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) {
1936 /* master quotactl */
1939 if (!nodemap_can_setquota(nodemap))
1944 RETURN(-EOPNOTSUPP);
1945 /* slave quotactl */
1950 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1954 /* map uid/gid for remote client */
1956 if (exp_connect_rmtclient(exp)) {
1957 struct lustre_idmap_table *idmap;
1959 idmap = exp->exp_mdt_data.med_idmap;
1961 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1962 oqctl->qc_cmd != Q_GETINFO))
1965 if (oqctl->qc_type == USRQUOTA)
1966 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1968 else if (oqctl->qc_type == GRPQUOTA)
1969 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1974 if (id == CFS_IDMAP_NOTFOUND) {
1975 CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
1980 if (oqctl->qc_type == USRQUOTA)
1981 id = nodemap_map_id(nodemap, NODEMAP_UID,
1982 NODEMAP_CLIENT_TO_FS, id);
1983 else if (oqctl->qc_type == GRPQUOTA)
1984 id = nodemap_map_id(nodemap, NODEMAP_UID,
1985 NODEMAP_CLIENT_TO_FS, id);
1987 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1989 RETURN(err_serious(-EFAULT));
1991 if (oqctl->qc_id != id)
1992 swap(oqctl->qc_id, id);
1994 switch (oqctl->qc_cmd) {
2000 /* forward quotactl request to QMT */
2001 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2006 /* slave quotactl */
2007 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2012 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2016 if (oqctl->qc_id != id)
2017 swap(oqctl->qc_id, id);
2023 /** clone llog ctxt from child (mdd)
2024 * This allows remote llog (replicator) access.
2025 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2026 * context was originally set up, or we can handle them directly.
2027 * I choose the latter, but that means I need any llog
2028 * contexts set up by child to be accessable by the mdt. So we clone the
2029 * context into our context list here.
2031 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2034 struct md_device *next = mdt->mdt_child;
2035 struct llog_ctxt *ctxt;
2038 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2041 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2042 if (rc || ctxt == NULL) {
2046 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2048 CERROR("Can't set mdt ctxt %d\n", rc);
2053 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2054 struct mdt_device *mdt, int idx)
2056 struct llog_ctxt *ctxt;
2058 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2061 /* Put once for the get we just did, and once for the clone */
2062 llog_ctxt_put(ctxt);
2063 llog_ctxt_put(ctxt);
2068 * sec context handlers
2070 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2074 rc = mdt_handle_idmap(tsi);
2076 struct ptlrpc_request *req = tgt_ses_req(tsi);
2079 opc = lustre_msg_get_opc(req->rq_reqmsg);
2080 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2081 sptlrpc_svc_ctx_invalidate(req);
2084 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2090 * quota request handlers
2092 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2094 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2095 struct lu_device *qmt = mdt->mdt_qmt_dev;
2100 RETURN(err_serious(-EOPNOTSUPP));
2102 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2106 struct mdt_object *mdt_object_new(const struct lu_env *env,
2107 struct mdt_device *d,
2108 const struct lu_fid *f)
2110 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2111 struct lu_object *o;
2112 struct mdt_object *m;
2115 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2116 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2117 if (unlikely(IS_ERR(o)))
2118 m = (struct mdt_object *)o;
2124 struct mdt_object *mdt_object_find(const struct lu_env *env,
2125 struct mdt_device *d,
2126 const struct lu_fid *f)
2128 struct lu_object *o;
2129 struct mdt_object *m;
2132 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2133 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2134 if (unlikely(IS_ERR(o)))
2135 m = (struct mdt_object *)o;
2143 * Asyncronous commit for mdt device.
2145 * Pass asynchonous commit call down the MDS stack.
2147 * \param env environment
2148 * \param mdt the mdt device
2150 static void mdt_device_commit_async(const struct lu_env *env,
2151 struct mdt_device *mdt)
2153 struct dt_device *dt = mdt->mdt_bottom;
2156 rc = dt->dd_ops->dt_commit_async(env, dt);
2157 if (unlikely(rc != 0))
2158 CWARN("%s: async commit start failed: rc = %d\n",
2159 mdt_obd_name(mdt), rc);
2163 * Mark the lock as "synchonous".
2165 * Mark the lock to deffer transaction commit to the unlock time.
2167 * \param lock the lock to mark as "synchonous"
2169 * \see mdt_is_lock_sync
2170 * \see mdt_save_lock
2172 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2174 lock->l_ast_data = (void*)1;
2178 * Check whehter the lock "synchonous" or not.
2180 * \param lock the lock to check
2181 * \retval 1 the lock is "synchonous"
2182 * \retval 0 the lock isn't "synchronous"
2184 * \see mdt_set_lock_sync
2185 * \see mdt_save_lock
2187 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2189 return lock->l_ast_data != NULL;
2193 * Blocking AST for mdt locks.
2195 * Starts transaction commit if in case of COS lock conflict or
2196 * deffers such a commit to the mdt_save_lock.
2198 * \param lock the lock which blocks a request or cancelling lock
2199 * \param desc unused
2200 * \param data unused
2201 * \param flag indicates whether this cancelling or blocking callback
2203 * \see ldlm_blocking_ast_nocheck
2205 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2206 void *data, int flag)
2208 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2209 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2213 if (flag == LDLM_CB_CANCELING)
2215 lock_res_and_lock(lock);
2216 if (lock->l_blocking_ast != mdt_blocking_ast) {
2217 unlock_res_and_lock(lock);
2220 if (mdt_cos_is_enabled(mdt) &&
2221 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2222 lock->l_blocking_lock != NULL &&
2223 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2224 mdt_set_lock_sync(lock);
2226 rc = ldlm_blocking_ast_nocheck(lock);
2228 /* There is no lock conflict if l_blocking_lock == NULL,
2229 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2230 * when the last reference to a local lock was released */
2231 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2234 rc = lu_env_init(&env, LCT_LOCAL);
2235 if (unlikely(rc != 0))
2236 CWARN("%s: lu_env initialization failed, cannot "
2237 "start asynchronous commit: rc = %d\n",
2240 mdt_device_commit_async(&env, mdt);
2246 /* Used for cross-MDT lock */
2247 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2248 void *data, int flag)
2250 struct lustre_handle lockh;
2254 case LDLM_CB_BLOCKING:
2255 ldlm_lock2handle(lock, &lockh);
2256 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
2258 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2262 case LDLM_CB_CANCELING:
2263 LDLM_DEBUG(lock, "Revoke remote lock\n");
2271 int mdt_check_resent_lock(struct mdt_thread_info *info,
2272 struct mdt_object *mo,
2273 struct mdt_lock_handle *lhc)
2275 /* the lock might already be gotten in ldlm_handle_enqueue() */
2276 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2277 struct ptlrpc_request *req = mdt_info_req(info);
2278 struct ldlm_lock *lock;
2280 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2281 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2283 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2284 * a resend case, however, it could be already destroyed
2285 * due to client eviction or a raced cancel RPC. */
2286 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
2287 lhc->mlh_reg_lh.cookie);
2291 if (!fid_res_name_eq(mdt_object_fid(mo),
2292 &lock->l_resource->lr_name)) {
2293 CWARN("%s: Although resent, but still not "
2294 "get child lock:"DFID"\n",
2295 info->mti_exp->exp_obd->obd_name,
2296 PFID(mdt_object_fid(mo)));
2297 LDLM_LOCK_PUT(lock);
2300 LDLM_LOCK_PUT(lock);
2306 int mdt_remote_object_lock(struct mdt_thread_info *mti,
2307 struct mdt_object *o, const struct lu_fid *fid,
2308 struct lustre_handle *lh, ldlm_mode_t mode,
2311 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2312 ldlm_policy_data_t *policy = &mti->mti_policy;
2313 struct ldlm_res_id *res_id = &mti->mti_res_id;
2317 LASSERT(mdt_object_remote(o));
2319 LASSERT(ibits == MDS_INODELOCK_UPDATE);
2321 fid_build_reg_res_name(fid, res_id);
2323 memset(einfo, 0, sizeof(*einfo));
2324 einfo->ei_type = LDLM_IBITS;
2325 einfo->ei_mode = mode;
2326 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2327 einfo->ei_cb_cp = ldlm_completion_ast;
2328 einfo->ei_enq_slave = 0;
2329 einfo->ei_res_id = res_id;
2331 memset(policy, 0, sizeof(*policy));
2332 policy->l_inodebits.bits = ibits;
2334 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2339 static int mdt_object_local_lock(struct mdt_thread_info *info,
2340 struct mdt_object *o,
2341 struct mdt_lock_handle *lh, __u64 ibits,
2342 bool nonblock, int locality)
2344 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2345 ldlm_policy_data_t *policy = &info->mti_policy;
2346 struct ldlm_res_id *res_id = &info->mti_res_id;
2351 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2352 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2353 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2354 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2356 /* Only enqueue LOOKUP lock for remote object */
2357 if (mdt_object_remote(o))
2358 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2360 if (lh->mlh_type == MDT_PDO_LOCK) {
2361 /* check for exists after object is locked */
2362 if (mdt_object_exists(o) == 0) {
2363 /* Non-existent object shouldn't have PDO lock */
2366 /* Non-dir object shouldn't have PDO lock */
2367 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2372 memset(policy, 0, sizeof(*policy));
2373 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2375 dlmflags = LDLM_FL_ATOMIC_CB;
2377 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2380 * Take PDO lock on whole directory and build correct @res_id for lock
2381 * on part of directory.
2383 if (lh->mlh_pdo_hash != 0) {
2384 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2385 mdt_lock_pdo_mode(info, o, lh);
2386 if (lh->mlh_pdo_mode != LCK_NL) {
2388 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2389 * is never going to be sent to client and we do not
2390 * want it slowed down due to possible cancels.
2392 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2393 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2394 policy, res_id, dlmflags,
2395 info->mti_exp == NULL ? NULL :
2396 &info->mti_exp->exp_handle.h_cookie);
2397 if (unlikely(rc != 0))
2398 GOTO(out_unlock, rc);
2402 * Finish res_id initializing by name hash marking part of
2403 * directory which is taking modification.
2405 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2408 policy->l_inodebits.bits = ibits;
2411 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2412 * going to be sent to client. If it is - mdt_intent_policy() path will
2413 * fix it up and turn FL_LOCAL flag off.
2415 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2416 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2417 info->mti_exp == NULL ? NULL :
2418 &info->mti_exp->exp_handle.h_cookie);
2421 mdt_object_unlock(info, o, lh, 1);
2422 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2423 lh->mlh_pdo_hash != 0 &&
2424 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2425 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2431 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2432 struct mdt_lock_handle *lh, __u64 ibits,
2433 bool nonblock, int locality)
2438 if (!mdt_object_remote(o))
2439 return mdt_object_local_lock(info, o, lh, ibits, nonblock,
2442 if (locality == MDT_LOCAL_LOCK) {
2443 CERROR("%s: try to get local lock for remote object"
2444 DFID".\n", mdt_obd_name(info->mti_mdt),
2445 PFID(mdt_object_fid(o)));
2449 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2450 ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2451 MDS_INODELOCK_XATTR);
2452 if (ibits & MDS_INODELOCK_UPDATE) {
2453 /* Sigh, PDO needs to enqueue 2 locks right now, but
2454 * enqueue RPC can only request 1 lock, to avoid extra
2455 * RPC, so it will instead enqueue EX lock for remote
2456 * object anyway XXX*/
2457 if (lh->mlh_type == MDT_PDO_LOCK &&
2458 lh->mlh_pdo_hash != 0) {
2459 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2460 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2461 PFID(mdt_object_fid(o)));
2462 lh->mlh_pdo_hash = 0;
2463 lh->mlh_rreg_mode = LCK_EX;
2464 lh->mlh_type = MDT_REG_LOCK;
2466 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2469 MDS_INODELOCK_UPDATE);
2474 /* Only enqueue LOOKUP lock for remote object */
2475 if (ibits & MDS_INODELOCK_LOOKUP) {
2476 rc = mdt_object_local_lock(info, o, lh,
2477 MDS_INODELOCK_LOOKUP,
2478 nonblock, locality);
2486 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2487 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2489 return mdt_object_lock_internal(info, o, lh, ibits, false, locality);
2492 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2493 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2495 struct mdt_lock_handle tmp = *lh;
2498 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true, locality);
2506 * Save a lock within request object.
2508 * Keep the lock referenced until whether client ACK or transaction
2509 * commit happens or release the lock immediately depending on input
2510 * parameters. If COS is ON, a write lock is converted to COS lock
2513 * \param info thead info object
2514 * \param h lock handle
2515 * \param mode lock mode
2516 * \param decref force immediate lock releasing
2519 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2520 ldlm_mode_t mode, int decref)
2524 if (lustre_handle_is_used(h)) {
2525 if (decref || !info->mti_has_trans ||
2526 !(mode & (LCK_PW | LCK_EX))){
2527 mdt_fid_unlock(h, mode);
2529 struct mdt_device *mdt = info->mti_mdt;
2530 struct ldlm_lock *lock = ldlm_handle2lock(h);
2531 struct ptlrpc_request *req = mdt_info_req(info);
2534 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2536 /* there is no request if mdt_object_unlock() is called
2537 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2538 if (likely(req != NULL)) {
2539 CDEBUG(D_HA, "request = %p reply state = %p"
2540 " transno = "LPD64"\n", req,
2541 req->rq_reply_state, req->rq_transno);
2542 if (mdt_cos_is_enabled(mdt)) {
2544 ldlm_lock_downgrade(lock, LCK_COS);
2547 ptlrpc_save_lock(req, h, mode, no_ack);
2549 ldlm_lock_decref(h, mode);
2551 if (mdt_is_lock_sync(lock)) {
2552 CDEBUG(D_HA, "found sync-lock,"
2553 " async commit started\n");
2554 mdt_device_commit_async(info->mti_env,
2557 LDLM_LOCK_PUT(lock);
2566 * Unlock mdt object.
2568 * Immeditely release the regular lock and the PDO lock or save the
2569 * lock in request and keep them referenced until client ACK or
2570 * transaction commit.
2572 * \param info thread info object
2573 * \param o mdt object
2574 * \param lh mdt lock handle referencing regular and PDO locks
2575 * \param decref force immediate lock releasing
2577 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2578 struct mdt_lock_handle *lh, int decref)
2582 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2583 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2585 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2586 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2591 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2592 const struct lu_fid *f,
2593 struct mdt_lock_handle *lh,
2596 struct mdt_object *o;
2598 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2602 rc = mdt_object_lock(info, o, lh, ibits,
2605 mdt_object_put(info->mti_env, o);
2612 void mdt_object_unlock_put(struct mdt_thread_info * info,
2613 struct mdt_object * o,
2614 struct mdt_lock_handle *lh,
2617 mdt_object_unlock(info, o, lh, decref);
2618 mdt_object_put(info->mti_env, o);
2622 * Generic code handling requests that have struct mdt_body passed in:
2624 * - extract mdt_body from request and save it in @info, if present;
2626 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2629 * - if HABEO_CORPUS flag is set for this request type check whether object
2630 * actually exists on storage (lu_object_exists()).
2633 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2635 const struct mdt_body *body;
2636 struct mdt_object *obj;
2637 const struct lu_env *env;
2638 struct req_capsule *pill;
2642 env = info->mti_env;
2643 pill = info->mti_pill;
2645 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2649 if (!(body->mbo_valid & OBD_MD_FLID))
2652 if (!fid_is_sane(&body->mbo_fid1)) {
2653 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2657 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2659 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2660 mdt_object_put(env, obj);
2663 info->mti_object = obj;
2672 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2674 struct req_capsule *pill = info->mti_pill;
2678 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2679 rc = mdt_body_unpack(info, flags);
2683 if (rc == 0 && (flags & HABEO_REFERO)) {
2685 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2686 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2688 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2689 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2692 rc = req_capsule_server_pack(pill);
2697 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2699 lh->mlh_type = MDT_NUL_LOCK;
2700 lh->mlh_reg_lh.cookie = 0ull;
2701 lh->mlh_reg_mode = LCK_MINMODE;
2702 lh->mlh_pdo_lh.cookie = 0ull;
2703 lh->mlh_pdo_mode = LCK_MINMODE;
2704 lh->mlh_rreg_lh.cookie = 0ull;
2705 lh->mlh_rreg_mode = LCK_MINMODE;
2708 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2710 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2711 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2715 * Initialize fields of struct mdt_thread_info. Other fields are left in
2716 * uninitialized state, because it's too expensive to zero out whole
2717 * mdt_thread_info (> 1K) on each request arrival.
2719 void mdt_thread_info_init(struct ptlrpc_request *req,
2720 struct mdt_thread_info *info)
2724 info->mti_pill = &req->rq_pill;
2727 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2728 mdt_lock_handle_init(&info->mti_lh[i]);
2730 /* mdt device: it can be NULL while CONNECT */
2731 if (req->rq_export) {
2732 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2733 info->mti_exp = req->rq_export;
2735 info->mti_mdt = NULL;
2736 info->mti_env = req->rq_svc_thread->t_env;
2737 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2739 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2740 info->mti_big_buf = LU_BUF_NULL;
2741 info->mti_body = NULL;
2742 info->mti_object = NULL;
2743 info->mti_dlm_req = NULL;
2744 info->mti_has_trans = 0;
2745 info->mti_cross_ref = 0;
2746 info->mti_opdata = 0;
2747 info->mti_big_lmm_used = 0;
2749 info->mti_spec.no_create = 0;
2750 info->mti_spec.sp_rm_entry = 0;
2751 info->mti_spec.sp_permitted = 0;
2753 info->mti_spec.u.sp_ea.eadata = NULL;
2754 info->mti_spec.u.sp_ea.eadatalen = 0;
2757 void mdt_thread_info_fini(struct mdt_thread_info *info)
2761 if (info->mti_object != NULL) {
2762 mdt_object_put(info->mti_env, info->mti_object);
2763 info->mti_object = NULL;
2766 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2767 mdt_lock_handle_fini(&info->mti_lh[i]);
2768 info->mti_env = NULL;
2769 info->mti_pill = NULL;
2770 info->mti_exp = NULL;
2772 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2773 lu_buf_free(&info->mti_big_buf);
2776 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2778 struct mdt_thread_info *mti;
2780 mti = mdt_th_info(tsi->tsi_env);
2781 LASSERT(mti != NULL);
2783 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2784 if (tsi->tsi_corpus != NULL) {
2785 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2786 lu_object_get(tsi->tsi_corpus);
2788 mti->mti_body = tsi->tsi_mdt_body;
2789 mti->mti_dlm_req = tsi->tsi_dlm_req;
2794 static int mdt_tgt_connect(struct tgt_session_info *tsi)
2796 struct ptlrpc_request *req = tgt_ses_req(tsi);
2801 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
2803 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
2804 set_current_state(TASK_UNINTERRUPTIBLE);
2805 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
2808 rc = tgt_connect(tsi);
2812 rc = mdt_init_idmap(tsi);
2817 obd_disconnect(class_export_get(req->rq_export));
2836 static int mdt_intent_getattr(enum mdt_it_code opcode,
2837 struct mdt_thread_info *info,
2838 struct ldlm_lock **,
2841 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2842 struct mdt_thread_info *info,
2843 struct ldlm_lock **lockp,
2846 static int mdt_intent_layout(enum mdt_it_code opcode,
2847 struct mdt_thread_info *info,
2848 struct ldlm_lock **,
2850 static int mdt_intent_reint(enum mdt_it_code opcode,
2851 struct mdt_thread_info *info,
2852 struct ldlm_lock **,
2855 static struct mdt_it_flavor {
2856 const struct req_format *it_fmt;
2858 int (*it_act)(enum mdt_it_code ,
2859 struct mdt_thread_info *,
2860 struct ldlm_lock **,
2863 } mdt_it_flavor[] = {
2865 .it_fmt = &RQF_LDLM_INTENT,
2866 /*.it_flags = HABEO_REFERO,*/
2868 .it_act = mdt_intent_reint,
2869 .it_reint = REINT_OPEN
2872 .it_fmt = &RQF_LDLM_INTENT,
2874 * OCREAT is not a MUTABOR request as if the file
2876 * We do the extra check of OBD_CONNECT_RDONLY in
2877 * mdt_reint_open() when we really need to create
2881 .it_act = mdt_intent_reint,
2882 .it_reint = REINT_OPEN
2885 .it_fmt = &RQF_LDLM_INTENT,
2886 .it_flags = MUTABOR,
2887 .it_act = mdt_intent_reint,
2888 .it_reint = REINT_CREATE
2890 [MDT_IT_GETATTR] = {
2891 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2892 .it_flags = HABEO_REFERO,
2893 .it_act = mdt_intent_getattr
2895 [MDT_IT_READDIR] = {
2901 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2902 .it_flags = HABEO_REFERO,
2903 .it_act = mdt_intent_getattr
2906 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2907 .it_flags = MUTABOR,
2909 .it_reint = REINT_UNLINK
2913 .it_flags = MUTABOR,
2916 [MDT_IT_GETXATTR] = {
2917 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2918 .it_flags = HABEO_CORPUS,
2919 .it_act = mdt_intent_getxattr
2922 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2924 .it_act = mdt_intent_layout
2929 mdt_intent_lock_replace(struct mdt_thread_info *info,
2930 struct ldlm_lock **lockp,
2931 struct mdt_lock_handle *lh,
2934 struct ptlrpc_request *req = mdt_info_req(info);
2935 struct ldlm_lock *lock = *lockp;
2936 struct ldlm_lock *new_lock;
2938 /* If possible resent found a lock, @lh is set to its handle */
2939 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2941 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2942 lh->mlh_reg_lh.cookie = 0;
2946 LASSERTF(new_lock != NULL,
2947 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2950 * If we've already given this lock to a client once, then we should
2951 * have no readers or writers. Otherwise, we should have one reader
2952 * _or_ writer ref (which will be zeroed below) before returning the
2955 if (new_lock->l_export == req->rq_export) {
2956 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2958 LASSERT(new_lock->l_export == NULL);
2959 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2964 if (new_lock->l_export == req->rq_export) {
2966 * Already gave this to the client, which means that we
2967 * reconstructed a reply.
2969 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2972 LDLM_LOCK_RELEASE(new_lock);
2973 lh->mlh_reg_lh.cookie = 0;
2974 RETURN(ELDLM_LOCK_REPLACED);
2978 * Fixup the lock to be given to the client.
2980 lock_res_and_lock(new_lock);
2981 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2982 * possible blocking AST. */
2983 while (new_lock->l_readers > 0) {
2984 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2985 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2986 new_lock->l_readers--;
2988 while (new_lock->l_writers > 0) {
2989 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2990 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2991 new_lock->l_writers--;
2994 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2995 new_lock->l_blocking_ast = lock->l_blocking_ast;
2996 new_lock->l_completion_ast = lock->l_completion_ast;
2997 new_lock->l_remote_handle = lock->l_remote_handle;
2998 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3000 unlock_res_and_lock(new_lock);
3002 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3003 &new_lock->l_remote_handle,
3004 &new_lock->l_exp_hash);
3006 LDLM_LOCK_RELEASE(new_lock);
3007 lh->mlh_reg_lh.cookie = 0;
3009 RETURN(ELDLM_LOCK_REPLACED);
3012 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3013 struct ldlm_lock *new_lock,
3014 struct mdt_lock_handle *lh,
3017 struct ptlrpc_request *req = mdt_info_req(info);
3018 struct ldlm_request *dlmreq;
3020 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3023 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3025 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3026 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3028 if (flags & LDLM_FL_RESENT) {
3029 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3030 lh->mlh_reg_mode = new_lock->l_granted_mode;
3032 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3033 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3034 lh->mlh_reg_lh.cookie);
3039 * If the xid matches, then we know this is a resent request, and allow
3040 * it. (It's probably an OPEN, for which we don't send a lock.
3042 if (req_xid_is_last(req))
3046 * This remote handle isn't enqueued, so we never received or processed
3047 * this request. Clear MSG_RESENT, because it can be handled like any
3048 * normal request now.
3050 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3052 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3053 dlmreq->lock_handle[0].cookie);
3056 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3057 struct mdt_thread_info *info,
3058 struct ldlm_lock **lockp,
3061 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3062 struct ldlm_reply *ldlm_rep = NULL;
3066 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3067 * (for the resend case) or a new lock. Below we will use it to
3068 * replace the original lock.
3070 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3071 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3072 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3073 rc = mdt_object_lock(info, info->mti_object, lhc,
3074 MDS_INODELOCK_XATTR,
3080 grc = mdt_getxattr(info);
3082 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3084 if (mdt_info_req(info)->rq_repmsg != NULL)
3085 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3086 if (ldlm_rep == NULL)
3087 RETURN(err_serious(-EFAULT));
3089 ldlm_rep->lock_policy_res2 = grc;
3094 static int mdt_intent_getattr(enum mdt_it_code opcode,
3095 struct mdt_thread_info *info,
3096 struct ldlm_lock **lockp,
3099 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3101 struct ldlm_reply *ldlm_rep;
3102 struct mdt_body *reqbody;
3103 struct mdt_body *repbody;
3107 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3110 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3113 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3114 repbody->mbo_eadatasize = 0;
3115 repbody->mbo_aclsize = 0;
3119 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3121 case MDT_IT_GETATTR:
3122 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3126 CERROR("Unsupported intent (%d)\n", opcode);
3127 GOTO(out_shrink, rc = -EINVAL);
3130 rc = mdt_init_ucred(info, reqbody);
3132 GOTO(out_shrink, rc);
3134 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3135 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3137 /* Get lock from request for possible resent case. */
3138 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3140 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3141 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3143 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3144 ldlm_rep->lock_policy_res2 = 0;
3145 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3146 ldlm_rep->lock_policy_res2) {
3147 lhc->mlh_reg_lh.cookie = 0ull;
3148 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3151 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3154 mdt_exit_ucred(info);
3156 mdt_client_compatibility(info);
3157 rc2 = mdt_fix_reply(info);
3163 static int mdt_intent_layout(enum mdt_it_code opcode,
3164 struct mdt_thread_info *info,
3165 struct ldlm_lock **lockp,
3168 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3169 struct layout_intent *layout;
3171 struct mdt_object *obj = NULL;
3175 if (opcode != MDT_IT_LAYOUT) {
3176 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3181 fid = &info->mti_tmp_fid2;
3182 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3184 /* Get lock from request for possible resent case. */
3185 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3187 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3189 RETURN(PTR_ERR(obj));
3191 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3192 /* get the length of lsm */
3193 rc = mdt_attr_get_eabuf_size(info, obj);
3195 mdt_object_put(info->mti_env, obj);
3199 if (rc > info->mti_mdt->mdt_max_mdsize)
3200 info->mti_mdt->mdt_max_mdsize = rc;
3203 mdt_object_put(info->mti_env, obj);
3205 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3206 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER, rc);
3207 rc = req_capsule_server_pack(info->mti_pill);
3211 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
3212 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3214 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3215 LASSERT(layout != NULL);
3216 if (layout->li_opc == LAYOUT_INTENT_ACCESS)
3217 /* return to normal/resent ldlm handling */
3220 CERROR("%s: Unsupported layout intent (%d)\n",
3221 mdt_obd_name(info->mti_mdt), layout->li_opc);
3225 static int mdt_intent_reint(enum mdt_it_code opcode,
3226 struct mdt_thread_info *info,
3227 struct ldlm_lock **lockp,
3230 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3231 struct ldlm_reply *rep = NULL;
3235 static const struct req_format *intent_fmts[REINT_MAX] = {
3236 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3237 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3242 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3246 if (mdt_it_flavor[opcode].it_reint != opc) {
3247 CERROR("Reint code %ld doesn't match intent: %d\n",
3249 RETURN(err_serious(-EPROTO));
3252 /* Get lock from request for possible resent case. */
3253 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3255 rc = mdt_reint_internal(info, lhc, opc);
3257 /* Check whether the reply has been packed successfully. */
3258 if (mdt_info_req(info)->rq_repmsg != NULL)
3259 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3261 RETURN(err_serious(-EFAULT));
3263 /* MDC expects this in any case */
3265 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3267 /* the open lock or the lock for cross-ref object should be
3268 * returned to the client */
3269 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3270 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3271 rep->lock_policy_res2 = 0;
3272 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3276 rep->lock_policy_res2 = clear_serious(rc);
3278 if (rep->lock_policy_res2 == -ENOENT &&
3279 mdt_get_disposition(rep, DISP_LOOKUP_NEG))
3280 rep->lock_policy_res2 = 0;
3282 lhc->mlh_reg_lh.cookie = 0ull;
3283 if (rc == -ENOTCONN || rc == -ENODEV ||
3284 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3286 * If it is the disconnect error (ENODEV & ENOCONN), the error
3287 * will be returned by rq_status, and client at ptlrpc layer
3288 * will detect this, then disconnect, reconnect the import
3289 * immediately, instead of impacting the following the rpc.
3294 * For other cases, the error will be returned by intent, and client
3295 * will retrieve the result from intent.
3297 RETURN(ELDLM_LOCK_ABORTED);
3300 static int mdt_intent_code(long itcode)
3308 case IT_OPEN|IT_CREAT:
3315 rc = MDT_IT_READDIR;
3318 rc = MDT_IT_GETATTR;
3330 rc = MDT_IT_GETXATTR;
3335 case IT_QUOTA_DQACQ:
3340 CERROR("Unknown intent opcode: %ld\n", itcode);
3347 static int mdt_intent_opc(long itopc, struct mdt_thread_info *info,
3348 struct ldlm_lock **lockp, __u64 flags)
3350 struct req_capsule *pill = info->mti_pill;
3351 struct ptlrpc_request *req = mdt_info_req(info);
3352 struct mdt_it_flavor *flv;
3357 opc = mdt_intent_code(itopc);
3361 if (opc == MDT_IT_QUOTA) {
3362 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3365 RETURN(-EOPNOTSUPP);
3367 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3368 /* pass the request to quota master */
3369 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3370 mdt_info_req(info), lockp,
3375 flv = &mdt_it_flavor[opc];
3376 if (flv->it_fmt != NULL)
3377 req_capsule_extend(pill, flv->it_fmt);
3379 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3383 if (flv->it_flags & MUTABOR &&
3384 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3387 if (flv->it_act != NULL) {
3388 struct ldlm_reply *rep;
3390 /* execute policy */
3391 rc = flv->it_act(opc, info, lockp, flags);
3393 /* Check whether the reply has been packed successfully. */
3394 if (req->rq_repmsg != NULL) {
3395 rep = req_capsule_server_get(info->mti_pill,
3397 rep->lock_policy_res2 =
3398 ptlrpc_status_hton(rep->lock_policy_res2);
3405 static int mdt_intent_policy(struct ldlm_namespace *ns,
3406 struct ldlm_lock **lockp, void *req_cookie,
3407 ldlm_mode_t mode, __u64 flags, void *data)
3409 struct tgt_session_info *tsi;
3410 struct mdt_thread_info *info;
3411 struct ptlrpc_request *req = req_cookie;
3412 struct ldlm_intent *it;
3413 struct req_capsule *pill;
3418 LASSERT(req != NULL);
3420 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3422 info = tsi2mdt_info(tsi);
3423 LASSERT(info != NULL);
3424 pill = info->mti_pill;
3425 LASSERT(pill->rc_req == req);
3427 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3428 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3429 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3431 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3435 /* Lock without inodebits makes no sense and will oops
3436 * later in ldlm. Let's check it now to see if we have
3437 * ibits corrupted somewhere in mdt_intent_opc().
3438 * The case for client miss to set ibits has been
3439 * processed by others. */
3440 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3441 lr_type == LDLM_IBITS,
3442 info->mti_dlm_req->lock_desc.\
3443 l_policy_data.l_inodebits.bits != 0));
3445 rc = err_serious(-EFAULT);
3447 /* No intent was provided */
3448 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3449 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3450 rc = req_capsule_server_pack(pill);
3452 rc = err_serious(rc);
3454 mdt_thread_info_fini(info);
3458 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3460 struct seq_server_site *ss = mdt_seq_site(mdt);
3462 if (ss->ss_node_id == 0)
3465 if (ss->ss_client_seq != NULL) {
3466 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3467 ss->ss_client_seq->lcs_exp = NULL;
3470 if (ss->ss_server_fld != NULL) {
3471 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3472 ss->ss_server_fld->lsf_control_exp = NULL;
3476 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3478 struct seq_server_site *ss = mdt_seq_site(mdt);
3483 if (ss->ss_server_seq != NULL)
3484 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3487 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3489 mdt_seq_fini_cli(mdt);
3490 mdt_deregister_seq_exp(mdt);
3492 return seq_site_fini(env, mdt_seq_site(mdt));
3496 * It will retrieve its FLDB entries from MDT0, and it only happens
3497 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3498 * and it needs to refresh FLDB from the MDT0.
3500 static int mdt_register_lwp_callback(void *data)
3503 struct mdt_device *mdt = data;
3504 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3508 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3510 if (!likely(fld->lsf_new))
3513 rc = lu_env_init(&env, LCT_MD_THREAD);
3515 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3519 rc = fld_update_from_controller(&env, fld);
3521 CERROR("%s: cannot update controller: rc = %d\n",
3522 mdt_obd_name(mdt), rc);
3530 static int mdt_register_seq_exp(struct mdt_device *mdt)
3532 struct seq_server_site *ss = mdt_seq_site(mdt);
3533 char *lwp_name = NULL;
3536 if (ss->ss_node_id == 0)
3539 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3540 if (lwp_name == NULL)
3541 GOTO(out_free, rc = -ENOMEM);
3543 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3547 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3552 rc = lustre_register_lwp_item(lwp_name,
3553 &ss->ss_server_fld->lsf_control_exp,
3554 mdt_register_lwp_callback, mdt);
3556 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3557 ss->ss_client_seq->lcs_exp = NULL;
3561 if (lwp_name != NULL)
3562 OBD_FREE(lwp_name, MAX_OBD_NAME);
3568 * Init client sequence manager which is used by local MDS to talk to sequence
3569 * controller on remote node.
3571 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3573 struct seq_server_site *ss = mdt_seq_site(mdt);
3578 /* check if this is adding the first MDC and controller is not yet
3580 OBD_ALLOC_PTR(ss->ss_client_seq);
3581 if (ss->ss_client_seq == NULL)
3584 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3585 if (prefix == NULL) {
3586 OBD_FREE_PTR(ss->ss_client_seq);
3587 ss->ss_client_seq = NULL;
3591 /* Note: seq_client_fini will be called in seq_site_fini */
3592 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3593 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3594 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3596 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3598 OBD_FREE_PTR(ss->ss_client_seq);
3599 ss->ss_client_seq = NULL;
3603 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3608 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3610 struct seq_server_site *ss;
3614 ss = mdt_seq_site(mdt);
3615 /* init sequence controller server(MDT0) */
3616 if (ss->ss_node_id == 0) {
3617 OBD_ALLOC_PTR(ss->ss_control_seq);
3618 if (ss->ss_control_seq == NULL)
3621 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3622 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3625 GOTO(out_seq_fini, rc);
3628 /* Init normal sequence server */
3629 OBD_ALLOC_PTR(ss->ss_server_seq);
3630 if (ss->ss_server_seq == NULL)
3631 GOTO(out_seq_fini, rc = -ENOMEM);
3633 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3634 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3636 GOTO(out_seq_fini, rc);
3638 /* init seq client for seq server to talk to seq controller(MDT0) */
3639 rc = mdt_seq_init_cli(env, mdt);
3641 GOTO(out_seq_fini, rc);
3643 if (ss->ss_node_id != 0)
3644 /* register controller export through lwp */
3645 rc = mdt_register_seq_exp(mdt);
3650 mdt_seq_fini(env, mdt);
3658 static int mdt_fld_fini(const struct lu_env *env,
3659 struct mdt_device *m)
3661 struct seq_server_site *ss = mdt_seq_site(m);
3664 if (ss && ss->ss_server_fld) {
3665 fld_server_fini(env, ss->ss_server_fld);
3666 OBD_FREE_PTR(ss->ss_server_fld);
3667 ss->ss_server_fld = NULL;
3673 static int mdt_fld_init(const struct lu_env *env,
3675 struct mdt_device *m)
3677 struct seq_server_site *ss;
3681 ss = mdt_seq_site(m);
3683 OBD_ALLOC_PTR(ss->ss_server_fld);
3684 if (ss->ss_server_fld == NULL)
3685 RETURN(rc = -ENOMEM);
3687 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3690 OBD_FREE_PTR(ss->ss_server_fld);
3691 ss->ss_server_fld = NULL;
3698 static void mdt_stack_pre_fini(const struct lu_env *env,
3699 struct mdt_device *m, struct lu_device *top)
3701 struct lustre_cfg_bufs *bufs;
3702 struct lustre_cfg *lcfg;
3703 struct mdt_thread_info *info;
3708 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3709 LASSERT(info != NULL);
3711 bufs = &info->mti_u.bufs;
3713 LASSERT(m->mdt_child_exp);
3714 LASSERT(m->mdt_child_exp->exp_obd);
3716 /* process cleanup, pass mdt obd name to get obd umount flags */
3717 /* XXX: this is needed because all layers are referenced by
3718 * objects (some of them are pinned by osd, for example *
3719 * the proper solution should be a model where object used
3720 * by osd only doesn't have mdt/mdd slices -bzzz */
3721 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3722 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3723 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3727 top->ld_ops->ldo_process_config(env, top, lcfg);
3728 lustre_cfg_free(lcfg);
3732 static void mdt_stack_fini(const struct lu_env *env,
3733 struct mdt_device *m, struct lu_device *top)
3735 struct obd_device *obd = mdt2obd_dev(m);
3736 struct lustre_cfg_bufs *bufs;
3737 struct lustre_cfg *lcfg;
3738 struct mdt_thread_info *info;
3742 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3743 LASSERT(info != NULL);
3745 lu_dev_del_linkage(top->ld_site, top);
3747 lu_site_purge(env, top->ld_site, -1);
3749 bufs = &info->mti_u.bufs;
3750 /* process cleanup, pass mdt obd name to get obd umount flags */
3751 /* another purpose is to let all layers to release their objects */
3752 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3757 lustre_cfg_bufs_set_string(bufs, 1, flags);
3758 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3763 top->ld_ops->ldo_process_config(env, top, lcfg);
3764 lustre_cfg_free(lcfg);
3766 lu_site_purge(env, top->ld_site, -1);
3768 m->mdt_child = NULL;
3769 m->mdt_bottom = NULL;
3771 obd_disconnect(m->mdt_child_exp);
3772 m->mdt_child_exp = NULL;
3774 obd_disconnect(m->mdt_bottom_exp);
3775 m->mdt_child_exp = NULL;
3778 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3779 const char *next, struct obd_export **exp)
3781 struct obd_connect_data *data = NULL;
3782 struct obd_device *obd;
3786 OBD_ALLOC_PTR(data);
3788 GOTO(out, rc = -ENOMEM);
3790 obd = class_name2obd(next);
3792 CERROR("%s: can't locate next device: %s\n",
3793 mdt_obd_name(m), next);
3794 GOTO(out, rc = -ENOTCONN);
3797 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3798 data->ocd_version = LUSTRE_VERSION_CODE;
3800 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3802 CERROR("%s: cannot connect to next dev %s (%d)\n",
3803 mdt_obd_name(m), next, rc);
3813 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3814 struct lustre_cfg *cfg)
3816 char *dev = lustre_cfg_string(cfg, 0);
3817 int rc, name_size, uuid_size;
3818 char *name, *uuid, *p;
3819 struct lustre_cfg_bufs *bufs;
3820 struct lustre_cfg *lcfg;
3821 struct obd_device *obd;
3822 struct lustre_profile *lprof;
3823 struct lu_site *site;
3826 /* in 1.8 we had the only device in the stack - MDS.
3827 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3828 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3829 * to generate names and setup MDT, MDD. MDT will be using
3830 * generated name to connect to MDD. for MDD the next device
3831 * will be LOD with name taken from so called "profile" which
3832 * is generated by mount_option line
3834 * 1.8 MGS generates config. commands like this:
3835 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3836 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3837 * 2.0 MGS generates config. commands like this:
3838 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3839 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3840 * 3:lustre-MDT0000-mdtlov 4:f
3842 * we generate MDD name from MDT one, just replacing T with D
3844 * after all the preparations, the logical equivalent will be
3845 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3846 * 3:lustre-MDT0000-mdtlov 4:f
3847 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3848 * 3:lustre-MDD0000 4:f
3850 * notice we build the stack from down to top: MDD first, then MDT */
3852 name_size = MAX_OBD_NAME;
3853 uuid_size = MAX_OBD_NAME;
3855 OBD_ALLOC(name, name_size);
3856 OBD_ALLOC(uuid, uuid_size);
3857 if (name == NULL || uuid == NULL)
3858 GOTO(cleanup_mem, rc = -ENOMEM);
3860 OBD_ALLOC_PTR(bufs);
3862 GOTO(cleanup_mem, rc = -ENOMEM);
3865 p = strstr(name, "-MDT");
3867 GOTO(free_bufs, rc = -ENOMEM);
3870 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3872 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3873 if (lprof == NULL || lprof->lp_dt == NULL) {
3874 CERROR("can't find the profile: %s\n",
3875 lustre_cfg_string(cfg, 0));
3876 GOTO(free_bufs, rc = -EINVAL);
3879 lustre_cfg_bufs_reset(bufs, name);
3880 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3881 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3882 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3884 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3886 GOTO(free_bufs, rc = -ENOMEM);
3888 rc = class_attach(lcfg);
3890 GOTO(lcfg_cleanup, rc);
3892 obd = class_name2obd(name);
3894 CERROR("Can not find obd %s (%s in config)\n",
3895 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3896 GOTO(lcfg_cleanup, rc = -EINVAL);
3899 lustre_cfg_free(lcfg);
3901 lustre_cfg_bufs_reset(bufs, name);
3902 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3903 lustre_cfg_bufs_set_string(bufs, 2, dev);
3904 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3906 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3908 GOTO(class_detach, rc = -ENOMEM);
3910 rc = class_setup(obd, lcfg);
3912 GOTO(class_detach, rc);
3914 /* connect to MDD we just setup */
3915 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3917 GOTO(class_detach, rc);
3919 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3921 LASSERT(mdt_lu_site(mdt) == NULL);
3922 mdt->mdt_lu_dev.ld_site = site;
3923 site->ls_top_dev = &mdt->mdt_lu_dev;
3924 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3926 /* now connect to bottom OSD */
3927 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3928 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3930 GOTO(class_detach, rc);
3932 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3934 rc = lu_env_refill((struct lu_env *)env);
3936 CERROR("Failure to refill session: '%d'\n", rc);
3938 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3943 class_detach(obd, lcfg);
3945 lustre_cfg_free(lcfg);
3950 OBD_FREE(name, name_size);
3952 OBD_FREE(uuid, uuid_size);
3956 /* setup quota master target on MDT0 */
3957 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3958 struct lustre_cfg *cfg)
3960 struct obd_device *obd;
3961 char *dev = lustre_cfg_string(cfg, 0);
3962 char *qmtname, *uuid, *p;
3963 struct lustre_cfg_bufs *bufs;
3964 struct lustre_cfg *lcfg;
3965 struct lustre_profile *lprof;
3966 struct obd_connect_data *data;
3970 LASSERT(mdt->mdt_qmt_exp == NULL);
3971 LASSERT(mdt->mdt_qmt_dev == NULL);
3973 /* quota master is on MDT0 only for now */
3974 if (mdt->mdt_seq_site.ss_node_id != 0)
3977 /* MGS generates config commands which look as follows:
3978 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3979 * 3:lustre-MDT0000-mdtlov 4:f
3981 * We generate the QMT name from the MDT one, just replacing MD with QM
3982 * after all the preparations, the logical equivalent will be:
3983 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
3984 * 3:lustre-MDT0000-osd 4:f */
3985 OBD_ALLOC(qmtname, MAX_OBD_NAME);
3986 OBD_ALLOC(uuid, UUID_MAX);
3987 OBD_ALLOC_PTR(bufs);
3988 OBD_ALLOC_PTR(data);
3989 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
3990 GOTO(cleanup_mem, rc = -ENOMEM);
3992 strcpy(qmtname, dev);
3993 p = strstr(qmtname, "-MDT");
3995 GOTO(cleanup_mem, rc = -ENOMEM);
3996 /* replace MD with QM */
4000 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4002 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4003 if (lprof == NULL || lprof->lp_dt == NULL) {
4004 CERROR("can't find profile for %s\n",
4005 lustre_cfg_string(cfg, 0));
4006 GOTO(cleanup_mem, rc = -EINVAL);
4009 lustre_cfg_bufs_reset(bufs, qmtname);
4010 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4011 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4012 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4014 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4016 GOTO(cleanup_mem, rc = -ENOMEM);
4018 rc = class_attach(lcfg);
4020 GOTO(lcfg_cleanup, rc);
4022 obd = class_name2obd(qmtname);
4024 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4025 lustre_cfg_string(cfg, 0));
4026 GOTO(lcfg_cleanup, rc = -EINVAL);
4029 lustre_cfg_free(lcfg);
4031 lustre_cfg_bufs_reset(bufs, qmtname);
4032 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4033 lustre_cfg_bufs_set_string(bufs, 2, dev);
4035 /* for quota, the next device should be the OSD device */
4036 lustre_cfg_bufs_set_string(bufs, 3,
4037 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4039 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4041 GOTO(class_detach, rc = -ENOMEM);
4043 rc = class_setup(obd, lcfg);
4045 GOTO(class_detach, rc);
4047 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4049 /* configure local quota objects */
4050 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4054 GOTO(class_cleanup, rc);
4056 /* connect to quota master target */
4057 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4058 data->ocd_version = LUSTRE_VERSION_CODE;
4059 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4062 CERROR("cannot connect to quota master device %s (%d)\n",
4064 GOTO(class_cleanup, rc);
4070 class_manual_cleanup(obd);
4071 mdt->mdt_qmt_dev = NULL;
4075 class_detach(obd, lcfg);
4077 lustre_cfg_free(lcfg);
4082 OBD_FREE(qmtname, MAX_OBD_NAME);
4084 OBD_FREE(uuid, UUID_MAX);
4090 /* Shutdown quota master target associated with mdt */
4091 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4095 if (mdt->mdt_qmt_exp == NULL)
4097 LASSERT(mdt->mdt_qmt_dev != NULL);
4099 /* the qmt automatically shuts down when the mdt disconnects */
4100 obd_disconnect(mdt->mdt_qmt_exp);
4101 mdt->mdt_qmt_exp = NULL;
4102 mdt->mdt_qmt_dev = NULL;
4106 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4107 * for now. This will be removed along with converting rest of MDT code
4108 * to use tgt_session_info */
4109 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4111 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4114 rc = mdt_getxattr(info);
4116 mdt_thread_info_fini(info);
4120 static struct tgt_handler mdt_tgt_handlers[] = {
4121 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4122 0, MDS_CONNECT, mdt_tgt_connect,
4123 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4124 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4125 0, MDS_DISCONNECT, tgt_disconnect,
4126 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4127 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4128 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4129 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4130 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4131 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4132 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4133 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4135 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4136 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4137 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4138 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4139 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4140 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4141 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4142 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4144 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4145 mdt_hsm_ct_register),
4146 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4147 mdt_hsm_ct_unregister),
4148 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4150 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4152 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4153 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4155 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4160 static struct tgt_handler mdt_sec_ctx_ops[] = {
4161 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4162 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4163 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4166 static struct tgt_handler mdt_quota_ops[] = {
4167 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4170 static struct tgt_opc_slice mdt_common_slice[] = {
4172 .tos_opc_start = MDS_FIRST_OPC,
4173 .tos_opc_end = MDS_LAST_OPC,
4174 .tos_hs = mdt_tgt_handlers
4177 .tos_opc_start = OBD_FIRST_OPC,
4178 .tos_opc_end = OBD_LAST_OPC,
4179 .tos_hs = tgt_obd_handlers
4182 .tos_opc_start = LDLM_FIRST_OPC,
4183 .tos_opc_end = LDLM_LAST_OPC,
4184 .tos_hs = tgt_dlm_handlers
4187 .tos_opc_start = SEC_FIRST_OPC,
4188 .tos_opc_end = SEC_LAST_OPC,
4189 .tos_hs = mdt_sec_ctx_ops
4192 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4193 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4194 .tos_hs = tgt_out_handlers
4197 .tos_opc_start = FLD_FIRST_OPC,
4198 .tos_opc_end = FLD_LAST_OPC,
4199 .tos_hs = fld_handlers
4202 .tos_opc_start = SEQ_FIRST_OPC,
4203 .tos_opc_end = SEQ_LAST_OPC,
4204 .tos_hs = seq_handlers
4207 .tos_opc_start = QUOTA_DQACQ,
4208 .tos_opc_end = QUOTA_LAST_OPC,
4209 .tos_hs = mdt_quota_ops
4212 .tos_opc_start = LLOG_FIRST_OPC,
4213 .tos_opc_end = LLOG_LAST_OPC,
4214 .tos_hs = tgt_llog_handlers
4217 .tos_opc_start = LFSCK_FIRST_OPC,
4218 .tos_opc_end = LFSCK_LAST_OPC,
4219 .tos_hs = tgt_lfsck_handlers
4227 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4229 struct md_device *next = m->mdt_child;
4230 struct lu_device *d = &m->mdt_lu_dev;
4231 struct obd_device *obd = mdt2obd_dev(m);
4232 struct lfsck_stop stop;
4235 stop.ls_status = LS_PAUSED;
4237 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4239 target_recovery_fini(obd);
4240 ping_evictor_stop();
4241 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4243 if (m->mdt_opts.mo_coordinator)
4244 mdt_hsm_cdt_stop(m);
4246 mdt_hsm_cdt_fini(m);
4248 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4249 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4250 obd_exports_barrier(obd);
4251 obd_zombie_barrier();
4255 tgt_fini(env, &m->mdt_lut);
4256 mdt_fs_cleanup(env, m);
4257 upcall_cache_cleanup(m->mdt_identity_cache);
4258 m->mdt_identity_cache = NULL;
4260 if (m->mdt_namespace != NULL) {
4261 ldlm_namespace_free(m->mdt_namespace, NULL,
4262 d->ld_obd->obd_force);
4263 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4266 mdt_quota_fini(env, m);
4268 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4270 mdt_seq_fini(env, m);
4271 mdt_fld_fini(env, m);
4276 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4278 LASSERT(atomic_read(&d->ld_ref) == 0);
4280 server_put_mount(mdt_obd_name(m), true);
4285 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4287 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4288 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4290 struct mdt_thread_info *info;
4291 struct obd_device *obd;
4292 const char *dev = lustre_cfg_string(cfg, 0);
4293 const char *num = lustre_cfg_string(cfg, 2);
4294 struct lustre_mount_info *lmi = NULL;
4295 struct lustre_sb_info *lsi;
4297 struct seq_server_site *ss_site;
4298 const char *identity_upcall = "NONE";
4299 struct md_device *next;
4305 lu_device_init(&m->mdt_lu_dev, ldt);
4307 * Environment (env) might be missing mdt_thread_key values at that
4308 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4311 * Usually device allocation path doesn't use module key values, but
4312 * mdt has to do a lot of work here, so allocate key value.
4314 rc = lu_env_refill((struct lu_env *)env);
4318 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4319 LASSERT(info != NULL);
4321 obd = class_name2obd(dev);
4322 LASSERT(obd != NULL);
4324 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4326 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4328 /* default is coordinator off, it is started through conf_param
4330 m->mdt_opts.mo_coordinator = 0;
4332 lmi = server_get_mount(dev);
4334 CERROR("Cannot get mount info for %s!\n", dev);
4337 lsi = s2lsi(lmi->lmi_sb);
4338 /* CMD is supported only in IAM mode */
4340 node_id = simple_strtol(num, NULL, 10);
4341 obd->u.obt.obt_magic = OBT_MAGIC;
4344 m->mdt_squash.rsi_uid = 0;
4345 m->mdt_squash.rsi_gid = 0;
4346 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4347 init_rwsem(&m->mdt_squash.rsi_sem);
4348 spin_lock_init(&m->mdt_osfs_lock);
4349 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4350 m->mdt_enable_remote_dir = 0;
4351 m->mdt_enable_remote_dir_gid = 0;
4353 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4354 m->mdt_lu_dev.ld_obd = obd;
4355 /* Set this lu_device to obd for error handling purposes. */
4356 obd->obd_lu_dev = &m->mdt_lu_dev;
4358 /* init the stack */
4359 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4361 CERROR("%s: Can't init device stack, rc %d\n",
4362 mdt_obd_name(m), rc);
4367 ss_site = mdt_seq_site(m);
4368 s->ld_seq_site = ss_site;
4371 /* set server index */
4372 ss_site->ss_node_id = node_id;
4374 /* failover is the default
4375 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4376 * assumed whose ss_node_id == 0 XXX
4378 obd->obd_replayable = 1;
4379 /* No connection accepted until configurations will finish */
4380 obd->obd_no_conn = 1;
4382 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4383 char *str = lustre_cfg_string(cfg, 4);
4384 if (strchr(str, 'n')) {
4385 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4386 obd->obd_replayable = 0;
4390 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4392 GOTO(err_fini_stack, rc);
4394 rc = mdt_seq_init(env, m);
4396 GOTO(err_fini_fld, rc);
4398 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4399 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4400 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4401 LDLM_NAMESPACE_SERVER,
4402 LDLM_NAMESPACE_GREEDY,
4404 if (m->mdt_namespace == NULL)
4405 GOTO(err_fini_seq, rc = -ENOMEM);
4407 m->mdt_namespace->ns_lvbp = m;
4408 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4410 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4411 /* set obd_namespace for compatibility with old code */
4412 obd->obd_namespace = m->mdt_namespace;
4414 rc = mdt_hsm_cdt_init(m);
4416 CERROR("%s: error initializing coordinator, rc %d\n",
4417 mdt_obd_name(m), rc);
4418 GOTO(err_free_ns, rc);
4421 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4422 OBD_FAIL_MDS_ALL_REQUEST_NET,
4423 OBD_FAIL_MDS_ALL_REPLY_NET);
4425 GOTO(err_free_hsm, rc);
4427 rc = mdt_fs_setup(env, m, obd, lsi);
4431 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4433 next = m->mdt_child;
4434 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4437 GOTO(err_fs_cleanup, rc);
4439 if (mntopts & MNTOPT_USERXATTR)
4440 m->mdt_opts.mo_user_xattr = 1;
4442 m->mdt_opts.mo_user_xattr = 0;
4444 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4446 GOTO(err_fs_cleanup, rc);
4448 if (mntopts & MNTOPT_ACL)
4449 m->mdt_opts.mo_acl = 1;
4451 m->mdt_opts.mo_acl = 0;
4453 /* XXX: to support suppgid for ACL, we enable identity_upcall
4454 * by default, otherwise, maybe got unexpected -EACCESS. */
4455 if (m->mdt_opts.mo_acl)
4456 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4458 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4460 &mdt_identity_upcall_cache_ops);
4461 if (IS_ERR(m->mdt_identity_cache)) {
4462 rc = PTR_ERR(m->mdt_identity_cache);
4463 m->mdt_identity_cache = NULL;
4464 GOTO(err_fs_cleanup, rc);
4467 rc = mdt_procfs_init(m, dev);
4469 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4470 GOTO(err_recovery, rc);
4473 rc = mdt_quota_init(env, m, cfg);
4475 GOTO(err_procfs, rc);
4477 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4478 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4479 "mdt_ldlm_client", m->mdt_ldlm_client);
4481 ping_evictor_start();
4483 /* recovery will be started upon mdt_prepare()
4484 * when the whole stack is complete and ready
4485 * to serve the requests */
4487 /* Reduce the initial timeout on an MDS because it doesn't need such
4488 * a long timeout as an OST does. Adaptive timeouts will adjust this
4489 * value appropriately. */
4490 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4491 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4497 target_recovery_fini(obd);
4498 upcall_cache_cleanup(m->mdt_identity_cache);
4499 m->mdt_identity_cache = NULL;
4501 mdt_fs_cleanup(env, m);
4503 tgt_fini(env, &m->mdt_lut);
4505 mdt_hsm_cdt_fini(m);
4507 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4508 obd->obd_namespace = m->mdt_namespace = NULL;
4510 mdt_seq_fini(env, m);
4512 mdt_fld_fini(env, m);
4514 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4517 server_put_mount(dev, true);
4521 /* For interoperability, the left element is old parameter, the right one
4522 * is the new version of the parameter, if some parameter is deprecated,
4523 * the new version should be set as NULL. */
4524 static struct cfg_interop_param mdt_interop_param[] = {
4525 { "mdt.group_upcall", NULL },
4526 { "mdt.quota_type", NULL },
4527 { "mdd.quota_type", NULL },
4528 { "mdt.rootsquash", "mdt.root_squash" },
4529 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4533 /* used by MGS to process specific configurations */
4534 static int mdt_process_config(const struct lu_env *env,
4535 struct lu_device *d, struct lustre_cfg *cfg)
4537 struct mdt_device *m = mdt_dev(d);
4538 struct md_device *md_next = m->mdt_child;
4539 struct lu_device *next = md2lu_dev(md_next);
4543 switch (cfg->lcfg_command) {
4545 struct obd_device *obd = d->ld_obd;
4547 /* For interoperability */
4548 struct cfg_interop_param *ptr = NULL;
4549 struct lustre_cfg *old_cfg = NULL;
4552 param = lustre_cfg_string(cfg, 1);
4553 if (param == NULL) {
4554 CERROR("param is empty\n");
4559 ptr = class_find_old_param(param, mdt_interop_param);
4561 if (ptr->new_param == NULL) {
4563 CWARN("For interoperability, skip this %s."
4564 " It is obsolete.\n", ptr->old_param);
4568 CWARN("Found old param %s, changed it to %s.\n",
4569 ptr->old_param, ptr->new_param);
4572 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4579 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4581 if (rc > 0 || rc == -ENOSYS) {
4582 /* is it an HSM var ? */
4583 rc = class_process_proc_param(PARAM_HSM,
4584 hsm_cdt_get_proc_vars(),
4586 if (rc > 0 || rc == -ENOSYS)
4587 /* we don't understand; pass it on */
4588 rc = next->ld_ops->ldo_process_config(env, next,
4592 if (old_cfg != NULL)
4593 lustre_cfg_free(cfg);
4598 /* others are passed further */
4599 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4605 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4606 const struct lu_object_header *hdr,
4607 struct lu_device *d)
4609 struct mdt_object *mo;
4613 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4615 struct lu_object *o;
4616 struct lu_object_header *h;
4619 h = &mo->mot_header;
4620 lu_object_header_init(h);
4621 lu_object_init(o, h, d);
4622 lu_object_add_top(h, o);
4623 o->lo_ops = &mdt_obj_ops;
4624 spin_lock_init(&mo->mot_write_lock);
4625 mutex_init(&mo->mot_lov_mutex);
4626 init_rwsem(&mo->mot_open_sem);
4632 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4633 const struct lu_object_conf *unused)
4635 struct mdt_device *d = mdt_dev(o->lo_dev);
4636 struct lu_device *under;
4637 struct lu_object *below;
4641 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4642 PFID(lu_object_fid(o)));
4644 under = &d->mdt_child->md_lu_dev;
4645 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4646 if (below != NULL) {
4647 lu_object_add(o, below);
4654 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4656 struct mdt_object *mo = mdt_obj(o);
4657 struct lu_object_header *h;
4661 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4662 PFID(lu_object_fid(o)));
4664 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4665 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4668 lu_object_header_fini(h);
4669 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4674 static int mdt_object_print(const struct lu_env *env, void *cookie,
4675 lu_printer_t p, const struct lu_object *o)
4677 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4679 return (*p)(env, cookie,
4680 LUSTRE_MDT_NAME"-object@%p(flags=%d, writecount=%d)",
4681 mdto, mdto->mot_flags, mdto->mot_write_count);
4684 static int mdt_prepare(const struct lu_env *env,
4685 struct lu_device *pdev,
4686 struct lu_device *cdev)
4688 struct mdt_device *mdt = mdt_dev(cdev);
4689 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4690 struct obd_device *obd = cdev->ld_obd;
4697 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4701 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4705 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4709 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4710 /* The LFSCK instance is registered just now, so it must be there when
4711 * register the namespace to such instance. */
4712 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4714 if (mdt->mdt_seq_site.ss_node_id == 0) {
4715 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4716 &mdt->mdt_md_root_fid);
4721 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4722 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4723 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4724 LASSERT(obd->obd_no_conn);
4725 spin_lock(&obd->obd_dev_lock);
4726 obd->obd_no_conn = 0;
4727 spin_unlock(&obd->obd_dev_lock);
4729 if (obd->obd_recovering == 0)
4730 mdt_postrecov(env, mdt);
4735 const struct lu_device_operations mdt_lu_ops = {
4736 .ldo_object_alloc = mdt_object_alloc,
4737 .ldo_process_config = mdt_process_config,
4738 .ldo_prepare = mdt_prepare,
4741 static const struct lu_object_operations mdt_obj_ops = {
4742 .loo_object_init = mdt_object_init,
4743 .loo_object_free = mdt_object_free,
4744 .loo_object_print = mdt_object_print
4747 static int mdt_obd_set_info_async(const struct lu_env *env,
4748 struct obd_export *exp,
4749 __u32 keylen, void *key,
4750 __u32 vallen, void *val,
4751 struct ptlrpc_request_set *set)
4757 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4758 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4766 * Match client and server connection feature flags.
4768 * Compute the compatibility flags for a connection request based on
4769 * features mutually supported by client and server.
4771 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4772 * must not be updated here, otherwise a partially initialized value may
4773 * be exposed. After the connection request is successfully processed,
4774 * the top-level MDT connect request handler atomically updates the export
4775 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4776 * reply. \see mdt_connect().
4778 * \param exp the obd_export associated with this client/target pair
4779 * \param mdt the target device for the connection
4780 * \param data stores data for this connect request
4783 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4784 * \retval -EBADE client and server feature requirements are incompatible
4786 static int mdt_connect_internal(struct obd_export *exp,
4787 struct mdt_device *mdt,
4788 struct obd_connect_data *data)
4790 LASSERT(data != NULL);
4792 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4793 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4795 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
4796 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
4797 CWARN("%s: client %s does not support ibits lock, either "
4798 "very old or an invalid client: flags "LPX64"\n",
4799 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
4800 data->ocd_connect_flags);
4804 if (!mdt->mdt_opts.mo_acl)
4805 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4807 if (!mdt->mdt_opts.mo_user_xattr)
4808 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4810 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4811 data->ocd_brw_size = min(data->ocd_brw_size,
4812 (__u32)MD_MAX_BRW_SIZE);
4813 if (data->ocd_brw_size == 0) {
4814 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4815 " ocd_version: %x ocd_grant: %d "
4816 "ocd_index: %u ocd_brw_size is "
4817 "unexpectedly zero, network data "
4818 "corruption? Refusing connection of this"
4821 exp->exp_client_uuid.uuid,
4822 exp, data->ocd_connect_flags, data->ocd_version,
4823 data->ocd_grant, data->ocd_index);
4828 /* NB: Disregard the rule against updating
4829 * exp_connect_data.ocd_connect_flags in this case, since
4830 * tgt_client_new() needs to know if this is a lightweight
4831 * connection, and it is safe to expose this flag before
4832 * connection processing completes. */
4833 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4834 spin_lock(&exp->exp_lock);
4835 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4836 spin_unlock(&exp->exp_lock);
4839 data->ocd_version = LUSTRE_VERSION_CODE;
4841 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4842 CWARN("%s: MDS requires FID support, but client not\n",
4847 if (OCD_HAS_FLAG(data, PINGLESS)) {
4848 if (ptlrpc_pinger_suppress_pings()) {
4849 spin_lock(&exp->exp_obd->obd_dev_lock);
4850 list_del_init(&exp->exp_obd_chain_timed);
4851 spin_unlock(&exp->exp_obd->obd_dev_lock);
4853 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4857 data->ocd_max_easize = mdt->mdt_max_ea_size;
4862 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
4863 struct mdt_thread_info *info,
4864 struct mdt_file_data *mfd)
4866 struct lu_context ses;
4870 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
4875 lu_context_enter(&ses);
4877 mdt_ucred(info)->uc_valid = UCRED_OLD;
4878 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
4880 lu_context_exit(&ses);
4881 lu_context_fini(&ses);
4887 static int mdt_export_cleanup(struct obd_export *exp)
4889 struct list_head closing_list;
4890 struct mdt_export_data *med = &exp->exp_mdt_data;
4891 struct obd_device *obd = exp->exp_obd;
4892 struct mdt_device *mdt;
4893 struct mdt_thread_info *info;
4895 struct mdt_file_data *mfd, *n;
4899 INIT_LIST_HEAD(&closing_list);
4900 spin_lock(&med->med_open_lock);
4901 while (!list_empty(&med->med_open_head)) {
4902 struct list_head *tmp = med->med_open_head.next;
4903 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
4905 /* Remove mfd handle so it can't be found again.
4906 * We are consuming the mfd_list reference here. */
4907 class_handle_unhash(&mfd->mfd_handle);
4908 list_move_tail(&mfd->mfd_list, &closing_list);
4910 spin_unlock(&med->med_open_lock);
4911 mdt = mdt_dev(obd->obd_lu_dev);
4912 LASSERT(mdt != NULL);
4914 rc = lu_env_init(&env, LCT_MD_THREAD);
4918 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
4919 LASSERT(info != NULL);
4920 memset(info, 0, sizeof *info);
4921 info->mti_env = &env;
4922 info->mti_mdt = mdt;
4923 info->mti_exp = exp;
4925 if (!list_empty(&closing_list)) {
4926 struct md_attr *ma = &info->mti_attr;
4928 /* Close any open files (which may also cause orphan
4930 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
4931 list_del_init(&mfd->mfd_list);
4932 ma->ma_need = ma->ma_valid = 0;
4934 /* This file is being closed due to an eviction, it
4935 * could have been modified and now dirty regarding to
4936 * HSM archive, check this!
4937 * The logic here is to mark a file dirty if there's a
4938 * chance it was dirtied before the client was evicted,
4939 * so that we don't have to wait for a release attempt
4940 * before finding out the file was actually dirty and
4941 * fail the release. Aggressively marking it dirty here
4942 * will cause the policy engine to attempt to
4943 * re-archive it; when rearchiving, we can compare the
4944 * current version to the HSM data_version and make the
4945 * archive request into a noop if it's not actually
4948 if (mfd->mfd_mode & FMODE_WRITE)
4949 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
4951 /* Don't unlink orphan on failover umount, LU-184 */
4952 if (exp->exp_flags & OBD_OPT_FAILOVER) {
4953 ma->ma_valid = MA_FLAGS;
4954 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
4956 mdt_mfd_close(info, mfd);
4959 info->mti_mdt = NULL;
4960 /* cleanup client slot early */
4961 /* Do not erase record for recoverable client. */
4962 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
4963 tgt_client_del(&env, exp);
4969 static int mdt_obd_disconnect(struct obd_export *exp)
4975 class_export_get(exp);
4977 nodemap_del_member(exp);
4978 rc = server_disconnect_export(exp);
4980 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
4982 rc = mdt_export_cleanup(exp);
4983 class_export_put(exp);
4987 /* mds_connect copy */
4988 static int mdt_obd_connect(const struct lu_env *env,
4989 struct obd_export **exp, struct obd_device *obd,
4990 struct obd_uuid *cluuid,
4991 struct obd_connect_data *data,
4994 struct obd_export *lexp;
4995 struct lustre_handle conn = { 0 };
4996 struct mdt_device *mdt;
4998 lnet_nid_t *client_nid = localdata;
5001 LASSERT(env != NULL);
5002 if (!exp || !obd || !cluuid)
5005 mdt = mdt_dev(obd->obd_lu_dev);
5008 * first, check whether the stack is ready to handle requests
5009 * XXX: probably not very appropriate method is used now
5010 * at some point we should find a better one
5012 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
5013 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
5014 rc = obd_get_info(env, mdt->mdt_child_exp,
5015 sizeof(KEY_OSP_CONNECTED),
5016 KEY_OSP_CONNECTED, NULL, NULL);
5019 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5022 rc = class_connect(&conn, obd, cluuid);
5026 lexp = class_conn2export(&conn);
5027 LASSERT(lexp != NULL);
5029 rc = mdt_connect_internal(lexp, mdt, data);
5031 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5034 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5035 rc = tgt_client_new(env, lexp);
5037 rc = nodemap_add_member(*client_nid, lexp);
5038 if (rc != 0 && rc != -EEXIST)
5041 mdt_export_stats_init(obd, lexp, localdata);
5044 /* For phase I, sync for cross-ref operation. */
5045 spin_lock(&lexp->exp_lock);
5046 lexp->exp_keep_sync = 1;
5047 spin_unlock(&lexp->exp_lock);
5051 class_disconnect(lexp);
5060 static int mdt_obd_reconnect(const struct lu_env *env,
5061 struct obd_export *exp, struct obd_device *obd,
5062 struct obd_uuid *cluuid,
5063 struct obd_connect_data *data,
5066 lnet_nid_t *client_nid = localdata;
5070 if (exp == NULL || obd == NULL || cluuid == NULL)
5073 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5075 rc = nodemap_add_member(*client_nid, exp);
5076 if (rc == 0 || rc == -EEXIST)
5077 mdt_export_stats_init(obd, exp, localdata);
5083 /* FIXME: Can we avoid using these two interfaces? */
5084 static int mdt_init_export(struct obd_export *exp)
5086 struct mdt_export_data *med = &exp->exp_mdt_data;
5090 INIT_LIST_HEAD(&med->med_open_head);
5091 spin_lock_init(&med->med_open_lock);
5092 mutex_init(&med->med_idmap_mutex);
5093 med->med_idmap = NULL;
5094 spin_lock(&exp->exp_lock);
5095 exp->exp_connecting = 1;
5096 spin_unlock(&exp->exp_lock);
5098 /* self-export doesn't need client data and ldlm initialization */
5099 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5100 &exp->exp_client_uuid)))
5103 rc = tgt_client_alloc(exp);
5107 rc = ldlm_init_export(exp);
5114 tgt_client_free(exp);
5116 CERROR("%s: Failed to initialize export: rc = %d\n",
5117 exp->exp_obd->obd_name, rc);
5121 static int mdt_destroy_export(struct obd_export *exp)
5125 if (exp_connect_rmtclient(exp))
5126 mdt_cleanup_idmap(&exp->exp_mdt_data);
5128 target_destroy_export(exp);
5129 /* destroy can be called from failed obd_setup, so
5130 * checking uuid is safer than obd_self_export */
5131 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5132 &exp->exp_client_uuid)))
5135 ldlm_destroy_export(exp);
5136 tgt_client_free(exp);
5138 LASSERT(list_empty(&exp->exp_outstanding_replies));
5139 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5144 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5145 struct linkea_data *ldata)
5149 LASSERT(ldata->ld_buf->lb_buf != NULL);
5151 if (!mdt_object_exists(mdt_obj))
5154 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5155 ldata->ld_buf, XATTR_NAME_LINK);
5156 if (rc == -ERANGE) {
5157 /* Buf was too small, figure out what we need. */
5158 lu_buf_free(ldata->ld_buf);
5159 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5160 ldata->ld_buf, XATTR_NAME_LINK);
5163 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5164 if (ldata->ld_buf->lb_buf == NULL)
5166 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5167 ldata->ld_buf, XATTR_NAME_LINK);
5172 return linkea_init(ldata);
5176 * Given an MDT object, try to look up the full path to the object.
5177 * Part of the MDT layer implementation of lfs fid2path.
5179 * \param[in] info Per-thread common data shared by MDT level handlers.
5180 * \param[in] obj Object to do path lookup of
5181 * \param[in,out] fp User-provided struct to store path information
5183 * \retval 0 Lookup successful, path information stored in fp
5184 * \retval -EAGAIN Lookup failed, usually because object is being moved
5185 * \retval negative errno if there was a problem
5187 static int mdt_path_current(struct mdt_thread_info *info,
5188 struct mdt_object *obj,
5189 struct getinfo_fid2path *fp)
5191 struct mdt_device *mdt = info->mti_mdt;
5192 struct mdt_object *mdt_obj;
5193 struct link_ea_header *leh;
5194 struct link_ea_entry *lee;
5195 struct lu_name *tmpname = &info->mti_name;
5196 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5197 struct lu_buf *buf = &info->mti_big_buf;
5200 struct linkea_data ldata = { NULL };
5205 /* temp buffer for path element, the buffer will be finally freed
5206 * in mdt_thread_info_fini */
5207 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5208 if (buf->lb_buf == NULL)
5212 ptr = fp->gf_path + fp->gf_pathlen - 1;
5215 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5217 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5218 * so checking root_fid can only happen on MDT0. */
5219 while (!lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid)) {
5220 struct lu_buf lmv_buf;
5222 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5223 if (IS_ERR(mdt_obj))
5224 GOTO(out, rc = PTR_ERR(mdt_obj));
5226 if (!mdt_object_exists(mdt_obj)) {
5227 mdt_object_put(info->mti_env, mdt_obj);
5228 GOTO(out, rc = -ENOENT);
5231 if (mdt_object_remote(mdt_obj)) {
5232 mdt_object_put(info->mti_env, mdt_obj);
5233 GOTO(remote_out, rc = -EREMOTE);
5236 rc = mdt_links_read(info, mdt_obj, &ldata);
5238 mdt_object_put(info->mti_env, mdt_obj);
5243 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5244 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5245 /* If set, use link #linkno for path lookup, otherwise use
5246 link #0. Only do this for the final path element. */
5247 if (first && fp->gf_linkno < leh->leh_reccount) {
5249 for (count = 0; count < fp->gf_linkno; count++) {
5250 lee = (struct link_ea_entry *)
5251 ((char *)lee + reclen);
5252 linkea_entry_unpack(lee, &reclen, tmpname,
5255 if (fp->gf_linkno < leh->leh_reccount - 1)
5256 /* indicate to user there are more links */
5260 lmv_buf.lb_buf = info->mti_xattr_buf;
5261 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5262 /* Check if it is slave stripes */
5263 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5264 &lmv_buf, XATTR_NAME_LMV);
5265 mdt_object_put(info->mti_env, mdt_obj);
5267 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5269 /* For slave stripes, get its master */
5270 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5271 fp->gf_fid = *tmpfid;
5274 } else if (rc < 0 && rc != -ENODATA) {
5280 /* Pack the name in the end of the buffer */
5281 ptr -= tmpname->ln_namelen;
5282 if (ptr - 1 <= fp->gf_path)
5283 GOTO(out, rc = -EOVERFLOW);
5284 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5287 /* keep the last resolved fid to the client, so the
5288 * client will build the left path on another MDT for
5290 fp->gf_fid = *tmpfid;
5296 ptr++; /* skip leading / */
5297 memmove(fp->gf_path, ptr, fp->gf_path + fp->gf_pathlen - ptr);
5304 * Given an MDT object, use mdt_path_current to get the path.
5305 * Essentially a wrapper to retry mdt_path_current a set number of times
5306 * if -EAGAIN is returned (usually because an object is being moved).
5308 * Part of the MDT layer implementation of lfs fid2path.
5310 * \param[in] info Per-thread common data shared by mdt level handlers.
5311 * \param[in] obj Object to do path lookup of
5312 * \param[in,out] fp User-provided struct for arguments and to store path
5315 * \retval 0 Lookup successful, path information stored in fp
5316 * \retval negative errno if there was a problem
5318 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5319 struct getinfo_fid2path *fp)
5321 struct mdt_device *mdt = info->mti_mdt;
5326 if (fp->gf_pathlen < 3)
5329 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5330 fp->gf_path[0] = '\0';
5334 /* Retry multiple times in case file is being moved */
5335 while (tries-- && rc == -EAGAIN)
5336 rc = mdt_path_current(info, obj, fp);
5342 * Get the full path of the provided FID, as of changelog record recno.
5344 * This checks sanity and looks up object for user provided FID
5345 * before calling the actual path lookup code.
5347 * Part of the MDT layer implementation of lfs fid2path.
5349 * \param[in] info Per-thread common data shared by mdt level handlers.
5350 * \param[in,out] fp User-provided struct for arguments and to store path
5353 * \retval 0 Lookup successful, path information and recno stored in fp
5354 * \retval -ENOENT, object does not exist
5355 * \retval negative errno if there was a problem
5357 static int mdt_fid2path(struct mdt_thread_info *info,
5358 struct getinfo_fid2path *fp)
5360 struct mdt_device *mdt = info->mti_mdt;
5361 struct mdt_object *obj;
5365 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5366 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5368 if (!fid_is_sane(&fp->gf_fid))
5371 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5372 CWARN("%s: "DFID" is invalid, sequence should be "
5373 ">= "LPX64"\n", mdt_obd_name(mdt),
5374 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5378 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5381 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5382 PFID(&fp->gf_fid), rc);
5386 if (mdt_object_remote(obj))
5388 else if (!mdt_object_exists(obj))
5394 mdt_object_put(info->mti_env, obj);
5395 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5396 PFID(&fp->gf_fid), rc);
5400 rc = mdt_path(info, obj, fp);
5402 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5403 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5405 mdt_object_put(info->mti_env, obj);
5410 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5411 void *val, int vallen)
5413 struct getinfo_fid2path *fpout, *fpin;
5416 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5419 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5420 lustre_swab_fid2path(fpin);
5422 memcpy(fpout, fpin, sizeof(*fpin));
5423 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5426 rc = mdt_fid2path(info, fpout);
5430 int mdt_get_info(struct tgt_session_info *tsi)
5440 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5442 CDEBUG(D_IOCTL, "No GETINFO key\n");
5443 RETURN(err_serious(-EFAULT));
5445 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5448 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5449 if (vallen == NULL) {
5450 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5451 tgt_name(tsi->tsi_tgt));
5452 RETURN(err_serious(-EFAULT));
5455 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5457 rc = req_capsule_server_pack(tsi->tsi_pill);
5459 RETURN(err_serious(rc));
5461 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5462 if (valout == NULL) {
5463 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5464 tgt_name(tsi->tsi_tgt));
5465 RETURN(err_serious(-EFAULT));
5468 if (KEY_IS(KEY_FID2PATH)) {
5469 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5471 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5472 mdt_thread_info_fini(info);
5479 /* Pass the ioc down */
5480 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5481 unsigned int cmd, int len, void *data)
5483 struct lu_context ioctl_session;
5484 struct md_device *next = mdt->mdt_child;
5488 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5491 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5492 lu_context_enter(&ioctl_session);
5493 env->le_ses = &ioctl_session;
5495 LASSERT(next->md_ops->mdo_iocontrol);
5496 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5498 lu_context_exit(&ioctl_session);
5499 lu_context_fini(&ioctl_session);
5503 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5505 struct obd_ioctl_data *data = karg;
5508 struct mdt_object *obj;
5509 struct mdt_lock_handle *lh;
5513 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5514 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5517 fid = (struct lu_fid *)data->ioc_inlbuf1;
5519 if (!fid_is_sane(fid))
5522 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5524 lh = &mti->mti_lh[MDT_LH_PARENT];
5525 mdt_lock_reg_init(lh, LCK_CR);
5527 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5529 RETURN(PTR_ERR(obj));
5531 if (mdt_object_remote(obj)) {
5534 * before calling version get the correct MDS should be
5535 * fid, this is error to find remote object here
5537 CERROR("nonlocal object "DFID"\n", PFID(fid));
5538 } else if (!mdt_object_exists(obj)) {
5539 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5542 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5543 *(__u64 *)data->ioc_inlbuf2 = version;
5546 mdt_object_unlock_put(mti, obj, lh, 1);
5550 /* ioctls on obd dev */
5551 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5552 void *karg, void *uarg)
5555 struct obd_device *obd = exp->exp_obd;
5556 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5557 struct dt_device *dt = mdt->mdt_bottom;
5561 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5562 rc = lu_env_init(&env, LCT_MD_THREAD);
5568 rc = mdt_device_sync(&env, mdt);
5570 case OBD_IOC_SET_READONLY:
5571 rc = dt->dd_ops->dt_ro(&env, dt);
5573 case OBD_IOC_ABORT_RECOVERY:
5574 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5575 target_stop_recovery_thread(obd);
5578 case OBD_IOC_CHANGELOG_REG:
5579 case OBD_IOC_CHANGELOG_DEREG:
5580 case OBD_IOC_CHANGELOG_CLEAR:
5581 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5583 case OBD_IOC_START_LFSCK: {
5584 struct md_device *next = mdt->mdt_child;
5585 struct obd_ioctl_data *data = karg;
5586 struct lfsck_start_param lsp;
5588 if (unlikely(data == NULL)) {
5593 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5594 lsp.lsp_index_valid = 0;
5595 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5598 case OBD_IOC_STOP_LFSCK: {
5599 struct md_device *next = mdt->mdt_child;
5600 struct obd_ioctl_data *data = karg;
5601 struct lfsck_stop stop;
5603 stop.ls_status = LS_STOPPED;
5604 /* Old lfsck utils may pass NULL @stop. */
5605 if (data->ioc_inlbuf1 == NULL)
5609 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5611 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5614 case OBD_IOC_GET_OBJ_VERSION: {
5615 struct mdt_thread_info *mti;
5616 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5617 memset(mti, 0, sizeof *mti);
5618 mti->mti_env = &env;
5622 rc = mdt_ioc_version_get(mti, karg);
5625 case OBD_IOC_CATLOGLIST: {
5626 struct mdt_thread_info *mti;
5628 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5629 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5630 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5631 &mti->mti_tmp_fid1);
5636 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5637 mdt_obd_name(mdt), cmd, rc);
5644 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5646 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5647 struct lfsck_start_param lsp;
5651 lsp.lsp_start = NULL;
5652 lsp.lsp_index_valid = 0;
5653 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
5654 OBD_IOC_START_LFSCK,
5656 if (rc != 0 && rc != -EALREADY)
5657 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
5658 mdt_obd_name(mdt), rc);
5660 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5664 static int mdt_obd_postrecov(struct obd_device *obd)
5669 rc = lu_env_init(&env, LCT_MD_THREAD);
5672 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5677 static struct obd_ops mdt_obd_device_ops = {
5678 .o_owner = THIS_MODULE,
5679 .o_set_info_async = mdt_obd_set_info_async,
5680 .o_connect = mdt_obd_connect,
5681 .o_reconnect = mdt_obd_reconnect,
5682 .o_disconnect = mdt_obd_disconnect,
5683 .o_init_export = mdt_init_export,
5684 .o_destroy_export = mdt_destroy_export,
5685 .o_iocontrol = mdt_iocontrol,
5686 .o_postrecov = mdt_obd_postrecov,
5689 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5690 struct lu_device *d)
5692 struct mdt_device *m = mdt_dev(d);
5699 static struct lu_device *mdt_device_free(const struct lu_env *env,
5700 struct lu_device *d)
5702 struct mdt_device *m = mdt_dev(d);
5705 lu_device_fini(&m->mdt_lu_dev);
5711 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5712 struct lu_device_type *t,
5713 struct lustre_cfg *cfg)
5715 struct lu_device *l;
5716 struct mdt_device *m;
5723 rc = mdt_init0(env, m, t, cfg);
5725 mdt_device_free(env, l);
5730 l = ERR_PTR(-ENOMEM);
5734 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5735 LU_KEY_INIT(mdt, struct mdt_thread_info);
5737 static void mdt_key_fini(const struct lu_context *ctx,
5738 struct lu_context_key *key, void* data)
5740 struct mdt_thread_info *info = data;
5742 if (info->mti_big_lmm) {
5743 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5744 info->mti_big_lmm = NULL;
5745 info->mti_big_lmmsize = 0;
5750 /* context key: mdt_thread_key */
5751 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5753 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5755 return lu_ucred(info->mti_env);
5758 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5760 return lu_ucred_check(info->mti_env);
5764 * Enable/disable COS (Commit On Sharing).
5766 * Set/Clear the COS flag in mdt options.
5768 * \param mdt mdt device
5769 * \param val 0 disables COS, other values enable COS
5771 void mdt_enable_cos(struct mdt_device *mdt, int val)
5776 mdt->mdt_opts.mo_cos = !!val;
5777 rc = lu_env_init(&env, LCT_LOCAL);
5778 if (unlikely(rc != 0)) {
5779 CWARN("%s: lu_env initialization failed, cannot "
5780 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
5783 mdt_device_sync(&env, mdt);
5788 * Check COS (Commit On Sharing) status.
5790 * Return COS flag status.
5792 * \param mdt mdt device
5794 int mdt_cos_is_enabled(struct mdt_device *mdt)
5796 return mdt->mdt_opts.mo_cos != 0;
5799 static struct lu_device_type_operations mdt_device_type_ops = {
5800 .ldto_device_alloc = mdt_device_alloc,
5801 .ldto_device_free = mdt_device_free,
5802 .ldto_device_fini = mdt_device_fini
5805 static struct lu_device_type mdt_device_type = {
5806 .ldt_tags = LU_DEVICE_MD,
5807 .ldt_name = LUSTRE_MDT_NAME,
5808 .ldt_ops = &mdt_device_type_ops,
5809 .ldt_ctx_tags = LCT_MD_THREAD
5812 static int __init mdt_mod_init(void)
5816 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5817 FID_NOBRACE_LEN + 1);
5818 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5820 rc = lu_kmem_init(mdt_caches);
5824 rc = mds_mod_init();
5828 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
5829 LUSTRE_MDT_NAME, &mdt_device_type);
5834 lu_kmem_fini(mdt_caches);
5841 static void __exit mdt_mod_exit(void)
5843 class_unregister_type(LUSTRE_MDT_NAME);
5845 lu_kmem_fini(mdt_caches);
5848 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
5849 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5850 MODULE_VERSION(LUSTRE_VERSION_STRING);
5851 MODULE_LICENSE("GPL");
5853 module_init(mdt_mod_init);
5854 module_exit(mdt_mod_exit);