4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_handler.c
38 * Lustre Metadata Target (mdt) request handler
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
43 * Author: Mike Shaver <shaver@clusterfs.com>
44 * Author: Nikita Danilov <nikita@clusterfs.com>
45 * Author: Huang Hua <huanghua@clusterfs.com>
46 * Author: Yury Umanets <umka@clusterfs.com>
49 #define DEBUG_SUBSYSTEM S_MDS
51 #include <linux/module.h>
53 * struct OBD_{ALLOC,FREE}*()
55 #include <obd_support.h>
56 #include <lustre_ioctl.h>
57 /* struct ptlrpc_request */
58 #include <lustre_net.h>
59 /* struct obd_export */
60 #include <lustre_export.h>
61 /* struct obd_device */
64 #include <dt_object.h>
65 #include <lustre_mds.h>
66 #include <lustre_log.h>
67 #include "mdt_internal.h"
68 #include <lustre_acl.h>
69 #include <lustre_param.h>
70 #include <lustre_quota.h>
71 #include <lustre_lfsck.h>
72 #include <lustre_nodemap.h>
74 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 ldlm_mode_t 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, ldlm_mode_t 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, ldlm_mode_t 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) && (ma->ma_hsm.mh_flags & HS_RELEASED) &&
901 mdt_hsm_restore_is_running(info, mdt_object_fid(o))) {
902 repbody->mbo_t_state = MS_RESTORE;
903 repbody->mbo_valid |= OBD_MD_TSTATE;
906 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
908 /* the Lustre protocol supposes to return default striping
909 * on the user-visible root if explicitly requested */
910 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
911 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
912 struct lu_fid rootfid;
913 struct mdt_object *root;
914 struct mdt_device *mdt = info->mti_mdt;
916 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
919 root = mdt_object_find(env, mdt, &rootfid);
921 RETURN(PTR_ERR(root));
922 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
923 mdt_object_put(info->mti_env, root);
925 CERROR("%s: getattr error for "DFID": rc = %d\n",
926 mdt_obd_name(info->mti_mdt),
927 PFID(mdt_object_fid(o)), rc);
932 if (likely(ma->ma_valid & MA_INODE))
933 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
937 if (mdt_body_has_lov(la, reqbody)) {
938 if (ma->ma_valid & MA_LOV) {
939 LASSERT(ma->ma_lmm_size);
940 repbody->mbo_eadatasize = ma->ma_lmm_size;
941 if (S_ISDIR(la->la_mode))
942 repbody->mbo_valid |= OBD_MD_FLDIREA;
944 repbody->mbo_valid |= OBD_MD_FLEASIZE;
945 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
947 if (ma->ma_valid & MA_LMV) {
948 /* Return -ENOTSUPP for old client */
949 if (!mdt_is_striped_client(req->rq_export))
952 LASSERT(S_ISDIR(la->la_mode));
953 mdt_dump_lmv(D_INFO, ma->ma_lmv);
954 repbody->mbo_eadatasize = ma->ma_lmv_size;
955 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
957 if (ma->ma_valid & MA_LMV_DEF) {
958 /* Return -ENOTSUPP for old client */
959 if (!mdt_is_striped_client(req->rq_export))
961 LASSERT(S_ISDIR(la->la_mode));
962 repbody->mbo_eadatasize = ma->ma_lmv_size;
963 repbody->mbo_valid |= (OBD_MD_FLDIREA |
966 } else if (S_ISLNK(la->la_mode) &&
967 reqbody->mbo_valid & OBD_MD_LINKNAME) {
968 buffer->lb_buf = ma->ma_lmm;
969 /* eadatasize from client includes NULL-terminator, so
970 * there is no need to read it */
971 buffer->lb_len = reqbody->mbo_eadatasize - 1;
972 rc = mo_readlink(env, next, buffer);
973 if (unlikely(rc <= 0)) {
974 CERROR("%s: readlink failed for "DFID": rc = %d\n",
975 mdt_obd_name(info->mti_mdt),
976 PFID(mdt_object_fid(o)), rc);
979 int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
981 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
983 repbody->mbo_valid |= OBD_MD_LINKNAME;
984 /* we need to report back size with NULL-terminator
985 * because client expects that */
986 repbody->mbo_eadatasize = rc + 1;
987 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
988 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
989 "on "DFID ", expected %d\n",
990 mdt_obd_name(info->mti_mdt),
991 rc, PFID(mdt_object_fid(o)),
992 reqbody->mbo_eadatasize - 1);
994 ((char *)ma->ma_lmm)[rc] = 0;
996 /* If the total CDEBUG() size is larger than a page, it
997 * will print a warning to the console, avoid this by
998 * printing just the last part of the symlink. */
999 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1000 print_limit < rc ? "..." : "", print_limit,
1001 (char *)ma->ma_lmm + rc - print_limit, rc);
1006 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1007 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1008 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1009 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1010 repbody->mbo_max_mdsize);
1013 if (exp_connect_rmtclient(info->mti_exp) &&
1014 reqbody->mbo_valid & OBD_MD_FLRMTPERM) {
1015 void *buf = req_capsule_server_get(pill, &RMF_ACL);
1017 /* mdt_getattr_lock only */
1018 rc = mdt_pack_remote_perm(info, o, buf);
1020 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
1021 repbody->mbo_aclsize = 0;
1024 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
1025 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
1028 #ifdef CONFIG_FS_POSIX_ACL
1029 else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1030 (reqbody->mbo_valid & OBD_MD_FLACL))
1031 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1036 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1041 static int mdt_getattr(struct tgt_session_info *tsi)
1043 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1044 struct mdt_object *obj = info->mti_object;
1045 struct req_capsule *pill = info->mti_pill;
1046 struct mdt_body *reqbody;
1047 struct mdt_body *repbody;
1051 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1053 LASSERT(obj != NULL);
1054 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1056 /* Unlike intent case where we need to pre-fill out buffers early on
1057 * in intent policy for ldlm reasons, here we can have a much better
1058 * guess at EA size by just reading it from disk.
1059 * Exceptions are readdir and (missing) directory striping */
1061 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1062 /* No easy way to know how long is the symlink, but it cannot
1063 * be more than PATH_MAX, so we allocate +1 */
1066 /* A special case for fs ROOT: getattr there might fetch
1067 * default EA for entire fs, not just for this dir!
1069 } else if (lu_fid_eq(mdt_object_fid(obj),
1070 &info->mti_mdt->mdt_md_root_fid) &&
1071 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1072 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1074 /* Should the default strping be bigger, mdt_fix_reply
1075 * will reallocate */
1076 rc = DEF_REP_MD_SIZE;
1078 /* Read the actual EA size from disk */
1079 rc = mdt_attr_get_eabuf_size(info, obj);
1083 GOTO(out, rc = err_serious(rc));
1085 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1087 rc = req_capsule_server_pack(pill);
1088 if (unlikely(rc != 0))
1089 GOTO(out, rc = err_serious(rc));
1091 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1092 LASSERT(repbody != NULL);
1093 repbody->mbo_eadatasize = 0;
1094 repbody->mbo_aclsize = 0;
1096 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1097 rc = mdt_init_ucred(info, reqbody);
1099 rc = mdt_check_ucred(info);
1101 GOTO(out_shrink, rc);
1103 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1105 rc = mdt_getattr_internal(info, obj, 0);
1106 if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1107 mdt_exit_ucred(info);
1110 mdt_client_compatibility(info);
1111 rc2 = mdt_fix_reply(info);
1115 mdt_thread_info_fini(info);
1120 * Exchange MOF_LOV_CREATED flags between two objects after a
1121 * layout swap. No assumption is made on whether o1 or o2 have
1122 * created objects or not.
1124 * \param[in,out] o1 First swap layout object
1125 * \param[in,out] o2 Second swap layout object
1127 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1131 mutex_lock(&o1->mot_lov_mutex);
1132 mutex_lock(&o2->mot_lov_mutex);
1134 o1_flags = o1->mot_flags;
1135 o1->mot_flags = (o1->mot_flags & ~MOF_LOV_CREATED) |
1136 (o2->mot_flags & MOF_LOV_CREATED);
1138 o2->mot_flags = (o2->mot_flags & ~MOF_LOV_CREATED) |
1139 (o1_flags & MOF_LOV_CREATED);
1141 mutex_unlock(&o2->mot_lov_mutex);
1142 mutex_unlock(&o1->mot_lov_mutex);
1145 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1147 struct mdt_thread_info *info;
1148 struct ptlrpc_request *req = tgt_ses_req(tsi);
1149 struct obd_export *exp = req->rq_export;
1150 struct mdt_object *o1, *o2, *o;
1151 struct mdt_lock_handle *lh1, *lh2;
1152 struct mdc_swap_layouts *msl;
1156 /* client does not support layout lock, so layout swaping
1158 * FIXME: there is a problem for old clients which don't support
1159 * layout lock yet. If those clients have already opened the file
1160 * they won't be notified at all so that old layout may still be
1161 * used to do IO. This can be fixed after file release is landed by
1162 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1163 if (!exp_connect_layout(exp))
1164 RETURN(-EOPNOTSUPP);
1166 info = tsi2mdt_info(tsi);
1168 if (info->mti_dlm_req != NULL)
1169 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1171 o1 = info->mti_object;
1172 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1173 &info->mti_body->mbo_fid2);
1175 GOTO(out, rc = PTR_ERR(o));
1177 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1178 GOTO(put, rc = -ENOENT);
1180 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1181 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1187 /* permission check. Make sure the calling process having permission
1188 * to write both files. */
1189 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1194 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1199 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1201 GOTO(put, rc = -EPROTO);
1203 lh1 = &info->mti_lh[MDT_LH_NEW];
1204 mdt_lock_reg_init(lh1, LCK_EX);
1205 lh2 = &info->mti_lh[MDT_LH_OLD];
1206 mdt_lock_reg_init(lh2, LCK_EX);
1208 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1209 MDS_INODELOCK_XATTR);
1213 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1214 MDS_INODELOCK_XATTR);
1218 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1219 mdt_object_child(o2), msl->msl_flags);
1223 mdt_swap_lov_flag(o1, o2);
1226 mdt_object_unlock(info, o2, lh2, rc);
1228 mdt_object_unlock(info, o1, lh1, rc);
1230 mdt_object_put(info->mti_env, o);
1232 mdt_thread_info_fini(info);
1236 static int mdt_raw_lookup(struct mdt_thread_info *info,
1237 struct mdt_object *parent,
1238 const struct lu_name *lname,
1239 struct ldlm_reply *ldlm_rep)
1241 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1245 LASSERT(!info->mti_cross_ref);
1247 /* Only got the fid of this obj by name */
1248 fid_zero(child_fid);
1249 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1250 lname, child_fid, &info->mti_spec);
1252 struct mdt_body *repbody;
1254 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1255 repbody->mbo_fid1 = *child_fid;
1256 repbody->mbo_valid = OBD_MD_FLID;
1257 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1258 } else if (rc == -ENOENT) {
1259 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1266 * UPDATE lock should be taken against parent, and be released before exit;
1267 * child_bits lock should be taken against child, and be returned back:
1268 * (1)normal request should release the child lock;
1269 * (2)intent request will grant the lock to client.
1271 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1272 struct mdt_lock_handle *lhc,
1274 struct ldlm_reply *ldlm_rep)
1276 struct ptlrpc_request *req = mdt_info_req(info);
1277 struct mdt_body *reqbody = NULL;
1278 struct mdt_object *parent = info->mti_object;
1279 struct mdt_object *child;
1280 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1281 struct lu_name *lname = NULL;
1282 struct mdt_lock_handle *lhp = NULL;
1283 struct ldlm_lock *lock;
1290 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1291 LASSERT(ergo(is_resent,
1292 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1297 if (info->mti_cross_ref) {
1298 /* Only getattr on the child. Parent is on another node. */
1299 mdt_set_disposition(info, ldlm_rep,
1300 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1302 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1304 PFID(mdt_object_fid(child)), ldlm_rep);
1306 rc = mdt_check_resent_lock(info, child, lhc);
1309 } else if (rc > 0) {
1310 mdt_lock_handle_init(lhc);
1311 mdt_lock_reg_init(lhc, LCK_PR);
1314 * Object's name is on another MDS, no lookup or layout
1315 * lock is needed here but update lock is.
1317 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1318 MDS_INODELOCK_LAYOUT);
1319 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1321 rc = mdt_object_lock(info, child, lhc, child_bits);
1326 /* Finally, we can get attr for child. */
1327 if (!mdt_object_exists(child)) {
1328 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1330 "remote object doesn't exist.\n");
1331 mdt_object_unlock(info, child, lhc, 1);
1335 rc = mdt_getattr_internal(info, child, 0);
1336 if (unlikely(rc != 0))
1337 mdt_object_unlock(info, child, lhc, 1);
1342 lname = &info->mti_name;
1343 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1345 if (lu_name_is_valid(lname)) {
1346 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1347 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1348 PNAME(lname), ldlm_rep);
1350 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1351 if (unlikely(reqbody == NULL))
1352 RETURN(err_serious(-EPROTO));
1354 *child_fid = reqbody->mbo_fid2;
1356 if (unlikely(!fid_is_sane(child_fid)))
1357 RETURN(err_serious(-EINVAL));
1359 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1361 PFID(mdt_object_fid(parent)),
1362 PFID(&reqbody->mbo_fid2), ldlm_rep);
1365 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1367 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1368 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1370 "Parent doesn't exist!\n");
1374 if (mdt_object_remote(parent)) {
1375 CERROR("%s: parent "DFID" is on remote target\n",
1376 mdt_obd_name(info->mti_mdt),
1377 PFID(mdt_object_fid(parent)));
1381 if (lu_name_is_valid(lname)) {
1382 /* Always allow to lookup ".." */
1383 if (unlikely(lname->ln_namelen == 2 &&
1384 lname->ln_name[0] == '.' &&
1385 lname->ln_name[1] == '.'))
1386 info->mti_spec.sp_permitted = 1;
1388 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1389 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1394 /* step 1: lock parent only if parent is a directory */
1395 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1396 lhp = &info->mti_lh[MDT_LH_PARENT];
1397 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1398 rc = mdt_object_lock(info, parent, lhp,
1399 MDS_INODELOCK_UPDATE);
1400 if (unlikely(rc != 0))
1404 /* step 2: lookup child's fid by name */
1405 fid_zero(child_fid);
1406 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1407 child_fid, &info->mti_spec);
1409 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1412 GOTO(out_parent, rc);
1415 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1418 *step 3: find the child object by fid & lock it.
1419 * regardless if it is local or remote.
1421 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1422 * set parent dir fid the same as child fid in getattr by fid case
1423 * we should not lu_object_find() the object again, could lead
1424 * to hung if there is a concurrent unlink destroyed the object.
1426 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1427 mdt_object_get(info->mti_env, parent);
1430 child = mdt_object_find(info->mti_env, info->mti_mdt,
1434 if (unlikely(IS_ERR(child)))
1435 GOTO(out_parent, rc = PTR_ERR(child));
1437 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1438 rc = mdt_check_resent_lock(info, child, lhc);
1440 GOTO(out_child, rc);
1441 } else if (rc > 0) {
1442 mdt_lock_handle_init(lhc);
1443 mdt_lock_reg_init(lhc, LCK_PR);
1446 if (!mdt_object_exists(child)) {
1447 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1449 "Object doesn't exist!\n");
1450 GOTO(out_child, rc = -ENOENT);
1453 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1454 mdt_object_exists(child) && !mdt_object_remote(child)) {
1455 struct md_attr *ma = &info->mti_attr;
1458 ma->ma_need = MA_INODE;
1459 rc = mdt_attr_get_complex(info, child, ma);
1460 if (unlikely(rc != 0))
1461 GOTO(out_child, rc);
1463 /* If the file has not been changed for some time, we
1464 * return not only a LOOKUP lock, but also an UPDATE
1465 * lock and this might save us RPC on later STAT. For
1466 * directories, it also let negative dentry cache start
1467 * working for this dir. */
1468 if (ma->ma_valid & MA_INODE &&
1469 ma->ma_attr.la_valid & LA_CTIME &&
1470 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1471 ma->ma_attr.la_ctime < cfs_time_current_sec())
1472 child_bits |= MDS_INODELOCK_UPDATE;
1475 /* layout lock must be granted in a best-effort way
1476 * for IT operations */
1477 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1478 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1479 exp_connect_layout(info->mti_exp) &&
1480 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1481 !mdt_object_remote(child) && ldlm_rep != NULL) {
1482 /* try to grant layout lock for regular file. */
1488 child_bits |= MDS_INODELOCK_LAYOUT;
1489 /* try layout lock, it may fail to be granted due to
1490 * contention at LOOKUP or UPDATE */
1491 if (!mdt_object_lock_try(info, child, lhc,
1493 child_bits &= ~MDS_INODELOCK_LAYOUT;
1494 LASSERT(child_bits != 0);
1495 rc = mdt_object_lock(info, child, lhc,
1501 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1502 * client will enqueue the lock to the remote MDT */
1503 if (mdt_object_remote(child))
1504 child_bits &= ~MDS_INODELOCK_UPDATE;
1505 rc = mdt_object_lock(info, child, lhc, child_bits);
1507 if (unlikely(rc != 0))
1508 GOTO(out_child, rc);
1511 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1513 /* finally, we can get attr for child. */
1514 rc = mdt_getattr_internal(info, child, ma_need);
1515 if (unlikely(rc != 0)) {
1516 mdt_object_unlock(info, child, lhc, 1);
1518 /* Debugging code. */
1519 LDLM_DEBUG(lock, "Returning lock to client");
1520 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1521 &lock->l_resource->lr_name),
1522 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1523 PLDLMRES(lock->l_resource),
1524 PFID(mdt_object_fid(child)));
1527 LDLM_LOCK_PUT(lock);
1531 mdt_object_put(info->mti_env, child);
1534 mdt_object_unlock(info, parent, lhp, 1);
1538 /* normal handler: should release the child lock */
1539 static int mdt_getattr_name(struct tgt_session_info *tsi)
1541 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1542 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1543 struct mdt_body *reqbody;
1544 struct mdt_body *repbody;
1548 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1549 LASSERT(reqbody != NULL);
1550 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1551 LASSERT(repbody != NULL);
1553 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1554 repbody->mbo_eadatasize = 0;
1555 repbody->mbo_aclsize = 0;
1557 rc = mdt_init_ucred(info, reqbody);
1559 GOTO(out_shrink, rc);
1561 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1562 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1563 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1564 lhc->mlh_reg_lh.cookie = 0;
1566 mdt_exit_ucred(info);
1569 mdt_client_compatibility(info);
1570 rc2 = mdt_fix_reply(info);
1573 mdt_thread_info_fini(info);
1577 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1578 void *karg, void *uarg);
1580 static int mdt_set_info(struct tgt_session_info *tsi)
1582 struct ptlrpc_request *req = tgt_ses_req(tsi);
1585 int keylen, vallen, rc = 0;
1589 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1591 DEBUG_REQ(D_HA, req, "no set_info key");
1592 RETURN(err_serious(-EFAULT));
1595 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1598 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1600 DEBUG_REQ(D_HA, req, "no set_info val");
1601 RETURN(err_serious(-EFAULT));
1604 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1607 /* Swab any part of val you need to here */
1608 if (KEY_IS(KEY_READ_ONLY)) {
1609 spin_lock(&req->rq_export->exp_lock);
1611 *exp_connect_flags_ptr(req->rq_export) |=
1614 *exp_connect_flags_ptr(req->rq_export) &=
1615 ~OBD_CONNECT_RDONLY;
1616 spin_unlock(&req->rq_export->exp_lock);
1617 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1618 struct changelog_setinfo *cs = val;
1620 if (vallen != sizeof(*cs)) {
1621 CERROR("%s: bad changelog_clear setinfo size %d\n",
1622 tgt_name(tsi->tsi_tgt), vallen);
1625 if (ptlrpc_req_need_swab(req)) {
1626 __swab64s(&cs->cs_recno);
1627 __swab32s(&cs->cs_id);
1630 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1638 static int mdt_readpage(struct tgt_session_info *tsi)
1640 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1641 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1642 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1643 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1644 struct mdt_body *repbody;
1650 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1651 RETURN(err_serious(-ENOMEM));
1653 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1654 if (repbody == NULL || reqbody == NULL)
1655 RETURN(err_serious(-EFAULT));
1658 * prepare @rdpg before calling lower layers and transfer itself. Here
1659 * reqbody->size contains offset of where to start to read and
1660 * reqbody->nlink contains number bytes to read.
1662 rdpg->rp_hash = reqbody->mbo_size;
1663 if (rdpg->rp_hash != reqbody->mbo_size) {
1664 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1665 rdpg->rp_hash, reqbody->mbo_size);
1669 rdpg->rp_attrs = reqbody->mbo_mode;
1670 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1671 rdpg->rp_attrs |= LUDA_64BITHASH;
1672 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
1673 exp_max_brw_size(tsi->tsi_exp));
1674 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1676 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1677 if (rdpg->rp_pages == NULL)
1680 for (i = 0; i < rdpg->rp_npages; ++i) {
1681 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1682 if (rdpg->rp_pages[i] == NULL)
1683 GOTO(free_rdpg, rc = -ENOMEM);
1686 /* call lower layers to fill allocated pages with directory data */
1687 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1689 GOTO(free_rdpg, rc);
1691 /* send pages to client */
1692 rc = tgt_sendpage(tsi, rdpg, rc);
1697 for (i = 0; i < rdpg->rp_npages; i++)
1698 if (rdpg->rp_pages[i] != NULL)
1699 __free_page(rdpg->rp_pages[i]);
1700 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1702 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1708 static int mdt_reint_internal(struct mdt_thread_info *info,
1709 struct mdt_lock_handle *lhc,
1712 struct req_capsule *pill = info->mti_pill;
1713 struct mdt_body *repbody;
1718 rc = mdt_reint_unpack(info, op);
1720 CERROR("Can't unpack reint, rc %d\n", rc);
1721 RETURN(err_serious(rc));
1724 /* for replay (no_create) lmm is not needed, client has it already */
1725 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1726 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1729 /* llog cookies are always 0, the field is kept for compatibility */
1730 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1731 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1733 rc = req_capsule_server_pack(pill);
1735 CERROR("Can't pack response, rc %d\n", rc);
1736 RETURN(err_serious(rc));
1739 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1740 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1742 repbody->mbo_eadatasize = 0;
1743 repbody->mbo_aclsize = 0;
1746 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1748 /* for replay no cookkie / lmm need, because client have this already */
1749 if (info->mti_spec.no_create)
1750 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1751 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1753 rc = mdt_init_ucred_reint(info);
1755 GOTO(out_shrink, rc);
1757 rc = mdt_fix_attr_ucred(info, op);
1759 GOTO(out_ucred, rc = err_serious(rc));
1761 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
1763 GOTO(out_ucred, rc);
1764 } else if (rc == 1) {
1765 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1766 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1767 GOTO(out_ucred, rc);
1769 rc = mdt_reint_rec(info, lhc);
1772 mdt_exit_ucred(info);
1774 mdt_client_compatibility(info);
1775 rc2 = mdt_fix_reply(info);
1781 static long mdt_reint_opcode(struct ptlrpc_request *req,
1782 const struct req_format **fmt)
1784 struct mdt_device *mdt;
1785 struct mdt_rec_reint *rec;
1788 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1790 opc = rec->rr_opcode;
1791 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1792 if (opc < REINT_MAX && fmt[opc] != NULL)
1793 req_capsule_extend(&req->rq_pill, fmt[opc]);
1795 mdt = mdt_exp2dev(req->rq_export);
1796 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1797 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1798 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1799 opc = err_serious(-EFAULT);
1802 opc = err_serious(-EFAULT);
1807 static int mdt_reint(struct tgt_session_info *tsi)
1811 static const struct req_format *reint_fmts[REINT_MAX] = {
1812 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1813 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1814 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1815 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1816 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1817 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1818 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1819 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
1820 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME
1825 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1827 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1829 * No lock possible here from client to pass it to reint code
1832 rc = mdt_reint_internal(info, NULL, opc);
1833 mdt_thread_info_fini(info);
1838 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1842 /* this should sync the whole device */
1843 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1845 struct dt_device *dt = mdt->mdt_bottom;
1849 rc = dt->dd_ops->dt_sync(env, dt);
1853 /* this should sync this object */
1854 static int mdt_object_sync(struct mdt_thread_info *info)
1856 struct md_object *next;
1860 if (!mdt_object_exists(info->mti_object)) {
1861 CWARN("%s: non existing object "DFID": rc = %d\n",
1862 mdt_obd_name(info->mti_mdt),
1863 PFID(mdt_object_fid(info->mti_object)), -ESTALE);
1866 next = mdt_object_child(info->mti_object);
1867 rc = mo_object_sync(info->mti_env, next);
1872 static int mdt_sync(struct tgt_session_info *tsi)
1874 struct ptlrpc_request *req = tgt_ses_req(tsi);
1875 struct req_capsule *pill = tsi->tsi_pill;
1876 struct mdt_body *body;
1881 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1882 RETURN(err_serious(-ENOMEM));
1884 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
1885 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
1887 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1889 /* sync an object */
1890 rc = mdt_object_sync(info);
1892 const struct lu_fid *fid;
1893 struct lu_attr *la = &info->mti_attr.ma_attr;
1895 info->mti_attr.ma_need = MA_INODE;
1896 info->mti_attr.ma_valid = 0;
1897 rc = mdt_attr_get_complex(info, info->mti_object,
1900 body = req_capsule_server_get(pill,
1902 fid = mdt_object_fid(info->mti_object);
1903 mdt_pack_attr2body(info, body, la, fid);
1906 mdt_thread_info_fini(info);
1909 mdt_counter_incr(req, LPROC_MDT_SYNC);
1915 * Handle quota control requests to consult current usage/limit, but also
1916 * to configure quota enforcement
1918 static int mdt_quotactl(struct tgt_session_info *tsi)
1920 struct obd_export *exp = tsi->tsi_exp;
1921 struct req_capsule *pill = tsi->tsi_pill;
1922 struct obd_quotactl *oqctl, *repoqc;
1924 struct mdt_device *mdt = mdt_exp2dev(exp);
1925 struct lu_device *qmt = mdt->mdt_qmt_dev;
1926 struct lu_nodemap *nodemap = exp->exp_target_data.ted_nodemap;
1929 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1931 RETURN(err_serious(-EPROTO));
1933 rc = req_capsule_server_pack(pill);
1935 RETURN(err_serious(rc));
1937 switch (oqctl->qc_cmd) {
1938 /* master quotactl */
1941 if (!nodemap_can_setquota(nodemap))
1946 RETURN(-EOPNOTSUPP);
1947 /* slave quotactl */
1952 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1956 /* map uid/gid for remote client */
1958 if (exp_connect_rmtclient(exp)) {
1959 struct lustre_idmap_table *idmap;
1961 idmap = exp->exp_mdt_data.med_idmap;
1963 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1964 oqctl->qc_cmd != Q_GETINFO))
1967 if (oqctl->qc_type == USRQUOTA)
1968 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1970 else if (oqctl->qc_type == GRPQUOTA)
1971 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1976 if (id == CFS_IDMAP_NOTFOUND) {
1977 CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
1982 if (oqctl->qc_type == USRQUOTA)
1983 id = nodemap_map_id(nodemap, NODEMAP_UID,
1984 NODEMAP_CLIENT_TO_FS, id);
1985 else if (oqctl->qc_type == GRPQUOTA)
1986 id = nodemap_map_id(nodemap, NODEMAP_UID,
1987 NODEMAP_CLIENT_TO_FS, id);
1989 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1991 RETURN(err_serious(-EFAULT));
1993 if (oqctl->qc_id != id)
1994 swap(oqctl->qc_id, id);
1996 switch (oqctl->qc_cmd) {
2002 /* forward quotactl request to QMT */
2003 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2008 /* slave quotactl */
2009 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2014 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2018 if (oqctl->qc_id != id)
2019 swap(oqctl->qc_id, id);
2025 /** clone llog ctxt from child (mdd)
2026 * This allows remote llog (replicator) access.
2027 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2028 * context was originally set up, or we can handle them directly.
2029 * I choose the latter, but that means I need any llog
2030 * contexts set up by child to be accessable by the mdt. So we clone the
2031 * context into our context list here.
2033 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2036 struct md_device *next = mdt->mdt_child;
2037 struct llog_ctxt *ctxt;
2040 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2043 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2044 if (rc || ctxt == NULL) {
2048 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2050 CERROR("Can't set mdt ctxt %d\n", rc);
2055 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2056 struct mdt_device *mdt, int idx)
2058 struct llog_ctxt *ctxt;
2060 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2063 /* Put once for the get we just did, and once for the clone */
2064 llog_ctxt_put(ctxt);
2065 llog_ctxt_put(ctxt);
2070 * sec context handlers
2072 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2076 rc = mdt_handle_idmap(tsi);
2078 struct ptlrpc_request *req = tgt_ses_req(tsi);
2081 opc = lustre_msg_get_opc(req->rq_reqmsg);
2082 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2083 sptlrpc_svc_ctx_invalidate(req);
2086 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2092 * quota request handlers
2094 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2096 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2097 struct lu_device *qmt = mdt->mdt_qmt_dev;
2102 RETURN(err_serious(-EOPNOTSUPP));
2104 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2108 struct mdt_object *mdt_object_new(const struct lu_env *env,
2109 struct mdt_device *d,
2110 const struct lu_fid *f)
2112 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2113 struct lu_object *o;
2114 struct mdt_object *m;
2117 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2118 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2119 if (unlikely(IS_ERR(o)))
2120 m = (struct mdt_object *)o;
2126 struct mdt_object *mdt_object_find(const struct lu_env *env,
2127 struct mdt_device *d,
2128 const struct lu_fid *f)
2130 struct lu_object *o;
2131 struct mdt_object *m;
2134 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2135 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2136 if (unlikely(IS_ERR(o)))
2137 m = (struct mdt_object *)o;
2145 * Asyncronous commit for mdt device.
2147 * Pass asynchonous commit call down the MDS stack.
2149 * \param env environment
2150 * \param mdt the mdt device
2152 static void mdt_device_commit_async(const struct lu_env *env,
2153 struct mdt_device *mdt)
2155 struct dt_device *dt = mdt->mdt_bottom;
2158 rc = dt->dd_ops->dt_commit_async(env, dt);
2159 if (unlikely(rc != 0))
2160 CWARN("%s: async commit start failed: rc = %d\n",
2161 mdt_obd_name(mdt), rc);
2165 * Mark the lock as "synchonous".
2167 * Mark the lock to deffer transaction commit to the unlock time.
2169 * \param lock the lock to mark as "synchonous"
2171 * \see mdt_is_lock_sync
2172 * \see mdt_save_lock
2174 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2176 lock->l_ast_data = (void*)1;
2180 * Check whehter the lock "synchonous" or not.
2182 * \param lock the lock to check
2183 * \retval 1 the lock is "synchonous"
2184 * \retval 0 the lock isn't "synchronous"
2186 * \see mdt_set_lock_sync
2187 * \see mdt_save_lock
2189 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2191 return lock->l_ast_data != NULL;
2195 * Blocking AST for mdt locks.
2197 * Starts transaction commit if in case of COS lock conflict or
2198 * deffers such a commit to the mdt_save_lock.
2200 * \param lock the lock which blocks a request or cancelling lock
2201 * \param desc unused
2202 * \param data unused
2203 * \param flag indicates whether this cancelling or blocking callback
2205 * \see ldlm_blocking_ast_nocheck
2207 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2208 void *data, int flag)
2210 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2211 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2215 if (flag == LDLM_CB_CANCELING)
2217 lock_res_and_lock(lock);
2218 if (lock->l_blocking_ast != mdt_blocking_ast) {
2219 unlock_res_and_lock(lock);
2222 if (mdt_cos_is_enabled(mdt) &&
2223 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2224 lock->l_blocking_lock != NULL &&
2225 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2226 mdt_set_lock_sync(lock);
2228 rc = ldlm_blocking_ast_nocheck(lock);
2230 /* There is no lock conflict if l_blocking_lock == NULL,
2231 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2232 * when the last reference to a local lock was released */
2233 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2236 rc = lu_env_init(&env, LCT_LOCAL);
2237 if (unlikely(rc != 0))
2238 CWARN("%s: lu_env initialization failed, cannot "
2239 "start asynchronous commit: rc = %d\n",
2242 mdt_device_commit_async(&env, mdt);
2248 /* Used for cross-MDT lock */
2249 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2250 void *data, int flag)
2252 struct lustre_handle lockh;
2256 case LDLM_CB_BLOCKING:
2257 ldlm_lock2handle(lock, &lockh);
2258 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
2260 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2264 case LDLM_CB_CANCELING:
2265 LDLM_DEBUG(lock, "Revoke remote lock\n");
2273 int mdt_check_resent_lock(struct mdt_thread_info *info,
2274 struct mdt_object *mo,
2275 struct mdt_lock_handle *lhc)
2277 /* the lock might already be gotten in ldlm_handle_enqueue() */
2278 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2279 struct ptlrpc_request *req = mdt_info_req(info);
2280 struct ldlm_lock *lock;
2282 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2283 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2285 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2286 * a resend case, however, it could be already destroyed
2287 * due to client eviction or a raced cancel RPC. */
2288 LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
2289 lhc->mlh_reg_lh.cookie);
2293 if (!fid_res_name_eq(mdt_object_fid(mo),
2294 &lock->l_resource->lr_name)) {
2295 CWARN("%s: Although resent, but still not "
2296 "get child lock:"DFID"\n",
2297 info->mti_exp->exp_obd->obd_name,
2298 PFID(mdt_object_fid(mo)));
2299 LDLM_LOCK_PUT(lock);
2302 LDLM_LOCK_PUT(lock);
2308 int mdt_remote_object_lock(struct mdt_thread_info *mti,
2309 struct mdt_object *o, const struct lu_fid *fid,
2310 struct lustre_handle *lh, ldlm_mode_t mode,
2313 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2314 ldlm_policy_data_t *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 memset(policy, 0, sizeof(*policy));
2332 policy->l_inodebits.bits = ibits;
2334 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2339 static int mdt_object_local_lock(struct mdt_thread_info *info,
2340 struct mdt_object *o,
2341 struct mdt_lock_handle *lh, __u64 ibits,
2344 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2345 ldlm_policy_data_t *policy = &info->mti_policy;
2346 struct ldlm_res_id *res_id = &info->mti_res_id;
2351 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2352 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2353 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2354 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2356 /* Only enqueue LOOKUP lock for remote object */
2357 if (mdt_object_remote(o))
2358 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2360 if (lh->mlh_type == MDT_PDO_LOCK) {
2361 /* check for exists after object is locked */
2362 if (mdt_object_exists(o) == 0) {
2363 /* Non-existent object shouldn't have PDO lock */
2366 /* Non-dir object shouldn't have PDO lock */
2367 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2372 memset(policy, 0, sizeof(*policy));
2373 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2375 dlmflags = LDLM_FL_ATOMIC_CB;
2377 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2380 * Take PDO lock on whole directory and build correct @res_id for lock
2381 * on part of directory.
2383 if (lh->mlh_pdo_hash != 0) {
2384 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2385 mdt_lock_pdo_mode(info, o, lh);
2386 if (lh->mlh_pdo_mode != LCK_NL) {
2388 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2389 * is never going to be sent to client and we do not
2390 * want it slowed down due to possible cancels.
2392 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2393 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2394 policy, res_id, dlmflags,
2395 info->mti_exp == NULL ? NULL :
2396 &info->mti_exp->exp_handle.h_cookie);
2397 if (unlikely(rc != 0))
2398 GOTO(out_unlock, rc);
2402 * Finish res_id initializing by name hash marking part of
2403 * directory which is taking modification.
2405 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2408 policy->l_inodebits.bits = ibits;
2411 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2412 * going to be sent to client. If it is - mdt_intent_policy() path will
2413 * fix it up and turn FL_LOCAL flag off.
2415 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2416 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2417 info->mti_exp == NULL ? NULL :
2418 &info->mti_exp->exp_handle.h_cookie);
2421 mdt_object_unlock(info, o, lh, 1);
2422 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2423 lh->mlh_pdo_hash != 0 &&
2424 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2425 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2431 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2432 struct mdt_lock_handle *lh, __u64 ibits,
2438 if (!mdt_object_remote(o))
2439 return mdt_object_local_lock(info, o, lh, ibits, nonblock);
2441 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2442 ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2443 MDS_INODELOCK_XATTR);
2444 if (ibits & MDS_INODELOCK_UPDATE) {
2445 /* Sigh, PDO needs to enqueue 2 locks right now, but
2446 * enqueue RPC can only request 1 lock, to avoid extra
2447 * RPC, so it will instead enqueue EX lock for remote
2448 * object anyway XXX*/
2449 if (lh->mlh_type == MDT_PDO_LOCK &&
2450 lh->mlh_pdo_hash != 0) {
2451 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2452 "EX lock.\n", mdt_obd_name(info->mti_mdt),
2453 PFID(mdt_object_fid(o)));
2454 lh->mlh_pdo_hash = 0;
2455 lh->mlh_rreg_mode = LCK_EX;
2456 lh->mlh_type = MDT_REG_LOCK;
2458 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2461 MDS_INODELOCK_UPDATE);
2466 /* Only enqueue LOOKUP lock for remote object */
2467 if (ibits & MDS_INODELOCK_LOOKUP) {
2468 rc = mdt_object_local_lock(info, o, lh,
2469 MDS_INODELOCK_LOOKUP,
2478 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2479 struct mdt_lock_handle *lh, __u64 ibits)
2481 return mdt_object_lock_internal(info, o, lh, ibits, false);
2484 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2485 struct mdt_lock_handle *lh, __u64 ibits)
2487 struct mdt_lock_handle tmp = *lh;
2490 rc = mdt_object_lock_internal(info, o, &tmp, ibits, true);
2498 * Save a lock within request object.
2500 * Keep the lock referenced until whether client ACK or transaction
2501 * commit happens or release the lock immediately depending on input
2502 * parameters. If COS is ON, a write lock is converted to COS lock
2505 * \param info thead info object
2506 * \param h lock handle
2507 * \param mode lock mode
2508 * \param decref force immediate lock releasing
2511 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2512 ldlm_mode_t mode, int decref)
2516 if (lustre_handle_is_used(h)) {
2517 if (decref || !info->mti_has_trans ||
2518 !(mode & (LCK_PW | LCK_EX))){
2519 mdt_fid_unlock(h, mode);
2521 struct mdt_device *mdt = info->mti_mdt;
2522 struct ldlm_lock *lock = ldlm_handle2lock(h);
2523 struct ptlrpc_request *req = mdt_info_req(info);
2526 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2528 /* there is no request if mdt_object_unlock() is called
2529 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2530 if (likely(req != NULL)) {
2531 CDEBUG(D_HA, "request = %p reply state = %p"
2532 " transno = "LPD64"\n", req,
2533 req->rq_reply_state, req->rq_transno);
2534 if (mdt_cos_is_enabled(mdt)) {
2536 ldlm_lock_downgrade(lock, LCK_COS);
2539 ptlrpc_save_lock(req, h, mode, no_ack);
2541 ldlm_lock_decref(h, mode);
2543 if (mdt_is_lock_sync(lock)) {
2544 CDEBUG(D_HA, "found sync-lock,"
2545 " async commit started\n");
2546 mdt_device_commit_async(info->mti_env,
2549 LDLM_LOCK_PUT(lock);
2558 * Unlock mdt object.
2560 * Immeditely release the regular lock and the PDO lock or save the
2561 * lock in request and keep them referenced until client ACK or
2562 * transaction commit.
2564 * \param info thread info object
2565 * \param o mdt object
2566 * \param lh mdt lock handle referencing regular and PDO locks
2567 * \param decref force immediate lock releasing
2569 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2570 struct mdt_lock_handle *lh, int decref)
2574 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2575 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2577 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2578 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2583 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2584 const struct lu_fid *f,
2585 struct mdt_lock_handle *lh,
2588 struct mdt_object *o;
2590 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2594 rc = mdt_object_lock(info, o, lh, ibits);
2596 mdt_object_put(info->mti_env, o);
2603 void mdt_object_unlock_put(struct mdt_thread_info * info,
2604 struct mdt_object * o,
2605 struct mdt_lock_handle *lh,
2608 mdt_object_unlock(info, o, lh, decref);
2609 mdt_object_put(info->mti_env, o);
2613 * Generic code handling requests that have struct mdt_body passed in:
2615 * - extract mdt_body from request and save it in @info, if present;
2617 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2620 * - if HABEO_CORPUS flag is set for this request type check whether object
2621 * actually exists on storage (lu_object_exists()).
2624 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2626 const struct mdt_body *body;
2627 struct mdt_object *obj;
2628 const struct lu_env *env;
2629 struct req_capsule *pill;
2633 env = info->mti_env;
2634 pill = info->mti_pill;
2636 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2640 if (!(body->mbo_valid & OBD_MD_FLID))
2643 if (!fid_is_sane(&body->mbo_fid1)) {
2644 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2648 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2650 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2651 mdt_object_put(env, obj);
2654 info->mti_object = obj;
2663 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2665 struct req_capsule *pill = info->mti_pill;
2669 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2670 rc = mdt_body_unpack(info, flags);
2674 if (rc == 0 && (flags & HABEO_REFERO)) {
2676 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2677 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2679 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2680 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2683 rc = req_capsule_server_pack(pill);
2688 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2690 lh->mlh_type = MDT_NUL_LOCK;
2691 lh->mlh_reg_lh.cookie = 0ull;
2692 lh->mlh_reg_mode = LCK_MINMODE;
2693 lh->mlh_pdo_lh.cookie = 0ull;
2694 lh->mlh_pdo_mode = LCK_MINMODE;
2695 lh->mlh_rreg_lh.cookie = 0ull;
2696 lh->mlh_rreg_mode = LCK_MINMODE;
2699 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2701 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2702 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2706 * Initialize fields of struct mdt_thread_info. Other fields are left in
2707 * uninitialized state, because it's too expensive to zero out whole
2708 * mdt_thread_info (> 1K) on each request arrival.
2710 void mdt_thread_info_init(struct ptlrpc_request *req,
2711 struct mdt_thread_info *info)
2715 info->mti_pill = &req->rq_pill;
2718 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2719 mdt_lock_handle_init(&info->mti_lh[i]);
2721 /* mdt device: it can be NULL while CONNECT */
2722 if (req->rq_export) {
2723 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2724 info->mti_exp = req->rq_export;
2726 info->mti_mdt = NULL;
2727 info->mti_env = req->rq_svc_thread->t_env;
2728 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2730 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2731 info->mti_big_buf = LU_BUF_NULL;
2732 info->mti_body = NULL;
2733 info->mti_object = NULL;
2734 info->mti_dlm_req = NULL;
2735 info->mti_has_trans = 0;
2736 info->mti_cross_ref = 0;
2737 info->mti_opdata = 0;
2738 info->mti_big_lmm_used = 0;
2740 info->mti_spec.no_create = 0;
2741 info->mti_spec.sp_rm_entry = 0;
2742 info->mti_spec.sp_permitted = 0;
2744 info->mti_spec.u.sp_ea.eadata = NULL;
2745 info->mti_spec.u.sp_ea.eadatalen = 0;
2748 void mdt_thread_info_fini(struct mdt_thread_info *info)
2752 if (info->mti_object != NULL) {
2753 mdt_object_put(info->mti_env, info->mti_object);
2754 info->mti_object = NULL;
2757 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2758 mdt_lock_handle_fini(&info->mti_lh[i]);
2759 info->mti_env = NULL;
2760 info->mti_pill = NULL;
2761 info->mti_exp = NULL;
2763 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2764 lu_buf_free(&info->mti_big_buf);
2767 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2769 struct mdt_thread_info *mti;
2771 mti = mdt_th_info(tsi->tsi_env);
2772 LASSERT(mti != NULL);
2774 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2775 if (tsi->tsi_corpus != NULL) {
2776 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2777 lu_object_get(tsi->tsi_corpus);
2779 mti->mti_body = tsi->tsi_mdt_body;
2780 mti->mti_dlm_req = tsi->tsi_dlm_req;
2785 static int mdt_tgt_connect(struct tgt_session_info *tsi)
2787 struct ptlrpc_request *req = tgt_ses_req(tsi);
2792 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
2794 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
2795 set_current_state(TASK_UNINTERRUPTIBLE);
2796 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
2799 rc = tgt_connect(tsi);
2803 rc = mdt_init_idmap(tsi);
2808 obd_disconnect(class_export_get(req->rq_export));
2827 static int mdt_intent_getattr(enum mdt_it_code opcode,
2828 struct mdt_thread_info *info,
2829 struct ldlm_lock **,
2832 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2833 struct mdt_thread_info *info,
2834 struct ldlm_lock **lockp,
2837 static int mdt_intent_layout(enum mdt_it_code opcode,
2838 struct mdt_thread_info *info,
2839 struct ldlm_lock **,
2841 static int mdt_intent_reint(enum mdt_it_code opcode,
2842 struct mdt_thread_info *info,
2843 struct ldlm_lock **,
2846 static struct mdt_it_flavor {
2847 const struct req_format *it_fmt;
2849 int (*it_act)(enum mdt_it_code ,
2850 struct mdt_thread_info *,
2851 struct ldlm_lock **,
2854 } mdt_it_flavor[] = {
2856 .it_fmt = &RQF_LDLM_INTENT,
2857 /*.it_flags = HABEO_REFERO,*/
2859 .it_act = mdt_intent_reint,
2860 .it_reint = REINT_OPEN
2863 .it_fmt = &RQF_LDLM_INTENT,
2865 * OCREAT is not a MUTABOR request as if the file
2867 * We do the extra check of OBD_CONNECT_RDONLY in
2868 * mdt_reint_open() when we really need to create
2872 .it_act = mdt_intent_reint,
2873 .it_reint = REINT_OPEN
2876 .it_fmt = &RQF_LDLM_INTENT,
2877 .it_flags = MUTABOR,
2878 .it_act = mdt_intent_reint,
2879 .it_reint = REINT_CREATE
2881 [MDT_IT_GETATTR] = {
2882 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2883 .it_flags = HABEO_REFERO,
2884 .it_act = mdt_intent_getattr
2886 [MDT_IT_READDIR] = {
2892 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2893 .it_flags = HABEO_REFERO,
2894 .it_act = mdt_intent_getattr
2897 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2898 .it_flags = MUTABOR,
2900 .it_reint = REINT_UNLINK
2904 .it_flags = MUTABOR,
2907 [MDT_IT_GETXATTR] = {
2908 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2909 .it_flags = HABEO_CORPUS,
2910 .it_act = mdt_intent_getxattr
2913 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2915 .it_act = mdt_intent_layout
2920 mdt_intent_lock_replace(struct mdt_thread_info *info,
2921 struct ldlm_lock **lockp,
2922 struct mdt_lock_handle *lh,
2925 struct ptlrpc_request *req = mdt_info_req(info);
2926 struct ldlm_lock *lock = *lockp;
2927 struct ldlm_lock *new_lock;
2929 /* If possible resent found a lock, @lh is set to its handle */
2930 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2932 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2933 lh->mlh_reg_lh.cookie = 0;
2937 LASSERTF(new_lock != NULL,
2938 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2941 * If we've already given this lock to a client once, then we should
2942 * have no readers or writers. Otherwise, we should have one reader
2943 * _or_ writer ref (which will be zeroed below) before returning the
2946 if (new_lock->l_export == req->rq_export) {
2947 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2949 LASSERT(new_lock->l_export == NULL);
2950 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2955 if (new_lock->l_export == req->rq_export) {
2957 * Already gave this to the client, which means that we
2958 * reconstructed a reply.
2960 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2963 LDLM_LOCK_RELEASE(new_lock);
2964 lh->mlh_reg_lh.cookie = 0;
2965 RETURN(ELDLM_LOCK_REPLACED);
2969 * Fixup the lock to be given to the client.
2971 lock_res_and_lock(new_lock);
2972 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2973 * possible blocking AST. */
2974 while (new_lock->l_readers > 0) {
2975 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2976 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2977 new_lock->l_readers--;
2979 while (new_lock->l_writers > 0) {
2980 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2981 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2982 new_lock->l_writers--;
2985 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2986 new_lock->l_blocking_ast = lock->l_blocking_ast;
2987 new_lock->l_completion_ast = lock->l_completion_ast;
2988 new_lock->l_remote_handle = lock->l_remote_handle;
2989 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2991 unlock_res_and_lock(new_lock);
2993 cfs_hash_add(new_lock->l_export->exp_lock_hash,
2994 &new_lock->l_remote_handle,
2995 &new_lock->l_exp_hash);
2997 LDLM_LOCK_RELEASE(new_lock);
2998 lh->mlh_reg_lh.cookie = 0;
3000 RETURN(ELDLM_LOCK_REPLACED);
3003 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3004 struct ldlm_lock *new_lock,
3005 struct mdt_lock_handle *lh,
3008 struct ptlrpc_request *req = mdt_info_req(info);
3009 struct ldlm_request *dlmreq;
3011 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3014 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3016 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3017 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3019 if (flags & LDLM_FL_RESENT) {
3020 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3021 lh->mlh_reg_mode = new_lock->l_granted_mode;
3023 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3024 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3025 lh->mlh_reg_lh.cookie);
3030 * If the xid matches, then we know this is a resent request, and allow
3031 * it. (It's probably an OPEN, for which we don't send a lock.
3033 if (req_can_reconstruct(req, NULL))
3037 * This remote handle isn't enqueued, so we never received or processed
3038 * this request. Clear MSG_RESENT, because it can be handled like any
3039 * normal request now.
3041 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3043 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3044 dlmreq->lock_handle[0].cookie);
3047 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3048 struct mdt_thread_info *info,
3049 struct ldlm_lock **lockp,
3052 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3053 struct ldlm_reply *ldlm_rep = NULL;
3057 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3058 * (for the resend case) or a new lock. Below we will use it to
3059 * replace the original lock.
3061 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3062 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3063 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3064 rc = mdt_object_lock(info, info->mti_object, lhc,
3065 MDS_INODELOCK_XATTR);
3070 grc = mdt_getxattr(info);
3072 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3074 if (mdt_info_req(info)->rq_repmsg != NULL)
3075 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3076 if (ldlm_rep == NULL)
3077 RETURN(err_serious(-EFAULT));
3079 ldlm_rep->lock_policy_res2 = grc;
3084 static int mdt_intent_getattr(enum mdt_it_code opcode,
3085 struct mdt_thread_info *info,
3086 struct ldlm_lock **lockp,
3089 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3091 struct ldlm_reply *ldlm_rep;
3092 struct mdt_body *reqbody;
3093 struct mdt_body *repbody;
3097 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3100 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3103 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3104 repbody->mbo_eadatasize = 0;
3105 repbody->mbo_aclsize = 0;
3109 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3111 case MDT_IT_GETATTR:
3112 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3116 CERROR("Unsupported intent (%d)\n", opcode);
3117 GOTO(out_shrink, rc = -EINVAL);
3120 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3122 GOTO(out_shrink, rc);
3124 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3125 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3127 /* Get lock from request for possible resent case. */
3128 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3130 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3131 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3133 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3134 ldlm_rep->lock_policy_res2 = 0;
3135 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3136 ldlm_rep->lock_policy_res2) {
3137 lhc->mlh_reg_lh.cookie = 0ull;
3138 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3141 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3144 mdt_exit_ucred(info);
3146 mdt_client_compatibility(info);
3147 rc2 = mdt_fix_reply(info);
3153 static int mdt_intent_layout(enum mdt_it_code opcode,
3154 struct mdt_thread_info *info,
3155 struct ldlm_lock **lockp,
3158 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3159 struct layout_intent *layout;
3161 struct mdt_object *obj = NULL;
3165 if (opcode != MDT_IT_LAYOUT) {
3166 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3171 fid = &info->mti_tmp_fid2;
3172 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3174 /* Get lock from request for possible resent case. */
3175 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3177 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3179 RETURN(PTR_ERR(obj));
3181 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3182 /* get the length of lsm */
3183 rc = mdt_attr_get_eabuf_size(info, obj);
3185 mdt_object_put(info->mti_env, obj);
3189 if (rc > info->mti_mdt->mdt_max_mdsize)
3190 info->mti_mdt->mdt_max_mdsize = rc;
3193 mdt_object_put(info->mti_env, obj);
3195 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3196 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER, rc);
3197 rc = req_capsule_server_pack(info->mti_pill);
3201 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
3202 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3204 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3205 LASSERT(layout != NULL);
3206 if (layout->li_opc == LAYOUT_INTENT_ACCESS)
3207 /* return to normal/resent ldlm handling */
3210 CERROR("%s: Unsupported layout intent (%d)\n",
3211 mdt_obd_name(info->mti_mdt), layout->li_opc);
3215 static int mdt_intent_reint(enum mdt_it_code opcode,
3216 struct mdt_thread_info *info,
3217 struct ldlm_lock **lockp,
3220 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3221 struct ldlm_reply *rep = NULL;
3225 static const struct req_format *intent_fmts[REINT_MAX] = {
3226 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3227 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3232 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3236 if (mdt_it_flavor[opcode].it_reint != opc) {
3237 CERROR("Reint code %ld doesn't match intent: %d\n",
3239 RETURN(err_serious(-EPROTO));
3242 /* Get lock from request for possible resent case. */
3243 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3245 rc = mdt_reint_internal(info, lhc, opc);
3247 /* Check whether the reply has been packed successfully. */
3248 if (mdt_info_req(info)->rq_repmsg != NULL)
3249 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3251 RETURN(err_serious(-EFAULT));
3253 /* MDC expects this in any case */
3255 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3257 /* the open lock or the lock for cross-ref object should be
3258 * returned to the client */
3259 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3260 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3261 rep->lock_policy_res2 = 0;
3262 rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3266 rep->lock_policy_res2 = clear_serious(rc);
3268 if (rep->lock_policy_res2 == -ENOENT &&
3269 mdt_get_disposition(rep, DISP_LOOKUP_NEG))
3270 rep->lock_policy_res2 = 0;
3272 lhc->mlh_reg_lh.cookie = 0ull;
3273 if (rc == -ENOTCONN || rc == -ENODEV ||
3274 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3276 * If it is the disconnect error (ENODEV & ENOCONN), the error
3277 * will be returned by rq_status, and client at ptlrpc layer
3278 * will detect this, then disconnect, reconnect the import
3279 * immediately, instead of impacting the following the rpc.
3284 * For other cases, the error will be returned by intent, and client
3285 * will retrieve the result from intent.
3287 RETURN(ELDLM_LOCK_ABORTED);
3290 static int mdt_intent_code(long itcode)
3298 case IT_OPEN|IT_CREAT:
3305 rc = MDT_IT_READDIR;
3308 rc = MDT_IT_GETATTR;
3320 rc = MDT_IT_GETXATTR;
3325 case IT_QUOTA_DQACQ:
3330 CERROR("Unknown intent opcode: %ld\n", itcode);
3337 static int mdt_intent_opc(long itopc, struct mdt_thread_info *info,
3338 struct ldlm_lock **lockp, __u64 flags)
3340 struct req_capsule *pill = info->mti_pill;
3341 struct ptlrpc_request *req = mdt_info_req(info);
3342 struct mdt_it_flavor *flv;
3347 opc = mdt_intent_code(itopc);
3351 if (opc == MDT_IT_QUOTA) {
3352 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3355 RETURN(-EOPNOTSUPP);
3357 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3358 /* pass the request to quota master */
3359 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3360 mdt_info_req(info), lockp,
3365 flv = &mdt_it_flavor[opc];
3366 if (flv->it_fmt != NULL)
3367 req_capsule_extend(pill, flv->it_fmt);
3369 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3373 if (flv->it_flags & MUTABOR &&
3374 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3377 if (flv->it_act != NULL) {
3378 struct ldlm_reply *rep;
3380 /* execute policy */
3381 rc = flv->it_act(opc, info, lockp, flags);
3383 /* Check whether the reply has been packed successfully. */
3384 if (req->rq_repmsg != NULL) {
3385 rep = req_capsule_server_get(info->mti_pill,
3387 rep->lock_policy_res2 =
3388 ptlrpc_status_hton(rep->lock_policy_res2);
3395 static int mdt_intent_policy(struct ldlm_namespace *ns,
3396 struct ldlm_lock **lockp, void *req_cookie,
3397 ldlm_mode_t mode, __u64 flags, void *data)
3399 struct tgt_session_info *tsi;
3400 struct mdt_thread_info *info;
3401 struct ptlrpc_request *req = req_cookie;
3402 struct ldlm_intent *it;
3403 struct req_capsule *pill;
3408 LASSERT(req != NULL);
3410 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3412 info = tsi2mdt_info(tsi);
3413 LASSERT(info != NULL);
3414 pill = info->mti_pill;
3415 LASSERT(pill->rc_req == req);
3417 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3418 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3419 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3421 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3425 /* Lock without inodebits makes no sense and will oops
3426 * later in ldlm. Let's check it now to see if we have
3427 * ibits corrupted somewhere in mdt_intent_opc().
3428 * The case for client miss to set ibits has been
3429 * processed by others. */
3430 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3431 lr_type == LDLM_IBITS,
3432 info->mti_dlm_req->lock_desc.\
3433 l_policy_data.l_inodebits.bits != 0));
3435 rc = err_serious(-EFAULT);
3437 /* No intent was provided */
3438 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3439 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3440 rc = req_capsule_server_pack(pill);
3442 rc = err_serious(rc);
3444 mdt_thread_info_fini(info);
3448 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3450 struct seq_server_site *ss = mdt_seq_site(mdt);
3452 if (ss->ss_node_id == 0)
3455 if (ss->ss_client_seq != NULL) {
3456 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3457 ss->ss_client_seq->lcs_exp = NULL;
3460 if (ss->ss_server_fld != NULL) {
3461 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3462 ss->ss_server_fld->lsf_control_exp = NULL;
3466 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3468 struct seq_server_site *ss = mdt_seq_site(mdt);
3473 if (ss->ss_server_seq != NULL)
3474 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3477 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3479 mdt_seq_fini_cli(mdt);
3480 mdt_deregister_seq_exp(mdt);
3482 return seq_site_fini(env, mdt_seq_site(mdt));
3486 * It will retrieve its FLDB entries from MDT0, and it only happens
3487 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3488 * and it needs to refresh FLDB from the MDT0.
3490 static int mdt_register_lwp_callback(void *data)
3493 struct mdt_device *mdt = data;
3494 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3498 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3500 rc = lu_env_init(&env, LCT_MD_THREAD);
3502 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3506 /* Allocate new sequence now to avoid creating local transaction
3507 * in the normal transaction process */
3508 rc = seq_server_check_and_alloc_super(&env,
3509 mdt_seq_site(mdt)->ss_server_seq);
3514 rc = fld_update_from_controller(&env, fld);
3516 CERROR("%s: cannot update controller: rc = %d\n",
3517 mdt_obd_name(mdt), rc);
3526 static int mdt_register_seq_exp(struct mdt_device *mdt)
3528 struct seq_server_site *ss = mdt_seq_site(mdt);
3529 char *lwp_name = NULL;
3532 if (ss->ss_node_id == 0)
3535 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3536 if (lwp_name == NULL)
3537 GOTO(out_free, rc = -ENOMEM);
3539 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3543 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3548 rc = lustre_register_lwp_item(lwp_name,
3549 &ss->ss_server_fld->lsf_control_exp,
3550 mdt_register_lwp_callback, mdt);
3552 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3553 ss->ss_client_seq->lcs_exp = NULL;
3557 if (lwp_name != NULL)
3558 OBD_FREE(lwp_name, MAX_OBD_NAME);
3564 * Init client sequence manager which is used by local MDS to talk to sequence
3565 * controller on remote node.
3567 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3569 struct seq_server_site *ss = mdt_seq_site(mdt);
3574 /* check if this is adding the first MDC and controller is not yet
3576 OBD_ALLOC_PTR(ss->ss_client_seq);
3577 if (ss->ss_client_seq == NULL)
3580 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3581 if (prefix == NULL) {
3582 OBD_FREE_PTR(ss->ss_client_seq);
3583 ss->ss_client_seq = NULL;
3587 /* Note: seq_client_fini will be called in seq_site_fini */
3588 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3589 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3590 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3592 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3594 OBD_FREE_PTR(ss->ss_client_seq);
3595 ss->ss_client_seq = NULL;
3599 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3604 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3606 struct seq_server_site *ss;
3610 ss = mdt_seq_site(mdt);
3611 /* init sequence controller server(MDT0) */
3612 if (ss->ss_node_id == 0) {
3613 OBD_ALLOC_PTR(ss->ss_control_seq);
3614 if (ss->ss_control_seq == NULL)
3617 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3618 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3621 GOTO(out_seq_fini, rc);
3624 /* Init normal sequence server */
3625 OBD_ALLOC_PTR(ss->ss_server_seq);
3626 if (ss->ss_server_seq == NULL)
3627 GOTO(out_seq_fini, rc = -ENOMEM);
3629 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3630 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3632 GOTO(out_seq_fini, rc);
3634 /* init seq client for seq server to talk to seq controller(MDT0) */
3635 rc = mdt_seq_init_cli(env, mdt);
3637 GOTO(out_seq_fini, rc);
3639 if (ss->ss_node_id != 0)
3640 /* register controller export through lwp */
3641 rc = mdt_register_seq_exp(mdt);
3646 mdt_seq_fini(env, mdt);
3654 static int mdt_fld_fini(const struct lu_env *env,
3655 struct mdt_device *m)
3657 struct seq_server_site *ss = mdt_seq_site(m);
3660 if (ss && ss->ss_server_fld) {
3661 fld_server_fini(env, ss->ss_server_fld);
3662 OBD_FREE_PTR(ss->ss_server_fld);
3663 ss->ss_server_fld = NULL;
3669 static int mdt_fld_init(const struct lu_env *env,
3671 struct mdt_device *m)
3673 struct seq_server_site *ss;
3677 ss = mdt_seq_site(m);
3679 OBD_ALLOC_PTR(ss->ss_server_fld);
3680 if (ss->ss_server_fld == NULL)
3681 RETURN(rc = -ENOMEM);
3683 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3686 OBD_FREE_PTR(ss->ss_server_fld);
3687 ss->ss_server_fld = NULL;
3694 static void mdt_stack_pre_fini(const struct lu_env *env,
3695 struct mdt_device *m, struct lu_device *top)
3697 struct lustre_cfg_bufs *bufs;
3698 struct lustre_cfg *lcfg;
3699 struct mdt_thread_info *info;
3704 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3705 LASSERT(info != NULL);
3707 bufs = &info->mti_u.bufs;
3709 LASSERT(m->mdt_child_exp);
3710 LASSERT(m->mdt_child_exp->exp_obd);
3712 /* process cleanup, pass mdt obd name to get obd umount flags */
3713 /* XXX: this is needed because all layers are referenced by
3714 * objects (some of them are pinned by osd, for example *
3715 * the proper solution should be a model where object used
3716 * by osd only doesn't have mdt/mdd slices -bzzz */
3717 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3718 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3719 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3723 top->ld_ops->ldo_process_config(env, top, lcfg);
3724 lustre_cfg_free(lcfg);
3728 static void mdt_stack_fini(const struct lu_env *env,
3729 struct mdt_device *m, struct lu_device *top)
3731 struct obd_device *obd = mdt2obd_dev(m);
3732 struct lustre_cfg_bufs *bufs;
3733 struct lustre_cfg *lcfg;
3734 struct mdt_thread_info *info;
3738 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3739 LASSERT(info != NULL);
3741 lu_dev_del_linkage(top->ld_site, top);
3743 lu_site_purge(env, top->ld_site, -1);
3745 bufs = &info->mti_u.bufs;
3746 /* process cleanup, pass mdt obd name to get obd umount flags */
3747 /* another purpose is to let all layers to release their objects */
3748 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3753 lustre_cfg_bufs_set_string(bufs, 1, flags);
3754 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3759 top->ld_ops->ldo_process_config(env, top, lcfg);
3760 lustre_cfg_free(lcfg);
3762 lu_site_purge(env, top->ld_site, -1);
3764 m->mdt_child = NULL;
3765 m->mdt_bottom = NULL;
3767 obd_disconnect(m->mdt_child_exp);
3768 m->mdt_child_exp = NULL;
3770 obd_disconnect(m->mdt_bottom_exp);
3771 m->mdt_child_exp = NULL;
3774 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3775 const char *next, struct obd_export **exp)
3777 struct obd_connect_data *data = NULL;
3778 struct obd_device *obd;
3782 OBD_ALLOC_PTR(data);
3784 GOTO(out, rc = -ENOMEM);
3786 obd = class_name2obd(next);
3788 CERROR("%s: can't locate next device: %s\n",
3789 mdt_obd_name(m), next);
3790 GOTO(out, rc = -ENOTCONN);
3793 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3794 data->ocd_version = LUSTRE_VERSION_CODE;
3796 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3798 CERROR("%s: cannot connect to next dev %s (%d)\n",
3799 mdt_obd_name(m), next, rc);
3809 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3810 struct lustre_cfg *cfg)
3812 char *dev = lustre_cfg_string(cfg, 0);
3813 int rc, name_size, uuid_size;
3814 char *name, *uuid, *p;
3815 struct lustre_cfg_bufs *bufs;
3816 struct lustre_cfg *lcfg;
3817 struct obd_device *obd;
3818 struct lustre_profile *lprof;
3819 struct lu_site *site;
3822 /* in 1.8 we had the only device in the stack - MDS.
3823 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3824 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3825 * to generate names and setup MDT, MDD. MDT will be using
3826 * generated name to connect to MDD. for MDD the next device
3827 * will be LOD with name taken from so called "profile" which
3828 * is generated by mount_option line
3830 * 1.8 MGS generates config. commands like this:
3831 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3832 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3833 * 2.0 MGS generates config. commands like this:
3834 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3835 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3836 * 3:lustre-MDT0000-mdtlov 4:f
3838 * we generate MDD name from MDT one, just replacing T with D
3840 * after all the preparations, the logical equivalent will be
3841 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3842 * 3:lustre-MDT0000-mdtlov 4:f
3843 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3844 * 3:lustre-MDD0000 4:f
3846 * notice we build the stack from down to top: MDD first, then MDT */
3848 name_size = MAX_OBD_NAME;
3849 uuid_size = MAX_OBD_NAME;
3851 OBD_ALLOC(name, name_size);
3852 OBD_ALLOC(uuid, uuid_size);
3853 if (name == NULL || uuid == NULL)
3854 GOTO(cleanup_mem, rc = -ENOMEM);
3856 OBD_ALLOC_PTR(bufs);
3858 GOTO(cleanup_mem, rc = -ENOMEM);
3861 p = strstr(name, "-MDT");
3863 GOTO(free_bufs, rc = -ENOMEM);
3866 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3868 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3869 if (lprof == NULL || lprof->lp_dt == NULL) {
3870 CERROR("can't find the profile: %s\n",
3871 lustre_cfg_string(cfg, 0));
3872 GOTO(free_bufs, rc = -EINVAL);
3875 lustre_cfg_bufs_reset(bufs, name);
3876 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3877 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3878 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3880 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3882 GOTO(free_bufs, rc = -ENOMEM);
3884 rc = class_attach(lcfg);
3886 GOTO(lcfg_cleanup, rc);
3888 obd = class_name2obd(name);
3890 CERROR("Can not find obd %s (%s in config)\n",
3891 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3892 GOTO(lcfg_cleanup, rc = -EINVAL);
3895 lustre_cfg_free(lcfg);
3897 lustre_cfg_bufs_reset(bufs, name);
3898 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3899 lustre_cfg_bufs_set_string(bufs, 2, dev);
3900 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3902 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3904 GOTO(class_detach, rc = -ENOMEM);
3906 rc = class_setup(obd, lcfg);
3908 GOTO(class_detach, rc);
3910 /* connect to MDD we just setup */
3911 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3913 GOTO(class_detach, rc);
3915 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3917 LASSERT(mdt_lu_site(mdt) == NULL);
3918 mdt->mdt_lu_dev.ld_site = site;
3919 site->ls_top_dev = &mdt->mdt_lu_dev;
3920 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3922 /* now connect to bottom OSD */
3923 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3924 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3926 GOTO(class_detach, rc);
3928 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3930 rc = lu_env_refill((struct lu_env *)env);
3932 CERROR("Failure to refill session: '%d'\n", rc);
3934 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3939 class_detach(obd, lcfg);
3941 lustre_cfg_free(lcfg);
3946 OBD_FREE(name, name_size);
3948 OBD_FREE(uuid, uuid_size);
3952 /* setup quota master target on MDT0 */
3953 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3954 struct lustre_cfg *cfg)
3956 struct obd_device *obd;
3957 char *dev = lustre_cfg_string(cfg, 0);
3958 char *qmtname, *uuid, *p;
3959 struct lustre_cfg_bufs *bufs;
3960 struct lustre_cfg *lcfg;
3961 struct lustre_profile *lprof;
3962 struct obd_connect_data *data;
3966 LASSERT(mdt->mdt_qmt_exp == NULL);
3967 LASSERT(mdt->mdt_qmt_dev == NULL);
3969 /* quota master is on MDT0 only for now */
3970 if (mdt->mdt_seq_site.ss_node_id != 0)
3973 /* MGS generates config commands which look as follows:
3974 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3975 * 3:lustre-MDT0000-mdtlov 4:f
3977 * We generate the QMT name from the MDT one, just replacing MD with QM
3978 * after all the preparations, the logical equivalent will be:
3979 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
3980 * 3:lustre-MDT0000-osd 4:f */
3981 OBD_ALLOC(qmtname, MAX_OBD_NAME);
3982 OBD_ALLOC(uuid, UUID_MAX);
3983 OBD_ALLOC_PTR(bufs);
3984 OBD_ALLOC_PTR(data);
3985 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
3986 GOTO(cleanup_mem, rc = -ENOMEM);
3988 strcpy(qmtname, dev);
3989 p = strstr(qmtname, "-MDT");
3991 GOTO(cleanup_mem, rc = -ENOMEM);
3992 /* replace MD with QM */
3996 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
3998 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3999 if (lprof == NULL || lprof->lp_dt == NULL) {
4000 CERROR("can't find profile for %s\n",
4001 lustre_cfg_string(cfg, 0));
4002 GOTO(cleanup_mem, rc = -EINVAL);
4005 lustre_cfg_bufs_reset(bufs, qmtname);
4006 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4007 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4008 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4010 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4012 GOTO(cleanup_mem, rc = -ENOMEM);
4014 rc = class_attach(lcfg);
4016 GOTO(lcfg_cleanup, rc);
4018 obd = class_name2obd(qmtname);
4020 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4021 lustre_cfg_string(cfg, 0));
4022 GOTO(lcfg_cleanup, rc = -EINVAL);
4025 lustre_cfg_free(lcfg);
4027 lustre_cfg_bufs_reset(bufs, qmtname);
4028 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4029 lustre_cfg_bufs_set_string(bufs, 2, dev);
4031 /* for quota, the next device should be the OSD device */
4032 lustre_cfg_bufs_set_string(bufs, 3,
4033 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4035 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4037 GOTO(class_detach, rc = -ENOMEM);
4039 rc = class_setup(obd, lcfg);
4041 GOTO(class_detach, rc);
4043 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4045 /* configure local quota objects */
4046 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4050 GOTO(class_cleanup, rc);
4052 /* connect to quota master target */
4053 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4054 data->ocd_version = LUSTRE_VERSION_CODE;
4055 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4058 CERROR("cannot connect to quota master device %s (%d)\n",
4060 GOTO(class_cleanup, rc);
4066 class_manual_cleanup(obd);
4067 mdt->mdt_qmt_dev = NULL;
4071 class_detach(obd, lcfg);
4073 lustre_cfg_free(lcfg);
4078 OBD_FREE(qmtname, MAX_OBD_NAME);
4080 OBD_FREE(uuid, UUID_MAX);
4086 /* Shutdown quota master target associated with mdt */
4087 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4091 if (mdt->mdt_qmt_exp == NULL)
4093 LASSERT(mdt->mdt_qmt_dev != NULL);
4095 /* the qmt automatically shuts down when the mdt disconnects */
4096 obd_disconnect(mdt->mdt_qmt_exp);
4097 mdt->mdt_qmt_exp = NULL;
4098 mdt->mdt_qmt_dev = NULL;
4102 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4103 * for now. This will be removed along with converting rest of MDT code
4104 * to use tgt_session_info */
4105 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4107 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4110 rc = mdt_getxattr(info);
4112 mdt_thread_info_fini(info);
4116 static struct tgt_handler mdt_tgt_handlers[] = {
4117 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4118 0, MDS_CONNECT, mdt_tgt_connect,
4119 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4120 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4121 0, MDS_DISCONNECT, tgt_disconnect,
4122 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4123 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4124 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4125 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4126 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4127 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4128 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4129 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4131 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4132 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4133 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4134 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4135 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4136 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4137 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4138 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4140 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4141 mdt_hsm_ct_register),
4142 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4143 mdt_hsm_ct_unregister),
4144 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4146 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4148 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4149 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4151 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4156 static struct tgt_handler mdt_sec_ctx_ops[] = {
4157 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4158 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4159 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4162 static struct tgt_handler mdt_quota_ops[] = {
4163 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4166 static struct tgt_opc_slice mdt_common_slice[] = {
4168 .tos_opc_start = MDS_FIRST_OPC,
4169 .tos_opc_end = MDS_LAST_OPC,
4170 .tos_hs = mdt_tgt_handlers
4173 .tos_opc_start = OBD_FIRST_OPC,
4174 .tos_opc_end = OBD_LAST_OPC,
4175 .tos_hs = tgt_obd_handlers
4178 .tos_opc_start = LDLM_FIRST_OPC,
4179 .tos_opc_end = LDLM_LAST_OPC,
4180 .tos_hs = tgt_dlm_handlers
4183 .tos_opc_start = SEC_FIRST_OPC,
4184 .tos_opc_end = SEC_LAST_OPC,
4185 .tos_hs = mdt_sec_ctx_ops
4188 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4189 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4190 .tos_hs = tgt_out_handlers
4193 .tos_opc_start = FLD_FIRST_OPC,
4194 .tos_opc_end = FLD_LAST_OPC,
4195 .tos_hs = fld_handlers
4198 .tos_opc_start = SEQ_FIRST_OPC,
4199 .tos_opc_end = SEQ_LAST_OPC,
4200 .tos_hs = seq_handlers
4203 .tos_opc_start = QUOTA_DQACQ,
4204 .tos_opc_end = QUOTA_LAST_OPC,
4205 .tos_hs = mdt_quota_ops
4208 .tos_opc_start = LLOG_FIRST_OPC,
4209 .tos_opc_end = LLOG_LAST_OPC,
4210 .tos_hs = tgt_llog_handlers
4213 .tos_opc_start = LFSCK_FIRST_OPC,
4214 .tos_opc_end = LFSCK_LAST_OPC,
4215 .tos_hs = tgt_lfsck_handlers
4223 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4225 struct md_device *next = m->mdt_child;
4226 struct lu_device *d = &m->mdt_lu_dev;
4227 struct obd_device *obd = mdt2obd_dev(m);
4228 struct lfsck_stop stop;
4231 stop.ls_status = LS_PAUSED;
4233 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4235 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4236 target_recovery_fini(obd);
4237 ping_evictor_stop();
4239 if (m->mdt_opts.mo_coordinator)
4240 mdt_hsm_cdt_stop(m);
4242 mdt_hsm_cdt_fini(m);
4244 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4245 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4246 obd_exports_barrier(obd);
4247 obd_zombie_barrier();
4251 tgt_fini(env, &m->mdt_lut);
4252 mdt_fs_cleanup(env, m);
4253 upcall_cache_cleanup(m->mdt_identity_cache);
4254 m->mdt_identity_cache = NULL;
4256 if (m->mdt_namespace != NULL) {
4257 ldlm_namespace_free(m->mdt_namespace, NULL,
4258 d->ld_obd->obd_force);
4259 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4262 mdt_quota_fini(env, m);
4264 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4266 mdt_seq_fini(env, m);
4267 mdt_fld_fini(env, m);
4272 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4274 LASSERT(atomic_read(&d->ld_ref) == 0);
4276 server_put_mount(mdt_obd_name(m), true);
4281 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4283 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4284 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4286 struct mdt_thread_info *info;
4287 struct obd_device *obd;
4288 const char *dev = lustre_cfg_string(cfg, 0);
4289 const char *num = lustre_cfg_string(cfg, 2);
4290 struct lustre_mount_info *lmi = NULL;
4291 struct lustre_sb_info *lsi;
4293 struct seq_server_site *ss_site;
4294 const char *identity_upcall = "NONE";
4295 struct md_device *next;
4301 lu_device_init(&m->mdt_lu_dev, ldt);
4303 * Environment (env) might be missing mdt_thread_key values at that
4304 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4307 * Usually device allocation path doesn't use module key values, but
4308 * mdt has to do a lot of work here, so allocate key value.
4310 rc = lu_env_refill((struct lu_env *)env);
4314 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4315 LASSERT(info != NULL);
4317 obd = class_name2obd(dev);
4318 LASSERT(obd != NULL);
4320 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4322 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4324 /* default is coordinator off, it is started through conf_param
4326 m->mdt_opts.mo_coordinator = 0;
4328 lmi = server_get_mount(dev);
4330 CERROR("Cannot get mount info for %s!\n", dev);
4333 lsi = s2lsi(lmi->lmi_sb);
4334 /* CMD is supported only in IAM mode */
4336 node_id = simple_strtol(num, NULL, 10);
4337 obd->u.obt.obt_magic = OBT_MAGIC;
4340 m->mdt_squash.rsi_uid = 0;
4341 m->mdt_squash.rsi_gid = 0;
4342 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
4343 init_rwsem(&m->mdt_squash.rsi_sem);
4344 spin_lock_init(&m->mdt_osfs_lock);
4345 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4346 m->mdt_enable_remote_dir = 0;
4347 m->mdt_enable_remote_dir_gid = 0;
4349 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4350 m->mdt_lu_dev.ld_obd = obd;
4351 /* Set this lu_device to obd for error handling purposes. */
4352 obd->obd_lu_dev = &m->mdt_lu_dev;
4354 /* init the stack */
4355 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4357 CERROR("%s: Can't init device stack, rc %d\n",
4358 mdt_obd_name(m), rc);
4363 ss_site = mdt_seq_site(m);
4364 s->ld_seq_site = ss_site;
4367 /* set server index */
4368 ss_site->ss_node_id = node_id;
4370 /* failover is the default
4371 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4372 * assumed whose ss_node_id == 0 XXX
4374 obd->obd_replayable = 1;
4375 /* No connection accepted until configurations will finish */
4376 obd->obd_no_conn = 1;
4378 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4379 char *str = lustre_cfg_string(cfg, 4);
4380 if (strchr(str, 'n')) {
4381 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4382 obd->obd_replayable = 0;
4386 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4388 GOTO(err_fini_stack, rc);
4390 rc = mdt_seq_init(env, m);
4392 GOTO(err_fini_fld, rc);
4394 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4395 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4396 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4397 LDLM_NAMESPACE_SERVER,
4398 LDLM_NAMESPACE_GREEDY,
4400 if (m->mdt_namespace == NULL)
4401 GOTO(err_fini_seq, rc = -ENOMEM);
4403 m->mdt_namespace->ns_lvbp = m;
4404 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4406 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4407 /* set obd_namespace for compatibility with old code */
4408 obd->obd_namespace = m->mdt_namespace;
4410 rc = mdt_hsm_cdt_init(m);
4412 CERROR("%s: error initializing coordinator, rc %d\n",
4413 mdt_obd_name(m), rc);
4414 GOTO(err_free_ns, rc);
4417 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4418 OBD_FAIL_MDS_ALL_REQUEST_NET,
4419 OBD_FAIL_MDS_ALL_REPLY_NET);
4421 GOTO(err_free_hsm, rc);
4423 rc = mdt_fs_setup(env, m, obd, lsi);
4427 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4429 next = m->mdt_child;
4430 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4433 GOTO(err_fs_cleanup, rc);
4435 if (mntopts & MNTOPT_USERXATTR)
4436 m->mdt_opts.mo_user_xattr = 1;
4438 m->mdt_opts.mo_user_xattr = 0;
4440 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4442 GOTO(err_fs_cleanup, rc);
4444 if (mntopts & MNTOPT_ACL)
4445 m->mdt_opts.mo_acl = 1;
4447 m->mdt_opts.mo_acl = 0;
4449 /* XXX: to support suppgid for ACL, we enable identity_upcall
4450 * by default, otherwise, maybe got unexpected -EACCESS. */
4451 if (m->mdt_opts.mo_acl)
4452 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4454 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4456 &mdt_identity_upcall_cache_ops);
4457 if (IS_ERR(m->mdt_identity_cache)) {
4458 rc = PTR_ERR(m->mdt_identity_cache);
4459 m->mdt_identity_cache = NULL;
4460 GOTO(err_fs_cleanup, rc);
4463 rc = mdt_procfs_init(m, dev);
4465 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4466 GOTO(err_recovery, rc);
4469 rc = mdt_quota_init(env, m, cfg);
4471 GOTO(err_procfs, rc);
4473 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4474 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4475 "mdt_ldlm_client", m->mdt_ldlm_client);
4477 ping_evictor_start();
4479 /* recovery will be started upon mdt_prepare()
4480 * when the whole stack is complete and ready
4481 * to serve the requests */
4483 /* Reduce the initial timeout on an MDS because it doesn't need such
4484 * a long timeout as an OST does. Adaptive timeouts will adjust this
4485 * value appropriately. */
4486 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4487 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4493 target_recovery_fini(obd);
4494 upcall_cache_cleanup(m->mdt_identity_cache);
4495 m->mdt_identity_cache = NULL;
4497 mdt_fs_cleanup(env, m);
4499 tgt_fini(env, &m->mdt_lut);
4501 mdt_hsm_cdt_fini(m);
4503 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4504 obd->obd_namespace = m->mdt_namespace = NULL;
4506 mdt_seq_fini(env, m);
4508 mdt_fld_fini(env, m);
4510 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4513 server_put_mount(dev, true);
4517 /* For interoperability, the left element is old parameter, the right one
4518 * is the new version of the parameter, if some parameter is deprecated,
4519 * the new version should be set as NULL. */
4520 static struct cfg_interop_param mdt_interop_param[] = {
4521 { "mdt.group_upcall", NULL },
4522 { "mdt.quota_type", NULL },
4523 { "mdd.quota_type", NULL },
4524 { "mdt.rootsquash", "mdt.root_squash" },
4525 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4529 /* used by MGS to process specific configurations */
4530 static int mdt_process_config(const struct lu_env *env,
4531 struct lu_device *d, struct lustre_cfg *cfg)
4533 struct mdt_device *m = mdt_dev(d);
4534 struct md_device *md_next = m->mdt_child;
4535 struct lu_device *next = md2lu_dev(md_next);
4539 switch (cfg->lcfg_command) {
4541 struct obd_device *obd = d->ld_obd;
4543 /* For interoperability */
4544 struct cfg_interop_param *ptr = NULL;
4545 struct lustre_cfg *old_cfg = NULL;
4548 param = lustre_cfg_string(cfg, 1);
4549 if (param == NULL) {
4550 CERROR("param is empty\n");
4555 ptr = class_find_old_param(param, mdt_interop_param);
4557 if (ptr->new_param == NULL) {
4559 CWARN("For interoperability, skip this %s."
4560 " It is obsolete.\n", ptr->old_param);
4564 CWARN("Found old param %s, changed it to %s.\n",
4565 ptr->old_param, ptr->new_param);
4568 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4575 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
4577 if (rc > 0 || rc == -ENOSYS) {
4578 /* is it an HSM var ? */
4579 rc = class_process_proc_param(PARAM_HSM,
4580 hsm_cdt_get_proc_vars(),
4582 if (rc > 0 || rc == -ENOSYS)
4583 /* we don't understand; pass it on */
4584 rc = next->ld_ops->ldo_process_config(env, next,
4588 if (old_cfg != NULL)
4589 lustre_cfg_free(cfg);
4594 /* others are passed further */
4595 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4601 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4602 const struct lu_object_header *hdr,
4603 struct lu_device *d)
4605 struct mdt_object *mo;
4609 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4611 struct lu_object *o;
4612 struct lu_object_header *h;
4615 h = &mo->mot_header;
4616 lu_object_header_init(h);
4617 lu_object_init(o, h, d);
4618 lu_object_add_top(h, o);
4619 o->lo_ops = &mdt_obj_ops;
4620 spin_lock_init(&mo->mot_write_lock);
4621 mutex_init(&mo->mot_lov_mutex);
4622 init_rwsem(&mo->mot_open_sem);
4628 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4629 const struct lu_object_conf *unused)
4631 struct mdt_device *d = mdt_dev(o->lo_dev);
4632 struct lu_device *under;
4633 struct lu_object *below;
4637 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4638 PFID(lu_object_fid(o)));
4640 under = &d->mdt_child->md_lu_dev;
4641 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4642 if (below != NULL) {
4643 lu_object_add(o, below);
4650 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4652 struct mdt_object *mo = mdt_obj(o);
4653 struct lu_object_header *h;
4657 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4658 PFID(lu_object_fid(o)));
4660 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4661 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4664 lu_object_header_fini(h);
4665 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4670 static int mdt_object_print(const struct lu_env *env, void *cookie,
4671 lu_printer_t p, const struct lu_object *o)
4673 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4675 return (*p)(env, cookie,
4676 LUSTRE_MDT_NAME"-object@%p(flags=%d, writecount=%d)",
4677 mdto, mdto->mot_flags, mdto->mot_write_count);
4680 static int mdt_prepare(const struct lu_env *env,
4681 struct lu_device *pdev,
4682 struct lu_device *cdev)
4684 struct mdt_device *mdt = mdt_dev(cdev);
4685 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4686 struct obd_device *obd = cdev->ld_obd;
4693 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4697 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4701 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4705 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4706 /* The LFSCK instance is registered just now, so it must be there when
4707 * register the namespace to such instance. */
4708 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4710 if (mdt->mdt_seq_site.ss_node_id == 0) {
4711 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4712 &mdt->mdt_md_root_fid);
4717 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4719 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4720 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4721 LASSERT(obd->obd_no_conn);
4722 spin_lock(&obd->obd_dev_lock);
4723 obd->obd_no_conn = 0;
4724 spin_unlock(&obd->obd_dev_lock);
4726 if (obd->obd_recovering == 0)
4727 mdt_postrecov(env, mdt);
4732 const struct lu_device_operations mdt_lu_ops = {
4733 .ldo_object_alloc = mdt_object_alloc,
4734 .ldo_process_config = mdt_process_config,
4735 .ldo_prepare = mdt_prepare,
4738 static const struct lu_object_operations mdt_obj_ops = {
4739 .loo_object_init = mdt_object_init,
4740 .loo_object_free = mdt_object_free,
4741 .loo_object_print = mdt_object_print
4744 static int mdt_obd_set_info_async(const struct lu_env *env,
4745 struct obd_export *exp,
4746 __u32 keylen, void *key,
4747 __u32 vallen, void *val,
4748 struct ptlrpc_request_set *set)
4754 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4755 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4763 * Match client and server connection feature flags.
4765 * Compute the compatibility flags for a connection request based on
4766 * features mutually supported by client and server.
4768 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4769 * must not be updated here, otherwise a partially initialized value may
4770 * be exposed. After the connection request is successfully processed,
4771 * the top-level MDT connect request handler atomically updates the export
4772 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4773 * reply. \see mdt_connect().
4775 * \param exp the obd_export associated with this client/target pair
4776 * \param mdt the target device for the connection
4777 * \param data stores data for this connect request
4780 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4781 * \retval -EBADE client and server feature requirements are incompatible
4783 static int mdt_connect_internal(struct obd_export *exp,
4784 struct mdt_device *mdt,
4785 struct obd_connect_data *data)
4787 LASSERT(data != NULL);
4789 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4790 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4792 if (!(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
4793 !(data->ocd_connect_flags & OBD_CONNECT_IBITS)) {
4794 CWARN("%s: client %s does not support ibits lock, either "
4795 "very old or an invalid client: flags "LPX64"\n",
4796 mdt_obd_name(mdt), exp->exp_client_uuid.uuid,
4797 data->ocd_connect_flags);
4801 if (!mdt->mdt_opts.mo_acl)
4802 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4804 if (!mdt->mdt_opts.mo_user_xattr)
4805 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4807 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4808 data->ocd_brw_size = min(data->ocd_brw_size,
4809 (__u32)MD_MAX_BRW_SIZE);
4810 if (data->ocd_brw_size == 0) {
4811 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4812 " ocd_version: %x ocd_grant: %d "
4813 "ocd_index: %u ocd_brw_size is "
4814 "unexpectedly zero, network data "
4815 "corruption? Refusing connection of this"
4818 exp->exp_client_uuid.uuid,
4819 exp, data->ocd_connect_flags, data->ocd_version,
4820 data->ocd_grant, data->ocd_index);
4825 /* NB: Disregard the rule against updating
4826 * exp_connect_data.ocd_connect_flags in this case, since
4827 * tgt_client_new() needs to know if this is a lightweight
4828 * connection, and it is safe to expose this flag before
4829 * connection processing completes. */
4830 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4831 spin_lock(&exp->exp_lock);
4832 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4833 spin_unlock(&exp->exp_lock);
4836 data->ocd_version = LUSTRE_VERSION_CODE;
4838 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4839 CWARN("%s: MDS requires FID support, but client not\n",
4844 if (OCD_HAS_FLAG(data, PINGLESS)) {
4845 if (ptlrpc_pinger_suppress_pings()) {
4846 spin_lock(&exp->exp_obd->obd_dev_lock);
4847 list_del_init(&exp->exp_obd_chain_timed);
4848 spin_unlock(&exp->exp_obd->obd_dev_lock);
4850 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4854 data->ocd_max_easize = mdt->mdt_max_ea_size;
4856 /* NB: Disregard the rule against updating
4857 * exp_connect_data.ocd_connect_flags in this case, since
4858 * tgt_client_new() needs to know if this is client supports
4859 * multiple modify RPCs, and it is safe to expose this flag before
4860 * connection processing completes. */
4861 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
4862 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
4863 spin_lock(&exp->exp_lock);
4864 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
4865 spin_unlock(&exp->exp_lock);
4871 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
4872 struct mdt_thread_info *info,
4873 struct mdt_file_data *mfd)
4875 struct lu_context ses;
4879 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
4884 lu_context_enter(&ses);
4886 mdt_ucred(info)->uc_valid = UCRED_OLD;
4887 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
4889 lu_context_exit(&ses);
4890 lu_context_fini(&ses);
4896 static int mdt_export_cleanup(struct obd_export *exp)
4898 struct list_head closing_list;
4899 struct mdt_export_data *med = &exp->exp_mdt_data;
4900 struct obd_device *obd = exp->exp_obd;
4901 struct mdt_device *mdt;
4902 struct mdt_thread_info *info;
4904 struct mdt_file_data *mfd, *n;
4908 INIT_LIST_HEAD(&closing_list);
4909 spin_lock(&med->med_open_lock);
4910 while (!list_empty(&med->med_open_head)) {
4911 struct list_head *tmp = med->med_open_head.next;
4912 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
4914 /* Remove mfd handle so it can't be found again.
4915 * We are consuming the mfd_list reference here. */
4916 class_handle_unhash(&mfd->mfd_handle);
4917 list_move_tail(&mfd->mfd_list, &closing_list);
4919 spin_unlock(&med->med_open_lock);
4920 mdt = mdt_dev(obd->obd_lu_dev);
4921 LASSERT(mdt != NULL);
4923 rc = lu_env_init(&env, LCT_MD_THREAD);
4927 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
4928 LASSERT(info != NULL);
4929 memset(info, 0, sizeof *info);
4930 info->mti_env = &env;
4931 info->mti_mdt = mdt;
4932 info->mti_exp = exp;
4934 if (!list_empty(&closing_list)) {
4935 struct md_attr *ma = &info->mti_attr;
4937 /* Close any open files (which may also cause orphan
4939 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
4940 list_del_init(&mfd->mfd_list);
4941 ma->ma_need = ma->ma_valid = 0;
4943 /* This file is being closed due to an eviction, it
4944 * could have been modified and now dirty regarding to
4945 * HSM archive, check this!
4946 * The logic here is to mark a file dirty if there's a
4947 * chance it was dirtied before the client was evicted,
4948 * so that we don't have to wait for a release attempt
4949 * before finding out the file was actually dirty and
4950 * fail the release. Aggressively marking it dirty here
4951 * will cause the policy engine to attempt to
4952 * re-archive it; when rearchiving, we can compare the
4953 * current version to the HSM data_version and make the
4954 * archive request into a noop if it's not actually
4957 if (mfd->mfd_mode & FMODE_WRITE)
4958 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
4960 /* Don't unlink orphan on failover umount, LU-184 */
4961 if (exp->exp_flags & OBD_OPT_FAILOVER) {
4962 ma->ma_valid = MA_FLAGS;
4963 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
4965 mdt_mfd_close(info, mfd);
4968 info->mti_mdt = NULL;
4969 /* cleanup client slot early */
4970 /* Do not erase record for recoverable client. */
4971 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
4972 tgt_client_del(&env, exp);
4978 static int mdt_obd_disconnect(struct obd_export *exp)
4984 class_export_get(exp);
4986 nodemap_del_member(exp);
4987 rc = server_disconnect_export(exp);
4989 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
4991 rc = mdt_export_cleanup(exp);
4992 class_export_put(exp);
4996 /* mds_connect copy */
4997 static int mdt_obd_connect(const struct lu_env *env,
4998 struct obd_export **exp, struct obd_device *obd,
4999 struct obd_uuid *cluuid,
5000 struct obd_connect_data *data,
5003 struct obd_export *lexp;
5004 struct lustre_handle conn = { 0 };
5005 struct mdt_device *mdt;
5007 lnet_nid_t *client_nid = localdata;
5010 LASSERT(env != NULL);
5011 if (!exp || !obd || !cluuid)
5014 mdt = mdt_dev(obd->obd_lu_dev);
5017 * first, check whether the stack is ready to handle requests
5018 * XXX: probably not very appropriate method is used now
5019 * at some point we should find a better one
5021 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
5022 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5023 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5024 rc = obd_get_info(env, mdt->mdt_child_exp,
5025 sizeof(KEY_OSP_CONNECTED),
5026 KEY_OSP_CONNECTED, NULL, NULL);
5029 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5032 rc = class_connect(&conn, obd, cluuid);
5036 lexp = class_conn2export(&conn);
5037 LASSERT(lexp != NULL);
5039 rc = mdt_connect_internal(lexp, mdt, data);
5041 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5044 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5045 rc = tgt_client_new(env, lexp);
5047 rc = nodemap_add_member(*client_nid, lexp);
5048 if (rc != 0 && rc != -EEXIST)
5051 mdt_export_stats_init(obd, lexp, localdata);
5056 class_disconnect(lexp);
5065 static int mdt_obd_reconnect(const struct lu_env *env,
5066 struct obd_export *exp, struct obd_device *obd,
5067 struct obd_uuid *cluuid,
5068 struct obd_connect_data *data,
5071 lnet_nid_t *client_nid = localdata;
5075 if (exp == NULL || obd == NULL || cluuid == NULL)
5078 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5080 rc = nodemap_add_member(*client_nid, exp);
5081 if (rc == 0 || rc == -EEXIST)
5082 mdt_export_stats_init(obd, exp, localdata);
5088 /* FIXME: Can we avoid using these two interfaces? */
5089 static int mdt_init_export(struct obd_export *exp)
5091 struct mdt_export_data *med = &exp->exp_mdt_data;
5095 INIT_LIST_HEAD(&med->med_open_head);
5096 spin_lock_init(&med->med_open_lock);
5097 mutex_init(&med->med_idmap_mutex);
5098 med->med_idmap = NULL;
5099 spin_lock(&exp->exp_lock);
5100 exp->exp_connecting = 1;
5101 spin_unlock(&exp->exp_lock);
5103 /* self-export doesn't need client data and ldlm initialization */
5104 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5105 &exp->exp_client_uuid)))
5108 rc = tgt_client_alloc(exp);
5112 rc = ldlm_init_export(exp);
5119 tgt_client_free(exp);
5121 CERROR("%s: Failed to initialize export: rc = %d\n",
5122 exp->exp_obd->obd_name, rc);
5126 static int mdt_destroy_export(struct obd_export *exp)
5130 if (exp_connect_rmtclient(exp))
5131 mdt_cleanup_idmap(&exp->exp_mdt_data);
5133 target_destroy_export(exp);
5134 /* destroy can be called from failed obd_setup, so
5135 * checking uuid is safer than obd_self_export */
5136 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5137 &exp->exp_client_uuid)))
5140 ldlm_destroy_export(exp);
5141 tgt_client_free(exp);
5143 LASSERT(list_empty(&exp->exp_outstanding_replies));
5144 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5149 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5150 struct linkea_data *ldata)
5154 LASSERT(ldata->ld_buf->lb_buf != NULL);
5156 if (!mdt_object_exists(mdt_obj))
5159 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5160 ldata->ld_buf, XATTR_NAME_LINK);
5161 if (rc == -ERANGE) {
5162 /* Buf was too small, figure out what we need. */
5163 lu_buf_free(ldata->ld_buf);
5164 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5165 ldata->ld_buf, XATTR_NAME_LINK);
5168 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5169 if (ldata->ld_buf->lb_buf == NULL)
5171 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5172 ldata->ld_buf, XATTR_NAME_LINK);
5177 return linkea_init(ldata);
5181 * Given an MDT object, try to look up the full path to the object.
5182 * Part of the MDT layer implementation of lfs fid2path.
5184 * \param[in] info Per-thread common data shared by MDT level handlers.
5185 * \param[in] obj Object to do path lookup of
5186 * \param[in,out] fp User-provided struct to store path information
5188 * \retval 0 Lookup successful, path information stored in fp
5189 * \retval -EAGAIN Lookup failed, usually because object is being moved
5190 * \retval negative errno if there was a problem
5192 static int mdt_path_current(struct mdt_thread_info *info,
5193 struct mdt_object *obj,
5194 struct getinfo_fid2path *fp)
5196 struct mdt_device *mdt = info->mti_mdt;
5197 struct mdt_object *mdt_obj;
5198 struct link_ea_header *leh;
5199 struct link_ea_entry *lee;
5200 struct lu_name *tmpname = &info->mti_name;
5201 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5202 struct lu_buf *buf = &info->mti_big_buf;
5205 struct linkea_data ldata = { NULL };
5210 /* temp buffer for path element, the buffer will be finally freed
5211 * in mdt_thread_info_fini */
5212 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5213 if (buf->lb_buf == NULL)
5217 ptr = fp->gf_path + fp->gf_pathlen - 1;
5220 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
5222 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5223 * so checking root_fid can only happen on MDT0. */
5224 while (!lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid)) {
5225 struct lu_buf lmv_buf;
5227 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
5228 if (IS_ERR(mdt_obj))
5229 GOTO(out, rc = PTR_ERR(mdt_obj));
5231 if (!mdt_object_exists(mdt_obj)) {
5232 mdt_object_put(info->mti_env, mdt_obj);
5233 GOTO(out, rc = -ENOENT);
5236 if (mdt_object_remote(mdt_obj)) {
5237 mdt_object_put(info->mti_env, mdt_obj);
5238 GOTO(remote_out, rc = -EREMOTE);
5241 rc = mdt_links_read(info, mdt_obj, &ldata);
5243 mdt_object_put(info->mti_env, mdt_obj);
5248 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5249 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5250 /* If set, use link #linkno for path lookup, otherwise use
5251 link #0. Only do this for the final path element. */
5252 if (first && fp->gf_linkno < leh->leh_reccount) {
5254 for (count = 0; count < fp->gf_linkno; count++) {
5255 lee = (struct link_ea_entry *)
5256 ((char *)lee + reclen);
5257 linkea_entry_unpack(lee, &reclen, tmpname,
5260 if (fp->gf_linkno < leh->leh_reccount - 1)
5261 /* indicate to user there are more links */
5265 lmv_buf.lb_buf = info->mti_xattr_buf;
5266 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
5267 /* Check if it is slave stripes */
5268 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5269 &lmv_buf, XATTR_NAME_LMV);
5270 mdt_object_put(info->mti_env, mdt_obj);
5272 union lmv_mds_md *lmm = lmv_buf.lb_buf;
5274 /* For slave stripes, get its master */
5275 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
5276 fp->gf_fid = *tmpfid;
5279 } else if (rc < 0 && rc != -ENODATA) {
5285 /* Pack the name in the end of the buffer */
5286 ptr -= tmpname->ln_namelen;
5287 if (ptr - 1 <= fp->gf_path)
5288 GOTO(out, rc = -EOVERFLOW);
5289 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5292 /* keep the last resolved fid to the client, so the
5293 * client will build the left path on another MDT for
5295 fp->gf_fid = *tmpfid;
5301 ptr++; /* skip leading / */
5302 memmove(fp->gf_path, ptr, fp->gf_path + fp->gf_pathlen - ptr);
5309 * Given an MDT object, use mdt_path_current to get the path.
5310 * Essentially a wrapper to retry mdt_path_current a set number of times
5311 * if -EAGAIN is returned (usually because an object is being moved).
5313 * Part of the MDT layer implementation of lfs fid2path.
5315 * \param[in] info Per-thread common data shared by mdt level handlers.
5316 * \param[in] obj Object to do path lookup of
5317 * \param[in,out] fp User-provided struct for arguments and to store path
5320 * \retval 0 Lookup successful, path information stored in fp
5321 * \retval negative errno if there was a problem
5323 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5324 struct getinfo_fid2path *fp)
5326 struct mdt_device *mdt = info->mti_mdt;
5331 if (fp->gf_pathlen < 3)
5334 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5335 fp->gf_path[0] = '\0';
5339 /* Retry multiple times in case file is being moved */
5340 while (tries-- && rc == -EAGAIN)
5341 rc = mdt_path_current(info, obj, fp);
5347 * Get the full path of the provided FID, as of changelog record recno.
5349 * This checks sanity and looks up object for user provided FID
5350 * before calling the actual path lookup code.
5352 * Part of the MDT layer implementation of lfs fid2path.
5354 * \param[in] info Per-thread common data shared by mdt level handlers.
5355 * \param[in,out] fp User-provided struct for arguments and to store path
5358 * \retval 0 Lookup successful, path information and recno stored in fp
5359 * \retval -ENOENT, object does not exist
5360 * \retval negative errno if there was a problem
5362 static int mdt_fid2path(struct mdt_thread_info *info,
5363 struct getinfo_fid2path *fp)
5365 struct mdt_device *mdt = info->mti_mdt;
5366 struct mdt_object *obj;
5370 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5371 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5373 if (!fid_is_sane(&fp->gf_fid))
5376 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5377 CWARN("%s: "DFID" is invalid, sequence should be "
5378 ">= "LPX64"\n", mdt_obd_name(mdt),
5379 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5383 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5386 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
5387 PFID(&fp->gf_fid), rc);
5391 if (mdt_object_remote(obj))
5393 else if (!mdt_object_exists(obj))
5399 mdt_object_put(info->mti_env, obj);
5400 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
5401 PFID(&fp->gf_fid), rc);
5405 rc = mdt_path(info, obj, fp);
5407 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5408 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5410 mdt_object_put(info->mti_env, obj);
5415 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5416 void *val, int vallen)
5418 struct getinfo_fid2path *fpout, *fpin;
5421 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5424 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5425 lustre_swab_fid2path(fpin);
5427 memcpy(fpout, fpin, sizeof(*fpin));
5428 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5431 rc = mdt_fid2path(info, fpout);
5435 int mdt_get_info(struct tgt_session_info *tsi)
5445 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5447 CDEBUG(D_IOCTL, "No GETINFO key\n");
5448 RETURN(err_serious(-EFAULT));
5450 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5453 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5454 if (vallen == NULL) {
5455 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
5456 tgt_name(tsi->tsi_tgt));
5457 RETURN(err_serious(-EFAULT));
5460 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5462 rc = req_capsule_server_pack(tsi->tsi_pill);
5464 RETURN(err_serious(rc));
5466 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5467 if (valout == NULL) {
5468 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
5469 tgt_name(tsi->tsi_tgt));
5470 RETURN(err_serious(-EFAULT));
5473 if (KEY_IS(KEY_FID2PATH)) {
5474 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5476 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5477 mdt_thread_info_fini(info);
5484 /* Pass the ioc down */
5485 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5486 unsigned int cmd, int len, void *data)
5488 struct lu_context ioctl_session;
5489 struct md_device *next = mdt->mdt_child;
5493 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5496 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5497 lu_context_enter(&ioctl_session);
5498 env->le_ses = &ioctl_session;
5500 LASSERT(next->md_ops->mdo_iocontrol);
5501 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5503 lu_context_exit(&ioctl_session);
5504 lu_context_fini(&ioctl_session);
5508 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5510 struct obd_ioctl_data *data = karg;
5513 struct mdt_object *obj;
5514 struct mdt_lock_handle *lh;
5518 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5519 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5522 fid = (struct lu_fid *)data->ioc_inlbuf1;
5524 if (!fid_is_sane(fid))
5527 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5529 lh = &mti->mti_lh[MDT_LH_PARENT];
5530 mdt_lock_reg_init(lh, LCK_CR);
5532 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5534 RETURN(PTR_ERR(obj));
5536 if (mdt_object_remote(obj)) {
5539 * before calling version get the correct MDS should be
5540 * fid, this is error to find remote object here
5542 CERROR("nonlocal object "DFID"\n", PFID(fid));
5543 } else if (!mdt_object_exists(obj)) {
5544 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5547 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5548 *(__u64 *)data->ioc_inlbuf2 = version;
5551 mdt_object_unlock_put(mti, obj, lh, 1);
5555 /* ioctls on obd dev */
5556 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5557 void *karg, void *uarg)
5560 struct obd_device *obd = exp->exp_obd;
5561 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5562 struct dt_device *dt = mdt->mdt_bottom;
5566 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5567 rc = lu_env_init(&env, LCT_MD_THREAD);
5573 rc = mdt_device_sync(&env, mdt);
5575 case OBD_IOC_SET_READONLY:
5576 rc = dt->dd_ops->dt_ro(&env, dt);
5578 case OBD_IOC_ABORT_RECOVERY:
5579 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5580 obd->obd_force_abort_recovery = 1;
5581 target_stop_recovery_thread(obd);
5584 case OBD_IOC_CHANGELOG_REG:
5585 case OBD_IOC_CHANGELOG_DEREG:
5586 case OBD_IOC_CHANGELOG_CLEAR:
5587 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5589 case OBD_IOC_START_LFSCK: {
5590 struct md_device *next = mdt->mdt_child;
5591 struct obd_ioctl_data *data = karg;
5592 struct lfsck_start_param lsp;
5594 if (unlikely(data == NULL)) {
5599 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5600 lsp.lsp_index_valid = 0;
5601 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5604 case OBD_IOC_STOP_LFSCK: {
5605 struct md_device *next = mdt->mdt_child;
5606 struct obd_ioctl_data *data = karg;
5607 struct lfsck_stop stop;
5609 stop.ls_status = LS_STOPPED;
5610 /* Old lfsck utils may pass NULL @stop. */
5611 if (data->ioc_inlbuf1 == NULL)
5615 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5617 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5620 case OBD_IOC_GET_OBJ_VERSION: {
5621 struct mdt_thread_info *mti;
5622 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5623 memset(mti, 0, sizeof *mti);
5624 mti->mti_env = &env;
5628 rc = mdt_ioc_version_get(mti, karg);
5631 case OBD_IOC_CATLOGLIST: {
5632 struct mdt_thread_info *mti;
5634 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5635 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5636 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5637 &mti->mti_tmp_fid1);
5642 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5643 mdt_obd_name(mdt), cmd, rc);
5650 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5652 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5653 struct lfsck_start_param lsp;
5657 lsp.lsp_start = NULL;
5658 lsp.lsp_index_valid = 0;
5659 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
5660 OBD_IOC_START_LFSCK,
5662 if (rc != 0 && rc != -EALREADY)
5663 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
5664 mdt_obd_name(mdt), rc);
5666 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5670 static int mdt_obd_postrecov(struct obd_device *obd)
5675 rc = lu_env_init(&env, LCT_MD_THREAD);
5678 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5683 static struct obd_ops mdt_obd_device_ops = {
5684 .o_owner = THIS_MODULE,
5685 .o_set_info_async = mdt_obd_set_info_async,
5686 .o_connect = mdt_obd_connect,
5687 .o_reconnect = mdt_obd_reconnect,
5688 .o_disconnect = mdt_obd_disconnect,
5689 .o_init_export = mdt_init_export,
5690 .o_destroy_export = mdt_destroy_export,
5691 .o_iocontrol = mdt_iocontrol,
5692 .o_postrecov = mdt_obd_postrecov,
5695 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5696 struct lu_device *d)
5698 struct mdt_device *m = mdt_dev(d);
5705 static struct lu_device *mdt_device_free(const struct lu_env *env,
5706 struct lu_device *d)
5708 struct mdt_device *m = mdt_dev(d);
5711 lu_device_fini(&m->mdt_lu_dev);
5717 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5718 struct lu_device_type *t,
5719 struct lustre_cfg *cfg)
5721 struct lu_device *l;
5722 struct mdt_device *m;
5729 rc = mdt_init0(env, m, t, cfg);
5731 mdt_device_free(env, l);
5736 l = ERR_PTR(-ENOMEM);
5740 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5741 LU_KEY_INIT(mdt, struct mdt_thread_info);
5743 static void mdt_key_fini(const struct lu_context *ctx,
5744 struct lu_context_key *key, void* data)
5746 struct mdt_thread_info *info = data;
5748 if (info->mti_big_lmm) {
5749 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5750 info->mti_big_lmm = NULL;
5751 info->mti_big_lmmsize = 0;
5756 /* context key: mdt_thread_key */
5757 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5759 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5761 return lu_ucred(info->mti_env);
5764 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5766 return lu_ucred_check(info->mti_env);
5770 * Enable/disable COS (Commit On Sharing).
5772 * Set/Clear the COS flag in mdt options.
5774 * \param mdt mdt device
5775 * \param val 0 disables COS, other values enable COS
5777 void mdt_enable_cos(struct mdt_device *mdt, int val)
5782 mdt->mdt_opts.mo_cos = !!val;
5783 rc = lu_env_init(&env, LCT_LOCAL);
5784 if (unlikely(rc != 0)) {
5785 CWARN("%s: lu_env initialization failed, cannot "
5786 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
5789 mdt_device_sync(&env, mdt);
5794 * Check COS (Commit On Sharing) status.
5796 * Return COS flag status.
5798 * \param mdt mdt device
5800 int mdt_cos_is_enabled(struct mdt_device *mdt)
5802 return mdt->mdt_opts.mo_cos != 0;
5805 static struct lu_device_type_operations mdt_device_type_ops = {
5806 .ldto_device_alloc = mdt_device_alloc,
5807 .ldto_device_free = mdt_device_free,
5808 .ldto_device_fini = mdt_device_fini
5811 static struct lu_device_type mdt_device_type = {
5812 .ldt_tags = LU_DEVICE_MD,
5813 .ldt_name = LUSTRE_MDT_NAME,
5814 .ldt_ops = &mdt_device_type_ops,
5815 .ldt_ctx_tags = LCT_MD_THREAD
5818 static int __init mdt_mod_init(void)
5822 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5823 FID_NOBRACE_LEN + 1);
5824 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5826 rc = lu_kmem_init(mdt_caches);
5830 rc = mds_mod_init();
5834 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
5835 LUSTRE_MDT_NAME, &mdt_device_type);
5840 lu_kmem_fini(mdt_caches);
5847 static void __exit mdt_mod_exit(void)
5849 class_unregister_type(LUSTRE_MDT_NAME);
5851 lu_kmem_fini(mdt_caches);
5854 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
5855 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5856 MODULE_VERSION(LUSTRE_VERSION_STRING);
5857 MODULE_LICENSE("GPL");
5859 module_init(mdt_mod_init);
5860 module_exit(mdt_mod_exit);