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_cookiesize = 0;
1004 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1005 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1006 CDEBUG(D_INODE, "I am going to change the MAX_MD_SIZE & "
1007 "MAX_COOKIE to : %d:%d\n", repbody->mbo_max_mdsize,
1008 repbody->mbo_max_cookiesize);
1011 if (exp_connect_rmtclient(info->mti_exp) &&
1012 reqbody->mbo_valid & OBD_MD_FLRMTPERM) {
1013 void *buf = req_capsule_server_get(pill, &RMF_ACL);
1015 /* mdt_getattr_lock only */
1016 rc = mdt_pack_remote_perm(info, o, buf);
1018 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
1019 repbody->mbo_aclsize = 0;
1022 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
1023 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
1026 #ifdef CONFIG_FS_POSIX_ACL
1027 else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1028 (reqbody->mbo_valid & OBD_MD_FLACL))
1029 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1034 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1039 static int mdt_getattr(struct tgt_session_info *tsi)
1041 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1042 struct mdt_object *obj = info->mti_object;
1043 struct req_capsule *pill = info->mti_pill;
1044 struct mdt_body *reqbody;
1045 struct mdt_body *repbody;
1049 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1051 LASSERT(obj != NULL);
1052 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1054 /* Unlike intent case where we need to pre-fill out buffers early on
1055 * in intent policy for ldlm reasons, here we can have a much better
1056 * guess at EA size by just reading it from disk.
1057 * Exceptions are readdir and (missing) directory striping */
1059 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1060 /* No easy way to know how long is the symlink, but it cannot
1061 * be more than PATH_MAX, so we allocate +1 */
1064 /* A special case for fs ROOT: getattr there might fetch
1065 * default EA for entire fs, not just for this dir!
1067 } else if (lu_fid_eq(mdt_object_fid(obj),
1068 &info->mti_mdt->mdt_md_root_fid) &&
1069 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1070 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1072 /* Should the default strping be bigger, mdt_fix_reply
1073 * will reallocate */
1074 rc = DEF_REP_MD_SIZE;
1076 /* Read the actual EA size from disk */
1077 rc = mdt_attr_get_eabuf_size(info, obj);
1081 GOTO(out, rc = err_serious(rc));
1083 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1085 rc = req_capsule_server_pack(pill);
1086 if (unlikely(rc != 0))
1087 GOTO(out, rc = err_serious(rc));
1089 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1090 LASSERT(repbody != NULL);
1091 repbody->mbo_eadatasize = 0;
1092 repbody->mbo_aclsize = 0;
1094 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1095 rc = mdt_init_ucred(info, reqbody);
1097 rc = mdt_check_ucred(info);
1099 GOTO(out_shrink, rc);
1101 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1103 rc = mdt_getattr_internal(info, obj, 0);
1104 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1105 mdt_exit_ucred(info);
1108 mdt_client_compatibility(info);
1109 rc2 = mdt_fix_reply(info);
1113 mdt_thread_info_fini(info);
1118 * Exchange MOF_LOV_CREATED flags between two objects after a
1119 * layout swap. No assumption is made on whether o1 or o2 have
1120 * created objects or not.
1122 * \param[in,out] o1 First swap layout object
1123 * \param[in,out] o2 Second swap layout object
1125 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1129 mutex_lock(&o1->mot_lov_mutex);
1130 mutex_lock(&o2->mot_lov_mutex);
1132 o1_flags = o1->mot_flags;
1133 o1->mot_flags = (o1->mot_flags & ~MOF_LOV_CREATED) |
1134 (o2->mot_flags & MOF_LOV_CREATED);
1136 o2->mot_flags = (o2->mot_flags & ~MOF_LOV_CREATED) |
1137 (o1_flags & MOF_LOV_CREATED);
1139 mutex_unlock(&o2->mot_lov_mutex);
1140 mutex_unlock(&o1->mot_lov_mutex);
1143 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1145 struct mdt_thread_info *info;
1146 struct ptlrpc_request *req = tgt_ses_req(tsi);
1147 struct obd_export *exp = req->rq_export;
1148 struct mdt_object *o1, *o2, *o;
1149 struct mdt_lock_handle *lh1, *lh2;
1150 struct mdc_swap_layouts *msl;
1154 /* client does not support layout lock, so layout swaping
1156 * FIXME: there is a problem for old clients which don't support
1157 * layout lock yet. If those clients have already opened the file
1158 * they won't be notified at all so that old layout may still be
1159 * used to do IO. This can be fixed after file release is landed by
1160 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1161 if (!exp_connect_layout(exp))
1162 RETURN(-EOPNOTSUPP);
1164 info = tsi2mdt_info(tsi);
1166 if (info->mti_dlm_req != NULL)
1167 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1169 o1 = info->mti_object;
1170 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1171 &info->mti_body->mbo_fid2);
1173 GOTO(out, rc = PTR_ERR(o));
1175 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1176 GOTO(put, rc = -ENOENT);
1178 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1179 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1185 /* permission check. Make sure the calling process having permission
1186 * to write both files. */
1187 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1192 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1197 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1199 GOTO(put, rc = -EPROTO);
1201 lh1 = &info->mti_lh[MDT_LH_NEW];
1202 mdt_lock_reg_init(lh1, LCK_EX);
1203 lh2 = &info->mti_lh[MDT_LH_OLD];
1204 mdt_lock_reg_init(lh2, LCK_EX);
1206 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1207 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1211 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1212 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1216 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1217 mdt_object_child(o2), msl->msl_flags);
1221 mdt_swap_lov_flag(o1, o2);
1224 mdt_object_unlock(info, o2, lh2, rc);
1226 mdt_object_unlock(info, o1, lh1, rc);
1228 mdt_object_put(info->mti_env, o);
1230 mdt_thread_info_fini(info);
1234 static int mdt_raw_lookup(struct mdt_thread_info *info,
1235 struct mdt_object *parent,
1236 const struct lu_name *lname,
1237 struct ldlm_reply *ldlm_rep)
1239 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1243 LASSERT(!info->mti_cross_ref);
1245 /* Only got the fid of this obj by name */
1246 fid_zero(child_fid);
1247 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1248 lname, child_fid, &info->mti_spec);
1250 struct mdt_body *repbody;
1252 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1253 repbody->mbo_fid1 = *child_fid;
1254 repbody->mbo_valid = OBD_MD_FLID;
1255 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1256 } else if (rc == -ENOENT) {
1257 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1264 * UPDATE lock should be taken against parent, and be released before exit;
1265 * child_bits lock should be taken against child, and be returned back:
1266 * (1)normal request should release the child lock;
1267 * (2)intent request will grant the lock to client.
1269 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1270 struct mdt_lock_handle *lhc,
1272 struct ldlm_reply *ldlm_rep)
1274 struct ptlrpc_request *req = mdt_info_req(info);
1275 struct mdt_body *reqbody = NULL;
1276 struct mdt_object *parent = info->mti_object;
1277 struct mdt_object *child;
1278 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1279 struct lu_name *lname = NULL;
1280 struct mdt_lock_handle *lhp = NULL;
1281 struct ldlm_lock *lock;
1288 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1289 LASSERT(ergo(is_resent,
1290 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1295 if (info->mti_cross_ref) {
1296 /* Only getattr on the child. Parent is on another node. */
1297 mdt_set_disposition(info, ldlm_rep,
1298 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1300 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1302 PFID(mdt_object_fid(child)), ldlm_rep);
1304 rc = mdt_check_resent_lock(info, child, lhc);
1307 } else if (rc > 0) {
1308 mdt_lock_handle_init(lhc);
1309 mdt_lock_reg_init(lhc, LCK_PR);
1312 * Object's name is on another MDS, no lookup or layout
1313 * lock is needed here but update lock is.
1315 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1316 MDS_INODELOCK_LAYOUT);
1317 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1319 rc = mdt_object_lock(info, child, lhc, child_bits,
1325 /* Finally, we can get attr for child. */
1326 if (!mdt_object_exists(child)) {
1327 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1329 "remote object doesn't exist.\n");
1330 mdt_object_unlock(info, child, lhc, 1);
1334 rc = mdt_getattr_internal(info, child, 0);
1335 if (unlikely(rc != 0))
1336 mdt_object_unlock(info, child, lhc, 1);
1341 lname = &info->mti_name;
1342 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1344 if (lu_name_is_valid(lname)) {
1345 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1346 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1347 PNAME(lname), ldlm_rep);
1349 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1350 if (unlikely(reqbody == NULL))
1351 RETURN(err_serious(-EPROTO));
1353 *child_fid = reqbody->mbo_fid2;
1355 if (unlikely(!fid_is_sane(child_fid)))
1356 RETURN(err_serious(-EINVAL));
1358 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1360 PFID(mdt_object_fid(parent)),
1361 PFID(&reqbody->mbo_fid2), ldlm_rep);
1364 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1366 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1367 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1369 "Parent doesn't exist!\n");
1373 if (mdt_object_remote(parent)) {
1374 CERROR("%s: parent "DFID" is on remote target\n",
1375 mdt_obd_name(info->mti_mdt),
1376 PFID(mdt_object_fid(parent)));
1380 if (lu_name_is_valid(lname)) {
1381 /* Always allow to lookup ".." */
1382 if (unlikely(lname->ln_namelen == 2 &&
1383 lname->ln_name[0] == '.' &&
1384 lname->ln_name[1] == '.'))
1385 info->mti_spec.sp_permitted = 1;
1387 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1388 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1393 /* step 1: lock parent only if parent is a directory */
1394 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1395 lhp = &info->mti_lh[MDT_LH_PARENT];
1396 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1397 rc = mdt_object_lock(info, parent, lhp,
1398 MDS_INODELOCK_UPDATE,
1400 if (unlikely(rc != 0))
1404 /* step 2: lookup child's fid by name */
1405 fid_zero(child_fid);
1406 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1407 child_fid, &info->mti_spec);
1409 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1412 GOTO(out_parent, rc);
1415 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1418 *step 3: find the child object by fid & lock it.
1419 * regardless if it is local or remote.
1421 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1422 * set parent dir fid the same as child fid in getattr by fid case
1423 * we should not lu_object_find() the object again, could lead
1424 * to hung if there is a concurrent unlink destroyed the object.
1426 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1427 mdt_object_get(info->mti_env, parent);
1430 child = mdt_object_find(info->mti_env, info->mti_mdt,
1434 if (unlikely(IS_ERR(child)))
1435 GOTO(out_parent, rc = PTR_ERR(child));
1437 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1438 rc = mdt_check_resent_lock(info, child, lhc);
1440 GOTO(out_child, rc);
1441 } else if (rc > 0) {
1442 mdt_lock_handle_init(lhc);
1443 mdt_lock_reg_init(lhc, LCK_PR);
1446 if (!mdt_object_exists(child)) {
1447 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1449 "Object doesn't exist!\n");
1450 GOTO(out_child, rc = -ENOENT);
1453 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1454 mdt_object_exists(child) && !mdt_object_remote(child)) {
1455 struct md_attr *ma = &info->mti_attr;
1458 ma->ma_need = MA_INODE;
1459 rc = mdt_attr_get_complex(info, child, ma);
1460 if (unlikely(rc != 0))
1461 GOTO(out_child, rc);
1463 /* If the file has not been changed for some time, we
1464 * return not only a LOOKUP lock, but also an UPDATE
1465 * lock and this might save us RPC on later STAT. For
1466 * directories, it also let negative dentry cache start
1467 * working for this dir. */
1468 if (ma->ma_valid & MA_INODE &&
1469 ma->ma_attr.la_valid & LA_CTIME &&
1470 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1471 ma->ma_attr.la_ctime < cfs_time_current_sec())
1472 child_bits |= MDS_INODELOCK_UPDATE;
1475 /* layout lock must be granted in a best-effort way
1476 * for IT operations */
1477 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1478 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1479 exp_connect_layout(info->mti_exp) &&
1480 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1481 !mdt_object_remote(child) && ldlm_rep != NULL) {
1482 /* try to grant layout lock for regular file. */
1488 child_bits |= MDS_INODELOCK_LAYOUT;
1489 /* try layout lock, it may fail to be granted due to
1490 * contention at LOOKUP or UPDATE */
1491 if (!mdt_object_lock_try(info, child, lhc, child_bits,
1493 child_bits &= ~MDS_INODELOCK_LAYOUT;
1494 LASSERT(child_bits != 0);
1495 rc = mdt_object_lock(info, child, lhc,
1496 child_bits, MDT_CROSS_LOCK);
1501 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1502 * client will enqueue the lock to the remote MDT */
1503 if (mdt_object_remote(child))
1504 child_bits &= ~MDS_INODELOCK_UPDATE;
1505 rc = mdt_object_lock(info, child, lhc, child_bits,
1508 if (unlikely(rc != 0))
1509 GOTO(out_child, rc);
1512 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1514 /* finally, we can get attr for child. */
1515 rc = mdt_getattr_internal(info, child, ma_need);
1516 if (unlikely(rc != 0)) {
1517 mdt_object_unlock(info, child, lhc, 1);
1519 /* Debugging code. */
1520 LDLM_DEBUG(lock, "Returning lock to client");
1521 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1522 &lock->l_resource->lr_name),
1523 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1524 PLDLMRES(lock->l_resource),
1525 PFID(mdt_object_fid(child)));
1528 LDLM_LOCK_PUT(lock);
1532 mdt_object_put(info->mti_env, child);
1535 mdt_object_unlock(info, parent, lhp, 1);
1539 /* normal handler: should release the child lock */
1540 static int mdt_getattr_name(struct tgt_session_info *tsi)
1542 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1543 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1544 struct mdt_body *reqbody;
1545 struct mdt_body *repbody;
1549 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1550 LASSERT(reqbody != NULL);
1551 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1552 LASSERT(repbody != NULL);
1554 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1555 repbody->mbo_eadatasize = 0;
1556 repbody->mbo_aclsize = 0;
1558 rc = mdt_init_ucred(info, reqbody);
1560 GOTO(out_shrink, rc);
1562 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1563 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1564 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1565 lhc->mlh_reg_lh.cookie = 0;
1567 mdt_exit_ucred(info);
1570 mdt_client_compatibility(info);
1571 rc2 = mdt_fix_reply(info);
1574 mdt_thread_info_fini(info);
1578 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1579 void *karg, void *uarg);
1581 static int mdt_set_info(struct tgt_session_info *tsi)
1583 struct ptlrpc_request *req = tgt_ses_req(tsi);
1586 int keylen, vallen, rc = 0;
1590 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1592 DEBUG_REQ(D_HA, req, "no set_info key");
1593 RETURN(err_serious(-EFAULT));
1596 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1599 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1601 DEBUG_REQ(D_HA, req, "no set_info val");
1602 RETURN(err_serious(-EFAULT));
1605 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1608 /* Swab any part of val you need to here */
1609 if (KEY_IS(KEY_READ_ONLY)) {
1610 spin_lock(&req->rq_export->exp_lock);
1612 *exp_connect_flags_ptr(req->rq_export) |=
1615 *exp_connect_flags_ptr(req->rq_export) &=
1616 ~OBD_CONNECT_RDONLY;
1617 spin_unlock(&req->rq_export->exp_lock);
1618 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1619 struct changelog_setinfo *cs = val;
1621 if (vallen != sizeof(*cs)) {
1622 CERROR("%s: bad changelog_clear setinfo size %d\n",
1623 tgt_name(tsi->tsi_tgt), vallen);
1626 if (ptlrpc_req_need_swab(req)) {
1627 __swab64s(&cs->cs_recno);
1628 __swab32s(&cs->cs_id);
1631 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1639 static int mdt_readpage(struct tgt_session_info *tsi)
1641 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1642 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1643 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1644 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1645 struct mdt_body *repbody;
1651 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1652 RETURN(err_serious(-ENOMEM));
1654 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1655 if (repbody == NULL || reqbody == NULL)
1656 RETURN(err_serious(-EFAULT));
1659 * prepare @rdpg before calling lower layers and transfer itself. Here
1660 * reqbody->size contains offset of where to start to read and
1661 * reqbody->nlink contains number bytes to read.
1663 rdpg->rp_hash = reqbody->mbo_size;
1664 if (rdpg->rp_hash != reqbody->mbo_size) {
1665 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1666 rdpg->rp_hash, reqbody->mbo_size);
1670 rdpg->rp_attrs = reqbody->mbo_mode;
1671 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1672 rdpg->rp_attrs |= LUDA_64BITHASH;
1673 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1674 exp_max_brw_size(tsi->tsi_exp));
1675 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1677 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1678 if (rdpg->rp_pages == NULL)
1681 for (i = 0; i < rdpg->rp_npages; ++i) {
1682 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1683 if (rdpg->rp_pages[i] == NULL)
1684 GOTO(free_rdpg, rc = -ENOMEM);
1687 /* call lower layers to fill allocated pages with directory data */
1688 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1690 GOTO(free_rdpg, rc);
1692 /* send pages to client */
1693 rc = tgt_sendpage(tsi, rdpg, rc);
1698 for (i = 0; i < rdpg->rp_npages; i++)
1699 if (rdpg->rp_pages[i] != NULL)
1700 __free_page(rdpg->rp_pages[i]);
1701 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1703 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1709 static int mdt_reint_internal(struct mdt_thread_info *info,
1710 struct mdt_lock_handle *lhc,
1713 struct req_capsule *pill = info->mti_pill;
1714 struct mdt_body *repbody;
1719 rc = mdt_reint_unpack(info, op);
1721 CERROR("Can't unpack reint, rc %d\n", rc);
1722 RETURN(err_serious(rc));
1725 /* for replay (no_create) lmm is not needed, client has it already */
1726 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1727 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1730 /* llog cookies are always 0, the field is kept for compatibility */
1731 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1732 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1734 rc = req_capsule_server_pack(pill);
1736 CERROR("Can't pack response, rc %d\n", rc);
1737 RETURN(err_serious(rc));
1740 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1741 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1743 repbody->mbo_eadatasize = 0;
1744 repbody->mbo_aclsize = 0;
1747 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1749 /* for replay no cookkie / lmm need, because client have this already */
1750 if (info->mti_spec.no_create)
1751 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1752 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1754 rc = mdt_init_ucred_reint(info);
1756 GOTO(out_shrink, rc);
1758 rc = mdt_fix_attr_ucred(info, op);
1760 GOTO(out_ucred, rc = err_serious(rc));
1762 if (mdt_check_resent(info, mdt_reconstruct, lhc)) {
1763 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1764 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1765 GOTO(out_ucred, rc);
1767 rc = mdt_reint_rec(info, lhc);
1770 mdt_exit_ucred(info);
1772 mdt_client_compatibility(info);
1773 rc2 = mdt_fix_reply(info);
1779 static long mdt_reint_opcode(struct ptlrpc_request *req,
1780 const struct req_format **fmt)
1782 struct mdt_device *mdt;
1783 struct mdt_rec_reint *rec;
1786 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1788 opc = rec->rr_opcode;
1789 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1790 if (opc < REINT_MAX && fmt[opc] != NULL)
1791 req_capsule_extend(&req->rq_pill, fmt[opc]);
1793 mdt = mdt_exp2dev(req->rq_export);
1794 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1795 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1796 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1797 opc = err_serious(-EFAULT);
1800 opc = err_serious(-EFAULT);
1805 static int mdt_reint(struct tgt_session_info *tsi)
1809 static const struct req_format *reint_fmts[REINT_MAX] = {
1810 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1811 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1812 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1813 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1814 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1815 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1816 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1817 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
1818 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME
1823 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1825 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1827 * No lock possible here from client to pass it to reint code
1830 rc = mdt_reint_internal(info, NULL, opc);
1831 mdt_thread_info_fini(info);
1836 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1840 /* this should sync the whole device */
1841 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1843 struct dt_device *dt = mdt->mdt_bottom;
1847 rc = dt->dd_ops->dt_sync(env, dt);
1851 /* this should sync this object */
1852 static int mdt_object_sync(struct mdt_thread_info *info)
1854 struct md_object *next;
1858 if (!mdt_object_exists(info->mti_object)) {
1859 CWARN("%s: non existing object "DFID": rc = %d\n",
1860 mdt_obd_name(info->mti_mdt),
1861 PFID(mdt_object_fid(info->mti_object)), -ESTALE);
1864 next = mdt_object_child(info->mti_object);
1865 rc = mo_object_sync(info->mti_env, next);
1870 static int mdt_sync(struct tgt_session_info *tsi)
1872 struct ptlrpc_request *req = tgt_ses_req(tsi);
1873 struct req_capsule *pill = tsi->tsi_pill;
1874 struct mdt_body *body;
1879 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1880 RETURN(err_serious(-ENOMEM));
1882 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
1883 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
1885 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1887 /* sync an object */
1888 rc = mdt_object_sync(info);
1890 const struct lu_fid *fid;
1891 struct lu_attr *la = &info->mti_attr.ma_attr;
1893 info->mti_attr.ma_need = MA_INODE;
1894 info->mti_attr.ma_valid = 0;
1895 rc = mdt_attr_get_complex(info, info->mti_object,
1898 body = req_capsule_server_get(pill,
1900 fid = mdt_object_fid(info->mti_object);
1901 mdt_pack_attr2body(info, body, la, fid);
1904 mdt_thread_info_fini(info);
1907 mdt_counter_incr(req, LPROC_MDT_SYNC);
1913 * Handle quota control requests to consult current usage/limit, but also
1914 * to configure quota enforcement
1916 static int mdt_quotactl(struct tgt_session_info *tsi)
1918 struct obd_export *exp = tsi->tsi_exp;
1919 struct req_capsule *pill = tsi->tsi_pill;
1920 struct obd_quotactl *oqctl, *repoqc;
1922 struct mdt_device *mdt = mdt_exp2dev(exp);
1923 struct lu_device *qmt = mdt->mdt_qmt_dev;
1924 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
1927 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1929 RETURN(err_serious(-EPROTO));
1931 rc = req_capsule_server_pack(pill);
1933 RETURN(err_serious(rc));
1935 switch (oqctl->qc_cmd) {
1937 case LUSTRE_Q_INVALIDATE:
1938 case LUSTRE_Q_FINVALIDATE:
1942 /* deprecated, not used any more */
1943 RETURN(-EOPNOTSUPP);
1944 /* master quotactl */
1947 if (!nodemap_can_setquota(nodemap))
1952 RETURN(-EOPNOTSUPP);
1953 /* slave quotactl */
1958 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1962 /* map uid/gid for remote client */
1964 if (exp_connect_rmtclient(exp)) {
1965 struct lustre_idmap_table *idmap;
1967 idmap = exp->exp_mdt_data.med_idmap;
1969 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1970 oqctl->qc_cmd != Q_GETINFO))
1973 if (oqctl->qc_type == USRQUOTA)
1974 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1976 else if (oqctl->qc_type == GRPQUOTA)
1977 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1982 if (id == CFS_IDMAP_NOTFOUND) {
1983 CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
1988 if (oqctl->qc_type == USRQUOTA)
1989 id = nodemap_map_id(nodemap, NODEMAP_UID,
1990 NODEMAP_CLIENT_TO_FS, id);
1991 else if (oqctl->qc_type == GRPQUOTA)
1992 id = nodemap_map_id(nodemap, NODEMAP_UID,
1993 NODEMAP_CLIENT_TO_FS, id);
1995 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1997 RETURN(err_serious(-EFAULT));
1999 if (oqctl->qc_id != id)
2000 swap(oqctl->qc_id, id);
2002 switch (oqctl->qc_cmd) {
2008 /* forward quotactl request to QMT */
2009 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2014 /* slave quotactl */
2015 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2020 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2024 if (oqctl->qc_id != id)
2025 swap(oqctl->qc_id, id);
2031 /** clone llog ctxt from child (mdd)
2032 * This allows remote llog (replicator) access.
2033 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2034 * context was originally set up, or we can handle them directly.
2035 * I choose the latter, but that means I need any llog
2036 * contexts set up by child to be accessable by the mdt. So we clone the
2037 * context into our context list here.
2039 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2042 struct md_device *next = mdt->mdt_child;
2043 struct llog_ctxt *ctxt;
2046 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2049 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2050 if (rc || ctxt == NULL) {
2054 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2056 CERROR("Can't set mdt ctxt %d\n", rc);
2061 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2062 struct mdt_device *mdt, int idx)
2064 struct llog_ctxt *ctxt;
2066 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2069 /* Put once for the get we just did, and once for the clone */
2070 llog_ctxt_put(ctxt);
2071 llog_ctxt_put(ctxt);
2076 * sec context handlers
2078 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2082 rc = mdt_handle_idmap(tsi);
2084 struct ptlrpc_request *req = tgt_ses_req(tsi);
2087 opc = lustre_msg_get_opc(req->rq_reqmsg);
2088 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2089 sptlrpc_svc_ctx_invalidate(req);
2092 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2098 * quota request handlers
2100 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2102 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2103 struct lu_device *qmt = mdt->mdt_qmt_dev;
2108 RETURN(err_serious(-EOPNOTSUPP));
2110 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2114 struct mdt_object *mdt_object_new(const struct lu_env *env,
2115 struct mdt_device *d,
2116 const struct lu_fid *f)
2118 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2119 struct lu_object *o;
2120 struct mdt_object *m;
2123 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2124 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2125 if (unlikely(IS_ERR(o)))
2126 m = (struct mdt_object *)o;
2132 struct mdt_object *mdt_object_find(const struct lu_env *env,
2133 struct mdt_device *d,
2134 const struct lu_fid *f)
2136 struct lu_object *o;
2137 struct mdt_object *m;
2140 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2141 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2142 if (unlikely(IS_ERR(o)))
2143 m = (struct mdt_object *)o;
2151 * Asyncronous commit for mdt device.
2153 * Pass asynchonous commit call down the MDS stack.
2155 * \param env environment
2156 * \param mdt the mdt device
2158 static void mdt_device_commit_async(const struct lu_env *env,
2159 struct mdt_device *mdt)
2161 struct dt_device *dt = mdt->mdt_bottom;
2164 rc = dt->dd_ops->dt_commit_async(env, dt);
2165 if (unlikely(rc != 0))
2166 CWARN("%s: async commit start failed: rc = %d\n",
2167 mdt_obd_name(mdt), rc);
2171 * Mark the lock as "synchonous".
2173 * Mark the lock to deffer transaction commit to the unlock time.
2175 * \param lock the lock to mark as "synchonous"
2177 * \see mdt_is_lock_sync
2178 * \see mdt_save_lock
2180 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2182 lock->l_ast_data = (void*)1;
2186 * Check whehter the lock "synchonous" or not.
2188 * \param lock the lock to check
2189 * \retval 1 the lock is "synchonous"
2190 * \retval 0 the lock isn't "synchronous"
2192 * \see mdt_set_lock_sync
2193 * \see mdt_save_lock
2195 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2197 return lock->l_ast_data != NULL;
2201 * Blocking AST for mdt locks.
2203 * Starts transaction commit if in case of COS lock conflict or
2204 * deffers such a commit to the mdt_save_lock.
2206 * \param lock the lock which blocks a request or cancelling lock
2207 * \param desc unused
2208 * \param data unused
2209 * \param flag indicates whether this cancelling or blocking callback
2211 * \see ldlm_blocking_ast_nocheck
2213 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2214 void *data, int flag)
2216 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2217 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2221 if (flag == LDLM_CB_CANCELING)
2223 lock_res_and_lock(lock);
2224 if (lock->l_blocking_ast != mdt_blocking_ast) {
2225 unlock_res_and_lock(lock);
2228 if (mdt_cos_is_enabled(mdt) &&
2229 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2230 lock->l_blocking_lock != NULL &&
2231 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2232 mdt_set_lock_sync(lock);
2234 rc = ldlm_blocking_ast_nocheck(lock);
2236 /* There is no lock conflict if l_blocking_lock == NULL,
2237 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2238 * when the last reference to a local lock was released */
2239 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2242 rc = lu_env_init(&env, LCT_LOCAL);
2243 if (unlikely(rc != 0))
2244 CWARN("%s: lu_env initialization failed, cannot "
2245 "start asynchronous commit: rc = %d\n",
2248 mdt_device_commit_async(&env, mdt);
2254 /* Used for cross-MDT lock */
2255 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2256 void *data, int flag)
2258 struct lustre_handle lockh;
2262 case LDLM_CB_BLOCKING:
2263 ldlm_lock2handle(lock, &lockh);
2264 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
2266 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2270 case LDLM_CB_CANCELING:
2271 LDLM_DEBUG(lock, "Revoke remote lock\n");
2279 int mdt_check_resent_lock(struct mdt_thread_info *info,
2280 struct mdt_object *mo,
2281 struct mdt_lock_handle *lhc)
2283 /* the lock might already be gotten in ldlm_handle_enqueue() */
2284 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2285 struct ptlrpc_request *req = mdt_info_req(info);
2286 struct ldlm_lock *lock;
2288 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2289 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2291 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2292 * a resend case, however, it could be already destroyed
2293 * due to client eviction or a raced cancel RPC. */
2294 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
2295 lhc->mlh_reg_lh.cookie);
2299 if (!fid_res_name_eq(mdt_object_fid(mo),
2300 &lock->l_resource->lr_name)) {
2301 CWARN("%s: Although resent, but still not "
2302 "get child lock:"DFID"\n",
2303 info->mti_exp->exp_obd->obd_name,
2304 PFID(mdt_object_fid(mo)));
2305 LDLM_LOCK_PUT(lock);
2308 LDLM_LOCK_PUT(lock);
2314 int mdt_remote_object_lock(struct mdt_thread_info *mti,
2315 struct mdt_object *o, const struct lu_fid *fid,
2316 struct lustre_handle *lh, ldlm_mode_t mode,
2319 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2320 ldlm_policy_data_t *policy = &mti->mti_policy;
2321 struct ldlm_res_id *res_id = &mti->mti_res_id;
2325 LASSERT(mdt_object_remote(o));
2327 LASSERT(ibits == MDS_INODELOCK_UPDATE);
2329 fid_build_reg_res_name(fid, res_id);
2331 memset(einfo, 0, sizeof(*einfo));
2332 einfo->ei_type = LDLM_IBITS;
2333 einfo->ei_mode = mode;
2334 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2335 einfo->ei_cb_cp = ldlm_completion_ast;
2336 einfo->ei_enq_slave = 0;
2337 einfo->ei_res_id = res_id;
2339 memset(policy, 0, sizeof(*policy));
2340 policy->l_inodebits.bits = ibits;
2342 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2347 static int mdt_object_local_lock(struct mdt_thread_info *info,
2348 struct mdt_object *o,
2349 struct mdt_lock_handle *lh, __u64 ibits,
2350 bool nonblock, int locality)
2352 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2353 ldlm_policy_data_t *policy = &info->mti_policy;
2354 struct ldlm_res_id *res_id = &info->mti_res_id;
2359 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2360 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2361 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2362 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2364 /* Only enqueue LOOKUP lock for remote object */
2365 if (mdt_object_remote(o))
2366 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2368 if (lh->mlh_type == MDT_PDO_LOCK) {
2369 /* check for exists after object is locked */
2370 if (mdt_object_exists(o) == 0) {
2371 /* Non-existent object shouldn't have PDO lock */
2374 /* Non-dir object shouldn't have PDO lock */
2375 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2380 memset(policy, 0, sizeof(*policy));
2381 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2383 dlmflags = LDLM_FL_ATOMIC_CB;
2385 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2388 * Take PDO lock on whole directory and build correct @res_id for lock
2389 * on part of directory.
2391 if (lh->mlh_pdo_hash != 0) {
2392 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2393 mdt_lock_pdo_mode(info, o, lh);
2394 if (lh->mlh_pdo_mode != LCK_NL) {
2396 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2397 * is never going to be sent to client and we do not
2398 * want it slowed down due to possible cancels.
2400 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2401 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2402 policy, res_id, dlmflags,
2403 info->mti_exp == NULL ? NULL :
2404 &info->mti_exp->exp_handle.h_cookie);
2405 if (unlikely(rc != 0))
2406 GOTO(out_unlock, rc);
2410 * Finish res_id initializing by name hash marking part of
2411 * directory which is taking modification.
2413 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2416 policy->l_inodebits.bits = ibits;
2419 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2420 * going to be sent to client. If it is - mdt_intent_policy() path will
2421 * fix it up and turn FL_LOCAL flag off.
2423 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2424 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2425 info->mti_exp == NULL ? NULL :
2426 &info->mti_exp->exp_handle.h_cookie);
2429 mdt_object_unlock(info, o, lh, 1);
2430 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2431 lh->mlh_pdo_hash != 0 &&
2432 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2433 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2439 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2440 struct mdt_lock_handle *lh, __u64 ibits,
2441 bool nonblock, int locality)
2446 if (!mdt_object_remote(o))
2447 return mdt_object_local_lock(info, o, lh, ibits, nonblock,
2450 if (locality == MDT_LOCAL_LOCK) {
2451 CERROR("%s: try to get local lock for remote object"
2452 DFID".\n", mdt_obd_name(info->mti_mdt),
2453 PFID(mdt_object_fid(o)));
2457 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2458 ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2459 MDS_INODELOCK_XATTR);
2460 if (ibits & MDS_INODELOCK_UPDATE) {
2461 /* Sigh, PDO needs to enqueue 2 locks right now, but
2462 * enqueue RPC can only request 1 lock, to avoid extra
2463 * RPC, so it will instead enqueue EX lock for remote
2464 * object anyway XXX*/
2465 if (lh->mlh_type == MDT_PDO_LOCK &&
2466 lh->mlh_pdo_hash != 0) {
2467 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2468 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2469 PFID(mdt_object_fid(o)));
2470 lh->mlh_pdo_hash = 0;
2471 lh->mlh_rreg_mode = LCK_EX;
2472 lh->mlh_type = MDT_REG_LOCK;
2474 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2477 MDS_INODELOCK_UPDATE);
2482 /* Only enqueue LOOKUP lock for remote object */
2483 if (ibits & MDS_INODELOCK_LOOKUP) {
2484 rc = mdt_object_local_lock(info, o, lh,
2485 MDS_INODELOCK_LOOKUP,
2486 nonblock, locality);
2494 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2495 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2497 return mdt_object_lock_internal(info, o, lh, ibits, false, locality);
2500 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2501 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2503 struct mdt_lock_handle tmp = *lh;
2506 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true, locality);
2514 * Save a lock within request object.
2516 * Keep the lock referenced until whether client ACK or transaction
2517 * commit happens or release the lock immediately depending on input
2518 * parameters. If COS is ON, a write lock is converted to COS lock
2521 * \param info thead info object
2522 * \param h lock handle
2523 * \param mode lock mode
2524 * \param decref force immediate lock releasing
2527 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2528 ldlm_mode_t mode, int decref)
2532 if (lustre_handle_is_used(h)) {
2533 if (decref || !info->mti_has_trans ||
2534 !(mode & (LCK_PW | LCK_EX))){
2535 mdt_fid_unlock(h, mode);
2537 struct mdt_device *mdt = info->mti_mdt;
2538 struct ldlm_lock *lock = ldlm_handle2lock(h);
2539 struct ptlrpc_request *req = mdt_info_req(info);
2542 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2544 /* there is no request if mdt_object_unlock() is called
2545 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2546 if (likely(req != NULL)) {
2547 CDEBUG(D_HA, "request = %p reply state = %p"
2548 " transno = "LPD64"\n", req,
2549 req->rq_reply_state, req->rq_transno);
2550 if (mdt_cos_is_enabled(mdt)) {
2552 ldlm_lock_downgrade(lock, LCK_COS);
2555 ptlrpc_save_lock(req, h, mode, no_ack);
2557 ldlm_lock_decref(h, mode);
2559 if (mdt_is_lock_sync(lock)) {
2560 CDEBUG(D_HA, "found sync-lock,"
2561 " async commit started\n");
2562 mdt_device_commit_async(info->mti_env,
2565 LDLM_LOCK_PUT(lock);
2574 * Unlock mdt object.
2576 * Immeditely release the regular lock and the PDO lock or save the
2577 * lock in request and keep them referenced until client ACK or
2578 * transaction commit.
2580 * \param info thread info object
2581 * \param o mdt object
2582 * \param lh mdt lock handle referencing regular and PDO locks
2583 * \param decref force immediate lock releasing
2585 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2586 struct mdt_lock_handle *lh, int decref)
2590 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2591 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2593 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2594 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2599 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2600 const struct lu_fid *f,
2601 struct mdt_lock_handle *lh,
2604 struct mdt_object *o;
2606 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2610 rc = mdt_object_lock(info, o, lh, ibits,
2613 mdt_object_put(info->mti_env, o);
2620 void mdt_object_unlock_put(struct mdt_thread_info * info,
2621 struct mdt_object * o,
2622 struct mdt_lock_handle *lh,
2625 mdt_object_unlock(info, o, lh, decref);
2626 mdt_object_put(info->mti_env, o);
2630 * Generic code handling requests that have struct mdt_body passed in:
2632 * - extract mdt_body from request and save it in @info, if present;
2634 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2637 * - if HABEO_CORPUS flag is set for this request type check whether object
2638 * actually exists on storage (lu_object_exists()).
2641 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2643 const struct mdt_body *body;
2644 struct mdt_object *obj;
2645 const struct lu_env *env;
2646 struct req_capsule *pill;
2650 env = info->mti_env;
2651 pill = info->mti_pill;
2653 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2657 if (!(body->mbo_valid & OBD_MD_FLID))
2660 if (!fid_is_sane(&body->mbo_fid1)) {
2661 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2665 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2667 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2668 mdt_object_put(env, obj);
2671 info->mti_object = obj;
2680 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2682 struct req_capsule *pill = info->mti_pill;
2686 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2687 rc = mdt_body_unpack(info, flags);
2691 if (rc == 0 && (flags & HABEO_REFERO)) {
2693 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2694 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2696 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2697 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2700 rc = req_capsule_server_pack(pill);
2705 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2707 lh->mlh_type = MDT_NUL_LOCK;
2708 lh->mlh_reg_lh.cookie = 0ull;
2709 lh->mlh_reg_mode = LCK_MINMODE;
2710 lh->mlh_pdo_lh.cookie = 0ull;
2711 lh->mlh_pdo_mode = LCK_MINMODE;
2712 lh->mlh_rreg_lh.cookie = 0ull;
2713 lh->mlh_rreg_mode = LCK_MINMODE;
2716 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2718 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2719 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2723 * Initialize fields of struct mdt_thread_info. Other fields are left in
2724 * uninitialized state, because it's too expensive to zero out whole
2725 * mdt_thread_info (> 1K) on each request arrival.
2727 void mdt_thread_info_init(struct ptlrpc_request *req,
2728 struct mdt_thread_info *info)
2732 info->mti_pill = &req->rq_pill;
2735 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2736 mdt_lock_handle_init(&info->mti_lh[i]);
2738 /* mdt device: it can be NULL while CONNECT */
2739 if (req->rq_export) {
2740 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2741 info->mti_exp = req->rq_export;
2743 info->mti_mdt = NULL;
2744 info->mti_env = req->rq_svc_thread->t_env;
2745 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2747 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2748 info->mti_big_buf = LU_BUF_NULL;
2749 info->mti_body = NULL;
2750 info->mti_object = NULL;
2751 info->mti_dlm_req = NULL;
2752 info->mti_has_trans = 0;
2753 info->mti_cross_ref = 0;
2754 info->mti_opdata = 0;
2755 info->mti_big_lmm_used = 0;
2757 info->mti_spec.no_create = 0;
2758 info->mti_spec.sp_rm_entry = 0;
2759 info->mti_spec.sp_permitted = 0;
2761 info->mti_spec.u.sp_ea.eadata = NULL;
2762 info->mti_spec.u.sp_ea.eadatalen = 0;
2765 void mdt_thread_info_fini(struct mdt_thread_info *info)
2769 if (info->mti_object != NULL) {
2770 mdt_object_put(info->mti_env, info->mti_object);
2771 info->mti_object = NULL;
2774 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2775 mdt_lock_handle_fini(&info->mti_lh[i]);
2776 info->mti_env = NULL;
2777 info->mti_pill = NULL;
2778 info->mti_exp = NULL;
2780 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2781 lu_buf_free(&info->mti_big_buf);
2784 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2786 struct mdt_thread_info *mti;
2788 mti = mdt_th_info(tsi->tsi_env);
2789 LASSERT(mti != NULL);
2791 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2792 if (tsi->tsi_corpus != NULL) {
2793 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2794 lu_object_get(tsi->tsi_corpus);
2796 mti->mti_body = tsi->tsi_mdt_body;
2797 mti->mti_dlm_req = tsi->tsi_dlm_req;
2802 static int mdt_tgt_connect(struct tgt_session_info *tsi)
2804 struct ptlrpc_request *req = tgt_ses_req(tsi);
2809 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
2811 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id)
2812 schedule_timeout_and_set_state(TASK_UNINTERRUPTIBLE,
2813 msecs_to_jiffies(3 * MSEC_PER_SEC));
2815 rc = tgt_connect(tsi);
2819 rc = mdt_init_idmap(tsi);
2824 obd_disconnect(class_export_get(req->rq_export));
2843 static int mdt_intent_getattr(enum mdt_it_code opcode,
2844 struct mdt_thread_info *info,
2845 struct ldlm_lock **,
2848 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2849 struct mdt_thread_info *info,
2850 struct ldlm_lock **lockp,
2853 static int mdt_intent_layout(enum mdt_it_code opcode,
2854 struct mdt_thread_info *info,
2855 struct ldlm_lock **,
2857 static int mdt_intent_reint(enum mdt_it_code opcode,
2858 struct mdt_thread_info *info,
2859 struct ldlm_lock **,
2862 static struct mdt_it_flavor {
2863 const struct req_format *it_fmt;
2865 int (*it_act)(enum mdt_it_code ,
2866 struct mdt_thread_info *,
2867 struct ldlm_lock **,
2870 } mdt_it_flavor[] = {
2872 .it_fmt = &RQF_LDLM_INTENT,
2873 /*.it_flags = HABEO_REFERO,*/
2875 .it_act = mdt_intent_reint,
2876 .it_reint = REINT_OPEN
2879 .it_fmt = &RQF_LDLM_INTENT,
2881 * OCREAT is not a MUTABOR request as if the file
2883 * We do the extra check of OBD_CONNECT_RDONLY in
2884 * mdt_reint_open() when we really need to create
2888 .it_act = mdt_intent_reint,
2889 .it_reint = REINT_OPEN
2892 .it_fmt = &RQF_LDLM_INTENT,
2893 .it_flags = MUTABOR,
2894 .it_act = mdt_intent_reint,
2895 .it_reint = REINT_CREATE
2897 [MDT_IT_GETATTR] = {
2898 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2899 .it_flags = HABEO_REFERO,
2900 .it_act = mdt_intent_getattr
2902 [MDT_IT_READDIR] = {
2908 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2909 .it_flags = HABEO_REFERO,
2910 .it_act = mdt_intent_getattr
2913 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2914 .it_flags = MUTABOR,
2916 .it_reint = REINT_UNLINK
2920 .it_flags = MUTABOR,
2923 [MDT_IT_GETXATTR] = {
2924 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2925 .it_flags = HABEO_CORPUS,
2926 .it_act = mdt_intent_getxattr
2929 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2931 .it_act = mdt_intent_layout
2936 mdt_intent_lock_replace(struct mdt_thread_info *info,
2937 struct ldlm_lock **lockp,
2938 struct mdt_lock_handle *lh,
2941 struct ptlrpc_request *req = mdt_info_req(info);
2942 struct ldlm_lock *lock = *lockp;
2943 struct ldlm_lock *new_lock;
2945 /* If possible resent found a lock, @lh is set to its handle */
2946 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2948 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2949 lh->mlh_reg_lh.cookie = 0;
2953 LASSERTF(new_lock != NULL,
2954 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2957 * If we've already given this lock to a client once, then we should
2958 * have no readers or writers. Otherwise, we should have one reader
2959 * _or_ writer ref (which will be zeroed below) before returning the
2962 if (new_lock->l_export == req->rq_export) {
2963 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2965 LASSERT(new_lock->l_export == NULL);
2966 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2971 if (new_lock->l_export == req->rq_export) {
2973 * Already gave this to the client, which means that we
2974 * reconstructed a reply.
2976 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2979 LDLM_LOCK_RELEASE(new_lock);
2980 lh->mlh_reg_lh.cookie = 0;
2981 RETURN(ELDLM_LOCK_REPLACED);
2985 * Fixup the lock to be given to the client.
2987 lock_res_and_lock(new_lock);
2988 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2989 * possible blocking AST. */
2990 while (new_lock->l_readers > 0) {
2991 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2992 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2993 new_lock->l_readers--;
2995 while (new_lock->l_writers > 0) {
2996 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2997 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2998 new_lock->l_writers--;
3001 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3002 new_lock->l_blocking_ast = lock->l_blocking_ast;
3003 new_lock->l_completion_ast = lock->l_completion_ast;
3004 new_lock->l_remote_handle = lock->l_remote_handle;
3005 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3007 unlock_res_and_lock(new_lock);
3009 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3010 &new_lock->l_remote_handle,
3011 &new_lock->l_exp_hash);
3013 LDLM_LOCK_RELEASE(new_lock);
3014 lh->mlh_reg_lh.cookie = 0;
3016 RETURN(ELDLM_LOCK_REPLACED);
3019 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3020 struct ldlm_lock *new_lock,
3021 struct mdt_lock_handle *lh,
3024 struct ptlrpc_request *req = mdt_info_req(info);
3025 struct ldlm_request *dlmreq;
3027 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3030 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3032 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3033 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3035 if (flags & LDLM_FL_RESENT) {
3036 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3037 lh->mlh_reg_mode = new_lock->l_granted_mode;
3039 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3040 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3041 lh->mlh_reg_lh.cookie);
3046 * If the xid matches, then we know this is a resent request, and allow
3047 * it. (It's probably an OPEN, for which we don't send a lock.
3049 if (req_xid_is_last(req))
3053 * This remote handle isn't enqueued, so we never received or processed
3054 * this request. Clear MSG_RESENT, because it can be handled like any
3055 * normal request now.
3057 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3059 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3060 dlmreq->lock_handle[0].cookie);
3063 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3064 struct mdt_thread_info *info,
3065 struct ldlm_lock **lockp,
3068 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3069 struct ldlm_reply *ldlm_rep = NULL;
3073 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3074 * (for the resend case) or a new lock. Below we will use it to
3075 * replace the original lock.
3077 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3078 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3079 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3080 rc = mdt_object_lock(info, info->mti_object, lhc,
3081 MDS_INODELOCK_XATTR,
3087 grc = mdt_getxattr(info);
3089 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3091 if (mdt_info_req(info)->rq_repmsg != NULL)
3092 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3093 if (ldlm_rep == NULL)
3094 RETURN(err_serious(-EFAULT));
3096 ldlm_rep->lock_policy_res2 = grc;
3101 static int mdt_intent_getattr(enum mdt_it_code opcode,
3102 struct mdt_thread_info *info,
3103 struct ldlm_lock **lockp,
3106 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3108 struct ldlm_reply *ldlm_rep;
3109 struct mdt_body *reqbody;
3110 struct mdt_body *repbody;
3114 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3117 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3120 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3121 repbody->mbo_eadatasize = 0;
3122 repbody->mbo_aclsize = 0;
3126 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3128 case MDT_IT_GETATTR:
3129 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3133 CERROR("Unsupported intent (%d)\n", opcode);
3134 GOTO(out_shrink, rc = -EINVAL);
3137 rc = mdt_init_ucred(info, reqbody);
3139 GOTO(out_shrink, rc);
3141 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3142 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3144 /* Get lock from request for possible resent case. */
3145 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3147 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3148 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3150 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3151 ldlm_rep->lock_policy_res2 = 0;
3152 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3153 ldlm_rep->lock_policy_res2) {
3154 lhc->mlh_reg_lh.cookie = 0ull;
3155 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3158 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3161 mdt_exit_ucred(info);
3163 mdt_client_compatibility(info);
3164 rc2 = mdt_fix_reply(info);
3170 static int mdt_intent_layout(enum mdt_it_code opcode,
3171 struct mdt_thread_info *info,
3172 struct ldlm_lock **lockp,
3175 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3176 struct layout_intent *layout;
3178 struct mdt_object *obj = NULL;
3182 if (opcode != MDT_IT_LAYOUT) {
3183 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3188 fid = &info->mti_tmp_fid2;
3189 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3191 /* Get lock from request for possible resent case. */
3192 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3194 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3196 RETURN(PTR_ERR(obj));
3198 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3199 /* get the length of lsm */
3200 rc = mdt_attr_get_eabuf_size(info, obj);
3202 mdt_object_put(info->mti_env, obj);
3206 if (rc > info->mti_mdt->mdt_max_mdsize)
3207 info->mti_mdt->mdt_max_mdsize = rc;
3210 mdt_object_put(info->mti_env, obj);
3212 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3213 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER, rc);
3214 rc = req_capsule_server_pack(info->mti_pill);
3218 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
3219 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3221 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3222 LASSERT(layout != NULL);
3223 if (layout->li_opc == LAYOUT_INTENT_ACCESS)
3224 /* return to normal/resent ldlm handling */
3227 CERROR("%s: Unsupported layout intent (%d)\n",
3228 mdt_obd_name(info->mti_mdt), layout->li_opc);
3232 static int mdt_intent_reint(enum mdt_it_code opcode,
3233 struct mdt_thread_info *info,
3234 struct ldlm_lock **lockp,
3237 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3238 struct ldlm_reply *rep = NULL;
3242 static const struct req_format *intent_fmts[REINT_MAX] = {
3243 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3244 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3249 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3253 if (mdt_it_flavor[opcode].it_reint != opc) {
3254 CERROR("Reint code %ld doesn't match intent: %d\n",
3256 RETURN(err_serious(-EPROTO));
3259 /* Get lock from request for possible resent case. */
3260 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3262 rc = mdt_reint_internal(info, lhc, opc);
3264 /* Check whether the reply has been packed successfully. */
3265 if (mdt_info_req(info)->rq_repmsg != NULL)
3266 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3268 RETURN(err_serious(-EFAULT));
3270 /* MDC expects this in any case */
3272 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3274 /* the open lock or the lock for cross-ref object should be
3275 * returned to the client */
3276 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3277 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3278 rep->lock_policy_res2 = 0;
3279 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3283 rep->lock_policy_res2 = clear_serious(rc);
3285 if (rep->lock_policy_res2 == -ENOENT &&
3286 mdt_get_disposition(rep, DISP_LOOKUP_NEG))
3287 rep->lock_policy_res2 = 0;
3289 lhc->mlh_reg_lh.cookie = 0ull;
3290 if (rc == -ENOTCONN || rc == -ENODEV ||
3291 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3293 * If it is the disconnect error (ENODEV & ENOCONN), the error
3294 * will be returned by rq_status, and client at ptlrpc layer
3295 * will detect this, then disconnect, reconnect the import
3296 * immediately, instead of impacting the following the rpc.
3301 * For other cases, the error will be returned by intent, and client
3302 * will retrieve the result from intent.
3304 RETURN(ELDLM_LOCK_ABORTED);
3307 static int mdt_intent_code(long itcode)
3315 case IT_OPEN|IT_CREAT:
3322 rc = MDT_IT_READDIR;
3325 rc = MDT_IT_GETATTR;
3337 rc = MDT_IT_GETXATTR;
3342 case IT_QUOTA_DQACQ:
3347 CERROR("Unknown intent opcode: %ld\n", itcode);
3354 static int mdt_intent_opc(long itopc, struct mdt_thread_info *info,
3355 struct ldlm_lock **lockp, __u64 flags)
3357 struct req_capsule *pill = info->mti_pill;
3358 struct ptlrpc_request *req = mdt_info_req(info);
3359 struct mdt_it_flavor *flv;
3364 opc = mdt_intent_code(itopc);
3368 if (opc == MDT_IT_QUOTA) {
3369 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3372 RETURN(-EOPNOTSUPP);
3374 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3375 /* pass the request to quota master */
3376 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3377 mdt_info_req(info), lockp,
3382 flv = &mdt_it_flavor[opc];
3383 if (flv->it_fmt != NULL)
3384 req_capsule_extend(pill, flv->it_fmt);
3386 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3390 if (flv->it_flags & MUTABOR &&
3391 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3394 if (flv->it_act != NULL) {
3395 struct ldlm_reply *rep;
3397 /* execute policy */
3398 rc = flv->it_act(opc, info, lockp, flags);
3400 /* Check whether the reply has been packed successfully. */
3401 if (req->rq_repmsg != NULL) {
3402 rep = req_capsule_server_get(info->mti_pill,
3404 rep->lock_policy_res2 =
3405 ptlrpc_status_hton(rep->lock_policy_res2);
3412 static int mdt_intent_policy(struct ldlm_namespace *ns,
3413 struct ldlm_lock **lockp, void *req_cookie,
3414 ldlm_mode_t mode, __u64 flags, void *data)
3416 struct tgt_session_info *tsi;
3417 struct mdt_thread_info *info;
3418 struct ptlrpc_request *req = req_cookie;
3419 struct ldlm_intent *it;
3420 struct req_capsule *pill;
3425 LASSERT(req != NULL);
3427 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3429 info = tsi2mdt_info(tsi);
3430 LASSERT(info != NULL);
3431 pill = info->mti_pill;
3432 LASSERT(pill->rc_req == req);
3434 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3435 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3436 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3438 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3442 /* Lock without inodebits makes no sense and will oops
3443 * later in ldlm. Let's check it now to see if we have
3444 * ibits corrupted somewhere in mdt_intent_opc().
3445 * The case for client miss to set ibits has been
3446 * processed by others. */
3447 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3448 lr_type == LDLM_IBITS,
3449 info->mti_dlm_req->lock_desc.\
3450 l_policy_data.l_inodebits.bits != 0));
3452 rc = err_serious(-EFAULT);
3454 /* No intent was provided */
3455 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3456 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3457 rc = req_capsule_server_pack(pill);
3459 rc = err_serious(rc);
3461 mdt_thread_info_fini(info);
3465 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3467 struct seq_server_site *ss = mdt_seq_site(mdt);
3469 if (ss->ss_node_id == 0)
3472 if (ss->ss_client_seq != NULL) {
3473 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3474 ss->ss_client_seq->lcs_exp = NULL;
3477 if (ss->ss_server_fld != NULL) {
3478 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3479 ss->ss_server_fld->lsf_control_exp = NULL;
3483 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3485 struct seq_server_site *ss = mdt_seq_site(mdt);
3490 if (ss->ss_server_seq != NULL)
3491 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3494 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3496 mdt_seq_fini_cli(mdt);
3497 mdt_deregister_seq_exp(mdt);
3499 return seq_site_fini(env, mdt_seq_site(mdt));
3503 * It will retrieve its FLDB entries from MDT0, and it only happens
3504 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3505 * and it needs to refresh FLDB from the MDT0.
3507 static int mdt_register_lwp_callback(void *data)
3510 struct mdt_device *mdt = data;
3511 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3515 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3517 if (!likely(fld->lsf_new))
3520 rc = lu_env_init(&env, LCT_MD_THREAD);
3522 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3526 rc = fld_update_from_controller(&env, fld);
3528 CERROR("%s: cannot update controller: rc = %d\n",
3529 mdt_obd_name(mdt), rc);
3537 static int mdt_register_seq_exp(struct mdt_device *mdt)
3539 struct seq_server_site *ss = mdt_seq_site(mdt);
3540 char *lwp_name = NULL;
3543 if (ss->ss_node_id == 0)
3546 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3547 if (lwp_name == NULL)
3548 GOTO(out_free, rc = -ENOMEM);
3550 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3554 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3559 rc = lustre_register_lwp_item(lwp_name,
3560 &ss->ss_server_fld->lsf_control_exp,
3561 mdt_register_lwp_callback, mdt);
3563 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3564 ss->ss_client_seq->lcs_exp = NULL;
3568 if (lwp_name != NULL)
3569 OBD_FREE(lwp_name, MAX_OBD_NAME);
3575 * Init client sequence manager which is used by local MDS to talk to sequence
3576 * controller on remote node.
3578 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3580 struct seq_server_site *ss = mdt_seq_site(mdt);
3585 /* check if this is adding the first MDC and controller is not yet
3587 OBD_ALLOC_PTR(ss->ss_client_seq);
3588 if (ss->ss_client_seq == NULL)
3591 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3592 if (prefix == NULL) {
3593 OBD_FREE_PTR(ss->ss_client_seq);
3594 ss->ss_client_seq = NULL;
3598 /* Note: seq_client_fini will be called in seq_site_fini */
3599 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3600 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3601 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3603 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3605 OBD_FREE_PTR(ss->ss_client_seq);
3606 ss->ss_client_seq = NULL;
3610 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3615 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3617 struct seq_server_site *ss;
3621 ss = mdt_seq_site(mdt);
3622 /* init sequence controller server(MDT0) */
3623 if (ss->ss_node_id == 0) {
3624 OBD_ALLOC_PTR(ss->ss_control_seq);
3625 if (ss->ss_control_seq == NULL)
3628 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3629 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3632 GOTO(out_seq_fini, rc);
3635 /* Init normal sequence server */
3636 OBD_ALLOC_PTR(ss->ss_server_seq);
3637 if (ss->ss_server_seq == NULL)
3638 GOTO(out_seq_fini, rc = -ENOMEM);
3640 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3641 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3643 GOTO(out_seq_fini, rc);
3645 /* init seq client for seq server to talk to seq controller(MDT0) */
3646 rc = mdt_seq_init_cli(env, mdt);
3648 GOTO(out_seq_fini, rc);
3650 if (ss->ss_node_id != 0)
3651 /* register controller export through lwp */
3652 rc = mdt_register_seq_exp(mdt);
3657 mdt_seq_fini(env, mdt);
3665 static int mdt_fld_fini(const struct lu_env *env,
3666 struct mdt_device *m)
3668 struct seq_server_site *ss = mdt_seq_site(m);
3671 if (ss && ss->ss_server_fld) {
3672 fld_server_fini(env, ss->ss_server_fld);
3673 OBD_FREE_PTR(ss->ss_server_fld);
3674 ss->ss_server_fld = NULL;
3680 static int mdt_fld_init(const struct lu_env *env,
3682 struct mdt_device *m)
3684 struct seq_server_site *ss;
3688 ss = mdt_seq_site(m);
3690 OBD_ALLOC_PTR(ss->ss_server_fld);
3691 if (ss->ss_server_fld == NULL)
3692 RETURN(rc = -ENOMEM);
3694 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3697 OBD_FREE_PTR(ss->ss_server_fld);
3698 ss->ss_server_fld = NULL;
3705 static void mdt_stack_pre_fini(const struct lu_env *env,
3706 struct mdt_device *m, struct lu_device *top)
3708 struct lustre_cfg_bufs *bufs;
3709 struct lustre_cfg *lcfg;
3710 struct mdt_thread_info *info;
3715 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3716 LASSERT(info != NULL);
3718 bufs = &info->mti_u.bufs;
3720 LASSERT(m->mdt_child_exp);
3721 LASSERT(m->mdt_child_exp->exp_obd);
3723 /* process cleanup, pass mdt obd name to get obd umount flags */
3724 /* XXX: this is needed because all layers are referenced by
3725 * objects (some of them are pinned by osd, for example *
3726 * the proper solution should be a model where object used
3727 * by osd only doesn't have mdt/mdd slices -bzzz */
3728 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3729 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3730 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3734 top->ld_ops->ldo_process_config(env, top, lcfg);
3735 lustre_cfg_free(lcfg);
3739 static void mdt_stack_fini(const struct lu_env *env,
3740 struct mdt_device *m, struct lu_device *top)
3742 struct obd_device *obd = mdt2obd_dev(m);
3743 struct lustre_cfg_bufs *bufs;
3744 struct lustre_cfg *lcfg;
3745 struct mdt_thread_info *info;
3749 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3750 LASSERT(info != NULL);
3752 lu_dev_del_linkage(top->ld_site, top);
3754 lu_site_purge(env, top->ld_site, -1);
3756 bufs = &info->mti_u.bufs;
3757 /* process cleanup, pass mdt obd name to get obd umount flags */
3758 /* another purpose is to let all layers to release their objects */
3759 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3764 lustre_cfg_bufs_set_string(bufs, 1, flags);
3765 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3770 top->ld_ops->ldo_process_config(env, top, lcfg);
3771 lustre_cfg_free(lcfg);
3773 lu_site_purge(env, top->ld_site, -1);
3775 m->mdt_child = NULL;
3776 m->mdt_bottom = NULL;
3778 obd_disconnect(m->mdt_child_exp);
3779 m->mdt_child_exp = NULL;
3781 obd_disconnect(m->mdt_bottom_exp);
3782 m->mdt_child_exp = NULL;
3785 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3786 const char *next, struct obd_export **exp)
3788 struct obd_connect_data *data = NULL;
3789 struct obd_device *obd;
3793 OBD_ALLOC_PTR(data);
3795 GOTO(out, rc = -ENOMEM);
3797 obd = class_name2obd(next);
3799 CERROR("%s: can't locate next device: %s\n",
3800 mdt_obd_name(m), next);
3801 GOTO(out, rc = -ENOTCONN);
3804 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3805 data->ocd_version = LUSTRE_VERSION_CODE;
3807 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3809 CERROR("%s: cannot connect to next dev %s (%d)\n",
3810 mdt_obd_name(m), next, rc);
3820 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3821 struct lustre_cfg *cfg)
3823 char *dev = lustre_cfg_string(cfg, 0);
3824 int rc, name_size, uuid_size;
3825 char *name, *uuid, *p;
3826 struct lustre_cfg_bufs *bufs;
3827 struct lustre_cfg *lcfg;
3828 struct obd_device *obd;
3829 struct lustre_profile *lprof;
3830 struct lu_site *site;
3833 /* in 1.8 we had the only device in the stack - MDS.
3834 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3835 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3836 * to generate names and setup MDT, MDD. MDT will be using
3837 * generated name to connect to MDD. for MDD the next device
3838 * will be LOD with name taken from so called "profile" which
3839 * is generated by mount_option line
3841 * 1.8 MGS generates config. commands like this:
3842 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3843 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3844 * 2.0 MGS generates config. commands like this:
3845 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3846 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3847 * 3:lustre-MDT0000-mdtlov 4:f
3849 * we generate MDD name from MDT one, just replacing T with D
3851 * after all the preparations, the logical equivalent will be
3852 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3853 * 3:lustre-MDT0000-mdtlov 4:f
3854 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3855 * 3:lustre-MDD0000 4:f
3857 * notice we build the stack from down to top: MDD first, then MDT */
3859 name_size = MAX_OBD_NAME;
3860 uuid_size = MAX_OBD_NAME;
3862 OBD_ALLOC(name, name_size);
3863 OBD_ALLOC(uuid, uuid_size);
3864 if (name == NULL || uuid == NULL)
3865 GOTO(cleanup_mem, rc = -ENOMEM);
3867 OBD_ALLOC_PTR(bufs);
3869 GOTO(cleanup_mem, rc = -ENOMEM);
3872 p = strstr(name, "-MDT");
3874 GOTO(free_bufs, rc = -ENOMEM);
3877 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3879 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3880 if (lprof == NULL || lprof->lp_dt == NULL) {
3881 CERROR("can't find the profile: %s\n",
3882 lustre_cfg_string(cfg, 0));
3883 GOTO(free_bufs, rc = -EINVAL);
3886 lustre_cfg_bufs_reset(bufs, name);
3887 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3888 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3889 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3891 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3893 GOTO(free_bufs, rc = -ENOMEM);
3895 rc = class_attach(lcfg);
3897 GOTO(lcfg_cleanup, rc);
3899 obd = class_name2obd(name);
3901 CERROR("Can not find obd %s (%s in config)\n",
3902 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3903 GOTO(lcfg_cleanup, rc = -EINVAL);
3906 lustre_cfg_free(lcfg);
3908 lustre_cfg_bufs_reset(bufs, name);
3909 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3910 lustre_cfg_bufs_set_string(bufs, 2, dev);
3911 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3913 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3915 GOTO(class_detach, rc = -ENOMEM);
3917 rc = class_setup(obd, lcfg);
3919 GOTO(class_detach, rc);
3921 /* connect to MDD we just setup */
3922 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3924 GOTO(class_detach, rc);
3926 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3928 LASSERT(mdt_lu_site(mdt) == NULL);
3929 mdt->mdt_lu_dev.ld_site = site;
3930 site->ls_top_dev = &mdt->mdt_lu_dev;
3931 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3933 /* now connect to bottom OSD */
3934 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3935 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3937 GOTO(class_detach, rc);
3939 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3941 rc = lu_env_refill((struct lu_env *)env);
3943 CERROR("Failure to refill session: '%d'\n", rc);
3945 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3950 class_detach(obd, lcfg);
3952 lustre_cfg_free(lcfg);
3957 OBD_FREE(name, name_size);
3959 OBD_FREE(uuid, uuid_size);
3963 /* setup quota master target on MDT0 */
3964 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3965 struct lustre_cfg *cfg)
3967 struct obd_device *obd;
3968 char *dev = lustre_cfg_string(cfg, 0);
3969 char *qmtname, *uuid, *p;
3970 struct lustre_cfg_bufs *bufs;
3971 struct lustre_cfg *lcfg;
3972 struct lustre_profile *lprof;
3973 struct obd_connect_data *data;
3977 LASSERT(mdt->mdt_qmt_exp == NULL);
3978 LASSERT(mdt->mdt_qmt_dev == NULL);
3980 /* quota master is on MDT0 only for now */
3981 if (mdt->mdt_seq_site.ss_node_id != 0)
3984 /* MGS generates config commands which look as follows:
3985 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3986 * 3:lustre-MDT0000-mdtlov 4:f
3988 * We generate the QMT name from the MDT one, just replacing MD with QM
3989 * after all the preparations, the logical equivalent will be:
3990 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
3991 * 3:lustre-MDT0000-osd 4:f */
3992 OBD_ALLOC(qmtname, MAX_OBD_NAME);
3993 OBD_ALLOC(uuid, UUID_MAX);
3994 OBD_ALLOC_PTR(bufs);
3995 OBD_ALLOC_PTR(data);
3996 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
3997 GOTO(cleanup_mem, rc = -ENOMEM);
3999 strcpy(qmtname, dev);
4000 p = strstr(qmtname, "-MDT");
4002 GOTO(cleanup_mem, rc = -ENOMEM);
4003 /* replace MD with QM */
4007 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4009 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4010 if (lprof == NULL || lprof->lp_dt == NULL) {
4011 CERROR("can't find profile for %s\n",
4012 lustre_cfg_string(cfg, 0));
4013 GOTO(cleanup_mem, rc = -EINVAL);
4016 lustre_cfg_bufs_reset(bufs, qmtname);
4017 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4018 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4019 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4021 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4023 GOTO(cleanup_mem, rc = -ENOMEM);
4025 rc = class_attach(lcfg);
4027 GOTO(lcfg_cleanup, rc);
4029 obd = class_name2obd(qmtname);
4031 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4032 lustre_cfg_string(cfg, 0));
4033 GOTO(lcfg_cleanup, rc = -EINVAL);
4036 lustre_cfg_free(lcfg);
4038 lustre_cfg_bufs_reset(bufs, qmtname);
4039 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4040 lustre_cfg_bufs_set_string(bufs, 2, dev);
4042 /* for quota, the next device should be the OSD device */
4043 lustre_cfg_bufs_set_string(bufs, 3,
4044 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4046 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4048 GOTO(class_detach, rc = -ENOMEM);
4050 rc = class_setup(obd, lcfg);
4052 GOTO(class_detach, rc);
4054 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4056 /* configure local quota objects */
4057 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4061 GOTO(class_cleanup, rc);
4063 /* connect to quota master target */
4064 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4065 data->ocd_version = LUSTRE_VERSION_CODE;
4066 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4069 CERROR("cannot connect to quota master device %s (%d)\n",
4071 GOTO(class_cleanup, rc);
4077 class_manual_cleanup(obd);
4078 mdt->mdt_qmt_dev = NULL;
4082 class_detach(obd, lcfg);
4084 lustre_cfg_free(lcfg);
4089 OBD_FREE(qmtname, MAX_OBD_NAME);
4091 OBD_FREE(uuid, UUID_MAX);
4097 /* Shutdown quota master target associated with mdt */
4098 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4102 if (mdt->mdt_qmt_exp == NULL)
4104 LASSERT(mdt->mdt_qmt_dev != NULL);
4106 /* the qmt automatically shuts down when the mdt disconnects */
4107 obd_disconnect(mdt->mdt_qmt_exp);
4108 mdt->mdt_qmt_exp = NULL;
4109 mdt->mdt_qmt_dev = NULL;
4113 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4114 * for now. This will be removed along with converting rest of MDT code
4115 * to use tgt_session_info */
4116 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4118 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4121 rc = mdt_getxattr(info);
4123 mdt_thread_info_fini(info);
4127 static struct tgt_handler mdt_tgt_handlers[] = {
4128 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4129 0, MDS_CONNECT, mdt_tgt_connect,
4130 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4131 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4132 0, MDS_DISCONNECT, tgt_disconnect,
4133 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4134 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4135 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4136 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4137 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4138 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4139 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4140 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4142 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4143 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4144 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4145 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4146 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4147 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4148 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4149 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4151 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4152 mdt_hsm_ct_register),
4153 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4154 mdt_hsm_ct_unregister),
4155 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4157 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4159 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4160 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4162 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4167 static struct tgt_handler mdt_sec_ctx_ops[] = {
4168 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4169 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4170 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4173 static struct tgt_handler mdt_quota_ops[] = {
4174 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4177 static struct tgt_opc_slice mdt_common_slice[] = {
4179 .tos_opc_start = MDS_FIRST_OPC,
4180 .tos_opc_end = MDS_LAST_OPC,
4181 .tos_hs = mdt_tgt_handlers
4184 .tos_opc_start = OBD_FIRST_OPC,
4185 .tos_opc_end = OBD_LAST_OPC,
4186 .tos_hs = tgt_obd_handlers
4189 .tos_opc_start = LDLM_FIRST_OPC,
4190 .tos_opc_end = LDLM_LAST_OPC,
4191 .tos_hs = tgt_dlm_handlers
4194 .tos_opc_start = SEC_FIRST_OPC,
4195 .tos_opc_end = SEC_LAST_OPC,
4196 .tos_hs = mdt_sec_ctx_ops
4199 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4200 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4201 .tos_hs = tgt_out_handlers
4204 .tos_opc_start = FLD_FIRST_OPC,
4205 .tos_opc_end = FLD_LAST_OPC,
4206 .tos_hs = fld_handlers
4209 .tos_opc_start = SEQ_FIRST_OPC,
4210 .tos_opc_end = SEQ_LAST_OPC,
4211 .tos_hs = seq_handlers
4214 .tos_opc_start = QUOTA_DQACQ,
4215 .tos_opc_end = QUOTA_LAST_OPC,
4216 .tos_hs = mdt_quota_ops
4219 .tos_opc_start = LLOG_FIRST_OPC,
4220 .tos_opc_end = LLOG_LAST_OPC,
4221 .tos_hs = tgt_llog_handlers
4224 .tos_opc_start = LFSCK_FIRST_OPC,
4225 .tos_opc_end = LFSCK_LAST_OPC,
4226 .tos_hs = tgt_lfsck_handlers
4234 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4236 struct md_device *next = m->mdt_child;
4237 struct lu_device *d = &m->mdt_lu_dev;
4238 struct obd_device *obd = mdt2obd_dev(m);
4239 struct lfsck_stop stop;
4242 stop.ls_status = LS_PAUSED;
4244 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4246 target_recovery_fini(obd);
4247 ping_evictor_stop();
4248 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4250 if (m->mdt_opts.mo_coordinator)
4251 mdt_hsm_cdt_stop(m);
4253 mdt_hsm_cdt_fini(m);
4255 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4256 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4257 obd_exports_barrier(obd);
4258 obd_zombie_barrier();
4262 tgt_fini(env, &m->mdt_lut);
4263 mdt_fs_cleanup(env, m);
4264 upcall_cache_cleanup(m->mdt_identity_cache);
4265 m->mdt_identity_cache = NULL;
4267 if (m->mdt_namespace != NULL) {
4268 ldlm_namespace_free(m->mdt_namespace, NULL,
4269 d->ld_obd->obd_force);
4270 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4273 mdt_quota_fini(env, m);
4275 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4277 mdt_seq_fini(env, m);
4278 mdt_fld_fini(env, m);
4283 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4285 LASSERT(atomic_read(&d->ld_ref) == 0);
4287 server_put_mount(mdt_obd_name(m), true);
4292 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4294 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4295 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4297 struct mdt_thread_info *info;
4298 struct obd_device *obd;
4299 const char *dev = lustre_cfg_string(cfg, 0);
4300 const char *num = lustre_cfg_string(cfg, 2);
4301 struct lustre_mount_info *lmi = NULL;
4302 struct lustre_sb_info *lsi;
4304 struct seq_server_site *ss_site;
4305 const char *identity_upcall = "NONE";
4306 struct md_device *next;
4312 lu_device_init(&m->mdt_lu_dev, ldt);
4314 * Environment (env) might be missing mdt_thread_key values at that
4315 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4318 * Usually device allocation path doesn't use module key values, but
4319 * mdt has to do a lot of work here, so allocate key value.
4321 rc = lu_env_refill((struct lu_env *)env);
4325 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4326 LASSERT(info != NULL);
4328 obd = class_name2obd(dev);
4329 LASSERT(obd != NULL);
4331 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4333 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4335 /* default is coordinator off, it is started through conf_param
4337 m->mdt_opts.mo_coordinator = 0;
4339 lmi = server_get_mount(dev);
4341 CERROR("Cannot get mount info for %s!\n", dev);
4344 lsi = s2lsi(lmi->lmi_sb);
4345 /* CMD is supported only in IAM mode */
4347 node_id = simple_strtol(num, NULL, 10);
4348 obd->u.obt.obt_magic = OBT_MAGIC;
4351 m->mdt_squash.rsi_uid = 0;
4352 m->mdt_squash.rsi_gid = 0;
4353 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4354 init_rwsem(&m->mdt_squash.rsi_sem);
4355 spin_lock_init(&m->mdt_osfs_lock);
4356 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4357 m->mdt_enable_remote_dir = 0;
4358 m->mdt_enable_remote_dir_gid = 0;
4360 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4361 m->mdt_lu_dev.ld_obd = obd;
4362 /* Set this lu_device to obd for error handling purposes. */
4363 obd->obd_lu_dev = &m->mdt_lu_dev;
4365 /* init the stack */
4366 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4368 CERROR("%s: Can't init device stack, rc %d\n",
4369 mdt_obd_name(m), rc);
4374 ss_site = mdt_seq_site(m);
4375 s->ld_seq_site = ss_site;
4378 /* set server index */
4379 ss_site->ss_node_id = node_id;
4381 /* failover is the default
4382 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4383 * assumed whose ss_node_id == 0 XXX
4385 obd->obd_replayable = 1;
4386 /* No connection accepted until configurations will finish */
4387 obd->obd_no_conn = 1;
4389 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4390 char *str = lustre_cfg_string(cfg, 4);
4391 if (strchr(str, 'n')) {
4392 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4393 obd->obd_replayable = 0;
4397 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4399 GOTO(err_fini_stack, rc);
4401 rc = mdt_seq_init(env, m);
4403 GOTO(err_fini_fld, rc);
4405 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4406 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4407 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4408 LDLM_NAMESPACE_SERVER,
4409 LDLM_NAMESPACE_GREEDY,
4411 if (m->mdt_namespace == NULL)
4412 GOTO(err_fini_seq, rc = -ENOMEM);
4414 m->mdt_namespace->ns_lvbp = m;
4415 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4417 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4418 /* set obd_namespace for compatibility with old code */
4419 obd->obd_namespace = m->mdt_namespace;
4421 rc = mdt_hsm_cdt_init(m);
4423 CERROR("%s: error initializing coordinator, rc %d\n",
4424 mdt_obd_name(m), rc);
4425 GOTO(err_free_ns, rc);
4428 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4429 OBD_FAIL_MDS_ALL_REQUEST_NET,
4430 OBD_FAIL_MDS_ALL_REPLY_NET);
4432 GOTO(err_free_hsm, rc);
4434 rc = mdt_fs_setup(env, m, obd, lsi);
4438 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4440 next = m->mdt_child;
4441 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4444 GOTO(err_fs_cleanup, rc);
4446 if (mntopts & MNTOPT_USERXATTR)
4447 m->mdt_opts.mo_user_xattr = 1;
4449 m->mdt_opts.mo_user_xattr = 0;
4451 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4453 GOTO(err_fs_cleanup, rc);
4455 if (mntopts & MNTOPT_ACL)
4456 m->mdt_opts.mo_acl = 1;
4458 m->mdt_opts.mo_acl = 0;
4460 /* XXX: to support suppgid for ACL, we enable identity_upcall
4461 * by default, otherwise, maybe got unexpected -EACCESS. */
4462 if (m->mdt_opts.mo_acl)
4463 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4465 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4467 &mdt_identity_upcall_cache_ops);
4468 if (IS_ERR(m->mdt_identity_cache)) {
4469 rc = PTR_ERR(m->mdt_identity_cache);
4470 m->mdt_identity_cache = NULL;
4471 GOTO(err_fs_cleanup, rc);
4474 rc = mdt_procfs_init(m, dev);
4476 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4477 GOTO(err_recovery, rc);
4480 rc = mdt_quota_init(env, m, cfg);
4482 GOTO(err_procfs, rc);
4484 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4485 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4486 "mdt_ldlm_client", m->mdt_ldlm_client);
4488 ping_evictor_start();
4490 /* recovery will be started upon mdt_prepare()
4491 * when the whole stack is complete and ready
4492 * to serve the requests */
4494 /* Reduce the initial timeout on an MDS because it doesn't need such
4495 * a long timeout as an OST does. Adaptive timeouts will adjust this
4496 * value appropriately. */
4497 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4498 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4504 target_recovery_fini(obd);
4505 upcall_cache_cleanup(m->mdt_identity_cache);
4506 m->mdt_identity_cache = NULL;
4508 mdt_fs_cleanup(env, m);
4510 tgt_fini(env, &m->mdt_lut);
4512 mdt_hsm_cdt_fini(m);
4514 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4515 obd->obd_namespace = m->mdt_namespace = NULL;
4517 mdt_seq_fini(env, m);
4519 mdt_fld_fini(env, m);
4521 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4524 server_put_mount(dev, true);
4528 /* For interoperability, the left element is old parameter, the right one
4529 * is the new version of the parameter, if some parameter is deprecated,
4530 * the new version should be set as NULL. */
4531 static struct cfg_interop_param mdt_interop_param[] = {
4532 { "mdt.group_upcall", NULL },
4533 { "mdt.quota_type", NULL },
4534 { "mdd.quota_type", NULL },
4535 { "mdt.rootsquash", "mdt.root_squash" },
4536 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4540 /* used by MGS to process specific configurations */
4541 static int mdt_process_config(const struct lu_env *env,
4542 struct lu_device *d, struct lustre_cfg *cfg)
4544 struct mdt_device *m = mdt_dev(d);
4545 struct md_device *md_next = m->mdt_child;
4546 struct lu_device *next = md2lu_dev(md_next);
4550 switch (cfg->lcfg_command) {
4552 struct obd_device *obd = d->ld_obd;
4554 /* For interoperability */
4555 struct cfg_interop_param *ptr = NULL;
4556 struct lustre_cfg *old_cfg = NULL;
4559 param = lustre_cfg_string(cfg, 1);
4560 if (param == NULL) {
4561 CERROR("param is empty\n");
4566 ptr = class_find_old_param(param, mdt_interop_param);
4568 if (ptr->new_param == NULL) {
4570 CWARN("For interoperability, skip this %s."
4571 " It is obsolete.\n", ptr->old_param);
4575 CWARN("Found old param %s, changed it to %s.\n",
4576 ptr->old_param, ptr->new_param);
4579 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4586 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4588 if (rc > 0 || rc == -ENOSYS) {
4589 /* is it an HSM var ? */
4590 rc = class_process_proc_param(PARAM_HSM,
4591 hsm_cdt_get_proc_vars(),
4593 if (rc > 0 || rc == -ENOSYS)
4594 /* we don't understand; pass it on */
4595 rc = next->ld_ops->ldo_process_config(env, next,
4599 if (old_cfg != NULL)
4600 lustre_cfg_free(cfg);
4605 /* others are passed further */
4606 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4612 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4613 const struct lu_object_header *hdr,
4614 struct lu_device *d)
4616 struct mdt_object *mo;
4620 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4622 struct lu_object *o;
4623 struct lu_object_header *h;
4626 h = &mo->mot_header;
4627 lu_object_header_init(h);
4628 lu_object_init(o, h, d);
4629 lu_object_add_top(h, o);
4630 o->lo_ops = &mdt_obj_ops;
4631 spin_lock_init(&mo->mot_write_lock);
4632 mutex_init(&mo->mot_lov_mutex);
4633 init_rwsem(&mo->mot_open_sem);
4639 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4640 const struct lu_object_conf *unused)
4642 struct mdt_device *d = mdt_dev(o->lo_dev);
4643 struct lu_device *under;
4644 struct lu_object *below;
4648 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4649 PFID(lu_object_fid(o)));
4651 under = &d->mdt_child->md_lu_dev;
4652 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4653 if (below != NULL) {
4654 lu_object_add(o, below);
4661 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4663 struct mdt_object *mo = mdt_obj(o);
4664 struct lu_object_header *h;
4668 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4669 PFID(lu_object_fid(o)));
4671 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4672 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4675 lu_object_header_fini(h);
4676 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4681 static int mdt_object_print(const struct lu_env *env, void *cookie,
4682 lu_printer_t p, const struct lu_object *o)
4684 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4686 return (*p)(env, cookie,
4687 LUSTRE_MDT_NAME"-object@%p(flags=%d, writecount=%d)",
4688 mdto, mdto->mot_flags, mdto->mot_write_count);
4691 static int mdt_prepare(const struct lu_env *env,
4692 struct lu_device *pdev,
4693 struct lu_device *cdev)
4695 struct mdt_device *mdt = mdt_dev(cdev);
4696 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4697 struct obd_device *obd = cdev->ld_obd;
4704 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4708 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4712 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4716 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4717 /* The LFSCK instance is registered just now, so it must be there when
4718 * register the namespace to such instance. */
4719 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4721 if (mdt->mdt_seq_site.ss_node_id == 0) {
4722 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4723 &mdt->mdt_md_root_fid);
4728 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4729 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4730 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4731 LASSERT(obd->obd_no_conn);
4732 spin_lock(&obd->obd_dev_lock);
4733 obd->obd_no_conn = 0;
4734 spin_unlock(&obd->obd_dev_lock);
4736 if (obd->obd_recovering == 0)
4737 mdt_postrecov(env, mdt);
4742 const struct lu_device_operations mdt_lu_ops = {
4743 .ldo_object_alloc = mdt_object_alloc,
4744 .ldo_process_config = mdt_process_config,
4745 .ldo_prepare = mdt_prepare,
4748 static const struct lu_object_operations mdt_obj_ops = {
4749 .loo_object_init = mdt_object_init,
4750 .loo_object_free = mdt_object_free,
4751 .loo_object_print = mdt_object_print
4754 static int mdt_obd_set_info_async(const struct lu_env *env,
4755 struct obd_export *exp,
4756 __u32 keylen, void *key,
4757 __u32 vallen, void *val,
4758 struct ptlrpc_request_set *set)
4764 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4765 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4773 * Match client and server connection feature flags.
4775 * Compute the compatibility flags for a connection request based on
4776 * features mutually supported by client and server.
4778 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4779 * must not be updated here, otherwise a partially initialized value may
4780 * be exposed. After the connection request is successfully processed,
4781 * the top-level MDT connect request handler atomically updates the export
4782 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4783 * reply. \see mdt_connect().
4785 * \param exp the obd_export associated with this client/target pair
4786 * \param mdt the target device for the connection
4787 * \param data stores data for this connect request
4790 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4791 * \retval -EBADE client and server feature requirements are incompatible
4793 static int mdt_connect_internal(struct obd_export *exp,
4794 struct mdt_device *mdt,
4795 struct obd_connect_data *data)
4797 LASSERT(data != NULL);
4799 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4800 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4802 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
4803 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
4804 CWARN("%s: client %s does not support ibits lock, either "
4805 "very old or an invalid client: flags "LPX64"\n",
4806 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
4807 data->ocd_connect_flags);
4811 if (!mdt->mdt_opts.mo_acl)
4812 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4814 if (!mdt->mdt_opts.mo_user_xattr)
4815 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4817 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4818 data->ocd_brw_size = min(data->ocd_brw_size,
4819 (__u32)MD_MAX_BRW_SIZE);
4820 if (data->ocd_brw_size == 0) {
4821 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4822 " ocd_version: %x ocd_grant: %d "
4823 "ocd_index: %u ocd_brw_size is "
4824 "unexpectedly zero, network data "
4825 "corruption? Refusing connection of this"
4828 exp->exp_client_uuid.uuid,
4829 exp, data->ocd_connect_flags, data->ocd_version,
4830 data->ocd_grant, data->ocd_index);
4835 /* NB: Disregard the rule against updating
4836 * exp_connect_data.ocd_connect_flags in this case, since
4837 * tgt_client_new() needs to know if this is a lightweight
4838 * connection, and it is safe to expose this flag before
4839 * connection processing completes. */
4840 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4841 spin_lock(&exp->exp_lock);
4842 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4843 spin_unlock(&exp->exp_lock);
4846 data->ocd_version = LUSTRE_VERSION_CODE;
4848 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4849 CWARN("%s: MDS requires FID support, but client not\n",
4854 if (OCD_HAS_FLAG(data, PINGLESS)) {
4855 if (ptlrpc_pinger_suppress_pings()) {
4856 spin_lock(&exp->exp_obd->obd_dev_lock);
4857 list_del_init(&exp->exp_obd_chain_timed);
4858 spin_unlock(&exp->exp_obd->obd_dev_lock);
4860 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4864 data->ocd_max_easize = mdt->mdt_max_ea_size;
4869 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
4870 struct mdt_thread_info *info,
4871 struct mdt_file_data *mfd)
4873 struct lu_context ses;
4877 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
4882 lu_context_enter(&ses);
4884 mdt_ucred(info)->uc_valid = UCRED_OLD;
4885 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
4887 lu_context_exit(&ses);
4888 lu_context_fini(&ses);
4894 static int mdt_export_cleanup(struct obd_export *exp)
4896 struct list_head closing_list;
4897 struct mdt_export_data *med = &exp->exp_mdt_data;
4898 struct obd_device *obd = exp->exp_obd;
4899 struct mdt_device *mdt;
4900 struct mdt_thread_info *info;
4902 struct mdt_file_data *mfd, *n;
4906 INIT_LIST_HEAD(&closing_list);
4907 spin_lock(&med->med_open_lock);
4908 while (!list_empty(&med->med_open_head)) {
4909 struct list_head *tmp = med->med_open_head.next;
4910 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
4912 /* Remove mfd handle so it can't be found again.
4913 * We are consuming the mfd_list reference here. */
4914 class_handle_unhash(&mfd->mfd_handle);
4915 list_move_tail(&mfd->mfd_list, &closing_list);
4917 spin_unlock(&med->med_open_lock);
4918 mdt = mdt_dev(obd->obd_lu_dev);
4919 LASSERT(mdt != NULL);
4921 rc = lu_env_init(&env, LCT_MD_THREAD);
4925 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
4926 LASSERT(info != NULL);
4927 memset(info, 0, sizeof *info);
4928 info->mti_env = &env;
4929 info->mti_mdt = mdt;
4930 info->mti_exp = exp;
4932 if (!list_empty(&closing_list)) {
4933 struct md_attr *ma = &info->mti_attr;
4935 /* Close any open files (which may also cause orphan
4937 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
4938 list_del_init(&mfd->mfd_list);
4939 ma->ma_need = ma->ma_valid = 0;
4941 /* This file is being closed due to an eviction, it
4942 * could have been modified and now dirty regarding to
4943 * HSM archive, check this!
4944 * The logic here is to mark a file dirty if there's a
4945 * chance it was dirtied before the client was evicted,
4946 * so that we don't have to wait for a release attempt
4947 * before finding out the file was actually dirty and
4948 * fail the release. Aggressively marking it dirty here
4949 * will cause the policy engine to attempt to
4950 * re-archive it; when rearchiving, we can compare the
4951 * current version to the HSM data_version and make the
4952 * archive request into a noop if it's not actually
4955 if (mfd->mfd_mode & FMODE_WRITE)
4956 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
4958 /* Don't unlink orphan on failover umount, LU-184 */
4959 if (exp->exp_flags & OBD_OPT_FAILOVER) {
4960 ma->ma_valid = MA_FLAGS;
4961 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
4963 mdt_mfd_close(info, mfd);
4966 info->mti_mdt = NULL;
4967 /* cleanup client slot early */
4968 /* Do not erase record for recoverable client. */
4969 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
4970 tgt_client_del(&env, exp);
4976 static int mdt_obd_disconnect(struct obd_export *exp)
4982 class_export_get(exp);
4984 nodemap_del_member(exp);
4985 rc = server_disconnect_export(exp);
4987 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
4989 rc = mdt_export_cleanup(exp);
4990 class_export_put(exp);
4994 /* mds_connect copy */
4995 static int mdt_obd_connect(const struct lu_env *env,
4996 struct obd_export **exp, struct obd_device *obd,
4997 struct obd_uuid *cluuid,
4998 struct obd_connect_data *data,
5001 struct obd_export *lexp;
5002 struct lustre_handle conn = { 0 };
5003 struct mdt_device *mdt;
5005 lnet_nid_t *client_nid = localdata;
5008 LASSERT(env != NULL);
5009 if (!exp || !obd || !cluuid)
5012 mdt = mdt_dev(obd->obd_lu_dev);
5015 * first, check whether the stack is ready to handle requests
5016 * XXX: probably not very appropriate method is used now
5017 * at some point we should find a better one
5019 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
5020 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
5021 rc = obd_get_info(env, mdt->mdt_child_exp,
5022 sizeof(KEY_OSP_CONNECTED),
5023 KEY_OSP_CONNECTED, NULL, NULL, NULL);
5026 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5029 rc = class_connect(&conn, obd, cluuid);
5033 lexp = class_conn2export(&conn);
5034 LASSERT(lexp != NULL);
5036 rc = mdt_connect_internal(lexp, mdt, data);
5038 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5041 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5042 rc = tgt_client_new(env, lexp);
5044 rc = nodemap_add_member(*client_nid, lexp);
5045 if (rc != 0 && rc != -EEXIST)
5048 mdt_export_stats_init(obd, lexp, localdata);
5051 /* For phase I, sync for cross-ref operation. */
5052 spin_lock(&lexp->exp_lock);
5053 lexp->exp_keep_sync = 1;
5054 spin_unlock(&lexp->exp_lock);
5058 class_disconnect(lexp);
5067 static int mdt_obd_reconnect(const struct lu_env *env,
5068 struct obd_export *exp, struct obd_device *obd,
5069 struct obd_uuid *cluuid,
5070 struct obd_connect_data *data,
5073 lnet_nid_t *client_nid = localdata;
5077 if (exp == NULL || obd == NULL || cluuid == NULL)
5080 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5082 rc = nodemap_add_member(*client_nid, exp);
5083 if (rc == 0 || rc == -EEXIST)
5084 mdt_export_stats_init(obd, exp, localdata);
5090 /* FIXME: Can we avoid using these two interfaces? */
5091 static int mdt_init_export(struct obd_export *exp)
5093 struct mdt_export_data *med = &exp->exp_mdt_data;
5097 INIT_LIST_HEAD(&med->med_open_head);
5098 spin_lock_init(&med->med_open_lock);
5099 mutex_init(&med->med_idmap_mutex);
5100 med->med_idmap = NULL;
5101 spin_lock(&exp->exp_lock);
5102 exp->exp_connecting = 1;
5103 spin_unlock(&exp->exp_lock);
5105 /* self-export doesn't need client data and ldlm initialization */
5106 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5107 &exp->exp_client_uuid)))
5110 rc = tgt_client_alloc(exp);
5114 rc = ldlm_init_export(exp);
5121 tgt_client_free(exp);
5123 CERROR("%s: Failed to initialize export: rc = %d\n",
5124 exp->exp_obd->obd_name, rc);
5128 static int mdt_destroy_export(struct obd_export *exp)
5132 if (exp_connect_rmtclient(exp))
5133 mdt_cleanup_idmap(&exp->exp_mdt_data);
5135 target_destroy_export(exp);
5136 /* destroy can be called from failed obd_setup, so
5137 * checking uuid is safer than obd_self_export */
5138 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5139 &exp->exp_client_uuid)))
5142 ldlm_destroy_export(exp);
5143 tgt_client_free(exp);
5145 LASSERT(list_empty(&exp->exp_outstanding_replies));
5146 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5151 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5152 struct linkea_data *ldata)
5156 LASSERT(ldata->ld_buf->lb_buf != NULL);
5158 if (!mdt_object_exists(mdt_obj))
5161 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5162 ldata->ld_buf, XATTR_NAME_LINK);
5163 if (rc == -ERANGE) {
5164 /* Buf was too small, figure out what we need. */
5165 lu_buf_free(ldata->ld_buf);
5166 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5167 ldata->ld_buf, XATTR_NAME_LINK);
5170 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5171 if (ldata->ld_buf->lb_buf == NULL)
5173 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5174 ldata->ld_buf, XATTR_NAME_LINK);
5179 return linkea_init(ldata);
5183 * Given an MDT object, try to look up the full path to the object.
5184 * Part of the MDT layer implementation of lfs fid2path.
5186 * \param[in] info Per-thread common data shared by MDT level handlers.
5187 * \param[in] obj Object to do path lookup of
5188 * \param[in,out] fp User-provided struct to store path information
5190 * \retval 0 Lookup successful, path information stored in fp
5191 * \retval -EAGAIN Lookup failed, usually because object is being moved
5192 * \retval negative errno if there was a problem
5194 static int mdt_path_current(struct mdt_thread_info *info,
5195 struct mdt_object *obj,
5196 struct getinfo_fid2path *fp)
5198 struct mdt_device *mdt = info->mti_mdt;
5199 struct mdt_object *mdt_obj;
5200 struct link_ea_header *leh;
5201 struct link_ea_entry *lee;
5202 struct lu_name *tmpname = &info->mti_name;
5203 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5204 struct lu_buf *buf = &info->mti_big_buf;
5207 struct linkea_data ldata = { NULL };
5212 /* temp buffer for path element, the buffer will be finally freed
5213 * in mdt_thread_info_fini */
5214 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5215 if (buf->lb_buf == NULL)
5219 ptr = fp->gf_path + fp->gf_pathlen - 1;
5222 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5224 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5225 * so checking root_fid can only happen on MDT0. */
5226 while (!lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid)) {
5227 struct lu_buf lmv_buf;
5229 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5230 if (IS_ERR(mdt_obj))
5231 GOTO(out, rc = PTR_ERR(mdt_obj));
5233 if (!mdt_object_exists(mdt_obj)) {
5234 mdt_object_put(info->mti_env, mdt_obj);
5235 GOTO(out, rc = -ENOENT);
5238 if (mdt_object_remote(mdt_obj)) {
5239 mdt_object_put(info->mti_env, mdt_obj);
5240 GOTO(remote_out, rc = -EREMOTE);
5243 rc = mdt_links_read(info, mdt_obj, &ldata);
5245 mdt_object_put(info->mti_env, mdt_obj);
5250 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5251 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5252 /* If set, use link #linkno for path lookup, otherwise use
5253 link #0. Only do this for the final path element. */
5254 if (first && fp->gf_linkno < leh->leh_reccount) {
5256 for (count = 0; count < fp->gf_linkno; count++) {
5257 lee = (struct link_ea_entry *)
5258 ((char *)lee + reclen);
5259 linkea_entry_unpack(lee, &reclen, tmpname,
5262 if (fp->gf_linkno < leh->leh_reccount - 1)
5263 /* indicate to user there are more links */
5267 lmv_buf.lb_buf = info->mti_xattr_buf;
5268 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5269 /* Check if it is slave stripes */
5270 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5271 &lmv_buf, XATTR_NAME_LMV);
5272 mdt_object_put(info->mti_env, mdt_obj);
5274 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5276 /* For slave stripes, get its master */
5277 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5278 fp->gf_fid = *tmpfid;
5281 } else if (rc < 0 && rc != -ENODATA) {
5287 /* Pack the name in the end of the buffer */
5288 ptr -= tmpname->ln_namelen;
5289 if (ptr - 1 <= fp->gf_path)
5290 GOTO(out, rc = -EOVERFLOW);
5291 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5294 /* keep the last resolved fid to the client, so the
5295 * client will build the left path on another MDT for
5297 fp->gf_fid = *tmpfid;
5303 ptr++; /* skip leading / */
5304 memmove(fp->gf_path, ptr, fp->gf_path + fp->gf_pathlen - ptr);
5311 * Given an MDT object, use mdt_path_current to get the path.
5312 * Essentially a wrapper to retry mdt_path_current a set number of times
5313 * if -EAGAIN is returned (usually because an object is being moved).
5315 * Part of the MDT layer implementation of lfs fid2path.
5317 * \param[in] info Per-thread common data shared by mdt level handlers.
5318 * \param[in] obj Object to do path lookup of
5319 * \param[in,out] fp User-provided struct for arguments and to store path
5322 * \retval 0 Lookup successful, path information stored in fp
5323 * \retval negative errno if there was a problem
5325 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5326 struct getinfo_fid2path *fp)
5328 struct mdt_device *mdt = info->mti_mdt;
5333 if (fp->gf_pathlen < 3)
5336 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5337 fp->gf_path[0] = '\0';
5341 /* Retry multiple times in case file is being moved */
5342 while (tries-- && rc == -EAGAIN)
5343 rc = mdt_path_current(info, obj, fp);
5349 * Get the full path of the provided FID, as of changelog record recno.
5351 * This checks sanity and looks up object for user provided FID
5352 * before calling the actual path lookup code.
5354 * Part of the MDT layer implementation of lfs fid2path.
5356 * \param[in] info Per-thread common data shared by mdt level handlers.
5357 * \param[in,out] fp User-provided struct for arguments and to store path
5360 * \retval 0 Lookup successful, path information and recno stored in fp
5361 * \retval -ENOENT, object does not exist
5362 * \retval negative errno if there was a problem
5364 static int mdt_fid2path(struct mdt_thread_info *info,
5365 struct getinfo_fid2path *fp)
5367 struct mdt_device *mdt = info->mti_mdt;
5368 struct mdt_object *obj;
5372 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5373 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5375 if (!fid_is_sane(&fp->gf_fid))
5378 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5379 CWARN("%s: "DFID" is invalid, sequence should be "
5380 ">= "LPX64"\n", mdt_obd_name(mdt),
5381 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5385 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5388 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5389 PFID(&fp->gf_fid), rc);
5393 if (mdt_object_remote(obj))
5395 else if (!mdt_object_exists(obj))
5401 mdt_object_put(info->mti_env, obj);
5402 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5403 PFID(&fp->gf_fid), rc);
5407 rc = mdt_path(info, obj, fp);
5409 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5410 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5412 mdt_object_put(info->mti_env, obj);
5417 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5418 void *val, int vallen)
5420 struct getinfo_fid2path *fpout, *fpin;
5423 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5426 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5427 lustre_swab_fid2path(fpin);
5429 memcpy(fpout, fpin, sizeof(*fpin));
5430 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5433 rc = mdt_fid2path(info, fpout);
5437 int mdt_get_info(struct tgt_session_info *tsi)
5447 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5449 CDEBUG(D_IOCTL, "No GETINFO key\n");
5450 RETURN(err_serious(-EFAULT));
5452 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5455 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5456 if (vallen == NULL) {
5457 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5458 tgt_name(tsi->tsi_tgt));
5459 RETURN(err_serious(-EFAULT));
5462 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5464 rc = req_capsule_server_pack(tsi->tsi_pill);
5466 RETURN(err_serious(rc));
5468 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5469 if (valout == NULL) {
5470 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5471 tgt_name(tsi->tsi_tgt));
5472 RETURN(err_serious(-EFAULT));
5475 if (KEY_IS(KEY_FID2PATH)) {
5476 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5478 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5479 mdt_thread_info_fini(info);
5486 /* Pass the ioc down */
5487 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5488 unsigned int cmd, int len, void *data)
5490 struct lu_context ioctl_session;
5491 struct md_device *next = mdt->mdt_child;
5495 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5498 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5499 lu_context_enter(&ioctl_session);
5500 env->le_ses = &ioctl_session;
5502 LASSERT(next->md_ops->mdo_iocontrol);
5503 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5505 lu_context_exit(&ioctl_session);
5506 lu_context_fini(&ioctl_session);
5510 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5512 struct obd_ioctl_data *data = karg;
5515 struct mdt_object *obj;
5516 struct mdt_lock_handle *lh;
5520 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5521 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5524 fid = (struct lu_fid *)data->ioc_inlbuf1;
5526 if (!fid_is_sane(fid))
5529 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5531 lh = &mti->mti_lh[MDT_LH_PARENT];
5532 mdt_lock_reg_init(lh, LCK_CR);
5534 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5536 RETURN(PTR_ERR(obj));
5538 if (mdt_object_remote(obj)) {
5541 * before calling version get the correct MDS should be
5542 * fid, this is error to find remote object here
5544 CERROR("nonlocal object "DFID"\n", PFID(fid));
5545 } else if (!mdt_object_exists(obj)) {
5546 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5549 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5550 *(__u64 *)data->ioc_inlbuf2 = version;
5553 mdt_object_unlock_put(mti, obj, lh, 1);
5557 /* ioctls on obd dev */
5558 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5559 void *karg, void *uarg)
5562 struct obd_device *obd = exp->exp_obd;
5563 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5564 struct dt_device *dt = mdt->mdt_bottom;
5568 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5569 rc = lu_env_init(&env, LCT_MD_THREAD);
5575 rc = mdt_device_sync(&env, mdt);
5577 case OBD_IOC_SET_READONLY:
5578 rc = dt->dd_ops->dt_ro(&env, dt);
5580 case OBD_IOC_ABORT_RECOVERY:
5581 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5582 target_stop_recovery_thread(obd);
5585 case OBD_IOC_CHANGELOG_REG:
5586 case OBD_IOC_CHANGELOG_DEREG:
5587 case OBD_IOC_CHANGELOG_CLEAR:
5588 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5590 case OBD_IOC_START_LFSCK: {
5591 struct md_device *next = mdt->mdt_child;
5592 struct obd_ioctl_data *data = karg;
5593 struct lfsck_start_param lsp;
5595 if (unlikely(data == NULL)) {
5600 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5601 lsp.lsp_index_valid = 0;
5602 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5605 case OBD_IOC_STOP_LFSCK: {
5606 struct md_device *next = mdt->mdt_child;
5607 struct obd_ioctl_data *data = karg;
5608 struct lfsck_stop stop;
5610 stop.ls_status = LS_STOPPED;
5611 /* Old lfsck utils may pass NULL @stop. */
5612 if (data->ioc_inlbuf1 == NULL)
5616 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5618 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5621 case OBD_IOC_GET_OBJ_VERSION: {
5622 struct mdt_thread_info *mti;
5623 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5624 memset(mti, 0, sizeof *mti);
5625 mti->mti_env = &env;
5629 rc = mdt_ioc_version_get(mti, karg);
5632 case OBD_IOC_CATLOGLIST: {
5633 struct mdt_thread_info *mti;
5635 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5636 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5637 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5638 &mti->mti_tmp_fid1);
5643 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5644 mdt_obd_name(mdt), cmd, rc);
5651 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5653 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5654 struct lfsck_start_param lsp;
5658 lsp.lsp_start = NULL;
5659 lsp.lsp_index_valid = 0;
5660 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
5661 OBD_IOC_START_LFSCK,
5663 if (rc != 0 && rc != -EALREADY)
5664 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
5665 mdt_obd_name(mdt), rc);
5667 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5671 static int mdt_obd_postrecov(struct obd_device *obd)
5676 rc = lu_env_init(&env, LCT_MD_THREAD);
5679 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5684 static struct obd_ops mdt_obd_device_ops = {
5685 .o_owner = THIS_MODULE,
5686 .o_set_info_async = mdt_obd_set_info_async,
5687 .o_connect = mdt_obd_connect,
5688 .o_reconnect = mdt_obd_reconnect,
5689 .o_disconnect = mdt_obd_disconnect,
5690 .o_init_export = mdt_init_export,
5691 .o_destroy_export = mdt_destroy_export,
5692 .o_iocontrol = mdt_iocontrol,
5693 .o_postrecov = mdt_obd_postrecov,
5696 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5697 struct lu_device *d)
5699 struct mdt_device *m = mdt_dev(d);
5706 static struct lu_device *mdt_device_free(const struct lu_env *env,
5707 struct lu_device *d)
5709 struct mdt_device *m = mdt_dev(d);
5712 lu_device_fini(&m->mdt_lu_dev);
5718 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5719 struct lu_device_type *t,
5720 struct lustre_cfg *cfg)
5722 struct lu_device *l;
5723 struct mdt_device *m;
5730 rc = mdt_init0(env, m, t, cfg);
5732 mdt_device_free(env, l);
5737 l = ERR_PTR(-ENOMEM);
5741 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5742 LU_KEY_INIT(mdt, struct mdt_thread_info);
5744 static void mdt_key_fini(const struct lu_context *ctx,
5745 struct lu_context_key *key, void* data)
5747 struct mdt_thread_info *info = data;
5749 if (info->mti_big_lmm) {
5750 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5751 info->mti_big_lmm = NULL;
5752 info->mti_big_lmmsize = 0;
5757 /* context key: mdt_thread_key */
5758 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5760 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5762 return lu_ucred(info->mti_env);
5765 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5767 return lu_ucred_check(info->mti_env);
5771 * Enable/disable COS (Commit On Sharing).
5773 * Set/Clear the COS flag in mdt options.
5775 * \param mdt mdt device
5776 * \param val 0 disables COS, other values enable COS
5778 void mdt_enable_cos(struct mdt_device *mdt, int val)
5783 mdt->mdt_opts.mo_cos = !!val;
5784 rc = lu_env_init(&env, LCT_LOCAL);
5785 if (unlikely(rc != 0)) {
5786 CWARN("%s: lu_env initialization failed, cannot "
5787 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
5790 mdt_device_sync(&env, mdt);
5795 * Check COS (Commit On Sharing) status.
5797 * Return COS flag status.
5799 * \param mdt mdt device
5801 int mdt_cos_is_enabled(struct mdt_device *mdt)
5803 return mdt->mdt_opts.mo_cos != 0;
5806 static struct lu_device_type_operations mdt_device_type_ops = {
5807 .ldto_device_alloc = mdt_device_alloc,
5808 .ldto_device_free = mdt_device_free,
5809 .ldto_device_fini = mdt_device_fini
5812 static struct lu_device_type mdt_device_type = {
5813 .ldt_tags = LU_DEVICE_MD,
5814 .ldt_name = LUSTRE_MDT_NAME,
5815 .ldt_ops = &mdt_device_type_ops,
5816 .ldt_ctx_tags = LCT_MD_THREAD
5819 static int __init mdt_mod_init(void)
5823 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5824 FID_NOBRACE_LEN + 1);
5825 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5827 rc = lu_kmem_init(mdt_caches);
5831 rc = mds_mod_init();
5835 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
5836 LUSTRE_MDT_NAME, &mdt_device_type);
5841 lu_kmem_fini(mdt_caches);
5848 static void __exit mdt_mod_exit(void)
5850 class_unregister_type(LUSTRE_MDT_NAME);
5852 lu_kmem_fini(mdt_caches);
5855 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
5856 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5857 MODULE_LICENSE("GPL");
5859 cfs_module(mdt, LUSTRE_VERSION_STRING, mdt_mod_init, mdt_mod_exit);