mdt->mdt_osfs_lock);
336 *osfs = info->mti_mdt->mdt_osfs;
337 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
341 mdt_counter_incr(req, LPROC_MDT_STATFS);
343 mdt_thread_info_fini(info);
347 #ifdef CONFIG_FS_POSIX_ACL
349 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
351 * \param info thread info object
352 * \param repbody reply to pack ACLs into
353 * \param o mdt object of file to examine
354 * \param nodemap nodemap of client to reply to
356 * \retval -errno error getting or parsing ACL from disk
358 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
359 struct mdt_object *o, struct lu_nodemap *nodemap)
361 const struct lu_env *env = info->mti_env;
362 struct md_object *next = mdt_object_child(o);
363 struct lu_buf *buf = &info->mti_buf;
366 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
367 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
369 if (buf->lb_len == 0)
372 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
374 if (rc == -ENODATA) {
375 repbody->mbo_aclsize = 0;
376 repbody->mbo_valid |= OBD_MD_FLACL;
378 } else if (rc == -EOPNOTSUPP) {
381 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
382 mdt_obd_name(info->mti_mdt),
383 PFID(mdt_object_fid(o)), rc);
386 rc = nodemap_map_acl(nodemap, buf->lb_buf,
387 rc, NODEMAP_FS_TO_CLIENT);
388 /* if all ACLs mapped out, rc is still >= 0 */
390 CERROR("%s: nodemap_map_acl unable to parse "DFID
391 " ACL: rc = %d\n", mdt_obd_name(info->mti_mdt),
392 PFID(mdt_object_fid(o)), rc);
394 repbody->mbo_aclsize = rc;
395 repbody->mbo_valid |= OBD_MD_FLACL;
403 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
404 const struct lu_attr *attr, const struct lu_fid *fid)
406 struct md_attr *ma = &info->mti_attr;
407 struct obd_export *exp = info->mti_exp;
408 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
410 LASSERT(ma->ma_valid & MA_INODE);
412 if (attr->la_valid & LA_ATIME) {
413 b->mbo_atime = attr->la_atime;
414 b->mbo_valid |= OBD_MD_FLATIME;
416 if (attr->la_valid & LA_MTIME) {
417 b->mbo_mtime = attr->la_mtime;
418 b->mbo_valid |= OBD_MD_FLMTIME;
420 if (attr->la_valid & LA_CTIME) {
421 b->mbo_ctime = attr->la_ctime;
422 b->mbo_valid |= OBD_MD_FLCTIME;
424 if (attr->la_valid & LA_FLAGS) {
425 b->mbo_flags = attr->la_flags;
426 b->mbo_valid |= OBD_MD_FLFLAGS;
428 if (attr->la_valid & LA_NLINK) {
429 b->mbo_nlink = attr->la_nlink;
430 b->mbo_valid |= OBD_MD_FLNLINK;
432 if (attr->la_valid & LA_UID) {
433 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
434 NODEMAP_FS_TO_CLIENT,
436 b->mbo_valid |= OBD_MD_FLUID;
438 if (attr->la_valid & LA_GID) {
439 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
440 NODEMAP_FS_TO_CLIENT,
442 b->mbo_valid |= OBD_MD_FLGID;
444 b->mbo_mode = attr->la_mode;
445 if (attr->la_valid & LA_MODE)
446 b->mbo_valid |= OBD_MD_FLMODE;
447 if (attr->la_valid & LA_TYPE)
448 b->mbo_valid |= OBD_MD_FLTYPE;
452 b->mbo_valid |= OBD_MD_FLID;
453 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid="LPX64"\n",
454 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
458 mdt_body_reverse_idmap(info, b);
460 if (!(attr->la_valid & LA_TYPE))
463 b->mbo_rdev = attr->la_rdev;
464 b->mbo_size = attr->la_size;
465 b->mbo_blocks = attr->la_blocks;
467 if (!S_ISREG(attr->la_mode)) {
468 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
469 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
470 /* means no objects are allocated on osts. */
471 LASSERT(!(ma->ma_valid & MA_LOV));
472 /* just ignore blocks occupied by extend attributes on MDS */
474 /* if no object is allocated on osts, the size on mds is valid.
476 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
477 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
478 ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
479 /* A released file stores its size on MDS. */
480 /* But return 1 block for released file, unless tools like tar
481 * will consider it fully sparse. (LU-3864)
483 if (unlikely(b->mbo_size == 0))
487 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
490 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
491 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
492 PFID(fid), (unsigned long long)b->mbo_size);
495 static inline int mdt_body_has_lov(const struct lu_attr *la,
496 const struct mdt_body *body)
498 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
499 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
502 void mdt_client_compatibility(struct mdt_thread_info *info)
504 struct mdt_body *body;
505 struct ptlrpc_request *req = mdt_info_req(info);
506 struct obd_export *exp = req->rq_export;
507 struct md_attr *ma = &info->mti_attr;
508 struct lu_attr *la = &ma->ma_attr;
511 if (exp_connect_layout(exp))
512 /* the client can deal with 16-bit lmm_stripe_count */
515 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
517 if (!mdt_body_has_lov(la, body))
520 /* now we have a reply with a lov for a client not compatible with the
521 * layout lock so we have to clean the layout generation number */
522 if (S_ISREG(la->la_mode))
523 ma->ma_lmm->lmm_layout_gen = 0;
527 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
528 struct mdt_object *o)
530 const struct lu_env *env = info->mti_env;
533 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
542 /* Is it a directory? Let's check for the LMV as well */
543 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
544 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
548 rc2 = mo_xattr_get(env, mdt_object_child(o),
550 XATTR_NAME_DEFAULT_LMV);
552 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
560 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
563 const struct lu_env *env = info->mti_env;
567 LASSERT(info->mti_big_lmm_used == 0);
568 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
572 /* big_lmm may need to be grown */
573 if (info->mti_big_lmmsize < rc) {
574 int size = size_roundup_power2(rc);
576 if (info->mti_big_lmmsize > 0) {
577 /* free old buffer */
578 LASSERT(info->mti_big_lmm);
579 OBD_FREE_LARGE(info->mti_big_lmm,
580 info->mti_big_lmmsize);
581 info->mti_big_lmm = NULL;
582 info->mti_big_lmmsize = 0;
585 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
586 if (info->mti_big_lmm == NULL)
588 info->mti_big_lmmsize = size;
590 LASSERT(info->mti_big_lmmsize >= rc);
592 info->mti_buf.lb_buf = info->mti_big_lmm;
593 info->mti_buf.lb_len = info->mti_big_lmmsize;
594 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
599 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
600 struct md_attr *ma, const char *name)
602 struct md_object *next = mdt_object_child(o);
603 struct lu_buf *buf = &info->mti_buf;
606 if (strcmp(name, XATTR_NAME_LOV) == 0) {
607 buf->lb_buf = ma->ma_lmm;
608 buf->lb_len = ma->ma_lmm_size;
609 LASSERT(!(ma->ma_valid & MA_LOV));
610 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
611 buf->lb_buf = ma->ma_lmv;
612 buf->lb_len = ma->ma_lmv_size;
613 LASSERT(!(ma->ma_valid & MA_LMV));
614 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
615 buf->lb_buf = ma->ma_lmv;
616 buf->lb_len = ma->ma_lmv_size;
617 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
622 rc = mo_xattr_get(info->mti_env, next, buf, name);
626 if (strcmp(name, XATTR_NAME_LOV) == 0) {
627 if (info->mti_big_lmm_used)
628 ma->ma_lmm = info->mti_big_lmm;
630 /* NOT return LOV EA with hole to old client. */
631 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
632 LOV_PATTERN_F_HOLE) &&
633 !(exp_connect_flags(info->mti_exp) &
634 OBD_CONNECT_LFSCK)) {
637 ma->ma_lmm_size = rc;
638 ma->ma_valid |= MA_LOV;
640 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
641 if (info->mti_big_lmm_used)
642 ma->ma_lmv = info->mti_big_lmm;
644 ma->ma_lmv_size = rc;
645 ma->ma_valid |= MA_LMV;
646 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
647 ma->ma_lmv_size = rc;
648 ma->ma_valid |= MA_LMV_DEF;
651 /* Update mdt_max_mdsize so all clients will be aware that */
652 if (info->mti_mdt->mdt_max_mdsize < rc)
653 info->mti_mdt->mdt_max_mdsize = rc;
656 } else if (rc == -ENODATA) {
659 } else if (rc == -ERANGE) {
660 /* Default LMV has fixed size, so it must be able to fit
661 * in the original buffer */
662 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
664 rc = mdt_big_xattr_get(info, o, name);
666 info->mti_big_lmm_used = 1;
674 static int mdt_attr_get_pfid(struct mdt_thread_info *info,
675 struct mdt_object *o, struct lu_fid *pfid)
677 struct lu_buf *buf = &info->mti_buf;
678 struct link_ea_header *leh;
679 struct link_ea_entry *lee;
683 buf->lb_buf = info->mti_big_lmm;
684 buf->lb_len = info->mti_big_lmmsize;
685 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
686 buf, XATTR_NAME_LINK);
687 /* ignore errors, MA_PFID won't be set and it is
688 * up to the caller to treat this as an error */
689 if (rc == -ERANGE || buf->lb_len == 0) {
690 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
691 buf->lb_buf = info->mti_big_lmm;
692 buf->lb_len = info->mti_big_lmmsize;
697 if (rc < sizeof(*leh)) {
698 CERROR("short LinkEA on "DFID": rc = %d\n",
699 PFID(mdt_object_fid(o)), rc);
703 leh = (struct link_ea_header *) buf->lb_buf;
704 lee = (struct link_ea_entry *)(leh + 1);
705 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
706 leh->leh_magic = LINK_EA_MAGIC;
707 leh->leh_reccount = __swab32(leh->leh_reccount);
708 leh->leh_len = __swab64(leh->leh_len);
710 if (leh->leh_magic != LINK_EA_MAGIC)
712 if (leh->leh_reccount == 0)
715 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
716 fid_be_to_cpu(pfid, pfid);
721 int mdt_attr_get_complex(struct mdt_thread_info *info,
722 struct mdt_object *o, struct md_attr *ma)
724 const struct lu_env *env = info->mti_env;
725 struct md_object *next = mdt_object_child(o);
726 struct lu_buf *buf = &info->mti_buf;
727 int need = ma->ma_need;
734 if (mdt_object_exists(o) == 0)
735 GOTO(out, rc = -ENOENT);
736 mode = lu_object_attr(&next->mo_lu);
738 if (need & MA_INODE) {
739 ma->ma_need = MA_INODE;
740 rc = mo_attr_get(env, next, ma);
743 ma->ma_valid |= MA_INODE;
746 if (need & MA_PFID) {
747 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
749 ma->ma_valid |= MA_PFID;
750 /* ignore this error, parent fid is not mandatory */
754 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
755 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
760 if (need & MA_LMV && S_ISDIR(mode)) {
761 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
766 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
767 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
772 if (need & MA_HSM && S_ISREG(mode)) {
773 buf->lb_buf = info->mti_xattr_buf;
774 buf->lb_len = sizeof(info->mti_xattr_buf);
775 CLASSERT(sizeof(struct hsm_attrs) <=
776 sizeof(info->mti_xattr_buf));
777 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
778 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
780 ma->ma_valid |= MA_HSM;
781 else if (rc2 < 0 && rc2 != -ENODATA)
785 #ifdef CONFIG_FS_POSIX_ACL
786 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
787 buf->lb_buf = ma->ma_acl;
788 buf->lb_len = ma->ma_acl_size;
789 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
791 ma->ma_acl_size = rc2;
792 ma->ma_valid |= MA_ACL_DEF;
793 } else if (rc2 == -ENODATA) {
802 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = "LPX64" ma_lmm=%p\n",
803 rc, ma->ma_valid, ma->ma_lmm);
807 static int mdt_getattr_internal(struct mdt_thread_info *info,
808 struct mdt_object *o, int ma_need)
810 struct md_object *next = mdt_object_child(o);
811 const struct mdt_body *reqbody = info->mti_body;
812 struct ptlrpc_request *req = mdt_info_req(info);
813 struct md_attr *ma = &info->mti_attr;
814 struct lu_attr *la = &ma->ma_attr;
815 struct req_capsule *pill = info->mti_pill;
816 const struct lu_env *env = info->mti_env;
817 struct mdt_body *repbody;
818 struct lu_buf *buffer = &info->mti_buf;
819 struct obd_export *exp = info->mti_exp;
820 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
825 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
826 RETURN(err_serious(-ENOMEM));
828 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
832 if (mdt_object_remote(o)) {
833 /* This object is located on remote node.*/
834 /* Return -ENOTSUPP for old client */
835 if (!mdt_is_dne_client(req->rq_export))
836 GOTO(out, rc = -ENOTSUPP);
838 repbody->mbo_fid1 = *mdt_object_fid(o);
839 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
843 if (reqbody->mbo_eadatasize > 0) {
844 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
845 if (buffer->lb_buf == NULL)
846 GOTO(out, rc = -EPROTO);
847 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
850 buffer->lb_buf = NULL;
852 ma_need &= ~(MA_LOV | MA_LMV);
853 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
854 mdt_obd_name(info->mti_mdt),
855 req->rq_export->exp_client_uuid.uuid);
858 /* If it is dir object and client require MEA, then we got MEA */
859 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
860 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
861 /* Assumption: MDT_MD size is enough for lmv size. */
862 ma->ma_lmv = buffer->lb_buf;
863 ma->ma_lmv_size = buffer->lb_len;
864 ma->ma_need = MA_INODE;
865 if (ma->ma_lmv_size > 0) {
866 if (reqbody->mbo_valid & OBD_MD_MEA)
867 ma->ma_need |= MA_LMV;
868 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
869 ma->ma_need |= MA_LMV_DEF;
872 ma->ma_lmm = buffer->lb_buf;
873 ma->ma_lmm_size = buffer->lb_len;
874 ma->ma_need = MA_INODE | MA_HSM;
875 if (ma->ma_lmm_size > 0)
876 ma->ma_need |= MA_LOV;
879 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
880 reqbody->mbo_valid & OBD_MD_FLDIREA &&
881 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
882 /* get default stripe info for this dir. */
883 ma->ma_need |= MA_LOV_DEF;
885 ma->ma_need |= ma_need;
887 rc = mdt_attr_get_complex(info, o, ma);
889 CERROR("%s: getattr error for "DFID": rc = %d\n",
890 mdt_obd_name(info->mti_mdt),
891 PFID(mdt_object_fid(o)), rc);
895 /* if file is released, check if a restore is running */
896 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_RELEASED) &&
897 mdt_hsm_restore_is_running(info, mdt_object_fid(o))) {
898 repbody->mbo_t_state = MS_RESTORE;
899 repbody->mbo_valid |= OBD_MD_TSTATE;
902 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
904 /* the Lustre protocol supposes to return default striping
905 * on the user-visible root if explicitly requested */
906 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
907 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
908 struct lu_fid rootfid;
909 struct mdt_object *root;
910 struct mdt_device *mdt = info->mti_mdt;
912 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
915 root = mdt_object_find(env, mdt, &rootfid);
917 RETURN(PTR_ERR(root));
918 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
919 mdt_object_put(info->mti_env, root);
921 CERROR("%s: getattr error for "DFID": rc = %d\n",
922 mdt_obd_name(info->mti_mdt),
923 PFID(mdt_object_fid(o)), rc);
928 if (likely(ma->ma_valid & MA_INODE))
929 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
933 if (mdt_body_has_lov(la, reqbody)) {
934 if (ma->ma_valid & MA_LOV) {
935 LASSERT(ma->ma_lmm_size);
936 repbody->mbo_eadatasize = ma->ma_lmm_size;
937 if (S_ISDIR(la->la_mode))
938 repbody->mbo_valid |= OBD_MD_FLDIREA;
940 repbody->mbo_valid |= OBD_MD_FLEASIZE;
941 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
943 if (ma->ma_valid & MA_LMV) {
944 /* Return -ENOTSUPP for old client */
945 if (!mdt_is_striped_client(req->rq_export))
948 LASSERT(S_ISDIR(la->la_mode));
949 mdt_dump_lmv(D_INFO, ma->ma_lmv);
950 repbody->mbo_eadatasize = ma->ma_lmv_size;
951 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
953 if (ma->ma_valid & MA_LMV_DEF) {
954 /* Return -ENOTSUPP for old client */
955 if (!mdt_is_striped_client(req->rq_export))
957 LASSERT(S_ISDIR(la->la_mode));
958 repbody->mbo_eadatasize = ma->ma_lmv_size;
959 repbody->mbo_valid |= (OBD_MD_FLDIREA |
962 } else if (S_ISLNK(la->la_mode) &&
963 reqbody->mbo_valid & OBD_MD_LINKNAME) {
964 buffer->lb_buf = ma->ma_lmm;
965 /* eadatasize from client includes NULL-terminator, so
966 * there is no need to read it */
967 buffer->lb_len = reqbody->mbo_eadatasize - 1;
968 rc = mo_readlink(env, next, buffer);
969 if (unlikely(rc <= 0)) {
970 CERROR("%s: readlink failed for "DFID": rc = %d\n",
971 mdt_obd_name(info->mti_mdt),
972 PFID(mdt_object_fid(o)), rc);
975 int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
977 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
979 repbody->mbo_valid |= OBD_MD_LINKNAME;
980 /* we need to report back size with NULL-terminator
981 * because client expects that */
982 repbody->mbo_eadatasize = rc + 1;
983 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
984 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
985 "on "DFID ", expected %d\n",
986 mdt_obd_name(info->mti_mdt),
987 rc, PFID(mdt_object_fid(o)),
988 reqbody->mbo_eadatasize - 1);
990 ((char *)ma->ma_lmm)[rc] = 0;
992 /* If the total CDEBUG() size is larger than a page, it
993 * will print a warning to the console, avoid this by
994 * printing just the last part of the symlink. */
995 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
996 print_limit < rc ? "..." : "", print_limit,
997 (char *)ma->ma_lmm + rc - print_limit, rc);
1002 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1003 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1004 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1005 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1006 repbody->mbo_max_mdsize);
1009 if (exp_connect_rmtclient(info->mti_exp) &&
1010 reqbody->mbo_valid & OBD_MD_FLRMTPERM) {
1011 void *buf = req_capsule_server_get(pill, &RMF_ACL);
1013 /* mdt_getattr_lock only */
1014 rc = mdt_pack_remote_perm(info, o, buf);
1016 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
1017 repbody->mbo_aclsize = 0;
1020 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
1021 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
1024 #ifdef CONFIG_FS_POSIX_ACL
1025 else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1026 (reqbody->mbo_valid & OBD_MD_FLACL))
1027 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1032 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1037 static int mdt_getattr(struct tgt_session_info *tsi)
1039 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1040 struct mdt_object *obj = info->mti_object;
1041 struct req_capsule *pill = info->mti_pill;
1042 struct mdt_body *reqbody;
1043 struct mdt_body *repbody;
1047 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1049 LASSERT(obj != NULL);
1050 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1052 /* Unlike intent case where we need to pre-fill out buffers early on
1053 * in intent policy for ldlm reasons, here we can have a much better
1054 * guess at EA size by just reading it from disk.
1055 * Exceptions are readdir and (missing) directory striping */
1057 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1058 /* No easy way to know how long is the symlink, but it cannot
1059 * be more than PATH_MAX, so we allocate +1 */
1062 /* A special case for fs ROOT: getattr there might fetch
1063 * default EA for entire fs, not just for this dir!
1065 } else if (lu_fid_eq(mdt_object_fid(obj),
1066 &info->mti_mdt->mdt_md_root_fid) &&
1067 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1068 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1070 /* Should the default strping be bigger, mdt_fix_reply
1071 * will reallocate */
1072 rc = DEF_REP_MD_SIZE;
1074 /* Read the actual EA size from disk */
1075 rc = mdt_attr_get_eabuf_size(info, obj);
1079 GOTO(out, rc = err_serious(rc));
1081 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1083 rc = req_capsule_server_pack(pill);
1084 if (unlikely(rc != 0))
1085 GOTO(out, rc = err_serious(rc));
1087 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1088 LASSERT(repbody != NULL);
1089 repbody->mbo_eadatasize = 0;
1090 repbody->mbo_aclsize = 0;
1092 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1093 rc = mdt_init_ucred(info, reqbody);
1095 rc = mdt_check_ucred(info);
1097 GOTO(out_shrink, rc);
1099 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1101 rc = mdt_getattr_internal(info, obj, 0);
1102 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1103 mdt_exit_ucred(info);
1106 mdt_client_compatibility(info);
1107 rc2 = mdt_fix_reply(info);
1111 mdt_thread_info_fini(info);
1116 * Exchange MOF_LOV_CREATED flags between two objects after a
1117 * layout swap. No assumption is made on whether o1 or o2 have
1118 * created objects or not.
1120 * \param[in,out] o1 First swap layout object
1121 * \param[in,out] o2 Second swap layout object
1123 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1127 mutex_lock(&o1->mot_lov_mutex);
1128 mutex_lock(&o2->mot_lov_mutex);
1130 o1_flags = o1->mot_flags;
1131 o1->mot_flags = (o1->mot_flags & ~MOF_LOV_CREATED) |
1132 (o2->mot_flags & MOF_LOV_CREATED);
1134 o2->mot_flags = (o2->mot_flags & ~MOF_LOV_CREATED) |
1135 (o1_flags & MOF_LOV_CREATED);
1137 mutex_unlock(&o2->mot_lov_mutex);
1138 mutex_unlock(&o1->mot_lov_mutex);
1141 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1143 struct mdt_thread_info *info;
1144 struct ptlrpc_request *req = tgt_ses_req(tsi);
1145 struct obd_export *exp = req->rq_export;
1146 struct mdt_object *o1, *o2, *o;
1147 struct mdt_lock_handle *lh1, *lh2;
1148 struct mdc_swap_layouts *msl;
1152 /* client does not support layout lock, so layout swaping
1154 * FIXME: there is a problem for old clients which don't support
1155 * layout lock yet. If those clients have already opened the file
1156 * they won't be notified at all so that old layout may still be
1157 * used to do IO. This can be fixed after file release is landed by
1158 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1159 if (!exp_connect_layout(exp))
1160 RETURN(-EOPNOTSUPP);
1162 info = tsi2mdt_info(tsi);
1164 if (info->mti_dlm_req != NULL)
1165 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1167 o1 = info->mti_object;
1168 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1169 &info->mti_body->mbo_fid2);
1171 GOTO(out, rc = PTR_ERR(o));
1173 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1174 GOTO(put, rc = -ENOENT);
1176 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1177 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1183 /* permission check. Make sure the calling process having permission
1184 * to write both files. */
1185 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1190 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1195 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1197 GOTO(put, rc = -EPROTO);
1199 lh1 = &info->mti_lh[MDT_LH_NEW];
1200 mdt_lock_reg_init(lh1, LCK_EX);
1201 lh2 = &info->mti_lh[MDT_LH_OLD];
1202 mdt_lock_reg_init(lh2, LCK_EX);
1204 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1205 MDS_INODELOCK_XATTR);
1209 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1210 MDS_INODELOCK_XATTR);
1214 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1215 mdt_object_child(o2), msl->msl_flags);
1219 mdt_swap_lov_flag(o1, o2);
1222 mdt_object_unlock(info, o2, lh2, rc);
1224 mdt_object_unlock(info, o1, lh1, rc);
1226 mdt_object_put(info->mti_env, o);
1228 mdt_thread_info_fini(info);
1232 static int mdt_raw_lookup(struct mdt_thread_info *info,
1233 struct mdt_object *parent,
1234 const struct lu_name *lname,
1235 struct ldlm_reply *ldlm_rep)
1237 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1241 LASSERT(!info->mti_cross_ref);
1243 /* Only got the fid of this obj by name */
1244 fid_zero(child_fid);
1245 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1246 lname, child_fid, &info->mti_spec);
1248 struct mdt_body *repbody;
1250 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1251 repbody->mbo_fid1 = *child_fid;
1252 repbody->mbo_valid = OBD_MD_FLID;
1253 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1254 } else if (rc == -ENOENT) {
1255 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1262 * UPDATE lock should be taken against parent, and be released before exit;
1263 * child_bits lock should be taken against child, and be returned back:
1264 * (1)normal request should release the child lock;
1265 * (2)intent request will grant the lock to client.
1267 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1268 struct mdt_lock_handle *lhc,
1270 struct ldlm_reply *ldlm_rep)
1272 struct ptlrpc_request *req = mdt_info_req(info);
1273 struct mdt_body *reqbody = NULL;
1274 struct mdt_object *parent = info->mti_object;
1275 struct mdt_object *child;
1276 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1277 struct lu_name *lname = NULL;
1278 struct mdt_lock_handle *lhp = NULL;
1279 struct ldlm_lock *lock;
1286 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1287 LASSERT(ergo(is_resent,
1288 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1293 if (info->mti_cross_ref) {
1294 /* Only getattr on the child. Parent is on another node. */
1295 mdt_set_disposition(info, ldlm_rep,
1296 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1298 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1300 PFID(mdt_object_fid(child)), ldlm_rep);
1302 rc = mdt_check_resent_lock(info, child, lhc);
1305 } else if (rc > 0) {
1306 mdt_lock_handle_init(lhc);
1307 mdt_lock_reg_init(lhc, LCK_PR);
1310 * Object's name is on another MDS, no lookup or layout
1311 * lock is needed here but update lock is.
1313 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1314 MDS_INODELOCK_LAYOUT);
1315 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1317 rc = mdt_object_lock(info, child, lhc, child_bits);
1322 /* Finally, we can get attr for child. */
1323 if (!mdt_object_exists(child)) {
1324 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1326 "remote object doesn't exist.\n");
1327 mdt_object_unlock(info, child, lhc, 1);
1331 rc = mdt_getattr_internal(info, child, 0);
1332 if (unlikely(rc != 0))
1333 mdt_object_unlock(info, child, lhc, 1);
1338 lname = &info->mti_name;
1339 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1341 if (lu_name_is_valid(lname)) {
1342 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1343 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1344 PNAME(lname), ldlm_rep);
1346 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1347 if (unlikely(reqbody == NULL))
1348 RETURN(err_serious(-EPROTO));
1350 *child_fid = reqbody->mbo_fid2;
1352 if (unlikely(!fid_is_sane(child_fid)))
1353 RETURN(err_serious(-EINVAL));
1355 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1357 PFID(mdt_object_fid(parent)),
1358 PFID(&reqbody->mbo_fid2), ldlm_rep);
1361 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1363 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1364 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1366 "Parent doesn't exist!\n");
1370 if (mdt_object_remote(parent)) {
1371 CERROR("%s: parent "DFID" is on remote target\n",
1372 mdt_obd_name(info->mti_mdt),
1373 PFID(mdt_object_fid(parent)));
1377 if (lu_name_is_valid(lname)) {
1378 /* Always allow to lookup ".." */
1379 if (unlikely(lname->ln_namelen == 2 &&
1380 lname->ln_name[0] == '.' &&
1381 lname->ln_name[1] == '.'))
1382 info->mti_spec.sp_permitted = 1;
1384 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1385 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1390 /* step 1: lock parent only if parent is a directory */
1391 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1392 lhp = &info->mti_lh[MDT_LH_PARENT];
1393 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1394 rc = mdt_object_lock(info, parent, lhp,
1395 MDS_INODELOCK_UPDATE);
1396 if (unlikely(rc != 0))
1400 /* step 2: lookup child's fid by name */
1401 fid_zero(child_fid);
1402 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1403 child_fid, &info->mti_spec);
1405 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1408 GOTO(out_parent, rc);
1411 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1414 *step 3: find the child object by fid & lock it.
1415 * regardless if it is local or remote.
1417 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1418 * set parent dir fid the same as child fid in getattr by fid case
1419 * we should not lu_object_find() the object again, could lead
1420 * to hung if there is a concurrent unlink destroyed the object.
1422 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1423 mdt_object_get(info->mti_env, parent);
1426 child = mdt_object_find(info->mti_env, info->mti_mdt,
1430 if (unlikely(IS_ERR(child)))
1431 GOTO(out_parent, rc = PTR_ERR(child));
1433 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1434 rc = mdt_check_resent_lock(info, child, lhc);
1436 GOTO(out_child, rc);
1437 } else if (rc > 0) {
1438 mdt_lock_handle_init(lhc);
1439 mdt_lock_reg_init(lhc, LCK_PR);
1442 if (!mdt_object_exists(child)) {
1443 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1445 "Object doesn't exist!\n");
1446 GOTO(out_child, rc = -ENOENT);
1449 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1450 mdt_object_exists(child) && !mdt_object_remote(child)) {
1451 struct md_attr *ma = &info->mti_attr;
1454 ma->ma_need = MA_INODE;
1455 rc = mdt_attr_get_complex(info, child, ma);
1456 if (unlikely(rc != 0))
1457 GOTO(out_child, rc);
1459 /* If the file has not been changed for some time, we
1460 * return not only a LOOKUP lock, but also an UPDATE
1461 * lock and this might save us RPC on later STAT. For
1462 * directories, it also let negative dentry cache start
1463 * working for this dir. */
1464 if (ma->ma_valid & MA_INODE &&
1465 ma->ma_attr.la_valid & LA_CTIME &&
1466 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1467 ma->ma_attr.la_ctime < cfs_time_current_sec())
1468 child_bits |= MDS_INODELOCK_UPDATE;
1471 /* layout lock must be granted in a best-effort way
1472 * for IT operations */
1473 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1474 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1475 exp_connect_layout(info->mti_exp) &&
1476 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1477 !mdt_object_remote(child) && ldlm_rep != NULL) {
1478 /* try to grant layout lock for regular file. */
1484 child_bits |= MDS_INODELOCK_LAYOUT;
1485 /* try layout lock, it may fail to be granted due to
1486 * contention at LOOKUP or UPDATE */
1487 if (!mdt_object_lock_try(info, child, lhc,
1489 child_bits &= ~MDS_INODELOCK_LAYOUT;
1490 LASSERT(child_bits != 0);
1491 rc = mdt_object_lock(info, child, lhc,
1497 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1498 * client will enqueue the lock to the remote MDT */
1499 if (mdt_object_remote(child))
1500 child_bits &= ~MDS_INODELOCK_UPDATE;
1501 rc = mdt_object_lock(info, child, lhc, child_bits);
1503 if (unlikely(rc != 0))
1504 GOTO(out_child, rc);
1507 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1509 /* finally, we can get attr for child. */
1510 rc = mdt_getattr_internal(info, child, ma_need);
1511 if (unlikely(rc != 0)) {
1512 mdt_object_unlock(info, child, lhc, 1);
1514 /* Debugging code. */
1515 LDLM_DEBUG(lock, "Returning lock to client");
1516 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1517 &lock->l_resource->lr_name),
1518 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1519 PLDLMRES(lock->l_resource),
1520 PFID(mdt_object_fid(child)));
1523 LDLM_LOCK_PUT(lock);
1527 mdt_object_put(info->mti_env, child);
1530 mdt_object_unlock(info, parent, lhp, 1);
1534 /* normal handler: should release the child lock */
1535 static int mdt_getattr_name(struct tgt_session_info *tsi)
1537 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1538 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1539 struct mdt_body *reqbody;
1540 struct mdt_body *repbody;
1544 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1545 LASSERT(reqbody != NULL);
1546 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1547 LASSERT(repbody != NULL);
1549 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1550 repbody->mbo_eadatasize = 0;
1551 repbody->mbo_aclsize = 0;
1553 rc = mdt_init_ucred(info, reqbody);
1555 GOTO(out_shrink, rc);
1557 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1558 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1559 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1560 lhc->mlh_reg_lh.cookie = 0;
1562 mdt_exit_ucred(info);
1565 mdt_client_compatibility(info);
1566 rc2 = mdt_fix_reply(info);
1569 mdt_thread_info_fini(info);
1573 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1574 void *karg, void *uarg);
1576 static int mdt_set_info(struct tgt_session_info *tsi)
1578 struct ptlrpc_request *req = tgt_ses_req(tsi);
1581 int keylen, vallen, rc = 0;
1585 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1587 DEBUG_REQ(D_HA, req, "no set_info key");
1588 RETURN(err_serious(-EFAULT));
1591 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1594 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1596 DEBUG_REQ(D_HA, req, "no set_info val");
1597 RETURN(err_serious(-EFAULT));
1600 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1603 /* Swab any part of val you need to here */
1604 if (KEY_IS(KEY_READ_ONLY)) {
1605 spin_lock(&req->rq_export->exp_lock);
1607 *exp_connect_flags_ptr(req->rq_export) |=
1610 *exp_connect_flags_ptr(req->rq_export) &=
1611 ~OBD_CONNECT_RDONLY;
1612 spin_unlock(&req->rq_export->exp_lock);
1613 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1614 struct changelog_setinfo *cs = val;
1616 if (vallen != sizeof(*cs)) {
1617 CERROR("%s: bad changelog_clear setinfo size %d\n",
1618 tgt_name(tsi->tsi_tgt), vallen);
1621 if (ptlrpc_req_need_swab(req)) {
1622 __swab64s(&cs->cs_recno);
1623 __swab32s(&cs->cs_id);
1626 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1634 static int mdt_readpage(struct tgt_session_info *tsi)
1636 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1637 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1638 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1639 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1640 struct mdt_body *repbody;
1646 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1647 RETURN(err_serious(-ENOMEM));
1649 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1650 if (repbody == NULL || reqbody == NULL)
1651 RETURN(err_serious(-EFAULT));
1654 * prepare @rdpg before calling lower layers and transfer itself. Here
1655 * reqbody->size contains offset of where to start to read and
1656 * reqbody->nlink contains number bytes to read.
1658 rdpg->rp_hash = reqbody->mbo_size;
1659 if (rdpg->rp_hash != reqbody->mbo_size) {
1660 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1661 rdpg->rp_hash, reqbody->mbo_size);
1665 rdpg->rp_attrs = reqbody->mbo_mode;
1666 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1667 rdpg->rp_attrs |= LUDA_64BITHASH;
1668 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1669 exp_max_brw_size(tsi->tsi_exp));
1670 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1672 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1673 if (rdpg->rp_pages == NULL)
1676 for (i = 0; i < rdpg->rp_npages; ++i) {
1677 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1678 if (rdpg->rp_pages[i] == NULL)
1679 GOTO(free_rdpg, rc = -ENOMEM);
1682 /* call lower layers to fill allocated pages with directory data */
1683 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1685 GOTO(free_rdpg, rc);
1687 /* send pages to client */
1688 rc = tgt_sendpage(tsi, rdpg, rc);
1693 for (i = 0; i < rdpg->rp_npages; i++)
1694 if (rdpg->rp_pages[i] != NULL)
1695 __free_page(rdpg->rp_pages[i]);
1696 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1698 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1704 static int mdt_reint_internal(struct mdt_thread_info *info,
1705 struct mdt_lock_handle *lhc,
1708 struct req_capsule *pill = info->mti_pill;
1709 struct mdt_body *repbody;
1714 rc = mdt_reint_unpack(info, op);
1716 CERROR("Can't unpack reint, rc %d\n", rc);
1717 RETURN(err_serious(rc));
1720 /* for replay (no_create) lmm is not needed, client has it already */
1721 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1722 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1725 /* llog cookies are always 0, the field is kept for compatibility */
1726 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1727 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1729 rc = req_capsule_server_pack(pill);
1731 CERROR("Can't pack response, rc %d\n", rc);
1732 RETURN(err_serious(rc));
1735 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1736 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1738 repbody->mbo_eadatasize = 0;
1739 repbody->mbo_aclsize = 0;
1742 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1744 /* for replay no cookkie / lmm need, because client have this already */
1745 if (info->mti_spec.no_create)
1746 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1747 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1749 rc = mdt_init_ucred_reint(info);
1751 GOTO(out_shrink, rc);
1753 rc = mdt_fix_attr_ucred(info, op);
1755 GOTO(out_ucred, rc = err_serious(rc));
1757 if (mdt_check_resent(info, mdt_reconstruct, lhc)) {
1758 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1759 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1760 GOTO(out_ucred, rc);
1762 rc = mdt_reint_rec(info, lhc);
1765 mdt_exit_ucred(info);
1767 mdt_client_compatibility(info);
1768 rc2 = mdt_fix_reply(info);
1774 static long mdt_reint_opcode(struct ptlrpc_request *req,
1775 const struct req_format **fmt)
1777 struct mdt_device *mdt;
1778 struct mdt_rec_reint *rec;
1781 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1783 opc = rec->rr_opcode;
1784 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1785 if (opc < REINT_MAX && fmt[opc] != NULL)
1786 req_capsule_extend(&req->rq_pill, fmt[opc]);
1788 mdt = mdt_exp2dev(req->rq_export);
1789 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1790 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1791 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1792 opc = err_serious(-EFAULT);
1795 opc = err_serious(-EFAULT);
1800 static int mdt_reint(struct tgt_session_info *tsi)
1804 static const struct req_format *reint_fmts[REINT_MAX] = {
1805 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1806 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1807 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1808 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1809 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1810 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1811 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1812 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
1813 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME
1818 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1820 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1822 * No lock possible here from client to pass it to reint code
1825 rc = mdt_reint_internal(info, NULL, opc);
1826 mdt_thread_info_fini(info);
1831 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1835 /* this should sync the whole device */
1836 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1838 struct dt_device *dt = mdt->mdt_bottom;
1842 rc = dt->dd_ops->dt_sync(env, dt);
1846 /* this should sync this object */
1847 static int mdt_object_sync(struct mdt_thread_info *info)
1849 struct md_object *next;
1853 if (!mdt_object_exists(info->mti_object)) {
1854 CWARN("%s: non existing object "DFID": rc = %d\n",
1855 mdt_obd_name(info->mti_mdt),
1856 PFID(mdt_object_fid(info->mti_object)), -ESTALE);
1859 next = mdt_object_child(info->mti_object);
1860 rc = mo_object_sync(info->mti_env, next);
1865 static int mdt_sync(struct tgt_session_info *tsi)
1867 struct ptlrpc_request *req = tgt_ses_req(tsi);
1868 struct req_capsule *pill = tsi->tsi_pill;
1869 struct mdt_body *body;
1874 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1875 RETURN(err_serious(-ENOMEM));
1877 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
1878 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
1880 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1882 /* sync an object */
1883 rc = mdt_object_sync(info);
1885 const struct lu_fid *fid;
1886 struct lu_attr *la = &info->mti_attr.ma_attr;
1888 info->mti_attr.ma_need = MA_INODE;
1889 info->mti_attr.ma_valid = 0;
1890 rc = mdt_attr_get_complex(info, info->mti_object,
1893 body = req_capsule_server_get(pill,
1895 fid = mdt_object_fid(info->mti_object);
1896 mdt_pack_attr2body(info, body, la, fid);
1899 mdt_thread_info_fini(info);
1902 mdt_counter_incr(req, LPROC_MDT_SYNC);
1908 * Handle quota control requests to consult current usage/limit, but also
1909 * to configure quota enforcement
1911 static int mdt_quotactl(struct tgt_session_info *tsi)
1913 struct obd_export *exp = tsi->tsi_exp;
1914 struct req_capsule *pill = tsi->tsi_pill;
1915 struct obd_quotactl *oqctl, *repoqc;
1917 struct mdt_device *mdt = mdt_exp2dev(exp);
1918 struct lu_device *qmt = mdt->mdt_qmt_dev;
1919 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
1922 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1924 RETURN(err_serious(-EPROTO));
1926 rc = req_capsule_server_pack(pill);
1928 RETURN(err_serious(rc));
1930 switch (oqctl->qc_cmd) {
1931 /* master quotactl */
1934 if (!nodemap_can_setquota(nodemap))
1939 RETURN(-EOPNOTSUPP);
1940 /* slave quotactl */
1945 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1949 /* map uid/gid for remote client */
1951 if (exp_connect_rmtclient(exp)) {
1952 struct lustre_idmap_table *idmap;
1954 idmap = exp->exp_mdt_data.med_idmap;
1956 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1957 oqctl->qc_cmd != Q_GETINFO))
1960 if (oqctl->qc_type == USRQUOTA)
1961 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1963 else if (oqctl->qc_type == GRPQUOTA)
1964 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1969 if (id == CFS_IDMAP_NOTFOUND) {
1970 CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
1975 if (oqctl->qc_type == USRQUOTA)
1976 id = nodemap_map_id(nodemap, NODEMAP_UID,
1977 NODEMAP_CLIENT_TO_FS, id);
1978 else if (oqctl->qc_type == GRPQUOTA)
1979 id = nodemap_map_id(nodemap, NODEMAP_UID,
1980 NODEMAP_CLIENT_TO_FS, id);
1982 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1984 RETURN(err_serious(-EFAULT));
1986 if (oqctl->qc_id != id)
1987 swap(oqctl->qc_id, id);
1989 switch (oqctl->qc_cmd) {
1995 /* forward quotactl request to QMT */
1996 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2001 /* slave quotactl */
2002 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2007 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2011 if (oqctl->qc_id != id)
2012 swap(oqctl->qc_id, id);
2018 /** clone llog ctxt from child (mdd)
2019 * This allows remote llog (replicator) access.
2020 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2021 * context was originally set up, or we can handle them directly.
2022 * I choose the latter, but that means I need any llog
2023 * contexts set up by child to be accessable by the mdt. So we clone the
2024 * context into our context list here.
2026 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2029 struct md_device *next = mdt->mdt_child;
2030 struct llog_ctxt *ctxt;
2033 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2036 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2037 if (rc || ctxt == NULL) {
2041 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2043 CERROR("Can't set mdt ctxt %d\n", rc);
2048 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2049 struct mdt_device *mdt, int idx)
2051 struct llog_ctxt *ctxt;
2053 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2056 /* Put once for the get we just did, and once for the clone */
2057 llog_ctxt_put(ctxt);
2058 llog_ctxt_put(ctxt);
2063 * sec context handlers
2065 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2069 rc = mdt_handle_idmap(tsi);
2071 struct ptlrpc_request *req = tgt_ses_req(tsi);
2074 opc = lustre_msg_get_opc(req->rq_reqmsg);
2075 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2076 sptlrpc_svc_ctx_invalidate(req);
2079 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2085 * quota request handlers
2087 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2089 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2090 struct lu_device *qmt = mdt->mdt_qmt_dev;
2095 RETURN(err_serious(-EOPNOTSUPP));
2097 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2101 struct mdt_object *mdt_object_new(const struct lu_env *env,
2102 struct mdt_device *d,
2103 const struct lu_fid *f)
2105 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2106 struct lu_object *o;
2107 struct mdt_object *m;
2110 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2111 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2112 if (unlikely(IS_ERR(o)))
2113 m = (struct mdt_object *)o;
2119 struct mdt_object *mdt_object_find(const struct lu_env *env,
2120 struct mdt_device *d,
2121 const struct lu_fid *f)
2123 struct lu_object *o;
2124 struct mdt_object *m;
2127 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2128 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2129 if (unlikely(IS_ERR(o)))
2130 m = (struct mdt_object *)o;
2138 * Asyncronous commit for mdt device.
2140 * Pass asynchonous commit call down the MDS stack.
2142 * \param env environment
2143 * \param mdt the mdt device
2145 static void mdt_device_commit_async(const struct lu_env *env,
2146 struct mdt_device *mdt)
2148 struct dt_device *dt = mdt->mdt_bottom;
2151 rc = dt->dd_ops->dt_commit_async(env, dt);
2152 if (unlikely(rc != 0))
2153 CWARN("%s: async commit start failed: rc = %d\n",
2154 mdt_obd_name(mdt), rc);
2158 * Mark the lock as "synchonous".
2160 * Mark the lock to deffer transaction commit to the unlock time.
2162 * \param lock the lock to mark as "synchonous"
2164 * \see mdt_is_lock_sync
2165 * \see mdt_save_lock
2167 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2169 lock->l_ast_data = (void*)1;
2173 * Check whehter the lock "synchonous" or not.
2175 * \param lock the lock to check
2176 * \retval 1 the lock is "synchonous"
2177 * \retval 0 the lock isn't "synchronous"
2179 * \see mdt_set_lock_sync
2180 * \see mdt_save_lock
2182 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2184 return lock->l_ast_data != NULL;
2188 * Blocking AST for mdt locks.
2190 * Starts transaction commit if in case of COS lock conflict or
2191 * deffers such a commit to the mdt_save_lock.
2193 * \param lock the lock which blocks a request or cancelling lock
2194 * \param desc unused
2195 * \param data unused
2196 * \param flag indicates whether this cancelling or blocking callback
2198 * \see ldlm_blocking_ast_nocheck
2200 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2201 void *data, int flag)
2203 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2204 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2208 if (flag == LDLM_CB_CANCELING)
2210 lock_res_and_lock(lock);
2211 if (lock->l_blocking_ast != mdt_blocking_ast) {
2212 unlock_res_and_lock(lock);
2215 if (mdt_cos_is_enabled(mdt) &&
2216 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2217 lock->l_blocking_lock != NULL &&
2218 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2219 mdt_set_lock_sync(lock);
2221 rc = ldlm_blocking_ast_nocheck(lock);
2223 /* There is no lock conflict if l_blocking_lock == NULL,
2224 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2225 * when the last reference to a local lock was released */
2226 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2229 rc = lu_env_init(&env, LCT_LOCAL);
2230 if (unlikely(rc != 0))
2231 CWARN("%s: lu_env initialization failed, cannot "
2232 "start asynchronous commit: rc = %d\n",
2235 mdt_device_commit_async(&env, mdt);
2241 /* Used for cross-MDT lock */
2242 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2243 void *data, int flag)
2245 struct lustre_handle lockh;
2249 case LDLM_CB_BLOCKING:
2250 ldlm_lock2handle(lock, &lockh);
2251 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
2253 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2257 case LDLM_CB_CANCELING:
2258 LDLM_DEBUG(lock, "Revoke remote lock\n");
2266 int mdt_check_resent_lock(struct mdt_thread_info *info,
2267 struct mdt_object *mo,
2268 struct mdt_lock_handle *lhc)
2270 /* the lock might already be gotten in ldlm_handle_enqueue() */
2271 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2272 struct ptlrpc_request *req = mdt_info_req(info);
2273 struct ldlm_lock *lock;
2275 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2276 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2278 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2279 * a resend case, however, it could be already destroyed
2280 * due to client eviction or a raced cancel RPC. */
2281 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
2282 lhc->mlh_reg_lh.cookie);
2286 if (!fid_res_name_eq(mdt_object_fid(mo),
2287 &lock->l_resource->lr_name)) {
2288 CWARN("%s: Although resent, but still not "
2289 "get child lock:"DFID"\n",
2290 info->mti_exp->exp_obd->obd_name,
2291 PFID(mdt_object_fid(mo)));
2292 LDLM_LOCK_PUT(lock);
2295 LDLM_LOCK_PUT(lock);
2301 int mdt_remote_object_lock(struct mdt_thread_info *mti,
2302 struct mdt_object *o, const struct lu_fid *fid,
2303 struct lustre_handle *lh, ldlm_mode_t mode,
2306 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2307 ldlm_policy_data_t *policy = &mti->mti_policy;
2308 struct ldlm_res_id *res_id = &mti->mti_res_id;
2312 LASSERT(mdt_object_remote(o));
2314 fid_build_reg_res_name(fid, res_id);
2316 memset(einfo, 0, sizeof(*einfo));
2317 einfo->ei_type = LDLM_IBITS;
2318 einfo->ei_mode = mode;
2319 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2320 einfo->ei_cb_cp = ldlm_completion_ast;
2321 einfo->ei_enq_slave = 0;
2322 einfo->ei_res_id = res_id;
2324 memset(policy, 0, sizeof(*policy));
2325 policy->l_inodebits.bits = ibits;
2327 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2332 static int mdt_object_local_lock(struct mdt_thread_info *info,
2333 struct mdt_object *o,
2334 struct mdt_lock_handle *lh, __u64 ibits,
2337 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2338 ldlm_policy_data_t *policy = &info->mti_policy;
2339 struct ldlm_res_id *res_id = &info->mti_res_id;
2344 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2345 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2346 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2347 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2349 /* Only enqueue LOOKUP lock for remote object */
2350 if (mdt_object_remote(o))
2351 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2353 if (lh->mlh_type == MDT_PDO_LOCK) {
2354 /* check for exists after object is locked */
2355 if (mdt_object_exists(o) == 0) {
2356 /* Non-existent object shouldn't have PDO lock */
2359 /* Non-dir object shouldn't have PDO lock */
2360 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2365 memset(policy, 0, sizeof(*policy));
2366 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2368 dlmflags = LDLM_FL_ATOMIC_CB;
2370 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2373 * Take PDO lock on whole directory and build correct @res_id for lock
2374 * on part of directory.
2376 if (lh->mlh_pdo_hash != 0) {
2377 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2378 mdt_lock_pdo_mode(info, o, lh);
2379 if (lh->mlh_pdo_mode != LCK_NL) {
2381 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2382 * is never going to be sent to client and we do not
2383 * want it slowed down due to possible cancels.
2385 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2386 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2387 policy, res_id, dlmflags,
2388 info->mti_exp == NULL ? NULL :
2389 &info->mti_exp->exp_handle.h_cookie);
2390 if (unlikely(rc != 0))
2391 GOTO(out_unlock, rc);
2395 * Finish res_id initializing by name hash marking part of
2396 * directory which is taking modification.
2398 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2401 policy->l_inodebits.bits = ibits;
2404 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2405 * going to be sent to client. If it is - mdt_intent_policy() path will
2406 * fix it up and turn FL_LOCAL flag off.
2408 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2409 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2410 info->mti_exp == NULL ? NULL :
2411 &info->mti_exp->exp_handle.h_cookie);
2414 mdt_object_unlock(info, o, lh, 1);
2415 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2416 lh->mlh_pdo_hash != 0 &&
2417 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2418 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2424 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2425 struct mdt_lock_handle *lh, __u64 ibits,
2431 if (!mdt_object_remote(o))
2432 return mdt_object_local_lock(info, o, lh, ibits, nonblock);
2434 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2435 ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2436 MDS_INODELOCK_XATTR);
2437 if (ibits & MDS_INODELOCK_UPDATE) {
2438 /* Sigh, PDO needs to enqueue 2 locks right now, but
2439 * enqueue RPC can only request 1 lock, to avoid extra
2440 * RPC, so it will instead enqueue EX lock for remote
2441 * object anyway XXX*/
2442 if (lh->mlh_type == MDT_PDO_LOCK &&
2443 lh->mlh_pdo_hash != 0) {
2444 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2445 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2446 PFID(mdt_object_fid(o)));
2447 lh->mlh_pdo_hash = 0;
2448 lh->mlh_rreg_mode = LCK_EX;
2449 lh->mlh_type = MDT_REG_LOCK;
2451 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2454 MDS_INODELOCK_UPDATE);
2459 /* Only enqueue LOOKUP lock for remote object */
2460 if (ibits & MDS_INODELOCK_LOOKUP) {
2461 rc = mdt_object_local_lock(info, o, lh,
2462 MDS_INODELOCK_LOOKUP,
2471 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2472 struct mdt_lock_handle *lh, __u64 ibits)
2474 return mdt_object_lock_internal(info, o, lh, ibits, false);
2477 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2478 struct mdt_lock_handle *lh, __u64 ibits)
2480 struct mdt_lock_handle tmp = *lh;
2483 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true);
2491 * Save a lock within request object.
2493 * Keep the lock referenced until whether client ACK or transaction
2494 * commit happens or release the lock immediately depending on input
2495 * parameters. If COS is ON, a write lock is converted to COS lock
2498 * \param info thead info object
2499 * \param h lock handle
2500 * \param mode lock mode
2501 * \param decref force immediate lock releasing
2504 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2505 ldlm_mode_t mode, int decref)
2509 if (lustre_handle_is_used(h)) {
2510 if (decref || !info->mti_has_trans ||
2511 !(mode & (LCK_PW | LCK_EX))){
2512 mdt_fid_unlock(h, mode);
2514 struct mdt_device *mdt = info->mti_mdt;
2515 struct ldlm_lock *lock = ldlm_handle2lock(h);
2516 struct ptlrpc_request *req = mdt_info_req(info);
2519 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2521 /* there is no request if mdt_object_unlock() is called
2522 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2523 if (likely(req != NULL)) {
2524 CDEBUG(D_HA, "request = %p reply state = %p"
2525 " transno = "LPD64"\n", req,
2526 req->rq_reply_state, req->rq_transno);
2527 if (mdt_cos_is_enabled(mdt)) {
2529 ldlm_lock_downgrade(lock, LCK_COS);
2532 ptlrpc_save_lock(req, h, mode, no_ack);
2534 ldlm_lock_decref(h, mode);
2536 if (mdt_is_lock_sync(lock)) {
2537 CDEBUG(D_HA, "found sync-lock,"
2538 " async commit started\n");
2539 mdt_device_commit_async(info->mti_env,
2542 LDLM_LOCK_PUT(lock);
2551 * Unlock mdt object.
2553 * Immeditely release the regular lock and the PDO lock or save the
2554 * lock in request and keep them referenced until client ACK or
2555 * transaction commit.
2557 * \param info thread info object
2558 * \param o mdt object
2559 * \param lh mdt lock handle referencing regular and PDO locks
2560 * \param decref force immediate lock releasing
2562 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2563 struct mdt_lock_handle *lh, int decref)
2567 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2568 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2570 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2571 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2576 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2577 const struct lu_fid *f,
2578 struct mdt_lock_handle *lh,
2581 struct mdt_object *o;
2583 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2587 rc = mdt_object_lock(info, o, lh, ibits);
2589 mdt_object_put(info->mti_env, o);
2596 void mdt_object_unlock_put(struct mdt_thread_info * info,
2597 struct mdt_object * o,
2598 struct mdt_lock_handle *lh,
2601 mdt_object_unlock(info, o, lh, decref);
2602 mdt_object_put(info->mti_env, o);
2606 * Generic code handling requests that have struct mdt_body passed in:
2608 * - extract mdt_body from request and save it in @info, if present;
2610 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2613 * - if HABEO_CORPUS flag is set for this request type check whether object
2614 * actually exists on storage (lu_object_exists()).
2617 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2619 const struct mdt_body *body;
2620 struct mdt_object *obj;
2621 const struct lu_env *env;
2622 struct req_capsule *pill;
2626 env = info->mti_env;
2627 pill = info->mti_pill;
2629 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2633 if (!(body->mbo_valid & OBD_MD_FLID))
2636 if (!fid_is_sane(&body->mbo_fid1)) {
2637 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2641 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2643 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2644 mdt_object_put(env, obj);
2647 info->mti_object = obj;
2656 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2658 struct req_capsule *pill = info->mti_pill;
2662 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2663 rc = mdt_body_unpack(info, flags);
2667 if (rc == 0 && (flags & HABEO_REFERO)) {
2669 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2670 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2672 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2673 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2676 rc = req_capsule_server_pack(pill);
2681 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2683 lh->mlh_type = MDT_NUL_LOCK;
2684 lh->mlh_reg_lh.cookie = 0ull;
2685 lh->mlh_reg_mode = LCK_MINMODE;
2686 lh->mlh_pdo_lh.cookie = 0ull;
2687 lh->mlh_pdo_mode = LCK_MINMODE;
2688 lh->mlh_rreg_lh.cookie = 0ull;
2689 lh->mlh_rreg_mode = LCK_MINMODE;
2692 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2694 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2695 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2699 * Initialize fields of struct mdt_thread_info. Other fields are left in
2700 * uninitialized state, because it's too expensive to zero out whole
2701 * mdt_thread_info (> 1K) on each request arrival.
2703 void mdt_thread_info_init(struct ptlrpc_request *req,
2704 struct mdt_thread_info *info)
2708 info->mti_pill = &req->rq_pill;
2711 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2712 mdt_lock_handle_init(&info->mti_lh[i]);
2714 /* mdt device: it can be NULL while CONNECT */
2715 if (req->rq_export) {
2716 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2717 info->mti_exp = req->rq_export;
2719 info->mti_mdt = NULL;
2720 info->mti_env = req->rq_svc_thread->t_env;
2721 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2723 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2724 info->mti_big_buf = LU_BUF_NULL;
2725 info->mti_body = NULL;
2726 info->mti_object = NULL;
2727 info->mti_dlm_req = NULL;
2728 info->mti_has_trans = 0;
2729 info->mti_cross_ref = 0;
2730 info->mti_opdata = 0;
2731 info->mti_big_lmm_used = 0;
2733 info->mti_spec.no_create = 0;
2734 info->mti_spec.sp_rm_entry = 0;
2735 info->mti_spec.sp_permitted = 0;
2737 info->mti_spec.u.sp_ea.eadata = NULL;
2738 info->mti_spec.u.sp_ea.eadatalen = 0;
2741 void mdt_thread_info_fini(struct mdt_thread_info *info)
2745 if (info->mti_object != NULL) {
2746 mdt_object_put(info->mti_env, info->mti_object);
2747 info->mti_object = NULL;
2750 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2751 mdt_lock_handle_fini(&info->mti_lh[i]);
2752 info->mti_env = NULL;
2753 info->mti_pill = NULL;
2754 info->mti_exp = NULL;
2756 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2757 lu_buf_free(&info->mti_big_buf);
2760 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2762 struct mdt_thread_info *mti;
2764 mti = mdt_th_info(tsi->tsi_env);
2765 LASSERT(mti != NULL);
2767 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2768 if (tsi->tsi_corpus != NULL) {
2769 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2770 lu_object_get(tsi->tsi_corpus);
2772 mti->mti_body = tsi->tsi_mdt_body;
2773 mti->mti_dlm_req = tsi->tsi_dlm_req;
2778 static int mdt_tgt_connect(struct tgt_session_info *tsi)
2780 struct ptlrpc_request *req = tgt_ses_req(tsi);
2785 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
2787 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
2788 set_current_state(TASK_UNINTERRUPTIBLE);
2789 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
2792 rc = tgt_connect(tsi);
2796 rc = mdt_init_idmap(tsi);
2801 obd_disconnect(class_export_get(req->rq_export));
2820 static int mdt_intent_getattr(enum mdt_it_code opcode,
2821 struct mdt_thread_info *info,
2822 struct ldlm_lock **,
2825 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2826 struct mdt_thread_info *info,
2827 struct ldlm_lock **lockp,
2830 static int mdt_intent_layout(enum mdt_it_code opcode,
2831 struct mdt_thread_info *info,
2832 struct ldlm_lock **,
2834 static int mdt_intent_reint(enum mdt_it_code opcode,
2835 struct mdt_thread_info *info,
2836 struct ldlm_lock **,
2839 static struct mdt_it_flavor {
2840 const struct req_format *it_fmt;
2842 int (*it_act)(enum mdt_it_code ,
2843 struct mdt_thread_info *,
2844 struct ldlm_lock **,
2847 } mdt_it_flavor[] = {
2849 .it_fmt = &RQF_LDLM_INTENT,
2850 /*.it_flags = HABEO_REFERO,*/
2852 .it_act = mdt_intent_reint,
2853 .it_reint = REINT_OPEN
2856 .it_fmt = &RQF_LDLM_INTENT,
2858 * OCREAT is not a MUTABOR request as if the file
2860 * We do the extra check of OBD_CONNECT_RDONLY in
2861 * mdt_reint_open() when we really need to create
2865 .it_act = mdt_intent_reint,
2866 .it_reint = REINT_OPEN
2869 .it_fmt = &RQF_LDLM_INTENT,
2870 .it_flags = MUTABOR,
2871 .it_act = mdt_intent_reint,
2872 .it_reint = REINT_CREATE
2874 [MDT_IT_GETATTR] = {
2875 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2876 .it_flags = HABEO_REFERO,
2877 .it_act = mdt_intent_getattr
2879 [MDT_IT_READDIR] = {
2885 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2886 .it_flags = HABEO_REFERO,
2887 .it_act = mdt_intent_getattr
2890 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2891 .it_flags = MUTABOR,
2893 .it_reint = REINT_UNLINK
2897 .it_flags = MUTABOR,
2900 [MDT_IT_GETXATTR] = {
2901 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2902 .it_flags = HABEO_CORPUS,
2903 .it_act = mdt_intent_getxattr
2906 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2908 .it_act = mdt_intent_layout
2913 mdt_intent_lock_replace(struct mdt_thread_info *info,
2914 struct ldlm_lock **lockp,
2915 struct mdt_lock_handle *lh,
2918 struct ptlrpc_request *req = mdt_info_req(info);
2919 struct ldlm_lock *lock = *lockp;
2920 struct ldlm_lock *new_lock;
2922 /* If possible resent found a lock, @lh is set to its handle */
2923 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2925 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2926 lh->mlh_reg_lh.cookie = 0;
2930 LASSERTF(new_lock != NULL,
2931 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2934 * If we've already given this lock to a client once, then we should
2935 * have no readers or writers. Otherwise, we should have one reader
2936 * _or_ writer ref (which will be zeroed below) before returning the
2939 if (new_lock->l_export == req->rq_export) {
2940 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2942 LASSERT(new_lock->l_export == NULL);
2943 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2948 if (new_lock->l_export == req->rq_export) {
2950 * Already gave this to the client, which means that we
2951 * reconstructed a reply.
2953 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2956 LDLM_LOCK_RELEASE(new_lock);
2957 lh->mlh_reg_lh.cookie = 0;
2958 RETURN(ELDLM_LOCK_REPLACED);
2962 * Fixup the lock to be given to the client.
2964 lock_res_and_lock(new_lock);
2965 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2966 * possible blocking AST. */
2967 while (new_lock->l_readers > 0) {
2968 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2969 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2970 new_lock->l_readers--;
2972 while (new_lock->l_writers > 0) {
2973 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2974 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2975 new_lock->l_writers--;
2978 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2979 new_lock->l_blocking_ast = lock->l_blocking_ast;
2980 new_lock->l_completion_ast = lock->l_completion_ast;
2981 new_lock->l_remote_handle = lock->l_remote_handle;
2982 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2984 unlock_res_and_lock(new_lock);
2986 cfs_hash_add(new_lock->l_export->exp_lock_hash,
2987 &new_lock->l_remote_handle,
2988 &new_lock->l_exp_hash);
2990 LDLM_LOCK_RELEASE(new_lock);
2991 lh->mlh_reg_lh.cookie = 0;
2993 RETURN(ELDLM_LOCK_REPLACED);
2996 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
2997 struct ldlm_lock *new_lock,
2998 struct mdt_lock_handle *lh,
3001 struct ptlrpc_request *req = mdt_info_req(info);
3002 struct ldlm_request *dlmreq;
3004 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3007 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3009 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3010 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3012 if (flags & LDLM_FL_RESENT) {
3013 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3014 lh->mlh_reg_mode = new_lock->l_granted_mode;
3016 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3017 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3018 lh->mlh_reg_lh.cookie);
3023 * If the xid matches, then we know this is a resent request, and allow
3024 * it. (It's probably an OPEN, for which we don't send a lock.
3026 if (req_xid_is_last(req))
3030 * This remote handle isn't enqueued, so we never received or processed
3031 * this request. Clear MSG_RESENT, because it can be handled like any
3032 * normal request now.
3034 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3036 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3037 dlmreq->lock_handle[0].cookie);
3040 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3041 struct mdt_thread_info *info,
3042 struct ldlm_lock **lockp,
3045 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3046 struct ldlm_reply *ldlm_rep = NULL;
3050 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3051 * (for the resend case) or a new lock. Below we will use it to
3052 * replace the original lock.
3054 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3055 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3056 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3057 rc = mdt_object_lock(info, info->mti_object, lhc,
3058 MDS_INODELOCK_XATTR);
3063 grc = mdt_getxattr(info);
3065 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3067 if (mdt_info_req(info)->rq_repmsg != NULL)
3068 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3069 if (ldlm_rep == NULL)
3070 RETURN(err_serious(-EFAULT));
3072 ldlm_rep->lock_policy_res2 = grc;
3077 static int mdt_intent_getattr(enum mdt_it_code opcode,
3078 struct mdt_thread_info *info,
3079 struct ldlm_lock **lockp,
3082 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3084 struct ldlm_reply *ldlm_rep;
3085 struct mdt_body *reqbody;
3086 struct mdt_body *repbody;
3090 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3093 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3096 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3097 repbody->mbo_eadatasize = 0;
3098 repbody->mbo_aclsize = 0;
3102 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3104 case MDT_IT_GETATTR:
3105 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3109 CERROR("Unsupported intent (%d)\n", opcode);
3110 GOTO(out_shrink, rc = -EINVAL);
3113 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3115 GOTO(out_shrink, rc);
3117 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3118 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3120 /* Get lock from request for possible resent case. */
3121 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3123 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3124 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3126 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3127 ldlm_rep->lock_policy_res2 = 0;
3128 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3129 ldlm_rep->lock_policy_res2) {
3130 lhc->mlh_reg_lh.cookie = 0ull;
3131 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3134 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3137 mdt_exit_ucred(info);
3139 mdt_client_compatibility(info);
3140 rc2 = mdt_fix_reply(info);
3146 static int mdt_intent_layout(enum mdt_it_code opcode,
3147 struct mdt_thread_info *info,
3148 struct ldlm_lock **lockp,
3151 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3152 struct layout_intent *layout;
3154 struct mdt_object *obj = NULL;
3158 if (opcode != MDT_IT_LAYOUT) {
3159 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3164 fid = &info->mti_tmp_fid2;
3165 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3167 /* Get lock from request for possible resent case. */
3168 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3170 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3172 RETURN(PTR_ERR(obj));
3174 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3175 /* get the length of lsm */
3176 rc = mdt_attr_get_eabuf_size(info, obj);
3178 mdt_object_put(info->mti_env, obj);
3182 if (rc > info->mti_mdt->mdt_max_mdsize)
3183 info->mti_mdt->mdt_max_mdsize = rc;
3186 mdt_object_put(info->mti_env, obj);
3188 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3189 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER, rc);
3190 rc = req_capsule_server_pack(info->mti_pill);
3194 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
3195 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3197 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3198 LASSERT(layout != NULL);
3199 if (layout->li_opc == LAYOUT_INTENT_ACCESS)
3200 /* return to normal/resent ldlm handling */
3203 CERROR("%s: Unsupported layout intent (%d)\n",
3204 mdt_obd_name(info->mti_mdt), layout->li_opc);
3208 static int mdt_intent_reint(enum mdt_it_code opcode,
3209 struct mdt_thread_info *info,
3210 struct ldlm_lock **lockp,
3213 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3214 struct ldlm_reply *rep = NULL;
3218 static const struct req_format *intent_fmts[REINT_MAX] = {
3219 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3220 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3225 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3229 if (mdt_it_flavor[opcode].it_reint != opc) {
3230 CERROR("Reint code %ld doesn't match intent: %d\n",
3232 RETURN(err_serious(-EPROTO));
3235 /* Get lock from request for possible resent case. */
3236 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3238 rc = mdt_reint_internal(info, lhc, opc);
3240 /* Check whether the reply has been packed successfully. */
3241 if (mdt_info_req(info)->rq_repmsg != NULL)
3242 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3244 RETURN(err_serious(-EFAULT));
3246 /* MDC expects this in any case */
3248 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3250 /* the open lock or the lock for cross-ref object should be
3251 * returned to the client */
3252 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3253 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3254 rep->lock_policy_res2 = 0;
3255 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3259 rep->lock_policy_res2 = clear_serious(rc);
3261 if (rep->lock_policy_res2 == -ENOENT &&
3262 mdt_get_disposition(rep, DISP_LOOKUP_NEG))
3263 rep->lock_policy_res2 = 0;
3265 lhc->mlh_reg_lh.cookie = 0ull;
3266 if (rc == -ENOTCONN || rc == -ENODEV ||
3267 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3269 * If it is the disconnect error (ENODEV & ENOCONN), the error
3270 * will be returned by rq_status, and client at ptlrpc layer
3271 * will detect this, then disconnect, reconnect the import
3272 * immediately, instead of impacting the following the rpc.
3277 * For other cases, the error will be returned by intent, and client
3278 * will retrieve the result from intent.
3280 RETURN(ELDLM_LOCK_ABORTED);
3283 static int mdt_intent_code(long itcode)
3291 case IT_OPEN|IT_CREAT:
3298 rc = MDT_IT_READDIR;
3301 rc = MDT_IT_GETATTR;
3313 rc = MDT_IT_GETXATTR;
3318 case IT_QUOTA_DQACQ:
3323 CERROR("Unknown intent opcode: %ld\n", itcode);
3330 static int mdt_intent_opc(long itopc, struct mdt_thread_info *info,
3331 struct ldlm_lock **lockp, __u64 flags)
3333 struct req_capsule *pill = info->mti_pill;
3334 struct ptlrpc_request *req = mdt_info_req(info);
3335 struct mdt_it_flavor *flv;
3340 opc = mdt_intent_code(itopc);
3344 if (opc == MDT_IT_QUOTA) {
3345 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3348 RETURN(-EOPNOTSUPP);
3350 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3351 /* pass the request to quota master */
3352 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3353 mdt_info_req(info), lockp,
3358 flv = &mdt_it_flavor[opc];
3359 if (flv->it_fmt != NULL)
3360 req_capsule_extend(pill, flv->it_fmt);
3362 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3366 if (flv->it_flags & MUTABOR &&
3367 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3370 if (flv->it_act != NULL) {
3371 struct ldlm_reply *rep;
3373 /* execute policy */
3374 rc = flv->it_act(opc, info, lockp, flags);
3376 /* Check whether the reply has been packed successfully. */
3377 if (req->rq_repmsg != NULL) {
3378 rep = req_capsule_server_get(info->mti_pill,
3380 rep->lock_policy_res2 =
3381 ptlrpc_status_hton(rep->lock_policy_res2);
3388 static int mdt_intent_policy(struct ldlm_namespace *ns,
3389 struct ldlm_lock **lockp, void *req_cookie,
3390 ldlm_mode_t mode, __u64 flags, void *data)
3392 struct tgt_session_info *tsi;
3393 struct mdt_thread_info *info;
3394 struct ptlrpc_request *req = req_cookie;
3395 struct ldlm_intent *it;
3396 struct req_capsule *pill;
3401 LASSERT(req != NULL);
3403 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3405 info = tsi2mdt_info(tsi);
3406 LASSERT(info != NULL);
3407 pill = info->mti_pill;
3408 LASSERT(pill->rc_req == req);
3410 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3411 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3412 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3414 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3418 /* Lock without inodebits makes no sense and will oops
3419 * later in ldlm. Let's check it now to see if we have
3420 * ibits corrupted somewhere in mdt_intent_opc().
3421 * The case for client miss to set ibits has been
3422 * processed by others. */
3423 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3424 lr_type == LDLM_IBITS,
3425 info->mti_dlm_req->lock_desc.\
3426 l_policy_data.l_inodebits.bits != 0));
3428 rc = err_serious(-EFAULT);
3430 /* No intent was provided */
3431 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3432 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3433 rc = req_capsule_server_pack(pill);
3435 rc = err_serious(rc);
3437 mdt_thread_info_fini(info);
3441 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3443 struct seq_server_site *ss = mdt_seq_site(mdt);
3445 if (ss->ss_node_id == 0)
3448 if (ss->ss_client_seq != NULL) {
3449 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3450 ss->ss_client_seq->lcs_exp = NULL;
3453 if (ss->ss_server_fld != NULL) {
3454 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3455 ss->ss_server_fld->lsf_control_exp = NULL;
3459 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3461 struct seq_server_site *ss = mdt_seq_site(mdt);
3466 if (ss->ss_server_seq != NULL)
3467 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3470 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3472 mdt_seq_fini_cli(mdt);
3473 mdt_deregister_seq_exp(mdt);
3475 return seq_site_fini(env, mdt_seq_site(mdt));
3479 * It will retrieve its FLDB entries from MDT0, and it only happens
3480 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3481 * and it needs to refresh FLDB from the MDT0.
3483 static int mdt_register_lwp_callback(void *data)
3486 struct mdt_device *mdt = data;
3487 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3491 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3493 rc = lu_env_init(&env, LCT_MD_THREAD);
3495 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3499 /* Allocate new sequence now to avoid creating local transaction
3500 * in the normal transaction process */
3501 rc = seq_server_check_and_alloc_super(&env,
3502 mdt_seq_site(mdt)->ss_server_seq);
3507 rc = fld_update_from_controller(&env, fld);
3509 CERROR("%s: cannot update controller: rc = %d\n",
3510 mdt_obd_name(mdt), rc);
3519 static int mdt_register_seq_exp(struct mdt_device *mdt)
3521 struct seq_server_site *ss = mdt_seq_site(mdt);
3522 char *lwp_name = NULL;
3525 if (ss->ss_node_id == 0)
3528 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3529 if (lwp_name == NULL)
3530 GOTO(out_free, rc = -ENOMEM);
3532 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3536 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3541 rc = lustre_register_lwp_item(lwp_name,
3542 &ss->ss_server_fld->lsf_control_exp,
3543 mdt_register_lwp_callback, mdt);
3545 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3546 ss->ss_client_seq->lcs_exp = NULL;
3550 if (lwp_name != NULL)
3551 OBD_FREE(lwp_name, MAX_OBD_NAME);
3557 * Init client sequence manager which is used by local MDS to talk to sequence
3558 * controller on remote node.
3560 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3562 struct seq_server_site *ss = mdt_seq_site(mdt);
3567 /* check if this is adding the first MDC and controller is not yet
3569 OBD_ALLOC_PTR(ss->ss_client_seq);
3570 if (ss->ss_client_seq == NULL)
3573 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3574 if (prefix == NULL) {
3575 OBD_FREE_PTR(ss->ss_client_seq);
3576 ss->ss_client_seq = NULL;
3580 /* Note: seq_client_fini will be called in seq_site_fini */
3581 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3582 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3583 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3585 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3587 OBD_FREE_PTR(ss->ss_client_seq);
3588 ss->ss_client_seq = NULL;
3592 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3597 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3599 struct seq_server_site *ss;
3603 ss = mdt_seq_site(mdt);
3604 /* init sequence controller server(MDT0) */
3605 if (ss->ss_node_id == 0) {
3606 OBD_ALLOC_PTR(ss->ss_control_seq);
3607 if (ss->ss_control_seq == NULL)
3610 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3611 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3614 GOTO(out_seq_fini, rc);
3617 /* Init normal sequence server */
3618 OBD_ALLOC_PTR(ss->ss_server_seq);
3619 if (ss->ss_server_seq == NULL)
3620 GOTO(out_seq_fini, rc = -ENOMEM);
3622 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3623 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3625 GOTO(out_seq_fini, rc);
3627 /* init seq client for seq server to talk to seq controller(MDT0) */
3628 rc = mdt_seq_init_cli(env, mdt);
3630 GOTO(out_seq_fini, rc);
3632 if (ss->ss_node_id != 0)
3633 /* register controller export through lwp */
3634 rc = mdt_register_seq_exp(mdt);
3639 mdt_seq_fini(env, mdt);
3647 static int mdt_fld_fini(const struct lu_env *env,
3648 struct mdt_device *m)
3650 struct seq_server_site *ss = mdt_seq_site(m);
3653 if (ss && ss->ss_server_fld) {
3654 fld_server_fini(env, ss->ss_server_fld);
3655 OBD_FREE_PTR(ss->ss_server_fld);
3656 ss->ss_server_fld = NULL;
3662 static int mdt_fld_init(const struct lu_env *env,
3664 struct mdt_device *m)
3666 struct seq_server_site *ss;
3670 ss = mdt_seq_site(m);
3672 OBD_ALLOC_PTR(ss->ss_server_fld);
3673 if (ss->ss_server_fld == NULL)
3674 RETURN(rc = -ENOMEM);
3676 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3679 OBD_FREE_PTR(ss->ss_server_fld);
3680 ss->ss_server_fld = NULL;
3687 static void mdt_stack_pre_fini(const struct lu_env *env,
3688 struct mdt_device *m, struct lu_device *top)
3690 struct lustre_cfg_bufs *bufs;
3691 struct lustre_cfg *lcfg;
3692 struct mdt_thread_info *info;
3697 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3698 LASSERT(info != NULL);
3700 bufs = &info->mti_u.bufs;
3702 LASSERT(m->mdt_child_exp);
3703 LASSERT(m->mdt_child_exp->exp_obd);
3705 /* process cleanup, pass mdt obd name to get obd umount flags */
3706 /* XXX: this is needed because all layers are referenced by
3707 * objects (some of them are pinned by osd, for example *
3708 * the proper solution should be a model where object used
3709 * by osd only doesn't have mdt/mdd slices -bzzz */
3710 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3711 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3712 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3716 top->ld_ops->ldo_process_config(env, top, lcfg);
3717 lustre_cfg_free(lcfg);
3721 static void mdt_stack_fini(const struct lu_env *env,
3722 struct mdt_device *m, struct lu_device *top)
3724 struct obd_device *obd = mdt2obd_dev(m);
3725 struct lustre_cfg_bufs *bufs;
3726 struct lustre_cfg *lcfg;
3727 struct mdt_thread_info *info;
3731 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3732 LASSERT(info != NULL);
3734 lu_dev_del_linkage(top->ld_site, top);
3736 lu_site_purge(env, top->ld_site, -1);
3738 bufs = &info->mti_u.bufs;
3739 /* process cleanup, pass mdt obd name to get obd umount flags */
3740 /* another purpose is to let all layers to release their objects */
3741 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3746 lustre_cfg_bufs_set_string(bufs, 1, flags);
3747 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3752 top->ld_ops->ldo_process_config(env, top, lcfg);
3753 lustre_cfg_free(lcfg);
3755 lu_site_purge(env, top->ld_site, -1);
3757 m->mdt_child = NULL;
3758 m->mdt_bottom = NULL;
3760 obd_disconnect(m->mdt_child_exp);
3761 m->mdt_child_exp = NULL;
3763 obd_disconnect(m->mdt_bottom_exp);
3764 m->mdt_child_exp = NULL;
3767 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3768 const char *next, struct obd_export **exp)
3770 struct obd_connect_data *data = NULL;
3771 struct obd_device *obd;
3775 OBD_ALLOC_PTR(data);
3777 GOTO(out, rc = -ENOMEM);
3779 obd = class_name2obd(next);
3781 CERROR("%s: can't locate next device: %s\n",
3782 mdt_obd_name(m), next);
3783 GOTO(out, rc = -ENOTCONN);
3786 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3787 data->ocd_version = LUSTRE_VERSION_CODE;
3789 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3791 CERROR("%s: cannot connect to next dev %s (%d)\n",
3792 mdt_obd_name(m), next, rc);
3802 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3803 struct lustre_cfg *cfg)
3805 char *dev = lustre_cfg_string(cfg, 0);
3806 int rc, name_size, uuid_size;
3807 char *name, *uuid, *p;
3808 struct lustre_cfg_bufs *bufs;
3809 struct lustre_cfg *lcfg;
3810 struct obd_device *obd;
3811 struct lustre_profile *lprof;
3812 struct lu_site *site;
3815 /* in 1.8 we had the only device in the stack - MDS.
3816 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3817 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3818 * to generate names and setup MDT, MDD. MDT will be using
3819 * generated name to connect to MDD. for MDD the next device
3820 * will be LOD with name taken from so called "profile" which
3821 * is generated by mount_option line
3823 * 1.8 MGS generates config. commands like this:
3824 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3825 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3826 * 2.0 MGS generates config. commands like this:
3827 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3828 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3829 * 3:lustre-MDT0000-mdtlov 4:f
3831 * we generate MDD name from MDT one, just replacing T with D
3833 * after all the preparations, the logical equivalent will be
3834 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3835 * 3:lustre-MDT0000-mdtlov 4:f
3836 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3837 * 3:lustre-MDD0000 4:f
3839 * notice we build the stack from down to top: MDD first, then MDT */
3841 name_size = MAX_OBD_NAME;
3842 uuid_size = MAX_OBD_NAME;
3844 OBD_ALLOC(name, name_size);
3845 OBD_ALLOC(uuid, uuid_size);
3846 if (name == NULL || uuid == NULL)
3847 GOTO(cleanup_mem, rc = -ENOMEM);
3849 OBD_ALLOC_PTR(bufs);
3851 GOTO(cleanup_mem, rc = -ENOMEM);
3854 p = strstr(name, "-MDT");
3856 GOTO(free_bufs, rc = -ENOMEM);
3859 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3861 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3862 if (lprof == NULL || lprof->lp_dt == NULL) {
3863 CERROR("can't find the profile: %s\n",
3864 lustre_cfg_string(cfg, 0));
3865 GOTO(free_bufs, rc = -EINVAL);
3868 lustre_cfg_bufs_reset(bufs, name);
3869 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3870 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3871 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3873 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3875 GOTO(free_bufs, rc = -ENOMEM);
3877 rc = class_attach(lcfg);
3879 GOTO(lcfg_cleanup, rc);
3881 obd = class_name2obd(name);
3883 CERROR("Can not find obd %s (%s in config)\n",
3884 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3885 GOTO(lcfg_cleanup, rc = -EINVAL);
3888 lustre_cfg_free(lcfg);
3890 lustre_cfg_bufs_reset(bufs, name);
3891 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3892 lustre_cfg_bufs_set_string(bufs, 2, dev);
3893 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3895 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3897 GOTO(class_detach, rc = -ENOMEM);
3899 rc = class_setup(obd, lcfg);
3901 GOTO(class_detach, rc);
3903 /* connect to MDD we just setup */
3904 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3906 GOTO(class_detach, rc);
3908 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3910 LASSERT(mdt_lu_site(mdt) == NULL);
3911 mdt->mdt_lu_dev.ld_site = site;
3912 site->ls_top_dev = &mdt->mdt_lu_dev;
3913 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3915 /* now connect to bottom OSD */
3916 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3917 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3919 GOTO(class_detach, rc);
3921 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3923 rc = lu_env_refill((struct lu_env *)env);
3925 CERROR("Failure to refill session: '%d'\n", rc);
3927 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3932 class_detach(obd, lcfg);
3934 lustre_cfg_free(lcfg);
3939 OBD_FREE(name, name_size);
3941 OBD_FREE(uuid, uuid_size);
3945 /* setup quota master target on MDT0 */
3946 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3947 struct lustre_cfg *cfg)
3949 struct obd_device *obd;
3950 char *dev = lustre_cfg_string(cfg, 0);
3951 char *qmtname, *uuid, *p;
3952 struct lustre_cfg_bufs *bufs;
3953 struct lustre_cfg *lcfg;
3954 struct lustre_profile *lprof;
3955 struct obd_connect_data *data;
3959 LASSERT(mdt->mdt_qmt_exp == NULL);
3960 LASSERT(mdt->mdt_qmt_dev == NULL);
3962 /* quota master is on MDT0 only for now */
3963 if (mdt->mdt_seq_site.ss_node_id != 0)
3966 /* MGS generates config commands which look as follows:
3967 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3968 * 3:lustre-MDT0000-mdtlov 4:f
3970 * We generate the QMT name from the MDT one, just replacing MD with QM
3971 * after all the preparations, the logical equivalent will be:
3972 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
3973 * 3:lustre-MDT0000-osd 4:f */
3974 OBD_ALLOC(qmtname, MAX_OBD_NAME);
3975 OBD_ALLOC(uuid, UUID_MAX);
3976 OBD_ALLOC_PTR(bufs);
3977 OBD_ALLOC_PTR(data);
3978 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
3979 GOTO(cleanup_mem, rc = -ENOMEM);
3981 strcpy(qmtname, dev);
3982 p = strstr(qmtname, "-MDT");
3984 GOTO(cleanup_mem, rc = -ENOMEM);
3985 /* replace MD with QM */
3989 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
3991 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3992 if (lprof == NULL || lprof->lp_dt == NULL) {
3993 CERROR("can't find profile for %s\n",
3994 lustre_cfg_string(cfg, 0));
3995 GOTO(cleanup_mem, rc = -EINVAL);
3998 lustre_cfg_bufs_reset(bufs, qmtname);
3999 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4000 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4001 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4003 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4005 GOTO(cleanup_mem, rc = -ENOMEM);
4007 rc = class_attach(lcfg);
4009 GOTO(lcfg_cleanup, rc);
4011 obd = class_name2obd(qmtname);
4013 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4014 lustre_cfg_string(cfg, 0));
4015 GOTO(lcfg_cleanup, rc = -EINVAL);
4018 lustre_cfg_free(lcfg);
4020 lustre_cfg_bufs_reset(bufs, qmtname);
4021 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4022 lustre_cfg_bufs_set_string(bufs, 2, dev);
4024 /* for quota, the next device should be the OSD device */
4025 lustre_cfg_bufs_set_string(bufs, 3,
4026 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4028 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4030 GOTO(class_detach, rc = -ENOMEM);
4032 rc = class_setup(obd, lcfg);
4034 GOTO(class_detach, rc);
4036 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4038 /* configure local quota objects */
4039 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4043 GOTO(class_cleanup, rc);
4045 /* connect to quota master target */
4046 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4047 data->ocd_version = LUSTRE_VERSION_CODE;
4048 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4051 CERROR("cannot connect to quota master device %s (%d)\n",
4053 GOTO(class_cleanup, rc);
4059 class_manual_cleanup(obd);
4060 mdt->mdt_qmt_dev = NULL;
4064 class_detach(obd, lcfg);
4066 lustre_cfg_free(lcfg);
4071 OBD_FREE(qmtname, MAX_OBD_NAME);
4073 OBD_FREE(uuid, UUID_MAX);
4079 /* Shutdown quota master target associated with mdt */
4080 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4084 if (mdt->mdt_qmt_exp == NULL)
4086 LASSERT(mdt->mdt_qmt_dev != NULL);
4088 /* the qmt automatically shuts down when the mdt disconnects */
4089 obd_disconnect(mdt->mdt_qmt_exp);
4090 mdt->mdt_qmt_exp = NULL;
4091 mdt->mdt_qmt_dev = NULL;
4095 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4096 * for now. This will be removed along with converting rest of MDT code
4097 * to use tgt_session_info */
4098 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4100 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4103 rc = mdt_getxattr(info);
4105 mdt_thread_info_fini(info);
4109 static struct tgt_handler mdt_tgt_handlers[] = {
4110 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4111 0, MDS_CONNECT, mdt_tgt_connect,
4112 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4113 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4114 0, MDS_DISCONNECT, tgt_disconnect,
4115 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4116 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4117 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4118 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4119 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4120 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4121 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4122 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4124 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4125 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4126 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4127 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4128 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4129 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4130 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4131 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4133 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4134 mdt_hsm_ct_register),
4135 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4136 mdt_hsm_ct_unregister),
4137 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4139 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4141 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4142 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4144 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4149 static struct tgt_handler mdt_sec_ctx_ops[] = {
4150 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4151 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4152 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4155 static struct tgt_handler mdt_quota_ops[] = {
4156 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4159 static struct tgt_opc_slice mdt_common_slice[] = {
4161 .tos_opc_start = MDS_FIRST_OPC,
4162 .tos_opc_end = MDS_LAST_OPC,
4163 .tos_hs = mdt_tgt_handlers
4166 .tos_opc_start = OBD_FIRST_OPC,
4167 .tos_opc_end = OBD_LAST_OPC,
4168 .tos_hs = tgt_obd_handlers
4171 .tos_opc_start = LDLM_FIRST_OPC,
4172 .tos_opc_end = LDLM_LAST_OPC,
4173 .tos_hs = tgt_dlm_handlers
4176 .tos_opc_start = SEC_FIRST_OPC,
4177 .tos_opc_end = SEC_LAST_OPC,
4178 .tos_hs = mdt_sec_ctx_ops
4181 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4182 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4183 .tos_hs = tgt_out_handlers
4186 .tos_opc_start = FLD_FIRST_OPC,
4187 .tos_opc_end = FLD_LAST_OPC,
4188 .tos_hs = fld_handlers
4191 .tos_opc_start = SEQ_FIRST_OPC,
4192 .tos_opc_end = SEQ_LAST_OPC,
4193 .tos_hs = seq_handlers
4196 .tos_opc_start = QUOTA_DQACQ,
4197 .tos_opc_end = QUOTA_LAST_OPC,
4198 .tos_hs = mdt_quota_ops
4201 .tos_opc_start = LLOG_FIRST_OPC,
4202 .tos_opc_end = LLOG_LAST_OPC,
4203 .tos_hs = tgt_llog_handlers
4206 .tos_opc_start = LFSCK_FIRST_OPC,
4207 .tos_opc_end = LFSCK_LAST_OPC,
4208 .tos_hs = tgt_lfsck_handlers
4216 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4218 struct md_device *next = m->mdt_child;
4219 struct lu_device *d = &m->mdt_lu_dev;
4220 struct obd_device *obd = mdt2obd_dev(m);
4221 struct lfsck_stop stop;
4224 stop.ls_status = LS_PAUSED;
4226 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4228 target_recovery_fini(obd);
4229 ping_evictor_stop();
4230 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4232 if (m->mdt_opts.mo_coordinator)
4233 mdt_hsm_cdt_stop(m);
4235 mdt_hsm_cdt_fini(m);
4237 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4238 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4239 obd_exports_barrier(obd);
4240 obd_zombie_barrier();
4244 tgt_fini(env, &m->mdt_lut);
4245 mdt_fs_cleanup(env, m);
4246 upcall_cache_cleanup(m->mdt_identity_cache);
4247 m->mdt_identity_cache = NULL;
4249 if (m->mdt_namespace != NULL) {
4250 ldlm_namespace_free(m->mdt_namespace, NULL,
4251 d->ld_obd->obd_force);
4252 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4255 mdt_quota_fini(env, m);
4257 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4259 mdt_seq_fini(env, m);
4260 mdt_fld_fini(env, m);
4265 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4267 LASSERT(atomic_read(&d->ld_ref) == 0);
4269 server_put_mount(mdt_obd_name(m), true);
4274 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4276 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4277 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4279 struct mdt_thread_info *info;
4280 struct obd_device *obd;
4281 const char *dev = lustre_cfg_string(cfg, 0);
4282 const char *num = lustre_cfg_string(cfg, 2);
4283 struct lustre_mount_info *lmi = NULL;
4284 struct lustre_sb_info *lsi;
4286 struct seq_server_site *ss_site;
4287 const char *identity_upcall = "NONE";
4288 struct md_device *next;
4294 lu_device_init(&m->mdt_lu_dev, ldt);
4296 * Environment (env) might be missing mdt_thread_key values at that
4297 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4300 * Usually device allocation path doesn't use module key values, but
4301 * mdt has to do a lot of work here, so allocate key value.
4303 rc = lu_env_refill((struct lu_env *)env);
4307 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4308 LASSERT(info != NULL);
4310 obd = class_name2obd(dev);
4311 LASSERT(obd != NULL);
4313 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4315 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4317 /* default is coordinator off, it is started through conf_param
4319 m->mdt_opts.mo_coordinator = 0;
4321 lmi = server_get_mount(dev);
4323 CERROR("Cannot get mount info for %s!\n", dev);
4326 lsi = s2lsi(lmi->lmi_sb);
4327 /* CMD is supported only in IAM mode */
4329 node_id = simple_strtol(num, NULL, 10);
4330 obd->u.obt.obt_magic = OBT_MAGIC;
4333 m->mdt_squash.rsi_uid = 0;
4334 m->mdt_squash.rsi_gid = 0;
4335 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4336 init_rwsem(&m->mdt_squash.rsi_sem);
4337 spin_lock_init(&m->mdt_osfs_lock);
4338 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4339 m->mdt_enable_remote_dir = 0;
4340 m->mdt_enable_remote_dir_gid = 0;
4342 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4343 m->mdt_lu_dev.ld_obd = obd;
4344 /* Set this lu_device to obd for error handling purposes. */
4345 obd->obd_lu_dev = &m->mdt_lu_dev;
4347 /* init the stack */
4348 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4350 CERROR("%s: Can't init device stack, rc %d\n",
4351 mdt_obd_name(m), rc);
4356 ss_site = mdt_seq_site(m);
4357 s->ld_seq_site = ss_site;
4360 /* set server index */
4361 ss_site->ss_node_id = node_id;
4363 /* failover is the default
4364 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4365 * assumed whose ss_node_id == 0 XXX
4367 obd->obd_replayable = 1;
4368 /* No connection accepted until configurations will finish */
4369 obd->obd_no_conn = 1;
4371 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4372 char *str = lustre_cfg_string(cfg, 4);
4373 if (strchr(str, 'n')) {
4374 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4375 obd->obd_replayable = 0;
4379 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4381 GOTO(err_fini_stack, rc);
4383 rc = mdt_seq_init(env, m);
4385 GOTO(err_fini_fld, rc);
4387 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4388 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4389 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4390 LDLM_NAMESPACE_SERVER,
4391 LDLM_NAMESPACE_GREEDY,
4393 if (m->mdt_namespace == NULL)
4394 GOTO(err_fini_seq, rc = -ENOMEM);
4396 m->mdt_namespace->ns_lvbp = m;
4397 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4399 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4400 /* set obd_namespace for compatibility with old code */
4401 obd->obd_namespace = m->mdt_namespace;
4403 rc = mdt_hsm_cdt_init(m);
4405 CERROR("%s: error initializing coordinator, rc %d\n",
4406 mdt_obd_name(m), rc);
4407 GOTO(err_free_ns, rc);
4410 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4411 OBD_FAIL_MDS_ALL_REQUEST_NET,
4412 OBD_FAIL_MDS_ALL_REPLY_NET);
4414 GOTO(err_free_hsm, rc);
4416 rc = mdt_fs_setup(env, m, obd, lsi);
4420 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4422 next = m->mdt_child;
4423 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4426 GOTO(err_fs_cleanup, rc);
4428 if (mntopts & MNTOPT_USERXATTR)
4429 m->mdt_opts.mo_user_xattr = 1;
4431 m->mdt_opts.mo_user_xattr = 0;
4433 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4435 GOTO(err_fs_cleanup, rc);
4437 if (mntopts & MNTOPT_ACL)
4438 m->mdt_opts.mo_acl = 1;
4440 m->mdt_opts.mo_acl = 0;
4442 /* XXX: to support suppgid for ACL, we enable identity_upcall
4443 * by default, otherwise, maybe got unexpected -EACCESS. */
4444 if (m->mdt_opts.mo_acl)
4445 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4447 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4449 &mdt_identity_upcall_cache_ops);
4450 if (IS_ERR(m->mdt_identity_cache)) {
4451 rc = PTR_ERR(m->mdt_identity_cache);
4452 m->mdt_identity_cache = NULL;
4453 GOTO(err_fs_cleanup, rc);
4456 rc = mdt_procfs_init(m, dev);
4458 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4459 GOTO(err_recovery, rc);
4462 rc = mdt_quota_init(env, m, cfg);
4464 GOTO(err_procfs, rc);
4466 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4467 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4468 "mdt_ldlm_client", m->mdt_ldlm_client);
4470 ping_evictor_start();
4472 /* recovery will be started upon mdt_prepare()
4473 * when the whole stack is complete and ready
4474 * to serve the requests */
4476 /* Reduce the initial timeout on an MDS because it doesn't need such
4477 * a long timeout as an OST does. Adaptive timeouts will adjust this
4478 * value appropriately. */
4479 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4480 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4486 target_recovery_fini(obd);
4487 upcall_cache_cleanup(m->mdt_identity_cache);
4488 m->mdt_identity_cache = NULL;
4490 mdt_fs_cleanup(env, m);
4492 tgt_fini(env, &m->mdt_lut);
4494 mdt_hsm_cdt_fini(m);
4496 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4497 obd->obd_namespace = m->mdt_namespace = NULL;
4499 mdt_seq_fini(env, m);
4501 mdt_fld_fini(env, m);
4503 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4506 server_put_mount(dev, true);
4510 /* For interoperability, the left element is old parameter, the right one
4511 * is the new version of the parameter, if some parameter is deprecated,
4512 * the new version should be set as NULL. */
4513 static struct cfg_interop_param mdt_interop_param[] = {
4514 { "mdt.group_upcall", NULL },
4515 { "mdt.quota_type", NULL },
4516 { "mdd.quota_type", NULL },
4517 { "mdt.rootsquash", "mdt.root_squash" },
4518 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4522 /* used by MGS to process specific configurations */
4523 static int mdt_process_config(const struct lu_env *env,
4524 struct lu_device *d, struct lustre_cfg *cfg)
4526 struct mdt_device *m = mdt_dev(d);
4527 struct md_device *md_next = m->mdt_child;
4528 struct lu_device *next = md2lu_dev(md_next);
4532 switch (cfg->lcfg_command) {
4534 struct obd_device *obd = d->ld_obd;
4536 /* For interoperability */
4537 struct cfg_interop_param *ptr = NULL;
4538 struct lustre_cfg *old_cfg = NULL;
4541 param = lustre_cfg_string(cfg, 1);
4542 if (param == NULL) {
4543 CERROR("param is empty\n");
4548 ptr = class_find_old_param(param, mdt_interop_param);
4550 if (ptr->new_param == NULL) {
4552 CWARN("For interoperability, skip this %s."
4553 " It is obsolete.\n", ptr->old_param);
4557 CWARN("Found old param %s, changed it to %s.\n",
4558 ptr->old_param, ptr->new_param);
4561 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4568 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4570 if (rc > 0 || rc == -ENOSYS) {
4571 /* is it an HSM var ? */
4572 rc = class_process_proc_param(PARAM_HSM,
4573 hsm_cdt_get_proc_vars(),
4575 if (rc > 0 || rc == -ENOSYS)
4576 /* we don't understand; pass it on */
4577 rc = next->ld_ops->ldo_process_config(env, next,
4581 if (old_cfg != NULL)
4582 lustre_cfg_free(cfg);
4587 /* others are passed further */
4588 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4594 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4595 const struct lu_object_header *hdr,
4596 struct lu_device *d)
4598 struct mdt_object *mo;
4602 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4604 struct lu_object *o;
4605 struct lu_object_header *h;
4608 h = &mo->mot_header;
4609 lu_object_header_init(h);
4610 lu_object_init(o, h, d);
4611 lu_object_add_top(h, o);
4612 o->lo_ops = &mdt_obj_ops;
4613 spin_lock_init(&mo->mot_write_lock);
4614 mutex_init(&mo->mot_lov_mutex);
4615 init_rwsem(&mo->mot_open_sem);
4621 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4622 const struct lu_object_conf *unused)
4624 struct mdt_device *d = mdt_dev(o->lo_dev);
4625 struct lu_device *under;
4626 struct lu_object *below;
4630 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4631 PFID(lu_object_fid(o)));
4633 under = &d->mdt_child->md_lu_dev;
4634 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4635 if (below != NULL) {
4636 lu_object_add(o, below);
4643 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4645 struct mdt_object *mo = mdt_obj(o);
4646 struct lu_object_header *h;
4650 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4651 PFID(lu_object_fid(o)));
4653 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4654 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4657 lu_object_header_fini(h);
4658 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4663 static int mdt_object_print(const struct lu_env *env, void *cookie,
4664 lu_printer_t p, const struct lu_object *o)
4666 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4668 return (*p)(env, cookie,
4669 LUSTRE_MDT_NAME"-object@%p(flags=%d, writecount=%d)",
4670 mdto, mdto->mot_flags, mdto->mot_write_count);
4673 static int mdt_prepare(const struct lu_env *env,
4674 struct lu_device *pdev,
4675 struct lu_device *cdev)
4677 struct mdt_device *mdt = mdt_dev(cdev);
4678 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4679 struct obd_device *obd = cdev->ld_obd;
4686 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4690 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4694 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4698 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4699 /* The LFSCK instance is registered just now, so it must be there when
4700 * register the namespace to such instance. */
4701 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4703 if (mdt->mdt_seq_site.ss_node_id == 0) {
4704 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4705 &mdt->mdt_md_root_fid);
4710 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4712 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4713 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4714 LASSERT(obd->obd_no_conn);
4715 spin_lock(&obd->obd_dev_lock);
4716 obd->obd_no_conn = 0;
4717 spin_unlock(&obd->obd_dev_lock);
4719 if (obd->obd_recovering == 0)
4720 mdt_postrecov(env, mdt);
4725 const struct lu_device_operations mdt_lu_ops = {
4726 .ldo_object_alloc = mdt_object_alloc,
4727 .ldo_process_config = mdt_process_config,
4728 .ldo_prepare = mdt_prepare,
4731 static const struct lu_object_operations mdt_obj_ops = {
4732 .loo_object_init = mdt_object_init,
4733 .loo_object_free = mdt_object_free,
4734 .loo_object_print = mdt_object_print
4737 static int mdt_obd_set_info_async(const struct lu_env *env,
4738 struct obd_export *exp,
4739 __u32 keylen, void *key,
4740 __u32 vallen, void *val,
4741 struct ptlrpc_request_set *set)
4747 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4748 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4756 * Match client and server connection feature flags.
4758 * Compute the compatibility flags for a connection request based on
4759 * features mutually supported by client and server.
4761 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4762 * must not be updated here, otherwise a partially initialized value may
4763 * be exposed. After the connection request is successfully processed,
4764 * the top-level MDT connect request handler atomically updates the export
4765 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4766 * reply. \see mdt_connect().
4768 * \param exp the obd_export associated with this client/target pair
4769 * \param mdt the target device for the connection
4770 * \param data stores data for this connect request
4773 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4774 * \retval -EBADE client and server feature requirements are incompatible
4776 static int mdt_connect_internal(struct obd_export *exp,
4777 struct mdt_device *mdt,
4778 struct obd_connect_data *data)
4780 LASSERT(data != NULL);
4782 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4783 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4785 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
4786 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
4787 CWARN("%s: client %s does not support ibits lock, either "
4788 "very old or an invalid client: flags "LPX64"\n",
4789 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
4790 data->ocd_connect_flags);
4794 if (!mdt->mdt_opts.mo_acl)
4795 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4797 if (!mdt->mdt_opts.mo_user_xattr)
4798 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4800 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4801 data->ocd_brw_size = min(data->ocd_brw_size,
4802 (__u32)MD_MAX_BRW_SIZE);
4803 if (data->ocd_brw_size == 0) {
4804 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4805 " ocd_version: %x ocd_grant: %d "
4806 "ocd_index: %u ocd_brw_size is "
4807 "unexpectedly zero, network data "
4808 "corruption? Refusing connection of this"
4811 exp->exp_client_uuid.uuid,
4812 exp, data->ocd_connect_flags, data->ocd_version,
4813 data->ocd_grant, data->ocd_index);
4818 /* NB: Disregard the rule against updating
4819 * exp_connect_data.ocd_connect_flags in this case, since
4820 * tgt_client_new() needs to know if this is a lightweight
4821 * connection, and it is safe to expose this flag before
4822 * connection processing completes. */
4823 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4824 spin_lock(&exp->exp_lock);
4825 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4826 spin_unlock(&exp->exp_lock);
4829 data->ocd_version = LUSTRE_VERSION_CODE;
4831 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4832 CWARN("%s: MDS requires FID support, but client not\n",
4837 if (OCD_HAS_FLAG(data, PINGLESS)) {
4838 if (ptlrpc_pinger_suppress_pings()) {
4839 spin_lock(&exp->exp_obd->obd_dev_lock);
4840 list_del_init(&exp->exp_obd_chain_timed);
4841 spin_unlock(&exp->exp_obd->obd_dev_lock);
4843 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4847 data->ocd_max_easize = mdt->mdt_max_ea_size;
4852 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
4853 struct mdt_thread_info *info,
4854 struct mdt_file_data *mfd)
4856 struct lu_context ses;
4860 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
4865 lu_context_enter(&ses);
4867 mdt_ucred(info)->uc_valid = UCRED_OLD;
4868 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
4870 lu_context_exit(&ses);
4871 lu_context_fini(&ses);
4877 static int mdt_export_cleanup(struct obd_export *exp)
4879 struct list_head closing_list;
4880 struct mdt_export_data *med = &exp->exp_mdt_data;
4881 struct obd_device *obd = exp->exp_obd;
4882 struct mdt_device *mdt;
4883 struct mdt_thread_info *info;
4885 struct mdt_file_data *mfd, *n;
4889 INIT_LIST_HEAD(&closing_list);
4890 spin_lock(&med->med_open_lock);
4891 while (!list_empty(&med->med_open_head)) {
4892 struct list_head *tmp = med->med_open_head.next;
4893 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
4895 /* Remove mfd handle so it can't be found again.
4896 * We are consuming the mfd_list reference here. */
4897 class_handle_unhash(&mfd->mfd_handle);
4898 list_move_tail(&mfd->mfd_list, &closing_list);
4900 spin_unlock(&med->med_open_lock);
4901 mdt = mdt_dev(obd->obd_lu_dev);
4902 LASSERT(mdt != NULL);
4904 rc = lu_env_init(&env, LCT_MD_THREAD);
4908 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
4909 LASSERT(info != NULL);
4910 memset(info, 0, sizeof *info);
4911 info->mti_env = &env;
4912 info->mti_mdt = mdt;
4913 info->mti_exp = exp;
4915 if (!list_empty(&closing_list)) {
4916 struct md_attr *ma = &info->mti_attr;
4918 /* Close any open files (which may also cause orphan
4920 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
4921 list_del_init(&mfd->mfd_list);
4922 ma->ma_need = ma->ma_valid = 0;
4924 /* This file is being closed due to an eviction, it
4925 * could have been modified and now dirty regarding to
4926 * HSM archive, check this!
4927 * The logic here is to mark a file dirty if there's a
4928 * chance it was dirtied before the client was evicted,
4929 * so that we don't have to wait for a release attempt
4930 * before finding out the file was actually dirty and
4931 * fail the release. Aggressively marking it dirty here
4932 * will cause the policy engine to attempt to
4933 * re-archive it; when rearchiving, we can compare the
4934 * current version to the HSM data_version and make the
4935 * archive request into a noop if it's not actually
4938 if (mfd->mfd_mode & FMODE_WRITE)
4939 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
4941 /* Don't unlink orphan on failover umount, LU-184 */
4942 if (exp->exp_flags & OBD_OPT_FAILOVER) {
4943 ma->ma_valid = MA_FLAGS;
4944 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
4946 mdt_mfd_close(info, mfd);
4949 info->mti_mdt = NULL;
4950 /* cleanup client slot early */
4951 /* Do not erase record for recoverable client. */
4952 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
4953 tgt_client_del(&env, exp);
4959 static int mdt_obd_disconnect(struct obd_export *exp)
4965 class_export_get(exp);
4967 nodemap_del_member(exp);
4968 rc = server_disconnect_export(exp);
4970 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
4972 rc = mdt_export_cleanup(exp);
4973 class_export_put(exp);
4977 /* mds_connect copy */
4978 static int mdt_obd_connect(const struct lu_env *env,
4979 struct obd_export **exp, struct obd_device *obd,
4980 struct obd_uuid *cluuid,
4981 struct obd_connect_data *data,
4984 struct obd_export *lexp;
4985 struct lustre_handle conn = { 0 };
4986 struct mdt_device *mdt;
4988 lnet_nid_t *client_nid = localdata;
4991 LASSERT(env != NULL);
4992 if (!exp || !obd || !cluuid)
4995 mdt = mdt_dev(obd->obd_lu_dev);
4998 * first, check whether the stack is ready to handle requests
4999 * XXX: probably not very appropriate method is used now
5000 * at some point we should find a better one
5002 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
5003 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5004 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5005 rc = obd_get_info(env, mdt->mdt_child_exp,
5006 sizeof(KEY_OSP_CONNECTED),
5007 KEY_OSP_CONNECTED, NULL, NULL);
5010 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5013 rc = class_connect(&conn, obd, cluuid);
5017 lexp = class_conn2export(&conn);
5018 LASSERT(lexp != NULL);
5020 rc = mdt_connect_internal(lexp, mdt, data);
5022 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5025 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5026 rc = tgt_client_new(env, lexp);
5028 rc = nodemap_add_member(*client_nid, lexp);
5029 if (rc != 0 && rc != -EEXIST)
5032 mdt_export_stats_init(obd, lexp, localdata);
5037 class_disconnect(lexp);
5046 static int mdt_obd_reconnect(const struct lu_env *env,
5047 struct obd_export *exp, struct obd_device *obd,
5048 struct obd_uuid *cluuid,
5049 struct obd_connect_data *data,
5052 lnet_nid_t *client_nid = localdata;
5056 if (exp == NULL || obd == NULL || cluuid == NULL)
5059 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5061 rc = nodemap_add_member(*client_nid, exp);
5062 if (rc == 0 || rc == -EEXIST)
5063 mdt_export_stats_init(obd, exp, localdata);
5069 /* FIXME: Can we avoid using these two interfaces? */
5070 static int mdt_init_export(struct obd_export *exp)
5072 struct mdt_export_data *med = &exp->exp_mdt_data;
5076 INIT_LIST_HEAD(&med->med_open_head);
5077 spin_lock_init(&med->med_open_lock);
5078 mutex_init(&med->med_idmap_mutex);
5079 med->med_idmap = NULL;
5080 spin_lock(&exp->exp_lock);
5081 exp->exp_connecting = 1;
5082 spin_unlock(&exp->exp_lock);
5084 /* self-export doesn't need client data and ldlm initialization */
5085 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5086 &exp->exp_client_uuid)))
5089 rc = tgt_client_alloc(exp);
5093 rc = ldlm_init_export(exp);
5100 tgt_client_free(exp);
5102 CERROR("%s: Failed to initialize export: rc = %d\n",
5103 exp->exp_obd->obd_name, rc);
5107 static int mdt_destroy_export(struct obd_export *exp)
5111 if (exp_connect_rmtclient(exp))
5112 mdt_cleanup_idmap(&exp->exp_mdt_data);
5114 target_destroy_export(exp);
5115 /* destroy can be called from failed obd_setup, so
5116 * checking uuid is safer than obd_self_export */
5117 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5118 &exp->exp_client_uuid)))
5121 ldlm_destroy_export(exp);
5122 tgt_client_free(exp);
5124 LASSERT(list_empty(&exp->exp_outstanding_replies));
5125 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5130 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5131 struct linkea_data *ldata)
5135 LASSERT(ldata->ld_buf->lb_buf != NULL);
5137 if (!mdt_object_exists(mdt_obj))
5140 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5141 ldata->ld_buf, XATTR_NAME_LINK);
5142 if (rc == -ERANGE) {
5143 /* Buf was too small, figure out what we need. */
5144 lu_buf_free(ldata->ld_buf);
5145 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5146 ldata->ld_buf, XATTR_NAME_LINK);
5149 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5150 if (ldata->ld_buf->lb_buf == NULL)
5152 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5153 ldata->ld_buf, XATTR_NAME_LINK);
5158 return linkea_init(ldata);
5162 * Given an MDT object, try to look up the full path to the object.
5163 * Part of the MDT layer implementation of lfs fid2path.
5165 * \param[in] info Per-thread common data shared by MDT level handlers.
5166 * \param[in] obj Object to do path lookup of
5167 * \param[in,out] fp User-provided struct to store path information
5169 * \retval 0 Lookup successful, path information stored in fp
5170 * \retval -EAGAIN Lookup failed, usually because object is being moved
5171 * \retval negative errno if there was a problem
5173 static int mdt_path_current(struct mdt_thread_info *info,
5174 struct mdt_object *obj,
5175 struct getinfo_fid2path *fp)
5177 struct mdt_device *mdt = info->mti_mdt;
5178 struct mdt_object *mdt_obj;
5179 struct link_ea_header *leh;
5180 struct link_ea_entry *lee;
5181 struct lu_name *tmpname = &info->mti_name;
5182 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5183 struct lu_buf *buf = &info->mti_big_buf;
5186 struct linkea_data ldata = { NULL };
5191 /* temp buffer for path element, the buffer will be finally freed
5192 * in mdt_thread_info_fini */
5193 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5194 if (buf->lb_buf == NULL)
5198 ptr = fp->gf_path + fp->gf_pathlen - 1;
5201 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5203 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5204 * so checking root_fid can only happen on MDT0. */
5205 while (!lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid)) {
5206 struct lu_buf lmv_buf;
5208 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5209 if (IS_ERR(mdt_obj))
5210 GOTO(out, rc = PTR_ERR(mdt_obj));
5212 if (!mdt_object_exists(mdt_obj)) {
5213 mdt_object_put(info->mti_env, mdt_obj);
5214 GOTO(out, rc = -ENOENT);
5217 if (mdt_object_remote(mdt_obj)) {
5218 mdt_object_put(info->mti_env, mdt_obj);
5219 GOTO(remote_out, rc = -EREMOTE);
5222 rc = mdt_links_read(info, mdt_obj, &ldata);
5224 mdt_object_put(info->mti_env, mdt_obj);
5229 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5230 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5231 /* If set, use link #linkno for path lookup, otherwise use
5232 link #0. Only do this for the final path element. */
5233 if (first && fp->gf_linkno < leh->leh_reccount) {
5235 for (count = 0; count < fp->gf_linkno; count++) {
5236 lee = (struct link_ea_entry *)
5237 ((char *)lee + reclen);
5238 linkea_entry_unpack(lee, &reclen, tmpname,
5241 if (fp->gf_linkno < leh->leh_reccount - 1)
5242 /* indicate to user there are more links */
5246 lmv_buf.lb_buf = info->mti_xattr_buf;
5247 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5248 /* Check if it is slave stripes */
5249 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5250 &lmv_buf, XATTR_NAME_LMV);
5251 mdt_object_put(info->mti_env, mdt_obj);
5253 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5255 /* For slave stripes, get its master */
5256 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5257 fp->gf_fid = *tmpfid;
5260 } else if (rc < 0 && rc != -ENODATA) {
5266 /* Pack the name in the end of the buffer */
5267 ptr -= tmpname->ln_namelen;
5268 if (ptr - 1 <= fp->gf_path)
5269 GOTO(out, rc = -EOVERFLOW);
5270 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5273 /* keep the last resolved fid to the client, so the
5274 * client will build the left path on another MDT for
5276 fp->gf_fid = *tmpfid;
5282 ptr++; /* skip leading / */
5283 memmove(fp->gf_path, ptr, fp->gf_path + fp->gf_pathlen - ptr);
5290 * Given an MDT object, use mdt_path_current to get the path.
5291 * Essentially a wrapper to retry mdt_path_current a set number of times
5292 * if -EAGAIN is returned (usually because an object is being moved).
5294 * Part of the MDT layer implementation of lfs fid2path.
5296 * \param[in] info Per-thread common data shared by mdt level handlers.
5297 * \param[in] obj Object to do path lookup of
5298 * \param[in,out] fp User-provided struct for arguments and to store path
5301 * \retval 0 Lookup successful, path information stored in fp
5302 * \retval negative errno if there was a problem
5304 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5305 struct getinfo_fid2path *fp)
5307 struct mdt_device *mdt = info->mti_mdt;
5312 if (fp->gf_pathlen < 3)
5315 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5316 fp->gf_path[0] = '\0';
5320 /* Retry multiple times in case file is being moved */
5321 while (tries-- && rc == -EAGAIN)
5322 rc = mdt_path_current(info, obj, fp);
5328 * Get the full path of the provided FID, as of changelog record recno.
5330 * This checks sanity and looks up object for user provided FID
5331 * before calling the actual path lookup code.
5333 * Part of the MDT layer implementation of lfs fid2path.
5335 * \param[in] info Per-thread common data shared by mdt level handlers.
5336 * \param[in,out] fp User-provided struct for arguments and to store path
5339 * \retval 0 Lookup successful, path information and recno stored in fp
5340 * \retval -ENOENT, object does not exist
5341 * \retval negative errno if there was a problem
5343 static int mdt_fid2path(struct mdt_thread_info *info,
5344 struct getinfo_fid2path *fp)
5346 struct mdt_device *mdt = info->mti_mdt;
5347 struct mdt_object *obj;
5351 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5352 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5354 if (!fid_is_sane(&fp->gf_fid))
5357 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5358 CWARN("%s: "DFID" is invalid, sequence should be "
5359 ">= "LPX64"\n", mdt_obd_name(mdt),
5360 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5364 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5367 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5368 PFID(&fp->gf_fid), rc);
5372 if (mdt_object_remote(obj))
5374 else if (!mdt_object_exists(obj))
5380 mdt_object_put(info->mti_env, obj);
5381 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5382 PFID(&fp->gf_fid), rc);
5386 rc = mdt_path(info, obj, fp);
5388 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5389 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5391 mdt_object_put(info->mti_env, obj);
5396 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5397 void *val, int vallen)
5399 struct getinfo_fid2path *fpout, *fpin;
5402 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5405 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5406 lustre_swab_fid2path(fpin);
5408 memcpy(fpout, fpin, sizeof(*fpin));
5409 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5412 rc = mdt_fid2path(info, fpout);
5416 int mdt_get_info(struct tgt_session_info *tsi)
5426 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5428 CDEBUG(D_IOCTL, "No GETINFO key\n");
5429 RETURN(err_serious(-EFAULT));
5431 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5434 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5435 if (vallen == NULL) {
5436 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5437 tgt_name(tsi->tsi_tgt));
5438 RETURN(err_serious(-EFAULT));
5441 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5443 rc = req_capsule_server_pack(tsi->tsi_pill);
5445 RETURN(err_serious(rc));
5447 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5448 if (valout == NULL) {
5449 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5450 tgt_name(tsi->tsi_tgt));
5451 RETURN(err_serious(-EFAULT));
5454 if (KEY_IS(KEY_FID2PATH)) {
5455 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5457 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5458 mdt_thread_info_fini(info);
5465 /* Pass the ioc down */
5466 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5467 unsigned int cmd, int len, void *data)
5469 struct lu_context ioctl_session;
5470 struct md_device *next = mdt->mdt_child;
5474 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5477 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5478 lu_context_enter(&ioctl_session);
5479 env->le_ses = &ioctl_session;
5481 LASSERT(next->md_ops->mdo_iocontrol);
5482 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5484 lu_context_exit(&ioctl_session);
5485 lu_context_fini(&ioctl_session);
5489 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5491 struct obd_ioctl_data *data = karg;
5494 struct mdt_object *obj;
5495 struct mdt_lock_handle *lh;
5499 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5500 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5503 fid = (struct lu_fid *)data->ioc_inlbuf1;
5505 if (!fid_is_sane(fid))
5508 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5510 lh = &mti->mti_lh[MDT_LH_PARENT];
5511 mdt_lock_reg_init(lh, LCK_CR);
5513 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5515 RETURN(PTR_ERR(obj));
5517 if (mdt_object_remote(obj)) {
5520 * before calling version get the correct MDS should be
5521 * fid, this is error to find remote object here
5523 CERROR("nonlocal object "DFID"\n", PFID(fid));
5524 } else if (!mdt_object_exists(obj)) {
5525 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5528 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5529 *(__u64 *)data->ioc_inlbuf2 = version;
5532 mdt_object_unlock_put(mti, obj, lh, 1);
5536 /* ioctls on obd dev */
5537 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5538 void *karg, void *uarg)
5541 struct obd_device *obd = exp->exp_obd;
5542 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5543 struct dt_device *dt = mdt->mdt_bottom;
5547 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5548 rc = lu_env_init(&env, LCT_MD_THREAD);
5554 rc = mdt_device_sync(&env, mdt);
5556 case OBD_IOC_SET_READONLY:
5557 rc = dt->dd_ops->dt_ro(&env, dt);
5559 case OBD_IOC_ABORT_RECOVERY:
5560 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5561 obd->obd_force_abort_recovery = 1;
5562 target_stop_recovery_thread(obd);
5565 case OBD_IOC_CHANGELOG_REG:
5566 case OBD_IOC_CHANGELOG_DEREG:
5567 case OBD_IOC_CHANGELOG_CLEAR:
5568 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5570 case OBD_IOC_START_LFSCK: {
5571 struct md_device *next = mdt->mdt_child;
5572 struct obd_ioctl_data *data = karg;
5573 struct lfsck_start_param lsp;
5575 if (unlikely(data == NULL)) {
5580 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5581 lsp.lsp_index_valid = 0;
5582 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5585 case OBD_IOC_STOP_LFSCK: {
5586 struct md_device *next = mdt->mdt_child;
5587 struct obd_ioctl_data *data = karg;
5588 struct lfsck_stop stop;
5590 stop.ls_status = LS_STOPPED;
5591 /* Old lfsck utils may pass NULL @stop. */
5592 if (data->ioc_inlbuf1 == NULL)
5596 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5598 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5601 case OBD_IOC_GET_OBJ_VERSION: {
5602 struct mdt_thread_info *mti;
5603 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5604 memset(mti, 0, sizeof *mti);
5605 mti->mti_env = &env;
5609 rc = mdt_ioc_version_get(mti, karg);
5612 case OBD_IOC_CATLOGLIST: {
5613 struct mdt_thread_info *mti;
5615 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5616 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5617 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5618 &mti->mti_tmp_fid1);
5623 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5624 mdt_obd_name(mdt), cmd, rc);
5631 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5633 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5634 struct lfsck_start_param lsp;
5638 lsp.lsp_start = NULL;
5639 lsp.lsp_index_valid = 0;
5640 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
5641 OBD_IOC_START_LFSCK,
5643 if (rc != 0 && rc != -EALREADY)
5644 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
5645 mdt_obd_name(mdt), rc);
5647 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5651 static int mdt_obd_postrecov(struct obd_device *obd)
5656 rc = lu_env_init(&env, LCT_MD_THREAD);
5659 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5664 static struct obd_ops mdt_obd_device_ops = {
5665 .o_owner = THIS_MODULE,
5666 .o_set_info_async = mdt_obd_set_info_async,
5667 .o_connect = mdt_obd_connect,
5668 .o_reconnect = mdt_obd_reconnect,
5669 .o_disconnect = mdt_obd_disconnect,
5670 .o_init_export = mdt_init_export,
5671 .o_destroy_export = mdt_destroy_export,
5672 .o_iocontrol = mdt_iocontrol,
5673 .o_postrecov = mdt_obd_postrecov,
5676 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5677 struct lu_device *d)
5679 struct mdt_device *m = mdt_dev(d);
5686 static struct lu_device *mdt_device_free(const struct lu_env *env,
5687 struct lu_device *d)
5689 struct mdt_device *m = mdt_dev(d);
5692 lu_device_fini(&m->mdt_lu_dev);
5698 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5699 struct lu_device_type *t,
5700 struct lustre_cfg *cfg)
5702 struct lu_device *l;
5703 struct mdt_device *m;
5710 rc = mdt_init0(env, m, t, cfg);
5712 mdt_device_free(env, l);
5717 l = ERR_PTR(-ENOMEM);
5721 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5722 LU_KEY_INIT(mdt, struct mdt_thread_info);
5724 static void mdt_key_fini(const struct lu_context *ctx,
5725 struct lu_context_key *key, void* data)
5727 struct mdt_thread_info *info = data;
5729 if (info->mti_big_lmm) {
5730 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5731 info->mti_big_lmm = NULL;
5732 info->mti_big_lmmsize = 0;
5737 /* context key: mdt_thread_key */
5738 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5740 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5742 return lu_ucred(info->mti_env);
5745 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5747 return lu_ucred_check(info->mti_env);
5751 * Enable/disable COS (Commit On Sharing).
5753 * Set/Clear the COS flag in mdt options.
5755 * \param mdt mdt device
5756 * \param val 0 disables COS, other values enable COS
5758 void mdt_enable_cos(struct mdt_device *mdt, int val)
5763 mdt->mdt_opts.mo_cos = !!val;
5764 rc = lu_env_init(&env, LCT_LOCAL);
5765 if (unlikely(rc != 0)) {
5766 CWARN("%s: lu_env initialization failed, cannot "
5767 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
5770 mdt_device_sync(&env, mdt);
5775 * Check COS (Commit On Sharing) status.
5777 * Return COS flag status.
5779 * \param mdt mdt device
5781 int mdt_cos_is_enabled(struct mdt_device *mdt)
5783 return mdt->mdt_opts.mo_cos != 0;
5786 static struct lu_device_type_operations mdt_device_type_ops = {
5787 .ldto_device_alloc = mdt_device_alloc,
5788 .ldto_device_free = mdt_device_free,
5789 .ldto_device_fini = mdt_device_fini
5792 static struct lu_device_type mdt_device_type = {
5793 .ldt_tags = LU_DEVICE_MD,
5794 .ldt_name = LUSTRE_MDT_NAME,
5795 .ldt_ops = &mdt_device_type_ops,
5796 .ldt_ctx_tags = LCT_MD_THREAD
5799 static int __init mdt_mod_init(void)
5803 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5804 FID_NOBRACE_LEN + 1);
5805 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5807 rc = lu_kmem_init(mdt_caches);
5811 rc = mds_mod_init();
5815 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
5816 LUSTRE_MDT_NAME, &mdt_device_type);
5821 lu_kmem_fini(mdt_caches);
5828 static void __exit mdt_mod_exit(void)
5830 class_unregister_type(LUSTRE_MDT_NAME);
5832 lu_kmem_fini(mdt_caches);
5835 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
5836 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5837 MODULE_VERSION(LUSTRE_VERSION_STRING);
5838 MODULE_LICENSE("GPL");
5840 module_init(mdt_mod_init);
5841 module_exit(mdt_mod_exit);