4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_handler.c
38 * Lustre Metadata Target (mdt) request handler
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
43 * Author: Mike Shaver <shaver@clusterfs.com>
44 * Author: Nikita Danilov <nikita@clusterfs.com>
45 * Author: Huang Hua <huanghua@clusterfs.com>
46 * Author: Yury Umanets <umka@clusterfs.com>
49 #define DEBUG_SUBSYSTEM S_MDS
51 #include <linux/module.h>
53 * struct OBD_{ALLOC,FREE}*()
55 #include <obd_support.h>
56 #include <lustre_ioctl.h>
57 /* struct ptlrpc_request */
58 #include <lustre_net.h>
59 /* struct obd_export */
60 #include <lustre_export.h>
61 /* struct obd_device */
64 #include <dt_object.h>
65 #include <lustre_mds.h>
66 #include <lustre_log.h>
67 #include "mdt_internal.h"
68 #include <lustre_acl.h>
69 #include <lustre_param.h>
70 #include <lustre_quota.h>
71 #include <lustre_lfsck.h>
72 #include <lustre_nodemap.h>
74 mdl_mode_t mdt_mdl_lock_modes[] = {
75 [LCK_MINMODE] = MDL_MINMODE,
82 [LCK_GROUP] = MDL_GROUP
85 ldlm_mode_t mdt_dlm_lock_modes[] = {
86 [MDL_MINMODE] = LCK_MINMODE,
93 [MDL_GROUP] = LCK_GROUP
96 static struct mdt_device *mdt_dev(struct lu_device *d);
97 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
99 static const struct lu_object_operations mdt_obj_ops;
101 /* Slab for MDT object allocation */
102 static struct kmem_cache *mdt_object_kmem;
104 /* For HSM restore handles */
105 struct kmem_cache *mdt_hsm_cdt_kmem;
107 /* For HSM request handles */
108 struct kmem_cache *mdt_hsm_car_kmem;
110 static struct lu_kmem_descr mdt_caches[] = {
112 .ckd_cache = &mdt_object_kmem,
113 .ckd_name = "mdt_obj",
114 .ckd_size = sizeof(struct mdt_object)
117 .ckd_cache = &mdt_hsm_cdt_kmem,
118 .ckd_name = "mdt_cdt_restore_handle",
119 .ckd_size = sizeof(struct cdt_restore_handle)
122 .ckd_cache = &mdt_hsm_car_kmem,
123 .ckd_name = "mdt_cdt_agent_req",
124 .ckd_size = sizeof(struct cdt_agent_req)
131 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
135 return rep->lock_policy_res1 & op_flag;
138 void mdt_clear_disposition(struct mdt_thread_info *info,
139 struct ldlm_reply *rep, __u64 op_flag)
142 info->mti_opdata &= ~op_flag;
143 tgt_opdata_clear(info->mti_env, op_flag);
146 rep->lock_policy_res1 &= ~op_flag;
149 void mdt_set_disposition(struct mdt_thread_info *info,
150 struct ldlm_reply *rep, __u64 op_flag)
153 info->mti_opdata |= op_flag;
154 tgt_opdata_set(info->mti_env, op_flag);
157 rep->lock_policy_res1 |= op_flag;
160 void mdt_lock_reg_init(struct mdt_lock_handle *lh, ldlm_mode_t lm)
162 lh->mlh_pdo_hash = 0;
163 lh->mlh_reg_mode = lm;
164 lh->mlh_rreg_mode = lm;
165 lh->mlh_type = MDT_REG_LOCK;
168 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, ldlm_mode_t lock_mode,
169 const struct lu_name *lname)
171 lh->mlh_reg_mode = lock_mode;
172 lh->mlh_rreg_mode = lock_mode;
173 lh->mlh_type = MDT_PDO_LOCK;
175 if (lu_name_is_valid(lname)) {
176 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
178 /* XXX Workaround for LU-2856
180 * Zero is a valid return value of full_name_hash, but
181 * several users of mlh_pdo_hash assume a non-zero
182 * hash value. We therefore map zero onto an
183 * arbitrary, but consistent value (1) to avoid
184 * problems further down the road. */
185 if (unlikely(lh->mlh_pdo_hash == 0))
186 lh->mlh_pdo_hash = 1;
188 lh->mlh_pdo_hash = 0;
192 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
193 struct mdt_lock_handle *lh)
199 * Any dir access needs couple of locks:
201 * 1) on part of dir we gonna take lookup/modify;
203 * 2) on whole dir to protect it from concurrent splitting and/or to
204 * flush client's cache for readdir().
206 * so, for a given mode and object this routine decides what lock mode
207 * to use for lock #2:
209 * 1) if caller's gonna lookup in dir then we need to protect dir from
210 * being splitted only - LCK_CR
212 * 2) if caller's gonna modify dir then we need to protect dir from
213 * being splitted and to flush cache - LCK_CW
215 * 3) if caller's gonna modify dir and that dir seems ready for
216 * splitting then we need to protect it from any type of access
217 * (lookup/modify/split) - LCK_EX --bzzz
220 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
221 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
224 * Ask underlaying level its opinion about preferable PDO lock mode
225 * having access type passed as regular lock mode:
227 * - MDL_MINMODE means that lower layer does not want to specify lock
230 * - MDL_NL means that no PDO lock should be taken. This is used in some
231 * cases. Say, for non-splittable directories no need to use PDO locks
234 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
235 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
237 if (mode != MDL_MINMODE) {
238 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
241 * Lower layer does not want to specify locking mode. We do it
242 * our selves. No special protection is needed, just flush
243 * client's cache on modification and allow concurrent
246 switch (lh->mlh_reg_mode) {
248 lh->mlh_pdo_mode = LCK_EX;
251 lh->mlh_pdo_mode = LCK_CR;
254 lh->mlh_pdo_mode = LCK_CW;
257 CERROR("Not expected lock type (0x%x)\n",
258 (int)lh->mlh_reg_mode);
263 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
267 static int mdt_getstatus(struct tgt_session_info *tsi)
269 struct mdt_thread_info *info = tsi2mdt_info(tsi);
270 struct mdt_device *mdt = info->mti_mdt;
271 struct mdt_body *repbody;
276 rc = mdt_check_ucred(info);
278 GOTO(out, rc = err_serious(rc));
280 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK))
281 GOTO(out, rc = err_serious(-ENOMEM));
283 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
284 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
285 repbody->mbo_valid |= OBD_MD_FLID;
289 mdt_thread_info_fini(info);
293 static int mdt_statfs(struct tgt_session_info *tsi)
295 struct ptlrpc_request *req = tgt_ses_req(tsi);
296 struct mdt_thread_info *info = tsi2mdt_info(tsi);
297 struct md_device *next = info->mti_mdt->mdt_child;
298 struct ptlrpc_service_part *svcpt;
299 struct obd_statfs *osfs;
304 svcpt = req->rq_rqbd->rqbd_svcpt;
306 /* This will trigger a watchdog timeout */
307 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
308 (MDT_SERVICE_WATCHDOG_FACTOR *
309 at_get(&svcpt->scp_at_estimate)) + 1);
311 rc = mdt_check_ucred(info);
313 GOTO(out, rc = err_serious(rc));
315 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
316 GOTO(out, rc = err_serious(-ENOMEM));
318 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
320 GOTO(out, rc = -EPROTO);
322 /** statfs information are cached in the mdt_device */
323 if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
324 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
325 /** statfs data is too old, get up-to-date one */
326 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
329 spin_lock(&info->mti_mdt->mdt_osfs_lock);
330 info->mti_mdt->mdt_osfs = *osfs;
331 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
332 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
334 /** use cached statfs data */
335 spin_lock(&info->mti_mdt->mdt_osfs_lock);
336 *osfs = info->mti_mdt->mdt_osfs;
337 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
341 mdt_counter_incr(req, LPROC_MDT_STATFS);
343 mdt_thread_info_fini(info);
347 #ifdef CONFIG_FS_POSIX_ACL
349 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
351 * \param info thread info object
352 * \param repbody reply to pack ACLs into
353 * \param o mdt object of file to examine
354 * \param nodemap nodemap of client to reply to
356 * \retval -errno error getting or parsing ACL from disk
358 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
359 struct mdt_object *o, struct lu_nodemap *nodemap)
361 const struct lu_env *env = info->mti_env;
362 struct md_object *next = mdt_object_child(o);
363 struct lu_buf *buf = &info->mti_buf;
366 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
367 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
369 if (buf->lb_len == 0)
372 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
374 if (rc == -ENODATA) {
375 repbody->mbo_aclsize = 0;
376 repbody->mbo_valid |= OBD_MD_FLACL;
378 } else if (rc == -EOPNOTSUPP) {
381 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
382 mdt_obd_name(info->mti_mdt),
383 PFID(mdt_object_fid(o)), rc);
386 rc = nodemap_map_acl(nodemap, buf->lb_buf,
387 rc, NODEMAP_FS_TO_CLIENT);
388 /* if all ACLs mapped out, rc is still >= 0 */
390 CERROR("%s: nodemap_map_acl unable to parse "DFID
391 " ACL: rc = %d\n", mdt_obd_name(info->mti_mdt),
392 PFID(mdt_object_fid(o)), rc);
394 repbody->mbo_aclsize = rc;
395 repbody->mbo_valid |= OBD_MD_FLACL;
403 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
404 const struct lu_attr *attr, const struct lu_fid *fid)
406 struct md_attr *ma = &info->mti_attr;
407 struct obd_export *exp = info->mti_exp;
408 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
410 LASSERT(ma->ma_valid & MA_INODE);
412 if (attr->la_valid & LA_ATIME) {
413 b->mbo_atime = attr->la_atime;
414 b->mbo_valid |= OBD_MD_FLATIME;
416 if (attr->la_valid & LA_MTIME) {
417 b->mbo_mtime = attr->la_mtime;
418 b->mbo_valid |= OBD_MD_FLMTIME;
420 if (attr->la_valid & LA_CTIME) {
421 b->mbo_ctime = attr->la_ctime;
422 b->mbo_valid |= OBD_MD_FLCTIME;
424 if (attr->la_valid & LA_FLAGS) {
425 b->mbo_flags = attr->la_flags;
426 b->mbo_valid |= OBD_MD_FLFLAGS;
428 if (attr->la_valid & LA_NLINK) {
429 b->mbo_nlink = attr->la_nlink;
430 b->mbo_valid |= OBD_MD_FLNLINK;
432 if (attr->la_valid & LA_UID) {
433 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
434 NODEMAP_FS_TO_CLIENT,
436 b->mbo_valid |= OBD_MD_FLUID;
438 if (attr->la_valid & LA_GID) {
439 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
440 NODEMAP_FS_TO_CLIENT,
442 b->mbo_valid |= OBD_MD_FLGID;
444 b->mbo_mode = attr->la_mode;
445 if (attr->la_valid & LA_MODE)
446 b->mbo_valid |= OBD_MD_FLMODE;
447 if (attr->la_valid & LA_TYPE)
448 b->mbo_valid |= OBD_MD_FLTYPE;
452 b->mbo_valid |= OBD_MD_FLID;
453 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid="LPX64"\n",
454 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
458 mdt_body_reverse_idmap(info, b);
460 if (!(attr->la_valid & LA_TYPE))
463 b->mbo_rdev = attr->la_rdev;
464 b->mbo_size = attr->la_size;
465 b->mbo_blocks = attr->la_blocks;
467 if (!S_ISREG(attr->la_mode)) {
468 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
469 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
470 /* means no objects are allocated on osts. */
471 LASSERT(!(ma->ma_valid & MA_LOV));
472 /* just ignore blocks occupied by extend attributes on MDS */
474 /* if no object is allocated on osts, the size on mds is valid.
476 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
477 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
478 ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
479 /* A released file stores its size on MDS. */
480 /* But return 1 block for released file, unless tools like tar
481 * will consider it fully sparse. (LU-3864)
483 if (unlikely(b->mbo_size == 0))
487 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
490 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
491 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
492 PFID(fid), (unsigned long long)b->mbo_size);
495 static inline int mdt_body_has_lov(const struct lu_attr *la,
496 const struct mdt_body *body)
498 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
499 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
502 void mdt_client_compatibility(struct mdt_thread_info *info)
504 struct mdt_body *body;
505 struct ptlrpc_request *req = mdt_info_req(info);
506 struct obd_export *exp = req->rq_export;
507 struct md_attr *ma = &info->mti_attr;
508 struct lu_attr *la = &ma->ma_attr;
511 if (exp_connect_layout(exp))
512 /* the client can deal with 16-bit lmm_stripe_count */
515 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
517 if (!mdt_body_has_lov(la, body))
520 /* now we have a reply with a lov for a client not compatible with the
521 * layout lock so we have to clean the layout generation number */
522 if (S_ISREG(la->la_mode))
523 ma->ma_lmm->lmm_layout_gen = 0;
527 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
528 struct mdt_object *o)
530 const struct lu_env *env = info->mti_env;
533 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
542 /* Is it a directory? Let's check for the LMV as well */
543 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
544 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
548 rc2 = mo_xattr_get(env, mdt_object_child(o),
550 XATTR_NAME_DEFAULT_LMV);
552 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
560 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
563 const struct lu_env *env = info->mti_env;
567 LASSERT(info->mti_big_lmm_used == 0);
568 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
572 /* big_lmm may need to be grown */
573 if (info->mti_big_lmmsize < rc) {
574 int size = size_roundup_power2(rc);
576 if (info->mti_big_lmmsize > 0) {
577 /* free old buffer */
578 LASSERT(info->mti_big_lmm);
579 OBD_FREE_LARGE(info->mti_big_lmm,
580 info->mti_big_lmmsize);
581 info->mti_big_lmm = NULL;
582 info->mti_big_lmmsize = 0;
585 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
586 if (info->mti_big_lmm == NULL)
588 info->mti_big_lmmsize = size;
590 LASSERT(info->mti_big_lmmsize >= rc);
592 info->mti_buf.lb_buf = info->mti_big_lmm;
593 info->mti_buf.lb_len = info->mti_big_lmmsize;
594 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
599 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
600 struct md_attr *ma, const char *name)
602 struct md_object *next = mdt_object_child(o);
603 struct lu_buf *buf = &info->mti_buf;
606 if (strcmp(name, XATTR_NAME_LOV) == 0) {
607 buf->lb_buf = ma->ma_lmm;
608 buf->lb_len = ma->ma_lmm_size;
609 LASSERT(!(ma->ma_valid & MA_LOV));
610 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
611 buf->lb_buf = ma->ma_lmv;
612 buf->lb_len = ma->ma_lmv_size;
613 LASSERT(!(ma->ma_valid & MA_LMV));
614 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
615 buf->lb_buf = ma->ma_lmv;
616 buf->lb_len = ma->ma_lmv_size;
617 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
622 rc = mo_xattr_get(info->mti_env, next, buf, name);
626 if (strcmp(name, XATTR_NAME_LOV) == 0) {
627 if (info->mti_big_lmm_used)
628 ma->ma_lmm = info->mti_big_lmm;
630 /* NOT return LOV EA with hole to old client. */
631 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
632 LOV_PATTERN_F_HOLE) &&
633 !(exp_connect_flags(info->mti_exp) &
634 OBD_CONNECT_LFSCK)) {
637 ma->ma_lmm_size = rc;
638 ma->ma_valid |= MA_LOV;
640 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
641 if (info->mti_big_lmm_used)
642 ma->ma_lmv = info->mti_big_lmm;
644 ma->ma_lmv_size = rc;
645 ma->ma_valid |= MA_LMV;
646 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
647 ma->ma_lmv_size = rc;
648 ma->ma_valid |= MA_LMV_DEF;
651 /* Update mdt_max_mdsize so all clients will be aware that */
652 if (info->mti_mdt->mdt_max_mdsize < rc)
653 info->mti_mdt->mdt_max_mdsize = rc;
656 } else if (rc == -ENODATA) {
659 } else if (rc == -ERANGE) {
660 /* Default LMV has fixed size, so it must be able to fit
661 * in the original buffer */
662 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
664 rc = mdt_big_xattr_get(info, o, name);
666 info->mti_big_lmm_used = 1;
674 static int mdt_attr_get_pfid(struct mdt_thread_info *info,
675 struct mdt_object *o, struct lu_fid *pfid)
677 struct lu_buf *buf = &info->mti_buf;
678 struct link_ea_header *leh;
679 struct link_ea_entry *lee;
683 buf->lb_buf = info->mti_big_lmm;
684 buf->lb_len = info->mti_big_lmmsize;
685 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
686 buf, XATTR_NAME_LINK);
687 /* ignore errors, MA_PFID won't be set and it is
688 * up to the caller to treat this as an error */
689 if (rc == -ERANGE || buf->lb_len == 0) {
690 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
691 buf->lb_buf = info->mti_big_lmm;
692 buf->lb_len = info->mti_big_lmmsize;
697 if (rc < sizeof(*leh)) {
698 CERROR("short LinkEA on "DFID": rc = %d\n",
699 PFID(mdt_object_fid(o)), rc);
703 leh = (struct link_ea_header *) buf->lb_buf;
704 lee = (struct link_ea_entry *)(leh + 1);
705 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
706 leh->leh_magic = LINK_EA_MAGIC;
707 leh->leh_reccount = __swab32(leh->leh_reccount);
708 leh->leh_len = __swab64(leh->leh_len);
710 if (leh->leh_magic != LINK_EA_MAGIC)
712 if (leh->leh_reccount == 0)
715 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
716 fid_be_to_cpu(pfid, pfid);
721 int mdt_attr_get_complex(struct mdt_thread_info *info,
722 struct mdt_object *o, struct md_attr *ma)
724 const struct lu_env *env = info->mti_env;
725 struct md_object *next = mdt_object_child(o);
726 struct lu_buf *buf = &info->mti_buf;
727 int need = ma->ma_need;
734 if (mdt_object_exists(o) == 0)
735 GOTO(out, rc = -ENOENT);
736 mode = lu_object_attr(&next->mo_lu);
738 if (need & MA_INODE) {
739 ma->ma_need = MA_INODE;
740 rc = mo_attr_get(env, next, ma);
743 ma->ma_valid |= MA_INODE;
746 if (need & MA_PFID) {
747 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
749 ma->ma_valid |= MA_PFID;
750 /* ignore this error, parent fid is not mandatory */
754 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
755 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
760 if (need & MA_LMV && S_ISDIR(mode)) {
761 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
766 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
767 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
772 if (need & MA_HSM && S_ISREG(mode)) {
773 buf->lb_buf = info->mti_xattr_buf;
774 buf->lb_len = sizeof(info->mti_xattr_buf);
775 CLASSERT(sizeof(struct hsm_attrs) <=
776 sizeof(info->mti_xattr_buf));
777 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
778 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
780 ma->ma_valid |= MA_HSM;
781 else if (rc2 < 0 && rc2 != -ENODATA)
785 #ifdef CONFIG_FS_POSIX_ACL
786 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
787 buf->lb_buf = ma->ma_acl;
788 buf->lb_len = ma->ma_acl_size;
789 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
791 ma->ma_acl_size = rc2;
792 ma->ma_valid |= MA_ACL_DEF;
793 } else if (rc2 == -ENODATA) {
802 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = "LPX64" ma_lmm=%p\n",
803 rc, ma->ma_valid, ma->ma_lmm);
807 static int mdt_getattr_internal(struct mdt_thread_info *info,
808 struct mdt_object *o, int ma_need)
810 struct md_object *next = mdt_object_child(o);
811 const struct mdt_body *reqbody = info->mti_body;
812 struct ptlrpc_request *req = mdt_info_req(info);
813 struct md_attr *ma = &info->mti_attr;
814 struct lu_attr *la = &ma->ma_attr;
815 struct req_capsule *pill = info->mti_pill;
816 const struct lu_env *env = info->mti_env;
817 struct mdt_body *repbody;
818 struct lu_buf *buffer = &info->mti_buf;
819 struct obd_export *exp = info->mti_exp;
820 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
825 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
826 RETURN(err_serious(-ENOMEM));
828 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
832 if (mdt_object_remote(o)) {
833 /* This object is located on remote node.*/
834 /* Return -ENOTSUPP for old client */
835 if (!mdt_is_dne_client(req->rq_export))
836 GOTO(out, rc = -ENOTSUPP);
838 repbody->mbo_fid1 = *mdt_object_fid(o);
839 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
843 if (reqbody->mbo_eadatasize > 0) {
844 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
845 if (buffer->lb_buf == NULL)
846 GOTO(out, rc = -EPROTO);
847 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
850 buffer->lb_buf = NULL;
852 ma_need &= ~(MA_LOV | MA_LMV);
853 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
854 mdt_obd_name(info->mti_mdt),
855 req->rq_export->exp_client_uuid.uuid);
858 /* If it is dir object and client require MEA, then we got MEA */
859 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
860 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
861 /* Assumption: MDT_MD size is enough for lmv size. */
862 ma->ma_lmv = buffer->lb_buf;
863 ma->ma_lmv_size = buffer->lb_len;
864 ma->ma_need = MA_INODE;
865 if (ma->ma_lmv_size > 0) {
866 if (reqbody->mbo_valid & OBD_MD_MEA)
867 ma->ma_need |= MA_LMV;
868 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
869 ma->ma_need |= MA_LMV_DEF;
872 ma->ma_lmm = buffer->lb_buf;
873 ma->ma_lmm_size = buffer->lb_len;
874 ma->ma_need = MA_INODE | MA_HSM;
875 if (ma->ma_lmm_size > 0)
876 ma->ma_need |= MA_LOV;
879 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
880 reqbody->mbo_valid & OBD_MD_FLDIREA &&
881 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
882 /* get default stripe info for this dir. */
883 ma->ma_need |= MA_LOV_DEF;
885 ma->ma_need |= ma_need;
887 rc = mdt_attr_get_complex(info, o, ma);
889 CERROR("%s: getattr error for "DFID": rc = %d\n",
890 mdt_obd_name(info->mti_mdt),
891 PFID(mdt_object_fid(o)), rc);
895 /* if file is released, check if a restore is running */
896 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_RELEASED) &&
897 mdt_hsm_restore_is_running(info, mdt_object_fid(o))) {
898 repbody->mbo_t_state = MS_RESTORE;
899 repbody->mbo_valid |= OBD_MD_TSTATE;
902 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
904 /* the Lustre protocol supposes to return default striping
905 * on the user-visible root if explicitly requested */
906 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
907 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
908 struct lu_fid rootfid;
909 struct mdt_object *root;
910 struct mdt_device *mdt = info->mti_mdt;
912 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
915 root = mdt_object_find(env, mdt, &rootfid);
917 RETURN(PTR_ERR(root));
918 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
919 mdt_object_put(info->mti_env, root);
921 CERROR("%s: getattr error for "DFID": rc = %d\n",
922 mdt_obd_name(info->mti_mdt),
923 PFID(mdt_object_fid(o)), rc);
928 if (likely(ma->ma_valid & MA_INODE))
929 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
933 if (mdt_body_has_lov(la, reqbody)) {
934 if (ma->ma_valid & MA_LOV) {
935 LASSERT(ma->ma_lmm_size);
936 repbody->mbo_eadatasize = ma->ma_lmm_size;
937 if (S_ISDIR(la->la_mode))
938 repbody->mbo_valid |= OBD_MD_FLDIREA;
940 repbody->mbo_valid |= OBD_MD_FLEASIZE;
941 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
943 if (ma->ma_valid & MA_LMV) {
944 /* Return -ENOTSUPP for old client */
945 if (!mdt_is_striped_client(req->rq_export))
948 LASSERT(S_ISDIR(la->la_mode));
949 mdt_dump_lmv(D_INFO, ma->ma_lmv);
950 repbody->mbo_eadatasize = ma->ma_lmv_size;
951 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
953 if (ma->ma_valid & MA_LMV_DEF) {
954 /* Return -ENOTSUPP for old client */
955 if (!mdt_is_striped_client(req->rq_export))
957 LASSERT(S_ISDIR(la->la_mode));
958 repbody->mbo_eadatasize = ma->ma_lmv_size;
959 repbody->mbo_valid |= (OBD_MD_FLDIREA |
962 } else if (S_ISLNK(la->la_mode) &&
963 reqbody->mbo_valid & OBD_MD_LINKNAME) {
964 buffer->lb_buf = ma->ma_lmm;
965 /* eadatasize from client includes NULL-terminator, so
966 * there is no need to read it */
967 buffer->lb_len = reqbody->mbo_eadatasize - 1;
968 rc = mo_readlink(env, next, buffer);
969 if (unlikely(rc <= 0)) {
970 CERROR("%s: readlink failed for "DFID": rc = %d\n",
971 mdt_obd_name(info->mti_mdt),
972 PFID(mdt_object_fid(o)), rc);
975 int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
977 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
979 repbody->mbo_valid |= OBD_MD_LINKNAME;
980 /* we need to report back size with NULL-terminator
981 * because client expects that */
982 repbody->mbo_eadatasize = rc + 1;
983 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
984 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
985 "on "DFID ", expected %d\n",
986 mdt_obd_name(info->mti_mdt),
987 rc, PFID(mdt_object_fid(o)),
988 reqbody->mbo_eadatasize - 1);
990 ((char *)ma->ma_lmm)[rc] = 0;
992 /* If the total CDEBUG() size is larger than a page, it
993 * will print a warning to the console, avoid this by
994 * printing just the last part of the symlink. */
995 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
996 print_limit < rc ? "..." : "", print_limit,
997 (char *)ma->ma_lmm + rc - print_limit, rc);
1002 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1003 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1004 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1005 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1006 repbody->mbo_max_mdsize);
1009 if (exp_connect_rmtclient(info->mti_exp) &&
1010 reqbody->mbo_valid & OBD_MD_FLRMTPERM) {
1011 void *buf = req_capsule_server_get(pill, &RMF_ACL);
1013 /* mdt_getattr_lock only */
1014 rc = mdt_pack_remote_perm(info, o, buf);
1016 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
1017 repbody->mbo_aclsize = 0;
1020 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
1021 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
1024 #ifdef CONFIG_FS_POSIX_ACL
1025 else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1026 (reqbody->mbo_valid & OBD_MD_FLACL))
1027 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1032 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1037 static int mdt_getattr(struct tgt_session_info *tsi)
1039 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1040 struct mdt_object *obj = info->mti_object;
1041 struct req_capsule *pill = info->mti_pill;
1042 struct mdt_body *reqbody;
1043 struct mdt_body *repbody;
1047 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1049 LASSERT(obj != NULL);
1050 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1052 /* Unlike intent case where we need to pre-fill out buffers early on
1053 * in intent policy for ldlm reasons, here we can have a much better
1054 * guess at EA size by just reading it from disk.
1055 * Exceptions are readdir and (missing) directory striping */
1057 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1058 /* No easy way to know how long is the symlink, but it cannot
1059 * be more than PATH_MAX, so we allocate +1 */
1062 /* A special case for fs ROOT: getattr there might fetch
1063 * default EA for entire fs, not just for this dir!
1065 } else if (lu_fid_eq(mdt_object_fid(obj),
1066 &info->mti_mdt->mdt_md_root_fid) &&
1067 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1068 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1070 /* Should the default strping be bigger, mdt_fix_reply
1071 * will reallocate */
1072 rc = DEF_REP_MD_SIZE;
1074 /* Read the actual EA size from disk */
1075 rc = mdt_attr_get_eabuf_size(info, obj);
1079 GOTO(out, rc = err_serious(rc));
1081 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1083 rc = req_capsule_server_pack(pill);
1084 if (unlikely(rc != 0))
1085 GOTO(out, rc = err_serious(rc));
1087 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1088 LASSERT(repbody != NULL);
1089 repbody->mbo_eadatasize = 0;
1090 repbody->mbo_aclsize = 0;
1092 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1093 rc = mdt_init_ucred(info, reqbody);
1095 rc = mdt_check_ucred(info);
1097 GOTO(out_shrink, rc);
1099 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1101 rc = mdt_getattr_internal(info, obj, 0);
1102 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1103 mdt_exit_ucred(info);
1106 mdt_client_compatibility(info);
1107 rc2 = mdt_fix_reply(info);
1111 mdt_thread_info_fini(info);
1116 * Exchange MOF_LOV_CREATED flags between two objects after a
1117 * layout swap. No assumption is made on whether o1 or o2 have
1118 * created objects or not.
1120 * \param[in,out] o1 First swap layout object
1121 * \param[in,out] o2 Second swap layout object
1123 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1127 mutex_lock(&o1->mot_lov_mutex);
1128 mutex_lock(&o2->mot_lov_mutex);
1130 o1_flags = o1->mot_flags;
1131 o1->mot_flags = (o1->mot_flags & ~MOF_LOV_CREATED) |
1132 (o2->mot_flags & MOF_LOV_CREATED);
1134 o2->mot_flags = (o2->mot_flags & ~MOF_LOV_CREATED) |
1135 (o1_flags & MOF_LOV_CREATED);
1137 mutex_unlock(&o2->mot_lov_mutex);
1138 mutex_unlock(&o1->mot_lov_mutex);
1141 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1143 struct mdt_thread_info *info;
1144 struct ptlrpc_request *req = tgt_ses_req(tsi);
1145 struct obd_export *exp = req->rq_export;
1146 struct mdt_object *o1, *o2, *o;
1147 struct mdt_lock_handle *lh1, *lh2;
1148 struct mdc_swap_layouts *msl;
1152 /* client does not support layout lock, so layout swaping
1154 * FIXME: there is a problem for old clients which don't support
1155 * layout lock yet. If those clients have already opened the file
1156 * they won't be notified at all so that old layout may still be
1157 * used to do IO. This can be fixed after file release is landed by
1158 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1159 if (!exp_connect_layout(exp))
1160 RETURN(-EOPNOTSUPP);
1162 info = tsi2mdt_info(tsi);
1164 if (info->mti_dlm_req != NULL)
1165 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1167 o1 = info->mti_object;
1168 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1169 &info->mti_body->mbo_fid2);
1171 GOTO(out, rc = PTR_ERR(o));
1173 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1174 GOTO(put, rc = -ENOENT);
1176 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1177 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1183 /* permission check. Make sure the calling process having permission
1184 * to write both files. */
1185 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1190 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1195 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1197 GOTO(put, rc = -EPROTO);
1199 lh1 = &info->mti_lh[MDT_LH_NEW];
1200 mdt_lock_reg_init(lh1, LCK_EX);
1201 lh2 = &info->mti_lh[MDT_LH_OLD];
1202 mdt_lock_reg_init(lh2, LCK_EX);
1204 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1205 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1209 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1210 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1214 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1215 mdt_object_child(o2), msl->msl_flags);
1219 mdt_swap_lov_flag(o1, o2);
1222 mdt_object_unlock(info, o2, lh2, rc);
1224 mdt_object_unlock(info, o1, lh1, rc);
1226 mdt_object_put(info->mti_env, o);
1228 mdt_thread_info_fini(info);
1232 static int mdt_raw_lookup(struct mdt_thread_info *info,
1233 struct mdt_object *parent,
1234 const struct lu_name *lname,
1235 struct ldlm_reply *ldlm_rep)
1237 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1241 LASSERT(!info->mti_cross_ref);
1243 /* Only got the fid of this obj by name */
1244 fid_zero(child_fid);
1245 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1246 lname, child_fid, &info->mti_spec);
1248 struct mdt_body *repbody;
1250 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1251 repbody->mbo_fid1 = *child_fid;
1252 repbody->mbo_valid = OBD_MD_FLID;
1253 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1254 } else if (rc == -ENOENT) {
1255 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1262 * UPDATE lock should be taken against parent, and be released before exit;
1263 * child_bits lock should be taken against child, and be returned back:
1264 * (1)normal request should release the child lock;
1265 * (2)intent request will grant the lock to client.
1267 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1268 struct mdt_lock_handle *lhc,
1270 struct ldlm_reply *ldlm_rep)
1272 struct ptlrpc_request *req = mdt_info_req(info);
1273 struct mdt_body *reqbody = NULL;
1274 struct mdt_object *parent = info->mti_object;
1275 struct mdt_object *child;
1276 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1277 struct lu_name *lname = NULL;
1278 struct mdt_lock_handle *lhp = NULL;
1279 struct ldlm_lock *lock;
1286 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1287 LASSERT(ergo(is_resent,
1288 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1293 if (info->mti_cross_ref) {
1294 /* Only getattr on the child. Parent is on another node. */
1295 mdt_set_disposition(info, ldlm_rep,
1296 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1298 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1300 PFID(mdt_object_fid(child)), ldlm_rep);
1302 rc = mdt_check_resent_lock(info, child, lhc);
1305 } else if (rc > 0) {
1306 mdt_lock_handle_init(lhc);
1307 mdt_lock_reg_init(lhc, LCK_PR);
1310 * Object's name is on another MDS, no lookup or layout
1311 * lock is needed here but update lock is.
1313 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1314 MDS_INODELOCK_LAYOUT);
1315 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1317 rc = mdt_object_lock(info, child, lhc, child_bits,
1323 /* Finally, we can get attr for child. */
1324 if (!mdt_object_exists(child)) {
1325 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1327 "remote object doesn't exist.\n");
1328 mdt_object_unlock(info, child, lhc, 1);
1332 rc = mdt_getattr_internal(info, child, 0);
1333 if (unlikely(rc != 0))
1334 mdt_object_unlock(info, child, lhc, 1);
1339 lname = &info->mti_name;
1340 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1342 if (lu_name_is_valid(lname)) {
1343 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1344 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1345 PNAME(lname), ldlm_rep);
1347 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1348 if (unlikely(reqbody == NULL))
1349 RETURN(err_serious(-EPROTO));
1351 *child_fid = reqbody->mbo_fid2;
1353 if (unlikely(!fid_is_sane(child_fid)))
1354 RETURN(err_serious(-EINVAL));
1356 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1358 PFID(mdt_object_fid(parent)),
1359 PFID(&reqbody->mbo_fid2), ldlm_rep);
1362 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1364 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1365 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1367 "Parent doesn't exist!\n");
1371 if (mdt_object_remote(parent)) {
1372 CERROR("%s: parent "DFID" is on remote target\n",
1373 mdt_obd_name(info->mti_mdt),
1374 PFID(mdt_object_fid(parent)));
1378 if (lu_name_is_valid(lname)) {
1379 /* Always allow to lookup ".." */
1380 if (unlikely(lname->ln_namelen == 2 &&
1381 lname->ln_name[0] == '.' &&
1382 lname->ln_name[1] == '.'))
1383 info->mti_spec.sp_permitted = 1;
1385 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1386 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1391 /* step 1: lock parent only if parent is a directory */
1392 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1393 lhp = &info->mti_lh[MDT_LH_PARENT];
1394 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1395 rc = mdt_object_lock(info, parent, lhp,
1396 MDS_INODELOCK_UPDATE,
1398 if (unlikely(rc != 0))
1402 /* step 2: lookup child's fid by name */
1403 fid_zero(child_fid);
1404 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1405 child_fid, &info->mti_spec);
1407 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1410 GOTO(out_parent, rc);
1413 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1416 *step 3: find the child object by fid & lock it.
1417 * regardless if it is local or remote.
1419 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1420 * set parent dir fid the same as child fid in getattr by fid case
1421 * we should not lu_object_find() the object again, could lead
1422 * to hung if there is a concurrent unlink destroyed the object.
1424 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1425 mdt_object_get(info->mti_env, parent);
1428 child = mdt_object_find(info->mti_env, info->mti_mdt,
1432 if (unlikely(IS_ERR(child)))
1433 GOTO(out_parent, rc = PTR_ERR(child));
1435 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1436 rc = mdt_check_resent_lock(info, child, lhc);
1438 GOTO(out_child, rc);
1439 } else if (rc > 0) {
1440 mdt_lock_handle_init(lhc);
1441 mdt_lock_reg_init(lhc, LCK_PR);
1444 if (!mdt_object_exists(child)) {
1445 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1447 "Object doesn't exist!\n");
1448 GOTO(out_child, rc = -ENOENT);
1451 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1452 mdt_object_exists(child) && !mdt_object_remote(child)) {
1453 struct md_attr *ma = &info->mti_attr;
1456 ma->ma_need = MA_INODE;
1457 rc = mdt_attr_get_complex(info, child, ma);
1458 if (unlikely(rc != 0))
1459 GOTO(out_child, rc);
1461 /* If the file has not been changed for some time, we
1462 * return not only a LOOKUP lock, but also an UPDATE
1463 * lock and this might save us RPC on later STAT. For
1464 * directories, it also let negative dentry cache start
1465 * working for this dir. */
1466 if (ma->ma_valid & MA_INODE &&
1467 ma->ma_attr.la_valid & LA_CTIME &&
1468 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1469 ma->ma_attr.la_ctime < cfs_time_current_sec())
1470 child_bits |= MDS_INODELOCK_UPDATE;
1473 /* layout lock must be granted in a best-effort way
1474 * for IT operations */
1475 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1476 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1477 exp_connect_layout(info->mti_exp) &&
1478 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1479 !mdt_object_remote(child) && ldlm_rep != NULL) {
1480 /* try to grant layout lock for regular file. */
1486 child_bits |= MDS_INODELOCK_LAYOUT;
1487 /* try layout lock, it may fail to be granted due to
1488 * contention at LOOKUP or UPDATE */
1489 if (!mdt_object_lock_try(info, child, lhc, child_bits,
1491 child_bits &= ~MDS_INODELOCK_LAYOUT;
1492 LASSERT(child_bits != 0);
1493 rc = mdt_object_lock(info, child, lhc,
1494 child_bits, MDT_CROSS_LOCK);
1499 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1500 * client will enqueue the lock to the remote MDT */
1501 if (mdt_object_remote(child))
1502 child_bits &= ~MDS_INODELOCK_UPDATE;
1503 rc = mdt_object_lock(info, child, lhc, child_bits,
1506 if (unlikely(rc != 0))
1507 GOTO(out_child, rc);
1510 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1512 /* finally, we can get attr for child. */
1513 rc = mdt_getattr_internal(info, child, ma_need);
1514 if (unlikely(rc != 0)) {
1515 mdt_object_unlock(info, child, lhc, 1);
1517 /* Debugging code. */
1518 LDLM_DEBUG(lock, "Returning lock to client");
1519 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1520 &lock->l_resource->lr_name),
1521 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1522 PLDLMRES(lock->l_resource),
1523 PFID(mdt_object_fid(child)));
1526 LDLM_LOCK_PUT(lock);
1530 mdt_object_put(info->mti_env, child);
1533 mdt_object_unlock(info, parent, lhp, 1);
1537 /* normal handler: should release the child lock */
1538 static int mdt_getattr_name(struct tgt_session_info *tsi)
1540 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1541 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1542 struct mdt_body *reqbody;
1543 struct mdt_body *repbody;
1547 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1548 LASSERT(reqbody != NULL);
1549 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1550 LASSERT(repbody != NULL);
1552 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1553 repbody->mbo_eadatasize = 0;
1554 repbody->mbo_aclsize = 0;
1556 rc = mdt_init_ucred(info, reqbody);
1558 GOTO(out_shrink, rc);
1560 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1561 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1562 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1563 lhc->mlh_reg_lh.cookie = 0;
1565 mdt_exit_ucred(info);
1568 mdt_client_compatibility(info);
1569 rc2 = mdt_fix_reply(info);
1572 mdt_thread_info_fini(info);
1576 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1577 void *karg, void *uarg);
1579 static int mdt_set_info(struct tgt_session_info *tsi)
1581 struct ptlrpc_request *req = tgt_ses_req(tsi);
1584 int keylen, vallen, rc = 0;
1588 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1590 DEBUG_REQ(D_HA, req, "no set_info key");
1591 RETURN(err_serious(-EFAULT));
1594 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1597 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1599 DEBUG_REQ(D_HA, req, "no set_info val");
1600 RETURN(err_serious(-EFAULT));
1603 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1606 /* Swab any part of val you need to here */
1607 if (KEY_IS(KEY_READ_ONLY)) {
1608 spin_lock(&req->rq_export->exp_lock);
1610 *exp_connect_flags_ptr(req->rq_export) |=
1613 *exp_connect_flags_ptr(req->rq_export) &=
1614 ~OBD_CONNECT_RDONLY;
1615 spin_unlock(&req->rq_export->exp_lock);
1616 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1617 struct changelog_setinfo *cs = val;
1619 if (vallen != sizeof(*cs)) {
1620 CERROR("%s: bad changelog_clear setinfo size %d\n",
1621 tgt_name(tsi->tsi_tgt), vallen);
1624 if (ptlrpc_req_need_swab(req)) {
1625 __swab64s(&cs->cs_recno);
1626 __swab32s(&cs->cs_id);
1629 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1637 static int mdt_readpage(struct tgt_session_info *tsi)
1639 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1640 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1641 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1642 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1643 struct mdt_body *repbody;
1649 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1650 RETURN(err_serious(-ENOMEM));
1652 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1653 if (repbody == NULL || reqbody == NULL)
1654 RETURN(err_serious(-EFAULT));
1657 * prepare @rdpg before calling lower layers and transfer itself. Here
1658 * reqbody->size contains offset of where to start to read and
1659 * reqbody->nlink contains number bytes to read.
1661 rdpg->rp_hash = reqbody->mbo_size;
1662 if (rdpg->rp_hash != reqbody->mbo_size) {
1663 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1664 rdpg->rp_hash, reqbody->mbo_size);
1668 rdpg->rp_attrs = reqbody->mbo_mode;
1669 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1670 rdpg->rp_attrs |= LUDA_64BITHASH;
1671 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1672 exp_max_brw_size(tsi->tsi_exp));
1673 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1675 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1676 if (rdpg->rp_pages == NULL)
1679 for (i = 0; i < rdpg->rp_npages; ++i) {
1680 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1681 if (rdpg->rp_pages[i] == NULL)
1682 GOTO(free_rdpg, rc = -ENOMEM);
1685 /* call lower layers to fill allocated pages with directory data */
1686 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1688 GOTO(free_rdpg, rc);
1690 /* send pages to client */
1691 rc = tgt_sendpage(tsi, rdpg, rc);
1696 for (i = 0; i < rdpg->rp_npages; i++)
1697 if (rdpg->rp_pages[i] != NULL)
1698 __free_page(rdpg->rp_pages[i]);
1699 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1701 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1707 static int mdt_reint_internal(struct mdt_thread_info *info,
1708 struct mdt_lock_handle *lhc,
1711 struct req_capsule *pill = info->mti_pill;
1712 struct mdt_body *repbody;
1717 rc = mdt_reint_unpack(info, op);
1719 CERROR("Can't unpack reint, rc %d\n", rc);
1720 RETURN(err_serious(rc));
1723 /* for replay (no_create) lmm is not needed, client has it already */
1724 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1725 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1728 /* llog cookies are always 0, the field is kept for compatibility */
1729 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1730 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1732 rc = req_capsule_server_pack(pill);
1734 CERROR("Can't pack response, rc %d\n", rc);
1735 RETURN(err_serious(rc));
1738 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1739 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1741 repbody->mbo_eadatasize = 0;
1742 repbody->mbo_aclsize = 0;
1745 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1747 /* for replay no cookkie / lmm need, because client have this already */
1748 if (info->mti_spec.no_create)
1749 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1750 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1752 rc = mdt_init_ucred_reint(info);
1754 GOTO(out_shrink, rc);
1756 rc = mdt_fix_attr_ucred(info, op);
1758 GOTO(out_ucred, rc = err_serious(rc));
1760 if (mdt_check_resent(info, mdt_reconstruct, lhc)) {
1761 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1762 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1763 GOTO(out_ucred, rc);
1765 rc = mdt_reint_rec(info, lhc);
1768 mdt_exit_ucred(info);
1770 mdt_client_compatibility(info);
1771 rc2 = mdt_fix_reply(info);
1777 static long mdt_reint_opcode(struct ptlrpc_request *req,
1778 const struct req_format **fmt)
1780 struct mdt_device *mdt;
1781 struct mdt_rec_reint *rec;
1784 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1786 opc = rec->rr_opcode;
1787 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1788 if (opc < REINT_MAX && fmt[opc] != NULL)
1789 req_capsule_extend(&req->rq_pill, fmt[opc]);
1791 mdt = mdt_exp2dev(req->rq_export);
1792 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1793 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1794 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1795 opc = err_serious(-EFAULT);
1798 opc = err_serious(-EFAULT);
1803 static int mdt_reint(struct tgt_session_info *tsi)
1807 static const struct req_format *reint_fmts[REINT_MAX] = {
1808 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1809 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1810 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1811 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1812 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1813 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1814 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1815 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
1816 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME
1821 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1823 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1825 * No lock possible here from client to pass it to reint code
1828 rc = mdt_reint_internal(info, NULL, opc);
1829 mdt_thread_info_fini(info);
1834 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1838 /* this should sync the whole device */
1839 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1841 struct dt_device *dt = mdt->mdt_bottom;
1845 rc = dt->dd_ops->dt_sync(env, dt);
1849 /* this should sync this object */
1850 static int mdt_object_sync(struct mdt_thread_info *info)
1852 struct md_object *next;
1856 if (!mdt_object_exists(info->mti_object)) {
1857 CWARN("%s: non existing object "DFID": rc = %d\n",
1858 mdt_obd_name(info->mti_mdt),
1859 PFID(mdt_object_fid(info->mti_object)), -ESTALE);
1862 next = mdt_object_child(info->mti_object);
1863 rc = mo_object_sync(info->mti_env, next);
1868 static int mdt_sync(struct tgt_session_info *tsi)
1870 struct ptlrpc_request *req = tgt_ses_req(tsi);
1871 struct req_capsule *pill = tsi->tsi_pill;
1872 struct mdt_body *body;
1877 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1878 RETURN(err_serious(-ENOMEM));
1880 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
1881 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
1883 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1885 /* sync an object */
1886 rc = mdt_object_sync(info);
1888 const struct lu_fid *fid;
1889 struct lu_attr *la = &info->mti_attr.ma_attr;
1891 info->mti_attr.ma_need = MA_INODE;
1892 info->mti_attr.ma_valid = 0;
1893 rc = mdt_attr_get_complex(info, info->mti_object,
1896 body = req_capsule_server_get(pill,
1898 fid = mdt_object_fid(info->mti_object);
1899 mdt_pack_attr2body(info, body, la, fid);
1902 mdt_thread_info_fini(info);
1905 mdt_counter_incr(req, LPROC_MDT_SYNC);
1911 * Handle quota control requests to consult current usage/limit, but also
1912 * to configure quota enforcement
1914 static int mdt_quotactl(struct tgt_session_info *tsi)
1916 struct obd_export *exp = tsi->tsi_exp;
1917 struct req_capsule *pill = tsi->tsi_pill;
1918 struct obd_quotactl *oqctl, *repoqc;
1920 struct mdt_device *mdt = mdt_exp2dev(exp);
1921 struct lu_device *qmt = mdt->mdt_qmt_dev;
1922 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
1925 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1927 RETURN(err_serious(-EPROTO));
1929 rc = req_capsule_server_pack(pill);
1931 RETURN(err_serious(rc));
1933 switch (oqctl->qc_cmd) {
1934 /* master quotactl */
1937 if (!nodemap_can_setquota(nodemap))
1942 RETURN(-EOPNOTSUPP);
1943 /* slave quotactl */
1948 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1952 /* map uid/gid for remote client */
1954 if (exp_connect_rmtclient(exp)) {
1955 struct lustre_idmap_table *idmap;
1957 idmap = exp->exp_mdt_data.med_idmap;
1959 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1960 oqctl->qc_cmd != Q_GETINFO))
1963 if (oqctl->qc_type == USRQUOTA)
1964 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1966 else if (oqctl->qc_type == GRPQUOTA)
1967 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1972 if (id == CFS_IDMAP_NOTFOUND) {
1973 CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
1978 if (oqctl->qc_type == USRQUOTA)
1979 id = nodemap_map_id(nodemap, NODEMAP_UID,
1980 NODEMAP_CLIENT_TO_FS, id);
1981 else if (oqctl->qc_type == GRPQUOTA)
1982 id = nodemap_map_id(nodemap, NODEMAP_UID,
1983 NODEMAP_CLIENT_TO_FS, id);
1985 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1987 RETURN(err_serious(-EFAULT));
1989 if (oqctl->qc_id != id)
1990 swap(oqctl->qc_id, id);
1992 switch (oqctl->qc_cmd) {
1998 /* forward quotactl request to QMT */
1999 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2004 /* slave quotactl */
2005 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2010 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2014 if (oqctl->qc_id != id)
2015 swap(oqctl->qc_id, id);
2021 /** clone llog ctxt from child (mdd)
2022 * This allows remote llog (replicator) access.
2023 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2024 * context was originally set up, or we can handle them directly.
2025 * I choose the latter, but that means I need any llog
2026 * contexts set up by child to be accessable by the mdt. So we clone the
2027 * context into our context list here.
2029 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2032 struct md_device *next = mdt->mdt_child;
2033 struct llog_ctxt *ctxt;
2036 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2039 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2040 if (rc || ctxt == NULL) {
2044 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2046 CERROR("Can't set mdt ctxt %d\n", rc);
2051 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2052 struct mdt_device *mdt, int idx)
2054 struct llog_ctxt *ctxt;
2056 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2059 /* Put once for the get we just did, and once for the clone */
2060 llog_ctxt_put(ctxt);
2061 llog_ctxt_put(ctxt);
2066 * sec context handlers
2068 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2072 rc = mdt_handle_idmap(tsi);
2074 struct ptlrpc_request *req = tgt_ses_req(tsi);
2077 opc = lustre_msg_get_opc(req->rq_reqmsg);
2078 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2079 sptlrpc_svc_ctx_invalidate(req);
2082 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2088 * quota request handlers
2090 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2092 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2093 struct lu_device *qmt = mdt->mdt_qmt_dev;
2098 RETURN(err_serious(-EOPNOTSUPP));
2100 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2104 struct mdt_object *mdt_object_new(const struct lu_env *env,
2105 struct mdt_device *d,
2106 const struct lu_fid *f)
2108 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2109 struct lu_object *o;
2110 struct mdt_object *m;
2113 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2114 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2115 if (unlikely(IS_ERR(o)))
2116 m = (struct mdt_object *)o;
2122 struct mdt_object *mdt_object_find(const struct lu_env *env,
2123 struct mdt_device *d,
2124 const struct lu_fid *f)
2126 struct lu_object *o;
2127 struct mdt_object *m;
2130 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2131 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2132 if (unlikely(IS_ERR(o)))
2133 m = (struct mdt_object *)o;
2141 * Asyncronous commit for mdt device.
2143 * Pass asynchonous commit call down the MDS stack.
2145 * \param env environment
2146 * \param mdt the mdt device
2148 static void mdt_device_commit_async(const struct lu_env *env,
2149 struct mdt_device *mdt)
2151 struct dt_device *dt = mdt->mdt_bottom;
2154 rc = dt->dd_ops->dt_commit_async(env, dt);
2155 if (unlikely(rc != 0))
2156 CWARN("%s: async commit start failed: rc = %d\n",
2157 mdt_obd_name(mdt), rc);
2161 * Mark the lock as "synchonous".
2163 * Mark the lock to deffer transaction commit to the unlock time.
2165 * \param lock the lock to mark as "synchonous"
2167 * \see mdt_is_lock_sync
2168 * \see mdt_save_lock
2170 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2172 lock->l_ast_data = (void*)1;
2176 * Check whehter the lock "synchonous" or not.
2178 * \param lock the lock to check
2179 * \retval 1 the lock is "synchonous"
2180 * \retval 0 the lock isn't "synchronous"
2182 * \see mdt_set_lock_sync
2183 * \see mdt_save_lock
2185 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2187 return lock->l_ast_data != NULL;
2191 * Blocking AST for mdt locks.
2193 * Starts transaction commit if in case of COS lock conflict or
2194 * deffers such a commit to the mdt_save_lock.
2196 * \param lock the lock which blocks a request or cancelling lock
2197 * \param desc unused
2198 * \param data unused
2199 * \param flag indicates whether this cancelling or blocking callback
2201 * \see ldlm_blocking_ast_nocheck
2203 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2204 void *data, int flag)
2206 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2207 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2211 if (flag == LDLM_CB_CANCELING)
2213 lock_res_and_lock(lock);
2214 if (lock->l_blocking_ast != mdt_blocking_ast) {
2215 unlock_res_and_lock(lock);
2218 if (mdt_cos_is_enabled(mdt) &&
2219 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2220 lock->l_blocking_lock != NULL &&
2221 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2222 mdt_set_lock_sync(lock);
2224 rc = ldlm_blocking_ast_nocheck(lock);
2226 /* There is no lock conflict if l_blocking_lock == NULL,
2227 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2228 * when the last reference to a local lock was released */
2229 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2232 rc = lu_env_init(&env, LCT_LOCAL);
2233 if (unlikely(rc != 0))
2234 CWARN("%s: lu_env initialization failed, cannot "
2235 "start asynchronous commit: rc = %d\n",
2238 mdt_device_commit_async(&env, mdt);
2244 /* Used for cross-MDT lock */
2245 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2246 void *data, int flag)
2248 struct lustre_handle lockh;
2252 case LDLM_CB_BLOCKING:
2253 ldlm_lock2handle(lock, &lockh);
2254 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
2256 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2260 case LDLM_CB_CANCELING:
2261 LDLM_DEBUG(lock, "Revoke remote lock\n");
2269 int mdt_check_resent_lock(struct mdt_thread_info *info,
2270 struct mdt_object *mo,
2271 struct mdt_lock_handle *lhc)
2273 /* the lock might already be gotten in ldlm_handle_enqueue() */
2274 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2275 struct ptlrpc_request *req = mdt_info_req(info);
2276 struct ldlm_lock *lock;
2278 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2279 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2281 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2282 * a resend case, however, it could be already destroyed
2283 * due to client eviction or a raced cancel RPC. */
2284 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
2285 lhc->mlh_reg_lh.cookie);
2289 if (!fid_res_name_eq(mdt_object_fid(mo),
2290 &lock->l_resource->lr_name)) {
2291 CWARN("%s: Although resent, but still not "
2292 "get child lock:"DFID"\n",
2293 info->mti_exp->exp_obd->obd_name,
2294 PFID(mdt_object_fid(mo)));
2295 LDLM_LOCK_PUT(lock);
2298 LDLM_LOCK_PUT(lock);
2304 int mdt_remote_object_lock(struct mdt_thread_info *mti,
2305 struct mdt_object *o, const struct lu_fid *fid,
2306 struct lustre_handle *lh, ldlm_mode_t mode,
2309 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2310 ldlm_policy_data_t *policy = &mti->mti_policy;
2311 struct ldlm_res_id *res_id = &mti->mti_res_id;
2315 LASSERT(mdt_object_remote(o));
2317 fid_build_reg_res_name(fid, res_id);
2319 memset(einfo, 0, sizeof(*einfo));
2320 einfo->ei_type = LDLM_IBITS;
2321 einfo->ei_mode = mode;
2322 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2323 einfo->ei_cb_cp = ldlm_completion_ast;
2324 einfo->ei_enq_slave = 0;
2325 einfo->ei_res_id = res_id;
2327 memset(policy, 0, sizeof(*policy));
2328 policy->l_inodebits.bits = ibits;
2330 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2335 static int mdt_object_local_lock(struct mdt_thread_info *info,
2336 struct mdt_object *o,
2337 struct mdt_lock_handle *lh, __u64 ibits,
2338 bool nonblock, int locality)
2340 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2341 ldlm_policy_data_t *policy = &info->mti_policy;
2342 struct ldlm_res_id *res_id = &info->mti_res_id;
2347 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2348 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2349 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2350 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2352 /* Only enqueue LOOKUP lock for remote object */
2353 if (mdt_object_remote(o))
2354 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2356 if (lh->mlh_type == MDT_PDO_LOCK) {
2357 /* check for exists after object is locked */
2358 if (mdt_object_exists(o) == 0) {
2359 /* Non-existent object shouldn't have PDO lock */
2362 /* Non-dir object shouldn't have PDO lock */
2363 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2368 memset(policy, 0, sizeof(*policy));
2369 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2371 dlmflags = LDLM_FL_ATOMIC_CB;
2373 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2376 * Take PDO lock on whole directory and build correct @res_id for lock
2377 * on part of directory.
2379 if (lh->mlh_pdo_hash != 0) {
2380 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2381 mdt_lock_pdo_mode(info, o, lh);
2382 if (lh->mlh_pdo_mode != LCK_NL) {
2384 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2385 * is never going to be sent to client and we do not
2386 * want it slowed down due to possible cancels.
2388 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2389 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2390 policy, res_id, dlmflags,
2391 info->mti_exp == NULL ? NULL :
2392 &info->mti_exp->exp_handle.h_cookie);
2393 if (unlikely(rc != 0))
2394 GOTO(out_unlock, rc);
2398 * Finish res_id initializing by name hash marking part of
2399 * directory which is taking modification.
2401 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2404 policy->l_inodebits.bits = ibits;
2407 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2408 * going to be sent to client. If it is - mdt_intent_policy() path will
2409 * fix it up and turn FL_LOCAL flag off.
2411 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2412 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2413 info->mti_exp == NULL ? NULL :
2414 &info->mti_exp->exp_handle.h_cookie);
2417 mdt_object_unlock(info, o, lh, 1);
2418 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2419 lh->mlh_pdo_hash != 0 &&
2420 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2421 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2427 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2428 struct mdt_lock_handle *lh, __u64 ibits,
2429 bool nonblock, int locality)
2434 if (!mdt_object_remote(o))
2435 return mdt_object_local_lock(info, o, lh, ibits, nonblock,
2438 if (locality == MDT_LOCAL_LOCK) {
2439 CERROR("%s: try to get local lock for remote object"
2440 DFID".\n", mdt_obd_name(info->mti_mdt),
2441 PFID(mdt_object_fid(o)));
2445 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2446 ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2447 MDS_INODELOCK_XATTR);
2448 if (ibits & MDS_INODELOCK_UPDATE) {
2449 /* Sigh, PDO needs to enqueue 2 locks right now, but
2450 * enqueue RPC can only request 1 lock, to avoid extra
2451 * RPC, so it will instead enqueue EX lock for remote
2452 * object anyway XXX*/
2453 if (lh->mlh_type == MDT_PDO_LOCK &&
2454 lh->mlh_pdo_hash != 0) {
2455 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2456 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2457 PFID(mdt_object_fid(o)));
2458 lh->mlh_pdo_hash = 0;
2459 lh->mlh_rreg_mode = LCK_EX;
2460 lh->mlh_type = MDT_REG_LOCK;
2462 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2465 MDS_INODELOCK_UPDATE);
2470 /* Only enqueue LOOKUP lock for remote object */
2471 if (ibits & MDS_INODELOCK_LOOKUP) {
2472 rc = mdt_object_local_lock(info, o, lh,
2473 MDS_INODELOCK_LOOKUP,
2474 nonblock, locality);
2482 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2483 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2485 return mdt_object_lock_internal(info, o, lh, ibits, false, locality);
2488 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2489 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2491 struct mdt_lock_handle tmp = *lh;
2494 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true, locality);
2502 * Save a lock within request object.
2504 * Keep the lock referenced until whether client ACK or transaction
2505 * commit happens or release the lock immediately depending on input
2506 * parameters. If COS is ON, a write lock is converted to COS lock
2509 * \param info thead info object
2510 * \param h lock handle
2511 * \param mode lock mode
2512 * \param decref force immediate lock releasing
2515 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2516 ldlm_mode_t mode, int decref)
2520 if (lustre_handle_is_used(h)) {
2521 if (decref || !info->mti_has_trans ||
2522 !(mode & (LCK_PW | LCK_EX))){
2523 mdt_fid_unlock(h, mode);
2525 struct mdt_device *mdt = info->mti_mdt;
2526 struct ldlm_lock *lock = ldlm_handle2lock(h);
2527 struct ptlrpc_request *req = mdt_info_req(info);
2530 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2532 /* there is no request if mdt_object_unlock() is called
2533 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2534 if (likely(req != NULL)) {
2535 CDEBUG(D_HA, "request = %p reply state = %p"
2536 " transno = "LPD64"\n", req,
2537 req->rq_reply_state, req->rq_transno);
2538 if (mdt_cos_is_enabled(mdt)) {
2540 ldlm_lock_downgrade(lock, LCK_COS);
2543 ptlrpc_save_lock(req, h, mode, no_ack);
2545 ldlm_lock_decref(h, mode);
2547 if (mdt_is_lock_sync(lock)) {
2548 CDEBUG(D_HA, "found sync-lock,"
2549 " async commit started\n");
2550 mdt_device_commit_async(info->mti_env,
2553 LDLM_LOCK_PUT(lock);
2562 * Unlock mdt object.
2564 * Immeditely release the regular lock and the PDO lock or save the
2565 * lock in request and keep them referenced until client ACK or
2566 * transaction commit.
2568 * \param info thread info object
2569 * \param o mdt object
2570 * \param lh mdt lock handle referencing regular and PDO locks
2571 * \param decref force immediate lock releasing
2573 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2574 struct mdt_lock_handle *lh, int decref)
2578 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2579 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2581 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2582 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2587 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2588 const struct lu_fid *f,
2589 struct mdt_lock_handle *lh,
2592 struct mdt_object *o;
2594 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2598 rc = mdt_object_lock(info, o, lh, ibits,
2601 mdt_object_put(info->mti_env, o);
2608 void mdt_object_unlock_put(struct mdt_thread_info * info,
2609 struct mdt_object * o,
2610 struct mdt_lock_handle *lh,
2613 mdt_object_unlock(info, o, lh, decref);
2614 mdt_object_put(info->mti_env, o);
2618 * Generic code handling requests that have struct mdt_body passed in:
2620 * - extract mdt_body from request and save it in @info, if present;
2622 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2625 * - if HABEO_CORPUS flag is set for this request type check whether object
2626 * actually exists on storage (lu_object_exists()).
2629 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2631 const struct mdt_body *body;
2632 struct mdt_object *obj;
2633 const struct lu_env *env;
2634 struct req_capsule *pill;
2638 env = info->mti_env;
2639 pill = info->mti_pill;
2641 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2645 if (!(body->mbo_valid & OBD_MD_FLID))
2648 if (!fid_is_sane(&body->mbo_fid1)) {
2649 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2653 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2655 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2656 mdt_object_put(env, obj);
2659 info->mti_object = obj;
2668 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2670 struct req_capsule *pill = info->mti_pill;
2674 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2675 rc = mdt_body_unpack(info, flags);
2679 if (rc == 0 && (flags & HABEO_REFERO)) {
2681 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2682 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2684 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2685 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2688 rc = req_capsule_server_pack(pill);
2693 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2695 lh->mlh_type = MDT_NUL_LOCK;
2696 lh->mlh_reg_lh.cookie = 0ull;
2697 lh->mlh_reg_mode = LCK_MINMODE;
2698 lh->mlh_pdo_lh.cookie = 0ull;
2699 lh->mlh_pdo_mode = LCK_MINMODE;
2700 lh->mlh_rreg_lh.cookie = 0ull;
2701 lh->mlh_rreg_mode = LCK_MINMODE;
2704 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2706 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2707 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2711 * Initialize fields of struct mdt_thread_info. Other fields are left in
2712 * uninitialized state, because it's too expensive to zero out whole
2713 * mdt_thread_info (> 1K) on each request arrival.
2715 void mdt_thread_info_init(struct ptlrpc_request *req,
2716 struct mdt_thread_info *info)
2720 info->mti_pill = &req->rq_pill;
2723 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2724 mdt_lock_handle_init(&info->mti_lh[i]);
2726 /* mdt device: it can be NULL while CONNECT */
2727 if (req->rq_export) {
2728 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2729 info->mti_exp = req->rq_export;
2731 info->mti_mdt = NULL;
2732 info->mti_env = req->rq_svc_thread->t_env;
2733 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2735 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2736 info->mti_big_buf = LU_BUF_NULL;
2737 info->mti_body = NULL;
2738 info->mti_object = NULL;
2739 info->mti_dlm_req = NULL;
2740 info->mti_has_trans = 0;
2741 info->mti_cross_ref = 0;
2742 info->mti_opdata = 0;
2743 info->mti_big_lmm_used = 0;
2745 info->mti_spec.no_create = 0;
2746 info->mti_spec.sp_rm_entry = 0;
2747 info->mti_spec.sp_permitted = 0;
2749 info->mti_spec.u.sp_ea.eadata = NULL;
2750 info->mti_spec.u.sp_ea.eadatalen = 0;
2753 void mdt_thread_info_fini(struct mdt_thread_info *info)
2757 if (info->mti_object != NULL) {
2758 mdt_object_put(info->mti_env, info->mti_object);
2759 info->mti_object = NULL;
2762 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2763 mdt_lock_handle_fini(&info->mti_lh[i]);
2764 info->mti_env = NULL;
2765 info->mti_pill = NULL;
2766 info->mti_exp = NULL;
2768 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2769 lu_buf_free(&info->mti_big_buf);
2772 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2774 struct mdt_thread_info *mti;
2776 mti = mdt_th_info(tsi->tsi_env);
2777 LASSERT(mti != NULL);
2779 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2780 if (tsi->tsi_corpus != NULL) {
2781 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2782 lu_object_get(tsi->tsi_corpus);
2784 mti->mti_body = tsi->tsi_mdt_body;
2785 mti->mti_dlm_req = tsi->tsi_dlm_req;
2790 static int mdt_tgt_connect(struct tgt_session_info *tsi)
2792 struct ptlrpc_request *req = tgt_ses_req(tsi);
2797 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
2799 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
2800 set_current_state(TASK_UNINTERRUPTIBLE);
2801 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
2804 rc = tgt_connect(tsi);
2808 rc = mdt_init_idmap(tsi);
2813 obd_disconnect(class_export_get(req->rq_export));
2832 static int mdt_intent_getattr(enum mdt_it_code opcode,
2833 struct mdt_thread_info *info,
2834 struct ldlm_lock **,
2837 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2838 struct mdt_thread_info *info,
2839 struct ldlm_lock **lockp,
2842 static int mdt_intent_layout(enum mdt_it_code opcode,
2843 struct mdt_thread_info *info,
2844 struct ldlm_lock **,
2846 static int mdt_intent_reint(enum mdt_it_code opcode,
2847 struct mdt_thread_info *info,
2848 struct ldlm_lock **,
2851 static struct mdt_it_flavor {
2852 const struct req_format *it_fmt;
2854 int (*it_act)(enum mdt_it_code ,
2855 struct mdt_thread_info *,
2856 struct ldlm_lock **,
2859 } mdt_it_flavor[] = {
2861 .it_fmt = &RQF_LDLM_INTENT,
2862 /*.it_flags = HABEO_REFERO,*/
2864 .it_act = mdt_intent_reint,
2865 .it_reint = REINT_OPEN
2868 .it_fmt = &RQF_LDLM_INTENT,
2870 * OCREAT is not a MUTABOR request as if the file
2872 * We do the extra check of OBD_CONNECT_RDONLY in
2873 * mdt_reint_open() when we really need to create
2877 .it_act = mdt_intent_reint,
2878 .it_reint = REINT_OPEN
2881 .it_fmt = &RQF_LDLM_INTENT,
2882 .it_flags = MUTABOR,
2883 .it_act = mdt_intent_reint,
2884 .it_reint = REINT_CREATE
2886 [MDT_IT_GETATTR] = {
2887 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2888 .it_flags = HABEO_REFERO,
2889 .it_act = mdt_intent_getattr
2891 [MDT_IT_READDIR] = {
2897 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2898 .it_flags = HABEO_REFERO,
2899 .it_act = mdt_intent_getattr
2902 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2903 .it_flags = MUTABOR,
2905 .it_reint = REINT_UNLINK
2909 .it_flags = MUTABOR,
2912 [MDT_IT_GETXATTR] = {
2913 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2914 .it_flags = HABEO_CORPUS,
2915 .it_act = mdt_intent_getxattr
2918 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2920 .it_act = mdt_intent_layout
2925 mdt_intent_lock_replace(struct mdt_thread_info *info,
2926 struct ldlm_lock **lockp,
2927 struct mdt_lock_handle *lh,
2930 struct ptlrpc_request *req = mdt_info_req(info);
2931 struct ldlm_lock *lock = *lockp;
2932 struct ldlm_lock *new_lock;
2934 /* If possible resent found a lock, @lh is set to its handle */
2935 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2937 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2938 lh->mlh_reg_lh.cookie = 0;
2942 LASSERTF(new_lock != NULL,
2943 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2946 * If we've already given this lock to a client once, then we should
2947 * have no readers or writers. Otherwise, we should have one reader
2948 * _or_ writer ref (which will be zeroed below) before returning the
2951 if (new_lock->l_export == req->rq_export) {
2952 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2954 LASSERT(new_lock->l_export == NULL);
2955 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2960 if (new_lock->l_export == req->rq_export) {
2962 * Already gave this to the client, which means that we
2963 * reconstructed a reply.
2965 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2968 LDLM_LOCK_RELEASE(new_lock);
2969 lh->mlh_reg_lh.cookie = 0;
2970 RETURN(ELDLM_LOCK_REPLACED);
2974 * Fixup the lock to be given to the client.
2976 lock_res_and_lock(new_lock);
2977 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2978 * possible blocking AST. */
2979 while (new_lock->l_readers > 0) {
2980 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2981 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2982 new_lock->l_readers--;
2984 while (new_lock->l_writers > 0) {
2985 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2986 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2987 new_lock->l_writers--;
2990 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2991 new_lock->l_blocking_ast = lock->l_blocking_ast;
2992 new_lock->l_completion_ast = lock->l_completion_ast;
2993 new_lock->l_remote_handle = lock->l_remote_handle;
2994 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2996 unlock_res_and_lock(new_lock);
2998 cfs_hash_add(new_lock->l_export->exp_lock_hash,
2999 &new_lock->l_remote_handle,
3000 &new_lock->l_exp_hash);
3002 LDLM_LOCK_RELEASE(new_lock);
3003 lh->mlh_reg_lh.cookie = 0;
3005 RETURN(ELDLM_LOCK_REPLACED);
3008 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3009 struct ldlm_lock *new_lock,
3010 struct mdt_lock_handle *lh,
3013 struct ptlrpc_request *req = mdt_info_req(info);
3014 struct ldlm_request *dlmreq;
3016 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3019 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3021 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3022 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3024 if (flags & LDLM_FL_RESENT) {
3025 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3026 lh->mlh_reg_mode = new_lock->l_granted_mode;
3028 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3029 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3030 lh->mlh_reg_lh.cookie);
3035 * If the xid matches, then we know this is a resent request, and allow
3036 * it. (It's probably an OPEN, for which we don't send a lock.
3038 if (req_xid_is_last(req))
3042 * This remote handle isn't enqueued, so we never received or processed
3043 * this request. Clear MSG_RESENT, because it can be handled like any
3044 * normal request now.
3046 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3048 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3049 dlmreq->lock_handle[0].cookie);
3052 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3053 struct mdt_thread_info *info,
3054 struct ldlm_lock **lockp,
3057 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3058 struct ldlm_reply *ldlm_rep = NULL;
3062 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3063 * (for the resend case) or a new lock. Below we will use it to
3064 * replace the original lock.
3066 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3067 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3068 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3069 rc = mdt_object_lock(info, info->mti_object, lhc,
3070 MDS_INODELOCK_XATTR,
3076 grc = mdt_getxattr(info);
3078 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3080 if (mdt_info_req(info)->rq_repmsg != NULL)
3081 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3082 if (ldlm_rep == NULL)
3083 RETURN(err_serious(-EFAULT));
3085 ldlm_rep->lock_policy_res2 = grc;
3090 static int mdt_intent_getattr(enum mdt_it_code opcode,
3091 struct mdt_thread_info *info,
3092 struct ldlm_lock **lockp,
3095 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3097 struct ldlm_reply *ldlm_rep;
3098 struct mdt_body *reqbody;
3099 struct mdt_body *repbody;
3103 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3106 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3109 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3110 repbody->mbo_eadatasize = 0;
3111 repbody->mbo_aclsize = 0;
3115 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3117 case MDT_IT_GETATTR:
3118 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3122 CERROR("Unsupported intent (%d)\n", opcode);
3123 GOTO(out_shrink, rc = -EINVAL);
3126 rc = mdt_init_ucred(info, reqbody);
3128 GOTO(out_shrink, rc);
3130 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3131 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3133 /* Get lock from request for possible resent case. */
3134 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3136 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3137 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3139 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3140 ldlm_rep->lock_policy_res2 = 0;
3141 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3142 ldlm_rep->lock_policy_res2) {
3143 lhc->mlh_reg_lh.cookie = 0ull;
3144 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3147 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3150 mdt_exit_ucred(info);
3152 mdt_client_compatibility(info);
3153 rc2 = mdt_fix_reply(info);
3159 static int mdt_intent_layout(enum mdt_it_code opcode,
3160 struct mdt_thread_info *info,
3161 struct ldlm_lock **lockp,
3164 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3165 struct layout_intent *layout;
3167 struct mdt_object *obj = NULL;
3171 if (opcode != MDT_IT_LAYOUT) {
3172 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3177 fid = &info->mti_tmp_fid2;
3178 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3180 /* Get lock from request for possible resent case. */
3181 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3183 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3185 RETURN(PTR_ERR(obj));
3187 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3188 /* get the length of lsm */
3189 rc = mdt_attr_get_eabuf_size(info, obj);
3191 mdt_object_put(info->mti_env, obj);
3195 if (rc > info->mti_mdt->mdt_max_mdsize)
3196 info->mti_mdt->mdt_max_mdsize = rc;
3199 mdt_object_put(info->mti_env, obj);
3201 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3202 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER, rc);
3203 rc = req_capsule_server_pack(info->mti_pill);
3207 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
3208 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3210 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3211 LASSERT(layout != NULL);
3212 if (layout->li_opc == LAYOUT_INTENT_ACCESS)
3213 /* return to normal/resent ldlm handling */
3216 CERROR("%s: Unsupported layout intent (%d)\n",
3217 mdt_obd_name(info->mti_mdt), layout->li_opc);
3221 static int mdt_intent_reint(enum mdt_it_code opcode,
3222 struct mdt_thread_info *info,
3223 struct ldlm_lock **lockp,
3226 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3227 struct ldlm_reply *rep = NULL;
3231 static const struct req_format *intent_fmts[REINT_MAX] = {
3232 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3233 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3238 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3242 if (mdt_it_flavor[opcode].it_reint != opc) {
3243 CERROR("Reint code %ld doesn't match intent: %d\n",
3245 RETURN(err_serious(-EPROTO));
3248 /* Get lock from request for possible resent case. */
3249 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3251 rc = mdt_reint_internal(info, lhc, opc);
3253 /* Check whether the reply has been packed successfully. */
3254 if (mdt_info_req(info)->rq_repmsg != NULL)
3255 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3257 RETURN(err_serious(-EFAULT));
3259 /* MDC expects this in any case */
3261 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3263 /* the open lock or the lock for cross-ref object should be
3264 * returned to the client */
3265 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3266 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3267 rep->lock_policy_res2 = 0;
3268 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3272 rep->lock_policy_res2 = clear_serious(rc);
3274 if (rep->lock_policy_res2 == -ENOENT &&
3275 mdt_get_disposition(rep, DISP_LOOKUP_NEG))
3276 rep->lock_policy_res2 = 0;
3278 lhc->mlh_reg_lh.cookie = 0ull;
3279 if (rc == -ENOTCONN || rc == -ENODEV ||
3280 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3282 * If it is the disconnect error (ENODEV & ENOCONN), the error
3283 * will be returned by rq_status, and client at ptlrpc layer
3284 * will detect this, then disconnect, reconnect the import
3285 * immediately, instead of impacting the following the rpc.
3290 * For other cases, the error will be returned by intent, and client
3291 * will retrieve the result from intent.
3293 RETURN(ELDLM_LOCK_ABORTED);
3296 static int mdt_intent_code(long itcode)
3304 case IT_OPEN|IT_CREAT:
3311 rc = MDT_IT_READDIR;
3314 rc = MDT_IT_GETATTR;
3326 rc = MDT_IT_GETXATTR;
3331 case IT_QUOTA_DQACQ:
3336 CERROR("Unknown intent opcode: %ld\n", itcode);
3343 static int mdt_intent_opc(long itopc, struct mdt_thread_info *info,
3344 struct ldlm_lock **lockp, __u64 flags)
3346 struct req_capsule *pill = info->mti_pill;
3347 struct ptlrpc_request *req = mdt_info_req(info);
3348 struct mdt_it_flavor *flv;
3353 opc = mdt_intent_code(itopc);
3357 if (opc == MDT_IT_QUOTA) {
3358 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3361 RETURN(-EOPNOTSUPP);
3363 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3364 /* pass the request to quota master */
3365 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3366 mdt_info_req(info), lockp,
3371 flv = &mdt_it_flavor[opc];
3372 if (flv->it_fmt != NULL)
3373 req_capsule_extend(pill, flv->it_fmt);
3375 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3379 if (flv->it_flags & MUTABOR &&
3380 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3383 if (flv->it_act != NULL) {
3384 struct ldlm_reply *rep;
3386 /* execute policy */
3387 rc = flv->it_act(opc, info, lockp, flags);
3389 /* Check whether the reply has been packed successfully. */
3390 if (req->rq_repmsg != NULL) {
3391 rep = req_capsule_server_get(info->mti_pill,
3393 rep->lock_policy_res2 =
3394 ptlrpc_status_hton(rep->lock_policy_res2);
3401 static int mdt_intent_policy(struct ldlm_namespace *ns,
3402 struct ldlm_lock **lockp, void *req_cookie,
3403 ldlm_mode_t mode, __u64 flags, void *data)
3405 struct tgt_session_info *tsi;
3406 struct mdt_thread_info *info;
3407 struct ptlrpc_request *req = req_cookie;
3408 struct ldlm_intent *it;
3409 struct req_capsule *pill;
3414 LASSERT(req != NULL);
3416 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3418 info = tsi2mdt_info(tsi);
3419 LASSERT(info != NULL);
3420 pill = info->mti_pill;
3421 LASSERT(pill->rc_req == req);
3423 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3424 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3425 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3427 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3431 /* Lock without inodebits makes no sense and will oops
3432 * later in ldlm. Let's check it now to see if we have
3433 * ibits corrupted somewhere in mdt_intent_opc().
3434 * The case for client miss to set ibits has been
3435 * processed by others. */
3436 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3437 lr_type == LDLM_IBITS,
3438 info->mti_dlm_req->lock_desc.\
3439 l_policy_data.l_inodebits.bits != 0));
3441 rc = err_serious(-EFAULT);
3443 /* No intent was provided */
3444 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3445 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3446 rc = req_capsule_server_pack(pill);
3448 rc = err_serious(rc);
3450 mdt_thread_info_fini(info);
3454 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3456 struct seq_server_site *ss = mdt_seq_site(mdt);
3458 if (ss->ss_node_id == 0)
3461 if (ss->ss_client_seq != NULL) {
3462 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3463 ss->ss_client_seq->lcs_exp = NULL;
3466 if (ss->ss_server_fld != NULL) {
3467 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3468 ss->ss_server_fld->lsf_control_exp = NULL;
3472 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3474 struct seq_server_site *ss = mdt_seq_site(mdt);
3479 if (ss->ss_server_seq != NULL)
3480 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3483 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3485 mdt_seq_fini_cli(mdt);
3486 mdt_deregister_seq_exp(mdt);
3488 return seq_site_fini(env, mdt_seq_site(mdt));
3492 * It will retrieve its FLDB entries from MDT0, and it only happens
3493 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3494 * and it needs to refresh FLDB from the MDT0.
3496 static int mdt_register_lwp_callback(void *data)
3499 struct mdt_device *mdt = data;
3500 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3504 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3506 rc = lu_env_init(&env, LCT_MD_THREAD);
3508 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3512 /* Allocate new sequence now to avoid creating local transaction
3513 * in the normal transaction process */
3514 rc = seq_server_check_and_alloc_super(&env,
3515 mdt_seq_site(mdt)->ss_server_seq);
3520 rc = fld_update_from_controller(&env, fld);
3522 CERROR("%s: cannot update controller: rc = %d\n",
3523 mdt_obd_name(mdt), rc);
3532 static int mdt_register_seq_exp(struct mdt_device *mdt)
3534 struct seq_server_site *ss = mdt_seq_site(mdt);
3535 char *lwp_name = NULL;
3538 if (ss->ss_node_id == 0)
3541 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3542 if (lwp_name == NULL)
3543 GOTO(out_free, rc = -ENOMEM);
3545 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3549 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3554 rc = lustre_register_lwp_item(lwp_name,
3555 &ss->ss_server_fld->lsf_control_exp,
3556 mdt_register_lwp_callback, mdt);
3558 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3559 ss->ss_client_seq->lcs_exp = NULL;
3563 if (lwp_name != NULL)
3564 OBD_FREE(lwp_name, MAX_OBD_NAME);
3570 * Init client sequence manager which is used by local MDS to talk to sequence
3571 * controller on remote node.
3573 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3575 struct seq_server_site *ss = mdt_seq_site(mdt);
3580 /* check if this is adding the first MDC and controller is not yet
3582 OBD_ALLOC_PTR(ss->ss_client_seq);
3583 if (ss->ss_client_seq == NULL)
3586 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3587 if (prefix == NULL) {
3588 OBD_FREE_PTR(ss->ss_client_seq);
3589 ss->ss_client_seq = NULL;
3593 /* Note: seq_client_fini will be called in seq_site_fini */
3594 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3595 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3596 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3598 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3600 OBD_FREE_PTR(ss->ss_client_seq);
3601 ss->ss_client_seq = NULL;
3605 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3610 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3612 struct seq_server_site *ss;
3616 ss = mdt_seq_site(mdt);
3617 /* init sequence controller server(MDT0) */
3618 if (ss->ss_node_id == 0) {
3619 OBD_ALLOC_PTR(ss->ss_control_seq);
3620 if (ss->ss_control_seq == NULL)
3623 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3624 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3627 GOTO(out_seq_fini, rc);
3630 /* Init normal sequence server */
3631 OBD_ALLOC_PTR(ss->ss_server_seq);
3632 if (ss->ss_server_seq == NULL)
3633 GOTO(out_seq_fini, rc = -ENOMEM);
3635 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3636 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3638 GOTO(out_seq_fini, rc);
3640 /* init seq client for seq server to talk to seq controller(MDT0) */
3641 rc = mdt_seq_init_cli(env, mdt);
3643 GOTO(out_seq_fini, rc);
3645 if (ss->ss_node_id != 0)
3646 /* register controller export through lwp */
3647 rc = mdt_register_seq_exp(mdt);
3652 mdt_seq_fini(env, mdt);
3660 static int mdt_fld_fini(const struct lu_env *env,
3661 struct mdt_device *m)
3663 struct seq_server_site *ss = mdt_seq_site(m);
3666 if (ss && ss->ss_server_fld) {
3667 fld_server_fini(env, ss->ss_server_fld);
3668 OBD_FREE_PTR(ss->ss_server_fld);
3669 ss->ss_server_fld = NULL;
3675 static int mdt_fld_init(const struct lu_env *env,
3677 struct mdt_device *m)
3679 struct seq_server_site *ss;
3683 ss = mdt_seq_site(m);
3685 OBD_ALLOC_PTR(ss->ss_server_fld);
3686 if (ss->ss_server_fld == NULL)
3687 RETURN(rc = -ENOMEM);
3689 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3692 OBD_FREE_PTR(ss->ss_server_fld);
3693 ss->ss_server_fld = NULL;
3700 static void mdt_stack_pre_fini(const struct lu_env *env,
3701 struct mdt_device *m, struct lu_device *top)
3703 struct lustre_cfg_bufs *bufs;
3704 struct lustre_cfg *lcfg;
3705 struct mdt_thread_info *info;
3710 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3711 LASSERT(info != NULL);
3713 bufs = &info->mti_u.bufs;
3715 LASSERT(m->mdt_child_exp);
3716 LASSERT(m->mdt_child_exp->exp_obd);
3718 /* process cleanup, pass mdt obd name to get obd umount flags */
3719 /* XXX: this is needed because all layers are referenced by
3720 * objects (some of them are pinned by osd, for example *
3721 * the proper solution should be a model where object used
3722 * by osd only doesn't have mdt/mdd slices -bzzz */
3723 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3724 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3725 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3729 top->ld_ops->ldo_process_config(env, top, lcfg);
3730 lustre_cfg_free(lcfg);
3734 static void mdt_stack_fini(const struct lu_env *env,
3735 struct mdt_device *m, struct lu_device *top)
3737 struct obd_device *obd = mdt2obd_dev(m);
3738 struct lustre_cfg_bufs *bufs;
3739 struct lustre_cfg *lcfg;
3740 struct mdt_thread_info *info;
3744 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3745 LASSERT(info != NULL);
3747 lu_dev_del_linkage(top->ld_site, top);
3749 lu_site_purge(env, top->ld_site, -1);
3751 bufs = &info->mti_u.bufs;
3752 /* process cleanup, pass mdt obd name to get obd umount flags */
3753 /* another purpose is to let all layers to release their objects */
3754 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3759 lustre_cfg_bufs_set_string(bufs, 1, flags);
3760 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3765 top->ld_ops->ldo_process_config(env, top, lcfg);
3766 lustre_cfg_free(lcfg);
3768 lu_site_purge(env, top->ld_site, -1);
3770 m->mdt_child = NULL;
3771 m->mdt_bottom = NULL;
3773 obd_disconnect(m->mdt_child_exp);
3774 m->mdt_child_exp = NULL;
3776 obd_disconnect(m->mdt_bottom_exp);
3777 m->mdt_child_exp = NULL;
3780 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3781 const char *next, struct obd_export **exp)
3783 struct obd_connect_data *data = NULL;
3784 struct obd_device *obd;
3788 OBD_ALLOC_PTR(data);
3790 GOTO(out, rc = -ENOMEM);
3792 obd = class_name2obd(next);
3794 CERROR("%s: can't locate next device: %s\n",
3795 mdt_obd_name(m), next);
3796 GOTO(out, rc = -ENOTCONN);
3799 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3800 data->ocd_version = LUSTRE_VERSION_CODE;
3802 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3804 CERROR("%s: cannot connect to next dev %s (%d)\n",
3805 mdt_obd_name(m), next, rc);
3815 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3816 struct lustre_cfg *cfg)
3818 char *dev = lustre_cfg_string(cfg, 0);
3819 int rc, name_size, uuid_size;
3820 char *name, *uuid, *p;
3821 struct lustre_cfg_bufs *bufs;
3822 struct lustre_cfg *lcfg;
3823 struct obd_device *obd;
3824 struct lustre_profile *lprof;
3825 struct lu_site *site;
3828 /* in 1.8 we had the only device in the stack - MDS.
3829 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3830 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3831 * to generate names and setup MDT, MDD. MDT will be using
3832 * generated name to connect to MDD. for MDD the next device
3833 * will be LOD with name taken from so called "profile" which
3834 * is generated by mount_option line
3836 * 1.8 MGS generates config. commands like this:
3837 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3838 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3839 * 2.0 MGS generates config. commands like this:
3840 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3841 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3842 * 3:lustre-MDT0000-mdtlov 4:f
3844 * we generate MDD name from MDT one, just replacing T with D
3846 * after all the preparations, the logical equivalent will be
3847 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3848 * 3:lustre-MDT0000-mdtlov 4:f
3849 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3850 * 3:lustre-MDD0000 4:f
3852 * notice we build the stack from down to top: MDD first, then MDT */
3854 name_size = MAX_OBD_NAME;
3855 uuid_size = MAX_OBD_NAME;
3857 OBD_ALLOC(name, name_size);
3858 OBD_ALLOC(uuid, uuid_size);
3859 if (name == NULL || uuid == NULL)
3860 GOTO(cleanup_mem, rc = -ENOMEM);
3862 OBD_ALLOC_PTR(bufs);
3864 GOTO(cleanup_mem, rc = -ENOMEM);
3867 p = strstr(name, "-MDT");
3869 GOTO(free_bufs, rc = -ENOMEM);
3872 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3874 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3875 if (lprof == NULL || lprof->lp_dt == NULL) {
3876 CERROR("can't find the profile: %s\n",
3877 lustre_cfg_string(cfg, 0));
3878 GOTO(free_bufs, rc = -EINVAL);
3881 lustre_cfg_bufs_reset(bufs, name);
3882 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3883 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3884 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3886 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3888 GOTO(free_bufs, rc = -ENOMEM);
3890 rc = class_attach(lcfg);
3892 GOTO(lcfg_cleanup, rc);
3894 obd = class_name2obd(name);
3896 CERROR("Can not find obd %s (%s in config)\n",
3897 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3898 GOTO(lcfg_cleanup, rc = -EINVAL);
3901 lustre_cfg_free(lcfg);
3903 lustre_cfg_bufs_reset(bufs, name);
3904 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3905 lustre_cfg_bufs_set_string(bufs, 2, dev);
3906 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3908 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3910 GOTO(class_detach, rc = -ENOMEM);
3912 rc = class_setup(obd, lcfg);
3914 GOTO(class_detach, rc);
3916 /* connect to MDD we just setup */
3917 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3919 GOTO(class_detach, rc);
3921 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3923 LASSERT(mdt_lu_site(mdt) == NULL);
3924 mdt->mdt_lu_dev.ld_site = site;
3925 site->ls_top_dev = &mdt->mdt_lu_dev;
3926 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3928 /* now connect to bottom OSD */
3929 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3930 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3932 GOTO(class_detach, rc);
3934 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3936 rc = lu_env_refill((struct lu_env *)env);
3938 CERROR("Failure to refill session: '%d'\n", rc);
3940 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3945 class_detach(obd, lcfg);
3947 lustre_cfg_free(lcfg);
3952 OBD_FREE(name, name_size);
3954 OBD_FREE(uuid, uuid_size);
3958 /* setup quota master target on MDT0 */
3959 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3960 struct lustre_cfg *cfg)
3962 struct obd_device *obd;
3963 char *dev = lustre_cfg_string(cfg, 0);
3964 char *qmtname, *uuid, *p;
3965 struct lustre_cfg_bufs *bufs;
3966 struct lustre_cfg *lcfg;
3967 struct lustre_profile *lprof;
3968 struct obd_connect_data *data;
3972 LASSERT(mdt->mdt_qmt_exp == NULL);
3973 LASSERT(mdt->mdt_qmt_dev == NULL);
3975 /* quota master is on MDT0 only for now */
3976 if (mdt->mdt_seq_site.ss_node_id != 0)
3979 /* MGS generates config commands which look as follows:
3980 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3981 * 3:lustre-MDT0000-mdtlov 4:f
3983 * We generate the QMT name from the MDT one, just replacing MD with QM
3984 * after all the preparations, the logical equivalent will be:
3985 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
3986 * 3:lustre-MDT0000-osd 4:f */
3987 OBD_ALLOC(qmtname, MAX_OBD_NAME);
3988 OBD_ALLOC(uuid, UUID_MAX);
3989 OBD_ALLOC_PTR(bufs);
3990 OBD_ALLOC_PTR(data);
3991 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
3992 GOTO(cleanup_mem, rc = -ENOMEM);
3994 strcpy(qmtname, dev);
3995 p = strstr(qmtname, "-MDT");
3997 GOTO(cleanup_mem, rc = -ENOMEM);
3998 /* replace MD with QM */
4002 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4004 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4005 if (lprof == NULL || lprof->lp_dt == NULL) {
4006 CERROR("can't find profile for %s\n",
4007 lustre_cfg_string(cfg, 0));
4008 GOTO(cleanup_mem, rc = -EINVAL);
4011 lustre_cfg_bufs_reset(bufs, qmtname);
4012 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4013 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4014 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4016 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4018 GOTO(cleanup_mem, rc = -ENOMEM);
4020 rc = class_attach(lcfg);
4022 GOTO(lcfg_cleanup, rc);
4024 obd = class_name2obd(qmtname);
4026 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4027 lustre_cfg_string(cfg, 0));
4028 GOTO(lcfg_cleanup, rc = -EINVAL);
4031 lustre_cfg_free(lcfg);
4033 lustre_cfg_bufs_reset(bufs, qmtname);
4034 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4035 lustre_cfg_bufs_set_string(bufs, 2, dev);
4037 /* for quota, the next device should be the OSD device */
4038 lustre_cfg_bufs_set_string(bufs, 3,
4039 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4041 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4043 GOTO(class_detach, rc = -ENOMEM);
4045 rc = class_setup(obd, lcfg);
4047 GOTO(class_detach, rc);
4049 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4051 /* configure local quota objects */
4052 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4056 GOTO(class_cleanup, rc);
4058 /* connect to quota master target */
4059 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4060 data->ocd_version = LUSTRE_VERSION_CODE;
4061 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4064 CERROR("cannot connect to quota master device %s (%d)\n",
4066 GOTO(class_cleanup, rc);
4072 class_manual_cleanup(obd);
4073 mdt->mdt_qmt_dev = NULL;
4077 class_detach(obd, lcfg);
4079 lustre_cfg_free(lcfg);
4084 OBD_FREE(qmtname, MAX_OBD_NAME);
4086 OBD_FREE(uuid, UUID_MAX);
4092 /* Shutdown quota master target associated with mdt */
4093 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4097 if (mdt->mdt_qmt_exp == NULL)
4099 LASSERT(mdt->mdt_qmt_dev != NULL);
4101 /* the qmt automatically shuts down when the mdt disconnects */
4102 obd_disconnect(mdt->mdt_qmt_exp);
4103 mdt->mdt_qmt_exp = NULL;
4104 mdt->mdt_qmt_dev = NULL;
4108 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4109 * for now. This will be removed along with converting rest of MDT code
4110 * to use tgt_session_info */
4111 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4113 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4116 rc = mdt_getxattr(info);
4118 mdt_thread_info_fini(info);
4122 static struct tgt_handler mdt_tgt_handlers[] = {
4123 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4124 0, MDS_CONNECT, mdt_tgt_connect,
4125 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4126 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4127 0, MDS_DISCONNECT, tgt_disconnect,
4128 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4129 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4130 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4131 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4132 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4133 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4134 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4135 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4137 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4138 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4139 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4140 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4141 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4142 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4143 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4144 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4146 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4147 mdt_hsm_ct_register),
4148 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4149 mdt_hsm_ct_unregister),
4150 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4152 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4154 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4155 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4157 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4162 static struct tgt_handler mdt_sec_ctx_ops[] = {
4163 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4164 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4165 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4168 static struct tgt_handler mdt_quota_ops[] = {
4169 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4172 static struct tgt_opc_slice mdt_common_slice[] = {
4174 .tos_opc_start = MDS_FIRST_OPC,
4175 .tos_opc_end = MDS_LAST_OPC,
4176 .tos_hs = mdt_tgt_handlers
4179 .tos_opc_start = OBD_FIRST_OPC,
4180 .tos_opc_end = OBD_LAST_OPC,
4181 .tos_hs = tgt_obd_handlers
4184 .tos_opc_start = LDLM_FIRST_OPC,
4185 .tos_opc_end = LDLM_LAST_OPC,
4186 .tos_hs = tgt_dlm_handlers
4189 .tos_opc_start = SEC_FIRST_OPC,
4190 .tos_opc_end = SEC_LAST_OPC,
4191 .tos_hs = mdt_sec_ctx_ops
4194 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4195 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4196 .tos_hs = tgt_out_handlers
4199 .tos_opc_start = FLD_FIRST_OPC,
4200 .tos_opc_end = FLD_LAST_OPC,
4201 .tos_hs = fld_handlers
4204 .tos_opc_start = SEQ_FIRST_OPC,
4205 .tos_opc_end = SEQ_LAST_OPC,
4206 .tos_hs = seq_handlers
4209 .tos_opc_start = QUOTA_DQACQ,
4210 .tos_opc_end = QUOTA_LAST_OPC,
4211 .tos_hs = mdt_quota_ops
4214 .tos_opc_start = LLOG_FIRST_OPC,
4215 .tos_opc_end = LLOG_LAST_OPC,
4216 .tos_hs = tgt_llog_handlers
4219 .tos_opc_start = LFSCK_FIRST_OPC,
4220 .tos_opc_end = LFSCK_LAST_OPC,
4221 .tos_hs = tgt_lfsck_handlers
4229 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4231 struct md_device *next = m->mdt_child;
4232 struct lu_device *d = &m->mdt_lu_dev;
4233 struct obd_device *obd = mdt2obd_dev(m);
4234 struct lfsck_stop stop;
4237 stop.ls_status = LS_PAUSED;
4239 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4241 target_recovery_fini(obd);
4242 ping_evictor_stop();
4243 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4245 if (m->mdt_opts.mo_coordinator)
4246 mdt_hsm_cdt_stop(m);
4248 mdt_hsm_cdt_fini(m);
4250 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4251 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4252 obd_exports_barrier(obd);
4253 obd_zombie_barrier();
4257 tgt_fini(env, &m->mdt_lut);
4258 mdt_fs_cleanup(env, m);
4259 upcall_cache_cleanup(m->mdt_identity_cache);
4260 m->mdt_identity_cache = NULL;
4262 if (m->mdt_namespace != NULL) {
4263 ldlm_namespace_free(m->mdt_namespace, NULL,
4264 d->ld_obd->obd_force);
4265 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4268 mdt_quota_fini(env, m);
4270 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4272 mdt_seq_fini(env, m);
4273 mdt_fld_fini(env, m);
4278 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4280 LASSERT(atomic_read(&d->ld_ref) == 0);
4282 server_put_mount(mdt_obd_name(m), true);
4287 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4289 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4290 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4292 struct mdt_thread_info *info;
4293 struct obd_device *obd;
4294 const char *dev = lustre_cfg_string(cfg, 0);
4295 const char *num = lustre_cfg_string(cfg, 2);
4296 struct lustre_mount_info *lmi = NULL;
4297 struct lustre_sb_info *lsi;
4299 struct seq_server_site *ss_site;
4300 const char *identity_upcall = "NONE";
4301 struct md_device *next;
4307 lu_device_init(&m->mdt_lu_dev, ldt);
4309 * Environment (env) might be missing mdt_thread_key values at that
4310 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4313 * Usually device allocation path doesn't use module key values, but
4314 * mdt has to do a lot of work here, so allocate key value.
4316 rc = lu_env_refill((struct lu_env *)env);
4320 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4321 LASSERT(info != NULL);
4323 obd = class_name2obd(dev);
4324 LASSERT(obd != NULL);
4326 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4328 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4330 /* default is coordinator off, it is started through conf_param
4332 m->mdt_opts.mo_coordinator = 0;
4334 lmi = server_get_mount(dev);
4336 CERROR("Cannot get mount info for %s!\n", dev);
4339 lsi = s2lsi(lmi->lmi_sb);
4340 /* CMD is supported only in IAM mode */
4342 node_id = simple_strtol(num, NULL, 10);
4343 obd->u.obt.obt_magic = OBT_MAGIC;
4346 m->mdt_squash.rsi_uid = 0;
4347 m->mdt_squash.rsi_gid = 0;
4348 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4349 init_rwsem(&m->mdt_squash.rsi_sem);
4350 spin_lock_init(&m->mdt_osfs_lock);
4351 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4352 m->mdt_enable_remote_dir = 0;
4353 m->mdt_enable_remote_dir_gid = 0;
4355 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4356 m->mdt_lu_dev.ld_obd = obd;
4357 /* Set this lu_device to obd for error handling purposes. */
4358 obd->obd_lu_dev = &m->mdt_lu_dev;
4360 /* init the stack */
4361 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4363 CERROR("%s: Can't init device stack, rc %d\n",
4364 mdt_obd_name(m), rc);
4369 ss_site = mdt_seq_site(m);
4370 s->ld_seq_site = ss_site;
4373 /* set server index */
4374 ss_site->ss_node_id = node_id;
4376 /* failover is the default
4377 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4378 * assumed whose ss_node_id == 0 XXX
4380 obd->obd_replayable = 1;
4381 /* No connection accepted until configurations will finish */
4382 obd->obd_no_conn = 1;
4384 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4385 char *str = lustre_cfg_string(cfg, 4);
4386 if (strchr(str, 'n')) {
4387 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4388 obd->obd_replayable = 0;
4392 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4394 GOTO(err_fini_stack, rc);
4396 rc = mdt_seq_init(env, m);
4398 GOTO(err_fini_fld, rc);
4400 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4401 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4402 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4403 LDLM_NAMESPACE_SERVER,
4404 LDLM_NAMESPACE_GREEDY,
4406 if (m->mdt_namespace == NULL)
4407 GOTO(err_fini_seq, rc = -ENOMEM);
4409 m->mdt_namespace->ns_lvbp = m;
4410 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4412 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4413 /* set obd_namespace for compatibility with old code */
4414 obd->obd_namespace = m->mdt_namespace;
4416 rc = mdt_hsm_cdt_init(m);
4418 CERROR("%s: error initializing coordinator, rc %d\n",
4419 mdt_obd_name(m), rc);
4420 GOTO(err_free_ns, rc);
4423 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4424 OBD_FAIL_MDS_ALL_REQUEST_NET,
4425 OBD_FAIL_MDS_ALL_REPLY_NET);
4427 GOTO(err_free_hsm, rc);
4429 rc = mdt_fs_setup(env, m, obd, lsi);
4433 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4435 next = m->mdt_child;
4436 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4439 GOTO(err_fs_cleanup, rc);
4441 if (mntopts & MNTOPT_USERXATTR)
4442 m->mdt_opts.mo_user_xattr = 1;
4444 m->mdt_opts.mo_user_xattr = 0;
4446 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4448 GOTO(err_fs_cleanup, rc);
4450 if (mntopts & MNTOPT_ACL)
4451 m->mdt_opts.mo_acl = 1;
4453 m->mdt_opts.mo_acl = 0;
4455 /* XXX: to support suppgid for ACL, we enable identity_upcall
4456 * by default, otherwise, maybe got unexpected -EACCESS. */
4457 if (m->mdt_opts.mo_acl)
4458 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4460 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4462 &mdt_identity_upcall_cache_ops);
4463 if (IS_ERR(m->mdt_identity_cache)) {
4464 rc = PTR_ERR(m->mdt_identity_cache);
4465 m->mdt_identity_cache = NULL;
4466 GOTO(err_fs_cleanup, rc);
4469 rc = mdt_procfs_init(m, dev);
4471 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4472 GOTO(err_recovery, rc);
4475 rc = mdt_quota_init(env, m, cfg);
4477 GOTO(err_procfs, rc);
4479 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4480 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4481 "mdt_ldlm_client", m->mdt_ldlm_client);
4483 ping_evictor_start();
4485 /* recovery will be started upon mdt_prepare()
4486 * when the whole stack is complete and ready
4487 * to serve the requests */
4489 /* Reduce the initial timeout on an MDS because it doesn't need such
4490 * a long timeout as an OST does. Adaptive timeouts will adjust this
4491 * value appropriately. */
4492 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4493 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4499 target_recovery_fini(obd);
4500 upcall_cache_cleanup(m->mdt_identity_cache);
4501 m->mdt_identity_cache = NULL;
4503 mdt_fs_cleanup(env, m);
4505 tgt_fini(env, &m->mdt_lut);
4507 mdt_hsm_cdt_fini(m);
4509 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4510 obd->obd_namespace = m->mdt_namespace = NULL;
4512 mdt_seq_fini(env, m);
4514 mdt_fld_fini(env, m);
4516 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4519 server_put_mount(dev, true);
4523 /* For interoperability, the left element is old parameter, the right one
4524 * is the new version of the parameter, if some parameter is deprecated,
4525 * the new version should be set as NULL. */
4526 static struct cfg_interop_param mdt_interop_param[] = {
4527 { "mdt.group_upcall", NULL },
4528 { "mdt.quota_type", NULL },
4529 { "mdd.quota_type", NULL },
4530 { "mdt.rootsquash", "mdt.root_squash" },
4531 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4535 /* used by MGS to process specific configurations */
4536 static int mdt_process_config(const struct lu_env *env,
4537 struct lu_device *d, struct lustre_cfg *cfg)
4539 struct mdt_device *m = mdt_dev(d);
4540 struct md_device *md_next = m->mdt_child;
4541 struct lu_device *next = md2lu_dev(md_next);
4545 switch (cfg->lcfg_command) {
4547 struct obd_device *obd = d->ld_obd;
4549 /* For interoperability */
4550 struct cfg_interop_param *ptr = NULL;
4551 struct lustre_cfg *old_cfg = NULL;
4554 param = lustre_cfg_string(cfg, 1);
4555 if (param == NULL) {
4556 CERROR("param is empty\n");
4561 ptr = class_find_old_param(param, mdt_interop_param);
4563 if (ptr->new_param == NULL) {
4565 CWARN("For interoperability, skip this %s."
4566 " It is obsolete.\n", ptr->old_param);
4570 CWARN("Found old param %s, changed it to %s.\n",
4571 ptr->old_param, ptr->new_param);
4574 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4581 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4583 if (rc > 0 || rc == -ENOSYS) {
4584 /* is it an HSM var ? */
4585 rc = class_process_proc_param(PARAM_HSM,
4586 hsm_cdt_get_proc_vars(),
4588 if (rc > 0 || rc == -ENOSYS)
4589 /* we don't understand; pass it on */
4590 rc = next->ld_ops->ldo_process_config(env, next,
4594 if (old_cfg != NULL)
4595 lustre_cfg_free(cfg);
4600 /* others are passed further */
4601 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4607 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4608 const struct lu_object_header *hdr,
4609 struct lu_device *d)
4611 struct mdt_object *mo;
4615 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4617 struct lu_object *o;
4618 struct lu_object_header *h;
4621 h = &mo->mot_header;
4622 lu_object_header_init(h);
4623 lu_object_init(o, h, d);
4624 lu_object_add_top(h, o);
4625 o->lo_ops = &mdt_obj_ops;
4626 spin_lock_init(&mo->mot_write_lock);
4627 mutex_init(&mo->mot_lov_mutex);
4628 init_rwsem(&mo->mot_open_sem);
4634 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4635 const struct lu_object_conf *unused)
4637 struct mdt_device *d = mdt_dev(o->lo_dev);
4638 struct lu_device *under;
4639 struct lu_object *below;
4643 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4644 PFID(lu_object_fid(o)));
4646 under = &d->mdt_child->md_lu_dev;
4647 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4648 if (below != NULL) {
4649 lu_object_add(o, below);
4656 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4658 struct mdt_object *mo = mdt_obj(o);
4659 struct lu_object_header *h;
4663 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4664 PFID(lu_object_fid(o)));
4666 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4667 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4670 lu_object_header_fini(h);
4671 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4676 static int mdt_object_print(const struct lu_env *env, void *cookie,
4677 lu_printer_t p, const struct lu_object *o)
4679 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4681 return (*p)(env, cookie,
4682 LUSTRE_MDT_NAME"-object@%p(flags=%d, writecount=%d)",
4683 mdto, mdto->mot_flags, mdto->mot_write_count);
4686 static int mdt_prepare(const struct lu_env *env,
4687 struct lu_device *pdev,
4688 struct lu_device *cdev)
4690 struct mdt_device *mdt = mdt_dev(cdev);
4691 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4692 struct obd_device *obd = cdev->ld_obd;
4699 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4703 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4707 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4711 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4712 /* The LFSCK instance is registered just now, so it must be there when
4713 * register the namespace to such instance. */
4714 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4716 if (mdt->mdt_seq_site.ss_node_id == 0) {
4717 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4718 &mdt->mdt_md_root_fid);
4723 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4724 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4725 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4726 LASSERT(obd->obd_no_conn);
4727 spin_lock(&obd->obd_dev_lock);
4728 obd->obd_no_conn = 0;
4729 spin_unlock(&obd->obd_dev_lock);
4731 if (obd->obd_recovering == 0)
4732 mdt_postrecov(env, mdt);
4737 const struct lu_device_operations mdt_lu_ops = {
4738 .ldo_object_alloc = mdt_object_alloc,
4739 .ldo_process_config = mdt_process_config,
4740 .ldo_prepare = mdt_prepare,
4743 static const struct lu_object_operations mdt_obj_ops = {
4744 .loo_object_init = mdt_object_init,
4745 .loo_object_free = mdt_object_free,
4746 .loo_object_print = mdt_object_print
4749 static int mdt_obd_set_info_async(const struct lu_env *env,
4750 struct obd_export *exp,
4751 __u32 keylen, void *key,
4752 __u32 vallen, void *val,
4753 struct ptlrpc_request_set *set)
4759 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4760 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4768 * Match client and server connection feature flags.
4770 * Compute the compatibility flags for a connection request based on
4771 * features mutually supported by client and server.
4773 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4774 * must not be updated here, otherwise a partially initialized value may
4775 * be exposed. After the connection request is successfully processed,
4776 * the top-level MDT connect request handler atomically updates the export
4777 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4778 * reply. \see mdt_connect().
4780 * \param exp the obd_export associated with this client/target pair
4781 * \param mdt the target device for the connection
4782 * \param data stores data for this connect request
4785 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4786 * \retval -EBADE client and server feature requirements are incompatible
4788 static int mdt_connect_internal(struct obd_export *exp,
4789 struct mdt_device *mdt,
4790 struct obd_connect_data *data)
4792 LASSERT(data != NULL);
4794 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4795 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4797 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
4798 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
4799 CWARN("%s: client %s does not support ibits lock, either "
4800 "very old or an invalid client: flags "LPX64"\n",
4801 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
4802 data->ocd_connect_flags);
4806 if (!mdt->mdt_opts.mo_acl)
4807 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4809 if (!mdt->mdt_opts.mo_user_xattr)
4810 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4812 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4813 data->ocd_brw_size = min(data->ocd_brw_size,
4814 (__u32)MD_MAX_BRW_SIZE);
4815 if (data->ocd_brw_size == 0) {
4816 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4817 " ocd_version: %x ocd_grant: %d "
4818 "ocd_index: %u ocd_brw_size is "
4819 "unexpectedly zero, network data "
4820 "corruption? Refusing connection of this"
4823 exp->exp_client_uuid.uuid,
4824 exp, data->ocd_connect_flags, data->ocd_version,
4825 data->ocd_grant, data->ocd_index);
4830 /* NB: Disregard the rule against updating
4831 * exp_connect_data.ocd_connect_flags in this case, since
4832 * tgt_client_new() needs to know if this is a lightweight
4833 * connection, and it is safe to expose this flag before
4834 * connection processing completes. */
4835 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4836 spin_lock(&exp->exp_lock);
4837 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4838 spin_unlock(&exp->exp_lock);
4841 data->ocd_version = LUSTRE_VERSION_CODE;
4843 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4844 CWARN("%s: MDS requires FID support, but client not\n",
4849 if (OCD_HAS_FLAG(data, PINGLESS)) {
4850 if (ptlrpc_pinger_suppress_pings()) {
4851 spin_lock(&exp->exp_obd->obd_dev_lock);
4852 list_del_init(&exp->exp_obd_chain_timed);
4853 spin_unlock(&exp->exp_obd->obd_dev_lock);
4855 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4859 data->ocd_max_easize = mdt->mdt_max_ea_size;
4864 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
4865 struct mdt_thread_info *info,
4866 struct mdt_file_data *mfd)
4868 struct lu_context ses;
4872 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
4877 lu_context_enter(&ses);
4879 mdt_ucred(info)->uc_valid = UCRED_OLD;
4880 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
4882 lu_context_exit(&ses);
4883 lu_context_fini(&ses);
4889 static int mdt_export_cleanup(struct obd_export *exp)
4891 struct list_head closing_list;
4892 struct mdt_export_data *med = &exp->exp_mdt_data;
4893 struct obd_device *obd = exp->exp_obd;
4894 struct mdt_device *mdt;
4895 struct mdt_thread_info *info;
4897 struct mdt_file_data *mfd, *n;
4901 INIT_LIST_HEAD(&closing_list);
4902 spin_lock(&med->med_open_lock);
4903 while (!list_empty(&med->med_open_head)) {
4904 struct list_head *tmp = med->med_open_head.next;
4905 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
4907 /* Remove mfd handle so it can't be found again.
4908 * We are consuming the mfd_list reference here. */
4909 class_handle_unhash(&mfd->mfd_handle);
4910 list_move_tail(&mfd->mfd_list, &closing_list);
4912 spin_unlock(&med->med_open_lock);
4913 mdt = mdt_dev(obd->obd_lu_dev);
4914 LASSERT(mdt != NULL);
4916 rc = lu_env_init(&env, LCT_MD_THREAD);
4920 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
4921 LASSERT(info != NULL);
4922 memset(info, 0, sizeof *info);
4923 info->mti_env = &env;
4924 info->mti_mdt = mdt;
4925 info->mti_exp = exp;
4927 if (!list_empty(&closing_list)) {
4928 struct md_attr *ma = &info->mti_attr;
4930 /* Close any open files (which may also cause orphan
4932 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
4933 list_del_init(&mfd->mfd_list);
4934 ma->ma_need = ma->ma_valid = 0;
4936 /* This file is being closed due to an eviction, it
4937 * could have been modified and now dirty regarding to
4938 * HSM archive, check this!
4939 * The logic here is to mark a file dirty if there's a
4940 * chance it was dirtied before the client was evicted,
4941 * so that we don't have to wait for a release attempt
4942 * before finding out the file was actually dirty and
4943 * fail the release. Aggressively marking it dirty here
4944 * will cause the policy engine to attempt to
4945 * re-archive it; when rearchiving, we can compare the
4946 * current version to the HSM data_version and make the
4947 * archive request into a noop if it's not actually
4950 if (mfd->mfd_mode & FMODE_WRITE)
4951 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
4953 /* Don't unlink orphan on failover umount, LU-184 */
4954 if (exp->exp_flags & OBD_OPT_FAILOVER) {
4955 ma->ma_valid = MA_FLAGS;
4956 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
4958 mdt_mfd_close(info, mfd);
4961 info->mti_mdt = NULL;
4962 /* cleanup client slot early */
4963 /* Do not erase record for recoverable client. */
4964 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
4965 tgt_client_del(&env, exp);
4971 static int mdt_obd_disconnect(struct obd_export *exp)
4977 class_export_get(exp);
4979 nodemap_del_member(exp);
4980 rc = server_disconnect_export(exp);
4982 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
4984 rc = mdt_export_cleanup(exp);
4985 class_export_put(exp);
4989 /* mds_connect copy */
4990 static int mdt_obd_connect(const struct lu_env *env,
4991 struct obd_export **exp, struct obd_device *obd,
4992 struct obd_uuid *cluuid,
4993 struct obd_connect_data *data,
4996 struct obd_export *lexp;
4997 struct lustre_handle conn = { 0 };
4998 struct mdt_device *mdt;
5000 lnet_nid_t *client_nid = localdata;
5003 LASSERT(env != NULL);
5004 if (!exp || !obd || !cluuid)
5007 mdt = mdt_dev(obd->obd_lu_dev);
5010 * first, check whether the stack is ready to handle requests
5011 * XXX: probably not very appropriate method is used now
5012 * at some point we should find a better one
5014 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
5015 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5016 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5017 rc = obd_get_info(env, mdt->mdt_child_exp,
5018 sizeof(KEY_OSP_CONNECTED),
5019 KEY_OSP_CONNECTED, NULL, NULL);
5022 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5025 rc = class_connect(&conn, obd, cluuid);
5029 lexp = class_conn2export(&conn);
5030 LASSERT(lexp != NULL);
5032 rc = mdt_connect_internal(lexp, mdt, data);
5034 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5037 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5038 rc = tgt_client_new(env, lexp);
5040 rc = nodemap_add_member(*client_nid, lexp);
5041 if (rc != 0 && rc != -EEXIST)
5044 mdt_export_stats_init(obd, lexp, localdata);
5047 /* For phase I, sync for cross-ref operation. */
5048 spin_lock(&lexp->exp_lock);
5049 lexp->exp_keep_sync = 1;
5050 spin_unlock(&lexp->exp_lock);
5054 class_disconnect(lexp);
5063 static int mdt_obd_reconnect(const struct lu_env *env,
5064 struct obd_export *exp, struct obd_device *obd,
5065 struct obd_uuid *cluuid,
5066 struct obd_connect_data *data,
5069 lnet_nid_t *client_nid = localdata;
5073 if (exp == NULL || obd == NULL || cluuid == NULL)
5076 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5078 rc = nodemap_add_member(*client_nid, exp);
5079 if (rc == 0 || rc == -EEXIST)
5080 mdt_export_stats_init(obd, exp, localdata);
5086 /* FIXME: Can we avoid using these two interfaces? */
5087 static int mdt_init_export(struct obd_export *exp)
5089 struct mdt_export_data *med = &exp->exp_mdt_data;
5093 INIT_LIST_HEAD(&med->med_open_head);
5094 spin_lock_init(&med->med_open_lock);
5095 mutex_init(&med->med_idmap_mutex);
5096 med->med_idmap = NULL;
5097 spin_lock(&exp->exp_lock);
5098 exp->exp_connecting = 1;
5099 spin_unlock(&exp->exp_lock);
5101 /* self-export doesn't need client data and ldlm initialization */
5102 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5103 &exp->exp_client_uuid)))
5106 rc = tgt_client_alloc(exp);
5110 rc = ldlm_init_export(exp);
5117 tgt_client_free(exp);
5119 CERROR("%s: Failed to initialize export: rc = %d\n",
5120 exp->exp_obd->obd_name, rc);
5124 static int mdt_destroy_export(struct obd_export *exp)
5128 if (exp_connect_rmtclient(exp))
5129 mdt_cleanup_idmap(&exp->exp_mdt_data);
5131 target_destroy_export(exp);
5132 /* destroy can be called from failed obd_setup, so
5133 * checking uuid is safer than obd_self_export */
5134 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5135 &exp->exp_client_uuid)))
5138 ldlm_destroy_export(exp);
5139 tgt_client_free(exp);
5141 LASSERT(list_empty(&exp->exp_outstanding_replies));
5142 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5147 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5148 struct linkea_data *ldata)
5152 LASSERT(ldata->ld_buf->lb_buf != NULL);
5154 if (!mdt_object_exists(mdt_obj))
5157 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5158 ldata->ld_buf, XATTR_NAME_LINK);
5159 if (rc == -ERANGE) {
5160 /* Buf was too small, figure out what we need. */
5161 lu_buf_free(ldata->ld_buf);
5162 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5163 ldata->ld_buf, XATTR_NAME_LINK);
5166 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5167 if (ldata->ld_buf->lb_buf == NULL)
5169 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5170 ldata->ld_buf, XATTR_NAME_LINK);
5175 return linkea_init(ldata);
5179 * Given an MDT object, try to look up the full path to the object.
5180 * Part of the MDT layer implementation of lfs fid2path.
5182 * \param[in] info Per-thread common data shared by MDT level handlers.
5183 * \param[in] obj Object to do path lookup of
5184 * \param[in,out] fp User-provided struct to store path information
5186 * \retval 0 Lookup successful, path information stored in fp
5187 * \retval -EAGAIN Lookup failed, usually because object is being moved
5188 * \retval negative errno if there was a problem
5190 static int mdt_path_current(struct mdt_thread_info *info,
5191 struct mdt_object *obj,
5192 struct getinfo_fid2path *fp)
5194 struct mdt_device *mdt = info->mti_mdt;
5195 struct mdt_object *mdt_obj;
5196 struct link_ea_header *leh;
5197 struct link_ea_entry *lee;
5198 struct lu_name *tmpname = &info->mti_name;
5199 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5200 struct lu_buf *buf = &info->mti_big_buf;
5203 struct linkea_data ldata = { NULL };
5208 /* temp buffer for path element, the buffer will be finally freed
5209 * in mdt_thread_info_fini */
5210 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5211 if (buf->lb_buf == NULL)
5215 ptr = fp->gf_path + fp->gf_pathlen - 1;
5218 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5220 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5221 * so checking root_fid can only happen on MDT0. */
5222 while (!lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid)) {
5223 struct lu_buf lmv_buf;
5225 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5226 if (IS_ERR(mdt_obj))
5227 GOTO(out, rc = PTR_ERR(mdt_obj));
5229 if (!mdt_object_exists(mdt_obj)) {
5230 mdt_object_put(info->mti_env, mdt_obj);
5231 GOTO(out, rc = -ENOENT);
5234 if (mdt_object_remote(mdt_obj)) {
5235 mdt_object_put(info->mti_env, mdt_obj);
5236 GOTO(remote_out, rc = -EREMOTE);
5239 rc = mdt_links_read(info, mdt_obj, &ldata);
5241 mdt_object_put(info->mti_env, mdt_obj);
5246 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5247 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5248 /* If set, use link #linkno for path lookup, otherwise use
5249 link #0. Only do this for the final path element. */
5250 if (first && fp->gf_linkno < leh->leh_reccount) {
5252 for (count = 0; count < fp->gf_linkno; count++) {
5253 lee = (struct link_ea_entry *)
5254 ((char *)lee + reclen);
5255 linkea_entry_unpack(lee, &reclen, tmpname,
5258 if (fp->gf_linkno < leh->leh_reccount - 1)
5259 /* indicate to user there are more links */
5263 lmv_buf.lb_buf = info->mti_xattr_buf;
5264 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5265 /* Check if it is slave stripes */
5266 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5267 &lmv_buf, XATTR_NAME_LMV);
5268 mdt_object_put(info->mti_env, mdt_obj);
5270 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5272 /* For slave stripes, get its master */
5273 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5274 fp->gf_fid = *tmpfid;
5277 } else if (rc < 0 && rc != -ENODATA) {
5283 /* Pack the name in the end of the buffer */
5284 ptr -= tmpname->ln_namelen;
5285 if (ptr - 1 <= fp->gf_path)
5286 GOTO(out, rc = -EOVERFLOW);
5287 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5290 /* keep the last resolved fid to the client, so the
5291 * client will build the left path on another MDT for
5293 fp->gf_fid = *tmpfid;
5299 ptr++; /* skip leading / */
5300 memmove(fp->gf_path, ptr, fp->gf_path + fp->gf_pathlen - ptr);
5307 * Given an MDT object, use mdt_path_current to get the path.
5308 * Essentially a wrapper to retry mdt_path_current a set number of times
5309 * if -EAGAIN is returned (usually because an object is being moved).
5311 * Part of the MDT layer implementation of lfs fid2path.
5313 * \param[in] info Per-thread common data shared by mdt level handlers.
5314 * \param[in] obj Object to do path lookup of
5315 * \param[in,out] fp User-provided struct for arguments and to store path
5318 * \retval 0 Lookup successful, path information stored in fp
5319 * \retval negative errno if there was a problem
5321 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5322 struct getinfo_fid2path *fp)
5324 struct mdt_device *mdt = info->mti_mdt;
5329 if (fp->gf_pathlen < 3)
5332 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5333 fp->gf_path[0] = '\0';
5337 /* Retry multiple times in case file is being moved */
5338 while (tries-- && rc == -EAGAIN)
5339 rc = mdt_path_current(info, obj, fp);
5345 * Get the full path of the provided FID, as of changelog record recno.
5347 * This checks sanity and looks up object for user provided FID
5348 * before calling the actual path lookup code.
5350 * Part of the MDT layer implementation of lfs fid2path.
5352 * \param[in] info Per-thread common data shared by mdt level handlers.
5353 * \param[in,out] fp User-provided struct for arguments and to store path
5356 * \retval 0 Lookup successful, path information and recno stored in fp
5357 * \retval -ENOENT, object does not exist
5358 * \retval negative errno if there was a problem
5360 static int mdt_fid2path(struct mdt_thread_info *info,
5361 struct getinfo_fid2path *fp)
5363 struct mdt_device *mdt = info->mti_mdt;
5364 struct mdt_object *obj;
5368 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5369 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5371 if (!fid_is_sane(&fp->gf_fid))
5374 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5375 CWARN("%s: "DFID" is invalid, sequence should be "
5376 ">= "LPX64"\n", mdt_obd_name(mdt),
5377 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5381 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5384 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5385 PFID(&fp->gf_fid), rc);
5389 if (mdt_object_remote(obj))
5391 else if (!mdt_object_exists(obj))
5397 mdt_object_put(info->mti_env, obj);
5398 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5399 PFID(&fp->gf_fid), rc);
5403 rc = mdt_path(info, obj, fp);
5405 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5406 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5408 mdt_object_put(info->mti_env, obj);
5413 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5414 void *val, int vallen)
5416 struct getinfo_fid2path *fpout, *fpin;
5419 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5422 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5423 lustre_swab_fid2path(fpin);
5425 memcpy(fpout, fpin, sizeof(*fpin));
5426 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5429 rc = mdt_fid2path(info, fpout);
5433 int mdt_get_info(struct tgt_session_info *tsi)
5443 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5445 CDEBUG(D_IOCTL, "No GETINFO key\n");
5446 RETURN(err_serious(-EFAULT));
5448 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5451 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5452 if (vallen == NULL) {
5453 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5454 tgt_name(tsi->tsi_tgt));
5455 RETURN(err_serious(-EFAULT));
5458 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5460 rc = req_capsule_server_pack(tsi->tsi_pill);
5462 RETURN(err_serious(rc));
5464 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5465 if (valout == NULL) {
5466 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5467 tgt_name(tsi->tsi_tgt));
5468 RETURN(err_serious(-EFAULT));
5471 if (KEY_IS(KEY_FID2PATH)) {
5472 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5474 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5475 mdt_thread_info_fini(info);
5482 /* Pass the ioc down */
5483 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5484 unsigned int cmd, int len, void *data)
5486 struct lu_context ioctl_session;
5487 struct md_device *next = mdt->mdt_child;
5491 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5494 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5495 lu_context_enter(&ioctl_session);
5496 env->le_ses = &ioctl_session;
5498 LASSERT(next->md_ops->mdo_iocontrol);
5499 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5501 lu_context_exit(&ioctl_session);
5502 lu_context_fini(&ioctl_session);
5506 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5508 struct obd_ioctl_data *data = karg;
5511 struct mdt_object *obj;
5512 struct mdt_lock_handle *lh;
5516 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5517 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5520 fid = (struct lu_fid *)data->ioc_inlbuf1;
5522 if (!fid_is_sane(fid))
5525 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5527 lh = &mti->mti_lh[MDT_LH_PARENT];
5528 mdt_lock_reg_init(lh, LCK_CR);
5530 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5532 RETURN(PTR_ERR(obj));
5534 if (mdt_object_remote(obj)) {
5537 * before calling version get the correct MDS should be
5538 * fid, this is error to find remote object here
5540 CERROR("nonlocal object "DFID"\n", PFID(fid));
5541 } else if (!mdt_object_exists(obj)) {
5542 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5545 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5546 *(__u64 *)data->ioc_inlbuf2 = version;
5549 mdt_object_unlock_put(mti, obj, lh, 1);
5553 /* ioctls on obd dev */
5554 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5555 void *karg, void *uarg)
5558 struct obd_device *obd = exp->exp_obd;
5559 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5560 struct dt_device *dt = mdt->mdt_bottom;
5564 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5565 rc = lu_env_init(&env, LCT_MD_THREAD);
5571 rc = mdt_device_sync(&env, mdt);
5573 case OBD_IOC_SET_READONLY:
5574 rc = dt->dd_ops->dt_ro(&env, dt);
5576 case OBD_IOC_ABORT_RECOVERY:
5577 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5578 target_stop_recovery_thread(obd);
5581 case OBD_IOC_CHANGELOG_REG:
5582 case OBD_IOC_CHANGELOG_DEREG:
5583 case OBD_IOC_CHANGELOG_CLEAR:
5584 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5586 case OBD_IOC_START_LFSCK: {
5587 struct md_device *next = mdt->mdt_child;
5588 struct obd_ioctl_data *data = karg;
5589 struct lfsck_start_param lsp;
5591 if (unlikely(data == NULL)) {
5596 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5597 lsp.lsp_index_valid = 0;
5598 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5601 case OBD_IOC_STOP_LFSCK: {
5602 struct md_device *next = mdt->mdt_child;
5603 struct obd_ioctl_data *data = karg;
5604 struct lfsck_stop stop;
5606 stop.ls_status = LS_STOPPED;
5607 /* Old lfsck utils may pass NULL @stop. */
5608 if (data->ioc_inlbuf1 == NULL)
5612 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5614 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5617 case OBD_IOC_GET_OBJ_VERSION: {
5618 struct mdt_thread_info *mti;
5619 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5620 memset(mti, 0, sizeof *mti);
5621 mti->mti_env = &env;
5625 rc = mdt_ioc_version_get(mti, karg);
5628 case OBD_IOC_CATLOGLIST: {
5629 struct mdt_thread_info *mti;
5631 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5632 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5633 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5634 &mti->mti_tmp_fid1);
5639 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5640 mdt_obd_name(mdt), cmd, rc);
5647 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5649 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5650 struct lfsck_start_param lsp;
5654 lsp.lsp_start = NULL;
5655 lsp.lsp_index_valid = 0;
5656 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
5657 OBD_IOC_START_LFSCK,
5659 if (rc != 0 && rc != -EALREADY)
5660 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
5661 mdt_obd_name(mdt), rc);
5663 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5667 static int mdt_obd_postrecov(struct obd_device *obd)
5672 rc = lu_env_init(&env, LCT_MD_THREAD);
5675 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5680 static struct obd_ops mdt_obd_device_ops = {
5681 .o_owner = THIS_MODULE,
5682 .o_set_info_async = mdt_obd_set_info_async,
5683 .o_connect = mdt_obd_connect,
5684 .o_reconnect = mdt_obd_reconnect,
5685 .o_disconnect = mdt_obd_disconnect,
5686 .o_init_export = mdt_init_export,
5687 .o_destroy_export = mdt_destroy_export,
5688 .o_iocontrol = mdt_iocontrol,
5689 .o_postrecov = mdt_obd_postrecov,
5692 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5693 struct lu_device *d)
5695 struct mdt_device *m = mdt_dev(d);
5702 static struct lu_device *mdt_device_free(const struct lu_env *env,
5703 struct lu_device *d)
5705 struct mdt_device *m = mdt_dev(d);
5708 lu_device_fini(&m->mdt_lu_dev);
5714 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5715 struct lu_device_type *t,
5716 struct lustre_cfg *cfg)
5718 struct lu_device *l;
5719 struct mdt_device *m;
5726 rc = mdt_init0(env, m, t, cfg);
5728 mdt_device_free(env, l);
5733 l = ERR_PTR(-ENOMEM);
5737 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5738 LU_KEY_INIT(mdt, struct mdt_thread_info);
5740 static void mdt_key_fini(const struct lu_context *ctx,
5741 struct lu_context_key *key, void* data)
5743 struct mdt_thread_info *info = data;
5745 if (info->mti_big_lmm) {
5746 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5747 info->mti_big_lmm = NULL;
5748 info->mti_big_lmmsize = 0;
5753 /* context key: mdt_thread_key */
5754 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5756 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5758 return lu_ucred(info->mti_env);
5761 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5763 return lu_ucred_check(info->mti_env);
5767 * Enable/disable COS (Commit On Sharing).
5769 * Set/Clear the COS flag in mdt options.
5771 * \param mdt mdt device
5772 * \param val 0 disables COS, other values enable COS
5774 void mdt_enable_cos(struct mdt_device *mdt, int val)
5779 mdt->mdt_opts.mo_cos = !!val;
5780 rc = lu_env_init(&env, LCT_LOCAL);
5781 if (unlikely(rc != 0)) {
5782 CWARN("%s: lu_env initialization failed, cannot "
5783 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
5786 mdt_device_sync(&env, mdt);
5791 * Check COS (Commit On Sharing) status.
5793 * Return COS flag status.
5795 * \param mdt mdt device
5797 int mdt_cos_is_enabled(struct mdt_device *mdt)
5799 return mdt->mdt_opts.mo_cos != 0;
5802 static struct lu_device_type_operations mdt_device_type_ops = {
5803 .ldto_device_alloc = mdt_device_alloc,
5804 .ldto_device_free = mdt_device_free,
5805 .ldto_device_fini = mdt_device_fini
5808 static struct lu_device_type mdt_device_type = {
5809 .ldt_tags = LU_DEVICE_MD,
5810 .ldt_name = LUSTRE_MDT_NAME,
5811 .ldt_ops = &mdt_device_type_ops,
5812 .ldt_ctx_tags = LCT_MD_THREAD
5815 static int __init mdt_mod_init(void)
5819 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5820 FID_NOBRACE_LEN + 1);
5821 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5823 rc = lu_kmem_init(mdt_caches);
5827 rc = mds_mod_init();
5831 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
5832 LUSTRE_MDT_NAME, &mdt_device_type);
5837 lu_kmem_fini(mdt_caches);
5844 static void __exit mdt_mod_exit(void)
5846 class_unregister_type(LUSTRE_MDT_NAME);
5848 lu_kmem_fini(mdt_caches);
5851 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
5852 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5853 MODULE_VERSION(LUSTRE_VERSION_STRING);
5854 MODULE_LICENSE("GPL");
5856 module_init(mdt_mod_init);
5857 module_exit(mdt_mod_exit);