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 * struct OBD_{ALLOC,FREE}*()
55 #include <obd_support.h>
56 #include <lustre_ioctl.h>
57 /* struct ptlrpc_request */
58 #include <lustre_net.h>
59 /* struct obd_export */
60 #include <lustre_export.h>
61 /* struct obd_device */
64 #include <dt_object.h>
65 #include <lustre_mds.h>
66 #include <lustre_log.h>
67 #include "mdt_internal.h"
68 #include <lustre_acl.h>
69 #include <lustre_param.h>
70 #include <lustre_quota.h>
71 #include <lustre_lfsck.h>
72 #include <lustre_nodemap.h>
74 static unsigned int max_mod_rpcs_per_client = 8;
75 CFS_MODULE_PARM(max_mod_rpcs_per_client, "i", uint, 0644,
76 "maximum number of modify RPCs in flight allowed per client");
78 mdl_mode_t mdt_mdl_lock_modes[] = {
79 [LCK_MINMODE] = MDL_MINMODE,
86 [LCK_GROUP] = MDL_GROUP
89 enum ldlm_mode mdt_dlm_lock_modes[] = {
90 [MDL_MINMODE] = LCK_MINMODE,
97 [MDL_GROUP] = LCK_GROUP
100 static struct mdt_device *mdt_dev(struct lu_device *d);
101 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
103 static const struct lu_object_operations mdt_obj_ops;
105 /* Slab for MDT object allocation */
106 static struct kmem_cache *mdt_object_kmem;
108 /* For HSM restore handles */
109 struct kmem_cache *mdt_hsm_cdt_kmem;
111 /* For HSM request handles */
112 struct kmem_cache *mdt_hsm_car_kmem;
114 static struct lu_kmem_descr mdt_caches[] = {
116 .ckd_cache = &mdt_object_kmem,
117 .ckd_name = "mdt_obj",
118 .ckd_size = sizeof(struct mdt_object)
121 .ckd_cache = &mdt_hsm_cdt_kmem,
122 .ckd_name = "mdt_cdt_restore_handle",
123 .ckd_size = sizeof(struct cdt_restore_handle)
126 .ckd_cache = &mdt_hsm_car_kmem,
127 .ckd_name = "mdt_cdt_agent_req",
128 .ckd_size = sizeof(struct cdt_agent_req)
135 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
139 return rep->lock_policy_res1 & op_flag;
142 void mdt_clear_disposition(struct mdt_thread_info *info,
143 struct ldlm_reply *rep, __u64 op_flag)
146 info->mti_opdata &= ~op_flag;
147 tgt_opdata_clear(info->mti_env, op_flag);
150 rep->lock_policy_res1 &= ~op_flag;
153 void mdt_set_disposition(struct mdt_thread_info *info,
154 struct ldlm_reply *rep, __u64 op_flag)
157 info->mti_opdata |= op_flag;
158 tgt_opdata_set(info->mti_env, op_flag);
161 rep->lock_policy_res1 |= op_flag;
164 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
166 lh->mlh_pdo_hash = 0;
167 lh->mlh_reg_mode = lm;
168 lh->mlh_rreg_mode = lm;
169 lh->mlh_type = MDT_REG_LOCK;
172 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
173 const struct lu_name *lname)
175 lh->mlh_reg_mode = lock_mode;
176 lh->mlh_rreg_mode = lock_mode;
177 lh->mlh_type = MDT_PDO_LOCK;
179 if (lu_name_is_valid(lname)) {
180 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
182 /* XXX Workaround for LU-2856
184 * Zero is a valid return value of full_name_hash, but
185 * several users of mlh_pdo_hash assume a non-zero
186 * hash value. We therefore map zero onto an
187 * arbitrary, but consistent value (1) to avoid
188 * problems further down the road. */
189 if (unlikely(lh->mlh_pdo_hash == 0))
190 lh->mlh_pdo_hash = 1;
192 lh->mlh_pdo_hash = 0;
196 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
197 struct mdt_lock_handle *lh)
203 * Any dir access needs couple of locks:
205 * 1) on part of dir we gonna take lookup/modify;
207 * 2) on whole dir to protect it from concurrent splitting and/or to
208 * flush client's cache for readdir().
210 * so, for a given mode and object this routine decides what lock mode
211 * to use for lock #2:
213 * 1) if caller's gonna lookup in dir then we need to protect dir from
214 * being splitted only - LCK_CR
216 * 2) if caller's gonna modify dir then we need to protect dir from
217 * being splitted and to flush cache - LCK_CW
219 * 3) if caller's gonna modify dir and that dir seems ready for
220 * splitting then we need to protect it from any type of access
221 * (lookup/modify/split) - LCK_EX --bzzz
224 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
225 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
228 * Ask underlaying level its opinion about preferable PDO lock mode
229 * having access type passed as regular lock mode:
231 * - MDL_MINMODE means that lower layer does not want to specify lock
234 * - MDL_NL means that no PDO lock should be taken. This is used in some
235 * cases. Say, for non-splittable directories no need to use PDO locks
238 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
239 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
241 if (mode != MDL_MINMODE) {
242 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
245 * Lower layer does not want to specify locking mode. We do it
246 * our selves. No special protection is needed, just flush
247 * client's cache on modification and allow concurrent
250 switch (lh->mlh_reg_mode) {
252 lh->mlh_pdo_mode = LCK_EX;
255 lh->mlh_pdo_mode = LCK_CR;
258 lh->mlh_pdo_mode = LCK_CW;
261 CERROR("Not expected lock type (0x%x)\n",
262 (int)lh->mlh_reg_mode);
267 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
271 static int mdt_getstatus(struct tgt_session_info *tsi)
273 struct mdt_thread_info *info = tsi2mdt_info(tsi);
274 struct mdt_device *mdt = info->mti_mdt;
275 struct mdt_body *repbody;
280 rc = mdt_check_ucred(info);
282 GOTO(out, rc = err_serious(rc));
284 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK))
285 GOTO(out, rc = err_serious(-ENOMEM));
287 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
288 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
289 repbody->mbo_valid |= OBD_MD_FLID;
293 mdt_thread_info_fini(info);
297 static int mdt_statfs(struct tgt_session_info *tsi)
299 struct ptlrpc_request *req = tgt_ses_req(tsi);
300 struct mdt_thread_info *info = tsi2mdt_info(tsi);
301 struct md_device *next = info->mti_mdt->mdt_child;
302 struct ptlrpc_service_part *svcpt;
303 struct obd_statfs *osfs;
308 svcpt = req->rq_rqbd->rqbd_svcpt;
310 /* This will trigger a watchdog timeout */
311 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
312 (MDT_SERVICE_WATCHDOG_FACTOR *
313 at_get(&svcpt->scp_at_estimate)) + 1);
315 rc = mdt_check_ucred(info);
317 GOTO(out, rc = err_serious(rc));
319 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
320 GOTO(out, rc = err_serious(-ENOMEM));
322 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
324 GOTO(out, rc = -EPROTO);
326 /** statfs information are cached in the mdt_device */
327 if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
328 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
329 /** statfs data is too old, get up-to-date one */
330 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
333 spin_lock(&info->mti_mdt->mdt_osfs_lock);
334 info->mti_mdt->mdt_osfs = *osfs;
335 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
336 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
338 /** use cached statfs data */
339 spin_lock(&info->mti_mdt->mdt_osfs_lock);
340 *osfs = info->mti_mdt->mdt_osfs;
341 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
345 mdt_counter_incr(req, LPROC_MDT_STATFS);
347 mdt_thread_info_fini(info);
351 #ifdef CONFIG_FS_POSIX_ACL
353 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
355 * \param info thread info object
356 * \param repbody reply to pack ACLs into
357 * \param o mdt object of file to examine
358 * \param nodemap nodemap of client to reply to
360 * \retval -errno error getting or parsing ACL from disk
362 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
363 struct mdt_object *o, struct lu_nodemap *nodemap)
365 const struct lu_env *env = info->mti_env;
366 struct md_object *next = mdt_object_child(o);
367 struct lu_buf *buf = &info->mti_buf;
370 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
371 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
373 if (buf->lb_len == 0)
376 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
378 if (rc == -ENODATA) {
379 repbody->mbo_aclsize = 0;
380 repbody->mbo_valid |= OBD_MD_FLACL;
382 } else if (rc == -EOPNOTSUPP) {
385 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
386 mdt_obd_name(info->mti_mdt),
387 PFID(mdt_object_fid(o)), rc);
390 rc = nodemap_map_acl(nodemap, buf->lb_buf,
391 rc, NODEMAP_FS_TO_CLIENT);
392 /* if all ACLs mapped out, rc is still >= 0 */
394 CERROR("%s: nodemap_map_acl unable to parse "DFID
395 " ACL: rc = %d\n", mdt_obd_name(info->mti_mdt),
396 PFID(mdt_object_fid(o)), rc);
398 repbody->mbo_aclsize = rc;
399 repbody->mbo_valid |= OBD_MD_FLACL;
407 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
408 const struct lu_attr *attr, const struct lu_fid *fid)
410 struct md_attr *ma = &info->mti_attr;
411 struct obd_export *exp = info->mti_exp;
412 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
414 LASSERT(ma->ma_valid & MA_INODE);
416 if (attr->la_valid & LA_ATIME) {
417 b->mbo_atime = attr->la_atime;
418 b->mbo_valid |= OBD_MD_FLATIME;
420 if (attr->la_valid & LA_MTIME) {
421 b->mbo_mtime = attr->la_mtime;
422 b->mbo_valid |= OBD_MD_FLMTIME;
424 if (attr->la_valid & LA_CTIME) {
425 b->mbo_ctime = attr->la_ctime;
426 b->mbo_valid |= OBD_MD_FLCTIME;
428 if (attr->la_valid & LA_FLAGS) {
429 b->mbo_flags = attr->la_flags;
430 b->mbo_valid |= OBD_MD_FLFLAGS;
432 if (attr->la_valid & LA_NLINK) {
433 b->mbo_nlink = attr->la_nlink;
434 b->mbo_valid |= OBD_MD_FLNLINK;
436 if (attr->la_valid & LA_UID) {
437 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
438 NODEMAP_FS_TO_CLIENT,
440 b->mbo_valid |= OBD_MD_FLUID;
442 if (attr->la_valid & LA_GID) {
443 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
444 NODEMAP_FS_TO_CLIENT,
446 b->mbo_valid |= OBD_MD_FLGID;
448 b->mbo_mode = attr->la_mode;
449 if (attr->la_valid & LA_MODE)
450 b->mbo_valid |= OBD_MD_FLMODE;
451 if (attr->la_valid & LA_TYPE)
452 b->mbo_valid |= OBD_MD_FLTYPE;
456 b->mbo_valid |= OBD_MD_FLID;
457 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid="LPX64"\n",
458 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
462 mdt_body_reverse_idmap(info, b);
464 if (!(attr->la_valid & LA_TYPE))
467 b->mbo_rdev = attr->la_rdev;
468 b->mbo_size = attr->la_size;
469 b->mbo_blocks = attr->la_blocks;
471 if (!S_ISREG(attr->la_mode)) {
472 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
473 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
474 /* means no objects are allocated on osts. */
475 LASSERT(!(ma->ma_valid & MA_LOV));
476 /* just ignore blocks occupied by extend attributes on MDS */
478 /* if no object is allocated on osts, the size on mds is valid.
480 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
481 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
482 ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
483 /* A released file stores its size on MDS. */
484 /* But return 1 block for released file, unless tools like tar
485 * will consider it fully sparse. (LU-3864)
487 if (unlikely(b->mbo_size == 0))
491 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
494 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
495 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
496 PFID(fid), (unsigned long long)b->mbo_size);
499 static inline int mdt_body_has_lov(const struct lu_attr *la,
500 const struct mdt_body *body)
502 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
503 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
506 void mdt_client_compatibility(struct mdt_thread_info *info)
508 struct mdt_body *body;
509 struct ptlrpc_request *req = mdt_info_req(info);
510 struct obd_export *exp = req->rq_export;
511 struct md_attr *ma = &info->mti_attr;
512 struct lu_attr *la = &ma->ma_attr;
515 if (exp_connect_layout(exp))
516 /* the client can deal with 16-bit lmm_stripe_count */
519 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
521 if (!mdt_body_has_lov(la, body))
524 /* now we have a reply with a lov for a client not compatible with the
525 * layout lock so we have to clean the layout generation number */
526 if (S_ISREG(la->la_mode))
527 ma->ma_lmm->lmm_layout_gen = 0;
531 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
532 struct mdt_object *o)
534 const struct lu_env *env = info->mti_env;
537 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
546 /* Is it a directory? Let's check for the LMV as well */
547 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
548 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
552 rc2 = mo_xattr_get(env, mdt_object_child(o),
554 XATTR_NAME_DEFAULT_LMV);
556 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
564 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
567 const struct lu_env *env = info->mti_env;
571 LASSERT(info->mti_big_lmm_used == 0);
572 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
576 /* big_lmm may need to be grown */
577 if (info->mti_big_lmmsize < rc) {
578 int size = size_roundup_power2(rc);
580 if (info->mti_big_lmmsize > 0) {
581 /* free old buffer */
582 LASSERT(info->mti_big_lmm);
583 OBD_FREE_LARGE(info->mti_big_lmm,
584 info->mti_big_lmmsize);
585 info->mti_big_lmm = NULL;
586 info->mti_big_lmmsize = 0;
589 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
590 if (info->mti_big_lmm == NULL)
592 info->mti_big_lmmsize = size;
594 LASSERT(info->mti_big_lmmsize >= rc);
596 info->mti_buf.lb_buf = info->mti_big_lmm;
597 info->mti_buf.lb_len = info->mti_big_lmmsize;
598 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
603 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
604 struct md_attr *ma, const char *name)
606 struct md_object *next = mdt_object_child(o);
607 struct lu_buf *buf = &info->mti_buf;
610 if (strcmp(name, XATTR_NAME_LOV) == 0) {
611 buf->lb_buf = ma->ma_lmm;
612 buf->lb_len = ma->ma_lmm_size;
613 LASSERT(!(ma->ma_valid & MA_LOV));
614 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
615 buf->lb_buf = ma->ma_lmv;
616 buf->lb_len = ma->ma_lmv_size;
617 LASSERT(!(ma->ma_valid & MA_LMV));
618 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
619 buf->lb_buf = ma->ma_lmv;
620 buf->lb_len = ma->ma_lmv_size;
621 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
626 rc = mo_xattr_get(info->mti_env, next, buf, name);
630 if (strcmp(name, XATTR_NAME_LOV) == 0) {
631 if (info->mti_big_lmm_used)
632 ma->ma_lmm = info->mti_big_lmm;
634 /* NOT return LOV EA with hole to old client. */
635 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
636 LOV_PATTERN_F_HOLE) &&
637 !(exp_connect_flags(info->mti_exp) &
638 OBD_CONNECT_LFSCK)) {
641 ma->ma_lmm_size = rc;
642 ma->ma_valid |= MA_LOV;
644 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
645 if (info->mti_big_lmm_used)
646 ma->ma_lmv = info->mti_big_lmm;
648 ma->ma_lmv_size = rc;
649 ma->ma_valid |= MA_LMV;
650 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
651 ma->ma_lmv_size = rc;
652 ma->ma_valid |= MA_LMV_DEF;
655 /* Update mdt_max_mdsize so all clients will be aware that */
656 if (info->mti_mdt->mdt_max_mdsize < rc)
657 info->mti_mdt->mdt_max_mdsize = rc;
660 } else if (rc == -ENODATA) {
663 } else if (rc == -ERANGE) {
664 /* Default LMV has fixed size, so it must be able to fit
665 * in the original buffer */
666 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
668 rc = mdt_big_xattr_get(info, o, name);
670 info->mti_big_lmm_used = 1;
678 static int mdt_attr_get_pfid(struct mdt_thread_info *info,
679 struct mdt_object *o, struct lu_fid *pfid)
681 struct lu_buf *buf = &info->mti_buf;
682 struct link_ea_header *leh;
683 struct link_ea_entry *lee;
687 buf->lb_buf = info->mti_big_lmm;
688 buf->lb_len = info->mti_big_lmmsize;
689 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
690 buf, XATTR_NAME_LINK);
691 /* ignore errors, MA_PFID won't be set and it is
692 * up to the caller to treat this as an error */
693 if (rc == -ERANGE || buf->lb_len == 0) {
694 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
695 buf->lb_buf = info->mti_big_lmm;
696 buf->lb_len = info->mti_big_lmmsize;
701 if (rc < sizeof(*leh)) {
702 CERROR("short LinkEA on "DFID": rc = %d\n",
703 PFID(mdt_object_fid(o)), rc);
707 leh = (struct link_ea_header *) buf->lb_buf;
708 lee = (struct link_ea_entry *)(leh + 1);
709 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
710 leh->leh_magic = LINK_EA_MAGIC;
711 leh->leh_reccount = __swab32(leh->leh_reccount);
712 leh->leh_len = __swab64(leh->leh_len);
714 if (leh->leh_magic != LINK_EA_MAGIC)
716 if (leh->leh_reccount == 0)
719 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
720 fid_be_to_cpu(pfid, pfid);
725 int mdt_attr_get_complex(struct mdt_thread_info *info,
726 struct mdt_object *o, struct md_attr *ma)
728 const struct lu_env *env = info->mti_env;
729 struct md_object *next = mdt_object_child(o);
730 struct lu_buf *buf = &info->mti_buf;
731 int need = ma->ma_need;
738 if (mdt_object_exists(o) == 0)
739 GOTO(out, rc = -ENOENT);
740 mode = lu_object_attr(&next->mo_lu);
742 if (need & MA_INODE) {
743 ma->ma_need = MA_INODE;
744 rc = mo_attr_get(env, next, ma);
747 ma->ma_valid |= MA_INODE;
750 if (need & MA_PFID) {
751 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
753 ma->ma_valid |= MA_PFID;
754 /* ignore this error, parent fid is not mandatory */
758 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
759 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
764 if (need & MA_LMV && S_ISDIR(mode)) {
765 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
770 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
771 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
776 if (need & MA_HSM && S_ISREG(mode)) {
777 buf->lb_buf = info->mti_xattr_buf;
778 buf->lb_len = sizeof(info->mti_xattr_buf);
779 CLASSERT(sizeof(struct hsm_attrs) <=
780 sizeof(info->mti_xattr_buf));
781 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
782 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
784 ma->ma_valid |= MA_HSM;
785 else if (rc2 < 0 && rc2 != -ENODATA)
789 #ifdef CONFIG_FS_POSIX_ACL
790 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
791 buf->lb_buf = ma->ma_acl;
792 buf->lb_len = ma->ma_acl_size;
793 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
795 ma->ma_acl_size = rc2;
796 ma->ma_valid |= MA_ACL_DEF;
797 } else if (rc2 == -ENODATA) {
806 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = "LPX64" ma_lmm=%p\n",
807 rc, ma->ma_valid, ma->ma_lmm);
811 static int mdt_getattr_internal(struct mdt_thread_info *info,
812 struct mdt_object *o, int ma_need)
814 struct md_object *next = mdt_object_child(o);
815 const struct mdt_body *reqbody = info->mti_body;
816 struct ptlrpc_request *req = mdt_info_req(info);
817 struct md_attr *ma = &info->mti_attr;
818 struct lu_attr *la = &ma->ma_attr;
819 struct req_capsule *pill = info->mti_pill;
820 const struct lu_env *env = info->mti_env;
821 struct mdt_body *repbody;
822 struct lu_buf *buffer = &info->mti_buf;
823 struct obd_export *exp = info->mti_exp;
824 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
829 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
830 RETURN(err_serious(-ENOMEM));
832 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
836 if (mdt_object_remote(o)) {
837 /* This object is located on remote node.*/
838 /* Return -ENOTSUPP for old client */
839 if (!mdt_is_dne_client(req->rq_export))
840 GOTO(out, rc = -ENOTSUPP);
842 repbody->mbo_fid1 = *mdt_object_fid(o);
843 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
847 if (reqbody->mbo_eadatasize > 0) {
848 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
849 if (buffer->lb_buf == NULL)
850 GOTO(out, rc = -EPROTO);
851 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
854 buffer->lb_buf = NULL;
856 ma_need &= ~(MA_LOV | MA_LMV);
857 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
858 mdt_obd_name(info->mti_mdt),
859 req->rq_export->exp_client_uuid.uuid);
862 /* If it is dir object and client require MEA, then we got MEA */
863 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
864 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
865 /* Assumption: MDT_MD size is enough for lmv size. */
866 ma->ma_lmv = buffer->lb_buf;
867 ma->ma_lmv_size = buffer->lb_len;
868 ma->ma_need = MA_INODE;
869 if (ma->ma_lmv_size > 0) {
870 if (reqbody->mbo_valid & OBD_MD_MEA)
871 ma->ma_need |= MA_LMV;
872 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
873 ma->ma_need |= MA_LMV_DEF;
876 ma->ma_lmm = buffer->lb_buf;
877 ma->ma_lmm_size = buffer->lb_len;
878 ma->ma_need = MA_INODE | MA_HSM;
879 if (ma->ma_lmm_size > 0)
880 ma->ma_need |= MA_LOV;
883 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
884 reqbody->mbo_valid & OBD_MD_FLDIREA &&
885 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
886 /* get default stripe info for this dir. */
887 ma->ma_need |= MA_LOV_DEF;
889 ma->ma_need |= ma_need;
891 rc = mdt_attr_get_complex(info, o, ma);
893 CERROR("%s: getattr error for "DFID": rc = %d\n",
894 mdt_obd_name(info->mti_mdt),
895 PFID(mdt_object_fid(o)), rc);
899 /* if file is released, check if a restore is running */
900 if (ma->ma_valid & MA_HSM) {
901 repbody->mbo_valid |= OBD_MD_TSTATE;
902 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
903 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
904 repbody->mbo_t_state = MS_RESTORE;
907 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
909 /* the Lustre protocol supposes to return default striping
910 * on the user-visible root if explicitly requested */
911 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
912 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
913 struct lu_fid rootfid;
914 struct mdt_object *root;
915 struct mdt_device *mdt = info->mti_mdt;
917 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
920 root = mdt_object_find(env, mdt, &rootfid);
922 RETURN(PTR_ERR(root));
923 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
924 mdt_object_put(info->mti_env, root);
926 CERROR("%s: getattr error for "DFID": rc = %d\n",
927 mdt_obd_name(info->mti_mdt),
928 PFID(mdt_object_fid(o)), rc);
933 if (likely(ma->ma_valid & MA_INODE))
934 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
938 if (mdt_body_has_lov(la, reqbody)) {
939 if (ma->ma_valid & MA_LOV) {
940 LASSERT(ma->ma_lmm_size);
941 repbody->mbo_eadatasize = ma->ma_lmm_size;
942 if (S_ISDIR(la->la_mode))
943 repbody->mbo_valid |= OBD_MD_FLDIREA;
945 repbody->mbo_valid |= OBD_MD_FLEASIZE;
946 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
948 if (ma->ma_valid & MA_LMV) {
949 /* Return -ENOTSUPP for old client */
950 if (!mdt_is_striped_client(req->rq_export))
953 LASSERT(S_ISDIR(la->la_mode));
954 mdt_dump_lmv(D_INFO, ma->ma_lmv);
955 repbody->mbo_eadatasize = ma->ma_lmv_size;
956 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
958 if (ma->ma_valid & MA_LMV_DEF) {
959 /* Return -ENOTSUPP for old client */
960 if (!mdt_is_striped_client(req->rq_export))
962 LASSERT(S_ISDIR(la->la_mode));
963 repbody->mbo_eadatasize = ma->ma_lmv_size;
964 repbody->mbo_valid |= (OBD_MD_FLDIREA |
967 } else if (S_ISLNK(la->la_mode) &&
968 reqbody->mbo_valid & OBD_MD_LINKNAME) {
969 buffer->lb_buf = ma->ma_lmm;
970 /* eadatasize from client includes NULL-terminator, so
971 * there is no need to read it */
972 buffer->lb_len = reqbody->mbo_eadatasize - 1;
973 rc = mo_readlink(env, next, buffer);
974 if (unlikely(rc <= 0)) {
975 CERROR("%s: readlink failed for "DFID": rc = %d\n",
976 mdt_obd_name(info->mti_mdt),
977 PFID(mdt_object_fid(o)), rc);
980 int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
982 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
984 repbody->mbo_valid |= OBD_MD_LINKNAME;
985 /* we need to report back size with NULL-terminator
986 * because client expects that */
987 repbody->mbo_eadatasize = rc + 1;
988 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
989 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
990 "on "DFID ", expected %d\n",
991 mdt_obd_name(info->mti_mdt),
992 rc, PFID(mdt_object_fid(o)),
993 reqbody->mbo_eadatasize - 1);
995 ((char *)ma->ma_lmm)[rc] = 0;
997 /* If the total CDEBUG() size is larger than a page, it
998 * will print a warning to the console, avoid this by
999 * printing just the last part of the symlink. */
1000 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1001 print_limit < rc ? "..." : "", print_limit,
1002 (char *)ma->ma_lmm + rc - print_limit, rc);
1007 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1008 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1009 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1010 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1011 repbody->mbo_max_mdsize);
1014 if (exp_connect_rmtclient(info->mti_exp) &&
1015 reqbody->mbo_valid & OBD_MD_FLRMTPERM) {
1016 void *buf = req_capsule_server_get(pill, &RMF_ACL);
1018 /* mdt_getattr_lock only */
1019 rc = mdt_pack_remote_perm(info, o, buf);
1021 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
1022 repbody->mbo_aclsize = 0;
1025 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
1026 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
1029 #ifdef CONFIG_FS_POSIX_ACL
1030 else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1031 (reqbody->mbo_valid & OBD_MD_FLACL))
1032 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1037 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1042 static int mdt_getattr(struct tgt_session_info *tsi)
1044 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1045 struct mdt_object *obj = info->mti_object;
1046 struct req_capsule *pill = info->mti_pill;
1047 struct mdt_body *reqbody;
1048 struct mdt_body *repbody;
1052 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1054 LASSERT(obj != NULL);
1055 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1057 /* Unlike intent case where we need to pre-fill out buffers early on
1058 * in intent policy for ldlm reasons, here we can have a much better
1059 * guess at EA size by just reading it from disk.
1060 * Exceptions are readdir and (missing) directory striping */
1062 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1063 /* No easy way to know how long is the symlink, but it cannot
1064 * be more than PATH_MAX, so we allocate +1 */
1067 /* A special case for fs ROOT: getattr there might fetch
1068 * default EA for entire fs, not just for this dir!
1070 } else if (lu_fid_eq(mdt_object_fid(obj),
1071 &info->mti_mdt->mdt_md_root_fid) &&
1072 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1073 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1075 /* Should the default strping be bigger, mdt_fix_reply
1076 * will reallocate */
1077 rc = DEF_REP_MD_SIZE;
1079 /* Read the actual EA size from disk */
1080 rc = mdt_attr_get_eabuf_size(info, obj);
1084 GOTO(out, rc = err_serious(rc));
1086 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1088 rc = req_capsule_server_pack(pill);
1089 if (unlikely(rc != 0))
1090 GOTO(out, rc = err_serious(rc));
1092 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1093 LASSERT(repbody != NULL);
1094 repbody->mbo_eadatasize = 0;
1095 repbody->mbo_aclsize = 0;
1097 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1098 rc = mdt_init_ucred(info, reqbody);
1100 rc = mdt_check_ucred(info);
1102 GOTO(out_shrink, rc);
1104 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1106 rc = mdt_getattr_internal(info, obj, 0);
1107 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1108 mdt_exit_ucred(info);
1111 mdt_client_compatibility(info);
1112 rc2 = mdt_fix_reply(info);
1116 mdt_thread_info_fini(info);
1121 * Exchange MOF_LOV_CREATED flags between two objects after a
1122 * layout swap. No assumption is made on whether o1 or o2 have
1123 * created objects or not.
1125 * \param[in,out] o1 First swap layout object
1126 * \param[in,out] o2 Second swap layout object
1128 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1132 mutex_lock(&o1->mot_lov_mutex);
1133 mutex_lock(&o2->mot_lov_mutex);
1135 o1_flags = o1->mot_flags;
1136 o1->mot_flags = (o1->mot_flags & ~MOF_LOV_CREATED) |
1137 (o2->mot_flags & MOF_LOV_CREATED);
1139 o2->mot_flags = (o2->mot_flags & ~MOF_LOV_CREATED) |
1140 (o1_flags & MOF_LOV_CREATED);
1142 mutex_unlock(&o2->mot_lov_mutex);
1143 mutex_unlock(&o1->mot_lov_mutex);
1146 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1148 struct mdt_thread_info *info;
1149 struct ptlrpc_request *req = tgt_ses_req(tsi);
1150 struct obd_export *exp = req->rq_export;
1151 struct mdt_object *o1, *o2, *o;
1152 struct mdt_lock_handle *lh1, *lh2;
1153 struct mdc_swap_layouts *msl;
1157 /* client does not support layout lock, so layout swaping
1159 * FIXME: there is a problem for old clients which don't support
1160 * layout lock yet. If those clients have already opened the file
1161 * they won't be notified at all so that old layout may still be
1162 * used to do IO. This can be fixed after file release is landed by
1163 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1164 if (!exp_connect_layout(exp))
1165 RETURN(-EOPNOTSUPP);
1167 info = tsi2mdt_info(tsi);
1169 if (info->mti_dlm_req != NULL)
1170 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1172 o1 = info->mti_object;
1173 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1174 &info->mti_body->mbo_fid2);
1176 GOTO(out, rc = PTR_ERR(o));
1178 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1179 GOTO(put, rc = -ENOENT);
1181 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1182 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1188 /* permission check. Make sure the calling process having permission
1189 * to write both files. */
1190 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1195 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1200 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1202 GOTO(put, rc = -EPROTO);
1204 lh1 = &info->mti_lh[MDT_LH_NEW];
1205 mdt_lock_reg_init(lh1, LCK_EX);
1206 lh2 = &info->mti_lh[MDT_LH_OLD];
1207 mdt_lock_reg_init(lh2, LCK_EX);
1209 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1210 MDS_INODELOCK_XATTR);
1214 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1215 MDS_INODELOCK_XATTR);
1219 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1220 mdt_object_child(o2), msl->msl_flags);
1224 mdt_swap_lov_flag(o1, o2);
1227 mdt_object_unlock(info, o2, lh2, rc);
1229 mdt_object_unlock(info, o1, lh1, rc);
1231 mdt_object_put(info->mti_env, o);
1233 mdt_thread_info_fini(info);
1237 static int mdt_raw_lookup(struct mdt_thread_info *info,
1238 struct mdt_object *parent,
1239 const struct lu_name *lname,
1240 struct ldlm_reply *ldlm_rep)
1242 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1246 LASSERT(!info->mti_cross_ref);
1248 /* Only got the fid of this obj by name */
1249 fid_zero(child_fid);
1250 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1251 lname, child_fid, &info->mti_spec);
1253 struct mdt_body *repbody;
1255 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1256 repbody->mbo_fid1 = *child_fid;
1257 repbody->mbo_valid = OBD_MD_FLID;
1258 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1259 } else if (rc == -ENOENT) {
1260 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1267 * UPDATE lock should be taken against parent, and be released before exit;
1268 * child_bits lock should be taken against child, and be returned back:
1269 * (1)normal request should release the child lock;
1270 * (2)intent request will grant the lock to client.
1272 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1273 struct mdt_lock_handle *lhc,
1275 struct ldlm_reply *ldlm_rep)
1277 struct ptlrpc_request *req = mdt_info_req(info);
1278 struct mdt_body *reqbody = NULL;
1279 struct mdt_object *parent = info->mti_object;
1280 struct mdt_object *child;
1281 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1282 struct lu_name *lname = NULL;
1283 struct mdt_lock_handle *lhp = NULL;
1284 struct ldlm_lock *lock;
1291 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1292 LASSERT(ergo(is_resent,
1293 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1298 if (info->mti_cross_ref) {
1299 /* Only getattr on the child. Parent is on another node. */
1300 mdt_set_disposition(info, ldlm_rep,
1301 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1303 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1305 PFID(mdt_object_fid(child)), ldlm_rep);
1307 rc = mdt_check_resent_lock(info, child, lhc);
1310 } else if (rc > 0) {
1311 mdt_lock_handle_init(lhc);
1312 mdt_lock_reg_init(lhc, LCK_PR);
1315 * Object's name is on another MDS, no lookup or layout
1316 * lock is needed here but update lock is.
1318 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1319 MDS_INODELOCK_LAYOUT);
1320 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1322 rc = mdt_object_lock(info, child, lhc, child_bits);
1327 /* Finally, we can get attr for child. */
1328 if (!mdt_object_exists(child)) {
1329 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1331 "remote object doesn't exist.\n");
1332 mdt_object_unlock(info, child, lhc, 1);
1336 rc = mdt_getattr_internal(info, child, 0);
1337 if (unlikely(rc != 0))
1338 mdt_object_unlock(info, child, lhc, 1);
1343 lname = &info->mti_name;
1344 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1346 if (lu_name_is_valid(lname)) {
1347 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1348 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1349 PNAME(lname), ldlm_rep);
1351 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1352 if (unlikely(reqbody == NULL))
1353 RETURN(err_serious(-EPROTO));
1355 *child_fid = reqbody->mbo_fid2;
1357 if (unlikely(!fid_is_sane(child_fid)))
1358 RETURN(err_serious(-EINVAL));
1360 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1362 PFID(mdt_object_fid(parent)),
1363 PFID(&reqbody->mbo_fid2), ldlm_rep);
1366 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1368 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1369 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1371 "Parent doesn't exist!\n");
1375 if (mdt_object_remote(parent)) {
1376 CERROR("%s: parent "DFID" is on remote target\n",
1377 mdt_obd_name(info->mti_mdt),
1378 PFID(mdt_object_fid(parent)));
1382 if (lu_name_is_valid(lname)) {
1383 /* Always allow to lookup ".." */
1384 if (unlikely(lname->ln_namelen == 2 &&
1385 lname->ln_name[0] == '.' &&
1386 lname->ln_name[1] == '.'))
1387 info->mti_spec.sp_permitted = 1;
1389 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1390 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1395 /* step 1: lock parent only if parent is a directory */
1396 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1397 lhp = &info->mti_lh[MDT_LH_PARENT];
1398 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1399 rc = mdt_object_lock(info, parent, lhp,
1400 MDS_INODELOCK_UPDATE);
1401 if (unlikely(rc != 0))
1405 /* step 2: lookup child's fid by name */
1406 fid_zero(child_fid);
1407 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1408 child_fid, &info->mti_spec);
1410 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1413 GOTO(out_parent, rc);
1416 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1419 *step 3: find the child object by fid & lock it.
1420 * regardless if it is local or remote.
1422 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1423 * set parent dir fid the same as child fid in getattr by fid case
1424 * we should not lu_object_find() the object again, could lead
1425 * to hung if there is a concurrent unlink destroyed the object.
1427 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1428 mdt_object_get(info->mti_env, parent);
1431 child = mdt_object_find(info->mti_env, info->mti_mdt,
1435 if (unlikely(IS_ERR(child)))
1436 GOTO(out_parent, rc = PTR_ERR(child));
1438 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1439 rc = mdt_check_resent_lock(info, child, lhc);
1441 GOTO(out_child, rc);
1442 } else if (rc > 0) {
1443 mdt_lock_handle_init(lhc);
1444 mdt_lock_reg_init(lhc, LCK_PR);
1447 if (!mdt_object_exists(child)) {
1448 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1450 "Object doesn't exist!\n");
1451 GOTO(out_child, rc = -ENOENT);
1454 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1455 mdt_object_exists(child) && !mdt_object_remote(child)) {
1456 struct md_attr *ma = &info->mti_attr;
1459 ma->ma_need = MA_INODE;
1460 rc = mdt_attr_get_complex(info, child, ma);
1461 if (unlikely(rc != 0))
1462 GOTO(out_child, rc);
1464 /* If the file has not been changed for some time, we
1465 * return not only a LOOKUP lock, but also an UPDATE
1466 * lock and this might save us RPC on later STAT. For
1467 * directories, it also let negative dentry cache start
1468 * working for this dir. */
1469 if (ma->ma_valid & MA_INODE &&
1470 ma->ma_attr.la_valid & LA_CTIME &&
1471 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1472 ma->ma_attr.la_ctime < cfs_time_current_sec())
1473 child_bits |= MDS_INODELOCK_UPDATE;
1476 /* layout lock must be granted in a best-effort way
1477 * for IT operations */
1478 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1479 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1480 exp_connect_layout(info->mti_exp) &&
1481 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1482 !mdt_object_remote(child) && ldlm_rep != NULL) {
1483 /* try to grant layout lock for regular file. */
1489 child_bits |= MDS_INODELOCK_LAYOUT;
1490 /* try layout lock, it may fail to be granted due to
1491 * contention at LOOKUP or UPDATE */
1492 if (!mdt_object_lock_try(info, child, lhc,
1494 child_bits &= ~MDS_INODELOCK_LAYOUT;
1495 LASSERT(child_bits != 0);
1496 rc = mdt_object_lock(info, child, lhc,
1502 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1503 * client will enqueue the lock to the remote MDT */
1504 if (mdt_object_remote(child))
1505 child_bits &= ~MDS_INODELOCK_UPDATE;
1506 rc = mdt_object_lock(info, child, lhc, child_bits);
1508 if (unlikely(rc != 0))
1509 GOTO(out_child, rc);
1512 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1514 /* finally, we can get attr for child. */
1515 rc = mdt_getattr_internal(info, child, ma_need);
1516 if (unlikely(rc != 0)) {
1517 mdt_object_unlock(info, child, lhc, 1);
1519 /* Debugging code. */
1520 LDLM_DEBUG(lock, "Returning lock to client");
1521 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1522 &lock->l_resource->lr_name),
1523 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1524 PLDLMRES(lock->l_resource),
1525 PFID(mdt_object_fid(child)));
1528 LDLM_LOCK_PUT(lock);
1532 mdt_object_put(info->mti_env, child);
1535 mdt_object_unlock(info, parent, lhp, 1);
1539 /* normal handler: should release the child lock */
1540 static int mdt_getattr_name(struct tgt_session_info *tsi)
1542 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1543 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1544 struct mdt_body *reqbody;
1545 struct mdt_body *repbody;
1549 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1550 LASSERT(reqbody != NULL);
1551 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1552 LASSERT(repbody != NULL);
1554 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1555 repbody->mbo_eadatasize = 0;
1556 repbody->mbo_aclsize = 0;
1558 rc = mdt_init_ucred(info, reqbody);
1560 GOTO(out_shrink, rc);
1562 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1563 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1564 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1565 lhc->mlh_reg_lh.cookie = 0;
1567 mdt_exit_ucred(info);
1570 mdt_client_compatibility(info);
1571 rc2 = mdt_fix_reply(info);
1574 mdt_thread_info_fini(info);
1578 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1579 void *karg, void *uarg);
1581 static int mdt_set_info(struct tgt_session_info *tsi)
1583 struct ptlrpc_request *req = tgt_ses_req(tsi);
1586 int keylen, vallen, rc = 0;
1590 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1592 DEBUG_REQ(D_HA, req, "no set_info key");
1593 RETURN(err_serious(-EFAULT));
1596 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1599 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1601 DEBUG_REQ(D_HA, req, "no set_info val");
1602 RETURN(err_serious(-EFAULT));
1605 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1608 /* Swab any part of val you need to here */
1609 if (KEY_IS(KEY_READ_ONLY)) {
1610 spin_lock(&req->rq_export->exp_lock);
1612 *exp_connect_flags_ptr(req->rq_export) |=
1615 *exp_connect_flags_ptr(req->rq_export) &=
1616 ~OBD_CONNECT_RDONLY;
1617 spin_unlock(&req->rq_export->exp_lock);
1618 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1619 struct changelog_setinfo *cs = val;
1621 if (vallen != sizeof(*cs)) {
1622 CERROR("%s: bad changelog_clear setinfo size %d\n",
1623 tgt_name(tsi->tsi_tgt), vallen);
1626 if (ptlrpc_req_need_swab(req)) {
1627 __swab64s(&cs->cs_recno);
1628 __swab32s(&cs->cs_id);
1631 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1639 static int mdt_readpage(struct tgt_session_info *tsi)
1641 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1642 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1643 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1644 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1645 struct mdt_body *repbody;
1651 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1652 RETURN(err_serious(-ENOMEM));
1654 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1655 if (repbody == NULL || reqbody == NULL)
1656 RETURN(err_serious(-EFAULT));
1659 * prepare @rdpg before calling lower layers and transfer itself. Here
1660 * reqbody->size contains offset of where to start to read and
1661 * reqbody->nlink contains number bytes to read.
1663 rdpg->rp_hash = reqbody->mbo_size;
1664 if (rdpg->rp_hash != reqbody->mbo_size) {
1665 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1666 rdpg->rp_hash, reqbody->mbo_size);
1670 rdpg->rp_attrs = reqbody->mbo_mode;
1671 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1672 rdpg->rp_attrs |= LUDA_64BITHASH;
1673 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1674 exp_max_brw_size(tsi->tsi_exp));
1675 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1677 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1678 if (rdpg->rp_pages == NULL)
1681 for (i = 0; i < rdpg->rp_npages; ++i) {
1682 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1683 if (rdpg->rp_pages[i] == NULL)
1684 GOTO(free_rdpg, rc = -ENOMEM);
1687 /* call lower layers to fill allocated pages with directory data */
1688 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1690 GOTO(free_rdpg, rc);
1692 /* send pages to client */
1693 rc = tgt_sendpage(tsi, rdpg, rc);
1698 for (i = 0; i < rdpg->rp_npages; i++)
1699 if (rdpg->rp_pages[i] != NULL)
1700 __free_page(rdpg->rp_pages[i]);
1701 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1703 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1709 static int mdt_reint_internal(struct mdt_thread_info *info,
1710 struct mdt_lock_handle *lhc,
1713 struct req_capsule *pill = info->mti_pill;
1714 struct mdt_body *repbody;
1719 rc = mdt_reint_unpack(info, op);
1721 CERROR("Can't unpack reint, rc %d\n", rc);
1722 RETURN(err_serious(rc));
1725 /* for replay (no_create) lmm is not needed, client has it already */
1726 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1727 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1730 /* llog cookies are always 0, the field is kept for compatibility */
1731 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1732 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1734 rc = req_capsule_server_pack(pill);
1736 CERROR("Can't pack response, rc %d\n", rc);
1737 RETURN(err_serious(rc));
1740 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1741 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1743 repbody->mbo_eadatasize = 0;
1744 repbody->mbo_aclsize = 0;
1747 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1749 /* for replay no cookkie / lmm need, because client have this already */
1750 if (info->mti_spec.no_create)
1751 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1752 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1754 rc = mdt_init_ucred_reint(info);
1756 GOTO(out_shrink, rc);
1758 rc = mdt_fix_attr_ucred(info, op);
1760 GOTO(out_ucred, rc = err_serious(rc));
1762 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, LCF_ASYNC);
2261 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2265 case LDLM_CB_CANCELING:
2266 LDLM_DEBUG(lock, "Revoke remote lock\n");
2274 int mdt_check_resent_lock(struct mdt_thread_info *info,
2275 struct mdt_object *mo,
2276 struct mdt_lock_handle *lhc)
2278 /* the lock might already be gotten in ldlm_handle_enqueue() */
2279 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2280 struct ptlrpc_request *req = mdt_info_req(info);
2281 struct ldlm_lock *lock;
2283 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2284 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2286 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2287 * a resend case, however, it could be already destroyed
2288 * due to client eviction or a raced cancel RPC. */
2289 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
2290 lhc->mlh_reg_lh.cookie);
2294 if (!fid_res_name_eq(mdt_object_fid(mo),
2295 &lock->l_resource->lr_name)) {
2296 CWARN("%s: Although resent, but still not "
2297 "get child lock:"DFID"\n",
2298 info->mti_exp->exp_obd->obd_name,
2299 PFID(mdt_object_fid(mo)));
2300 LDLM_LOCK_PUT(lock);
2303 LDLM_LOCK_PUT(lock);
2309 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
2310 const struct lu_fid *fid, struct lustre_handle *lh,
2311 enum ldlm_mode mode, __u64 ibits, bool nonblock)
2313 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2314 union ldlm_policy_data *policy = &mti->mti_policy;
2315 struct ldlm_res_id *res_id = &mti->mti_res_id;
2319 LASSERT(mdt_object_remote(o));
2321 fid_build_reg_res_name(fid, res_id);
2323 memset(einfo, 0, sizeof(*einfo));
2324 einfo->ei_type = LDLM_IBITS;
2325 einfo->ei_mode = mode;
2326 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2327 einfo->ei_cb_cp = ldlm_completion_ast;
2328 einfo->ei_enq_slave = 0;
2329 einfo->ei_res_id = res_id;
2331 einfo->ei_nonblock = 1;
2333 memset(policy, 0, sizeof(*policy));
2334 policy->l_inodebits.bits = ibits;
2336 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2341 static int mdt_object_local_lock(struct mdt_thread_info *info,
2342 struct mdt_object *o,
2343 struct mdt_lock_handle *lh, __u64 ibits,
2346 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2347 union ldlm_policy_data *policy = &info->mti_policy;
2348 struct ldlm_res_id *res_id = &info->mti_res_id;
2353 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2354 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2355 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2356 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2358 /* Only enqueue LOOKUP lock for remote object */
2359 if (mdt_object_remote(o))
2360 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2362 if (lh->mlh_type == MDT_PDO_LOCK) {
2363 /* check for exists after object is locked */
2364 if (mdt_object_exists(o) == 0) {
2365 /* Non-existent object shouldn't have PDO lock */
2368 /* Non-dir object shouldn't have PDO lock */
2369 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2374 memset(policy, 0, sizeof(*policy));
2375 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2377 dlmflags = LDLM_FL_ATOMIC_CB;
2379 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2382 * Take PDO lock on whole directory and build correct @res_id for lock
2383 * on part of directory.
2385 if (lh->mlh_pdo_hash != 0) {
2386 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2387 mdt_lock_pdo_mode(info, o, lh);
2388 if (lh->mlh_pdo_mode != LCK_NL) {
2390 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2391 * is never going to be sent to client and we do not
2392 * want it slowed down due to possible cancels.
2394 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2395 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2396 policy, res_id, dlmflags,
2397 info->mti_exp == NULL ? NULL :
2398 &info->mti_exp->exp_handle.h_cookie);
2399 if (unlikely(rc != 0))
2400 GOTO(out_unlock, rc);
2404 * Finish res_id initializing by name hash marking part of
2405 * directory which is taking modification.
2407 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2410 policy->l_inodebits.bits = ibits;
2413 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2414 * going to be sent to client. If it is - mdt_intent_policy() path will
2415 * fix it up and turn FL_LOCAL flag off.
2417 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2418 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2419 info->mti_exp == NULL ? NULL :
2420 &info->mti_exp->exp_handle.h_cookie);
2423 mdt_object_unlock(info, o, lh, 1);
2424 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2425 lh->mlh_pdo_hash != 0 &&
2426 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2427 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2433 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2434 struct mdt_lock_handle *lh, __u64 ibits,
2437 struct mdt_lock_handle *local_lh = NULL;
2441 if (!mdt_object_remote(o))
2442 return mdt_object_local_lock(info, o, lh, ibits, nonblock);
2444 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2445 ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2446 MDS_INODELOCK_XATTR);
2448 /* Only enqueue LOOKUP lock for remote object */
2449 if (ibits & MDS_INODELOCK_LOOKUP) {
2450 rc = mdt_object_local_lock(info, o, lh,
2451 MDS_INODELOCK_LOOKUP,
2459 if (ibits & MDS_INODELOCK_UPDATE) {
2460 /* Sigh, PDO needs to enqueue 2 locks right now, but
2461 * enqueue RPC can only request 1 lock, to avoid extra
2462 * RPC, so it will instead enqueue EX lock for remote
2463 * object anyway XXX*/
2464 if (lh->mlh_type == MDT_PDO_LOCK &&
2465 lh->mlh_pdo_hash != 0) {
2466 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2467 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2468 PFID(mdt_object_fid(o)));
2469 lh->mlh_pdo_hash = 0;
2470 lh->mlh_rreg_mode = LCK_EX;
2471 lh->mlh_type = MDT_REG_LOCK;
2473 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2476 MDS_INODELOCK_UPDATE, nonblock);
2477 if (rc != ELDLM_OK) {
2478 if (local_lh != NULL)
2479 mdt_object_unlock(info, o, local_lh, rc);
2487 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2488 struct mdt_lock_handle *lh, __u64 ibits)
2490 return mdt_object_lock_internal(info, o, lh, ibits, false);
2493 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2494 struct mdt_lock_handle *lh, __u64 ibits)
2496 struct mdt_lock_handle tmp = *lh;
2499 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true);
2507 * Save a lock within request object.
2509 * Keep the lock referenced until whether client ACK or transaction
2510 * commit happens or release the lock immediately depending on input
2511 * parameters. If COS is ON, a write lock is converted to COS lock
2514 * \param info thead info object
2515 * \param h lock handle
2516 * \param mode lock mode
2517 * \param decref force immediate lock releasing
2519 static void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2520 enum ldlm_mode mode, int decref)
2524 if (lustre_handle_is_used(h)) {
2525 if (decref || !info->mti_has_trans ||
2526 !(mode & (LCK_PW | LCK_EX))) {
2527 mdt_fid_unlock(h, mode);
2529 struct mdt_device *mdt = info->mti_mdt;
2530 struct ldlm_lock *lock = ldlm_handle2lock(h);
2531 struct ptlrpc_request *req = mdt_info_req(info);
2534 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2536 /* there is no request if mdt_object_unlock() is called
2537 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2538 if (likely(req != NULL)) {
2539 CDEBUG(D_HA, "request = %p reply state = %p"
2540 " transno = "LPD64"\n", req,
2541 req->rq_reply_state, req->rq_transno);
2542 if (mdt_cos_is_enabled(mdt)) {
2544 ldlm_lock_downgrade(lock, LCK_COS);
2547 ptlrpc_save_lock(req, h, mode, no_ack);
2549 ldlm_lock_decref(h, mode);
2551 if (mdt_is_lock_sync(lock)) {
2552 CDEBUG(D_HA, "found sync-lock,"
2553 " async commit started\n");
2554 mdt_device_commit_async(info->mti_env,
2557 LDLM_LOCK_PUT(lock);
2566 * Unlock mdt object.
2568 * Immeditely release the regular lock and the PDO lock or save the
2569 * lock in request and keep them referenced until client ACK or
2570 * transaction commit.
2572 * \param info thread info object
2573 * \param o mdt object
2574 * \param lh mdt lock handle referencing regular and PDO locks
2575 * \param decref force immediate lock releasing
2577 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2578 struct mdt_lock_handle *lh, int decref)
2582 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2583 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2585 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2586 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2591 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2592 const struct lu_fid *f,
2593 struct mdt_lock_handle *lh,
2596 struct mdt_object *o;
2598 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2602 rc = mdt_object_lock(info, o, lh, ibits);
2604 mdt_object_put(info->mti_env, o);
2611 void mdt_object_unlock_put(struct mdt_thread_info * info,
2612 struct mdt_object * o,
2613 struct mdt_lock_handle *lh,
2616 mdt_object_unlock(info, o, lh, decref);
2617 mdt_object_put(info->mti_env, o);
2621 * Generic code handling requests that have struct mdt_body passed in:
2623 * - extract mdt_body from request and save it in @info, if present;
2625 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2628 * - if HABEO_CORPUS flag is set for this request type check whether object
2629 * actually exists on storage (lu_object_exists()).
2632 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2634 const struct mdt_body *body;
2635 struct mdt_object *obj;
2636 const struct lu_env *env;
2637 struct req_capsule *pill;
2641 env = info->mti_env;
2642 pill = info->mti_pill;
2644 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2648 if (!(body->mbo_valid & OBD_MD_FLID))
2651 if (!fid_is_sane(&body->mbo_fid1)) {
2652 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2656 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2658 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2659 mdt_object_put(env, obj);
2662 info->mti_object = obj;
2671 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2673 struct req_capsule *pill = info->mti_pill;
2677 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2678 rc = mdt_body_unpack(info, flags);
2682 if (rc == 0 && (flags & HABEO_REFERO)) {
2684 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2685 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2687 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2688 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2691 rc = req_capsule_server_pack(pill);
2696 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2698 lh->mlh_type = MDT_NUL_LOCK;
2699 lh->mlh_reg_lh.cookie = 0ull;
2700 lh->mlh_reg_mode = LCK_MINMODE;
2701 lh->mlh_pdo_lh.cookie = 0ull;
2702 lh->mlh_pdo_mode = LCK_MINMODE;
2703 lh->mlh_rreg_lh.cookie = 0ull;
2704 lh->mlh_rreg_mode = LCK_MINMODE;
2707 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2709 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2710 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2714 * Initialize fields of struct mdt_thread_info. Other fields are left in
2715 * uninitialized state, because it's too expensive to zero out whole
2716 * mdt_thread_info (> 1K) on each request arrival.
2718 void mdt_thread_info_init(struct ptlrpc_request *req,
2719 struct mdt_thread_info *info)
2723 info->mti_pill = &req->rq_pill;
2726 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2727 mdt_lock_handle_init(&info->mti_lh[i]);
2729 /* mdt device: it can be NULL while CONNECT */
2730 if (req->rq_export) {
2731 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2732 info->mti_exp = req->rq_export;
2734 info->mti_mdt = NULL;
2735 info->mti_env = req->rq_svc_thread->t_env;
2736 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2738 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2739 info->mti_big_buf = LU_BUF_NULL;
2740 info->mti_body = NULL;
2741 info->mti_object = NULL;
2742 info->mti_dlm_req = NULL;
2743 info->mti_has_trans = 0;
2744 info->mti_cross_ref = 0;
2745 info->mti_opdata = 0;
2746 info->mti_big_lmm_used = 0;
2748 info->mti_spec.no_create = 0;
2749 info->mti_spec.sp_rm_entry = 0;
2750 info->mti_spec.sp_permitted = 0;
2751 info->mti_spec.sp_migrate_close = 0;
2753 info->mti_spec.u.sp_ea.eadata = NULL;
2754 info->mti_spec.u.sp_ea.eadatalen = 0;
2757 void mdt_thread_info_fini(struct mdt_thread_info *info)
2761 if (info->mti_object != NULL) {
2762 mdt_object_put(info->mti_env, info->mti_object);
2763 info->mti_object = NULL;
2766 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2767 mdt_lock_handle_fini(&info->mti_lh[i]);
2768 info->mti_env = NULL;
2769 info->mti_pill = NULL;
2770 info->mti_exp = NULL;
2772 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2773 lu_buf_free(&info->mti_big_buf);
2776 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2778 struct mdt_thread_info *mti;
2780 mti = mdt_th_info(tsi->tsi_env);
2781 LASSERT(mti != NULL);
2783 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2784 if (tsi->tsi_corpus != NULL) {
2785 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2786 lu_object_get(tsi->tsi_corpus);
2788 mti->mti_body = tsi->tsi_mdt_body;
2789 mti->mti_dlm_req = tsi->tsi_dlm_req;
2794 static int mdt_tgt_connect(struct tgt_session_info *tsi)
2796 struct ptlrpc_request *req = tgt_ses_req(tsi);
2801 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
2803 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
2804 set_current_state(TASK_UNINTERRUPTIBLE);
2805 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
2808 rc = tgt_connect(tsi);
2812 rc = mdt_init_idmap(tsi);
2817 obd_disconnect(class_export_get(req->rq_export));
2836 static int mdt_intent_getattr(enum mdt_it_code opcode,
2837 struct mdt_thread_info *info,
2838 struct ldlm_lock **,
2841 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2842 struct mdt_thread_info *info,
2843 struct ldlm_lock **lockp,
2846 static int mdt_intent_layout(enum mdt_it_code opcode,
2847 struct mdt_thread_info *info,
2848 struct ldlm_lock **,
2850 static int mdt_intent_reint(enum mdt_it_code opcode,
2851 struct mdt_thread_info *info,
2852 struct ldlm_lock **,
2855 static struct mdt_it_flavor {
2856 const struct req_format *it_fmt;
2858 int (*it_act)(enum mdt_it_code ,
2859 struct mdt_thread_info *,
2860 struct ldlm_lock **,
2863 } mdt_it_flavor[] = {
2865 .it_fmt = &RQF_LDLM_INTENT,
2866 /*.it_flags = HABEO_REFERO,*/
2868 .it_act = mdt_intent_reint,
2869 .it_reint = REINT_OPEN
2872 .it_fmt = &RQF_LDLM_INTENT,
2874 * OCREAT is not a MUTABOR request as if the file
2876 * We do the extra check of OBD_CONNECT_RDONLY in
2877 * mdt_reint_open() when we really need to create
2881 .it_act = mdt_intent_reint,
2882 .it_reint = REINT_OPEN
2885 .it_fmt = &RQF_LDLM_INTENT,
2886 .it_flags = MUTABOR,
2887 .it_act = mdt_intent_reint,
2888 .it_reint = REINT_CREATE
2890 [MDT_IT_GETATTR] = {
2891 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2892 .it_flags = HABEO_REFERO,
2893 .it_act = mdt_intent_getattr
2895 [MDT_IT_READDIR] = {
2901 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2902 .it_flags = HABEO_REFERO,
2903 .it_act = mdt_intent_getattr
2906 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2907 .it_flags = MUTABOR,
2909 .it_reint = REINT_UNLINK
2913 .it_flags = MUTABOR,
2916 [MDT_IT_GETXATTR] = {
2917 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2918 .it_flags = HABEO_CORPUS,
2919 .it_act = mdt_intent_getxattr
2922 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2924 .it_act = mdt_intent_layout
2929 mdt_intent_lock_replace(struct mdt_thread_info *info,
2930 struct ldlm_lock **lockp,
2931 struct mdt_lock_handle *lh,
2934 struct ptlrpc_request *req = mdt_info_req(info);
2935 struct ldlm_lock *lock = *lockp;
2936 struct ldlm_lock *new_lock;
2938 /* If possible resent found a lock, @lh is set to its handle */
2939 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2941 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2942 lh->mlh_reg_lh.cookie = 0;
2946 LASSERTF(new_lock != NULL,
2947 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2950 * If we've already given this lock to a client once, then we should
2951 * have no readers or writers. Otherwise, we should have one reader
2952 * _or_ writer ref (which will be zeroed below) before returning the
2955 if (new_lock->l_export == req->rq_export) {
2956 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2958 LASSERT(new_lock->l_export == NULL);
2959 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2964 if (new_lock->l_export == req->rq_export) {
2966 * Already gave this to the client, which means that we
2967 * reconstructed a reply.
2969 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2972 LDLM_LOCK_RELEASE(new_lock);
2973 lh->mlh_reg_lh.cookie = 0;
2974 RETURN(ELDLM_LOCK_REPLACED);
2978 * Fixup the lock to be given to the client.
2980 lock_res_and_lock(new_lock);
2981 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2982 * possible blocking AST. */
2983 while (new_lock->l_readers > 0) {
2984 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2985 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2986 new_lock->l_readers--;
2988 while (new_lock->l_writers > 0) {
2989 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2990 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2991 new_lock->l_writers--;
2994 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2995 new_lock->l_blocking_ast = lock->l_blocking_ast;
2996 new_lock->l_completion_ast = lock->l_completion_ast;
2997 new_lock->l_remote_handle = lock->l_remote_handle;
2998 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3000 unlock_res_and_lock(new_lock);
3002 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3003 &new_lock->l_remote_handle,
3004 &new_lock->l_exp_hash);
3006 LDLM_LOCK_RELEASE(new_lock);
3007 lh->mlh_reg_lh.cookie = 0;
3009 RETURN(ELDLM_LOCK_REPLACED);
3012 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3013 struct ldlm_lock *new_lock,
3014 struct mdt_lock_handle *lh,
3017 struct ptlrpc_request *req = mdt_info_req(info);
3018 struct ldlm_request *dlmreq;
3020 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3023 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3025 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3026 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3028 if (flags & LDLM_FL_RESENT) {
3029 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3030 lh->mlh_reg_mode = new_lock->l_granted_mode;
3032 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3033 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3034 lh->mlh_reg_lh.cookie);
3039 * If the xid matches, then we know this is a resent request, and allow
3040 * it. (It's probably an OPEN, for which we don't send a lock.
3042 if (req_can_reconstruct(req, NULL))
3046 * This remote handle isn't enqueued, so we never received or processed
3047 * this request. Clear MSG_RESENT, because it can be handled like any
3048 * normal request now.
3050 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3052 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3053 dlmreq->lock_handle[0].cookie);
3056 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3057 struct mdt_thread_info *info,
3058 struct ldlm_lock **lockp,
3061 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3062 struct ldlm_reply *ldlm_rep = NULL;
3066 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3067 * (for the resend case) or a new lock. Below we will use it to
3068 * replace the original lock.
3070 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3071 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3072 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3073 rc = mdt_object_lock(info, info->mti_object, lhc,
3074 MDS_INODELOCK_XATTR);
3079 grc = mdt_getxattr(info);
3081 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3083 if (mdt_info_req(info)->rq_repmsg != NULL)
3084 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3085 if (ldlm_rep == NULL)
3086 RETURN(err_serious(-EFAULT));
3088 ldlm_rep->lock_policy_res2 = grc;
3093 static int mdt_intent_getattr(enum mdt_it_code opcode,
3094 struct mdt_thread_info *info,
3095 struct ldlm_lock **lockp,
3098 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3100 struct ldlm_reply *ldlm_rep;
3101 struct mdt_body *reqbody;
3102 struct mdt_body *repbody;
3106 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3109 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3112 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3113 repbody->mbo_eadatasize = 0;
3114 repbody->mbo_aclsize = 0;
3118 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3120 case MDT_IT_GETATTR:
3121 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3125 CERROR("Unsupported intent (%d)\n", opcode);
3126 GOTO(out_shrink, rc = -EINVAL);
3129 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3131 GOTO(out_shrink, rc);
3133 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3134 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3136 /* Get lock from request for possible resent case. */
3137 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3139 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3140 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3142 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3143 ldlm_rep->lock_policy_res2 = 0;
3144 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3145 ldlm_rep->lock_policy_res2) {
3146 lhc->mlh_reg_lh.cookie = 0ull;
3147 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3150 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3153 mdt_exit_ucred(info);
3155 mdt_client_compatibility(info);
3156 rc2 = mdt_fix_reply(info);
3162 static int mdt_intent_layout(enum mdt_it_code opcode,
3163 struct mdt_thread_info *info,
3164 struct ldlm_lock **lockp,
3167 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3168 struct layout_intent *layout;
3170 struct mdt_object *obj = NULL;
3174 if (opcode != MDT_IT_LAYOUT) {
3175 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3180 fid = &info->mti_tmp_fid2;
3181 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3183 /* Get lock from request for possible resent case. */
3184 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3186 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3188 RETURN(PTR_ERR(obj));
3190 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3191 /* get the length of lsm */
3192 rc = mdt_attr_get_eabuf_size(info, obj);
3194 mdt_object_put(info->mti_env, obj);
3198 if (rc > info->mti_mdt->mdt_max_mdsize)
3199 info->mti_mdt->mdt_max_mdsize = rc;
3202 mdt_object_put(info->mti_env, obj);
3204 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3205 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER, rc);
3206 rc = req_capsule_server_pack(info->mti_pill);
3210 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
3211 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3213 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3214 LASSERT(layout != NULL);
3215 if (layout->li_opc == LAYOUT_INTENT_ACCESS)
3216 /* return to normal/resent ldlm handling */
3219 CERROR("%s: Unsupported layout intent (%d)\n",
3220 mdt_obd_name(info->mti_mdt), layout->li_opc);
3224 static int mdt_intent_reint(enum mdt_it_code opcode,
3225 struct mdt_thread_info *info,
3226 struct ldlm_lock **lockp,
3229 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3230 struct ldlm_reply *rep = NULL;
3234 static const struct req_format *intent_fmts[REINT_MAX] = {
3235 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3236 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3241 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3245 if (mdt_it_flavor[opcode].it_reint != opc) {
3246 CERROR("Reint code %ld doesn't match intent: %d\n",
3248 RETURN(err_serious(-EPROTO));
3251 /* Get lock from request for possible resent case. */
3252 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3254 rc = mdt_reint_internal(info, lhc, opc);
3256 /* Check whether the reply has been packed successfully. */
3257 if (mdt_info_req(info)->rq_repmsg != NULL)
3258 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3260 RETURN(err_serious(-EFAULT));
3262 /* MDC expects this in any case */
3264 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3266 /* the open lock or the lock for cross-ref object should be
3267 * returned to the client */
3268 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3269 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3270 rep->lock_policy_res2 = 0;
3271 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3275 rep->lock_policy_res2 = clear_serious(rc);
3277 if (rep->lock_policy_res2 == -ENOENT &&
3278 mdt_get_disposition(rep, DISP_LOOKUP_NEG))
3279 rep->lock_policy_res2 = 0;
3281 lhc->mlh_reg_lh.cookie = 0ull;
3282 if (rc == -ENOTCONN || rc == -ENODEV ||
3283 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3285 * If it is the disconnect error (ENODEV & ENOCONN), the error
3286 * will be returned by rq_status, and client at ptlrpc layer
3287 * will detect this, then disconnect, reconnect the import
3288 * immediately, instead of impacting the following the rpc.
3293 * For other cases, the error will be returned by intent, and client
3294 * will retrieve the result from intent.
3296 RETURN(ELDLM_LOCK_ABORTED);
3299 static int mdt_intent_code(enum ldlm_intent_flags itcode)
3307 case IT_OPEN|IT_CREAT:
3314 rc = MDT_IT_READDIR;
3317 rc = MDT_IT_GETATTR;
3329 rc = MDT_IT_GETXATTR;
3334 case IT_QUOTA_DQACQ:
3339 CERROR("Unknown intent opcode: 0x%08x\n", itcode);
3346 static int mdt_intent_opc(enum ldlm_intent_flags itopc,
3347 struct mdt_thread_info *info,
3348 struct ldlm_lock **lockp, __u64 flags)
3350 struct req_capsule *pill = info->mti_pill;
3351 struct ptlrpc_request *req = mdt_info_req(info);
3352 struct mdt_it_flavor *flv;
3357 opc = mdt_intent_code(itopc);
3361 if (opc == MDT_IT_QUOTA) {
3362 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3365 RETURN(-EOPNOTSUPP);
3367 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3368 /* pass the request to quota master */
3369 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3370 mdt_info_req(info), lockp,
3375 flv = &mdt_it_flavor[opc];
3376 if (flv->it_fmt != NULL)
3377 req_capsule_extend(pill, flv->it_fmt);
3379 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3383 if (flv->it_flags & MUTABOR &&
3384 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3387 if (flv->it_act != NULL) {
3388 struct ldlm_reply *rep;
3390 /* execute policy */
3391 rc = flv->it_act(opc, info, lockp, flags);
3393 /* Check whether the reply has been packed successfully. */
3394 if (req->rq_repmsg != NULL) {
3395 rep = req_capsule_server_get(info->mti_pill,
3397 rep->lock_policy_res2 =
3398 ptlrpc_status_hton(rep->lock_policy_res2);
3405 static int mdt_intent_policy(struct ldlm_namespace *ns,
3406 struct ldlm_lock **lockp, void *req_cookie,
3407 enum ldlm_mode mode, __u64 flags, void *data)
3409 struct tgt_session_info *tsi;
3410 struct mdt_thread_info *info;
3411 struct ptlrpc_request *req = req_cookie;
3412 struct ldlm_intent *it;
3413 struct req_capsule *pill;
3418 LASSERT(req != NULL);
3420 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3422 info = tsi2mdt_info(tsi);
3423 LASSERT(info != NULL);
3424 pill = info->mti_pill;
3425 LASSERT(pill->rc_req == req);
3427 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3428 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3429 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3431 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3435 /* Lock without inodebits makes no sense and will oops
3436 * later in ldlm. Let's check it now to see if we have
3437 * ibits corrupted somewhere in mdt_intent_opc().
3438 * The case for client miss to set ibits has been
3439 * processed by others. */
3440 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3441 lr_type == LDLM_IBITS,
3442 info->mti_dlm_req->lock_desc.\
3443 l_policy_data.l_inodebits.bits != 0));
3445 rc = err_serious(-EFAULT);
3447 /* No intent was provided */
3448 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3449 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3450 rc = req_capsule_server_pack(pill);
3452 rc = err_serious(rc);
3454 mdt_thread_info_fini(info);
3458 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3460 struct seq_server_site *ss = mdt_seq_site(mdt);
3462 if (ss->ss_node_id == 0)
3465 if (ss->ss_client_seq != NULL) {
3466 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3467 ss->ss_client_seq->lcs_exp = NULL;
3470 if (ss->ss_server_fld != NULL) {
3471 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3472 ss->ss_server_fld->lsf_control_exp = NULL;
3476 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3478 struct seq_server_site *ss = mdt_seq_site(mdt);
3483 if (ss->ss_server_seq != NULL)
3484 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3487 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3489 mdt_seq_fini_cli(mdt);
3490 mdt_deregister_seq_exp(mdt);
3492 return seq_site_fini(env, mdt_seq_site(mdt));
3496 * It will retrieve its FLDB entries from MDT0, and it only happens
3497 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3498 * and it needs to refresh FLDB from the MDT0.
3500 static int mdt_register_lwp_callback(void *data)
3503 struct mdt_device *mdt = data;
3504 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3508 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3510 rc = lu_env_init(&env, LCT_MD_THREAD);
3512 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3516 /* Allocate new sequence now to avoid creating local transaction
3517 * in the normal transaction process */
3518 rc = seq_server_check_and_alloc_super(&env,
3519 mdt_seq_site(mdt)->ss_server_seq);
3524 rc = fld_update_from_controller(&env, fld);
3526 CERROR("%s: cannot update controller: rc = %d\n",
3527 mdt_obd_name(mdt), rc);
3536 static int mdt_register_seq_exp(struct mdt_device *mdt)
3538 struct seq_server_site *ss = mdt_seq_site(mdt);
3539 char *lwp_name = NULL;
3542 if (ss->ss_node_id == 0)
3545 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3546 if (lwp_name == NULL)
3547 GOTO(out_free, rc = -ENOMEM);
3549 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3553 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3558 rc = lustre_register_lwp_item(lwp_name,
3559 &ss->ss_server_fld->lsf_control_exp,
3560 mdt_register_lwp_callback, mdt);
3562 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3563 ss->ss_client_seq->lcs_exp = NULL;
3567 if (lwp_name != NULL)
3568 OBD_FREE(lwp_name, MAX_OBD_NAME);
3574 * Init client sequence manager which is used by local MDS to talk to sequence
3575 * controller on remote node.
3577 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3579 struct seq_server_site *ss = mdt_seq_site(mdt);
3584 /* check if this is adding the first MDC and controller is not yet
3586 OBD_ALLOC_PTR(ss->ss_client_seq);
3587 if (ss->ss_client_seq == NULL)
3590 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3591 if (prefix == NULL) {
3592 OBD_FREE_PTR(ss->ss_client_seq);
3593 ss->ss_client_seq = NULL;
3597 /* Note: seq_client_fini will be called in seq_site_fini */
3598 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3599 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3600 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3602 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3604 OBD_FREE_PTR(ss->ss_client_seq);
3605 ss->ss_client_seq = NULL;
3609 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3614 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3616 struct seq_server_site *ss;
3620 ss = mdt_seq_site(mdt);
3621 /* init sequence controller server(MDT0) */
3622 if (ss->ss_node_id == 0) {
3623 OBD_ALLOC_PTR(ss->ss_control_seq);
3624 if (ss->ss_control_seq == NULL)
3627 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3628 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3631 GOTO(out_seq_fini, rc);
3634 /* Init normal sequence server */
3635 OBD_ALLOC_PTR(ss->ss_server_seq);
3636 if (ss->ss_server_seq == NULL)
3637 GOTO(out_seq_fini, rc = -ENOMEM);
3639 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3640 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3642 GOTO(out_seq_fini, rc);
3644 /* init seq client for seq server to talk to seq controller(MDT0) */
3645 rc = mdt_seq_init_cli(env, mdt);
3647 GOTO(out_seq_fini, rc);
3649 if (ss->ss_node_id != 0)
3650 /* register controller export through lwp */
3651 rc = mdt_register_seq_exp(mdt);
3656 mdt_seq_fini(env, mdt);
3664 static int mdt_fld_fini(const struct lu_env *env,
3665 struct mdt_device *m)
3667 struct seq_server_site *ss = mdt_seq_site(m);
3670 if (ss && ss->ss_server_fld) {
3671 fld_server_fini(env, ss->ss_server_fld);
3672 OBD_FREE_PTR(ss->ss_server_fld);
3673 ss->ss_server_fld = NULL;
3679 static int mdt_fld_init(const struct lu_env *env,
3681 struct mdt_device *m)
3683 struct seq_server_site *ss;
3687 ss = mdt_seq_site(m);
3689 OBD_ALLOC_PTR(ss->ss_server_fld);
3690 if (ss->ss_server_fld == NULL)
3691 RETURN(rc = -ENOMEM);
3693 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3696 OBD_FREE_PTR(ss->ss_server_fld);
3697 ss->ss_server_fld = NULL;
3704 static void mdt_stack_pre_fini(const struct lu_env *env,
3705 struct mdt_device *m, struct lu_device *top)
3707 struct lustre_cfg_bufs *bufs;
3708 struct lustre_cfg *lcfg;
3709 struct mdt_thread_info *info;
3714 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3715 LASSERT(info != NULL);
3717 bufs = &info->mti_u.bufs;
3719 LASSERT(m->mdt_child_exp);
3720 LASSERT(m->mdt_child_exp->exp_obd);
3722 /* process cleanup, pass mdt obd name to get obd umount flags */
3723 /* XXX: this is needed because all layers are referenced by
3724 * objects (some of them are pinned by osd, for example *
3725 * the proper solution should be a model where object used
3726 * by osd only doesn't have mdt/mdd slices -bzzz */
3727 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3728 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3729 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3733 top->ld_ops->ldo_process_config(env, top, lcfg);
3734 lustre_cfg_free(lcfg);
3738 static void mdt_stack_fini(const struct lu_env *env,
3739 struct mdt_device *m, struct lu_device *top)
3741 struct obd_device *obd = mdt2obd_dev(m);
3742 struct lustre_cfg_bufs *bufs;
3743 struct lustre_cfg *lcfg;
3744 struct mdt_thread_info *info;
3748 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3749 LASSERT(info != NULL);
3751 lu_dev_del_linkage(top->ld_site, top);
3753 lu_site_purge(env, top->ld_site, -1);
3755 bufs = &info->mti_u.bufs;
3756 /* process cleanup, pass mdt obd name to get obd umount flags */
3757 /* another purpose is to let all layers to release their objects */
3758 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3763 lustre_cfg_bufs_set_string(bufs, 1, flags);
3764 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3769 top->ld_ops->ldo_process_config(env, top, lcfg);
3770 lustre_cfg_free(lcfg);
3772 lu_site_purge(env, top->ld_site, -1);
3774 m->mdt_child = NULL;
3775 m->mdt_bottom = NULL;
3777 obd_disconnect(m->mdt_child_exp);
3778 m->mdt_child_exp = NULL;
3780 obd_disconnect(m->mdt_bottom_exp);
3781 m->mdt_child_exp = NULL;
3784 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3785 const char *next, struct obd_export **exp)
3787 struct obd_connect_data *data = NULL;
3788 struct obd_device *obd;
3792 OBD_ALLOC_PTR(data);
3794 GOTO(out, rc = -ENOMEM);
3796 obd = class_name2obd(next);
3798 CERROR("%s: can't locate next device: %s\n",
3799 mdt_obd_name(m), next);
3800 GOTO(out, rc = -ENOTCONN);
3803 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3804 data->ocd_version = LUSTRE_VERSION_CODE;
3806 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3808 CERROR("%s: cannot connect to next dev %s (%d)\n",
3809 mdt_obd_name(m), next, rc);
3819 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3820 struct lustre_cfg *cfg)
3822 char *dev = lustre_cfg_string(cfg, 0);
3823 int rc, name_size, uuid_size;
3824 char *name, *uuid, *p;
3825 struct lustre_cfg_bufs *bufs;
3826 struct lustre_cfg *lcfg;
3827 struct obd_device *obd;
3828 struct lustre_profile *lprof;
3829 struct lu_site *site;
3832 /* in 1.8 we had the only device in the stack - MDS.
3833 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3834 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3835 * to generate names and setup MDT, MDD. MDT will be using
3836 * generated name to connect to MDD. for MDD the next device
3837 * will be LOD with name taken from so called "profile" which
3838 * is generated by mount_option line
3840 * 1.8 MGS generates config. commands like this:
3841 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3842 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3843 * 2.0 MGS generates config. commands like this:
3844 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3845 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3846 * 3:lustre-MDT0000-mdtlov 4:f
3848 * we generate MDD name from MDT one, just replacing T with D
3850 * after all the preparations, the logical equivalent will be
3851 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3852 * 3:lustre-MDT0000-mdtlov 4:f
3853 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3854 * 3:lustre-MDD0000 4:f
3856 * notice we build the stack from down to top: MDD first, then MDT */
3858 name_size = MAX_OBD_NAME;
3859 uuid_size = MAX_OBD_NAME;
3861 OBD_ALLOC(name, name_size);
3862 OBD_ALLOC(uuid, uuid_size);
3863 if (name == NULL || uuid == NULL)
3864 GOTO(cleanup_mem, rc = -ENOMEM);
3866 OBD_ALLOC_PTR(bufs);
3868 GOTO(cleanup_mem, rc = -ENOMEM);
3871 p = strstr(name, "-MDT");
3873 GOTO(free_bufs, rc = -ENOMEM);
3876 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3878 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3879 if (lprof == NULL || lprof->lp_dt == NULL) {
3880 CERROR("can't find the profile: %s\n",
3881 lustre_cfg_string(cfg, 0));
3882 GOTO(free_bufs, rc = -EINVAL);
3885 lustre_cfg_bufs_reset(bufs, name);
3886 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3887 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3888 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3890 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3892 GOTO(put_profile, rc = -ENOMEM);
3894 rc = class_attach(lcfg);
3896 GOTO(lcfg_cleanup, rc);
3898 obd = class_name2obd(name);
3900 CERROR("Can not find obd %s (%s in config)\n",
3901 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3902 GOTO(lcfg_cleanup, rc = -EINVAL);
3905 lustre_cfg_free(lcfg);
3907 lustre_cfg_bufs_reset(bufs, name);
3908 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3909 lustre_cfg_bufs_set_string(bufs, 2, dev);
3910 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3912 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3914 GOTO(class_detach, rc = -ENOMEM);
3916 rc = class_setup(obd, lcfg);
3918 GOTO(class_detach, rc);
3920 /* connect to MDD we just setup */
3921 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3923 GOTO(class_detach, rc);
3925 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3927 LASSERT(mdt_lu_site(mdt) == NULL);
3928 mdt->mdt_lu_dev.ld_site = site;
3929 site->ls_top_dev = &mdt->mdt_lu_dev;
3930 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3932 /* now connect to bottom OSD */
3933 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3934 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3936 GOTO(class_detach, rc);
3938 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3940 rc = lu_env_refill((struct lu_env *)env);
3942 CERROR("Failure to refill session: '%d'\n", rc);
3944 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3949 class_detach(obd, lcfg);
3951 lustre_cfg_free(lcfg);
3953 class_put_profile(lprof);
3958 OBD_FREE(name, name_size);
3960 OBD_FREE(uuid, uuid_size);
3964 /* setup quota master target on MDT0 */
3965 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3966 struct lustre_cfg *cfg)
3968 struct obd_device *obd;
3969 char *dev = lustre_cfg_string(cfg, 0);
3970 char *qmtname, *uuid, *p;
3971 struct lustre_cfg_bufs *bufs;
3972 struct lustre_cfg *lcfg;
3973 struct lustre_profile *lprof;
3974 struct obd_connect_data *data;
3978 LASSERT(mdt->mdt_qmt_exp == NULL);
3979 LASSERT(mdt->mdt_qmt_dev == NULL);
3981 /* quota master is on MDT0 only for now */
3982 if (mdt->mdt_seq_site.ss_node_id != 0)
3985 /* MGS generates config commands which look as follows:
3986 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3987 * 3:lustre-MDT0000-mdtlov 4:f
3989 * We generate the QMT name from the MDT one, just replacing MD with QM
3990 * after all the preparations, the logical equivalent will be:
3991 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
3992 * 3:lustre-MDT0000-osd 4:f */
3993 OBD_ALLOC(qmtname, MAX_OBD_NAME);
3994 OBD_ALLOC(uuid, UUID_MAX);
3995 OBD_ALLOC_PTR(bufs);
3996 OBD_ALLOC_PTR(data);
3997 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
3998 GOTO(cleanup_mem, rc = -ENOMEM);
4000 strcpy(qmtname, dev);
4001 p = strstr(qmtname, "-MDT");
4003 GOTO(cleanup_mem, rc = -ENOMEM);
4004 /* replace MD with QM */
4008 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4010 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4011 if (lprof == NULL || lprof->lp_dt == NULL) {
4012 CERROR("can't find profile for %s\n",
4013 lustre_cfg_string(cfg, 0));
4014 GOTO(cleanup_mem, rc = -EINVAL);
4017 lustre_cfg_bufs_reset(bufs, qmtname);
4018 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4019 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4020 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4022 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4024 GOTO(put_profile, rc = -ENOMEM);
4026 rc = class_attach(lcfg);
4028 GOTO(lcfg_cleanup, rc);
4030 obd = class_name2obd(qmtname);
4032 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4033 lustre_cfg_string(cfg, 0));
4034 GOTO(lcfg_cleanup, rc = -EINVAL);
4037 lustre_cfg_free(lcfg);
4039 lustre_cfg_bufs_reset(bufs, qmtname);
4040 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4041 lustre_cfg_bufs_set_string(bufs, 2, dev);
4043 /* for quota, the next device should be the OSD device */
4044 lustre_cfg_bufs_set_string(bufs, 3,
4045 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4047 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4049 GOTO(class_detach, rc = -ENOMEM);
4051 rc = class_setup(obd, lcfg);
4053 GOTO(class_detach, rc);
4055 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4057 /* configure local quota objects */
4058 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4062 GOTO(class_cleanup, rc);
4064 /* connect to quota master target */
4065 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4066 data->ocd_version = LUSTRE_VERSION_CODE;
4067 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4070 CERROR("cannot connect to quota master device %s (%d)\n",
4072 GOTO(class_cleanup, rc);
4078 class_manual_cleanup(obd);
4079 mdt->mdt_qmt_dev = NULL;
4083 class_detach(obd, lcfg);
4085 lustre_cfg_free(lcfg);
4087 class_put_profile(lprof);
4092 OBD_FREE(qmtname, MAX_OBD_NAME);
4094 OBD_FREE(uuid, UUID_MAX);
4100 /* Shutdown quota master target associated with mdt */
4101 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4105 if (mdt->mdt_qmt_exp == NULL)
4107 LASSERT(mdt->mdt_qmt_dev != NULL);
4109 /* the qmt automatically shuts down when the mdt disconnects */
4110 obd_disconnect(mdt->mdt_qmt_exp);
4111 mdt->mdt_qmt_exp = NULL;
4112 mdt->mdt_qmt_dev = NULL;
4116 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4117 * for now. This will be removed along with converting rest of MDT code
4118 * to use tgt_session_info */
4119 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4121 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4124 rc = mdt_getxattr(info);
4126 mdt_thread_info_fini(info);
4130 static struct tgt_handler mdt_tgt_handlers[] = {
4131 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4132 0, MDS_CONNECT, mdt_tgt_connect,
4133 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4134 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4135 0, MDS_DISCONNECT, tgt_disconnect,
4136 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4137 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4138 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4139 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4140 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4141 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4142 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4143 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4145 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4146 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4147 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4148 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4149 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4150 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4151 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4152 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4154 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4155 mdt_hsm_ct_register),
4156 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4157 mdt_hsm_ct_unregister),
4158 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4160 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4162 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4163 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4165 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4170 static struct tgt_handler mdt_sec_ctx_ops[] = {
4171 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4172 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4173 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4176 static struct tgt_handler mdt_quota_ops[] = {
4177 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4180 static struct tgt_opc_slice mdt_common_slice[] = {
4182 .tos_opc_start = MDS_FIRST_OPC,
4183 .tos_opc_end = MDS_LAST_OPC,
4184 .tos_hs = mdt_tgt_handlers
4187 .tos_opc_start = OBD_FIRST_OPC,
4188 .tos_opc_end = OBD_LAST_OPC,
4189 .tos_hs = tgt_obd_handlers
4192 .tos_opc_start = LDLM_FIRST_OPC,
4193 .tos_opc_end = LDLM_LAST_OPC,
4194 .tos_hs = tgt_dlm_handlers
4197 .tos_opc_start = SEC_FIRST_OPC,
4198 .tos_opc_end = SEC_LAST_OPC,
4199 .tos_hs = mdt_sec_ctx_ops
4202 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4203 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4204 .tos_hs = tgt_out_handlers
4207 .tos_opc_start = FLD_FIRST_OPC,
4208 .tos_opc_end = FLD_LAST_OPC,
4209 .tos_hs = fld_handlers
4212 .tos_opc_start = SEQ_FIRST_OPC,
4213 .tos_opc_end = SEQ_LAST_OPC,
4214 .tos_hs = seq_handlers
4217 .tos_opc_start = QUOTA_DQACQ,
4218 .tos_opc_end = QUOTA_LAST_OPC,
4219 .tos_hs = mdt_quota_ops
4222 .tos_opc_start = LLOG_FIRST_OPC,
4223 .tos_opc_end = LLOG_LAST_OPC,
4224 .tos_hs = tgt_llog_handlers
4227 .tos_opc_start = LFSCK_FIRST_OPC,
4228 .tos_opc_end = LFSCK_LAST_OPC,
4229 .tos_hs = tgt_lfsck_handlers
4237 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4239 struct md_device *next = m->mdt_child;
4240 struct lu_device *d = &m->mdt_lu_dev;
4241 struct obd_device *obd = mdt2obd_dev(m);
4242 struct lfsck_stop stop;
4245 stop.ls_status = LS_PAUSED;
4247 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4249 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4250 target_recovery_fini(obd);
4251 ping_evictor_stop();
4253 if (m->mdt_opts.mo_coordinator)
4254 mdt_hsm_cdt_stop(m);
4256 mdt_hsm_cdt_fini(m);
4258 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4259 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4261 if (m->mdt_namespace != NULL)
4262 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
4263 d->ld_obd->obd_force);
4265 obd_exports_barrier(obd);
4266 obd_zombie_barrier();
4270 tgt_fini(env, &m->mdt_lut);
4271 mdt_fs_cleanup(env, m);
4272 upcall_cache_cleanup(m->mdt_identity_cache);
4273 m->mdt_identity_cache = NULL;
4275 if (m->mdt_namespace != NULL) {
4276 ldlm_namespace_free_post(m->mdt_namespace);
4277 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4280 mdt_quota_fini(env, m);
4282 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4284 mdt_seq_fini(env, m);
4285 mdt_fld_fini(env, m);
4290 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4292 LASSERT(atomic_read(&d->ld_ref) == 0);
4294 server_put_mount(mdt_obd_name(m), true);
4299 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4301 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4302 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4304 struct mdt_thread_info *info;
4305 struct obd_device *obd;
4306 const char *dev = lustre_cfg_string(cfg, 0);
4307 const char *num = lustre_cfg_string(cfg, 2);
4308 struct lustre_mount_info *lmi = NULL;
4309 struct lustre_sb_info *lsi;
4311 struct seq_server_site *ss_site;
4312 const char *identity_upcall = "NONE";
4313 struct md_device *next;
4319 lu_device_init(&m->mdt_lu_dev, ldt);
4321 * Environment (env) might be missing mdt_thread_key values at that
4322 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4325 * Usually device allocation path doesn't use module key values, but
4326 * mdt has to do a lot of work here, so allocate key value.
4328 rc = lu_env_refill((struct lu_env *)env);
4332 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4333 LASSERT(info != NULL);
4335 obd = class_name2obd(dev);
4336 LASSERT(obd != NULL);
4338 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4340 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4342 /* default is coordinator off, it is started through conf_param
4344 m->mdt_opts.mo_coordinator = 0;
4346 lmi = server_get_mount(dev);
4348 CERROR("Cannot get mount info for %s!\n", dev);
4351 lsi = s2lsi(lmi->lmi_sb);
4352 /* CMD is supported only in IAM mode */
4354 node_id = simple_strtol(num, NULL, 10);
4355 obd->u.obt.obt_magic = OBT_MAGIC;
4358 m->mdt_squash.rsi_uid = 0;
4359 m->mdt_squash.rsi_gid = 0;
4360 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4361 init_rwsem(&m->mdt_squash.rsi_sem);
4362 spin_lock_init(&m->mdt_osfs_lock);
4363 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4364 m->mdt_enable_remote_dir = 0;
4365 m->mdt_enable_remote_dir_gid = 0;
4367 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4368 m->mdt_lu_dev.ld_obd = obd;
4369 /* Set this lu_device to obd for error handling purposes. */
4370 obd->obd_lu_dev = &m->mdt_lu_dev;
4372 /* init the stack */
4373 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4375 CERROR("%s: Can't init device stack, rc %d\n",
4376 mdt_obd_name(m), rc);
4381 ss_site = mdt_seq_site(m);
4382 s->ld_seq_site = ss_site;
4385 /* set server index */
4386 ss_site->ss_node_id = node_id;
4388 /* failover is the default
4389 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4390 * assumed whose ss_node_id == 0 XXX
4392 obd->obd_replayable = 1;
4393 /* No connection accepted until configurations will finish */
4394 obd->obd_no_conn = 1;
4396 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4397 char *str = lustre_cfg_string(cfg, 4);
4398 if (strchr(str, 'n')) {
4399 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4400 obd->obd_replayable = 0;
4404 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4406 GOTO(err_fini_stack, rc);
4408 rc = mdt_seq_init(env, m);
4410 GOTO(err_fini_fld, rc);
4412 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4413 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4414 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4415 LDLM_NAMESPACE_SERVER,
4416 LDLM_NAMESPACE_GREEDY,
4418 if (m->mdt_namespace == NULL)
4419 GOTO(err_fini_seq, rc = -ENOMEM);
4421 m->mdt_namespace->ns_lvbp = m;
4422 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4424 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4425 /* set obd_namespace for compatibility with old code */
4426 obd->obd_namespace = m->mdt_namespace;
4428 rc = mdt_hsm_cdt_init(m);
4430 CERROR("%s: error initializing coordinator, rc %d\n",
4431 mdt_obd_name(m), rc);
4432 GOTO(err_free_ns, rc);
4435 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4436 OBD_FAIL_MDS_ALL_REQUEST_NET,
4437 OBD_FAIL_MDS_ALL_REPLY_NET);
4439 GOTO(err_free_hsm, rc);
4441 rc = mdt_fs_setup(env, m, obd, lsi);
4445 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4447 next = m->mdt_child;
4448 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4451 GOTO(err_fs_cleanup, rc);
4453 if (mntopts & MNTOPT_USERXATTR)
4454 m->mdt_opts.mo_user_xattr = 1;
4456 m->mdt_opts.mo_user_xattr = 0;
4458 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4460 GOTO(err_fs_cleanup, rc);
4462 if (mntopts & MNTOPT_ACL)
4463 m->mdt_opts.mo_acl = 1;
4465 m->mdt_opts.mo_acl = 0;
4467 /* XXX: to support suppgid for ACL, we enable identity_upcall
4468 * by default, otherwise, maybe got unexpected -EACCESS. */
4469 if (m->mdt_opts.mo_acl)
4470 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4472 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4474 &mdt_identity_upcall_cache_ops);
4475 if (IS_ERR(m->mdt_identity_cache)) {
4476 rc = PTR_ERR(m->mdt_identity_cache);
4477 m->mdt_identity_cache = NULL;
4478 GOTO(err_fs_cleanup, rc);
4481 rc = mdt_procfs_init(m, dev);
4483 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4484 GOTO(err_recovery, rc);
4487 rc = mdt_quota_init(env, m, cfg);
4489 GOTO(err_procfs, rc);
4491 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4492 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4493 "mdt_ldlm_client", m->mdt_ldlm_client);
4495 ping_evictor_start();
4497 /* recovery will be started upon mdt_prepare()
4498 * when the whole stack is complete and ready
4499 * to serve the requests */
4501 /* Reduce the initial timeout on an MDS because it doesn't need such
4502 * a long timeout as an OST does. Adaptive timeouts will adjust this
4503 * value appropriately. */
4504 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4505 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4511 target_recovery_fini(obd);
4512 upcall_cache_cleanup(m->mdt_identity_cache);
4513 m->mdt_identity_cache = NULL;
4515 mdt_fs_cleanup(env, m);
4517 tgt_fini(env, &m->mdt_lut);
4519 mdt_hsm_cdt_fini(m);
4521 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4522 obd->obd_namespace = m->mdt_namespace = NULL;
4524 mdt_seq_fini(env, m);
4526 mdt_fld_fini(env, m);
4528 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4531 server_put_mount(dev, true);
4535 /* For interoperability, the left element is old parameter, the right one
4536 * is the new version of the parameter, if some parameter is deprecated,
4537 * the new version should be set as NULL. */
4538 static struct cfg_interop_param mdt_interop_param[] = {
4539 { "mdt.group_upcall", NULL },
4540 { "mdt.quota_type", NULL },
4541 { "mdd.quota_type", NULL },
4542 { "mdt.rootsquash", "mdt.root_squash" },
4543 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4547 /* used by MGS to process specific configurations */
4548 static int mdt_process_config(const struct lu_env *env,
4549 struct lu_device *d, struct lustre_cfg *cfg)
4551 struct mdt_device *m = mdt_dev(d);
4552 struct md_device *md_next = m->mdt_child;
4553 struct lu_device *next = md2lu_dev(md_next);
4557 switch (cfg->lcfg_command) {
4559 struct obd_device *obd = d->ld_obd;
4561 /* For interoperability */
4562 struct cfg_interop_param *ptr = NULL;
4563 struct lustre_cfg *old_cfg = NULL;
4566 param = lustre_cfg_string(cfg, 1);
4567 if (param == NULL) {
4568 CERROR("param is empty\n");
4573 ptr = class_find_old_param(param, mdt_interop_param);
4575 if (ptr->new_param == NULL) {
4577 CWARN("For interoperability, skip this %s."
4578 " It is obsolete.\n", ptr->old_param);
4582 CWARN("Found old param %s, changed it to %s.\n",
4583 ptr->old_param, ptr->new_param);
4586 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4593 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4595 if (rc > 0 || rc == -ENOSYS) {
4596 /* is it an HSM var ? */
4597 rc = class_process_proc_param(PARAM_HSM,
4598 hsm_cdt_get_proc_vars(),
4600 if (rc > 0 || rc == -ENOSYS)
4601 /* we don't understand; pass it on */
4602 rc = next->ld_ops->ldo_process_config(env, next,
4606 if (old_cfg != NULL)
4607 lustre_cfg_free(cfg);
4612 /* others are passed further */
4613 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4619 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4620 const struct lu_object_header *hdr,
4621 struct lu_device *d)
4623 struct mdt_object *mo;
4627 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4629 struct lu_object *o;
4630 struct lu_object_header *h;
4633 h = &mo->mot_header;
4634 lu_object_header_init(h);
4635 lu_object_init(o, h, d);
4636 lu_object_add_top(h, o);
4637 o->lo_ops = &mdt_obj_ops;
4638 spin_lock_init(&mo->mot_write_lock);
4639 mutex_init(&mo->mot_lov_mutex);
4640 init_rwsem(&mo->mot_open_sem);
4646 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4647 const struct lu_object_conf *unused)
4649 struct mdt_device *d = mdt_dev(o->lo_dev);
4650 struct lu_device *under;
4651 struct lu_object *below;
4655 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4656 PFID(lu_object_fid(o)));
4658 under = &d->mdt_child->md_lu_dev;
4659 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4660 if (below != NULL) {
4661 lu_object_add(o, below);
4668 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4670 struct mdt_object *mo = mdt_obj(o);
4671 struct lu_object_header *h;
4675 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4676 PFID(lu_object_fid(o)));
4678 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4679 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4682 lu_object_header_fini(h);
4683 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4688 static int mdt_object_print(const struct lu_env *env, void *cookie,
4689 lu_printer_t p, const struct lu_object *o)
4691 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4693 return (*p)(env, cookie,
4694 LUSTRE_MDT_NAME"-object@%p(flags=%d, writecount=%d)",
4695 mdto, mdto->mot_flags, mdto->mot_write_count);
4698 static int mdt_prepare(const struct lu_env *env,
4699 struct lu_device *pdev,
4700 struct lu_device *cdev)
4702 struct mdt_device *mdt = mdt_dev(cdev);
4703 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4704 struct obd_device *obd = cdev->ld_obd;
4711 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4715 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4719 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4723 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4724 /* The LFSCK instance is registered just now, so it must be there when
4725 * register the namespace to such instance. */
4726 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4728 if (mdt->mdt_seq_site.ss_node_id == 0) {
4729 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4730 &mdt->mdt_md_root_fid);
4735 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4737 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4738 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4739 LASSERT(obd->obd_no_conn);
4740 spin_lock(&obd->obd_dev_lock);
4741 obd->obd_no_conn = 0;
4742 spin_unlock(&obd->obd_dev_lock);
4744 if (obd->obd_recovering == 0)
4745 mdt_postrecov(env, mdt);
4750 const struct lu_device_operations mdt_lu_ops = {
4751 .ldo_object_alloc = mdt_object_alloc,
4752 .ldo_process_config = mdt_process_config,
4753 .ldo_prepare = mdt_prepare,
4756 static const struct lu_object_operations mdt_obj_ops = {
4757 .loo_object_init = mdt_object_init,
4758 .loo_object_free = mdt_object_free,
4759 .loo_object_print = mdt_object_print
4762 static int mdt_obd_set_info_async(const struct lu_env *env,
4763 struct obd_export *exp,
4764 __u32 keylen, void *key,
4765 __u32 vallen, void *val,
4766 struct ptlrpc_request_set *set)
4772 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4773 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4781 * Match client and server connection feature flags.
4783 * Compute the compatibility flags for a connection request based on
4784 * features mutually supported by client and server.
4786 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4787 * must not be updated here, otherwise a partially initialized value may
4788 * be exposed. After the connection request is successfully processed,
4789 * the top-level MDT connect request handler atomically updates the export
4790 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4791 * reply. \see mdt_connect().
4793 * \param exp the obd_export associated with this client/target pair
4794 * \param mdt the target device for the connection
4795 * \param data stores data for this connect request
4798 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4799 * \retval -EBADE client and server feature requirements are incompatible
4801 static int mdt_connect_internal(struct obd_export *exp,
4802 struct mdt_device *mdt,
4803 struct obd_connect_data *data)
4805 LASSERT(data != NULL);
4807 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4808 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4810 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
4811 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
4812 CWARN("%s: client %s does not support ibits lock, either "
4813 "very old or an invalid client: flags "LPX64"\n",
4814 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
4815 data->ocd_connect_flags);
4819 if (!mdt->mdt_opts.mo_acl)
4820 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4822 if (!mdt->mdt_opts.mo_user_xattr)
4823 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4825 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4826 data->ocd_brw_size = min(data->ocd_brw_size,
4827 (__u32)MD_MAX_BRW_SIZE);
4828 if (data->ocd_brw_size == 0) {
4829 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4830 " ocd_version: %x ocd_grant: %d "
4831 "ocd_index: %u ocd_brw_size is "
4832 "unexpectedly zero, network data "
4833 "corruption? Refusing connection of this"
4836 exp->exp_client_uuid.uuid,
4837 exp, data->ocd_connect_flags, data->ocd_version,
4838 data->ocd_grant, data->ocd_index);
4843 /* NB: Disregard the rule against updating
4844 * exp_connect_data.ocd_connect_flags in this case, since
4845 * tgt_client_new() needs to know if this is a lightweight
4846 * connection, and it is safe to expose this flag before
4847 * connection processing completes. */
4848 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4849 spin_lock(&exp->exp_lock);
4850 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4851 spin_unlock(&exp->exp_lock);
4854 data->ocd_version = LUSTRE_VERSION_CODE;
4856 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4857 CWARN("%s: MDS requires FID support, but client not\n",
4862 if (OCD_HAS_FLAG(data, PINGLESS)) {
4863 if (ptlrpc_pinger_suppress_pings()) {
4864 spin_lock(&exp->exp_obd->obd_dev_lock);
4865 list_del_init(&exp->exp_obd_chain_timed);
4866 spin_unlock(&exp->exp_obd->obd_dev_lock);
4868 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4872 data->ocd_max_easize = mdt->mdt_max_ea_size;
4874 /* NB: Disregard the rule against updating
4875 * exp_connect_data.ocd_connect_flags in this case, since
4876 * tgt_client_new() needs to know if this is client supports
4877 * multiple modify RPCs, and it is safe to expose this flag before
4878 * connection processing completes. */
4879 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
4880 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
4881 spin_lock(&exp->exp_lock);
4882 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
4883 spin_unlock(&exp->exp_lock);
4889 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
4890 struct mdt_thread_info *info,
4891 struct mdt_file_data *mfd)
4893 struct lu_context ses;
4897 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
4902 lu_context_enter(&ses);
4904 mdt_ucred(info)->uc_valid = UCRED_OLD;
4905 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
4907 lu_context_exit(&ses);
4908 lu_context_fini(&ses);
4914 static int mdt_export_cleanup(struct obd_export *exp)
4916 struct list_head closing_list;
4917 struct mdt_export_data *med = &exp->exp_mdt_data;
4918 struct obd_device *obd = exp->exp_obd;
4919 struct mdt_device *mdt;
4920 struct mdt_thread_info *info;
4922 struct mdt_file_data *mfd, *n;
4926 INIT_LIST_HEAD(&closing_list);
4927 spin_lock(&med->med_open_lock);
4928 while (!list_empty(&med->med_open_head)) {
4929 struct list_head *tmp = med->med_open_head.next;
4930 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
4932 /* Remove mfd handle so it can't be found again.
4933 * We are consuming the mfd_list reference here. */
4934 class_handle_unhash(&mfd->mfd_handle);
4935 list_move_tail(&mfd->mfd_list, &closing_list);
4937 spin_unlock(&med->med_open_lock);
4938 mdt = mdt_dev(obd->obd_lu_dev);
4939 LASSERT(mdt != NULL);
4941 rc = lu_env_init(&env, LCT_MD_THREAD);
4945 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
4946 LASSERT(info != NULL);
4947 memset(info, 0, sizeof *info);
4948 info->mti_env = &env;
4949 info->mti_mdt = mdt;
4950 info->mti_exp = exp;
4952 if (!list_empty(&closing_list)) {
4953 struct md_attr *ma = &info->mti_attr;
4955 /* Close any open files (which may also cause orphan
4957 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
4958 list_del_init(&mfd->mfd_list);
4959 ma->ma_need = ma->ma_valid = 0;
4961 /* This file is being closed due to an eviction, it
4962 * could have been modified and now dirty regarding to
4963 * HSM archive, check this!
4964 * The logic here is to mark a file dirty if there's a
4965 * chance it was dirtied before the client was evicted,
4966 * so that we don't have to wait for a release attempt
4967 * before finding out the file was actually dirty and
4968 * fail the release. Aggressively marking it dirty here
4969 * will cause the policy engine to attempt to
4970 * re-archive it; when rearchiving, we can compare the
4971 * current version to the HSM data_version and make the
4972 * archive request into a noop if it's not actually
4975 if (mfd->mfd_mode & FMODE_WRITE)
4976 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
4978 /* Don't unlink orphan on failover umount, LU-184 */
4979 if (exp->exp_flags & OBD_OPT_FAILOVER) {
4980 ma->ma_valid = MA_FLAGS;
4981 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
4983 mdt_mfd_close(info, mfd);
4986 info->mti_mdt = NULL;
4987 /* cleanup client slot early */
4988 /* Do not erase record for recoverable client. */
4989 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
4990 tgt_client_del(&env, exp);
4996 static int mdt_obd_disconnect(struct obd_export *exp)
5002 class_export_get(exp);
5004 nodemap_del_member(exp);
5005 rc = server_disconnect_export(exp);
5007 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
5009 rc = mdt_export_cleanup(exp);
5010 class_export_put(exp);
5014 /* mds_connect copy */
5015 static int mdt_obd_connect(const struct lu_env *env,
5016 struct obd_export **exp, struct obd_device *obd,
5017 struct obd_uuid *cluuid,
5018 struct obd_connect_data *data,
5021 struct obd_export *lexp;
5022 struct lustre_handle conn = { 0 };
5023 struct mdt_device *mdt;
5025 lnet_nid_t *client_nid = localdata;
5028 LASSERT(env != NULL);
5029 if (!exp || !obd || !cluuid)
5032 mdt = mdt_dev(obd->obd_lu_dev);
5035 * first, check whether the stack is ready to handle requests
5036 * XXX: probably not very appropriate method is used now
5037 * at some point we should find a better one
5039 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
5040 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5041 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5042 rc = obd_get_info(env, mdt->mdt_child_exp,
5043 sizeof(KEY_OSP_CONNECTED),
5044 KEY_OSP_CONNECTED, NULL, NULL);
5047 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5050 rc = class_connect(&conn, obd, cluuid);
5054 lexp = class_conn2export(&conn);
5055 LASSERT(lexp != NULL);
5057 rc = mdt_connect_internal(lexp, mdt, data);
5059 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5062 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5063 rc = tgt_client_new(env, lexp);
5065 rc = nodemap_add_member(*client_nid, lexp);
5066 if (rc != 0 && rc != -EEXIST)
5069 mdt_export_stats_init(obd, lexp, localdata);
5074 class_disconnect(lexp);
5083 static int mdt_obd_reconnect(const struct lu_env *env,
5084 struct obd_export *exp, struct obd_device *obd,
5085 struct obd_uuid *cluuid,
5086 struct obd_connect_data *data,
5089 lnet_nid_t *client_nid = localdata;
5093 if (exp == NULL || obd == NULL || cluuid == NULL)
5096 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5098 rc = nodemap_add_member(*client_nid, exp);
5099 if (rc == 0 || rc == -EEXIST)
5100 mdt_export_stats_init(obd, exp, localdata);
5106 /* FIXME: Can we avoid using these two interfaces? */
5107 static int mdt_init_export(struct obd_export *exp)
5109 struct mdt_export_data *med = &exp->exp_mdt_data;
5113 INIT_LIST_HEAD(&med->med_open_head);
5114 spin_lock_init(&med->med_open_lock);
5115 mutex_init(&med->med_idmap_mutex);
5116 med->med_idmap = NULL;
5117 spin_lock(&exp->exp_lock);
5118 exp->exp_connecting = 1;
5119 spin_unlock(&exp->exp_lock);
5121 /* self-export doesn't need client data and ldlm initialization */
5122 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5123 &exp->exp_client_uuid)))
5126 rc = tgt_client_alloc(exp);
5130 rc = ldlm_init_export(exp);
5137 tgt_client_free(exp);
5139 CERROR("%s: Failed to initialize export: rc = %d\n",
5140 exp->exp_obd->obd_name, rc);
5144 static int mdt_destroy_export(struct obd_export *exp)
5148 if (exp_connect_rmtclient(exp))
5149 mdt_cleanup_idmap(&exp->exp_mdt_data);
5151 target_destroy_export(exp);
5152 /* destroy can be called from failed obd_setup, so
5153 * checking uuid is safer than obd_self_export */
5154 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5155 &exp->exp_client_uuid)))
5158 ldlm_destroy_export(exp);
5159 tgt_client_free(exp);
5161 LASSERT(list_empty(&exp->exp_outstanding_replies));
5162 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5167 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5168 struct linkea_data *ldata)
5172 LASSERT(ldata->ld_buf->lb_buf != NULL);
5174 if (!mdt_object_exists(mdt_obj))
5177 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5178 ldata->ld_buf, XATTR_NAME_LINK);
5179 if (rc == -ERANGE) {
5180 /* Buf was too small, figure out what we need. */
5181 lu_buf_free(ldata->ld_buf);
5182 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5183 ldata->ld_buf, XATTR_NAME_LINK);
5186 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5187 if (ldata->ld_buf->lb_buf == NULL)
5189 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5190 ldata->ld_buf, XATTR_NAME_LINK);
5195 return linkea_init(ldata);
5199 * Given an MDT object, try to look up the full path to the object.
5200 * Part of the MDT layer implementation of lfs fid2path.
5202 * \param[in] info Per-thread common data shared by MDT level handlers.
5203 * \param[in] obj Object to do path lookup of
5204 * \param[in,out] fp User-provided struct to store path information
5206 * \retval 0 Lookup successful, path information stored in fp
5207 * \retval -EAGAIN Lookup failed, usually because object is being moved
5208 * \retval negative errno if there was a problem
5210 static int mdt_path_current(struct mdt_thread_info *info,
5211 struct mdt_object *obj,
5212 struct getinfo_fid2path *fp)
5214 struct mdt_device *mdt = info->mti_mdt;
5215 struct mdt_object *mdt_obj;
5216 struct link_ea_header *leh;
5217 struct link_ea_entry *lee;
5218 struct lu_name *tmpname = &info->mti_name;
5219 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5220 struct lu_buf *buf = &info->mti_big_buf;
5223 struct linkea_data ldata = { NULL };
5228 /* temp buffer for path element, the buffer will be finally freed
5229 * in mdt_thread_info_fini */
5230 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5231 if (buf->lb_buf == NULL)
5235 ptr = fp->gf_path + fp->gf_pathlen - 1;
5238 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5240 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5241 * so checking root_fid can only happen on MDT0. */
5242 while (!lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid)) {
5243 struct lu_buf lmv_buf;
5245 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5246 if (IS_ERR(mdt_obj))
5247 GOTO(out, rc = PTR_ERR(mdt_obj));
5249 if (!mdt_object_exists(mdt_obj)) {
5250 mdt_object_put(info->mti_env, mdt_obj);
5251 GOTO(out, rc = -ENOENT);
5254 if (mdt_object_remote(mdt_obj)) {
5255 mdt_object_put(info->mti_env, mdt_obj);
5256 GOTO(remote_out, rc = -EREMOTE);
5259 rc = mdt_links_read(info, mdt_obj, &ldata);
5261 mdt_object_put(info->mti_env, mdt_obj);
5266 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5267 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5268 /* If set, use link #linkno for path lookup, otherwise use
5269 link #0. Only do this for the final path element. */
5270 if (first && fp->gf_linkno < leh->leh_reccount) {
5272 for (count = 0; count < fp->gf_linkno; count++) {
5273 lee = (struct link_ea_entry *)
5274 ((char *)lee + reclen);
5275 linkea_entry_unpack(lee, &reclen, tmpname,
5278 if (fp->gf_linkno < leh->leh_reccount - 1)
5279 /* indicate to user there are more links */
5283 lmv_buf.lb_buf = info->mti_xattr_buf;
5284 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5285 /* Check if it is slave stripes */
5286 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5287 &lmv_buf, XATTR_NAME_LMV);
5288 mdt_object_put(info->mti_env, mdt_obj);
5290 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5292 /* For slave stripes, get its master */
5293 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5294 fp->gf_fid = *tmpfid;
5297 } else if (rc < 0 && rc != -ENODATA) {
5303 /* Pack the name in the end of the buffer */
5304 ptr -= tmpname->ln_namelen;
5305 if (ptr - 1 <= fp->gf_path)
5306 GOTO(out, rc = -EOVERFLOW);
5307 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5310 /* keep the last resolved fid to the client, so the
5311 * client will build the left path on another MDT for
5313 fp->gf_fid = *tmpfid;
5319 ptr++; /* skip leading / */
5320 memmove(fp->gf_path, ptr, fp->gf_path + fp->gf_pathlen - ptr);
5327 * Given an MDT object, use mdt_path_current to get the path.
5328 * Essentially a wrapper to retry mdt_path_current a set number of times
5329 * if -EAGAIN is returned (usually because an object is being moved).
5331 * Part of the MDT layer implementation of lfs fid2path.
5333 * \param[in] info Per-thread common data shared by mdt level handlers.
5334 * \param[in] obj Object to do path lookup of
5335 * \param[in,out] fp User-provided struct for arguments and to store path
5338 * \retval 0 Lookup successful, path information stored in fp
5339 * \retval negative errno if there was a problem
5341 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5342 struct getinfo_fid2path *fp)
5344 struct mdt_device *mdt = info->mti_mdt;
5349 if (fp->gf_pathlen < 3)
5352 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5353 fp->gf_path[0] = '\0';
5357 /* Retry multiple times in case file is being moved */
5358 while (tries-- && rc == -EAGAIN)
5359 rc = mdt_path_current(info, obj, fp);
5365 * Get the full path of the provided FID, as of changelog record recno.
5367 * This checks sanity and looks up object for user provided FID
5368 * before calling the actual path lookup code.
5370 * Part of the MDT layer implementation of lfs fid2path.
5372 * \param[in] info Per-thread common data shared by mdt level handlers.
5373 * \param[in,out] fp User-provided struct for arguments and to store path
5376 * \retval 0 Lookup successful, path information and recno stored in fp
5377 * \retval -ENOENT, object does not exist
5378 * \retval negative errno if there was a problem
5380 static int mdt_fid2path(struct mdt_thread_info *info,
5381 struct getinfo_fid2path *fp)
5383 struct mdt_device *mdt = info->mti_mdt;
5384 struct mdt_object *obj;
5388 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5389 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5391 if (!fid_is_sane(&fp->gf_fid))
5394 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5395 CWARN("%s: "DFID" is invalid, sequence should be "
5396 ">= "LPX64"\n", mdt_obd_name(mdt),
5397 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5401 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5404 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5405 PFID(&fp->gf_fid), rc);
5409 if (mdt_object_remote(obj))
5411 else if (!mdt_object_exists(obj))
5417 mdt_object_put(info->mti_env, obj);
5418 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5419 PFID(&fp->gf_fid), rc);
5423 rc = mdt_path(info, obj, fp);
5425 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5426 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5428 mdt_object_put(info->mti_env, obj);
5433 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5434 void *val, int vallen)
5436 struct getinfo_fid2path *fpout, *fpin;
5439 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5442 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5443 lustre_swab_fid2path(fpin);
5445 memcpy(fpout, fpin, sizeof(*fpin));
5446 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5449 rc = mdt_fid2path(info, fpout);
5453 int mdt_get_info(struct tgt_session_info *tsi)
5463 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5465 CDEBUG(D_IOCTL, "No GETINFO key\n");
5466 RETURN(err_serious(-EFAULT));
5468 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5471 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5472 if (vallen == NULL) {
5473 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5474 tgt_name(tsi->tsi_tgt));
5475 RETURN(err_serious(-EFAULT));
5478 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5480 rc = req_capsule_server_pack(tsi->tsi_pill);
5482 RETURN(err_serious(rc));
5484 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5485 if (valout == NULL) {
5486 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5487 tgt_name(tsi->tsi_tgt));
5488 RETURN(err_serious(-EFAULT));
5491 if (KEY_IS(KEY_FID2PATH)) {
5492 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5494 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5495 mdt_thread_info_fini(info);
5502 /* Pass the ioc down */
5503 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5504 unsigned int cmd, int len, void *data)
5506 struct lu_context ioctl_session;
5507 struct md_device *next = mdt->mdt_child;
5511 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5514 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5515 lu_context_enter(&ioctl_session);
5516 env->le_ses = &ioctl_session;
5518 LASSERT(next->md_ops->mdo_iocontrol);
5519 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5521 lu_context_exit(&ioctl_session);
5522 lu_context_fini(&ioctl_session);
5526 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5528 struct obd_ioctl_data *data = karg;
5531 struct mdt_object *obj;
5532 struct mdt_lock_handle *lh;
5536 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5537 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5540 fid = (struct lu_fid *)data->ioc_inlbuf1;
5542 if (!fid_is_sane(fid))
5545 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5547 lh = &mti->mti_lh[MDT_LH_PARENT];
5548 mdt_lock_reg_init(lh, LCK_CR);
5550 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5552 RETURN(PTR_ERR(obj));
5554 if (mdt_object_remote(obj)) {
5557 * before calling version get the correct MDS should be
5558 * fid, this is error to find remote object here
5560 CERROR("nonlocal object "DFID"\n", PFID(fid));
5561 } else if (!mdt_object_exists(obj)) {
5562 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5565 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5566 *(__u64 *)data->ioc_inlbuf2 = version;
5569 mdt_object_unlock_put(mti, obj, lh, 1);
5573 /* ioctls on obd dev */
5574 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5575 void *karg, void *uarg)
5578 struct obd_device *obd = exp->exp_obd;
5579 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5580 struct dt_device *dt = mdt->mdt_bottom;
5584 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5585 rc = lu_env_init(&env, LCT_MD_THREAD);
5591 rc = mdt_device_sync(&env, mdt);
5593 case OBD_IOC_SET_READONLY:
5594 rc = dt->dd_ops->dt_ro(&env, dt);
5596 case OBD_IOC_ABORT_RECOVERY:
5597 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5598 obd->obd_force_abort_recovery = 1;
5599 target_stop_recovery_thread(obd);
5602 case OBD_IOC_CHANGELOG_REG:
5603 case OBD_IOC_CHANGELOG_DEREG:
5604 case OBD_IOC_CHANGELOG_CLEAR:
5605 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5607 case OBD_IOC_START_LFSCK: {
5608 struct md_device *next = mdt->mdt_child;
5609 struct obd_ioctl_data *data = karg;
5610 struct lfsck_start_param lsp;
5612 if (unlikely(data == NULL)) {
5617 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5618 lsp.lsp_index_valid = 0;
5619 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5622 case OBD_IOC_STOP_LFSCK: {
5623 struct md_device *next = mdt->mdt_child;
5624 struct obd_ioctl_data *data = karg;
5625 struct lfsck_stop stop;
5627 stop.ls_status = LS_STOPPED;
5628 /* Old lfsck utils may pass NULL @stop. */
5629 if (data->ioc_inlbuf1 == NULL)
5633 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5635 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5638 case OBD_IOC_GET_OBJ_VERSION: {
5639 struct mdt_thread_info *mti;
5640 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5641 memset(mti, 0, sizeof *mti);
5642 mti->mti_env = &env;
5646 rc = mdt_ioc_version_get(mti, karg);
5649 case OBD_IOC_CATLOGLIST: {
5650 struct mdt_thread_info *mti;
5652 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5653 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5654 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5655 &mti->mti_tmp_fid1);
5660 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5661 mdt_obd_name(mdt), cmd, rc);
5668 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5670 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5671 struct lfsck_start_param lsp;
5675 lsp.lsp_start = NULL;
5676 lsp.lsp_index_valid = 0;
5677 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
5678 OBD_IOC_START_LFSCK,
5680 if (rc != 0 && rc != -EALREADY)
5681 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
5682 mdt_obd_name(mdt), rc);
5684 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5688 static int mdt_obd_postrecov(struct obd_device *obd)
5693 rc = lu_env_init(&env, LCT_MD_THREAD);
5696 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5701 static struct obd_ops mdt_obd_device_ops = {
5702 .o_owner = THIS_MODULE,
5703 .o_set_info_async = mdt_obd_set_info_async,
5704 .o_connect = mdt_obd_connect,
5705 .o_reconnect = mdt_obd_reconnect,
5706 .o_disconnect = mdt_obd_disconnect,
5707 .o_init_export = mdt_init_export,
5708 .o_destroy_export = mdt_destroy_export,
5709 .o_iocontrol = mdt_iocontrol,
5710 .o_postrecov = mdt_obd_postrecov,
5713 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5714 struct lu_device *d)
5716 struct mdt_device *m = mdt_dev(d);
5723 static struct lu_device *mdt_device_free(const struct lu_env *env,
5724 struct lu_device *d)
5726 struct mdt_device *m = mdt_dev(d);
5729 lu_device_fini(&m->mdt_lu_dev);
5735 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5736 struct lu_device_type *t,
5737 struct lustre_cfg *cfg)
5739 struct lu_device *l;
5740 struct mdt_device *m;
5747 rc = mdt_init0(env, m, t, cfg);
5749 mdt_device_free(env, l);
5754 l = ERR_PTR(-ENOMEM);
5758 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5759 LU_KEY_INIT(mdt, struct mdt_thread_info);
5761 static void mdt_key_fini(const struct lu_context *ctx,
5762 struct lu_context_key *key, void* data)
5764 struct mdt_thread_info *info = data;
5766 if (info->mti_big_lmm) {
5767 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5768 info->mti_big_lmm = NULL;
5769 info->mti_big_lmmsize = 0;
5774 /* context key: mdt_thread_key */
5775 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5777 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5779 return lu_ucred(info->mti_env);
5782 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5784 return lu_ucred_check(info->mti_env);
5788 * Enable/disable COS (Commit On Sharing).
5790 * Set/Clear the COS flag in mdt options.
5792 * \param mdt mdt device
5793 * \param val 0 disables COS, other values enable COS
5795 void mdt_enable_cos(struct mdt_device *mdt, int val)
5800 mdt->mdt_opts.mo_cos = !!val;
5801 rc = lu_env_init(&env, LCT_LOCAL);
5802 if (unlikely(rc != 0)) {
5803 CWARN("%s: lu_env initialization failed, cannot "
5804 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
5807 mdt_device_sync(&env, mdt);
5812 * Check COS (Commit On Sharing) status.
5814 * Return COS flag status.
5816 * \param mdt mdt device
5818 int mdt_cos_is_enabled(struct mdt_device *mdt)
5820 return mdt->mdt_opts.mo_cos != 0;
5823 static struct lu_device_type_operations mdt_device_type_ops = {
5824 .ldto_device_alloc = mdt_device_alloc,
5825 .ldto_device_free = mdt_device_free,
5826 .ldto_device_fini = mdt_device_fini
5829 static struct lu_device_type mdt_device_type = {
5830 .ldt_tags = LU_DEVICE_MD,
5831 .ldt_name = LUSTRE_MDT_NAME,
5832 .ldt_ops = &mdt_device_type_ops,
5833 .ldt_ctx_tags = LCT_MD_THREAD
5836 static int __init mdt_mod_init(void)
5840 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5841 FID_NOBRACE_LEN + 1);
5842 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5844 rc = lu_kmem_init(mdt_caches);
5848 rc = mds_mod_init();
5852 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
5853 LUSTRE_MDT_NAME, &mdt_device_type);
5858 lu_kmem_fini(mdt_caches);
5865 static void __exit mdt_mod_exit(void)
5867 class_unregister_type(LUSTRE_MDT_NAME);
5869 lu_kmem_fini(mdt_caches);
5872 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
5873 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5874 MODULE_VERSION(LUSTRE_VERSION_STRING);
5875 MODULE_LICENSE("GPL");
5877 module_init(mdt_mod_init);
5878 module_exit(mdt_mod_exit);