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, 2015, 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 #include <dt_object.h>
54 #include <lustre_acl.h>
55 #include <lustre_export.h>
56 #include <lustre_ioctl.h>
57 #include <lustre_lfsck.h>
58 #include <lustre_log.h>
59 #include <lustre_net.h>
60 #include <lustre_nodemap.h>
61 #include <lustre_mds.h>
62 #include <lustre_param.h>
63 #include <lustre_quota.h>
64 #include <lustre_swab.h>
66 #include <obd_support.h>
68 #include "mdt_internal.h"
71 static unsigned int max_mod_rpcs_per_client = 8;
72 CFS_MODULE_PARM(max_mod_rpcs_per_client, "i", uint, 0644,
73 "maximum number of modify RPCs in flight allowed per client");
75 mdl_mode_t mdt_mdl_lock_modes[] = {
76 [LCK_MINMODE] = MDL_MINMODE,
83 [LCK_GROUP] = MDL_GROUP
86 enum ldlm_mode mdt_dlm_lock_modes[] = {
87 [MDL_MINMODE] = LCK_MINMODE,
94 [MDL_GROUP] = LCK_GROUP
97 static struct mdt_device *mdt_dev(struct lu_device *d);
98 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
100 static const struct lu_object_operations mdt_obj_ops;
102 /* Slab for MDT object allocation */
103 static struct kmem_cache *mdt_object_kmem;
105 /* For HSM restore handles */
106 struct kmem_cache *mdt_hsm_cdt_kmem;
108 /* For HSM request handles */
109 struct kmem_cache *mdt_hsm_car_kmem;
111 static struct lu_kmem_descr mdt_caches[] = {
113 .ckd_cache = &mdt_object_kmem,
114 .ckd_name = "mdt_obj",
115 .ckd_size = sizeof(struct mdt_object)
118 .ckd_cache = &mdt_hsm_cdt_kmem,
119 .ckd_name = "mdt_cdt_restore_handle",
120 .ckd_size = sizeof(struct cdt_restore_handle)
123 .ckd_cache = &mdt_hsm_car_kmem,
124 .ckd_name = "mdt_cdt_agent_req",
125 .ckd_size = sizeof(struct cdt_agent_req)
132 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
136 return rep->lock_policy_res1 & op_flag;
139 void mdt_clear_disposition(struct mdt_thread_info *info,
140 struct ldlm_reply *rep, __u64 op_flag)
143 info->mti_opdata &= ~op_flag;
144 tgt_opdata_clear(info->mti_env, op_flag);
147 rep->lock_policy_res1 &= ~op_flag;
150 void mdt_set_disposition(struct mdt_thread_info *info,
151 struct ldlm_reply *rep, __u64 op_flag)
154 info->mti_opdata |= op_flag;
155 tgt_opdata_set(info->mti_env, op_flag);
158 rep->lock_policy_res1 |= op_flag;
161 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
163 lh->mlh_pdo_hash = 0;
164 lh->mlh_reg_mode = lm;
165 lh->mlh_rreg_mode = lm;
166 lh->mlh_type = MDT_REG_LOCK;
169 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
170 const struct lu_name *lname)
172 lh->mlh_reg_mode = lock_mode;
173 lh->mlh_rreg_mode = lock_mode;
174 lh->mlh_type = MDT_PDO_LOCK;
176 if (lu_name_is_valid(lname)) {
177 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
179 /* XXX Workaround for LU-2856
181 * Zero is a valid return value of full_name_hash, but
182 * several users of mlh_pdo_hash assume a non-zero
183 * hash value. We therefore map zero onto an
184 * arbitrary, but consistent value (1) to avoid
185 * problems further down the road. */
186 if (unlikely(lh->mlh_pdo_hash == 0))
187 lh->mlh_pdo_hash = 1;
189 lh->mlh_pdo_hash = 0;
193 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
194 struct mdt_lock_handle *lh)
200 * Any dir access needs couple of locks:
202 * 1) on part of dir we gonna take lookup/modify;
204 * 2) on whole dir to protect it from concurrent splitting and/or to
205 * flush client's cache for readdir().
207 * so, for a given mode and object this routine decides what lock mode
208 * to use for lock #2:
210 * 1) if caller's gonna lookup in dir then we need to protect dir from
211 * being splitted only - LCK_CR
213 * 2) if caller's gonna modify dir then we need to protect dir from
214 * being splitted and to flush cache - LCK_CW
216 * 3) if caller's gonna modify dir and that dir seems ready for
217 * splitting then we need to protect it from any type of access
218 * (lookup/modify/split) - LCK_EX --bzzz
221 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
222 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
225 * Ask underlaying level its opinion about preferable PDO lock mode
226 * having access type passed as regular lock mode:
228 * - MDL_MINMODE means that lower layer does not want to specify lock
231 * - MDL_NL means that no PDO lock should be taken. This is used in some
232 * cases. Say, for non-splittable directories no need to use PDO locks
235 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
236 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
238 if (mode != MDL_MINMODE) {
239 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
242 * Lower layer does not want to specify locking mode. We do it
243 * our selves. No special protection is needed, just flush
244 * client's cache on modification and allow concurrent
247 switch (lh->mlh_reg_mode) {
249 lh->mlh_pdo_mode = LCK_EX;
252 lh->mlh_pdo_mode = LCK_CR;
255 lh->mlh_pdo_mode = LCK_CW;
258 CERROR("Not expected lock type (0x%x)\n",
259 (int)lh->mlh_reg_mode);
264 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
268 static int mdt_getstatus(struct tgt_session_info *tsi)
270 struct mdt_thread_info *info = tsi2mdt_info(tsi);
271 struct mdt_device *mdt = info->mti_mdt;
272 struct mdt_body *repbody;
277 rc = mdt_check_ucred(info);
279 GOTO(out, rc = err_serious(rc));
281 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK))
282 GOTO(out, rc = err_serious(-ENOMEM));
284 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
285 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
286 repbody->mbo_valid |= OBD_MD_FLID;
290 mdt_thread_info_fini(info);
294 static int mdt_statfs(struct tgt_session_info *tsi)
296 struct ptlrpc_request *req = tgt_ses_req(tsi);
297 struct mdt_thread_info *info = tsi2mdt_info(tsi);
298 struct md_device *next = info->mti_mdt->mdt_child;
299 struct ptlrpc_service_part *svcpt;
300 struct obd_statfs *osfs;
305 svcpt = req->rq_rqbd->rqbd_svcpt;
307 /* This will trigger a watchdog timeout */
308 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
309 (MDT_SERVICE_WATCHDOG_FACTOR *
310 at_get(&svcpt->scp_at_estimate)) + 1);
312 rc = mdt_check_ucred(info);
314 GOTO(out, rc = err_serious(rc));
316 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
317 GOTO(out, rc = err_serious(-ENOMEM));
319 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
321 GOTO(out, rc = -EPROTO);
323 /** statfs information are cached in the mdt_device */
324 if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
325 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
326 /** statfs data is too old, get up-to-date one */
327 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
330 spin_lock(&info->mti_mdt->mdt_osfs_lock);
331 info->mti_mdt->mdt_osfs = *osfs;
332 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
333 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
335 /** use cached statfs data */
336 spin_lock(&info->mti_mdt->mdt_osfs_lock);
337 *osfs = info->mti_mdt->mdt_osfs;
338 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
342 mdt_counter_incr(req, LPROC_MDT_STATFS);
344 mdt_thread_info_fini(info);
348 #ifdef CONFIG_FS_POSIX_ACL
350 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
352 * \param info thread info object
353 * \param repbody reply to pack ACLs into
354 * \param o mdt object of file to examine
355 * \param nodemap nodemap of client to reply to
357 * \retval -errno error getting or parsing ACL from disk
359 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
360 struct mdt_object *o, struct lu_nodemap *nodemap)
362 const struct lu_env *env = info->mti_env;
363 struct md_object *next = mdt_object_child(o);
364 struct lu_buf *buf = &info->mti_buf;
367 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
368 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
370 if (buf->lb_len == 0)
373 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
375 if (rc == -ENODATA) {
376 repbody->mbo_aclsize = 0;
377 repbody->mbo_valid |= OBD_MD_FLACL;
379 } else if (rc == -EOPNOTSUPP) {
382 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
383 mdt_obd_name(info->mti_mdt),
384 PFID(mdt_object_fid(o)), rc);
387 rc = nodemap_map_acl(nodemap, buf->lb_buf,
388 rc, NODEMAP_FS_TO_CLIENT);
389 /* if all ACLs mapped out, rc is still >= 0 */
391 CERROR("%s: nodemap_map_acl unable to parse "DFID
392 " ACL: rc = %d\n", mdt_obd_name(info->mti_mdt),
393 PFID(mdt_object_fid(o)), rc);
395 repbody->mbo_aclsize = rc;
396 repbody->mbo_valid |= OBD_MD_FLACL;
404 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
405 const struct lu_attr *attr, const struct lu_fid *fid)
407 struct md_attr *ma = &info->mti_attr;
408 struct obd_export *exp = info->mti_exp;
409 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
411 LASSERT(ma->ma_valid & MA_INODE);
413 if (attr->la_valid & LA_ATIME) {
414 b->mbo_atime = attr->la_atime;
415 b->mbo_valid |= OBD_MD_FLATIME;
417 if (attr->la_valid & LA_MTIME) {
418 b->mbo_mtime = attr->la_mtime;
419 b->mbo_valid |= OBD_MD_FLMTIME;
421 if (attr->la_valid & LA_CTIME) {
422 b->mbo_ctime = attr->la_ctime;
423 b->mbo_valid |= OBD_MD_FLCTIME;
425 if (attr->la_valid & LA_FLAGS) {
426 b->mbo_flags = attr->la_flags;
427 b->mbo_valid |= OBD_MD_FLFLAGS;
429 if (attr->la_valid & LA_NLINK) {
430 b->mbo_nlink = attr->la_nlink;
431 b->mbo_valid |= OBD_MD_FLNLINK;
433 if (attr->la_valid & LA_UID) {
434 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
435 NODEMAP_FS_TO_CLIENT,
437 b->mbo_valid |= OBD_MD_FLUID;
439 if (attr->la_valid & LA_GID) {
440 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
441 NODEMAP_FS_TO_CLIENT,
443 b->mbo_valid |= OBD_MD_FLGID;
445 b->mbo_mode = attr->la_mode;
446 if (attr->la_valid & LA_MODE)
447 b->mbo_valid |= OBD_MD_FLMODE;
448 if (attr->la_valid & LA_TYPE)
449 b->mbo_valid |= OBD_MD_FLTYPE;
453 b->mbo_valid |= OBD_MD_FLID;
454 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid="LPX64"\n",
455 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
459 mdt_body_reverse_idmap(info, b);
461 if (!(attr->la_valid & LA_TYPE))
464 b->mbo_rdev = attr->la_rdev;
465 b->mbo_size = attr->la_size;
466 b->mbo_blocks = attr->la_blocks;
468 if (!S_ISREG(attr->la_mode)) {
469 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
470 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
471 /* means no objects are allocated on osts. */
472 LASSERT(!(ma->ma_valid & MA_LOV));
473 /* just ignore blocks occupied by extend attributes on MDS */
475 /* if no object is allocated on osts, the size on mds is valid.
477 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
478 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
479 ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
480 /* A released file stores its size on MDS. */
481 /* But return 1 block for released file, unless tools like tar
482 * will consider it fully sparse. (LU-3864)
484 if (unlikely(b->mbo_size == 0))
488 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
491 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
492 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
493 PFID(fid), (unsigned long long)b->mbo_size);
496 static inline int mdt_body_has_lov(const struct lu_attr *la,
497 const struct mdt_body *body)
499 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
500 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
503 void mdt_client_compatibility(struct mdt_thread_info *info)
505 struct mdt_body *body;
506 struct ptlrpc_request *req = mdt_info_req(info);
507 struct obd_export *exp = req->rq_export;
508 struct md_attr *ma = &info->mti_attr;
509 struct lu_attr *la = &ma->ma_attr;
512 if (exp_connect_layout(exp))
513 /* the client can deal with 16-bit lmm_stripe_count */
516 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
518 if (!mdt_body_has_lov(la, body))
521 /* now we have a reply with a lov for a client not compatible with the
522 * layout lock so we have to clean the layout generation number */
523 if (S_ISREG(la->la_mode))
524 ma->ma_lmm->lmm_layout_gen = 0;
528 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
529 struct mdt_object *o)
531 const struct lu_env *env = info->mti_env;
534 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
543 /* Is it a directory? Let's check for the LMV as well */
544 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
545 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
549 rc2 = mo_xattr_get(env, mdt_object_child(o),
551 XATTR_NAME_DEFAULT_LMV);
553 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
561 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
564 const struct lu_env *env = info->mti_env;
568 LASSERT(info->mti_big_lmm_used == 0);
569 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
573 /* big_lmm may need to be grown */
574 if (info->mti_big_lmmsize < rc) {
575 int size = size_roundup_power2(rc);
577 if (info->mti_big_lmmsize > 0) {
578 /* free old buffer */
579 LASSERT(info->mti_big_lmm);
580 OBD_FREE_LARGE(info->mti_big_lmm,
581 info->mti_big_lmmsize);
582 info->mti_big_lmm = NULL;
583 info->mti_big_lmmsize = 0;
586 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
587 if (info->mti_big_lmm == NULL)
589 info->mti_big_lmmsize = size;
591 LASSERT(info->mti_big_lmmsize >= rc);
593 info->mti_buf.lb_buf = info->mti_big_lmm;
594 info->mti_buf.lb_len = info->mti_big_lmmsize;
595 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
600 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
601 struct md_attr *ma, const char *name)
603 struct md_object *next = mdt_object_child(o);
604 struct lu_buf *buf = &info->mti_buf;
607 if (strcmp(name, XATTR_NAME_LOV) == 0) {
608 buf->lb_buf = ma->ma_lmm;
609 buf->lb_len = ma->ma_lmm_size;
610 LASSERT(!(ma->ma_valid & MA_LOV));
611 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
612 buf->lb_buf = ma->ma_lmv;
613 buf->lb_len = ma->ma_lmv_size;
614 LASSERT(!(ma->ma_valid & MA_LMV));
615 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
616 buf->lb_buf = ma->ma_lmv;
617 buf->lb_len = ma->ma_lmv_size;
618 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
623 rc = mo_xattr_get(info->mti_env, next, buf, name);
627 if (strcmp(name, XATTR_NAME_LOV) == 0) {
628 if (info->mti_big_lmm_used)
629 ma->ma_lmm = info->mti_big_lmm;
631 /* NOT return LOV EA with hole to old client. */
632 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
633 LOV_PATTERN_F_HOLE) &&
634 !(exp_connect_flags(info->mti_exp) &
635 OBD_CONNECT_LFSCK)) {
638 ma->ma_lmm_size = rc;
639 ma->ma_valid |= MA_LOV;
641 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
642 if (info->mti_big_lmm_used)
643 ma->ma_lmv = info->mti_big_lmm;
645 ma->ma_lmv_size = rc;
646 ma->ma_valid |= MA_LMV;
647 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
648 ma->ma_lmv_size = rc;
649 ma->ma_valid |= MA_LMV_DEF;
652 /* Update mdt_max_mdsize so all clients will be aware that */
653 if (info->mti_mdt->mdt_max_mdsize < rc)
654 info->mti_mdt->mdt_max_mdsize = rc;
657 } else if (rc == -ENODATA) {
660 } else if (rc == -ERANGE) {
661 /* Default LMV has fixed size, so it must be able to fit
662 * in the original buffer */
663 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
665 rc = mdt_big_xattr_get(info, o, name);
667 info->mti_big_lmm_used = 1;
675 static int mdt_attr_get_pfid(struct mdt_thread_info *info,
676 struct mdt_object *o, struct lu_fid *pfid)
678 struct lu_buf *buf = &info->mti_buf;
679 struct link_ea_header *leh;
680 struct link_ea_entry *lee;
684 buf->lb_buf = info->mti_big_lmm;
685 buf->lb_len = info->mti_big_lmmsize;
686 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
687 buf, XATTR_NAME_LINK);
688 /* ignore errors, MA_PFID won't be set and it is
689 * up to the caller to treat this as an error */
690 if (rc == -ERANGE || buf->lb_len == 0) {
691 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
692 buf->lb_buf = info->mti_big_lmm;
693 buf->lb_len = info->mti_big_lmmsize;
698 if (rc < sizeof(*leh)) {
699 CERROR("short LinkEA on "DFID": rc = %d\n",
700 PFID(mdt_object_fid(o)), rc);
704 leh = (struct link_ea_header *) buf->lb_buf;
705 lee = (struct link_ea_entry *)(leh + 1);
706 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
707 leh->leh_magic = LINK_EA_MAGIC;
708 leh->leh_reccount = __swab32(leh->leh_reccount);
709 leh->leh_len = __swab64(leh->leh_len);
711 if (leh->leh_magic != LINK_EA_MAGIC)
713 if (leh->leh_reccount == 0)
716 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
717 fid_be_to_cpu(pfid, pfid);
722 int mdt_attr_get_complex(struct mdt_thread_info *info,
723 struct mdt_object *o, struct md_attr *ma)
725 const struct lu_env *env = info->mti_env;
726 struct md_object *next = mdt_object_child(o);
727 struct lu_buf *buf = &info->mti_buf;
728 int need = ma->ma_need;
735 if (mdt_object_exists(o) == 0)
736 GOTO(out, rc = -ENOENT);
737 mode = lu_object_attr(&next->mo_lu);
739 if (need & MA_INODE) {
740 ma->ma_need = MA_INODE;
741 rc = mo_attr_get(env, next, ma);
744 ma->ma_valid |= MA_INODE;
747 if (need & MA_PFID) {
748 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
750 ma->ma_valid |= MA_PFID;
751 /* ignore this error, parent fid is not mandatory */
755 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
756 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
761 if (need & MA_LMV && S_ISDIR(mode)) {
762 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
767 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
768 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
773 if (need & MA_HSM && S_ISREG(mode)) {
774 buf->lb_buf = info->mti_xattr_buf;
775 buf->lb_len = sizeof(info->mti_xattr_buf);
776 CLASSERT(sizeof(struct hsm_attrs) <=
777 sizeof(info->mti_xattr_buf));
778 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
779 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
781 ma->ma_valid |= MA_HSM;
782 else if (rc2 < 0 && rc2 != -ENODATA)
786 #ifdef CONFIG_FS_POSIX_ACL
787 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
788 buf->lb_buf = ma->ma_acl;
789 buf->lb_len = ma->ma_acl_size;
790 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
792 ma->ma_acl_size = rc2;
793 ma->ma_valid |= MA_ACL_DEF;
794 } else if (rc2 == -ENODATA) {
803 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = "LPX64" ma_lmm=%p\n",
804 rc, ma->ma_valid, ma->ma_lmm);
808 static int mdt_getattr_internal(struct mdt_thread_info *info,
809 struct mdt_object *o, int ma_need)
811 struct md_object *next = mdt_object_child(o);
812 const struct mdt_body *reqbody = info->mti_body;
813 struct ptlrpc_request *req = mdt_info_req(info);
814 struct md_attr *ma = &info->mti_attr;
815 struct lu_attr *la = &ma->ma_attr;
816 struct req_capsule *pill = info->mti_pill;
817 const struct lu_env *env = info->mti_env;
818 struct mdt_body *repbody;
819 struct lu_buf *buffer = &info->mti_buf;
820 struct obd_export *exp = info->mti_exp;
821 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
826 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
827 RETURN(err_serious(-ENOMEM));
829 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
833 if (mdt_object_remote(o)) {
834 /* This object is located on remote node.*/
835 /* Return -ENOTSUPP for old client */
836 if (!mdt_is_dne_client(req->rq_export))
837 GOTO(out, rc = -ENOTSUPP);
839 repbody->mbo_fid1 = *mdt_object_fid(o);
840 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
844 if (reqbody->mbo_eadatasize > 0) {
845 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
846 if (buffer->lb_buf == NULL)
847 GOTO(out, rc = -EPROTO);
848 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
851 buffer->lb_buf = NULL;
853 ma_need &= ~(MA_LOV | MA_LMV);
854 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
855 mdt_obd_name(info->mti_mdt),
856 req->rq_export->exp_client_uuid.uuid);
859 /* If it is dir object and client require MEA, then we got MEA */
860 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
861 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
862 /* Assumption: MDT_MD size is enough for lmv size. */
863 ma->ma_lmv = buffer->lb_buf;
864 ma->ma_lmv_size = buffer->lb_len;
865 ma->ma_need = MA_INODE;
866 if (ma->ma_lmv_size > 0) {
867 if (reqbody->mbo_valid & OBD_MD_MEA)
868 ma->ma_need |= MA_LMV;
869 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
870 ma->ma_need |= MA_LMV_DEF;
873 ma->ma_lmm = buffer->lb_buf;
874 ma->ma_lmm_size = buffer->lb_len;
875 ma->ma_need = MA_INODE | MA_HSM;
876 if (ma->ma_lmm_size > 0)
877 ma->ma_need |= MA_LOV;
880 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
881 reqbody->mbo_valid & OBD_MD_FLDIREA &&
882 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
883 /* get default stripe info for this dir. */
884 ma->ma_need |= MA_LOV_DEF;
886 ma->ma_need |= ma_need;
888 rc = mdt_attr_get_complex(info, o, ma);
890 CERROR("%s: getattr error for "DFID": rc = %d\n",
891 mdt_obd_name(info->mti_mdt),
892 PFID(mdt_object_fid(o)), rc);
896 /* if file is released, check if a restore is running */
897 if (ma->ma_valid & MA_HSM) {
898 repbody->mbo_valid |= OBD_MD_TSTATE;
899 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
900 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
901 repbody->mbo_t_state = MS_RESTORE;
904 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
906 /* the Lustre protocol supposes to return default striping
907 * on the user-visible root if explicitly requested */
908 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
909 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
910 struct lu_fid rootfid;
911 struct mdt_object *root;
912 struct mdt_device *mdt = info->mti_mdt;
914 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
917 root = mdt_object_find(env, mdt, &rootfid);
919 RETURN(PTR_ERR(root));
920 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
921 mdt_object_put(info->mti_env, root);
923 CERROR("%s: getattr error for "DFID": rc = %d\n",
924 mdt_obd_name(info->mti_mdt),
925 PFID(mdt_object_fid(o)), rc);
930 if (likely(ma->ma_valid & MA_INODE))
931 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
935 if (mdt_body_has_lov(la, reqbody)) {
936 if (ma->ma_valid & MA_LOV) {
937 LASSERT(ma->ma_lmm_size);
938 repbody->mbo_eadatasize = ma->ma_lmm_size;
939 if (S_ISDIR(la->la_mode))
940 repbody->mbo_valid |= OBD_MD_FLDIREA;
942 repbody->mbo_valid |= OBD_MD_FLEASIZE;
943 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
945 if (ma->ma_valid & MA_LMV) {
946 /* Return -ENOTSUPP for old client */
947 if (!mdt_is_striped_client(req->rq_export))
950 LASSERT(S_ISDIR(la->la_mode));
951 mdt_dump_lmv(D_INFO, ma->ma_lmv);
952 repbody->mbo_eadatasize = ma->ma_lmv_size;
953 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
955 if (ma->ma_valid & MA_LMV_DEF) {
956 /* Return -ENOTSUPP for old client */
957 if (!mdt_is_striped_client(req->rq_export))
959 LASSERT(S_ISDIR(la->la_mode));
960 repbody->mbo_eadatasize = ma->ma_lmv_size;
961 repbody->mbo_valid |= (OBD_MD_FLDIREA |
964 } else if (S_ISLNK(la->la_mode) &&
965 reqbody->mbo_valid & OBD_MD_LINKNAME) {
966 buffer->lb_buf = ma->ma_lmm;
967 /* eadatasize from client includes NULL-terminator, so
968 * there is no need to read it */
969 buffer->lb_len = reqbody->mbo_eadatasize - 1;
970 rc = mo_readlink(env, next, buffer);
971 if (unlikely(rc <= 0)) {
972 CERROR("%s: readlink failed for "DFID": rc = %d\n",
973 mdt_obd_name(info->mti_mdt),
974 PFID(mdt_object_fid(o)), rc);
977 int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
979 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
981 repbody->mbo_valid |= OBD_MD_LINKNAME;
982 /* we need to report back size with NULL-terminator
983 * because client expects that */
984 repbody->mbo_eadatasize = rc + 1;
985 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
986 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
987 "on "DFID ", expected %d\n",
988 mdt_obd_name(info->mti_mdt),
989 rc, PFID(mdt_object_fid(o)),
990 reqbody->mbo_eadatasize - 1);
992 ((char *)ma->ma_lmm)[rc] = 0;
994 /* If the total CDEBUG() size is larger than a page, it
995 * will print a warning to the console, avoid this by
996 * printing just the last part of the symlink. */
997 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
998 print_limit < rc ? "..." : "", print_limit,
999 (char *)ma->ma_lmm + rc - print_limit, rc);
1004 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1005 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1006 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1007 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1008 repbody->mbo_max_mdsize);
1011 if (exp_connect_rmtclient(info->mti_exp) &&
1012 reqbody->mbo_valid & OBD_MD_FLRMTPERM) {
1013 void *buf = req_capsule_server_get(pill, &RMF_ACL);
1015 /* mdt_getattr_lock only */
1016 rc = mdt_pack_remote_perm(info, o, buf);
1018 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
1019 repbody->mbo_aclsize = 0;
1022 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
1023 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
1026 #ifdef CONFIG_FS_POSIX_ACL
1027 else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1028 (reqbody->mbo_valid & OBD_MD_FLACL))
1029 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1034 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1039 static int mdt_getattr(struct tgt_session_info *tsi)
1041 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1042 struct mdt_object *obj = info->mti_object;
1043 struct req_capsule *pill = info->mti_pill;
1044 struct mdt_body *reqbody;
1045 struct mdt_body *repbody;
1049 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1051 LASSERT(obj != NULL);
1052 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1054 /* Unlike intent case where we need to pre-fill out buffers early on
1055 * in intent policy for ldlm reasons, here we can have a much better
1056 * guess at EA size by just reading it from disk.
1057 * Exceptions are readdir and (missing) directory striping */
1059 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1060 /* No easy way to know how long is the symlink, but it cannot
1061 * be more than PATH_MAX, so we allocate +1 */
1064 /* A special case for fs ROOT: getattr there might fetch
1065 * default EA for entire fs, not just for this dir!
1067 } else if (lu_fid_eq(mdt_object_fid(obj),
1068 &info->mti_mdt->mdt_md_root_fid) &&
1069 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1070 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1072 /* Should the default strping be bigger, mdt_fix_reply
1073 * will reallocate */
1074 rc = DEF_REP_MD_SIZE;
1076 /* Read the actual EA size from disk */
1077 rc = mdt_attr_get_eabuf_size(info, obj);
1081 GOTO(out, rc = err_serious(rc));
1083 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1085 rc = req_capsule_server_pack(pill);
1086 if (unlikely(rc != 0))
1087 GOTO(out, rc = err_serious(rc));
1089 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1090 LASSERT(repbody != NULL);
1091 repbody->mbo_eadatasize = 0;
1092 repbody->mbo_aclsize = 0;
1094 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1095 rc = mdt_init_ucred(info, reqbody);
1097 rc = mdt_check_ucred(info);
1099 GOTO(out_shrink, rc);
1101 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1103 rc = mdt_getattr_internal(info, obj, 0);
1104 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1105 mdt_exit_ucred(info);
1108 mdt_client_compatibility(info);
1109 rc2 = mdt_fix_reply(info);
1113 mdt_thread_info_fini(info);
1118 * Exchange MOF_LOV_CREATED flags between two objects after a
1119 * layout swap. No assumption is made on whether o1 or o2 have
1120 * created objects or not.
1122 * \param[in,out] o1 First swap layout object
1123 * \param[in,out] o2 Second swap layout object
1125 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1129 mutex_lock(&o1->mot_lov_mutex);
1130 mutex_lock(&o2->mot_lov_mutex);
1132 o1_flags = o1->mot_flags;
1133 o1->mot_flags = (o1->mot_flags & ~MOF_LOV_CREATED) |
1134 (o2->mot_flags & MOF_LOV_CREATED);
1136 o2->mot_flags = (o2->mot_flags & ~MOF_LOV_CREATED) |
1137 (o1_flags & MOF_LOV_CREATED);
1139 mutex_unlock(&o2->mot_lov_mutex);
1140 mutex_unlock(&o1->mot_lov_mutex);
1143 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1145 struct mdt_thread_info *info;
1146 struct ptlrpc_request *req = tgt_ses_req(tsi);
1147 struct obd_export *exp = req->rq_export;
1148 struct mdt_object *o1, *o2, *o;
1149 struct mdt_lock_handle *lh1, *lh2;
1150 struct mdc_swap_layouts *msl;
1154 /* client does not support layout lock, so layout swaping
1156 * FIXME: there is a problem for old clients which don't support
1157 * layout lock yet. If those clients have already opened the file
1158 * they won't be notified at all so that old layout may still be
1159 * used to do IO. This can be fixed after file release is landed by
1160 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1161 if (!exp_connect_layout(exp))
1162 RETURN(-EOPNOTSUPP);
1164 info = tsi2mdt_info(tsi);
1166 if (info->mti_dlm_req != NULL)
1167 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1169 o1 = info->mti_object;
1170 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1171 &info->mti_body->mbo_fid2);
1173 GOTO(out, rc = PTR_ERR(o));
1175 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1176 GOTO(put, rc = -ENOENT);
1178 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1179 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1185 /* permission check. Make sure the calling process having permission
1186 * to write both files. */
1187 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1192 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1197 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1199 GOTO(put, rc = -EPROTO);
1201 lh1 = &info->mti_lh[MDT_LH_NEW];
1202 mdt_lock_reg_init(lh1, LCK_EX);
1203 lh2 = &info->mti_lh[MDT_LH_OLD];
1204 mdt_lock_reg_init(lh2, LCK_EX);
1206 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1207 MDS_INODELOCK_XATTR);
1211 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1212 MDS_INODELOCK_XATTR);
1216 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1217 mdt_object_child(o2), msl->msl_flags);
1221 mdt_swap_lov_flag(o1, o2);
1224 mdt_object_unlock(info, o2, lh2, rc);
1226 mdt_object_unlock(info, o1, lh1, rc);
1228 mdt_object_put(info->mti_env, o);
1230 mdt_thread_info_fini(info);
1234 static int mdt_raw_lookup(struct mdt_thread_info *info,
1235 struct mdt_object *parent,
1236 const struct lu_name *lname,
1237 struct ldlm_reply *ldlm_rep)
1239 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1243 LASSERT(!info->mti_cross_ref);
1245 /* Only got the fid of this obj by name */
1246 fid_zero(child_fid);
1247 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1248 lname, child_fid, &info->mti_spec);
1250 struct mdt_body *repbody;
1252 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1253 repbody->mbo_fid1 = *child_fid;
1254 repbody->mbo_valid = OBD_MD_FLID;
1255 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1256 } else if (rc == -ENOENT) {
1257 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1264 * UPDATE lock should be taken against parent, and be released before exit;
1265 * child_bits lock should be taken against child, and be returned back:
1266 * (1)normal request should release the child lock;
1267 * (2)intent request will grant the lock to client.
1269 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1270 struct mdt_lock_handle *lhc,
1272 struct ldlm_reply *ldlm_rep)
1274 struct ptlrpc_request *req = mdt_info_req(info);
1275 struct mdt_body *reqbody = NULL;
1276 struct mdt_object *parent = info->mti_object;
1277 struct mdt_object *child;
1278 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1279 struct lu_name *lname = NULL;
1280 struct mdt_lock_handle *lhp = NULL;
1281 struct ldlm_lock *lock;
1288 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1289 LASSERT(ergo(is_resent,
1290 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1295 if (info->mti_cross_ref) {
1296 /* Only getattr on the child. Parent is on another node. */
1297 mdt_set_disposition(info, ldlm_rep,
1298 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1300 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1302 PFID(mdt_object_fid(child)), ldlm_rep);
1304 rc = mdt_check_resent_lock(info, child, lhc);
1307 } else if (rc > 0) {
1308 mdt_lock_handle_init(lhc);
1309 mdt_lock_reg_init(lhc, LCK_PR);
1312 * Object's name is on another MDS, no lookup or layout
1313 * lock is needed here but update lock is.
1315 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1316 MDS_INODELOCK_LAYOUT);
1317 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1319 rc = mdt_object_lock(info, child, lhc, child_bits);
1324 /* Finally, we can get attr for child. */
1325 if (!mdt_object_exists(child)) {
1326 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1328 "remote object doesn't exist.\n");
1329 mdt_object_unlock(info, child, lhc, 1);
1333 rc = mdt_getattr_internal(info, child, 0);
1334 if (unlikely(rc != 0))
1335 mdt_object_unlock(info, child, lhc, 1);
1340 lname = &info->mti_name;
1341 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1343 if (lu_name_is_valid(lname)) {
1344 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1345 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1346 PNAME(lname), ldlm_rep);
1348 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1349 if (unlikely(reqbody == NULL))
1350 RETURN(err_serious(-EPROTO));
1352 *child_fid = reqbody->mbo_fid2;
1354 if (unlikely(!fid_is_sane(child_fid)))
1355 RETURN(err_serious(-EINVAL));
1357 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1359 PFID(mdt_object_fid(parent)),
1360 PFID(&reqbody->mbo_fid2), ldlm_rep);
1363 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1365 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1366 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1368 "Parent doesn't exist!\n");
1372 if (mdt_object_remote(parent)) {
1373 CERROR("%s: parent "DFID" is on remote target\n",
1374 mdt_obd_name(info->mti_mdt),
1375 PFID(mdt_object_fid(parent)));
1379 if (lu_name_is_valid(lname)) {
1380 /* Always allow to lookup ".." */
1381 if (unlikely(lname->ln_namelen == 2 &&
1382 lname->ln_name[0] == '.' &&
1383 lname->ln_name[1] == '.'))
1384 info->mti_spec.sp_permitted = 1;
1386 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1387 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1392 /* step 1: lock parent only if parent is a directory */
1393 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1394 lhp = &info->mti_lh[MDT_LH_PARENT];
1395 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1396 rc = mdt_object_lock(info, parent, lhp,
1397 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 if (!mdt_object_exists(child)) {
1437 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1439 "Object doesn't exist!\n");
1440 GOTO(out_child, rc = -ENOENT);
1443 rc = mdt_check_resent_lock(info, child, lhc);
1445 GOTO(out_child, rc);
1446 } else if (rc > 0) {
1447 mdt_lock_handle_init(lhc);
1448 mdt_lock_reg_init(lhc, LCK_PR);
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,
1491 child_bits &= ~MDS_INODELOCK_LAYOUT;
1492 LASSERT(child_bits != 0);
1493 rc = mdt_object_lock(info, child, lhc,
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);
1505 if (unlikely(rc != 0))
1506 GOTO(out_child, rc);
1509 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1511 /* finally, we can get attr for child. */
1512 rc = mdt_getattr_internal(info, child, ma_need);
1513 if (unlikely(rc != 0)) {
1514 mdt_object_unlock(info, child, lhc, 1);
1516 /* Debugging code. */
1517 LDLM_DEBUG(lock, "Returning lock to client");
1518 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1519 &lock->l_resource->lr_name),
1520 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1521 PLDLMRES(lock->l_resource),
1522 PFID(mdt_object_fid(child)));
1525 LDLM_LOCK_PUT(lock);
1529 mdt_object_put(info->mti_env, child);
1532 mdt_object_unlock(info, parent, lhp, 1);
1536 /* normal handler: should release the child lock */
1537 static int mdt_getattr_name(struct tgt_session_info *tsi)
1539 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1540 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1541 struct mdt_body *reqbody;
1542 struct mdt_body *repbody;
1546 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1547 LASSERT(reqbody != NULL);
1548 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1549 LASSERT(repbody != NULL);
1551 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1552 repbody->mbo_eadatasize = 0;
1553 repbody->mbo_aclsize = 0;
1555 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1557 GOTO(out_shrink, rc);
1559 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1560 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1561 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1562 lhc->mlh_reg_lh.cookie = 0;
1564 mdt_exit_ucred(info);
1567 mdt_client_compatibility(info);
1568 rc2 = mdt_fix_reply(info);
1571 mdt_thread_info_fini(info);
1575 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1576 void *karg, void *uarg);
1578 static int mdt_set_info(struct tgt_session_info *tsi)
1580 struct ptlrpc_request *req = tgt_ses_req(tsi);
1583 int keylen, vallen, rc = 0;
1587 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1589 DEBUG_REQ(D_HA, req, "no set_info key");
1590 RETURN(err_serious(-EFAULT));
1593 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1596 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1598 DEBUG_REQ(D_HA, req, "no set_info val");
1599 RETURN(err_serious(-EFAULT));
1602 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1605 /* Swab any part of val you need to here */
1606 if (KEY_IS(KEY_READ_ONLY)) {
1607 spin_lock(&req->rq_export->exp_lock);
1609 *exp_connect_flags_ptr(req->rq_export) |=
1612 *exp_connect_flags_ptr(req->rq_export) &=
1613 ~OBD_CONNECT_RDONLY;
1614 spin_unlock(&req->rq_export->exp_lock);
1615 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1616 struct changelog_setinfo *cs = val;
1618 if (vallen != sizeof(*cs)) {
1619 CERROR("%s: bad changelog_clear setinfo size %d\n",
1620 tgt_name(tsi->tsi_tgt), vallen);
1623 if (ptlrpc_req_need_swab(req)) {
1624 __swab64s(&cs->cs_recno);
1625 __swab32s(&cs->cs_id);
1628 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1630 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
1632 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
1639 static int mdt_readpage(struct tgt_session_info *tsi)
1641 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1642 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1643 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1644 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1645 struct mdt_body *repbody;
1651 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1652 RETURN(err_serious(-ENOMEM));
1654 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1655 if (repbody == NULL || reqbody == NULL)
1656 RETURN(err_serious(-EFAULT));
1659 * prepare @rdpg before calling lower layers and transfer itself. Here
1660 * reqbody->size contains offset of where to start to read and
1661 * reqbody->nlink contains number bytes to read.
1663 rdpg->rp_hash = reqbody->mbo_size;
1664 if (rdpg->rp_hash != reqbody->mbo_size) {
1665 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1666 rdpg->rp_hash, reqbody->mbo_size);
1670 rdpg->rp_attrs = reqbody->mbo_mode;
1671 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1672 rdpg->rp_attrs |= LUDA_64BITHASH;
1673 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1674 exp_max_brw_size(tsi->tsi_exp));
1675 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1677 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1678 if (rdpg->rp_pages == NULL)
1681 for (i = 0; i < rdpg->rp_npages; ++i) {
1682 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1683 if (rdpg->rp_pages[i] == NULL)
1684 GOTO(free_rdpg, rc = -ENOMEM);
1687 /* call lower layers to fill allocated pages with directory data */
1688 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1690 GOTO(free_rdpg, rc);
1692 /* send pages to client */
1693 rc = tgt_sendpage(tsi, rdpg, rc);
1698 for (i = 0; i < rdpg->rp_npages; i++)
1699 if (rdpg->rp_pages[i] != NULL)
1700 __free_page(rdpg->rp_pages[i]);
1701 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1703 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1709 static int mdt_reint_internal(struct mdt_thread_info *info,
1710 struct mdt_lock_handle *lhc,
1713 struct req_capsule *pill = info->mti_pill;
1714 struct mdt_body *repbody;
1719 rc = mdt_reint_unpack(info, op);
1721 CERROR("Can't unpack reint, rc %d\n", rc);
1722 RETURN(err_serious(rc));
1725 /* for replay (no_create) lmm is not needed, client has it already */
1726 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1727 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1730 /* llog cookies are always 0, the field is kept for compatibility */
1731 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1732 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1734 rc = req_capsule_server_pack(pill);
1736 CERROR("Can't pack response, rc %d\n", rc);
1737 RETURN(err_serious(rc));
1740 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1741 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1743 repbody->mbo_eadatasize = 0;
1744 repbody->mbo_aclsize = 0;
1747 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1749 /* for replay no cookkie / lmm need, because client have this already */
1750 if (info->mti_spec.no_create)
1751 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1752 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1754 rc = mdt_init_ucred_reint(info);
1756 GOTO(out_shrink, rc);
1758 rc = mdt_fix_attr_ucred(info, op);
1760 GOTO(out_ucred, rc = err_serious(rc));
1762 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
1764 GOTO(out_ucred, rc);
1765 } else if (rc == 1) {
1766 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1767 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1768 GOTO(out_ucred, rc);
1770 rc = mdt_reint_rec(info, lhc);
1773 mdt_exit_ucred(info);
1775 mdt_client_compatibility(info);
1776 rc2 = mdt_fix_reply(info);
1782 static long mdt_reint_opcode(struct ptlrpc_request *req,
1783 const struct req_format **fmt)
1785 struct mdt_device *mdt;
1786 struct mdt_rec_reint *rec;
1789 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1791 opc = rec->rr_opcode;
1792 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1793 if (opc < REINT_MAX && fmt[opc] != NULL)
1794 req_capsule_extend(&req->rq_pill, fmt[opc]);
1796 mdt = mdt_exp2dev(req->rq_export);
1797 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1798 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1799 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1800 opc = err_serious(-EFAULT);
1803 opc = err_serious(-EFAULT);
1808 static int mdt_reint(struct tgt_session_info *tsi)
1812 static const struct req_format *reint_fmts[REINT_MAX] = {
1813 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1814 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1815 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1816 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1817 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1818 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1819 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1820 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
1821 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME
1826 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1828 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1830 * No lock possible here from client to pass it to reint code
1833 rc = mdt_reint_internal(info, NULL, opc);
1834 mdt_thread_info_fini(info);
1839 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1843 /* this should sync the whole device */
1844 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1846 struct dt_device *dt = mdt->mdt_bottom;
1850 rc = dt->dd_ops->dt_sync(env, dt);
1854 /* this should sync this object */
1855 static int mdt_object_sync(struct mdt_thread_info *info)
1857 struct md_object *next;
1861 if (!mdt_object_exists(info->mti_object)) {
1862 CWARN("%s: non existing object "DFID": rc = %d\n",
1863 mdt_obd_name(info->mti_mdt),
1864 PFID(mdt_object_fid(info->mti_object)), -ESTALE);
1867 next = mdt_object_child(info->mti_object);
1868 rc = mo_object_sync(info->mti_env, next);
1873 static int mdt_sync(struct tgt_session_info *tsi)
1875 struct ptlrpc_request *req = tgt_ses_req(tsi);
1876 struct req_capsule *pill = tsi->tsi_pill;
1877 struct mdt_body *body;
1882 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1883 RETURN(err_serious(-ENOMEM));
1885 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
1886 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
1888 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1890 /* sync an object */
1891 rc = mdt_object_sync(info);
1893 const struct lu_fid *fid;
1894 struct lu_attr *la = &info->mti_attr.ma_attr;
1896 info->mti_attr.ma_need = MA_INODE;
1897 info->mti_attr.ma_valid = 0;
1898 rc = mdt_attr_get_complex(info, info->mti_object,
1901 body = req_capsule_server_get(pill,
1903 fid = mdt_object_fid(info->mti_object);
1904 mdt_pack_attr2body(info, body, la, fid);
1907 mdt_thread_info_fini(info);
1910 mdt_counter_incr(req, LPROC_MDT_SYNC);
1916 * Handle quota control requests to consult current usage/limit, but also
1917 * to configure quota enforcement
1919 static int mdt_quotactl(struct tgt_session_info *tsi)
1921 struct obd_export *exp = tsi->tsi_exp;
1922 struct req_capsule *pill = tsi->tsi_pill;
1923 struct obd_quotactl *oqctl, *repoqc;
1925 struct mdt_device *mdt = mdt_exp2dev(exp);
1926 struct lu_device *qmt = mdt->mdt_qmt_dev;
1927 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
1930 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1932 RETURN(err_serious(-EPROTO));
1934 rc = req_capsule_server_pack(pill);
1936 RETURN(err_serious(rc));
1938 switch (oqctl->qc_cmd) {
1939 /* master quotactl */
1942 if (!nodemap_can_setquota(nodemap))
1947 RETURN(-EOPNOTSUPP);
1948 /* slave quotactl */
1953 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1957 /* map uid/gid for remote client */
1959 if (exp_connect_rmtclient(exp)) {
1960 struct lustre_idmap_table *idmap;
1962 idmap = exp->exp_mdt_data.med_idmap;
1964 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1965 oqctl->qc_cmd != Q_GETINFO))
1968 if (oqctl->qc_type == USRQUOTA)
1969 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1971 else if (oqctl->qc_type == GRPQUOTA)
1972 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1977 if (id == CFS_IDMAP_NOTFOUND) {
1978 CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
1983 if (oqctl->qc_type == USRQUOTA)
1984 id = nodemap_map_id(nodemap, NODEMAP_UID,
1985 NODEMAP_CLIENT_TO_FS, id);
1986 else if (oqctl->qc_type == GRPQUOTA)
1987 id = nodemap_map_id(nodemap, NODEMAP_UID,
1988 NODEMAP_CLIENT_TO_FS, id);
1990 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1992 RETURN(err_serious(-EFAULT));
1994 if (oqctl->qc_id != id)
1995 swap(oqctl->qc_id, id);
1997 switch (oqctl->qc_cmd) {
2003 /* forward quotactl request to QMT */
2004 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2009 /* slave quotactl */
2010 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2015 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2019 if (oqctl->qc_id != id)
2020 swap(oqctl->qc_id, id);
2026 /** clone llog ctxt from child (mdd)
2027 * This allows remote llog (replicator) access.
2028 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2029 * context was originally set up, or we can handle them directly.
2030 * I choose the latter, but that means I need any llog
2031 * contexts set up by child to be accessable by the mdt. So we clone the
2032 * context into our context list here.
2034 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2037 struct md_device *next = mdt->mdt_child;
2038 struct llog_ctxt *ctxt;
2041 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2044 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2045 if (rc || ctxt == NULL) {
2049 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2051 CERROR("Can't set mdt ctxt %d\n", rc);
2056 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2057 struct mdt_device *mdt, int idx)
2059 struct llog_ctxt *ctxt;
2061 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2064 /* Put once for the get we just did, and once for the clone */
2065 llog_ctxt_put(ctxt);
2066 llog_ctxt_put(ctxt);
2071 * sec context handlers
2073 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2077 rc = mdt_handle_idmap(tsi);
2079 struct ptlrpc_request *req = tgt_ses_req(tsi);
2082 opc = lustre_msg_get_opc(req->rq_reqmsg);
2083 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2084 sptlrpc_svc_ctx_invalidate(req);
2087 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2093 * quota request handlers
2095 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2097 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2098 struct lu_device *qmt = mdt->mdt_qmt_dev;
2103 RETURN(err_serious(-EOPNOTSUPP));
2105 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2109 struct mdt_object *mdt_object_new(const struct lu_env *env,
2110 struct mdt_device *d,
2111 const struct lu_fid *f)
2113 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2114 struct lu_object *o;
2115 struct mdt_object *m;
2118 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2119 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2120 if (unlikely(IS_ERR(o)))
2121 m = (struct mdt_object *)o;
2127 struct mdt_object *mdt_object_find(const struct lu_env *env,
2128 struct mdt_device *d,
2129 const struct lu_fid *f)
2131 struct lu_object *o;
2132 struct mdt_object *m;
2135 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2136 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2137 if (unlikely(IS_ERR(o)))
2138 m = (struct mdt_object *)o;
2146 * Asyncronous commit for mdt device.
2148 * Pass asynchonous commit call down the MDS stack.
2150 * \param env environment
2151 * \param mdt the mdt device
2153 static void mdt_device_commit_async(const struct lu_env *env,
2154 struct mdt_device *mdt)
2156 struct dt_device *dt = mdt->mdt_bottom;
2159 rc = dt->dd_ops->dt_commit_async(env, dt);
2160 if (unlikely(rc != 0))
2161 CWARN("%s: async commit start failed: rc = %d\n",
2162 mdt_obd_name(mdt), rc);
2166 * Mark the lock as "synchonous".
2168 * Mark the lock to deffer transaction commit to the unlock time.
2170 * \param lock the lock to mark as "synchonous"
2172 * \see mdt_is_lock_sync
2173 * \see mdt_save_lock
2175 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2177 lock->l_ast_data = (void*)1;
2181 * Check whehter the lock "synchonous" or not.
2183 * \param lock the lock to check
2184 * \retval 1 the lock is "synchonous"
2185 * \retval 0 the lock isn't "synchronous"
2187 * \see mdt_set_lock_sync
2188 * \see mdt_save_lock
2190 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2192 return lock->l_ast_data != NULL;
2196 * Blocking AST for mdt locks.
2198 * Starts transaction commit if in case of COS lock conflict or
2199 * deffers such a commit to the mdt_save_lock.
2201 * \param lock the lock which blocks a request or cancelling lock
2202 * \param desc unused
2203 * \param data unused
2204 * \param flag indicates whether this cancelling or blocking callback
2206 * \see ldlm_blocking_ast_nocheck
2208 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2209 void *data, int flag)
2211 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2212 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2216 if (flag == LDLM_CB_CANCELING)
2218 lock_res_and_lock(lock);
2219 if (lock->l_blocking_ast != mdt_blocking_ast) {
2220 unlock_res_and_lock(lock);
2223 if (mdt_cos_is_enabled(mdt) &&
2224 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2225 lock->l_blocking_lock != NULL &&
2226 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2227 mdt_set_lock_sync(lock);
2229 rc = ldlm_blocking_ast_nocheck(lock);
2231 /* There is no lock conflict if l_blocking_lock == NULL,
2232 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2233 * when the last reference to a local lock was released */
2234 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2237 rc = lu_env_init(&env, LCT_LOCAL);
2238 if (unlikely(rc != 0))
2239 CWARN("%s: lu_env initialization failed, cannot "
2240 "start asynchronous commit: rc = %d\n",
2243 mdt_device_commit_async(&env, mdt);
2249 /* Used for cross-MDT lock */
2250 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2251 void *data, int flag)
2253 struct lustre_handle lockh;
2257 case LDLM_CB_BLOCKING:
2258 ldlm_lock2handle(lock, &lockh);
2259 rc = ldlm_cli_cancel(&lockh,
2260 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
2262 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2266 case LDLM_CB_CANCELING:
2267 LDLM_DEBUG(lock, "Revoke remote lock\n");
2275 int mdt_check_resent_lock(struct mdt_thread_info *info,
2276 struct mdt_object *mo,
2277 struct mdt_lock_handle *lhc)
2279 /* the lock might already be gotten in ldlm_handle_enqueue() */
2280 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2281 struct ptlrpc_request *req = mdt_info_req(info);
2282 struct ldlm_lock *lock;
2284 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2285 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2287 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2288 * a resend case, however, it could be already destroyed
2289 * due to client eviction or a raced cancel RPC. */
2290 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
2291 lhc->mlh_reg_lh.cookie);
2295 if (!fid_res_name_eq(mdt_object_fid(mo),
2296 &lock->l_resource->lr_name)) {
2297 CWARN("%s: Although resent, but still not "
2298 "get child lock:"DFID"\n",
2299 info->mti_exp->exp_obd->obd_name,
2300 PFID(mdt_object_fid(mo)));
2301 LDLM_LOCK_PUT(lock);
2304 LDLM_LOCK_PUT(lock);
2310 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
2311 const struct lu_fid *fid, struct lustre_handle *lh,
2312 enum ldlm_mode mode, __u64 ibits, bool nonblock)
2314 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2315 union ldlm_policy_data *policy = &mti->mti_policy;
2316 struct ldlm_res_id *res_id = &mti->mti_res_id;
2320 LASSERT(mdt_object_remote(o));
2322 fid_build_reg_res_name(fid, res_id);
2324 memset(einfo, 0, sizeof(*einfo));
2325 einfo->ei_type = LDLM_IBITS;
2326 einfo->ei_mode = mode;
2327 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2328 einfo->ei_cb_cp = ldlm_completion_ast;
2329 einfo->ei_enq_slave = 0;
2330 einfo->ei_res_id = res_id;
2332 einfo->ei_nonblock = 1;
2334 memset(policy, 0, sizeof(*policy));
2335 policy->l_inodebits.bits = ibits;
2337 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2342 static int mdt_object_local_lock(struct mdt_thread_info *info,
2343 struct mdt_object *o,
2344 struct mdt_lock_handle *lh, __u64 ibits,
2347 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2348 union ldlm_policy_data *policy = &info->mti_policy;
2349 struct ldlm_res_id *res_id = &info->mti_res_id;
2354 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2355 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2356 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2357 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2359 /* Only enqueue LOOKUP lock for remote object */
2360 if (mdt_object_remote(o))
2361 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2363 if (lh->mlh_type == MDT_PDO_LOCK) {
2364 /* check for exists after object is locked */
2365 if (mdt_object_exists(o) == 0) {
2366 /* Non-existent object shouldn't have PDO lock */
2369 /* Non-dir object shouldn't have PDO lock */
2370 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2375 memset(policy, 0, sizeof(*policy));
2376 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2378 dlmflags = LDLM_FL_ATOMIC_CB;
2380 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2383 * Take PDO lock on whole directory and build correct @res_id for lock
2384 * on part of directory.
2386 if (lh->mlh_pdo_hash != 0) {
2387 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2388 mdt_lock_pdo_mode(info, o, lh);
2389 if (lh->mlh_pdo_mode != LCK_NL) {
2391 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2392 * is never going to be sent to client and we do not
2393 * want it slowed down due to possible cancels.
2395 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2396 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2397 policy, res_id, dlmflags,
2398 info->mti_exp == NULL ? NULL :
2399 &info->mti_exp->exp_handle.h_cookie);
2400 if (unlikely(rc != 0))
2401 GOTO(out_unlock, rc);
2405 * Finish res_id initializing by name hash marking part of
2406 * directory which is taking modification.
2408 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2411 policy->l_inodebits.bits = ibits;
2414 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2415 * going to be sent to client. If it is - mdt_intent_policy() path will
2416 * fix it up and turn FL_LOCAL flag off.
2418 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2419 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2420 info->mti_exp == NULL ? NULL :
2421 &info->mti_exp->exp_handle.h_cookie);
2424 mdt_object_unlock(info, o, lh, 1);
2425 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2426 lh->mlh_pdo_hash != 0 &&
2427 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2428 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2434 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2435 struct mdt_lock_handle *lh, __u64 ibits,
2438 struct mdt_lock_handle *local_lh = NULL;
2442 if (!mdt_object_remote(o))
2443 return mdt_object_local_lock(info, o, lh, ibits, nonblock);
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);
2449 /* Only enqueue LOOKUP lock for remote object */
2450 if (ibits & MDS_INODELOCK_LOOKUP) {
2451 rc = mdt_object_local_lock(info, o, lh,
2452 MDS_INODELOCK_LOOKUP,
2460 if (ibits & MDS_INODELOCK_UPDATE) {
2461 /* Sigh, PDO needs to enqueue 2 locks right now, but
2462 * enqueue RPC can only request 1 lock, to avoid extra
2463 * RPC, so it will instead enqueue EX lock for remote
2464 * object anyway XXX*/
2465 if (lh->mlh_type == MDT_PDO_LOCK &&
2466 lh->mlh_pdo_hash != 0) {
2467 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2468 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2469 PFID(mdt_object_fid(o)));
2470 lh->mlh_pdo_hash = 0;
2471 lh->mlh_rreg_mode = LCK_EX;
2472 lh->mlh_type = MDT_REG_LOCK;
2474 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2477 MDS_INODELOCK_UPDATE, nonblock);
2478 if (rc != ELDLM_OK) {
2479 if (local_lh != NULL)
2480 mdt_object_unlock(info, o, local_lh, rc);
2488 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2489 struct mdt_lock_handle *lh, __u64 ibits)
2491 return mdt_object_lock_internal(info, o, lh, ibits, false);
2494 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2495 struct mdt_lock_handle *lh, __u64 ibits)
2497 struct mdt_lock_handle tmp = *lh;
2500 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true);
2508 * Save a lock within request object.
2510 * Keep the lock referenced until whether client ACK or transaction
2511 * commit happens or release the lock immediately depending on input
2512 * parameters. If COS is ON, a write lock is converted to COS lock
2515 * \param info thead info object
2516 * \param h lock handle
2517 * \param mode lock mode
2518 * \param decref force immediate lock releasing
2520 static void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2521 enum ldlm_mode mode, int decref)
2525 if (lustre_handle_is_used(h)) {
2526 if (decref || !info->mti_has_trans ||
2527 !(mode & (LCK_PW | LCK_EX))) {
2528 mdt_fid_unlock(h, mode);
2530 struct mdt_device *mdt = info->mti_mdt;
2531 struct ldlm_lock *lock = ldlm_handle2lock(h);
2532 struct ptlrpc_request *req = mdt_info_req(info);
2535 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2537 /* there is no request if mdt_object_unlock() is called
2538 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2539 if (likely(req != NULL)) {
2540 CDEBUG(D_HA, "request = %p reply state = %p"
2541 " transno = "LPD64"\n", req,
2542 req->rq_reply_state, req->rq_transno);
2543 if (mdt_cos_is_enabled(mdt)) {
2545 ldlm_lock_downgrade(lock, LCK_COS);
2548 ptlrpc_save_lock(req, h, mode, no_ack);
2550 ldlm_lock_decref(h, mode);
2552 if (mdt_is_lock_sync(lock)) {
2553 CDEBUG(D_HA, "found sync-lock,"
2554 " async commit started\n");
2555 mdt_device_commit_async(info->mti_env,
2558 LDLM_LOCK_PUT(lock);
2567 * Unlock mdt object.
2569 * Immeditely release the regular lock and the PDO lock or save the
2570 * lock in request and keep them referenced until client ACK or
2571 * transaction commit.
2573 * \param info thread info object
2574 * \param o mdt object
2575 * \param lh mdt lock handle referencing regular and PDO locks
2576 * \param decref force immediate lock releasing
2578 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2579 struct mdt_lock_handle *lh, int decref)
2583 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2584 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2586 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2587 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2592 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2593 const struct lu_fid *f,
2594 struct mdt_lock_handle *lh,
2597 struct mdt_object *o;
2599 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2603 rc = mdt_object_lock(info, o, lh, ibits);
2605 mdt_object_put(info->mti_env, o);
2612 void mdt_object_unlock_put(struct mdt_thread_info * info,
2613 struct mdt_object * o,
2614 struct mdt_lock_handle *lh,
2617 mdt_object_unlock(info, o, lh, decref);
2618 mdt_object_put(info->mti_env, o);
2622 * Generic code handling requests that have struct mdt_body passed in:
2624 * - extract mdt_body from request and save it in @info, if present;
2626 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2629 * - if HABEO_CORPUS flag is set for this request type check whether object
2630 * actually exists on storage (lu_object_exists()).
2633 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2635 const struct mdt_body *body;
2636 struct mdt_object *obj;
2637 const struct lu_env *env;
2638 struct req_capsule *pill;
2642 env = info->mti_env;
2643 pill = info->mti_pill;
2645 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2649 if (!(body->mbo_valid & OBD_MD_FLID))
2652 if (!fid_is_sane(&body->mbo_fid1)) {
2653 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2657 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2659 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2660 mdt_object_put(env, obj);
2663 info->mti_object = obj;
2672 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2674 struct req_capsule *pill = info->mti_pill;
2678 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2679 rc = mdt_body_unpack(info, flags);
2683 if (rc == 0 && (flags & HABEO_REFERO)) {
2685 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2686 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2688 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2689 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2692 rc = req_capsule_server_pack(pill);
2697 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2699 lh->mlh_type = MDT_NUL_LOCK;
2700 lh->mlh_reg_lh.cookie = 0ull;
2701 lh->mlh_reg_mode = LCK_MINMODE;
2702 lh->mlh_pdo_lh.cookie = 0ull;
2703 lh->mlh_pdo_mode = LCK_MINMODE;
2704 lh->mlh_rreg_lh.cookie = 0ull;
2705 lh->mlh_rreg_mode = LCK_MINMODE;
2708 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2710 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2711 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2715 * Initialize fields of struct mdt_thread_info. Other fields are left in
2716 * uninitialized state, because it's too expensive to zero out whole
2717 * mdt_thread_info (> 1K) on each request arrival.
2719 void mdt_thread_info_init(struct ptlrpc_request *req,
2720 struct mdt_thread_info *info)
2724 info->mti_pill = &req->rq_pill;
2727 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2728 mdt_lock_handle_init(&info->mti_lh[i]);
2730 /* mdt device: it can be NULL while CONNECT */
2731 if (req->rq_export) {
2732 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2733 info->mti_exp = req->rq_export;
2735 info->mti_mdt = NULL;
2736 info->mti_env = req->rq_svc_thread->t_env;
2737 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2739 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2740 info->mti_big_buf = LU_BUF_NULL;
2741 info->mti_body = NULL;
2742 info->mti_object = NULL;
2743 info->mti_dlm_req = NULL;
2744 info->mti_has_trans = 0;
2745 info->mti_cross_ref = 0;
2746 info->mti_opdata = 0;
2747 info->mti_big_lmm_used = 0;
2749 info->mti_spec.no_create = 0;
2750 info->mti_spec.sp_rm_entry = 0;
2751 info->mti_spec.sp_permitted = 0;
2752 info->mti_spec.sp_migrate_close = 0;
2754 info->mti_spec.u.sp_ea.eadata = NULL;
2755 info->mti_spec.u.sp_ea.eadatalen = 0;
2758 void mdt_thread_info_fini(struct mdt_thread_info *info)
2762 if (info->mti_object != NULL) {
2763 mdt_object_put(info->mti_env, info->mti_object);
2764 info->mti_object = NULL;
2767 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2768 mdt_lock_handle_fini(&info->mti_lh[i]);
2769 info->mti_env = NULL;
2770 info->mti_pill = NULL;
2771 info->mti_exp = NULL;
2773 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2774 lu_buf_free(&info->mti_big_buf);
2777 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2779 struct mdt_thread_info *mti;
2781 mti = mdt_th_info(tsi->tsi_env);
2782 LASSERT(mti != NULL);
2784 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2785 if (tsi->tsi_corpus != NULL) {
2786 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2787 lu_object_get(tsi->tsi_corpus);
2789 mti->mti_body = tsi->tsi_mdt_body;
2790 mti->mti_dlm_req = tsi->tsi_dlm_req;
2795 static int mdt_tgt_connect(struct tgt_session_info *tsi)
2797 struct ptlrpc_request *req = tgt_ses_req(tsi);
2802 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
2804 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
2805 set_current_state(TASK_UNINTERRUPTIBLE);
2806 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
2809 rc = tgt_connect(tsi);
2813 rc = mdt_init_idmap(tsi);
2818 obd_disconnect(class_export_get(req->rq_export));
2837 static int mdt_intent_getattr(enum mdt_it_code opcode,
2838 struct mdt_thread_info *info,
2839 struct ldlm_lock **,
2842 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2843 struct mdt_thread_info *info,
2844 struct ldlm_lock **lockp,
2847 static int mdt_intent_layout(enum mdt_it_code opcode,
2848 struct mdt_thread_info *info,
2849 struct ldlm_lock **,
2851 static int mdt_intent_reint(enum mdt_it_code opcode,
2852 struct mdt_thread_info *info,
2853 struct ldlm_lock **,
2856 static struct mdt_it_flavor {
2857 const struct req_format *it_fmt;
2859 int (*it_act)(enum mdt_it_code ,
2860 struct mdt_thread_info *,
2861 struct ldlm_lock **,
2864 } mdt_it_flavor[] = {
2866 .it_fmt = &RQF_LDLM_INTENT,
2867 /*.it_flags = HABEO_REFERO,*/
2869 .it_act = mdt_intent_reint,
2870 .it_reint = REINT_OPEN
2873 .it_fmt = &RQF_LDLM_INTENT,
2875 * OCREAT is not a MUTABOR request as if the file
2877 * We do the extra check of OBD_CONNECT_RDONLY in
2878 * mdt_reint_open() when we really need to create
2882 .it_act = mdt_intent_reint,
2883 .it_reint = REINT_OPEN
2886 .it_fmt = &RQF_LDLM_INTENT,
2887 .it_flags = MUTABOR,
2888 .it_act = mdt_intent_reint,
2889 .it_reint = REINT_CREATE
2891 [MDT_IT_GETATTR] = {
2892 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2893 .it_flags = HABEO_REFERO,
2894 .it_act = mdt_intent_getattr
2896 [MDT_IT_READDIR] = {
2902 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2903 .it_flags = HABEO_REFERO,
2904 .it_act = mdt_intent_getattr
2907 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2908 .it_flags = MUTABOR,
2910 .it_reint = REINT_UNLINK
2914 .it_flags = MUTABOR,
2917 [MDT_IT_GETXATTR] = {
2918 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2919 .it_flags = HABEO_CORPUS,
2920 .it_act = mdt_intent_getxattr
2923 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2925 .it_act = mdt_intent_layout
2930 mdt_intent_lock_replace(struct mdt_thread_info *info,
2931 struct ldlm_lock **lockp,
2932 struct mdt_lock_handle *lh,
2935 struct ptlrpc_request *req = mdt_info_req(info);
2936 struct ldlm_lock *lock = *lockp;
2937 struct ldlm_lock *new_lock;
2939 /* If possible resent found a lock, @lh is set to its handle */
2940 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2942 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2943 lh->mlh_reg_lh.cookie = 0;
2947 LASSERTF(new_lock != NULL,
2948 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2951 * If we've already given this lock to a client once, then we should
2952 * have no readers or writers. Otherwise, we should have one reader
2953 * _or_ writer ref (which will be zeroed below) before returning the
2956 if (new_lock->l_export == req->rq_export) {
2957 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2959 LASSERT(new_lock->l_export == NULL);
2960 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2965 if (new_lock->l_export == req->rq_export) {
2967 * Already gave this to the client, which means that we
2968 * reconstructed a reply.
2970 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2973 LDLM_LOCK_RELEASE(new_lock);
2974 lh->mlh_reg_lh.cookie = 0;
2975 RETURN(ELDLM_LOCK_REPLACED);
2979 * Fixup the lock to be given to the client.
2981 lock_res_and_lock(new_lock);
2982 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2983 * possible blocking AST. */
2984 while (new_lock->l_readers > 0) {
2985 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2986 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2987 new_lock->l_readers--;
2989 while (new_lock->l_writers > 0) {
2990 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2991 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2992 new_lock->l_writers--;
2995 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2996 new_lock->l_blocking_ast = lock->l_blocking_ast;
2997 new_lock->l_completion_ast = lock->l_completion_ast;
2998 new_lock->l_remote_handle = lock->l_remote_handle;
2999 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3001 unlock_res_and_lock(new_lock);
3003 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3004 &new_lock->l_remote_handle,
3005 &new_lock->l_exp_hash);
3007 LDLM_LOCK_RELEASE(new_lock);
3008 lh->mlh_reg_lh.cookie = 0;
3010 RETURN(ELDLM_LOCK_REPLACED);
3013 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3014 struct ldlm_lock *new_lock,
3015 struct mdt_lock_handle *lh,
3018 struct ptlrpc_request *req = mdt_info_req(info);
3019 struct ldlm_request *dlmreq;
3021 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3024 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3026 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3027 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3029 if (flags & LDLM_FL_RESENT) {
3030 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3031 lh->mlh_reg_mode = new_lock->l_granted_mode;
3033 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3034 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3035 lh->mlh_reg_lh.cookie);
3040 * If the xid matches, then we know this is a resent request, and allow
3041 * it. (It's probably an OPEN, for which we don't send a lock.
3043 if (req_can_reconstruct(req, NULL))
3047 * This remote handle isn't enqueued, so we never received or processed
3048 * this request. Clear MSG_RESENT, because it can be handled like any
3049 * normal request now.
3051 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3053 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3054 dlmreq->lock_handle[0].cookie);
3057 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3058 struct mdt_thread_info *info,
3059 struct ldlm_lock **lockp,
3062 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3063 struct ldlm_reply *ldlm_rep = NULL;
3067 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3068 * (for the resend case) or a new lock. Below we will use it to
3069 * replace the original lock.
3071 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3072 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3073 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3074 rc = mdt_object_lock(info, info->mti_object, lhc,
3075 MDS_INODELOCK_XATTR);
3080 grc = mdt_getxattr(info);
3082 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3084 if (mdt_info_req(info)->rq_repmsg != NULL)
3085 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3086 if (ldlm_rep == NULL)
3087 RETURN(err_serious(-EFAULT));
3089 ldlm_rep->lock_policy_res2 = grc;
3094 static int mdt_intent_getattr(enum mdt_it_code opcode,
3095 struct mdt_thread_info *info,
3096 struct ldlm_lock **lockp,
3099 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3101 struct ldlm_reply *ldlm_rep;
3102 struct mdt_body *reqbody;
3103 struct mdt_body *repbody;
3107 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3110 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3113 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3114 repbody->mbo_eadatasize = 0;
3115 repbody->mbo_aclsize = 0;
3119 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3121 case MDT_IT_GETATTR:
3122 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3126 CERROR("Unsupported intent (%d)\n", opcode);
3127 GOTO(out_shrink, rc = -EINVAL);
3130 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3132 GOTO(out_shrink, rc);
3134 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3135 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3137 /* Get lock from request for possible resent case. */
3138 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3140 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3141 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3143 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3144 ldlm_rep->lock_policy_res2 = 0;
3145 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3146 ldlm_rep->lock_policy_res2) {
3147 lhc->mlh_reg_lh.cookie = 0ull;
3148 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3151 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3154 mdt_exit_ucred(info);
3156 mdt_client_compatibility(info);
3157 rc2 = mdt_fix_reply(info);
3163 static int mdt_intent_layout(enum mdt_it_code opcode,
3164 struct mdt_thread_info *info,
3165 struct ldlm_lock **lockp,
3168 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3169 struct layout_intent *layout;
3171 struct mdt_object *obj = NULL;
3172 int layout_size = 0;
3176 if (opcode != MDT_IT_LAYOUT) {
3177 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3182 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3186 if (layout->li_opc != LAYOUT_INTENT_ACCESS) {
3187 CERROR("%s: Unsupported layout intent opc %d\n",
3188 mdt_obd_name(info->mti_mdt), layout->li_opc);
3192 fid = &info->mti_tmp_fid2;
3193 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3195 /* Get lock from request for possible resent case. */
3196 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3198 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3200 GOTO(out, rc = PTR_ERR(obj));
3202 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3203 layout_size = mdt_attr_get_eabuf_size(info, obj);
3204 if (layout_size < 0)
3205 GOTO(out_obj, rc = layout_size);
3207 if (layout_size > info->mti_mdt->mdt_max_mdsize)
3208 info->mti_mdt->mdt_max_mdsize = layout_size;
3211 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3212 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3214 rc = req_capsule_server_pack(info->mti_pill);
3218 mdt_object_put(info->mti_env, obj);
3220 if (rc == 0 && lustre_handle_is_used(&lhc->mlh_reg_lh))
3221 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3224 lhc->mlh_reg_lh.cookie = 0;
3229 static int mdt_intent_reint(enum mdt_it_code opcode,
3230 struct mdt_thread_info *info,
3231 struct ldlm_lock **lockp,
3234 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3235 struct ldlm_reply *rep = NULL;
3239 static const struct req_format *intent_fmts[REINT_MAX] = {
3240 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3241 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3246 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3250 if (mdt_it_flavor[opcode].it_reint != opc) {
3251 CERROR("Reint code %ld doesn't match intent: %d\n",
3253 RETURN(err_serious(-EPROTO));
3256 /* Get lock from request for possible resent case. */
3257 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3259 rc = mdt_reint_internal(info, lhc, opc);
3261 /* Check whether the reply has been packed successfully. */
3262 if (mdt_info_req(info)->rq_repmsg != NULL)
3263 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3265 RETURN(err_serious(-EFAULT));
3267 /* MDC expects this in any case */
3269 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3271 /* the open lock or the lock for cross-ref object should be
3272 * returned to the client */
3273 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3274 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3275 rep->lock_policy_res2 = 0;
3276 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3280 rep->lock_policy_res2 = clear_serious(rc);
3282 if (rep->lock_policy_res2 == -ENOENT &&
3283 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
3284 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
3285 rep->lock_policy_res2 = 0;
3287 lhc->mlh_reg_lh.cookie = 0ull;
3288 if (rc == -ENOTCONN || rc == -ENODEV ||
3289 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3291 * If it is the disconnect error (ENODEV & ENOCONN), the error
3292 * will be returned by rq_status, and client at ptlrpc layer
3293 * will detect this, then disconnect, reconnect the import
3294 * immediately, instead of impacting the following the rpc.
3299 * For other cases, the error will be returned by intent, and client
3300 * will retrieve the result from intent.
3302 RETURN(ELDLM_LOCK_ABORTED);
3305 static int mdt_intent_code(enum ldlm_intent_flags itcode)
3313 case IT_OPEN|IT_CREAT:
3320 rc = MDT_IT_READDIR;
3323 rc = MDT_IT_GETATTR;
3335 rc = MDT_IT_GETXATTR;
3340 case IT_QUOTA_DQACQ:
3345 CERROR("Unknown intent opcode: 0x%08x\n", itcode);
3352 static int mdt_intent_opc(enum ldlm_intent_flags itopc,
3353 struct mdt_thread_info *info,
3354 struct ldlm_lock **lockp, __u64 flags)
3356 struct req_capsule *pill = info->mti_pill;
3357 struct ptlrpc_request *req = mdt_info_req(info);
3358 struct mdt_it_flavor *flv;
3363 opc = mdt_intent_code(itopc);
3367 if (opc == MDT_IT_QUOTA) {
3368 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3371 RETURN(-EOPNOTSUPP);
3373 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3374 /* pass the request to quota master */
3375 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3376 mdt_info_req(info), lockp,
3381 flv = &mdt_it_flavor[opc];
3382 if (flv->it_fmt != NULL)
3383 req_capsule_extend(pill, flv->it_fmt);
3385 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3389 if (flv->it_flags & MUTABOR &&
3390 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3393 if (flv->it_act != NULL) {
3394 struct ldlm_reply *rep;
3396 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
3398 /* execute policy */
3399 rc = flv->it_act(opc, info, lockp, flags);
3401 /* Check whether the reply has been packed successfully. */
3402 if (req->rq_repmsg != NULL) {
3403 rep = req_capsule_server_get(info->mti_pill,
3405 rep->lock_policy_res2 =
3406 ptlrpc_status_hton(rep->lock_policy_res2);
3413 static int mdt_intent_policy(struct ldlm_namespace *ns,
3414 struct ldlm_lock **lockp, void *req_cookie,
3415 enum ldlm_mode mode, __u64 flags, void *data)
3417 struct tgt_session_info *tsi;
3418 struct mdt_thread_info *info;
3419 struct ptlrpc_request *req = req_cookie;
3420 struct ldlm_intent *it;
3421 struct req_capsule *pill;
3426 LASSERT(req != NULL);
3428 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3430 info = tsi2mdt_info(tsi);
3431 LASSERT(info != NULL);
3432 pill = info->mti_pill;
3433 LASSERT(pill->rc_req == req);
3435 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3436 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3437 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3439 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3443 /* Lock without inodebits makes no sense and will oops
3444 * later in ldlm. Let's check it now to see if we have
3445 * ibits corrupted somewhere in mdt_intent_opc().
3446 * The case for client miss to set ibits has been
3447 * processed by others. */
3448 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3449 lr_type == LDLM_IBITS,
3450 info->mti_dlm_req->lock_desc.\
3451 l_policy_data.l_inodebits.bits != 0));
3453 rc = err_serious(-EFAULT);
3455 /* No intent was provided */
3456 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3457 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3458 rc = req_capsule_server_pack(pill);
3460 rc = err_serious(rc);
3462 mdt_thread_info_fini(info);
3466 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3468 struct seq_server_site *ss = mdt_seq_site(mdt);
3470 if (ss->ss_node_id == 0)
3473 if (ss->ss_client_seq != NULL) {
3474 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3475 ss->ss_client_seq->lcs_exp = NULL;
3478 if (ss->ss_server_fld != NULL) {
3479 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3480 ss->ss_server_fld->lsf_control_exp = NULL;
3484 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3486 struct seq_server_site *ss = mdt_seq_site(mdt);
3491 if (ss->ss_server_seq != NULL)
3492 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3495 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3497 mdt_seq_fini_cli(mdt);
3498 mdt_deregister_seq_exp(mdt);
3500 return seq_site_fini(env, mdt_seq_site(mdt));
3504 * It will retrieve its FLDB entries from MDT0, and it only happens
3505 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3506 * and it needs to refresh FLDB from the MDT0.
3508 static int mdt_register_lwp_callback(void *data)
3511 struct mdt_device *mdt = data;
3512 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3516 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3518 rc = lu_env_init(&env, LCT_MD_THREAD);
3520 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3524 /* Allocate new sequence now to avoid creating local transaction
3525 * in the normal transaction process */
3526 rc = seq_server_check_and_alloc_super(&env,
3527 mdt_seq_site(mdt)->ss_server_seq);
3532 rc = fld_update_from_controller(&env, fld);
3534 CERROR("%s: cannot update controller: rc = %d\n",
3535 mdt_obd_name(mdt), rc);
3544 static int mdt_register_seq_exp(struct mdt_device *mdt)
3546 struct seq_server_site *ss = mdt_seq_site(mdt);
3547 char *lwp_name = NULL;
3550 if (ss->ss_node_id == 0)
3553 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3554 if (lwp_name == NULL)
3555 GOTO(out_free, rc = -ENOMEM);
3557 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3561 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3566 rc = lustre_register_lwp_item(lwp_name,
3567 &ss->ss_server_fld->lsf_control_exp,
3568 mdt_register_lwp_callback, mdt);
3570 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3571 ss->ss_client_seq->lcs_exp = NULL;
3575 if (lwp_name != NULL)
3576 OBD_FREE(lwp_name, MAX_OBD_NAME);
3582 * Init client sequence manager which is used by local MDS to talk to sequence
3583 * controller on remote node.
3585 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3587 struct seq_server_site *ss = mdt_seq_site(mdt);
3592 /* check if this is adding the first MDC and controller is not yet
3594 OBD_ALLOC_PTR(ss->ss_client_seq);
3595 if (ss->ss_client_seq == NULL)
3598 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3599 if (prefix == NULL) {
3600 OBD_FREE_PTR(ss->ss_client_seq);
3601 ss->ss_client_seq = NULL;
3605 /* Note: seq_client_fini will be called in seq_site_fini */
3606 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3607 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3608 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3610 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3612 OBD_FREE_PTR(ss->ss_client_seq);
3613 ss->ss_client_seq = NULL;
3617 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3622 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3624 struct seq_server_site *ss;
3628 ss = mdt_seq_site(mdt);
3629 /* init sequence controller server(MDT0) */
3630 if (ss->ss_node_id == 0) {
3631 OBD_ALLOC_PTR(ss->ss_control_seq);
3632 if (ss->ss_control_seq == NULL)
3635 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3636 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3639 GOTO(out_seq_fini, rc);
3642 /* Init normal sequence server */
3643 OBD_ALLOC_PTR(ss->ss_server_seq);
3644 if (ss->ss_server_seq == NULL)
3645 GOTO(out_seq_fini, rc = -ENOMEM);
3647 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3648 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3650 GOTO(out_seq_fini, rc);
3652 /* init seq client for seq server to talk to seq controller(MDT0) */
3653 rc = mdt_seq_init_cli(env, mdt);
3655 GOTO(out_seq_fini, rc);
3657 if (ss->ss_node_id != 0)
3658 /* register controller export through lwp */
3659 rc = mdt_register_seq_exp(mdt);
3664 mdt_seq_fini(env, mdt);
3672 static int mdt_fld_fini(const struct lu_env *env,
3673 struct mdt_device *m)
3675 struct seq_server_site *ss = mdt_seq_site(m);
3678 if (ss && ss->ss_server_fld) {
3679 fld_server_fini(env, ss->ss_server_fld);
3680 OBD_FREE_PTR(ss->ss_server_fld);
3681 ss->ss_server_fld = NULL;
3687 static int mdt_fld_init(const struct lu_env *env,
3689 struct mdt_device *m)
3691 struct seq_server_site *ss;
3695 ss = mdt_seq_site(m);
3697 OBD_ALLOC_PTR(ss->ss_server_fld);
3698 if (ss->ss_server_fld == NULL)
3699 RETURN(rc = -ENOMEM);
3701 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3704 OBD_FREE_PTR(ss->ss_server_fld);
3705 ss->ss_server_fld = NULL;
3712 static void mdt_stack_pre_fini(const struct lu_env *env,
3713 struct mdt_device *m, struct lu_device *top)
3715 struct lustre_cfg_bufs *bufs;
3716 struct lustre_cfg *lcfg;
3717 struct mdt_thread_info *info;
3722 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3723 LASSERT(info != NULL);
3725 bufs = &info->mti_u.bufs;
3727 LASSERT(m->mdt_child_exp);
3728 LASSERT(m->mdt_child_exp->exp_obd);
3730 /* process cleanup, pass mdt obd name to get obd umount flags */
3731 /* XXX: this is needed because all layers are referenced by
3732 * objects (some of them are pinned by osd, for example *
3733 * the proper solution should be a model where object used
3734 * by osd only doesn't have mdt/mdd slices -bzzz */
3735 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3736 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3737 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3741 top->ld_ops->ldo_process_config(env, top, lcfg);
3742 lustre_cfg_free(lcfg);
3746 static void mdt_stack_fini(const struct lu_env *env,
3747 struct mdt_device *m, struct lu_device *top)
3749 struct obd_device *obd = mdt2obd_dev(m);
3750 struct lustre_cfg_bufs *bufs;
3751 struct lustre_cfg *lcfg;
3752 struct mdt_thread_info *info;
3756 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3757 LASSERT(info != NULL);
3759 lu_dev_del_linkage(top->ld_site, top);
3761 lu_site_purge(env, top->ld_site, -1);
3763 bufs = &info->mti_u.bufs;
3764 /* process cleanup, pass mdt obd name to get obd umount flags */
3765 /* another purpose is to let all layers to release their objects */
3766 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3771 lustre_cfg_bufs_set_string(bufs, 1, flags);
3772 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3777 top->ld_ops->ldo_process_config(env, top, lcfg);
3778 lustre_cfg_free(lcfg);
3780 lu_site_purge(env, top->ld_site, -1);
3782 m->mdt_child = NULL;
3783 m->mdt_bottom = NULL;
3785 obd_disconnect(m->mdt_child_exp);
3786 m->mdt_child_exp = NULL;
3788 obd_disconnect(m->mdt_bottom_exp);
3789 m->mdt_child_exp = NULL;
3792 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3793 const char *next, struct obd_export **exp)
3795 struct obd_connect_data *data = NULL;
3796 struct obd_device *obd;
3800 OBD_ALLOC_PTR(data);
3802 GOTO(out, rc = -ENOMEM);
3804 obd = class_name2obd(next);
3806 CERROR("%s: can't locate next device: %s\n",
3807 mdt_obd_name(m), next);
3808 GOTO(out, rc = -ENOTCONN);
3811 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3812 data->ocd_version = LUSTRE_VERSION_CODE;
3814 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3816 CERROR("%s: cannot connect to next dev %s (%d)\n",
3817 mdt_obd_name(m), next, rc);
3827 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3828 struct lustre_cfg *cfg)
3830 char *dev = lustre_cfg_string(cfg, 0);
3831 int rc, name_size, uuid_size;
3832 char *name, *uuid, *p;
3833 struct lustre_cfg_bufs *bufs;
3834 struct lustre_cfg *lcfg;
3835 struct obd_device *obd;
3836 struct lustre_profile *lprof;
3837 struct lu_site *site;
3840 /* in 1.8 we had the only device in the stack - MDS.
3841 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3842 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3843 * to generate names and setup MDT, MDD. MDT will be using
3844 * generated name to connect to MDD. for MDD the next device
3845 * will be LOD with name taken from so called "profile" which
3846 * is generated by mount_option line
3848 * 1.8 MGS generates config. commands like this:
3849 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3850 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3851 * 2.0 MGS generates config. commands like this:
3852 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3853 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3854 * 3:lustre-MDT0000-mdtlov 4:f
3856 * we generate MDD name from MDT one, just replacing T with D
3858 * after all the preparations, the logical equivalent will be
3859 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3860 * 3:lustre-MDT0000-mdtlov 4:f
3861 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3862 * 3:lustre-MDD0000 4:f
3864 * notice we build the stack from down to top: MDD first, then MDT */
3866 name_size = MAX_OBD_NAME;
3867 uuid_size = MAX_OBD_NAME;
3869 OBD_ALLOC(name, name_size);
3870 OBD_ALLOC(uuid, uuid_size);
3871 if (name == NULL || uuid == NULL)
3872 GOTO(cleanup_mem, rc = -ENOMEM);
3874 OBD_ALLOC_PTR(bufs);
3876 GOTO(cleanup_mem, rc = -ENOMEM);
3879 p = strstr(name, "-MDT");
3881 GOTO(free_bufs, rc = -ENOMEM);
3884 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3886 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3887 if (lprof == NULL || lprof->lp_dt == NULL) {
3888 CERROR("can't find the profile: %s\n",
3889 lustre_cfg_string(cfg, 0));
3890 GOTO(free_bufs, rc = -EINVAL);
3893 lustre_cfg_bufs_reset(bufs, name);
3894 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3895 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3896 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3898 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3900 GOTO(put_profile, rc = -ENOMEM);
3902 rc = class_attach(lcfg);
3904 GOTO(lcfg_cleanup, rc);
3906 obd = class_name2obd(name);
3908 CERROR("Can not find obd %s (%s in config)\n",
3909 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3910 GOTO(lcfg_cleanup, rc = -EINVAL);
3913 lustre_cfg_free(lcfg);
3915 lustre_cfg_bufs_reset(bufs, name);
3916 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3917 lustre_cfg_bufs_set_string(bufs, 2, dev);
3918 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3920 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3922 GOTO(class_detach, rc = -ENOMEM);
3924 rc = class_setup(obd, lcfg);
3926 GOTO(class_detach, rc);
3928 /* connect to MDD we just setup */
3929 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3931 GOTO(class_detach, rc);
3933 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3935 LASSERT(mdt_lu_site(mdt) == NULL);
3936 mdt->mdt_lu_dev.ld_site = site;
3937 site->ls_top_dev = &mdt->mdt_lu_dev;
3938 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3940 /* now connect to bottom OSD */
3941 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3942 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3944 GOTO(class_detach, rc);
3946 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3948 rc = lu_env_refill((struct lu_env *)env);
3950 CERROR("Failure to refill session: '%d'\n", rc);
3952 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3957 class_detach(obd, lcfg);
3959 lustre_cfg_free(lcfg);
3961 class_put_profile(lprof);
3966 OBD_FREE(name, name_size);
3968 OBD_FREE(uuid, uuid_size);
3972 /* setup quota master target on MDT0 */
3973 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3974 struct lustre_cfg *cfg)
3976 struct obd_device *obd;
3977 char *dev = lustre_cfg_string(cfg, 0);
3978 char *qmtname, *uuid, *p;
3979 struct lustre_cfg_bufs *bufs;
3980 struct lustre_cfg *lcfg;
3981 struct lustre_profile *lprof;
3982 struct obd_connect_data *data;
3986 LASSERT(mdt->mdt_qmt_exp == NULL);
3987 LASSERT(mdt->mdt_qmt_dev == NULL);
3989 /* quota master is on MDT0 only for now */
3990 if (mdt->mdt_seq_site.ss_node_id != 0)
3993 /* MGS generates config commands which look as follows:
3994 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3995 * 3:lustre-MDT0000-mdtlov 4:f
3997 * We generate the QMT name from the MDT one, just replacing MD with QM
3998 * after all the preparations, the logical equivalent will be:
3999 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4000 * 3:lustre-MDT0000-osd 4:f */
4001 OBD_ALLOC(qmtname, MAX_OBD_NAME);
4002 OBD_ALLOC(uuid, UUID_MAX);
4003 OBD_ALLOC_PTR(bufs);
4004 OBD_ALLOC_PTR(data);
4005 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
4006 GOTO(cleanup_mem, rc = -ENOMEM);
4008 strcpy(qmtname, dev);
4009 p = strstr(qmtname, "-MDT");
4011 GOTO(cleanup_mem, rc = -ENOMEM);
4012 /* replace MD with QM */
4016 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4018 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4019 if (lprof == NULL || lprof->lp_dt == NULL) {
4020 CERROR("can't find profile for %s\n",
4021 lustre_cfg_string(cfg, 0));
4022 GOTO(cleanup_mem, rc = -EINVAL);
4025 lustre_cfg_bufs_reset(bufs, qmtname);
4026 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4027 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4028 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4030 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4032 GOTO(put_profile, rc = -ENOMEM);
4034 rc = class_attach(lcfg);
4036 GOTO(lcfg_cleanup, rc);
4038 obd = class_name2obd(qmtname);
4040 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4041 lustre_cfg_string(cfg, 0));
4042 GOTO(lcfg_cleanup, rc = -EINVAL);
4045 lustre_cfg_free(lcfg);
4047 lustre_cfg_bufs_reset(bufs, qmtname);
4048 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4049 lustre_cfg_bufs_set_string(bufs, 2, dev);
4051 /* for quota, the next device should be the OSD device */
4052 lustre_cfg_bufs_set_string(bufs, 3,
4053 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4055 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4057 GOTO(class_detach, rc = -ENOMEM);
4059 rc = class_setup(obd, lcfg);
4061 GOTO(class_detach, rc);
4063 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4065 /* configure local quota objects */
4066 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4070 GOTO(class_cleanup, rc);
4072 /* connect to quota master target */
4073 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4074 data->ocd_version = LUSTRE_VERSION_CODE;
4075 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4078 CERROR("cannot connect to quota master device %s (%d)\n",
4080 GOTO(class_cleanup, rc);
4086 class_manual_cleanup(obd);
4087 mdt->mdt_qmt_dev = NULL;
4091 class_detach(obd, lcfg);
4093 lustre_cfg_free(lcfg);
4095 class_put_profile(lprof);
4100 OBD_FREE(qmtname, MAX_OBD_NAME);
4102 OBD_FREE(uuid, UUID_MAX);
4108 /* Shutdown quota master target associated with mdt */
4109 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4113 if (mdt->mdt_qmt_exp == NULL)
4115 LASSERT(mdt->mdt_qmt_dev != NULL);
4117 /* the qmt automatically shuts down when the mdt disconnects */
4118 obd_disconnect(mdt->mdt_qmt_exp);
4119 mdt->mdt_qmt_exp = NULL;
4120 mdt->mdt_qmt_dev = NULL;
4124 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4125 * for now. This will be removed along with converting rest of MDT code
4126 * to use tgt_session_info */
4127 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4129 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4132 rc = mdt_getxattr(info);
4134 mdt_thread_info_fini(info);
4138 static struct tgt_handler mdt_tgt_handlers[] = {
4139 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4140 0, MDS_CONNECT, mdt_tgt_connect,
4141 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4142 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4143 0, MDS_DISCONNECT, tgt_disconnect,
4144 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4145 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4146 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4147 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4148 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4149 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4150 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4151 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4153 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4154 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4155 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4156 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4157 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4158 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4159 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4160 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4162 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4163 mdt_hsm_ct_register),
4164 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4165 mdt_hsm_ct_unregister),
4166 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4168 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4170 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4171 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4173 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4178 static struct tgt_handler mdt_sec_ctx_ops[] = {
4179 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4180 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4181 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4184 static struct tgt_handler mdt_quota_ops[] = {
4185 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4188 static struct tgt_opc_slice mdt_common_slice[] = {
4190 .tos_opc_start = MDS_FIRST_OPC,
4191 .tos_opc_end = MDS_LAST_OPC,
4192 .tos_hs = mdt_tgt_handlers
4195 .tos_opc_start = OBD_FIRST_OPC,
4196 .tos_opc_end = OBD_LAST_OPC,
4197 .tos_hs = tgt_obd_handlers
4200 .tos_opc_start = LDLM_FIRST_OPC,
4201 .tos_opc_end = LDLM_LAST_OPC,
4202 .tos_hs = tgt_dlm_handlers
4205 .tos_opc_start = SEC_FIRST_OPC,
4206 .tos_opc_end = SEC_LAST_OPC,
4207 .tos_hs = mdt_sec_ctx_ops
4210 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4211 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4212 .tos_hs = tgt_out_handlers
4215 .tos_opc_start = FLD_FIRST_OPC,
4216 .tos_opc_end = FLD_LAST_OPC,
4217 .tos_hs = fld_handlers
4220 .tos_opc_start = SEQ_FIRST_OPC,
4221 .tos_opc_end = SEQ_LAST_OPC,
4222 .tos_hs = seq_handlers
4225 .tos_opc_start = QUOTA_DQACQ,
4226 .tos_opc_end = QUOTA_LAST_OPC,
4227 .tos_hs = mdt_quota_ops
4230 .tos_opc_start = LLOG_FIRST_OPC,
4231 .tos_opc_end = LLOG_LAST_OPC,
4232 .tos_hs = tgt_llog_handlers
4235 .tos_opc_start = LFSCK_FIRST_OPC,
4236 .tos_opc_end = LFSCK_LAST_OPC,
4237 .tos_hs = tgt_lfsck_handlers
4245 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4247 struct md_device *next = m->mdt_child;
4248 struct lu_device *d = &m->mdt_lu_dev;
4249 struct obd_device *obd = mdt2obd_dev(m);
4250 struct lfsck_stop stop;
4253 stop.ls_status = LS_PAUSED;
4255 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4257 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4258 target_recovery_fini(obd);
4259 ping_evictor_stop();
4261 if (m->mdt_opts.mo_coordinator)
4262 mdt_hsm_cdt_stop(m);
4264 mdt_hsm_cdt_fini(m);
4266 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4267 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4269 if (m->mdt_namespace != NULL)
4270 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
4271 d->ld_obd->obd_force);
4273 obd_exports_barrier(obd);
4274 obd_zombie_barrier();
4278 tgt_fini(env, &m->mdt_lut);
4279 mdt_fs_cleanup(env, m);
4280 upcall_cache_cleanup(m->mdt_identity_cache);
4281 m->mdt_identity_cache = NULL;
4283 if (m->mdt_namespace != NULL) {
4284 ldlm_namespace_free_post(m->mdt_namespace);
4285 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4288 mdt_quota_fini(env, m);
4290 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4292 mdt_seq_fini(env, m);
4293 mdt_fld_fini(env, m);
4298 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4300 LASSERT(atomic_read(&d->ld_ref) == 0);
4302 server_put_mount(mdt_obd_name(m), true);
4307 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4309 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4310 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4312 struct mdt_thread_info *info;
4313 struct obd_device *obd;
4314 const char *dev = lustre_cfg_string(cfg, 0);
4315 const char *num = lustre_cfg_string(cfg, 2);
4316 struct lustre_mount_info *lmi = NULL;
4317 struct lustre_sb_info *lsi;
4319 struct seq_server_site *ss_site;
4320 const char *identity_upcall = "NONE";
4321 struct md_device *next;
4327 lu_device_init(&m->mdt_lu_dev, ldt);
4329 * Environment (env) might be missing mdt_thread_key values at that
4330 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4333 * Usually device allocation path doesn't use module key values, but
4334 * mdt has to do a lot of work here, so allocate key value.
4336 rc = lu_env_refill((struct lu_env *)env);
4340 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4341 LASSERT(info != NULL);
4343 obd = class_name2obd(dev);
4344 LASSERT(obd != NULL);
4346 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4347 m->mdt_opts.mo_evict_tgt_nids = 1;
4348 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4350 /* default is coordinator off, it is started through conf_param
4352 m->mdt_opts.mo_coordinator = 0;
4354 lmi = server_get_mount(dev);
4356 CERROR("Cannot get mount info for %s!\n", dev);
4359 lsi = s2lsi(lmi->lmi_sb);
4360 /* CMD is supported only in IAM mode */
4362 node_id = simple_strtol(num, NULL, 10);
4363 obd->u.obt.obt_magic = OBT_MAGIC;
4364 if (lsi->lsi_lmd != NULL &&
4365 lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
4366 m->mdt_skip_lfsck = 1;
4369 m->mdt_squash.rsi_uid = 0;
4370 m->mdt_squash.rsi_gid = 0;
4371 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4372 init_rwsem(&m->mdt_squash.rsi_sem);
4373 spin_lock_init(&m->mdt_osfs_lock);
4374 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4375 m->mdt_enable_remote_dir = 0;
4376 m->mdt_enable_remote_dir_gid = 0;
4378 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4379 m->mdt_lu_dev.ld_obd = obd;
4380 /* Set this lu_device to obd for error handling purposes. */
4381 obd->obd_lu_dev = &m->mdt_lu_dev;
4383 /* init the stack */
4384 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4386 CERROR("%s: Can't init device stack, rc %d\n",
4387 mdt_obd_name(m), rc);
4392 ss_site = mdt_seq_site(m);
4393 s->ld_seq_site = ss_site;
4396 /* set server index */
4397 ss_site->ss_node_id = node_id;
4399 /* failover is the default
4400 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4401 * assumed whose ss_node_id == 0 XXX
4403 obd->obd_replayable = 1;
4404 /* No connection accepted until configurations will finish */
4405 obd->obd_no_conn = 1;
4407 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4408 char *str = lustre_cfg_string(cfg, 4);
4409 if (strchr(str, 'n')) {
4410 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4411 obd->obd_replayable = 0;
4415 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4417 GOTO(err_fini_stack, rc);
4419 rc = mdt_seq_init(env, m);
4421 GOTO(err_fini_fld, rc);
4423 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4424 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4425 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4426 LDLM_NAMESPACE_SERVER,
4427 LDLM_NAMESPACE_GREEDY,
4429 if (m->mdt_namespace == NULL)
4430 GOTO(err_fini_seq, rc = -ENOMEM);
4432 m->mdt_namespace->ns_lvbp = m;
4433 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4435 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4436 /* set obd_namespace for compatibility with old code */
4437 obd->obd_namespace = m->mdt_namespace;
4439 rc = mdt_hsm_cdt_init(m);
4441 CERROR("%s: error initializing coordinator, rc %d\n",
4442 mdt_obd_name(m), rc);
4443 GOTO(err_free_ns, rc);
4446 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4447 OBD_FAIL_MDS_ALL_REQUEST_NET,
4448 OBD_FAIL_MDS_ALL_REPLY_NET);
4450 GOTO(err_free_hsm, rc);
4452 rc = mdt_fs_setup(env, m, obd, lsi);
4456 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4458 next = m->mdt_child;
4459 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4462 GOTO(err_fs_cleanup, rc);
4464 if (mntopts & MNTOPT_USERXATTR)
4465 m->mdt_opts.mo_user_xattr = 1;
4467 m->mdt_opts.mo_user_xattr = 0;
4469 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4471 GOTO(err_fs_cleanup, rc);
4473 if (mntopts & MNTOPT_ACL)
4474 m->mdt_opts.mo_acl = 1;
4476 m->mdt_opts.mo_acl = 0;
4478 /* XXX: to support suppgid for ACL, we enable identity_upcall
4479 * by default, otherwise, maybe got unexpected -EACCESS. */
4480 if (m->mdt_opts.mo_acl)
4481 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4483 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4485 &mdt_identity_upcall_cache_ops);
4486 if (IS_ERR(m->mdt_identity_cache)) {
4487 rc = PTR_ERR(m->mdt_identity_cache);
4488 m->mdt_identity_cache = NULL;
4489 GOTO(err_fs_cleanup, rc);
4492 rc = mdt_procfs_init(m, dev);
4494 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4495 GOTO(err_recovery, rc);
4498 rc = mdt_quota_init(env, m, cfg);
4500 GOTO(err_procfs, rc);
4502 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4503 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4504 "mdt_ldlm_client", m->mdt_ldlm_client);
4506 ping_evictor_start();
4508 /* recovery will be started upon mdt_prepare()
4509 * when the whole stack is complete and ready
4510 * to serve the requests */
4512 /* Reduce the initial timeout on an MDS because it doesn't need such
4513 * a long timeout as an OST does. Adaptive timeouts will adjust this
4514 * value appropriately. */
4515 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4516 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4522 target_recovery_fini(obd);
4523 upcall_cache_cleanup(m->mdt_identity_cache);
4524 m->mdt_identity_cache = NULL;
4526 mdt_fs_cleanup(env, m);
4528 tgt_fini(env, &m->mdt_lut);
4530 mdt_hsm_cdt_fini(m);
4532 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4533 obd->obd_namespace = m->mdt_namespace = NULL;
4535 mdt_seq_fini(env, m);
4537 mdt_fld_fini(env, m);
4539 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4542 server_put_mount(dev, true);
4546 /* For interoperability, the left element is old parameter, the right one
4547 * is the new version of the parameter, if some parameter is deprecated,
4548 * the new version should be set as NULL. */
4549 static struct cfg_interop_param mdt_interop_param[] = {
4550 { "mdt.group_upcall", NULL },
4551 { "mdt.quota_type", NULL },
4552 { "mdd.quota_type", NULL },
4553 { "mdt.rootsquash", "mdt.root_squash" },
4554 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4558 /* used by MGS to process specific configurations */
4559 static int mdt_process_config(const struct lu_env *env,
4560 struct lu_device *d, struct lustre_cfg *cfg)
4562 struct mdt_device *m = mdt_dev(d);
4563 struct md_device *md_next = m->mdt_child;
4564 struct lu_device *next = md2lu_dev(md_next);
4568 switch (cfg->lcfg_command) {
4570 struct obd_device *obd = d->ld_obd;
4572 /* For interoperability */
4573 struct cfg_interop_param *ptr = NULL;
4574 struct lustre_cfg *old_cfg = NULL;
4577 param = lustre_cfg_string(cfg, 1);
4578 if (param == NULL) {
4579 CERROR("param is empty\n");
4584 ptr = class_find_old_param(param, mdt_interop_param);
4586 if (ptr->new_param == NULL) {
4588 CWARN("For interoperability, skip this %s."
4589 " It is obsolete.\n", ptr->old_param);
4593 CWARN("Found old param %s, changed it to %s.\n",
4594 ptr->old_param, ptr->new_param);
4597 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4604 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4606 if (rc > 0 || rc == -ENOSYS) {
4607 /* is it an HSM var ? */
4608 rc = class_process_proc_param(PARAM_HSM,
4609 hsm_cdt_get_proc_vars(),
4611 if (rc > 0 || rc == -ENOSYS)
4612 /* we don't understand; pass it on */
4613 rc = next->ld_ops->ldo_process_config(env, next,
4617 if (old_cfg != NULL)
4618 lustre_cfg_free(cfg);
4623 /* others are passed further */
4624 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4630 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4631 const struct lu_object_header *hdr,
4632 struct lu_device *d)
4634 struct mdt_object *mo;
4638 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4640 struct lu_object *o;
4641 struct lu_object_header *h;
4644 h = &mo->mot_header;
4645 lu_object_header_init(h);
4646 lu_object_init(o, h, d);
4647 lu_object_add_top(h, o);
4648 o->lo_ops = &mdt_obj_ops;
4649 spin_lock_init(&mo->mot_write_lock);
4650 mutex_init(&mo->mot_lov_mutex);
4651 init_rwsem(&mo->mot_open_sem);
4657 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4658 const struct lu_object_conf *unused)
4660 struct mdt_device *d = mdt_dev(o->lo_dev);
4661 struct lu_device *under;
4662 struct lu_object *below;
4666 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4667 PFID(lu_object_fid(o)));
4669 under = &d->mdt_child->md_lu_dev;
4670 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4671 if (below != NULL) {
4672 lu_object_add(o, below);
4679 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4681 struct mdt_object *mo = mdt_obj(o);
4682 struct lu_object_header *h;
4686 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4687 PFID(lu_object_fid(o)));
4689 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4690 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4693 lu_object_header_fini(h);
4694 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4699 static int mdt_object_print(const struct lu_env *env, void *cookie,
4700 lu_printer_t p, const struct lu_object *o)
4702 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4704 return (*p)(env, cookie,
4705 LUSTRE_MDT_NAME"-object@%p(flags=%d, writecount=%d)",
4706 mdto, mdto->mot_flags, mdto->mot_write_count);
4709 static int mdt_prepare(const struct lu_env *env,
4710 struct lu_device *pdev,
4711 struct lu_device *cdev)
4713 struct mdt_device *mdt = mdt_dev(cdev);
4714 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4715 struct obd_device *obd = cdev->ld_obd;
4722 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4726 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4730 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4734 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4735 /* The LFSCK instance is registered just now, so it must be there when
4736 * register the namespace to such instance. */
4737 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4739 if (mdt->mdt_seq_site.ss_node_id == 0) {
4740 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4741 &mdt->mdt_md_root_fid);
4746 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4748 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4749 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4750 LASSERT(obd->obd_no_conn);
4751 spin_lock(&obd->obd_dev_lock);
4752 obd->obd_no_conn = 0;
4753 spin_unlock(&obd->obd_dev_lock);
4755 if (obd->obd_recovering == 0)
4756 mdt_postrecov(env, mdt);
4761 const struct lu_device_operations mdt_lu_ops = {
4762 .ldo_object_alloc = mdt_object_alloc,
4763 .ldo_process_config = mdt_process_config,
4764 .ldo_prepare = mdt_prepare,
4767 static const struct lu_object_operations mdt_obj_ops = {
4768 .loo_object_init = mdt_object_init,
4769 .loo_object_free = mdt_object_free,
4770 .loo_object_print = mdt_object_print
4773 static int mdt_obd_set_info_async(const struct lu_env *env,
4774 struct obd_export *exp,
4775 __u32 keylen, void *key,
4776 __u32 vallen, void *val,
4777 struct ptlrpc_request_set *set)
4783 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4784 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4792 * Match client and server connection feature flags.
4794 * Compute the compatibility flags for a connection request based on
4795 * features mutually supported by client and server.
4797 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4798 * must not be updated here, otherwise a partially initialized value may
4799 * be exposed. After the connection request is successfully processed,
4800 * the top-level MDT connect request handler atomically updates the export
4801 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4802 * reply. \see mdt_connect().
4804 * \param exp the obd_export associated with this client/target pair
4805 * \param mdt the target device for the connection
4806 * \param data stores data for this connect request
4809 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4810 * \retval -EBADE client and server feature requirements are incompatible
4812 static int mdt_connect_internal(struct obd_export *exp,
4813 struct mdt_device *mdt,
4814 struct obd_connect_data *data)
4816 LASSERT(data != NULL);
4818 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4819 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4821 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
4822 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
4823 CWARN("%s: client %s does not support ibits lock, either "
4824 "very old or an invalid client: flags "LPX64"\n",
4825 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
4826 data->ocd_connect_flags);
4830 if (!mdt->mdt_opts.mo_acl)
4831 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4833 if (!mdt->mdt_opts.mo_user_xattr)
4834 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4836 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4837 data->ocd_brw_size = min(data->ocd_brw_size,
4838 (__u32)MD_MAX_BRW_SIZE);
4839 if (data->ocd_brw_size == 0) {
4840 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4841 " ocd_version: %x ocd_grant: %d "
4842 "ocd_index: %u ocd_brw_size is "
4843 "unexpectedly zero, network data "
4844 "corruption? Refusing connection of this"
4847 exp->exp_client_uuid.uuid,
4848 exp, data->ocd_connect_flags, data->ocd_version,
4849 data->ocd_grant, data->ocd_index);
4854 /* NB: Disregard the rule against updating
4855 * exp_connect_data.ocd_connect_flags in this case, since
4856 * tgt_client_new() needs to know if this is a lightweight
4857 * connection, and it is safe to expose this flag before
4858 * connection processing completes. */
4859 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4860 spin_lock(&exp->exp_lock);
4861 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4862 spin_unlock(&exp->exp_lock);
4865 data->ocd_version = LUSTRE_VERSION_CODE;
4867 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4868 CWARN("%s: MDS requires FID support, but client not\n",
4873 if (OCD_HAS_FLAG(data, PINGLESS)) {
4874 if (ptlrpc_pinger_suppress_pings()) {
4875 spin_lock(&exp->exp_obd->obd_dev_lock);
4876 list_del_init(&exp->exp_obd_chain_timed);
4877 spin_unlock(&exp->exp_obd->obd_dev_lock);
4879 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4883 data->ocd_max_easize = mdt->mdt_max_ea_size;
4885 /* NB: Disregard the rule against updating
4886 * exp_connect_data.ocd_connect_flags in this case, since
4887 * tgt_client_new() needs to know if this is client supports
4888 * multiple modify RPCs, and it is safe to expose this flag before
4889 * connection processing completes. */
4890 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
4891 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
4892 spin_lock(&exp->exp_lock);
4893 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
4894 spin_unlock(&exp->exp_lock);
4900 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
4901 struct mdt_thread_info *info,
4902 struct mdt_file_data *mfd)
4904 struct lu_context ses;
4908 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
4913 lu_context_enter(&ses);
4915 mdt_ucred(info)->uc_valid = UCRED_OLD;
4916 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
4918 lu_context_exit(&ses);
4919 lu_context_fini(&ses);
4925 static int mdt_export_cleanup(struct obd_export *exp)
4927 struct list_head closing_list;
4928 struct mdt_export_data *med = &exp->exp_mdt_data;
4929 struct obd_device *obd = exp->exp_obd;
4930 struct mdt_device *mdt;
4931 struct mdt_thread_info *info;
4933 struct mdt_file_data *mfd, *n;
4937 INIT_LIST_HEAD(&closing_list);
4938 spin_lock(&med->med_open_lock);
4939 while (!list_empty(&med->med_open_head)) {
4940 struct list_head *tmp = med->med_open_head.next;
4941 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
4943 /* Remove mfd handle so it can't be found again.
4944 * We are consuming the mfd_list reference here. */
4945 class_handle_unhash(&mfd->mfd_handle);
4946 list_move_tail(&mfd->mfd_list, &closing_list);
4948 spin_unlock(&med->med_open_lock);
4949 mdt = mdt_dev(obd->obd_lu_dev);
4950 LASSERT(mdt != NULL);
4952 rc = lu_env_init(&env, LCT_MD_THREAD);
4956 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
4957 LASSERT(info != NULL);
4958 memset(info, 0, sizeof *info);
4959 info->mti_env = &env;
4960 info->mti_mdt = mdt;
4961 info->mti_exp = exp;
4963 if (!list_empty(&closing_list)) {
4964 struct md_attr *ma = &info->mti_attr;
4966 /* Close any open files (which may also cause orphan
4968 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
4969 list_del_init(&mfd->mfd_list);
4970 ma->ma_need = ma->ma_valid = 0;
4972 /* This file is being closed due to an eviction, it
4973 * could have been modified and now dirty regarding to
4974 * HSM archive, check this!
4975 * The logic here is to mark a file dirty if there's a
4976 * chance it was dirtied before the client was evicted,
4977 * so that we don't have to wait for a release attempt
4978 * before finding out the file was actually dirty and
4979 * fail the release. Aggressively marking it dirty here
4980 * will cause the policy engine to attempt to
4981 * re-archive it; when rearchiving, we can compare the
4982 * current version to the HSM data_version and make the
4983 * archive request into a noop if it's not actually
4986 if (mfd->mfd_mode & FMODE_WRITE)
4987 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
4989 /* Don't unlink orphan on failover umount, LU-184 */
4990 if (exp->exp_flags & OBD_OPT_FAILOVER) {
4991 ma->ma_valid = MA_FLAGS;
4992 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
4994 mdt_mfd_close(info, mfd);
4997 info->mti_mdt = NULL;
4998 /* cleanup client slot early */
4999 /* Do not erase record for recoverable client. */
5000 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5001 tgt_client_del(&env, exp);
5007 static int mdt_obd_disconnect(struct obd_export *exp)
5013 class_export_get(exp);
5015 rc = server_disconnect_export(exp);
5017 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
5019 rc = mdt_export_cleanup(exp);
5020 nodemap_del_member(exp);
5021 class_export_put(exp);
5025 /* mds_connect copy */
5026 static int mdt_obd_connect(const struct lu_env *env,
5027 struct obd_export **exp, struct obd_device *obd,
5028 struct obd_uuid *cluuid,
5029 struct obd_connect_data *data,
5032 struct obd_export *lexp;
5033 struct lustre_handle conn = { 0 };
5034 struct mdt_device *mdt;
5036 lnet_nid_t *client_nid = localdata;
5039 LASSERT(env != NULL);
5040 if (!exp || !obd || !cluuid)
5043 mdt = mdt_dev(obd->obd_lu_dev);
5046 * first, check whether the stack is ready to handle requests
5047 * XXX: probably not very appropriate method is used now
5048 * at some point we should find a better one
5050 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
5051 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5052 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5053 rc = obd_get_info(env, mdt->mdt_child_exp,
5054 sizeof(KEY_OSP_CONNECTED),
5055 KEY_OSP_CONNECTED, NULL, NULL);
5058 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5061 rc = class_connect(&conn, obd, cluuid);
5065 lexp = class_conn2export(&conn);
5066 LASSERT(lexp != NULL);
5068 rc = nodemap_add_member(*client_nid, lexp);
5069 if (rc != 0 && rc != -EEXIST)
5072 rc = mdt_connect_internal(lexp, mdt, data);
5074 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5077 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5078 rc = tgt_client_new(env, lexp);
5080 mdt_export_stats_init(obd, lexp, localdata);
5084 class_disconnect(lexp);
5085 nodemap_del_member(lexp);
5094 static int mdt_obd_reconnect(const struct lu_env *env,
5095 struct obd_export *exp, struct obd_device *obd,
5096 struct obd_uuid *cluuid,
5097 struct obd_connect_data *data,
5100 lnet_nid_t *client_nid = localdata;
5104 if (exp == NULL || obd == NULL || cluuid == NULL)
5107 rc = nodemap_add_member(*client_nid, exp);
5108 if (rc != 0 && rc != -EEXIST)
5111 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5113 mdt_export_stats_init(obd, exp, localdata);
5115 nodemap_del_member(exp);
5120 /* FIXME: Can we avoid using these two interfaces? */
5121 static int mdt_init_export(struct obd_export *exp)
5123 struct mdt_export_data *med = &exp->exp_mdt_data;
5127 INIT_LIST_HEAD(&med->med_open_head);
5128 spin_lock_init(&med->med_open_lock);
5129 mutex_init(&med->med_idmap_mutex);
5130 med->med_idmap = NULL;
5131 spin_lock(&exp->exp_lock);
5132 exp->exp_connecting = 1;
5133 spin_unlock(&exp->exp_lock);
5135 /* self-export doesn't need client data and ldlm initialization */
5136 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5137 &exp->exp_client_uuid)))
5140 rc = tgt_client_alloc(exp);
5144 rc = ldlm_init_export(exp);
5151 tgt_client_free(exp);
5153 CERROR("%s: Failed to initialize export: rc = %d\n",
5154 exp->exp_obd->obd_name, rc);
5158 static int mdt_destroy_export(struct obd_export *exp)
5162 if (exp_connect_rmtclient(exp))
5163 mdt_cleanup_idmap(&exp->exp_mdt_data);
5165 target_destroy_export(exp);
5166 /* destroy can be called from failed obd_setup, so
5167 * checking uuid is safer than obd_self_export */
5168 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5169 &exp->exp_client_uuid)))
5172 ldlm_destroy_export(exp);
5173 tgt_client_free(exp);
5175 LASSERT(list_empty(&exp->exp_outstanding_replies));
5176 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5181 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5182 struct linkea_data *ldata)
5186 LASSERT(ldata->ld_buf->lb_buf != NULL);
5188 if (!mdt_object_exists(mdt_obj))
5191 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5192 ldata->ld_buf, XATTR_NAME_LINK);
5193 if (rc == -ERANGE) {
5194 /* Buf was too small, figure out what we need. */
5195 lu_buf_free(ldata->ld_buf);
5196 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5197 ldata->ld_buf, XATTR_NAME_LINK);
5200 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5201 if (ldata->ld_buf->lb_buf == NULL)
5203 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5204 ldata->ld_buf, XATTR_NAME_LINK);
5209 return linkea_init(ldata);
5213 * Given an MDT object, try to look up the full path to the object.
5214 * Part of the MDT layer implementation of lfs fid2path.
5216 * \param[in] info Per-thread common data shared by MDT level handlers.
5217 * \param[in] obj Object to do path lookup of
5218 * \param[in,out] fp User-provided struct to store path information
5220 * \retval 0 Lookup successful, path information stored in fp
5221 * \retval -EAGAIN Lookup failed, usually because object is being moved
5222 * \retval negative errno if there was a problem
5224 static int mdt_path_current(struct mdt_thread_info *info,
5225 struct mdt_object *obj,
5226 struct getinfo_fid2path *fp)
5228 struct mdt_device *mdt = info->mti_mdt;
5229 struct mdt_object *mdt_obj;
5230 struct link_ea_header *leh;
5231 struct link_ea_entry *lee;
5232 struct lu_name *tmpname = &info->mti_name;
5233 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5234 struct lu_buf *buf = &info->mti_big_buf;
5237 struct linkea_data ldata = { NULL };
5242 /* temp buffer for path element, the buffer will be finally freed
5243 * in mdt_thread_info_fini */
5244 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5245 if (buf->lb_buf == NULL)
5249 ptr = fp->gf_path + fp->gf_pathlen - 1;
5252 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5254 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5255 * so checking root_fid can only happen on MDT0. */
5256 while (!lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid)) {
5257 struct lu_buf lmv_buf;
5259 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5260 if (IS_ERR(mdt_obj))
5261 GOTO(out, rc = PTR_ERR(mdt_obj));
5263 if (!mdt_object_exists(mdt_obj)) {
5264 mdt_object_put(info->mti_env, mdt_obj);
5265 GOTO(out, rc = -ENOENT);
5268 if (mdt_object_remote(mdt_obj)) {
5269 mdt_object_put(info->mti_env, mdt_obj);
5270 GOTO(remote_out, rc = -EREMOTE);
5273 rc = mdt_links_read(info, mdt_obj, &ldata);
5275 mdt_object_put(info->mti_env, mdt_obj);
5280 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5281 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5282 /* If set, use link #linkno for path lookup, otherwise use
5283 link #0. Only do this for the final path element. */
5284 if (first && fp->gf_linkno < leh->leh_reccount) {
5286 for (count = 0; count < fp->gf_linkno; count++) {
5287 lee = (struct link_ea_entry *)
5288 ((char *)lee + reclen);
5289 linkea_entry_unpack(lee, &reclen, tmpname,
5292 if (fp->gf_linkno < leh->leh_reccount - 1)
5293 /* indicate to user there are more links */
5297 lmv_buf.lb_buf = info->mti_xattr_buf;
5298 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5299 /* Check if it is slave stripes */
5300 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5301 &lmv_buf, XATTR_NAME_LMV);
5302 mdt_object_put(info->mti_env, mdt_obj);
5304 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5306 /* For slave stripes, get its master */
5307 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5308 fp->gf_fid = *tmpfid;
5311 } else if (rc < 0 && rc != -ENODATA) {
5317 /* Pack the name in the end of the buffer */
5318 ptr -= tmpname->ln_namelen;
5319 if (ptr - 1 <= fp->gf_path)
5320 GOTO(out, rc = -EOVERFLOW);
5321 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5324 /* keep the last resolved fid to the client, so the
5325 * client will build the left path on another MDT for
5327 fp->gf_fid = *tmpfid;
5333 ptr++; /* skip leading / */
5334 memmove(fp->gf_path, ptr, fp->gf_path + fp->gf_pathlen - ptr);
5341 * Given an MDT object, use mdt_path_current to get the path.
5342 * Essentially a wrapper to retry mdt_path_current a set number of times
5343 * if -EAGAIN is returned (usually because an object is being moved).
5345 * Part of the MDT layer implementation of lfs fid2path.
5347 * \param[in] info Per-thread common data shared by mdt level handlers.
5348 * \param[in] obj Object to do path lookup of
5349 * \param[in,out] fp User-provided struct for arguments and to store path
5352 * \retval 0 Lookup successful, path information stored in fp
5353 * \retval negative errno if there was a problem
5355 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5356 struct getinfo_fid2path *fp)
5358 struct mdt_device *mdt = info->mti_mdt;
5363 if (fp->gf_pathlen < 3)
5366 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5367 fp->gf_path[0] = '\0';
5371 /* Retry multiple times in case file is being moved */
5372 while (tries-- && rc == -EAGAIN)
5373 rc = mdt_path_current(info, obj, fp);
5379 * Get the full path of the provided FID, as of changelog record recno.
5381 * This checks sanity and looks up object for user provided FID
5382 * before calling the actual path lookup code.
5384 * Part of the MDT layer implementation of lfs fid2path.
5386 * \param[in] info Per-thread common data shared by mdt level handlers.
5387 * \param[in,out] fp User-provided struct for arguments and to store path
5390 * \retval 0 Lookup successful, path information and recno stored in fp
5391 * \retval -ENOENT, object does not exist
5392 * \retval negative errno if there was a problem
5394 static int mdt_fid2path(struct mdt_thread_info *info,
5395 struct getinfo_fid2path *fp)
5397 struct mdt_device *mdt = info->mti_mdt;
5398 struct mdt_object *obj;
5402 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5403 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5405 if (!fid_is_sane(&fp->gf_fid))
5408 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5409 CWARN("%s: "DFID" is invalid, sequence should be "
5410 ">= "LPX64"\n", mdt_obd_name(mdt),
5411 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5415 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5418 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5419 PFID(&fp->gf_fid), rc);
5423 if (mdt_object_remote(obj))
5425 else if (!mdt_object_exists(obj))
5431 mdt_object_put(info->mti_env, obj);
5432 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5433 PFID(&fp->gf_fid), rc);
5437 rc = mdt_path(info, obj, fp);
5439 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5440 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5442 mdt_object_put(info->mti_env, obj);
5447 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5448 void *val, int vallen)
5450 struct getinfo_fid2path *fpout, *fpin;
5453 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5456 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5457 lustre_swab_fid2path(fpin);
5459 memcpy(fpout, fpin, sizeof(*fpin));
5460 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5463 rc = mdt_fid2path(info, fpout);
5467 int mdt_get_info(struct tgt_session_info *tsi)
5477 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5479 CDEBUG(D_IOCTL, "No GETINFO key\n");
5480 RETURN(err_serious(-EFAULT));
5482 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5485 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5486 if (vallen == NULL) {
5487 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5488 tgt_name(tsi->tsi_tgt));
5489 RETURN(err_serious(-EFAULT));
5492 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5494 rc = req_capsule_server_pack(tsi->tsi_pill);
5496 RETURN(err_serious(rc));
5498 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5499 if (valout == NULL) {
5500 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5501 tgt_name(tsi->tsi_tgt));
5502 RETURN(err_serious(-EFAULT));
5505 if (KEY_IS(KEY_FID2PATH)) {
5506 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5508 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5509 mdt_thread_info_fini(info);
5516 /* Pass the ioc down */
5517 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5518 unsigned int cmd, int len, void *data)
5520 struct lu_context ioctl_session;
5521 struct md_device *next = mdt->mdt_child;
5525 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5528 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5529 lu_context_enter(&ioctl_session);
5530 env->le_ses = &ioctl_session;
5532 LASSERT(next->md_ops->mdo_iocontrol);
5533 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5535 lu_context_exit(&ioctl_session);
5536 lu_context_fini(&ioctl_session);
5540 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5542 struct obd_ioctl_data *data = karg;
5545 struct mdt_object *obj;
5546 struct mdt_lock_handle *lh;
5550 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5551 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5554 fid = (struct lu_fid *)data->ioc_inlbuf1;
5556 if (!fid_is_sane(fid))
5559 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5561 lh = &mti->mti_lh[MDT_LH_PARENT];
5562 mdt_lock_reg_init(lh, LCK_CR);
5564 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5566 RETURN(PTR_ERR(obj));
5568 if (mdt_object_remote(obj)) {
5571 * before calling version get the correct MDS should be
5572 * fid, this is error to find remote object here
5574 CERROR("nonlocal object "DFID"\n", PFID(fid));
5575 } else if (!mdt_object_exists(obj)) {
5576 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5579 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5580 *(__u64 *)data->ioc_inlbuf2 = version;
5583 mdt_object_unlock_put(mti, obj, lh, 1);
5587 /* ioctls on obd dev */
5588 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5589 void *karg, void *uarg)
5592 struct obd_device *obd = exp->exp_obd;
5593 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5594 struct dt_device *dt = mdt->mdt_bottom;
5598 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5599 rc = lu_env_init(&env, LCT_MD_THREAD);
5605 rc = mdt_device_sync(&env, mdt);
5607 case OBD_IOC_SET_READONLY:
5608 rc = dt->dd_ops->dt_ro(&env, dt);
5610 case OBD_IOC_ABORT_RECOVERY:
5611 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5612 obd->obd_abort_recovery = 1;
5613 target_stop_recovery_thread(obd);
5616 case OBD_IOC_CHANGELOG_REG:
5617 case OBD_IOC_CHANGELOG_DEREG:
5618 case OBD_IOC_CHANGELOG_CLEAR:
5619 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5621 case OBD_IOC_START_LFSCK: {
5622 struct md_device *next = mdt->mdt_child;
5623 struct obd_ioctl_data *data = karg;
5624 struct lfsck_start_param lsp;
5626 if (unlikely(data == NULL)) {
5631 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5632 lsp.lsp_index_valid = 0;
5633 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5636 case OBD_IOC_STOP_LFSCK: {
5637 struct md_device *next = mdt->mdt_child;
5638 struct obd_ioctl_data *data = karg;
5639 struct lfsck_stop stop;
5641 stop.ls_status = LS_STOPPED;
5642 /* Old lfsck utils may pass NULL @stop. */
5643 if (data->ioc_inlbuf1 == NULL)
5647 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5649 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5652 case OBD_IOC_GET_OBJ_VERSION: {
5653 struct mdt_thread_info *mti;
5654 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5655 memset(mti, 0, sizeof *mti);
5656 mti->mti_env = &env;
5660 rc = mdt_ioc_version_get(mti, karg);
5663 case OBD_IOC_CATLOGLIST: {
5664 struct mdt_thread_info *mti;
5666 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5667 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5668 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5669 &mti->mti_tmp_fid1);
5674 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5675 mdt_obd_name(mdt), cmd, rc);
5682 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5684 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5688 if (!mdt->mdt_skip_lfsck) {
5689 struct lfsck_start_param lsp;
5691 lsp.lsp_start = NULL;
5692 lsp.lsp_index_valid = 0;
5693 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
5694 OBD_IOC_START_LFSCK,
5696 if (rc != 0 && rc != -EALREADY)
5697 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
5698 mdt_obd_name(mdt), rc);
5701 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5705 static int mdt_obd_postrecov(struct obd_device *obd)
5710 rc = lu_env_init(&env, LCT_MD_THREAD);
5713 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5718 static struct obd_ops mdt_obd_device_ops = {
5719 .o_owner = THIS_MODULE,
5720 .o_set_info_async = mdt_obd_set_info_async,
5721 .o_connect = mdt_obd_connect,
5722 .o_reconnect = mdt_obd_reconnect,
5723 .o_disconnect = mdt_obd_disconnect,
5724 .o_init_export = mdt_init_export,
5725 .o_destroy_export = mdt_destroy_export,
5726 .o_iocontrol = mdt_iocontrol,
5727 .o_postrecov = mdt_obd_postrecov,
5730 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5731 struct lu_device *d)
5733 struct mdt_device *m = mdt_dev(d);
5740 static struct lu_device *mdt_device_free(const struct lu_env *env,
5741 struct lu_device *d)
5743 struct mdt_device *m = mdt_dev(d);
5746 lu_device_fini(&m->mdt_lu_dev);
5752 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5753 struct lu_device_type *t,
5754 struct lustre_cfg *cfg)
5756 struct lu_device *l;
5757 struct mdt_device *m;
5764 rc = mdt_init0(env, m, t, cfg);
5766 mdt_device_free(env, l);
5771 l = ERR_PTR(-ENOMEM);
5775 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5776 LU_KEY_INIT(mdt, struct mdt_thread_info);
5778 static void mdt_key_fini(const struct lu_context *ctx,
5779 struct lu_context_key *key, void* data)
5781 struct mdt_thread_info *info = data;
5783 if (info->mti_big_lmm) {
5784 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5785 info->mti_big_lmm = NULL;
5786 info->mti_big_lmmsize = 0;
5791 /* context key: mdt_thread_key */
5792 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5794 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5796 return lu_ucred(info->mti_env);
5799 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5801 return lu_ucred_check(info->mti_env);
5805 * Enable/disable COS (Commit On Sharing).
5807 * Set/Clear the COS flag in mdt options.
5809 * \param mdt mdt device
5810 * \param val 0 disables COS, other values enable COS
5812 void mdt_enable_cos(struct mdt_device *mdt, int val)
5817 mdt->mdt_opts.mo_cos = !!val;
5818 rc = lu_env_init(&env, LCT_LOCAL);
5819 if (unlikely(rc != 0)) {
5820 CWARN("%s: lu_env initialization failed, cannot "
5821 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
5824 mdt_device_sync(&env, mdt);
5829 * Check COS (Commit On Sharing) status.
5831 * Return COS flag status.
5833 * \param mdt mdt device
5835 int mdt_cos_is_enabled(struct mdt_device *mdt)
5837 return mdt->mdt_opts.mo_cos != 0;
5840 static struct lu_device_type_operations mdt_device_type_ops = {
5841 .ldto_device_alloc = mdt_device_alloc,
5842 .ldto_device_free = mdt_device_free,
5843 .ldto_device_fini = mdt_device_fini
5846 static struct lu_device_type mdt_device_type = {
5847 .ldt_tags = LU_DEVICE_MD,
5848 .ldt_name = LUSTRE_MDT_NAME,
5849 .ldt_ops = &mdt_device_type_ops,
5850 .ldt_ctx_tags = LCT_MD_THREAD
5853 static int __init mdt_init(void)
5857 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5858 FID_NOBRACE_LEN + 1);
5859 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5861 rc = lu_kmem_init(mdt_caches);
5865 rc = mds_mod_init();
5869 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
5870 LUSTRE_MDT_NAME, &mdt_device_type);
5875 lu_kmem_fini(mdt_caches);
5882 static void __exit mdt_exit(void)
5884 class_unregister_type(LUSTRE_MDT_NAME);
5886 lu_kmem_fini(mdt_caches);
5889 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
5890 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5891 MODULE_VERSION(LUSTRE_VERSION_STRING);
5892 MODULE_LICENSE("GPL");
5894 module_init(mdt_init);
5895 module_exit(mdt_exit);