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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdt/mdt_handler.c
34 * Lustre Metadata Target (mdt) request handler
36 * Author: Peter Braam <braam@clusterfs.com>
37 * Author: Andreas Dilger <adilger@clusterfs.com>
38 * Author: Phil Schwan <phil@clusterfs.com>
39 * Author: Mike Shaver <shaver@clusterfs.com>
40 * Author: Nikita Danilov <nikita@clusterfs.com>
41 * Author: Huang Hua <huanghua@clusterfs.com>
42 * Author: Yury Umanets <umka@clusterfs.com>
45 #define DEBUG_SUBSYSTEM S_MDS
47 #include <linux/module.h>
48 #include <linux/pagemap.h>
50 #include <dt_object.h>
51 #include <lustre_acl.h>
52 #include <lustre_export.h>
53 #include <uapi/linux/lustre/lustre_ioctl.h>
54 #include <lustre_lfsck.h>
55 #include <lustre_log.h>
56 #include <lustre_nodemap.h>
57 #include <lustre_mds.h>
58 #include <uapi/linux/lustre/lustre_param.h>
59 #include <lustre_quota.h>
60 #include <lustre_swab.h>
62 #include <obd_support.h>
63 #include <lustre_barrier.h>
64 #include <obd_cksum.h>
65 #include <llog_swab.h>
67 #include "mdt_internal.h"
69 static unsigned int max_mod_rpcs_per_client = 8;
70 module_param(max_mod_rpcs_per_client, uint, 0644);
71 MODULE_PARM_DESC(max_mod_rpcs_per_client, "maximum number of modify RPCs in flight allowed per client");
73 mdl_mode_t mdt_mdl_lock_modes[] = {
74 [LCK_MINMODE] = MDL_MINMODE,
81 [LCK_GROUP] = MDL_GROUP
84 enum ldlm_mode mdt_dlm_lock_modes[] = {
85 [MDL_MINMODE] = LCK_MINMODE,
92 [MDL_GROUP] = LCK_GROUP
95 static struct mdt_device *mdt_dev(struct lu_device *d);
97 static const struct lu_object_operations mdt_obj_ops;
99 /* Slab for MDT object allocation */
100 static struct kmem_cache *mdt_object_kmem;
102 /* For HSM restore handles */
103 struct kmem_cache *mdt_hsm_cdt_kmem;
105 /* For HSM request handles */
106 struct kmem_cache *mdt_hsm_car_kmem;
108 static struct lu_kmem_descr mdt_caches[] = {
110 .ckd_cache = &mdt_object_kmem,
111 .ckd_name = "mdt_obj",
112 .ckd_size = sizeof(struct mdt_object)
115 .ckd_cache = &mdt_hsm_cdt_kmem,
116 .ckd_name = "mdt_cdt_restore_handle",
117 .ckd_size = sizeof(struct cdt_restore_handle)
120 .ckd_cache = &mdt_hsm_car_kmem,
121 .ckd_name = "mdt_cdt_agent_req",
122 .ckd_size = sizeof(struct cdt_agent_req)
129 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
133 return rep->lock_policy_res1 & op_flag;
136 void mdt_clear_disposition(struct mdt_thread_info *info,
137 struct ldlm_reply *rep, __u64 op_flag)
140 info->mti_opdata &= ~op_flag;
141 tgt_opdata_clear(info->mti_env, op_flag);
144 rep->lock_policy_res1 &= ~op_flag;
147 void mdt_set_disposition(struct mdt_thread_info *info,
148 struct ldlm_reply *rep, __u64 op_flag)
151 info->mti_opdata |= op_flag;
152 tgt_opdata_set(info->mti_env, op_flag);
155 rep->lock_policy_res1 |= op_flag;
158 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
160 lh->mlh_pdo_hash = 0;
161 lh->mlh_reg_mode = lm;
162 lh->mlh_rreg_mode = lm;
163 lh->mlh_type = MDT_REG_LOCK;
166 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
167 const struct lu_name *lname)
169 lh->mlh_reg_mode = lock_mode;
170 lh->mlh_pdo_mode = LCK_MINMODE;
171 lh->mlh_rreg_mode = lock_mode;
172 lh->mlh_type = MDT_PDO_LOCK;
174 if (lu_name_is_valid(lname)) {
175 lh->mlh_pdo_hash = ll_full_name_hash(NULL, lname->ln_name,
177 /* XXX Workaround for LU-2856
179 * Zero is a valid return value of full_name_hash, but
180 * several users of mlh_pdo_hash assume a non-zero
181 * hash value. We therefore map zero onto an
182 * arbitrary, but consistent value (1) to avoid
183 * problems further down the road. */
184 if (unlikely(lh->mlh_pdo_hash == 0))
185 lh->mlh_pdo_hash = 1;
187 lh->mlh_pdo_hash = 0;
191 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
192 struct mdt_lock_handle *lh)
198 * Any dir access needs couple of locks:
200 * 1) on part of dir we gonna take lookup/modify;
202 * 2) on whole dir to protect it from concurrent splitting and/or to
203 * flush client's cache for readdir().
205 * so, for a given mode and object this routine decides what lock mode
206 * to use for lock #2:
208 * 1) if caller's gonna lookup in dir then we need to protect dir from
209 * being splitted only - LCK_CR
211 * 2) if caller's gonna modify dir then we need to protect dir from
212 * being splitted and to flush cache - LCK_CW
214 * 3) if caller's gonna modify dir and that dir seems ready for
215 * splitting then we need to protect it from any type of access
216 * (lookup/modify/split) - LCK_EX --bzzz
219 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
220 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
223 * Ask underlaying level its opinion about preferable PDO lock mode
224 * having access type passed as regular lock mode:
226 * - MDL_MINMODE means that lower layer does not want to specify lock
229 * - MDL_NL means that no PDO lock should be taken. This is used in some
230 * cases. Say, for non-splittable directories no need to use PDO locks
233 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
234 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
236 if (mode != MDL_MINMODE) {
237 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
240 * Lower layer does not want to specify locking mode. We do it
241 * our selves. No special protection is needed, just flush
242 * client's cache on modification and allow concurrent
245 switch (lh->mlh_reg_mode) {
247 lh->mlh_pdo_mode = LCK_EX;
250 lh->mlh_pdo_mode = LCK_CR;
253 lh->mlh_pdo_mode = LCK_CW;
256 CERROR("Not expected lock type (0x%x)\n",
257 (int)lh->mlh_reg_mode);
262 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
266 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
269 struct mdt_device *mdt = info->mti_mdt;
270 struct lu_name *lname = &info->mti_name;
271 char *filename = info->mti_filename;
272 struct mdt_object *parent;
276 LASSERT(!info->mti_cross_ref);
279 * We may want to allow this to mount a completely separate
280 * fileset from the MDT in the future, but keeping it to
281 * ROOT/ only for now avoid potential security issues.
283 *fid = mdt->mdt_md_root_fid;
285 while (rc == 0 && fileset != NULL && *fileset != '\0') {
286 const char *s1 = fileset;
292 while (*s2 != '/' && *s2 != '\0')
300 lname->ln_namelen = s2 - s1;
301 if (lname->ln_namelen > NAME_MAX) {
306 /* reject .. as a path component */
307 if (lname->ln_namelen == 2 &&
308 strncmp(s1, "..", 2) == 0) {
313 strncpy(filename, s1, lname->ln_namelen);
314 filename[lname->ln_namelen] = '\0';
315 lname->ln_name = filename;
317 parent = mdt_object_find(info->mti_env, mdt, fid);
318 if (IS_ERR(parent)) {
319 rc = PTR_ERR(parent);
322 /* Only got the fid of this obj by name */
324 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
325 fid, &info->mti_spec);
326 mdt_object_put(info->mti_env, parent);
329 parent = mdt_object_find(info->mti_env, mdt, fid);
331 rc = PTR_ERR(parent);
333 mode = lu_object_attr(&parent->mot_obj);
334 mdt_object_put(info->mti_env, parent);
343 static int mdt_get_root(struct tgt_session_info *tsi)
345 struct mdt_thread_info *info = tsi2mdt_info(tsi);
346 struct mdt_device *mdt = info->mti_mdt;
347 struct mdt_body *repbody;
348 char *fileset = NULL, *buffer = NULL;
350 struct obd_export *exp = info->mti_exp;
351 char *nodemap_fileset;
355 rc = mdt_check_ucred(info);
357 GOTO(out, rc = err_serious(rc));
359 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
360 GOTO(out, rc = err_serious(-ENOMEM));
362 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
363 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
364 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
366 GOTO(out, rc = err_serious(-EFAULT));
369 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
370 if (nodemap_fileset && nodemap_fileset[0]) {
371 CDEBUG(D_INFO, "nodemap fileset is %s\n", nodemap_fileset);
373 /* consider fileset from client as a sub-fileset
374 * of the nodemap one */
375 OBD_ALLOC(buffer, PATH_MAX + 1);
377 GOTO(out, rc = err_serious(-ENOMEM));
378 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
379 nodemap_fileset, fileset) >= PATH_MAX + 1)
380 GOTO(out, rc = err_serious(-EINVAL));
383 /* enforce fileset as specified in the nodemap */
384 fileset = nodemap_fileset;
389 CDEBUG(D_INFO, "Getting fileset %s\n", fileset);
390 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
392 GOTO(out, rc = err_serious(rc));
394 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
396 repbody->mbo_valid |= OBD_MD_FLID;
400 mdt_thread_info_fini(info);
402 OBD_FREE(buffer, PATH_MAX+1);
406 static int mdt_statfs(struct tgt_session_info *tsi)
408 struct ptlrpc_request *req = tgt_ses_req(tsi);
409 struct mdt_thread_info *info = tsi2mdt_info(tsi);
410 struct mdt_device *mdt = info->mti_mdt;
411 struct tg_grants_data *tgd = &mdt->mdt_lut.lut_tgd;
412 struct md_device *next = mdt->mdt_child;
413 struct ptlrpc_service_part *svcpt;
414 struct obd_statfs *osfs;
415 struct mdt_body *reqbody = NULL;
416 struct mdt_statfs_cache *msf;
421 svcpt = req->rq_rqbd->rqbd_svcpt;
423 /* This will trigger a watchdog timeout */
424 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
425 (MDT_SERVICE_WATCHDOG_FACTOR *
426 at_get(&svcpt->scp_at_estimate)) + 1);
428 rc = mdt_check_ucred(info);
430 GOTO(out, rc = err_serious(rc));
432 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
433 GOTO(out, rc = err_serious(-ENOMEM));
435 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
437 GOTO(out, rc = -EPROTO);
439 if (mdt_is_sum_statfs_client(req->rq_export))
440 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
442 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
443 msf = &mdt->mdt_sum_osfs;
445 msf = &mdt->mdt_osfs;
447 if (msf->msf_age + OBD_STATFS_CACHE_SECONDS <= ktime_get_seconds()) {
448 /** statfs data is too old, get up-to-date one */
449 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
450 rc = next->md_ops->mdo_statfs(info->mti_env,
453 rc = dt_statfs(info->mti_env, mdt->mdt_bottom,
457 spin_lock(&mdt->mdt_lock);
458 msf->msf_osfs = *osfs;
459 msf->msf_age = ktime_get_seconds();
460 spin_unlock(&mdt->mdt_lock);
462 /** use cached statfs data */
463 spin_lock(&mdt->mdt_lock);
464 *osfs = msf->msf_osfs;
465 spin_unlock(&mdt->mdt_lock);
468 /* at least try to account for cached pages. its still racy and
469 * might be under-reporting if clients haven't announced their
470 * caches with brw recently */
471 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
472 " pending %llu free %llu avail %llu\n",
473 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
474 tgd->tgd_tot_pending,
475 osfs->os_bfree << tgd->tgd_blockbits,
476 osfs->os_bavail << tgd->tgd_blockbits);
478 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
479 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
480 osfs->os_bsize - 1) >> tgd->tgd_blockbits));
482 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
483 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
484 "%llu objects: %llu free; state %x\n",
485 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
486 osfs->os_files, osfs->os_ffree, osfs->os_state);
488 if (!exp_grant_param_supp(tsi->tsi_exp) &&
489 tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT) {
490 /* clients which don't support OBD_CONNECT_GRANT_PARAM
491 * should not see a block size > page size, otherwise
492 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
493 * block size which is the biggest block size known to work
494 * with all client's page size. */
495 osfs->os_blocks <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
496 osfs->os_bfree <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
497 osfs->os_bavail <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
498 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
501 mdt_counter_incr(req, LPROC_MDT_STATFS);
503 mdt_thread_info_fini(info);
508 * Pack size attributes into the reply.
510 int mdt_pack_size2body(struct mdt_thread_info *info,
511 const struct lu_fid *fid, struct lustre_handle *lh)
514 struct md_attr *ma = &info->mti_attr;
516 bool dom_lock = false;
520 LASSERT(ma->ma_attr.la_valid & LA_MODE);
522 if (!S_ISREG(ma->ma_attr.la_mode) ||
523 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
526 dom_stripe = mdt_lmm_dom_entry(ma->ma_lmm);
527 /* no DoM stripe, no size in reply */
528 if (dom_stripe == LMM_NO_DOM)
531 if (lustre_handle_is_used(lh)) {
532 struct ldlm_lock *lock;
534 lock = ldlm_handle2lock(lh);
536 dom_lock = ldlm_has_dom(lock);
541 /* no DoM lock, no size in reply */
545 /* Either DoM lock exists or LMM has only DoM stripe then
546 * return size on body. */
547 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
549 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
553 #ifdef CONFIG_FS_POSIX_ACL
555 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
557 * \param info thread info object
558 * \param repbody reply to pack ACLs into
559 * \param o mdt object of file to examine
560 * \param nodemap nodemap of client to reply to
562 * \retval -errno error getting or parsing ACL from disk
564 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
565 struct mdt_object *o, struct lu_nodemap *nodemap)
567 const struct lu_env *env = info->mti_env;
568 struct md_object *next = mdt_object_child(o);
569 struct lu_buf *buf = &info->mti_buf;
570 struct mdt_device *mdt = info->mti_mdt;
571 struct req_capsule *pill = info->mti_pill;
576 buf->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
577 buf->lb_len = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
578 if (buf->lb_len == 0)
582 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
584 if (rc == -ENODATA) {
585 repbody->mbo_aclsize = 0;
586 repbody->mbo_valid |= OBD_MD_FLACL;
588 } else if (rc == -EOPNOTSUPP) {
592 exp_connect_large_acl(info->mti_exp) &&
593 buf->lb_buf != info->mti_big_acl) {
594 if (info->mti_big_acl == NULL) {
595 OBD_ALLOC_LARGE(info->mti_big_acl,
596 mdt->mdt_max_ea_size);
597 if (info->mti_big_acl == NULL) {
598 CERROR("%s: unable to grow "
601 PFID(mdt_object_fid(o)));
605 info->mti_big_aclsize =
606 mdt->mdt_max_ea_size;
609 CDEBUG(D_INODE, "%s: grow the "DFID
610 " ACL buffer to size %d\n",
612 PFID(mdt_object_fid(o)),
613 mdt->mdt_max_ea_size);
615 buf->lb_buf = info->mti_big_acl;
616 buf->lb_len = info->mti_big_aclsize;
621 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
622 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
631 acl_buflen = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
632 if (acl_buflen >= rc)
635 /* If LOV/LMA EA is small, we can reuse part of their buffer */
636 client = ptlrpc_req_get_repsize(pill->rc_req);
637 server = lustre_packed_msg_size(pill->rc_req->rq_repmsg);
638 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER)) {
639 lmm_buflen = req_capsule_get_size(pill, &RMF_MDT_MD,
641 lmmsize = repbody->mbo_eadatasize;
644 if (client < server - acl_buflen - lmm_buflen + rc + lmmsize) {
645 CDEBUG(D_INODE, "%s: client prepared buffer size %d "
646 "is not big enough with the ACL size %d (%d)\n",
647 mdt_obd_name(mdt), client, rc,
648 server - acl_buflen - lmm_buflen + rc + lmmsize);
649 repbody->mbo_aclsize = 0;
650 repbody->mbo_valid &= ~OBD_MD_FLACL;
655 if (buf->lb_buf == info->mti_big_acl)
656 info->mti_big_acl_used = 1;
658 rc = nodemap_map_acl(nodemap, buf->lb_buf,
659 rc, NODEMAP_FS_TO_CLIENT);
660 /* if all ACLs mapped out, rc is still >= 0 */
662 CERROR("%s: nodemap_map_acl unable to parse "DFID
663 " ACL: rc = %d\n", mdt_obd_name(mdt),
664 PFID(mdt_object_fid(o)), rc);
665 repbody->mbo_aclsize = 0;
666 repbody->mbo_valid &= ~OBD_MD_FLACL;
668 repbody->mbo_aclsize = rc;
669 repbody->mbo_valid |= OBD_MD_FLACL;
678 /* XXX Look into layout in MDT layer. */
679 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
681 struct lov_comp_md_v1 *comp_v1;
682 struct lov_mds_md *v1;
685 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
686 comp_v1 = (struct lov_comp_md_v1 *)lmm;
688 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
689 v1 = (struct lov_mds_md *)((char *)comp_v1 +
690 comp_v1->lcm_entries[i].lcme_offset);
691 /* We don't support partial release for now */
692 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
697 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
702 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
703 const struct lu_attr *attr, const struct lu_fid *fid)
705 struct md_attr *ma = &info->mti_attr;
706 struct obd_export *exp = info->mti_exp;
707 struct lu_nodemap *nodemap = NULL;
709 LASSERT(ma->ma_valid & MA_INODE);
711 if (attr->la_valid & LA_ATIME) {
712 b->mbo_atime = attr->la_atime;
713 b->mbo_valid |= OBD_MD_FLATIME;
715 if (attr->la_valid & LA_MTIME) {
716 b->mbo_mtime = attr->la_mtime;
717 b->mbo_valid |= OBD_MD_FLMTIME;
719 if (attr->la_valid & LA_CTIME) {
720 b->mbo_ctime = attr->la_ctime;
721 b->mbo_valid |= OBD_MD_FLCTIME;
723 if (attr->la_valid & LA_FLAGS) {
724 b->mbo_flags = attr->la_flags;
725 b->mbo_valid |= OBD_MD_FLFLAGS;
727 if (attr->la_valid & LA_NLINK) {
728 b->mbo_nlink = attr->la_nlink;
729 b->mbo_valid |= OBD_MD_FLNLINK;
731 if (attr->la_valid & (LA_UID|LA_GID)) {
732 nodemap = nodemap_get_from_exp(exp);
736 if (attr->la_valid & LA_UID) {
737 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
738 NODEMAP_FS_TO_CLIENT,
740 b->mbo_valid |= OBD_MD_FLUID;
742 if (attr->la_valid & LA_GID) {
743 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
744 NODEMAP_FS_TO_CLIENT,
746 b->mbo_valid |= OBD_MD_FLGID;
749 if (attr->la_valid & LA_PROJID) {
750 /* TODO, nodemap for project id */
751 b->mbo_projid = attr->la_projid;
752 b->mbo_valid |= OBD_MD_FLPROJID;
755 b->mbo_mode = attr->la_mode;
756 if (attr->la_valid & LA_MODE)
757 b->mbo_valid |= OBD_MD_FLMODE;
758 if (attr->la_valid & LA_TYPE)
759 b->mbo_valid |= OBD_MD_FLTYPE;
763 b->mbo_valid |= OBD_MD_FLID;
764 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
765 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
768 if (!(attr->la_valid & LA_TYPE))
771 b->mbo_rdev = attr->la_rdev;
772 b->mbo_size = attr->la_size;
773 b->mbo_blocks = attr->la_blocks;
775 if (!S_ISREG(attr->la_mode)) {
776 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
777 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
778 /* means no objects are allocated on osts. */
779 LASSERT(!(ma->ma_valid & MA_LOV));
780 /* just ignore blocks occupied by extend attributes on MDS */
782 /* if no object is allocated on osts, the size on mds is valid.
784 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
785 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
786 if (mdt_hsm_is_released(ma->ma_lmm)) {
787 /* A released file stores its size on MDS. */
788 /* But return 1 block for released file, unless tools
789 * like tar will consider it fully sparse. (LU-3864)
791 if (unlikely(b->mbo_size == 0))
795 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
796 } else if (info->mti_som_valid) { /* som is valid */
797 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
801 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
802 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
803 PFID(fid), (unsigned long long)b->mbo_size);
806 if (!IS_ERR_OR_NULL(nodemap))
807 nodemap_putref(nodemap);
810 static inline int mdt_body_has_lov(const struct lu_attr *la,
811 const struct mdt_body *body)
813 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
814 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
817 void mdt_client_compatibility(struct mdt_thread_info *info)
819 struct mdt_body *body;
820 struct ptlrpc_request *req = mdt_info_req(info);
821 struct obd_export *exp = req->rq_export;
822 struct md_attr *ma = &info->mti_attr;
823 struct lu_attr *la = &ma->ma_attr;
826 if (exp_connect_layout(exp))
827 /* the client can deal with 16-bit lmm_stripe_count */
830 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
832 if (!mdt_body_has_lov(la, body))
835 /* now we have a reply with a lov for a client not compatible with the
836 * layout lock so we have to clean the layout generation number */
837 if (S_ISREG(la->la_mode))
838 ma->ma_lmm->lmm_layout_gen = 0;
842 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
843 struct mdt_object *o)
845 const struct lu_env *env = info->mti_env;
848 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
857 /* Is it a directory? Let's check for the LMV as well */
858 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
859 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
863 rc2 = mo_xattr_get(env, mdt_object_child(o),
865 XATTR_NAME_DEFAULT_LMV);
867 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
875 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
878 const struct lu_env *env = info->mti_env;
882 LASSERT(info->mti_big_lmm_used == 0);
883 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
887 /* big_lmm may need to be grown */
888 if (info->mti_big_lmmsize < rc) {
889 int size = size_roundup_power2(rc);
891 if (info->mti_big_lmmsize > 0) {
892 /* free old buffer */
893 LASSERT(info->mti_big_lmm);
894 OBD_FREE_LARGE(info->mti_big_lmm,
895 info->mti_big_lmmsize);
896 info->mti_big_lmm = NULL;
897 info->mti_big_lmmsize = 0;
900 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
901 if (info->mti_big_lmm == NULL)
903 info->mti_big_lmmsize = size;
905 LASSERT(info->mti_big_lmmsize >= rc);
907 info->mti_buf.lb_buf = info->mti_big_lmm;
908 info->mti_buf.lb_len = info->mti_big_lmmsize;
909 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
914 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
915 struct md_attr *ma, const char *name)
917 struct md_object *next = mdt_object_child(o);
918 struct lu_buf *buf = &info->mti_buf;
921 if (strcmp(name, XATTR_NAME_LOV) == 0) {
922 buf->lb_buf = ma->ma_lmm;
923 buf->lb_len = ma->ma_lmm_size;
924 LASSERT(!(ma->ma_valid & MA_LOV));
925 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
926 buf->lb_buf = ma->ma_lmv;
927 buf->lb_len = ma->ma_lmv_size;
928 LASSERT(!(ma->ma_valid & MA_LMV));
929 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
930 buf->lb_buf = ma->ma_lmv;
931 buf->lb_len = ma->ma_lmv_size;
932 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
937 LASSERT(buf->lb_buf);
939 rc = mo_xattr_get(info->mti_env, next, buf, name);
943 if (strcmp(name, XATTR_NAME_LOV) == 0) {
944 if (info->mti_big_lmm_used)
945 ma->ma_lmm = info->mti_big_lmm;
947 /* NOT return LOV EA with hole to old client. */
948 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
949 LOV_PATTERN_F_HOLE) &&
950 !(exp_connect_flags(info->mti_exp) &
951 OBD_CONNECT_LFSCK)) {
954 ma->ma_lmm_size = rc;
955 ma->ma_valid |= MA_LOV;
957 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
958 if (info->mti_big_lmm_used)
959 ma->ma_lmv = info->mti_big_lmm;
961 ma->ma_lmv_size = rc;
962 ma->ma_valid |= MA_LMV;
963 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
964 ma->ma_lmv_size = rc;
965 ma->ma_valid |= MA_LMV_DEF;
968 /* Update mdt_max_mdsize so all clients will be aware that */
969 if (info->mti_mdt->mdt_max_mdsize < rc)
970 info->mti_mdt->mdt_max_mdsize = rc;
973 } else if (rc == -ENODATA) {
976 } else if (rc == -ERANGE) {
977 /* Default LMV has fixed size, so it must be able to fit
978 * in the original buffer */
979 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
981 rc = mdt_big_xattr_get(info, o, name);
983 info->mti_big_lmm_used = 1;
991 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
994 struct lu_buf *buf = &info->mti_buf;
995 struct link_ea_header *leh;
996 struct link_ea_entry *lee;
1000 buf->lb_buf = info->mti_big_lmm;
1001 buf->lb_len = info->mti_big_lmmsize;
1002 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1003 buf, XATTR_NAME_LINK);
1004 /* ignore errors, MA_PFID won't be set and it is
1005 * up to the caller to treat this as an error */
1006 if (rc == -ERANGE || buf->lb_len == 0) {
1007 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1008 buf->lb_buf = info->mti_big_lmm;
1009 buf->lb_len = info->mti_big_lmmsize;
1014 if (rc < sizeof(*leh)) {
1015 CERROR("short LinkEA on "DFID": rc = %d\n",
1016 PFID(mdt_object_fid(o)), rc);
1020 leh = (struct link_ea_header *) buf->lb_buf;
1021 lee = (struct link_ea_entry *)(leh + 1);
1022 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1023 leh->leh_magic = LINK_EA_MAGIC;
1024 leh->leh_reccount = __swab32(leh->leh_reccount);
1025 leh->leh_len = __swab64(leh->leh_len);
1027 if (leh->leh_magic != LINK_EA_MAGIC)
1029 if (leh->leh_reccount == 0)
1032 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1033 fid_be_to_cpu(pfid, pfid);
1038 int mdt_attr_get_complex(struct mdt_thread_info *info,
1039 struct mdt_object *o, struct md_attr *ma)
1041 const struct lu_env *env = info->mti_env;
1042 struct md_object *next = mdt_object_child(o);
1043 struct lu_buf *buf = &info->mti_buf;
1044 int need = ma->ma_need;
1051 if (mdt_object_exists(o) == 0)
1052 GOTO(out, rc = -ENOENT);
1053 mode = lu_object_attr(&next->mo_lu);
1055 if (need & MA_INODE) {
1056 ma->ma_need = MA_INODE;
1057 rc = mo_attr_get(env, next, ma);
1062 (void) mdt_get_som(info, o, ma);
1063 ma->ma_valid |= MA_INODE;
1066 if (need & MA_PFID) {
1067 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1069 ma->ma_valid |= MA_PFID;
1070 /* ignore this error, parent fid is not mandatory */
1074 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1075 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1080 if (need & MA_LMV && S_ISDIR(mode)) {
1081 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1086 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1087 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1093 * In the handle of MA_INODE, we may already get the SOM attr.
1095 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1096 rc = mdt_get_som(info, o, ma);
1101 if (need & MA_HSM && S_ISREG(mode)) {
1102 buf->lb_buf = info->mti_xattr_buf;
1103 buf->lb_len = sizeof(info->mti_xattr_buf);
1104 CLASSERT(sizeof(struct hsm_attrs) <=
1105 sizeof(info->mti_xattr_buf));
1106 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1107 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1109 ma->ma_valid |= MA_HSM;
1110 else if (rc2 < 0 && rc2 != -ENODATA)
1111 GOTO(out, rc = rc2);
1114 #ifdef CONFIG_FS_POSIX_ACL
1115 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1116 buf->lb_buf = ma->ma_acl;
1117 buf->lb_len = ma->ma_acl_size;
1118 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1120 ma->ma_acl_size = rc2;
1121 ma->ma_valid |= MA_ACL_DEF;
1122 } else if (rc2 == -ENODATA) {
1124 ma->ma_acl_size = 0;
1126 GOTO(out, rc = rc2);
1131 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1132 rc, ma->ma_valid, ma->ma_lmm);
1136 static int mdt_getattr_internal(struct mdt_thread_info *info,
1137 struct mdt_object *o, int ma_need)
1139 struct md_object *next = mdt_object_child(o);
1140 const struct mdt_body *reqbody = info->mti_body;
1141 struct ptlrpc_request *req = mdt_info_req(info);
1142 struct md_attr *ma = &info->mti_attr;
1143 struct lu_attr *la = &ma->ma_attr;
1144 struct req_capsule *pill = info->mti_pill;
1145 const struct lu_env *env = info->mti_env;
1146 struct mdt_body *repbody;
1147 struct lu_buf *buffer = &info->mti_buf;
1148 struct obd_export *exp = info->mti_exp;
1152 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1153 RETURN(err_serious(-ENOMEM));
1155 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1159 if (mdt_object_remote(o)) {
1160 /* This object is located on remote node.*/
1161 /* Return -ENOTSUPP for old client */
1162 if (!mdt_is_dne_client(req->rq_export))
1163 GOTO(out, rc = -ENOTSUPP);
1165 repbody->mbo_fid1 = *mdt_object_fid(o);
1166 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1170 if (reqbody->mbo_eadatasize > 0) {
1171 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1172 if (buffer->lb_buf == NULL)
1173 GOTO(out, rc = -EPROTO);
1174 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1177 buffer->lb_buf = NULL;
1179 ma_need &= ~(MA_LOV | MA_LMV);
1180 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1181 mdt_obd_name(info->mti_mdt),
1182 req->rq_export->exp_client_uuid.uuid);
1185 /* If it is dir object and client require MEA, then we got MEA */
1186 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1187 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1188 /* Assumption: MDT_MD size is enough for lmv size. */
1189 ma->ma_lmv = buffer->lb_buf;
1190 ma->ma_lmv_size = buffer->lb_len;
1191 ma->ma_need = MA_INODE;
1192 if (ma->ma_lmv_size > 0) {
1193 if (reqbody->mbo_valid & OBD_MD_MEA)
1194 ma->ma_need |= MA_LMV;
1195 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
1196 ma->ma_need |= MA_LMV_DEF;
1199 ma->ma_lmm = buffer->lb_buf;
1200 ma->ma_lmm_size = buffer->lb_len;
1201 ma->ma_need = MA_INODE | MA_HSM;
1202 if (ma->ma_lmm_size > 0)
1203 ma->ma_need |= MA_LOV;
1206 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1207 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1208 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1209 /* get default stripe info for this dir. */
1210 ma->ma_need |= MA_LOV_DEF;
1212 ma->ma_need |= ma_need;
1214 rc = mdt_attr_get_complex(info, o, ma);
1216 CDEBUG(rc == -ENOENT ? D_OTHER : D_ERROR,
1217 "%s: getattr error for "DFID": rc = %d\n",
1218 mdt_obd_name(info->mti_mdt),
1219 PFID(mdt_object_fid(o)), rc);
1223 /* if file is released, check if a restore is running */
1224 if (ma->ma_valid & MA_HSM) {
1225 repbody->mbo_valid |= OBD_MD_TSTATE;
1226 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1227 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1228 repbody->mbo_t_state = MS_RESTORE;
1231 if (likely(ma->ma_valid & MA_INODE))
1232 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1236 if (mdt_body_has_lov(la, reqbody)) {
1237 if (ma->ma_valid & MA_LOV) {
1238 LASSERT(ma->ma_lmm_size);
1239 repbody->mbo_eadatasize = ma->ma_lmm_size;
1240 if (S_ISDIR(la->la_mode))
1241 repbody->mbo_valid |= OBD_MD_FLDIREA;
1243 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1244 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1246 if (ma->ma_valid & MA_LMV) {
1247 /* Return -ENOTSUPP for old client */
1248 if (!mdt_is_striped_client(req->rq_export))
1251 LASSERT(S_ISDIR(la->la_mode));
1252 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1253 repbody->mbo_eadatasize = ma->ma_lmv_size;
1254 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1256 if (ma->ma_valid & MA_LMV_DEF) {
1257 /* Return -ENOTSUPP for old client */
1258 if (!mdt_is_striped_client(req->rq_export))
1260 LASSERT(S_ISDIR(la->la_mode));
1261 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1262 repbody->mbo_eadatasize = ma->ma_lmv_size;
1263 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1264 OBD_MD_DEFAULT_MEA);
1266 } else if (S_ISLNK(la->la_mode) &&
1267 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1268 buffer->lb_buf = ma->ma_lmm;
1269 /* eadatasize from client includes NULL-terminator, so
1270 * there is no need to read it */
1271 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1272 rc = mo_readlink(env, next, buffer);
1273 if (unlikely(rc <= 0)) {
1274 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1275 mdt_obd_name(info->mti_mdt),
1276 PFID(mdt_object_fid(o)), rc);
1279 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1281 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1283 repbody->mbo_valid |= OBD_MD_LINKNAME;
1284 /* we need to report back size with NULL-terminator
1285 * because client expects that */
1286 repbody->mbo_eadatasize = rc + 1;
1287 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1288 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1289 "on "DFID ", expected %d\n",
1290 mdt_obd_name(info->mti_mdt),
1291 rc, PFID(mdt_object_fid(o)),
1292 reqbody->mbo_eadatasize - 1);
1293 /* NULL terminate */
1294 ((char *)ma->ma_lmm)[rc] = 0;
1296 /* If the total CDEBUG() size is larger than a page, it
1297 * will print a warning to the console, avoid this by
1298 * printing just the last part of the symlink. */
1299 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1300 print_limit < rc ? "..." : "", print_limit,
1301 (char *)ma->ma_lmm + rc - print_limit, rc);
1306 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1307 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1308 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1309 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1310 repbody->mbo_max_mdsize);
1313 #ifdef CONFIG_FS_POSIX_ACL
1314 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1315 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1316 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1317 if (IS_ERR(nodemap))
1318 RETURN(PTR_ERR(nodemap));
1320 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1321 nodemap_putref(nodemap);
1327 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1332 static int mdt_getattr(struct tgt_session_info *tsi)
1334 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1335 struct mdt_object *obj = info->mti_object;
1336 struct req_capsule *pill = info->mti_pill;
1337 struct mdt_body *reqbody;
1338 struct mdt_body *repbody;
1342 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1344 LASSERT(obj != NULL);
1345 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1347 /* Special case for Data-on-MDT files to get data version */
1348 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1349 rc = mdt_data_version_get(tsi);
1353 /* Unlike intent case where we need to pre-fill out buffers early on
1354 * in intent policy for ldlm reasons, here we can have a much better
1355 * guess at EA size by just reading it from disk.
1356 * Exceptions are readdir and (missing) directory striping */
1358 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1359 /* No easy way to know how long is the symlink, but it cannot
1360 * be more than PATH_MAX, so we allocate +1 */
1362 /* A special case for fs ROOT: getattr there might fetch
1363 * default EA for entire fs, not just for this dir!
1365 } else if (lu_fid_eq(mdt_object_fid(obj),
1366 &info->mti_mdt->mdt_md_root_fid) &&
1367 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1368 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1370 /* Should the default strping be bigger, mdt_fix_reply
1371 * will reallocate */
1372 rc = DEF_REP_MD_SIZE;
1374 /* Read the actual EA size from disk */
1375 rc = mdt_attr_get_eabuf_size(info, obj);
1379 GOTO(out, rc = err_serious(rc));
1381 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1383 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1384 * by default. If the target object has more ACL entries, then
1385 * enlarge the buffer when necessary. */
1386 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1387 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1389 rc = req_capsule_server_pack(pill);
1390 if (unlikely(rc != 0))
1391 GOTO(out, rc = err_serious(rc));
1393 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1394 LASSERT(repbody != NULL);
1395 repbody->mbo_eadatasize = 0;
1396 repbody->mbo_aclsize = 0;
1398 rc = mdt_check_ucred(info);
1400 GOTO(out_shrink, rc);
1402 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1404 rc = mdt_getattr_internal(info, obj, 0);
1407 mdt_client_compatibility(info);
1408 rc2 = mdt_fix_reply(info);
1412 mdt_thread_info_fini(info);
1417 * Handler of layout intent RPC requiring the layout modification
1419 * \param[in] info thread environment
1420 * \param[in] obj object
1421 * \param[in] layout layout change descriptor
1423 * \retval 0 on success
1424 * \retval < 0 error code
1426 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1427 struct md_layout_change *layout)
1429 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1433 if (!mdt_object_exists(obj))
1434 GOTO(out, rc = -ENOENT);
1436 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1437 GOTO(out, rc = -EINVAL);
1439 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1444 /* take layout lock to prepare layout change */
1445 mdt_lock_reg_init(lh, LCK_EX);
1446 rc = mdt_object_lock(info, obj, lh, MDS_INODELOCK_LAYOUT);
1450 mutex_lock(&obj->mot_som_mutex);
1451 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1452 mutex_unlock(&obj->mot_som_mutex);
1453 mdt_object_unlock(info, obj, lh, 1);
1459 * Exchange MOF_LOV_CREATED flags between two objects after a
1460 * layout swap. No assumption is made on whether o1 or o2 have
1461 * created objects or not.
1463 * \param[in,out] o1 First swap layout object
1464 * \param[in,out] o2 Second swap layout object
1466 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1468 unsigned int o1_lov_created = o1->mot_lov_created;
1470 mutex_lock(&o1->mot_lov_mutex);
1471 mutex_lock(&o2->mot_lov_mutex);
1473 o1->mot_lov_created = o2->mot_lov_created;
1474 o2->mot_lov_created = o1_lov_created;
1476 mutex_unlock(&o2->mot_lov_mutex);
1477 mutex_unlock(&o1->mot_lov_mutex);
1480 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1482 struct mdt_thread_info *info;
1483 struct ptlrpc_request *req = tgt_ses_req(tsi);
1484 struct obd_export *exp = req->rq_export;
1485 struct mdt_object *o1, *o2, *o;
1486 struct mdt_lock_handle *lh1, *lh2;
1487 struct mdc_swap_layouts *msl;
1491 /* client does not support layout lock, so layout swaping
1493 * FIXME: there is a problem for old clients which don't support
1494 * layout lock yet. If those clients have already opened the file
1495 * they won't be notified at all so that old layout may still be
1496 * used to do IO. This can be fixed after file release is landed by
1497 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1498 if (!exp_connect_layout(exp))
1499 RETURN(-EOPNOTSUPP);
1501 info = tsi2mdt_info(tsi);
1503 if (info->mti_dlm_req != NULL)
1504 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1506 o1 = info->mti_object;
1507 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1508 &info->mti_body->mbo_fid2);
1510 GOTO(out, rc = PTR_ERR(o));
1512 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1513 GOTO(put, rc = -ENOENT);
1515 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1516 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1522 /* permission check. Make sure the calling process having permission
1523 * to write both files. */
1524 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1529 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1534 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1536 GOTO(put, rc = -EPROTO);
1538 lh1 = &info->mti_lh[MDT_LH_NEW];
1539 mdt_lock_reg_init(lh1, LCK_EX);
1540 lh2 = &info->mti_lh[MDT_LH_OLD];
1541 mdt_lock_reg_init(lh2, LCK_EX);
1543 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1544 MDS_INODELOCK_XATTR);
1548 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1549 MDS_INODELOCK_XATTR);
1553 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1554 mdt_object_child(o2), msl->msl_flags);
1558 mdt_swap_lov_flag(o1, o2);
1561 mdt_object_unlock(info, o2, lh2, rc);
1563 mdt_object_unlock(info, o1, lh1, rc);
1565 mdt_object_put(info->mti_env, o);
1567 mdt_thread_info_fini(info);
1571 static int mdt_raw_lookup(struct mdt_thread_info *info,
1572 struct mdt_object *parent,
1573 const struct lu_name *lname,
1574 struct ldlm_reply *ldlm_rep)
1576 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1580 LASSERT(!info->mti_cross_ref);
1582 /* Only got the fid of this obj by name */
1583 fid_zero(child_fid);
1584 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1585 lname, child_fid, &info->mti_spec);
1587 struct mdt_body *repbody;
1589 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1590 repbody->mbo_fid1 = *child_fid;
1591 repbody->mbo_valid = OBD_MD_FLID;
1592 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1593 } else if (rc == -ENOENT) {
1594 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1601 * UPDATE lock should be taken against parent, and be released before exit;
1602 * child_bits lock should be taken against child, and be returned back:
1603 * (1)normal request should release the child lock;
1604 * (2)intent request will grant the lock to client.
1606 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1607 struct mdt_lock_handle *lhc,
1609 struct ldlm_reply *ldlm_rep)
1611 struct ptlrpc_request *req = mdt_info_req(info);
1612 struct mdt_body *reqbody = NULL;
1613 struct mdt_object *parent = info->mti_object;
1614 struct mdt_object *child;
1615 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1616 struct lu_name *lname = NULL;
1617 struct mdt_lock_handle *lhp = NULL;
1618 struct ldlm_lock *lock;
1626 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1627 LASSERT(ergo(is_resent,
1628 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1633 if (info->mti_cross_ref) {
1634 /* Only getattr on the child. Parent is on another node. */
1635 mdt_set_disposition(info, ldlm_rep,
1636 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1638 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1640 PFID(mdt_object_fid(child)), ldlm_rep);
1642 rc = mdt_check_resent_lock(info, child, lhc);
1645 } else if (rc > 0) {
1646 mdt_lock_handle_init(lhc);
1647 mdt_lock_reg_init(lhc, LCK_PR);
1650 * Object's name is on another MDS, no lookup or layout
1651 * lock is needed here but update lock is.
1653 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1654 MDS_INODELOCK_LAYOUT);
1655 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1657 rc = mdt_object_lock(info, child, lhc, child_bits);
1662 /* Finally, we can get attr for child. */
1663 if (!mdt_object_exists(child)) {
1664 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1666 "remote object doesn't exist.");
1667 mdt_object_unlock(info, child, lhc, 1);
1671 rc = mdt_getattr_internal(info, child, 0);
1672 if (unlikely(rc != 0))
1673 mdt_object_unlock(info, child, lhc, 1);
1678 lname = &info->mti_name;
1679 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1681 if (lu_name_is_valid(lname)) {
1682 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1683 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1684 PNAME(lname), ldlm_rep);
1686 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1687 if (unlikely(reqbody == NULL))
1688 RETURN(err_serious(-EPROTO));
1690 *child_fid = reqbody->mbo_fid2;
1692 if (unlikely(!fid_is_sane(child_fid)))
1693 RETURN(err_serious(-EINVAL));
1695 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1697 PFID(mdt_object_fid(parent)),
1698 PFID(&reqbody->mbo_fid2), ldlm_rep);
1701 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1703 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1704 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1706 "Parent doesn't exist!");
1710 if (mdt_object_remote(parent)) {
1711 CERROR("%s: parent "DFID" is on remote target\n",
1712 mdt_obd_name(info->mti_mdt),
1713 PFID(mdt_object_fid(parent)));
1717 if (lu_name_is_valid(lname)) {
1718 /* Always allow to lookup ".." */
1719 if (unlikely(lname->ln_namelen == 2 &&
1720 lname->ln_name[0] == '.' &&
1721 lname->ln_name[1] == '.'))
1722 info->mti_spec.sp_permitted = 1;
1724 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1725 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1730 /* step 1: lock parent only if parent is a directory */
1731 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1732 lhp = &info->mti_lh[MDT_LH_PARENT];
1733 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1734 rc = mdt_object_lock(info, parent, lhp,
1735 MDS_INODELOCK_UPDATE);
1736 if (unlikely(rc != 0))
1740 /* step 2: lookup child's fid by name */
1741 fid_zero(child_fid);
1742 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1743 child_fid, &info->mti_spec);
1745 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1748 GOTO(out_parent, rc);
1751 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1754 *step 3: find the child object by fid & lock it.
1755 * regardless if it is local or remote.
1757 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1758 * set parent dir fid the same as child fid in getattr by fid case
1759 * we should not lu_object_find() the object again, could lead
1760 * to hung if there is a concurrent unlink destroyed the object.
1762 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1763 mdt_object_get(info->mti_env, parent);
1766 child = mdt_object_find(info->mti_env, info->mti_mdt,
1770 if (unlikely(IS_ERR(child)))
1771 GOTO(out_parent, rc = PTR_ERR(child));
1773 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1774 if (!mdt_object_exists(child)) {
1775 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1777 "Object doesn't exist!");
1778 GOTO(out_child, rc = -ENOENT);
1781 rc = mdt_check_resent_lock(info, child, lhc);
1783 GOTO(out_child, rc);
1784 } else if (rc > 0) {
1785 mdt_lock_handle_init(lhc);
1786 mdt_lock_reg_init(lhc, LCK_PR);
1788 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1789 mdt_object_exists(child) && !mdt_object_remote(child)) {
1790 struct md_attr *ma = &info->mti_attr;
1793 ma->ma_need = MA_INODE;
1794 rc = mdt_attr_get_complex(info, child, ma);
1795 if (unlikely(rc != 0))
1796 GOTO(out_child, rc);
1798 /* If the file has not been changed for some time, we
1799 * return not only a LOOKUP lock, but also an UPDATE
1800 * lock and this might save us RPC on later STAT. For
1801 * directories, it also let negative dentry cache start
1802 * working for this dir. */
1803 if (ma->ma_valid & MA_INODE &&
1804 ma->ma_attr.la_valid & LA_CTIME &&
1805 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1806 ma->ma_attr.la_ctime < ktime_get_real_seconds())
1807 child_bits |= MDS_INODELOCK_UPDATE;
1810 /* layout lock must be granted in a best-effort way
1811 * for IT operations */
1812 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1813 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1814 !mdt_object_remote(child) && ldlm_rep != NULL) {
1815 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1816 exp_connect_layout(info->mti_exp)) {
1817 /* try to grant layout lock for regular file. */
1818 try_bits = MDS_INODELOCK_LAYOUT;
1820 /* Acquire DOM lock in advance for data-on-mdt file */
1821 if (child != parent)
1822 try_bits |= MDS_INODELOCK_DOM;
1825 if (try_bits != 0) {
1826 /* try layout lock, it may fail to be granted due to
1827 * contention at LOOKUP or UPDATE */
1828 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
1830 if (child_bits & MDS_INODELOCK_LAYOUT)
1833 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1834 * client will enqueue the lock to the remote MDT */
1835 if (mdt_object_remote(child))
1836 child_bits &= ~MDS_INODELOCK_UPDATE;
1837 rc = mdt_object_lock(info, child, lhc, child_bits);
1839 if (unlikely(rc != 0))
1840 GOTO(out_child, rc);
1843 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1845 /* finally, we can get attr for child. */
1846 rc = mdt_getattr_internal(info, child, ma_need);
1847 if (unlikely(rc != 0)) {
1848 mdt_object_unlock(info, child, lhc, 1);
1850 /* Debugging code. */
1851 LDLM_DEBUG(lock, "Returning lock to client");
1852 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1853 &lock->l_resource->lr_name),
1854 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1855 PLDLMRES(lock->l_resource),
1856 PFID(mdt_object_fid(child)));
1858 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1859 mdt_object_exists(child) && !mdt_object_remote(child) &&
1861 LDLM_LOCK_PUT(lock);
1862 mdt_object_put(info->mti_env, child);
1863 /* NB: call the mdt_pack_size2body always after
1864 * mdt_object_put(), that is why this special
1865 * exit path is used. */
1866 rc = mdt_pack_size2body(info, child_fid,
1868 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
1869 /* DOM lock was taken in advance but this is
1870 * not DoM file. Drop the lock. */
1871 lock_res_and_lock(lock);
1872 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
1873 unlock_res_and_lock(lock);
1876 GOTO(out_parent, rc = 0);
1880 LDLM_LOCK_PUT(lock);
1884 mdt_object_put(info->mti_env, child);
1887 mdt_object_unlock(info, parent, lhp, 1);
1891 /* normal handler: should release the child lock */
1892 static int mdt_getattr_name(struct tgt_session_info *tsi)
1894 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1895 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1896 struct mdt_body *reqbody;
1897 struct mdt_body *repbody;
1901 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1902 LASSERT(reqbody != NULL);
1903 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1904 LASSERT(repbody != NULL);
1906 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1907 repbody->mbo_eadatasize = 0;
1908 repbody->mbo_aclsize = 0;
1910 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1912 GOTO(out_shrink, rc);
1914 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1915 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1916 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1917 lhc->mlh_reg_lh.cookie = 0;
1919 mdt_exit_ucred(info);
1922 mdt_client_compatibility(info);
1923 rc2 = mdt_fix_reply(info);
1926 mdt_thread_info_fini(info);
1930 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1931 void *karg, void __user *uarg);
1933 static int mdt_set_info(struct tgt_session_info *tsi)
1935 struct ptlrpc_request *req = tgt_ses_req(tsi);
1938 int keylen, vallen, rc = 0;
1942 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1944 DEBUG_REQ(D_HA, req, "no set_info key");
1945 RETURN(err_serious(-EFAULT));
1948 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1951 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1953 DEBUG_REQ(D_HA, req, "no set_info val");
1954 RETURN(err_serious(-EFAULT));
1957 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1960 /* Swab any part of val you need to here */
1961 if (KEY_IS(KEY_READ_ONLY)) {
1962 spin_lock(&req->rq_export->exp_lock);
1964 *exp_connect_flags_ptr(req->rq_export) |=
1967 *exp_connect_flags_ptr(req->rq_export) &=
1968 ~OBD_CONNECT_RDONLY;
1969 spin_unlock(&req->rq_export->exp_lock);
1970 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1971 struct changelog_setinfo *cs = val;
1973 if (vallen != sizeof(*cs)) {
1974 CERROR("%s: bad changelog_clear setinfo size %d\n",
1975 tgt_name(tsi->tsi_tgt), vallen);
1978 if (ptlrpc_req_need_swab(req)) {
1979 __swab64s(&cs->cs_recno);
1980 __swab32s(&cs->cs_id);
1983 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1985 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
1987 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
1994 static int mdt_readpage(struct tgt_session_info *tsi)
1996 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1997 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1998 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1999 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2000 struct mdt_body *repbody;
2006 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2007 RETURN(err_serious(-ENOMEM));
2009 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2010 if (repbody == NULL || reqbody == NULL)
2011 RETURN(err_serious(-EFAULT));
2014 * prepare @rdpg before calling lower layers and transfer itself. Here
2015 * reqbody->size contains offset of where to start to read and
2016 * reqbody->nlink contains number bytes to read.
2018 rdpg->rp_hash = reqbody->mbo_size;
2019 if (rdpg->rp_hash != reqbody->mbo_size) {
2020 CERROR("Invalid hash: %#llx != %#llx\n",
2021 rdpg->rp_hash, reqbody->mbo_size);
2025 rdpg->rp_attrs = reqbody->mbo_mode;
2026 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2027 rdpg->rp_attrs |= LUDA_64BITHASH;
2028 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2029 exp_max_brw_size(tsi->tsi_exp));
2030 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2032 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2033 if (rdpg->rp_pages == NULL)
2036 for (i = 0; i < rdpg->rp_npages; ++i) {
2037 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2038 if (rdpg->rp_pages[i] == NULL)
2039 GOTO(free_rdpg, rc = -ENOMEM);
2042 /* call lower layers to fill allocated pages with directory data */
2043 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2045 GOTO(free_rdpg, rc);
2047 /* send pages to client */
2048 rc = tgt_sendpage(tsi, rdpg, rc);
2053 for (i = 0; i < rdpg->rp_npages; i++)
2054 if (rdpg->rp_pages[i] != NULL)
2055 __free_page(rdpg->rp_pages[i]);
2056 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2058 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2064 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2066 struct lu_ucred *uc = mdt_ucred_check(info);
2067 struct lu_attr *attr = &info->mti_attr.ma_attr;
2072 if (op != REINT_SETATTR) {
2073 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2074 attr->la_uid = uc->uc_fsuid;
2075 /* for S_ISGID, inherit gid from his parent, such work will be
2076 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2077 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2078 attr->la_gid = uc->uc_fsgid;
2084 static int mdt_reint_internal(struct mdt_thread_info *info,
2085 struct mdt_lock_handle *lhc,
2088 struct req_capsule *pill = info->mti_pill;
2089 struct mdt_body *repbody;
2094 rc = mdt_reint_unpack(info, op);
2096 CERROR("Can't unpack reint, rc %d\n", rc);
2097 RETURN(err_serious(rc));
2100 /* for replay (no_create) lmm is not needed, client has it already */
2101 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2102 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2105 /* llog cookies are always 0, the field is kept for compatibility */
2106 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2107 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
2109 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
2110 * by default. If the target object has more ACL entries, then
2111 * enlarge the buffer when necessary. */
2112 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
2113 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
2114 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
2116 rc = req_capsule_server_pack(pill);
2118 CERROR("Can't pack response, rc %d\n", rc);
2119 RETURN(err_serious(rc));
2122 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
2123 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
2125 repbody->mbo_eadatasize = 0;
2126 repbody->mbo_aclsize = 0;
2129 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
2131 /* for replay no cookkie / lmm need, because client have this already */
2132 if (info->mti_spec.no_create)
2133 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2134 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
2136 rc = mdt_init_ucred_reint(info);
2138 GOTO(out_shrink, rc);
2140 rc = mdt_fix_attr_ucred(info, op);
2142 GOTO(out_ucred, rc = err_serious(rc));
2144 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
2146 GOTO(out_ucred, rc);
2147 } else if (rc == 1) {
2148 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
2149 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
2150 GOTO(out_ucred, rc);
2152 rc = mdt_reint_rec(info, lhc);
2155 mdt_exit_ucred(info);
2157 mdt_client_compatibility(info);
2159 rc2 = mdt_fix_reply(info);
2164 * Data-on-MDT optimization - read data along with OPEN and return it
2165 * in reply. Do that only if we have both DOM and LAYOUT locks.
2167 if (rc == 0 && op == REINT_OPEN &&
2168 info->mti_attr.ma_lmm != NULL &&
2169 mdt_lmm_dom_entry(info->mti_attr.ma_lmm) == LMM_DOM_ONLY) {
2170 rc = mdt_dom_read_on_open(info, info->mti_mdt,
2177 static long mdt_reint_opcode(struct ptlrpc_request *req,
2178 const struct req_format **fmt)
2180 struct mdt_device *mdt;
2181 struct mdt_rec_reint *rec;
2184 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
2186 opc = rec->rr_opcode;
2187 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
2188 if (opc < REINT_MAX && fmt[opc] != NULL)
2189 req_capsule_extend(&req->rq_pill, fmt[opc]);
2191 mdt = mdt_exp2dev(req->rq_export);
2192 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
2193 " rc = %d\n", req->rq_export->exp_obd->obd_name,
2194 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
2195 opc = err_serious(-EFAULT);
2198 opc = err_serious(-EFAULT);
2203 static int mdt_reint(struct tgt_session_info *tsi)
2207 static const struct req_format *reint_fmts[REINT_MAX] = {
2208 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
2209 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
2210 [REINT_LINK] = &RQF_MDS_REINT_LINK,
2211 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
2212 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
2213 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
2214 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
2215 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
2216 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
2217 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
2222 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
2224 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2226 * No lock possible here from client to pass it to reint code
2229 rc = mdt_reint_internal(info, NULL, opc);
2230 mdt_thread_info_fini(info);
2235 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
2239 /* this should sync the whole device */
2240 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
2242 struct dt_device *dt = mdt->mdt_bottom;
2246 rc = dt->dd_ops->dt_sync(env, dt);
2250 /* this should sync this object */
2251 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
2252 struct mdt_object *mo)
2258 if (!mdt_object_exists(mo)) {
2259 CWARN("%s: non existing object "DFID": rc = %d\n",
2260 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
2265 rc = mo_object_sync(env, mdt_object_child(mo));
2270 static int mdt_sync(struct tgt_session_info *tsi)
2272 struct ptlrpc_request *req = tgt_ses_req(tsi);
2273 struct req_capsule *pill = tsi->tsi_pill;
2274 struct mdt_body *body;
2279 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2280 RETURN(err_serious(-ENOMEM));
2282 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2283 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2285 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2287 /* sync an object */
2288 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
2291 const struct lu_fid *fid;
2292 struct lu_attr *la = &info->mti_attr.ma_attr;
2294 info->mti_attr.ma_need = MA_INODE;
2295 info->mti_attr.ma_valid = 0;
2296 rc = mdt_attr_get_complex(info, info->mti_object,
2299 body = req_capsule_server_get(pill,
2301 fid = mdt_object_fid(info->mti_object);
2302 mdt_pack_attr2body(info, body, la, fid);
2305 mdt_thread_info_fini(info);
2308 mdt_counter_incr(req, LPROC_MDT_SYNC);
2313 static int mdt_data_sync(struct tgt_session_info *tsi)
2315 struct mdt_thread_info *info;
2316 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2317 struct ost_body *body = tsi->tsi_ost_body;
2318 struct ost_body *repbody;
2319 struct mdt_object *mo = NULL;
2325 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2327 /* if no fid is specified then do nothing,
2328 * device sync is done via MDS_SYNC */
2329 if (fid_is_zero(&tsi->tsi_fid))
2332 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
2334 RETURN(PTR_ERR(mo));
2336 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
2340 repbody->oa.o_oi = body->oa.o_oi;
2341 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2343 info = tsi2mdt_info(tsi);
2344 ma = &info->mti_attr;
2345 ma->ma_need = MA_INODE;
2347 rc = mdt_attr_get_complex(info, mo, ma);
2349 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
2352 mdt_thread_info_fini(info);
2357 mdt_object_put(tsi->tsi_env, mo);
2362 * Handle quota control requests to consult current usage/limit, but also
2363 * to configure quota enforcement
2365 static int mdt_quotactl(struct tgt_session_info *tsi)
2367 struct obd_export *exp = tsi->tsi_exp;
2368 struct req_capsule *pill = tsi->tsi_pill;
2369 struct obd_quotactl *oqctl, *repoqc;
2371 struct mdt_device *mdt = mdt_exp2dev(exp);
2372 struct lu_device *qmt = mdt->mdt_qmt_dev;
2373 struct lu_nodemap *nodemap;
2376 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2378 RETURN(err_serious(-EPROTO));
2380 rc = req_capsule_server_pack(pill);
2382 RETURN(err_serious(rc));
2384 nodemap = nodemap_get_from_exp(exp);
2385 if (IS_ERR(nodemap))
2386 RETURN(PTR_ERR(nodemap));
2388 switch (oqctl->qc_cmd) {
2389 /* master quotactl */
2392 case LUSTRE_Q_SETDEFAULT:
2393 if (!nodemap_can_setquota(nodemap))
2394 GOTO(out_nodemap, rc = -EPERM);
2397 case LUSTRE_Q_GETDEFAULT:
2399 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2400 /* slave quotactl */
2405 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2406 GOTO(out_nodemap, rc = -EFAULT);
2410 switch (oqctl->qc_type) {
2412 id = nodemap_map_id(nodemap, NODEMAP_UID,
2413 NODEMAP_CLIENT_TO_FS, id);
2416 id = nodemap_map_id(nodemap, NODEMAP_GID,
2417 NODEMAP_CLIENT_TO_FS, id);
2420 /* todo: check/map project id */
2424 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2426 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2428 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2430 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
2431 barrier_exit(tsi->tsi_tgt->lut_bottom);
2433 if (oqctl->qc_id != id)
2434 swap(oqctl->qc_id, id);
2436 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
2437 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
2438 RETURN(-EINPROGRESS);
2441 switch (oqctl->qc_cmd) {
2447 case LUSTRE_Q_SETDEFAULT:
2448 case LUSTRE_Q_GETDEFAULT:
2449 /* forward quotactl request to QMT */
2450 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2455 /* slave quotactl */
2456 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2461 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2462 GOTO(out_nodemap, rc = -EFAULT);
2465 if (oqctl->qc_id != id)
2466 swap(oqctl->qc_id, id);
2473 nodemap_putref(nodemap);
2478 /** clone llog ctxt from child (mdd)
2479 * This allows remote llog (replicator) access.
2480 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2481 * context was originally set up, or we can handle them directly.
2482 * I choose the latter, but that means I need any llog
2483 * contexts set up by child to be accessable by the mdt. So we clone the
2484 * context into our context list here.
2486 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2489 struct md_device *next = mdt->mdt_child;
2490 struct llog_ctxt *ctxt;
2493 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2496 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2497 if (rc || ctxt == NULL) {
2501 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2503 CERROR("Can't set mdt ctxt %d\n", rc);
2508 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2509 struct mdt_device *mdt, int idx)
2511 struct llog_ctxt *ctxt;
2513 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2516 /* Put once for the get we just did, and once for the clone */
2517 llog_ctxt_put(ctxt);
2518 llog_ctxt_put(ctxt);
2523 * sec context handlers
2525 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2527 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2533 * quota request handlers
2535 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2537 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2538 struct lu_device *qmt = mdt->mdt_qmt_dev;
2543 RETURN(err_serious(-EOPNOTSUPP));
2545 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2549 struct mdt_object *mdt_object_new(const struct lu_env *env,
2550 struct mdt_device *d,
2551 const struct lu_fid *f)
2553 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2554 struct lu_object *o;
2555 struct mdt_object *m;
2558 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2559 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2560 if (unlikely(IS_ERR(o)))
2561 m = (struct mdt_object *)o;
2567 struct mdt_object *mdt_object_find(const struct lu_env *env,
2568 struct mdt_device *d,
2569 const struct lu_fid *f)
2571 struct lu_object *o;
2572 struct mdt_object *m;
2575 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2576 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2577 if (unlikely(IS_ERR(o)))
2578 m = (struct mdt_object *)o;
2586 * Asyncronous commit for mdt device.
2588 * Pass asynchonous commit call down the MDS stack.
2590 * \param env environment
2591 * \param mdt the mdt device
2593 static void mdt_device_commit_async(const struct lu_env *env,
2594 struct mdt_device *mdt)
2596 struct dt_device *dt = mdt->mdt_bottom;
2600 rc = dt->dd_ops->dt_commit_async(env, dt);
2601 if (unlikely(rc != 0))
2602 CWARN("%s: async commit start failed: rc = %d\n",
2603 mdt_obd_name(mdt), rc);
2604 atomic_inc(&mdt->mdt_async_commit_count);
2609 * Mark the lock as "synchonous".
2611 * Mark the lock to deffer transaction commit to the unlock time.
2613 * \param lock the lock to mark as "synchonous"
2615 * \see mdt_is_lock_sync
2616 * \see mdt_save_lock
2618 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2620 lock->l_ast_data = (void*)1;
2624 * Check whehter the lock "synchonous" or not.
2626 * \param lock the lock to check
2627 * \retval 1 the lock is "synchonous"
2628 * \retval 0 the lock isn't "synchronous"
2630 * \see mdt_set_lock_sync
2631 * \see mdt_save_lock
2633 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2635 return lock->l_ast_data != NULL;
2639 * Blocking AST for mdt locks.
2641 * Starts transaction commit if in case of COS lock conflict or
2642 * deffers such a commit to the mdt_save_lock.
2644 * \param lock the lock which blocks a request or cancelling lock
2645 * \param desc unused
2646 * \param data unused
2647 * \param flag indicates whether this cancelling or blocking callback
2649 * \see ldlm_blocking_ast_nocheck
2651 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2652 void *data, int flag)
2654 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2655 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2656 bool commit_async = false;
2660 if (flag == LDLM_CB_CANCELING)
2663 lock_res_and_lock(lock);
2664 if (lock->l_blocking_ast != mdt_blocking_ast) {
2665 unlock_res_and_lock(lock);
2668 /* There is no lock conflict if l_blocking_lock == NULL,
2669 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2670 * when the last reference to a local lock was released */
2671 if (lock->l_req_mode & (LCK_PW | LCK_EX) &&
2672 lock->l_blocking_lock != NULL) {
2673 if (mdt_cos_is_enabled(mdt)) {
2674 if (lock->l_client_cookie !=
2675 lock->l_blocking_lock->l_client_cookie)
2676 mdt_set_lock_sync(lock);
2677 } else if (mdt_slc_is_enabled(mdt) &&
2678 ldlm_is_cos_incompat(lock->l_blocking_lock)) {
2679 mdt_set_lock_sync(lock);
2681 * we may do extra commit here, but there is a small
2682 * window to miss a commit: lock was unlocked (saved),
2683 * then a conflict lock queued and we come here, but
2684 * REP-ACK not received, so lock was not converted to
2686 * Fortunately this window is quite small, so the
2687 * extra commit should be rare (not to say distributed
2688 * operation is rare too).
2690 commit_async = true;
2692 } else if (lock->l_req_mode == LCK_COS &&
2693 lock->l_blocking_lock != NULL) {
2694 commit_async = true;
2697 rc = ldlm_blocking_ast_nocheck(lock);
2702 rc = lu_env_init(&env, LCT_LOCAL);
2703 if (unlikely(rc != 0))
2704 CWARN("%s: lu_env initialization failed, cannot "
2705 "start asynchronous commit: rc = %d\n",
2708 mdt_device_commit_async(&env, mdt);
2715 * Blocking AST for cross-MDT lock
2717 * Discard lock from uncommitted_slc_locks and cancel it.
2719 * \param lock the lock which blocks a request or cancelling lock
2720 * \param desc unused
2721 * \param data unused
2722 * \param flag indicates whether this cancelling or blocking callback
2723 * \retval 0 on success
2724 * \retval negative number on error
2726 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2727 void *data, int flag)
2733 case LDLM_CB_BLOCKING: {
2734 struct lustre_handle lockh;
2736 ldlm_lock2handle(lock, &lockh);
2737 rc = ldlm_cli_cancel(&lockh,
2738 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
2740 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2745 case LDLM_CB_CANCELING: {
2746 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2747 struct mdt_device *mdt =
2748 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
2750 LDLM_DEBUG(lock, "Revoke remote lock\n");
2752 /* discard slc lock here so that it can be cleaned anytime,
2753 * especially for cleanup_resource() */
2754 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
2756 /* once we cache lock, l_ast_data is set to mdt_object */
2757 if (lock->l_ast_data != NULL) {
2758 struct mdt_object *mo = lock->l_ast_data;
2761 rc = lu_env_init(&env, LCT_MD_THREAD);
2762 if (unlikely(rc != 0)) {
2763 CWARN("%s: lu_env initialization failed, object"
2764 "%p "DFID" is leaked!\n",
2766 PFID(mdt_object_fid(mo)));
2770 if (lock->l_policy_data.l_inodebits.bits &
2771 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
2772 rc = mo_invalidate(&env, mdt_object_child(mo));
2773 mo->mot_cache_attr = 0;
2775 mdt_object_put(&env, mo);
2787 int mdt_check_resent_lock(struct mdt_thread_info *info,
2788 struct mdt_object *mo,
2789 struct mdt_lock_handle *lhc)
2791 /* the lock might already be gotten in ldlm_handle_enqueue() */
2792 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2793 struct ptlrpc_request *req = mdt_info_req(info);
2794 struct ldlm_lock *lock;
2796 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2797 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2799 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2800 * a resend case, however, it could be already destroyed
2801 * due to client eviction or a raced cancel RPC. */
2802 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
2803 lhc->mlh_reg_lh.cookie);
2807 if (!fid_res_name_eq(mdt_object_fid(mo),
2808 &lock->l_resource->lr_name)) {
2809 CWARN("%s: Although resent, but still not "
2810 "get child lock:"DFID"\n",
2811 info->mti_exp->exp_obd->obd_name,
2812 PFID(mdt_object_fid(mo)));
2813 LDLM_LOCK_PUT(lock);
2816 LDLM_LOCK_PUT(lock);
2822 int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
2823 struct mdt_object *o, const struct lu_fid *fid,
2824 struct lustre_handle *lh, enum ldlm_mode mode,
2825 __u64 *ibits, __u64 trybits, bool cache)
2827 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
2828 union ldlm_policy_data *policy = &mti->mti_policy;
2829 struct ldlm_res_id *res_id = &mti->mti_res_id;
2833 LASSERT(mdt_object_remote(o));
2835 fid_build_reg_res_name(fid, res_id);
2837 memset(einfo, 0, sizeof(*einfo));
2838 einfo->ei_type = LDLM_IBITS;
2839 einfo->ei_mode = mode;
2840 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2841 einfo->ei_cb_cp = ldlm_completion_ast;
2842 einfo->ei_enq_slave = 0;
2843 einfo->ei_res_id = res_id;
2847 * if we cache lock, couple lock with mdt_object, so that object
2848 * can be easily found in lock ASTs.
2850 mdt_object_get(mti->mti_env, o);
2851 einfo->ei_cbdata = o;
2854 memset(policy, 0, sizeof(*policy));
2855 policy->l_inodebits.bits = *ibits;
2856 policy->l_inodebits.try_bits = trybits;
2858 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2860 if (rc < 0 && cache)
2861 mdt_object_put(mti->mti_env, o);
2863 /* Return successfully acquired bits to a caller */
2865 struct ldlm_lock *lock = ldlm_handle2lock(lh);
2868 *ibits = lock->l_policy_data.l_inodebits.bits;
2869 LDLM_LOCK_PUT(lock);
2874 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
2875 const struct lu_fid *fid, struct lustre_handle *lh,
2876 enum ldlm_mode mode, __u64 ibits, bool cache)
2878 return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0,
2882 static int mdt_object_local_lock(struct mdt_thread_info *info,
2883 struct mdt_object *o,
2884 struct mdt_lock_handle *lh, __u64 *ibits,
2885 __u64 trybits, bool cos_incompat)
2887 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2888 union ldlm_policy_data *policy = &info->mti_policy;
2889 struct ldlm_res_id *res_id = &info->mti_res_id;
2894 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2895 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2896 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2897 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2900 LASSERT(lh->mlh_reg_mode == LCK_PW ||
2901 lh->mlh_reg_mode == LCK_EX);
2902 dlmflags |= LDLM_FL_COS_INCOMPAT;
2903 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
2904 dlmflags |= LDLM_FL_COS_ENABLED;
2907 /* Only enqueue LOOKUP lock for remote object */
2908 LASSERT(ergo(mdt_object_remote(o), *ibits == MDS_INODELOCK_LOOKUP));
2910 if (lh->mlh_type == MDT_PDO_LOCK) {
2911 /* check for exists after object is locked */
2912 if (mdt_object_exists(o) == 0) {
2913 /* Non-existent object shouldn't have PDO lock */
2916 /* Non-dir object shouldn't have PDO lock */
2917 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2923 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2924 dlmflags |= LDLM_FL_ATOMIC_CB;
2927 * Take PDO lock on whole directory and build correct @res_id for lock
2928 * on part of directory.
2930 if (lh->mlh_pdo_hash != 0) {
2931 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2932 mdt_lock_pdo_mode(info, o, lh);
2933 if (lh->mlh_pdo_mode != LCK_NL) {
2935 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2936 * is never going to be sent to client and we do not
2937 * want it slowed down due to possible cancels.
2939 policy->l_inodebits.bits =
2940 *ibits & MDS_INODELOCK_UPDATE;
2941 policy->l_inodebits.try_bits =
2942 trybits & MDS_INODELOCK_UPDATE;
2943 /* at least one of them should be set */
2944 LASSERT(policy->l_inodebits.bits |
2945 policy->l_inodebits.try_bits);
2946 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2947 policy, res_id, dlmflags,
2948 info->mti_exp == NULL ? NULL :
2949 &info->mti_exp->exp_handle.h_cookie);
2950 if (unlikely(rc != 0))
2951 GOTO(out_unlock, rc);
2955 * Finish res_id initializing by name hash marking part of
2956 * directory which is taking modification.
2958 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2961 policy->l_inodebits.bits = *ibits;
2962 policy->l_inodebits.try_bits = trybits;
2965 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2966 * going to be sent to client. If it is - mdt_intent_policy() path will
2967 * fix it up and turn FL_LOCAL flag off.
2969 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2970 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2971 info->mti_exp == NULL ? NULL :
2972 &info->mti_exp->exp_handle.h_cookie);
2975 mdt_object_unlock(info, o, lh, 1);
2976 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2977 lh->mlh_pdo_hash != 0 &&
2978 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2979 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2981 /* Return successfully acquired bits to a caller */
2983 struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh);
2986 *ibits = lock->l_policy_data.l_inodebits.bits;
2987 LDLM_LOCK_PUT(lock);
2993 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2994 struct mdt_lock_handle *lh, __u64 *ibits,
2995 __u64 trybits, bool cos_incompat)
2997 struct mdt_lock_handle *local_lh = NULL;
3001 if (!mdt_object_remote(o)) {
3002 rc = mdt_object_local_lock(info, o, lh, ibits, trybits,
3007 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
3008 *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
3009 MDS_INODELOCK_XATTR);
3011 /* Only enqueue LOOKUP lock for remote object */
3012 if (*ibits & MDS_INODELOCK_LOOKUP) {
3013 __u64 local = MDS_INODELOCK_LOOKUP;
3015 rc = mdt_object_local_lock(info, o, lh, &local, 0,
3023 if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) {
3024 /* Sigh, PDO needs to enqueue 2 locks right now, but
3025 * enqueue RPC can only request 1 lock, to avoid extra
3026 * RPC, so it will instead enqueue EX lock for remote
3027 * object anyway XXX*/
3028 if (lh->mlh_type == MDT_PDO_LOCK &&
3029 lh->mlh_pdo_hash != 0) {
3030 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
3031 "EX lock.\n", mdt_obd_name(info->mti_mdt),
3032 PFID(mdt_object_fid(o)));
3033 lh->mlh_pdo_hash = 0;
3034 lh->mlh_rreg_mode = LCK_EX;
3035 lh->mlh_type = MDT_REG_LOCK;
3038 rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o),
3041 ibits, trybits, false);
3042 if (rc != ELDLM_OK) {
3043 if (local_lh != NULL)
3044 mdt_object_unlock(info, o, local_lh, rc);
3052 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3053 struct mdt_lock_handle *lh, __u64 ibits)
3055 return mdt_object_lock_internal(info, o, lh, &ibits, 0, false);
3058 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3059 struct mdt_lock_handle *lh, __u64 ibits,
3062 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
3063 return mdt_object_lock_internal(info, o, lh, &ibits, 0,
3067 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
3068 struct mdt_lock_handle *lh, __u64 *ibits,
3069 __u64 trybits, bool cos_incompat)
3071 bool trylock_only = *ibits == 0;
3074 LASSERT(!(*ibits & trybits));
3075 rc = mdt_object_lock_internal(info, o, lh, ibits, trybits,
3077 if (rc && trylock_only) { /* clear error for try ibits lock only */
3078 LASSERT(*ibits == 0);
3085 * Save a lock within request object.
3087 * Keep the lock referenced until whether client ACK or transaction
3088 * commit happens or release the lock immediately depending on input
3089 * parameters. If COS is ON, a write lock is converted to COS lock
3092 * \param info thead info object
3093 * \param h lock handle
3094 * \param mode lock mode
3095 * \param decref force immediate lock releasing
3097 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
3098 enum ldlm_mode mode, int decref)
3102 if (lustre_handle_is_used(h)) {
3103 if (decref || !info->mti_has_trans ||
3104 !(mode & (LCK_PW | LCK_EX))) {
3105 mdt_fid_unlock(h, mode);
3107 struct mdt_device *mdt = info->mti_mdt;
3108 struct ldlm_lock *lock = ldlm_handle2lock(h);
3109 struct ptlrpc_request *req = mdt_info_req(info);
3110 bool cos = mdt_cos_is_enabled(mdt);
3111 bool convert_lock = !cos && mdt_slc_is_enabled(mdt);
3113 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
3116 /* there is no request if mdt_object_unlock() is called
3117 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
3118 if (likely(req != NULL)) {
3119 LDLM_DEBUG(lock, "save lock request %p reply "
3120 "state %p transno %lld\n", req,
3121 req->rq_reply_state, req->rq_transno);
3123 ldlm_lock_mode_downgrade(lock, LCK_COS);
3126 if (req->rq_export->exp_disconnected)
3127 mdt_fid_unlock(h, mode);
3129 ptlrpc_save_lock(req, h, mode, cos,
3132 mdt_fid_unlock(h, mode);
3134 if (mdt_is_lock_sync(lock)) {
3135 CDEBUG(D_HA, "found sync-lock,"
3136 " async commit started\n");
3137 mdt_device_commit_async(info->mti_env,
3140 LDLM_LOCK_PUT(lock);
3149 * Save cross-MDT lock in uncommitted_slc_locks
3151 * Keep the lock referenced until transaction commit happens or release the lock
3152 * immediately depending on input parameters.
3154 * \param info thead info object
3155 * \param h lock handle
3156 * \param mode lock mode
3157 * \param decref force immediate lock releasing
3159 static void mdt_save_remote_lock(struct mdt_thread_info *info,
3160 struct mdt_object *o, struct lustre_handle *h,
3161 enum ldlm_mode mode, int decref)
3165 if (lustre_handle_is_used(h)) {
3166 struct ldlm_lock *lock = ldlm_handle2lock(h);
3169 (lock->l_policy_data.l_inodebits.bits &
3170 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
3171 mo_invalidate(info->mti_env, mdt_object_child(o));
3173 if (decref || !info->mti_has_trans ||
3174 !(mode & (LCK_PW | LCK_EX))) {
3175 ldlm_lock_decref_and_cancel(h, mode);
3176 LDLM_LOCK_PUT(lock);
3178 struct ptlrpc_request *req = mdt_info_req(info);
3180 LASSERT(req != NULL);
3181 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
3183 ldlm_lock_decref(h, mode);
3192 * Unlock mdt object.
3194 * Immeditely release the regular lock and the PDO lock or save the
3195 * lock in request and keep them referenced until client ACK or
3196 * transaction commit.
3198 * \param info thread info object
3199 * \param o mdt object
3200 * \param lh mdt lock handle referencing regular and PDO locks
3201 * \param decref force immediate lock releasing
3203 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
3205 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
3206 struct mdt_lock_handle *lh, int decref)
3210 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
3211 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
3212 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
3218 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
3219 const struct lu_fid *f,
3220 struct mdt_lock_handle *lh,
3223 struct mdt_object *o;
3225 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
3229 rc = mdt_object_lock(info, o, lh, ibits);
3231 mdt_object_put(info->mti_env, o);
3238 void mdt_object_unlock_put(struct mdt_thread_info * info,
3239 struct mdt_object * o,
3240 struct mdt_lock_handle *lh,
3243 mdt_object_unlock(info, o, lh, decref);
3244 mdt_object_put(info->mti_env, o);
3248 * Generic code handling requests that have struct mdt_body passed in:
3250 * - extract mdt_body from request and save it in @info, if present;
3252 * - create lu_object, corresponding to the fid in mdt_body, and save it in
3255 * - if HABEO_CORPUS flag is set for this request type check whether object
3256 * actually exists on storage (lu_object_exists()).
3259 static int mdt_body_unpack(struct mdt_thread_info *info,
3260 enum tgt_handler_flags flags)
3262 const struct mdt_body *body;
3263 struct mdt_object *obj;
3264 const struct lu_env *env;
3265 struct req_capsule *pill;
3269 env = info->mti_env;
3270 pill = info->mti_pill;
3272 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
3276 if (!(body->mbo_valid & OBD_MD_FLID))
3279 if (!fid_is_sane(&body->mbo_fid1)) {
3280 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
3284 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
3286 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
3287 mdt_object_put(env, obj);
3290 info->mti_object = obj;
3299 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
3300 enum tgt_handler_flags flags)
3302 struct req_capsule *pill = info->mti_pill;
3306 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
3307 rc = mdt_body_unpack(info, flags);
3311 if (rc == 0 && (flags & HABEO_REFERO)) {
3313 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3314 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
3316 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3317 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
3320 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3321 * by default. If the target object has more ACL entries, then
3322 * enlarge the buffer when necessary. */
3323 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3324 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3325 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3327 rc = req_capsule_server_pack(pill);
3332 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
3334 lh->mlh_type = MDT_NUL_LOCK;
3335 lh->mlh_reg_lh.cookie = 0ull;
3336 lh->mlh_reg_mode = LCK_MINMODE;
3337 lh->mlh_pdo_lh.cookie = 0ull;
3338 lh->mlh_pdo_mode = LCK_MINMODE;
3339 lh->mlh_rreg_lh.cookie = 0ull;
3340 lh->mlh_rreg_mode = LCK_MINMODE;
3343 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
3345 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3346 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3350 * Initialize fields of struct mdt_thread_info. Other fields are left in
3351 * uninitialized state, because it's too expensive to zero out whole
3352 * mdt_thread_info (> 1K) on each request arrival.
3354 void mdt_thread_info_init(struct ptlrpc_request *req,
3355 struct mdt_thread_info *info)
3359 info->mti_pill = &req->rq_pill;
3362 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3363 mdt_lock_handle_init(&info->mti_lh[i]);
3365 /* mdt device: it can be NULL while CONNECT */
3366 if (req->rq_export) {
3367 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
3368 info->mti_exp = req->rq_export;
3370 info->mti_mdt = NULL;
3371 info->mti_env = req->rq_svc_thread->t_env;
3372 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
3374 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
3375 info->mti_big_buf = LU_BUF_NULL;
3376 info->mti_body = NULL;
3377 info->mti_object = NULL;
3378 info->mti_dlm_req = NULL;
3379 info->mti_has_trans = 0;
3380 info->mti_cross_ref = 0;
3381 info->mti_opdata = 0;
3382 info->mti_big_lmm_used = 0;
3383 info->mti_big_acl_used = 0;
3384 info->mti_som_valid = 0;
3386 info->mti_spec.no_create = 0;
3387 info->mti_spec.sp_rm_entry = 0;
3388 info->mti_spec.sp_permitted = 0;
3389 info->mti_spec.sp_migrate_close = 0;
3391 info->mti_spec.u.sp_ea.eadata = NULL;
3392 info->mti_spec.u.sp_ea.eadatalen = 0;
3395 void mdt_thread_info_fini(struct mdt_thread_info *info)
3399 if (info->mti_object != NULL) {
3400 mdt_object_put(info->mti_env, info->mti_object);
3401 info->mti_object = NULL;
3404 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3405 mdt_lock_handle_fini(&info->mti_lh[i]);
3406 info->mti_env = NULL;
3407 info->mti_pill = NULL;
3408 info->mti_exp = NULL;
3410 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3411 lu_buf_free(&info->mti_big_buf);
3414 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3416 struct mdt_thread_info *mti;
3418 mti = mdt_th_info(tsi->tsi_env);
3419 LASSERT(mti != NULL);
3421 mdt_thread_info_init(tgt_ses_req(tsi), mti);
3422 if (tsi->tsi_corpus != NULL) {
3423 mti->mti_object = mdt_obj(tsi->tsi_corpus);
3424 lu_object_get(tsi->tsi_corpus);
3426 mti->mti_body = tsi->tsi_mdt_body;
3427 mti->mti_dlm_req = tsi->tsi_dlm_req;
3432 static int mdt_tgt_connect(struct tgt_session_info *tsi)
3434 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
3436 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
3437 set_current_state(TASK_UNINTERRUPTIBLE);
3438 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
3441 return tgt_connect(tsi);
3444 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
3445 struct mdt_thread_info *info,
3446 struct ldlm_lock **lockp, __u64 flags)
3448 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
3451 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
3452 struct mdt_thread_info *info,
3453 struct ldlm_lock **lockp, __u64 flags)
3455 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
3459 int mdt_intent_lock_replace(struct mdt_thread_info *info,
3460 struct ldlm_lock **lockp,
3461 struct mdt_lock_handle *lh,
3462 __u64 flags, int result)
3464 struct ptlrpc_request *req = mdt_info_req(info);
3465 struct ldlm_lock *lock = *lockp;
3466 struct ldlm_lock *new_lock;
3468 /* If possible resent found a lock, @lh is set to its handle */
3469 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3471 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3472 lh->mlh_reg_lh.cookie = 0;
3476 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
3477 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3478 * a resend case, however, it could be already destroyed
3479 * due to client eviction or a raced cancel RPC. */
3480 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
3481 lh->mlh_reg_lh.cookie);
3482 lh->mlh_reg_lh.cookie = 0;
3486 LASSERTF(new_lock != NULL,
3487 "lockh %#llx flags %#llx : rc = %d\n",
3488 lh->mlh_reg_lh.cookie, flags, result);
3491 * If we've already given this lock to a client once, then we should
3492 * have no readers or writers. Otherwise, we should have one reader
3493 * _or_ writer ref (which will be zeroed below) before returning the
3496 if (new_lock->l_export == req->rq_export) {
3497 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3499 LASSERT(new_lock->l_export == NULL);
3500 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3505 if (new_lock->l_export == req->rq_export) {
3507 * Already gave this to the client, which means that we
3508 * reconstructed a reply.
3510 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3513 LDLM_LOCK_RELEASE(new_lock);
3514 lh->mlh_reg_lh.cookie = 0;
3515 RETURN(ELDLM_LOCK_REPLACED);
3519 * Fixup the lock to be given to the client.
3521 lock_res_and_lock(new_lock);
3522 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3523 * possible blocking AST. */
3524 while (new_lock->l_readers > 0) {
3525 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3526 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3527 new_lock->l_readers--;
3529 while (new_lock->l_writers > 0) {
3530 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3531 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3532 new_lock->l_writers--;
3535 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3536 new_lock->l_blocking_ast = lock->l_blocking_ast;
3537 new_lock->l_completion_ast = lock->l_completion_ast;
3538 if (ldlm_has_dom(new_lock))
3539 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
3540 new_lock->l_remote_handle = lock->l_remote_handle;
3541 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3543 unlock_res_and_lock(new_lock);
3545 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3546 &new_lock->l_remote_handle,
3547 &new_lock->l_exp_hash);
3549 LDLM_LOCK_RELEASE(new_lock);
3550 lh->mlh_reg_lh.cookie = 0;
3552 RETURN(ELDLM_LOCK_REPLACED);
3555 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3556 struct ldlm_lock *new_lock,
3557 struct mdt_lock_handle *lh, __u64 flags)
3559 struct ptlrpc_request *req = mdt_info_req(info);
3560 struct ldlm_request *dlmreq;
3562 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3565 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3567 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3568 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3570 if (flags & LDLM_FL_RESENT) {
3571 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3572 lh->mlh_reg_mode = new_lock->l_granted_mode;
3574 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3575 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
3576 lh->mlh_reg_lh.cookie);
3581 * If the xid matches, then we know this is a resent request, and allow
3582 * it. (It's probably an OPEN, for which we don't send a lock.
3584 if (req_can_reconstruct(req, NULL))
3588 * This remote handle isn't enqueued, so we never received or processed
3589 * this request. Clear MSG_RESENT, because it can be handled like any
3590 * normal request now.
3592 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3594 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
3595 dlmreq->lock_handle[0].cookie);
3598 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
3599 struct mdt_thread_info *info,
3600 struct ldlm_lock **lockp,
3603 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3604 struct ldlm_reply *ldlm_rep = NULL;
3609 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3610 * (for the resend case) or a new lock. Below we will use it to
3611 * replace the original lock.
3613 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3614 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3615 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3616 rc = mdt_object_lock(info, info->mti_object, lhc,
3617 MDS_INODELOCK_XATTR);
3622 rc = mdt_getxattr(info);
3624 if (mdt_info_req(info)->rq_repmsg != NULL)
3625 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3627 if (ldlm_rep == NULL ||
3628 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
3629 mdt_object_unlock(info, info->mti_object, lhc, 1);
3630 RETURN(err_serious(-EFAULT));
3633 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3635 /* This is left for interop instead of adding a new interop flag.
3637 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
3638 if (ldlm_rep->lock_policy_res2) {
3639 mdt_object_unlock(info, info->mti_object, lhc, 1);
3640 RETURN(ELDLM_LOCK_ABORTED);
3644 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3648 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
3649 struct mdt_thread_info *info,
3650 struct ldlm_lock **lockp,
3653 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3655 struct ldlm_reply *ldlm_rep;
3656 struct mdt_body *reqbody;
3657 struct mdt_body *repbody;
3661 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3664 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3667 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3668 repbody->mbo_eadatasize = 0;
3669 repbody->mbo_aclsize = 0;
3673 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3676 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3680 CERROR("%s: unsupported intent %#x\n",
3681 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
3682 GOTO(out_shrink, rc = -EINVAL);
3685 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3687 GOTO(out_shrink, rc);
3689 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3690 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3692 /* Get lock from request for possible resent case. */
3693 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3695 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3696 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3698 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3699 ldlm_rep->lock_policy_res2 = 0;
3700 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3701 ldlm_rep->lock_policy_res2) {
3702 lhc->mlh_reg_lh.cookie = 0ull;
3703 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3706 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3709 mdt_exit_ucred(info);
3711 mdt_client_compatibility(info);
3712 rc2 = mdt_fix_reply(info);
3718 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
3719 struct mdt_thread_info *info,
3720 struct ldlm_lock **lockp,
3723 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3724 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
3725 struct layout_intent *intent;
3726 struct lu_fid *fid = &info->mti_tmp_fid2;
3727 struct mdt_object *obj = NULL;
3728 int layout_size = 0;
3732 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3734 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3738 CDEBUG(D_INFO, DFID "got layout change request from client: "
3739 "opc:%u flags:%#x extent "DEXT"\n",
3740 PFID(fid), intent->li_opc, intent->li_flags,
3741 PEXT(&intent->li_extent));
3743 switch (intent->li_opc) {
3744 case LAYOUT_INTENT_TRUNC:
3745 case LAYOUT_INTENT_WRITE:
3746 layout.mlc_opc = MD_LAYOUT_WRITE;
3747 layout.mlc_intent = intent;
3749 case LAYOUT_INTENT_ACCESS:
3751 case LAYOUT_INTENT_READ:
3752 case LAYOUT_INTENT_GLIMPSE:
3753 case LAYOUT_INTENT_RELEASE:
3754 case LAYOUT_INTENT_RESTORE:
3755 CERROR("%s: Unsupported layout intent opc %d\n",
3756 mdt_obd_name(info->mti_mdt), intent->li_opc);
3760 CERROR("%s: Unknown layout intent opc %d\n",
3761 mdt_obd_name(info->mti_mdt), intent->li_opc);
3768 /* Get lock from request for possible resent case. */
3769 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3771 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3773 GOTO(out, rc = PTR_ERR(obj));
3776 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3777 /* if layout is going to be changed don't use the current EA
3778 * size but the maximum one. That buffer will be shrinked
3779 * to the actual size in req_capsule_shrink() before reply.
3781 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
3782 layout_size = info->mti_mdt->mdt_max_mdsize;
3784 layout_size = mdt_attr_get_eabuf_size(info, obj);
3785 if (layout_size < 0)
3786 GOTO(out_obj, rc = layout_size);
3788 if (layout_size > info->mti_mdt->mdt_max_mdsize)
3789 info->mti_mdt->mdt_max_mdsize = layout_size;
3794 * set reply buffer size, so that ldlm_handle_enqueue0()->
3795 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
3797 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3798 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3800 rc = req_capsule_server_pack(info->mti_pill);
3805 if (layout.mlc_opc != MD_LAYOUT_NOP) {
3806 struct lu_buf *buf = &layout.mlc_buf;
3809 * mdt_layout_change is a reint operation, when the request
3810 * is resent, layout write shouldn't reprocess it again.
3812 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
3814 GOTO(out_obj, rc = rc < 0 ? rc : 0);
3817 * There is another resent case: the client's job has been
3818 * done by another client, referring lod_declare_layout_change
3819 * -EALREADY case, and it became a operation w/o transaction,
3820 * so we should not do the layout change, otherwise
3821 * mdt_layout_change() will try to cancel the granted server
3822 * CR lock whose remote counterpart is still in hold on the
3823 * client, and a deadlock ensues.
3825 rc = mdt_check_resent_lock(info, obj, lhc);
3831 if (unlikely(req_is_replay(mdt_info_req(info)))) {
3832 buf->lb_buf = req_capsule_client_get(info->mti_pill,
3834 buf->lb_len = req_capsule_get_size(info->mti_pill,
3835 &RMF_EADATA, RCL_CLIENT);
3837 * If it's a replay of layout write intent RPC, the
3838 * client has saved the extended lovea when
3839 * it get reply then.
3841 if (buf->lb_len > 0)
3842 mdt_fix_lov_magic(info, buf->lb_buf);
3845 * Instantiate some layout components, if @buf contains
3846 * lovea, then it's a replay of the layout intent write
3849 rc = mdt_layout_change(info, obj, &layout);
3854 mdt_object_put(info->mti_env, obj);
3856 if (rc == 0 && lustre_handle_is_used(&lhc->mlh_reg_lh))
3857 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3860 lhc->mlh_reg_lh.cookie = 0;
3865 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
3866 struct mdt_thread_info *info,
3867 struct ldlm_lock **lockp,
3870 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3871 struct ldlm_reply *rep = NULL;
3875 static const struct req_format *intent_fmts[REINT_MAX] = {
3876 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3877 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3882 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3886 /* Get lock from request for possible resent case. */
3887 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3889 rc = mdt_reint_internal(info, lhc, opc);
3891 /* Check whether the reply has been packed successfully. */
3892 if (mdt_info_req(info)->rq_repmsg != NULL)
3893 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3895 RETURN(err_serious(-EFAULT));
3897 /* MDC expects this in any case */
3899 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3901 /* the open lock or the lock for cross-ref object should be
3902 * returned to the client */
3903 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3904 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3905 rep->lock_policy_res2 = 0;
3906 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3910 rep->lock_policy_res2 = clear_serious(rc);
3912 if (rep->lock_policy_res2 == -ENOENT &&
3913 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
3914 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
3915 rep->lock_policy_res2 = 0;
3917 lhc->mlh_reg_lh.cookie = 0ull;
3918 if (rc == -ENOTCONN || rc == -ENODEV ||
3919 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3921 * If it is the disconnect error (ENODEV & ENOCONN), the error
3922 * will be returned by rq_status, and client at ptlrpc layer
3923 * will detect this, then disconnect, reconnect the import
3924 * immediately, instead of impacting the following the rpc.
3929 * For other cases, the error will be returned by intent, and client
3930 * will retrieve the result from intent.
3932 RETURN(ELDLM_LOCK_ABORTED);
3935 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
3936 struct mdt_thread_info *info,
3937 struct ldlm_lock **lockp,
3938 u64 flags /* LDLM_FL_* */)
3940 struct req_capsule *pill = info->mti_pill;
3941 struct ptlrpc_request *req = mdt_info_req(info);
3942 const struct req_format *it_format;
3943 int (*it_handler)(enum ldlm_intent_flags,
3944 struct mdt_thread_info *,
3945 struct ldlm_lock **,
3947 enum tgt_handler_flags it_handler_flags = 0;
3948 struct ldlm_reply *rep;
3954 case IT_OPEN|IT_CREAT:
3956 * OCREAT is not a MUTABOR request since the file may
3957 * already exist. We do the extra check of
3958 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
3959 * really need to create the object.
3961 it_format = &RQF_LDLM_INTENT;
3962 it_handler = &mdt_intent_open;
3966 it_format = &RQF_LDLM_INTENT_GETATTR;
3967 it_handler = &mdt_intent_getattr;
3968 it_handler_flags = HABEO_REFERO;
3971 it_format = &RQF_LDLM_INTENT_GETXATTR;
3972 it_handler = &mdt_intent_getxattr;
3973 it_handler_flags = HABEO_CORPUS;
3976 it_format = &RQF_LDLM_INTENT_LAYOUT;
3977 it_handler = &mdt_intent_layout;
3980 it_format = &RQF_LDLM_INTENT;
3981 it_handler = &mdt_intent_glimpse;
3984 it_format = &RQF_LDLM_INTENT;
3985 it_handler = &mdt_intent_brw;
3987 case IT_QUOTA_DQACQ:
3988 case IT_QUOTA_CONN: {
3989 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3992 RETURN(-EOPNOTSUPP);
3994 if (mdt_rdonly(req->rq_export))
3997 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3998 /* pass the request to quota master */
3999 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
4000 mdt_info_req(info), lockp,
4005 CERROR("%s: unknown intent code %#x\n",
4006 mdt_obd_name(info->mti_mdt), it_opc);
4010 req_capsule_extend(pill, it_format);
4012 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
4016 if (it_handler_flags & MUTABOR && mdt_rdonly(req->rq_export))
4019 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
4021 /* execute policy */
4022 rc = (*it_handler)(it_opc, info, lockp, flags);
4024 /* Check whether the reply has been packed successfully. */
4025 if (req->rq_repmsg != NULL) {
4026 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4027 rep->lock_policy_res2 =
4028 ptlrpc_status_hton(rep->lock_policy_res2);
4034 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
4035 enum ldlm_intent_flags it_opc)
4037 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
4039 /* update stats when IT code is known */
4040 if (srv_stats != NULL)
4041 lprocfs_counter_incr(srv_stats,
4042 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
4043 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
4046 static int mdt_intent_policy(struct ldlm_namespace *ns,
4047 struct ldlm_lock **lockp, void *req_cookie,
4048 enum ldlm_mode mode, __u64 flags, void *data)
4050 struct tgt_session_info *tsi;
4051 struct mdt_thread_info *info;
4052 struct ptlrpc_request *req = req_cookie;
4053 struct ldlm_intent *it;
4054 struct req_capsule *pill;
4055 const struct ldlm_lock_desc *ldesc;
4060 LASSERT(req != NULL);
4062 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
4064 info = tsi2mdt_info(tsi);
4065 LASSERT(info != NULL);
4066 pill = info->mti_pill;
4067 LASSERT(pill->rc_req == req);
4068 ldesc = &info->mti_dlm_req->lock_desc;
4070 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
4071 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
4072 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
4074 mdt_ptlrpc_stats_update(req, it->opc);
4075 rc = mdt_intent_opc(it->opc, info, lockp, flags);
4079 /* Lock without inodebits makes no sense and will oops
4080 * later in ldlm. Let's check it now to see if we have
4081 * ibits corrupted somewhere in mdt_intent_opc().
4082 * The case for client miss to set ibits has been
4083 * processed by others. */
4084 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
4085 ldesc->l_policy_data.l_inodebits.bits != 0));
4087 rc = err_serious(-EFAULT);
4090 /* No intent was provided */
4091 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
4092 rc = req_capsule_server_pack(pill);
4094 rc = err_serious(rc);
4096 mdt_thread_info_fini(info);
4100 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
4102 struct seq_server_site *ss = mdt_seq_site(mdt);
4104 if (ss->ss_node_id == 0)
4107 if (ss->ss_client_seq != NULL) {
4108 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4109 ss->ss_client_seq->lcs_exp = NULL;
4112 if (ss->ss_server_fld != NULL) {
4113 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
4114 ss->ss_server_fld->lsf_control_exp = NULL;
4118 static void mdt_seq_fini_cli(struct mdt_device *mdt)
4120 struct seq_server_site *ss = mdt_seq_site(mdt);
4125 if (ss->ss_server_seq != NULL)
4126 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
4129 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
4131 mdt_seq_fini_cli(mdt);
4132 mdt_deregister_seq_exp(mdt);
4134 return seq_site_fini(env, mdt_seq_site(mdt));
4138 * It will retrieve its FLDB entries from MDT0, and it only happens
4139 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
4140 * and it needs to refresh FLDB from the MDT0.
4142 static int mdt_register_lwp_callback(void *data)
4145 struct mdt_device *mdt = data;
4146 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
4150 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
4152 rc = lu_env_init(&env, LCT_MD_THREAD);
4154 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
4158 /* Allocate new sequence now to avoid creating local transaction
4159 * in the normal transaction process */
4160 rc = seq_server_check_and_alloc_super(&env,
4161 mdt_seq_site(mdt)->ss_server_seq);
4166 rc = fld_update_from_controller(&env, fld);
4168 CERROR("%s: cannot update controller: rc = %d\n",
4169 mdt_obd_name(mdt), rc);
4178 static int mdt_register_seq_exp(struct mdt_device *mdt)
4180 struct seq_server_site *ss = mdt_seq_site(mdt);
4181 char *lwp_name = NULL;
4184 if (ss->ss_node_id == 0)
4187 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
4188 if (lwp_name == NULL)
4189 GOTO(out_free, rc = -ENOMEM);
4191 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
4195 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
4200 rc = lustre_register_lwp_item(lwp_name,
4201 &ss->ss_server_fld->lsf_control_exp,
4202 mdt_register_lwp_callback, mdt);
4204 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4205 ss->ss_client_seq->lcs_exp = NULL;
4209 if (lwp_name != NULL)
4210 OBD_FREE(lwp_name, MAX_OBD_NAME);
4216 * Init client sequence manager which is used by local MDS to talk to sequence
4217 * controller on remote node.
4219 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
4221 struct seq_server_site *ss = mdt_seq_site(mdt);
4226 /* check if this is adding the first MDC and controller is not yet
4228 OBD_ALLOC_PTR(ss->ss_client_seq);
4229 if (ss->ss_client_seq == NULL)
4232 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
4233 if (prefix == NULL) {
4234 OBD_FREE_PTR(ss->ss_client_seq);
4235 ss->ss_client_seq = NULL;
4239 /* Note: seq_client_fini will be called in seq_site_fini */
4240 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
4241 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
4242 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
4244 OBD_FREE(prefix, MAX_OBD_NAME + 5);
4246 OBD_FREE_PTR(ss->ss_client_seq);
4247 ss->ss_client_seq = NULL;
4251 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
4256 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
4258 struct seq_server_site *ss;
4262 ss = mdt_seq_site(mdt);
4263 /* init sequence controller server(MDT0) */
4264 if (ss->ss_node_id == 0) {
4265 OBD_ALLOC_PTR(ss->ss_control_seq);
4266 if (ss->ss_control_seq == NULL)
4269 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
4270 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
4273 GOTO(out_seq_fini, rc);
4276 /* Init normal sequence server */
4277 OBD_ALLOC_PTR(ss->ss_server_seq);
4278 if (ss->ss_server_seq == NULL)
4279 GOTO(out_seq_fini, rc = -ENOMEM);
4281 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
4282 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
4284 GOTO(out_seq_fini, rc);
4286 /* init seq client for seq server to talk to seq controller(MDT0) */
4287 rc = mdt_seq_init_cli(env, mdt);
4289 GOTO(out_seq_fini, rc);
4291 if (ss->ss_node_id != 0)
4292 /* register controller export through lwp */
4293 rc = mdt_register_seq_exp(mdt);
4298 mdt_seq_fini(env, mdt);
4306 static int mdt_fld_fini(const struct lu_env *env,
4307 struct mdt_device *m)
4309 struct seq_server_site *ss = mdt_seq_site(m);
4312 if (ss && ss->ss_server_fld) {
4313 fld_server_fini(env, ss->ss_server_fld);
4314 OBD_FREE_PTR(ss->ss_server_fld);
4315 ss->ss_server_fld = NULL;
4321 static int mdt_fld_init(const struct lu_env *env,
4323 struct mdt_device *m)
4325 struct seq_server_site *ss;
4329 ss = mdt_seq_site(m);
4331 OBD_ALLOC_PTR(ss->ss_server_fld);
4332 if (ss->ss_server_fld == NULL)
4333 RETURN(rc = -ENOMEM);
4335 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
4338 OBD_FREE_PTR(ss->ss_server_fld);
4339 ss->ss_server_fld = NULL;
4346 static void mdt_stack_pre_fini(const struct lu_env *env,
4347 struct mdt_device *m, struct lu_device *top)
4349 struct lustre_cfg_bufs *bufs;
4350 struct lustre_cfg *lcfg;
4351 struct mdt_thread_info *info;
4356 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4357 LASSERT(info != NULL);
4359 bufs = &info->mti_u.bufs;
4361 LASSERT(m->mdt_child_exp);
4362 LASSERT(m->mdt_child_exp->exp_obd);
4364 /* process cleanup, pass mdt obd name to get obd umount flags */
4365 /* XXX: this is needed because all layers are referenced by
4366 * objects (some of them are pinned by osd, for example *
4367 * the proper solution should be a model where object used
4368 * by osd only doesn't have mdt/mdd slices -bzzz */
4369 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4370 lustre_cfg_bufs_set_string(bufs, 1, NULL);
4371 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4374 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
4376 top->ld_ops->ldo_process_config(env, top, lcfg);
4377 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4381 static void mdt_stack_fini(const struct lu_env *env,
4382 struct mdt_device *m, struct lu_device *top)
4384 struct obd_device *obd = mdt2obd_dev(m);
4385 struct lustre_cfg_bufs *bufs;
4386 struct lustre_cfg *lcfg;
4387 struct mdt_thread_info *info;
4391 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4392 LASSERT(info != NULL);
4394 lu_dev_del_linkage(top->ld_site, top);
4396 lu_site_purge(env, top->ld_site, -1);
4398 bufs = &info->mti_u.bufs;
4399 /* process cleanup, pass mdt obd name to get obd umount flags */
4400 /* another purpose is to let all layers to release their objects */
4401 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4406 lustre_cfg_bufs_set_string(bufs, 1, flags);
4407 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4410 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
4413 top->ld_ops->ldo_process_config(env, top, lcfg);
4414 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4416 lu_site_purge(env, top->ld_site, -1);
4418 m->mdt_child = NULL;
4419 m->mdt_bottom = NULL;
4421 obd_disconnect(m->mdt_child_exp);
4422 m->mdt_child_exp = NULL;
4424 obd_disconnect(m->mdt_bottom_exp);
4425 m->mdt_child_exp = NULL;
4428 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
4429 const char *next, struct obd_export **exp)
4431 struct obd_connect_data *data = NULL;
4432 struct obd_device *obd;
4436 OBD_ALLOC_PTR(data);
4438 GOTO(out, rc = -ENOMEM);
4440 obd = class_name2obd(next);
4442 CERROR("%s: can't locate next device: %s\n",
4443 mdt_obd_name(m), next);
4444 GOTO(out, rc = -ENOTCONN);
4447 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4448 data->ocd_version = LUSTRE_VERSION_CODE;
4450 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
4452 CERROR("%s: cannot connect to next dev %s (%d)\n",
4453 mdt_obd_name(m), next, rc);
4463 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
4464 struct lustre_cfg *cfg)
4466 char *dev = lustre_cfg_string(cfg, 0);
4467 int rc, name_size, uuid_size;
4468 char *name, *uuid, *p;
4469 struct lustre_cfg_bufs *bufs;
4470 struct lustre_cfg *lcfg;
4471 struct obd_device *obd;
4472 struct lustre_profile *lprof;
4473 struct lu_site *site;
4476 /* in 1.8 we had the only device in the stack - MDS.
4477 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
4478 * in 2.3 OSD is instantiated by obd_mount.c, so we need
4479 * to generate names and setup MDT, MDD. MDT will be using
4480 * generated name to connect to MDD. for MDD the next device
4481 * will be LOD with name taken from so called "profile" which
4482 * is generated by mount_option line
4484 * 1.8 MGS generates config. commands like this:
4485 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
4486 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
4487 * 2.0 MGS generates config. commands like this:
4488 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
4489 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4490 * 3:lustre-MDT0000-mdtlov 4:f
4492 * we generate MDD name from MDT one, just replacing T with D
4494 * after all the preparations, the logical equivalent will be
4495 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
4496 * 3:lustre-MDT0000-mdtlov 4:f
4497 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4498 * 3:lustre-MDD0000 4:f
4500 * notice we build the stack from down to top: MDD first, then MDT */
4502 name_size = MAX_OBD_NAME;
4503 uuid_size = MAX_OBD_NAME;
4505 OBD_ALLOC(name, name_size);
4506 OBD_ALLOC(uuid, uuid_size);
4507 if (name == NULL || uuid == NULL)
4508 GOTO(cleanup_mem, rc = -ENOMEM);
4510 OBD_ALLOC_PTR(bufs);
4512 GOTO(cleanup_mem, rc = -ENOMEM);
4515 p = strstr(name, "-MDT");
4517 GOTO(free_bufs, rc = -ENOMEM);
4520 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
4522 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4523 if (lprof == NULL || lprof->lp_dt == NULL) {
4524 CERROR("can't find the profile: %s\n",
4525 lustre_cfg_string(cfg, 0));
4526 GOTO(free_bufs, rc = -EINVAL);
4529 lustre_cfg_bufs_reset(bufs, name);
4530 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
4531 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4532 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4534 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4536 GOTO(put_profile, rc = -ENOMEM);
4537 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4539 rc = class_attach(lcfg);
4541 GOTO(lcfg_cleanup, rc);
4543 obd = class_name2obd(name);
4545 CERROR("Can not find obd %s (%s in config)\n",
4546 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
4547 GOTO(lcfg_cleanup, rc = -EINVAL);
4550 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4552 lustre_cfg_bufs_reset(bufs, name);
4553 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4554 lustre_cfg_bufs_set_string(bufs, 2, dev);
4555 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4557 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4559 GOTO(class_detach, rc = -ENOMEM);
4560 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4562 rc = class_setup(obd, lcfg);
4564 GOTO(class_detach, rc);
4566 /* connect to MDD we just setup */
4567 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
4569 GOTO(class_detach, rc);
4571 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
4573 LASSERT(mdt_lu_site(mdt) == NULL);
4574 mdt->mdt_lu_dev.ld_site = site;
4575 site->ls_top_dev = &mdt->mdt_lu_dev;
4576 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
4578 /* now connect to bottom OSD */
4579 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
4580 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
4582 GOTO(class_detach, rc);
4584 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
4586 rc = lu_env_refill((struct lu_env *)env);
4588 CERROR("Failure to refill session: '%d'\n", rc);
4590 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
4595 class_detach(obd, lcfg);
4597 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4599 class_put_profile(lprof);
4604 OBD_FREE(name, name_size);
4606 OBD_FREE(uuid, uuid_size);
4610 /* setup quota master target on MDT0 */
4611 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
4612 struct lustre_cfg *cfg)
4614 struct obd_device *obd;
4615 char *dev = lustre_cfg_string(cfg, 0);
4616 char *qmtname, *uuid, *p;
4617 struct lustre_cfg_bufs *bufs;
4618 struct lustre_cfg *lcfg;
4619 struct lustre_profile *lprof;
4620 struct obd_connect_data *data;
4624 LASSERT(mdt->mdt_qmt_exp == NULL);
4625 LASSERT(mdt->mdt_qmt_dev == NULL);
4627 /* quota master is on MDT0 only for now */
4628 if (mdt->mdt_seq_site.ss_node_id != 0)
4631 /* MGS generates config commands which look as follows:
4632 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4633 * 3:lustre-MDT0000-mdtlov 4:f
4635 * We generate the QMT name from the MDT one, just replacing MD with QM
4636 * after all the preparations, the logical equivalent will be:
4637 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4638 * 3:lustre-MDT0000-osd 4:f */
4639 OBD_ALLOC(qmtname, MAX_OBD_NAME);
4640 OBD_ALLOC(uuid, UUID_MAX);
4641 OBD_ALLOC_PTR(bufs);
4642 OBD_ALLOC_PTR(data);
4643 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
4644 GOTO(cleanup_mem, rc = -ENOMEM);
4646 strcpy(qmtname, dev);
4647 p = strstr(qmtname, "-MDT");
4649 GOTO(cleanup_mem, rc = -ENOMEM);
4650 /* replace MD with QM */
4654 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4656 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4657 if (lprof == NULL || lprof->lp_dt == NULL) {
4658 CERROR("can't find profile for %s\n",
4659 lustre_cfg_string(cfg, 0));
4660 GOTO(cleanup_mem, rc = -EINVAL);
4663 lustre_cfg_bufs_reset(bufs, qmtname);
4664 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4665 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4666 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4668 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4670 GOTO(put_profile, rc = -ENOMEM);
4671 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4673 rc = class_attach(lcfg);
4675 GOTO(lcfg_cleanup, rc);
4677 obd = class_name2obd(qmtname);
4679 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4680 lustre_cfg_string(cfg, 0));
4681 GOTO(lcfg_cleanup, rc = -EINVAL);
4684 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4686 lustre_cfg_bufs_reset(bufs, qmtname);
4687 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4688 lustre_cfg_bufs_set_string(bufs, 2, dev);
4690 /* for quota, the next device should be the OSD device */
4691 lustre_cfg_bufs_set_string(bufs, 3,
4692 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4694 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4696 GOTO(class_detach, rc = -ENOMEM);
4697 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4699 rc = class_setup(obd, lcfg);
4701 GOTO(class_detach, rc);
4703 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4705 /* configure local quota objects */
4706 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4710 GOTO(class_cleanup, rc);
4712 /* connect to quota master target */
4713 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4714 data->ocd_version = LUSTRE_VERSION_CODE;
4715 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4718 CERROR("cannot connect to quota master device %s (%d)\n",
4720 GOTO(class_cleanup, rc);
4726 class_manual_cleanup(obd);
4727 mdt->mdt_qmt_dev = NULL;
4731 class_detach(obd, lcfg);
4733 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4735 class_put_profile(lprof);
4740 OBD_FREE(qmtname, MAX_OBD_NAME);
4742 OBD_FREE(uuid, UUID_MAX);
4748 /* Shutdown quota master target associated with mdt */
4749 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4753 if (mdt->mdt_qmt_exp == NULL)
4755 LASSERT(mdt->mdt_qmt_dev != NULL);
4757 /* the qmt automatically shuts down when the mdt disconnects */
4758 obd_disconnect(mdt->mdt_qmt_exp);
4759 mdt->mdt_qmt_exp = NULL;
4760 mdt->mdt_qmt_dev = NULL;
4764 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4765 * for now. This will be removed along with converting rest of MDT code
4766 * to use tgt_session_info */
4767 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4769 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4772 rc = mdt_getxattr(info);
4774 mdt_thread_info_fini(info);
4778 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
4779 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
4780 #define OST_BRW_READ OST_READ
4781 #define OST_BRW_WRITE OST_WRITE
4783 static struct tgt_handler mdt_tgt_handlers[] = {
4784 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4785 0, MDS_CONNECT, mdt_tgt_connect,
4786 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4787 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4788 0, MDS_DISCONNECT, tgt_disconnect,
4789 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4790 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4791 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4792 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4793 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4794 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GET_ROOT, mdt_get_root),
4795 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4796 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4798 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4799 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4800 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4801 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4802 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4803 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4804 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4805 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4807 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4808 mdt_hsm_ct_register),
4809 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4810 mdt_hsm_ct_unregister),
4811 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4813 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4815 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4816 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4818 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4823 static struct tgt_handler mdt_io_ops[] = {
4824 TGT_OST_HDL_HP(HABEO_CORPUS | HABEO_REFERO, OST_BRW_READ, tgt_brw_read,
4826 TGT_OST_HDL_HP(HABEO_CORPUS | MUTABOR, OST_BRW_WRITE, tgt_brw_write,
4828 TGT_OST_HDL_HP(HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4829 OST_PUNCH, mdt_punch_hdl,
4831 TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO, OST_SYNC, mdt_data_sync),
4834 static struct tgt_handler mdt_sec_ctx_ops[] = {
4835 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4836 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4837 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4840 static struct tgt_handler mdt_quota_ops[] = {
4841 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4844 static struct tgt_opc_slice mdt_common_slice[] = {
4846 .tos_opc_start = MDS_FIRST_OPC,
4847 .tos_opc_end = MDS_LAST_OPC,
4848 .tos_hs = mdt_tgt_handlers
4851 .tos_opc_start = OBD_FIRST_OPC,
4852 .tos_opc_end = OBD_LAST_OPC,
4853 .tos_hs = tgt_obd_handlers
4856 .tos_opc_start = LDLM_FIRST_OPC,
4857 .tos_opc_end = LDLM_LAST_OPC,
4858 .tos_hs = tgt_dlm_handlers
4861 .tos_opc_start = SEC_FIRST_OPC,
4862 .tos_opc_end = SEC_LAST_OPC,
4863 .tos_hs = mdt_sec_ctx_ops
4866 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4867 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4868 .tos_hs = tgt_out_handlers
4871 .tos_opc_start = FLD_FIRST_OPC,
4872 .tos_opc_end = FLD_LAST_OPC,
4873 .tos_hs = fld_handlers
4876 .tos_opc_start = SEQ_FIRST_OPC,
4877 .tos_opc_end = SEQ_LAST_OPC,
4878 .tos_hs = seq_handlers
4881 .tos_opc_start = QUOTA_DQACQ,
4882 .tos_opc_end = QUOTA_LAST_OPC,
4883 .tos_hs = mdt_quota_ops
4886 .tos_opc_start = LLOG_FIRST_OPC,
4887 .tos_opc_end = LLOG_LAST_OPC,
4888 .tos_hs = tgt_llog_handlers
4891 .tos_opc_start = LFSCK_FIRST_OPC,
4892 .tos_opc_end = LFSCK_LAST_OPC,
4893 .tos_hs = tgt_lfsck_handlers
4896 .tos_opc_start = OST_FIRST_OPC,
4897 .tos_opc_end = OST_LAST_OPC,
4898 .tos_hs = mdt_io_ops
4905 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4907 struct md_device *next = m->mdt_child;
4908 struct lu_device *d = &m->mdt_lu_dev;
4909 struct obd_device *obd = mdt2obd_dev(m);
4910 struct lfsck_stop stop;
4913 stop.ls_status = LS_PAUSED;
4915 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4917 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4918 ping_evictor_stop();
4920 /* Remove the HSM /proc entry so the coordinator cannot be
4921 * restarted by a user while it's shutting down. */
4922 hsm_cdt_procfs_fini(m);
4923 mdt_hsm_cdt_stop(m);
4925 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4926 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4928 if (m->mdt_namespace != NULL)
4929 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
4930 d->ld_obd->obd_force);
4932 obd_exports_barrier(obd);
4933 obd_zombie_barrier();
4935 mdt_quota_fini(env, m);
4937 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4939 /* Calling the cleanup functions in the same order as in the mdt_init0
4944 target_recovery_fini(obd);
4945 upcall_cache_cleanup(m->mdt_identity_cache);
4946 m->mdt_identity_cache = NULL;
4948 mdt_fs_cleanup(env, m);
4950 tgt_fini(env, &m->mdt_lut);
4952 mdt_hsm_cdt_fini(m);
4954 if (m->mdt_los != NULL) {
4955 local_oid_storage_fini(env, m->mdt_los);
4959 if (m->mdt_namespace != NULL) {
4960 ldlm_namespace_free_post(m->mdt_namespace);
4961 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4964 if (m->mdt_md_root != NULL) {
4965 mdt_object_put(env, m->mdt_md_root);
4966 m->mdt_md_root = NULL;
4969 mdt_seq_fini(env, m);
4971 mdt_fld_fini(env, m);
4976 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4978 LASSERT(atomic_read(&d->ld_ref) == 0);
4980 server_put_mount(mdt_obd_name(m), true);
4985 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4987 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4988 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4990 const struct dt_device_param *dt_conf;
4991 struct mdt_thread_info *info;
4992 struct obd_device *obd;
4993 const char *dev = lustre_cfg_string(cfg, 0);
4994 const char *num = lustre_cfg_string(cfg, 2);
4995 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
4996 struct lustre_mount_info *lmi = NULL;
4997 struct lustre_sb_info *lsi;
4999 struct seq_server_site *ss_site;
5000 const char *identity_upcall = "NONE";
5001 struct md_device *next;
5008 lu_device_init(&m->mdt_lu_dev, ldt);
5010 * Environment (env) might be missing mdt_thread_key values at that
5011 * point, if device is allocated when mdt_thread_key is in QUIESCENT
5014 * Usually device allocation path doesn't use module key values, but
5015 * mdt has to do a lot of work here, so allocate key value.
5017 rc = lu_env_refill((struct lu_env *)env);
5021 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5022 LASSERT(info != NULL);
5024 obd = class_name2obd(dev);
5025 LASSERT(obd != NULL);
5027 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
5028 m->mdt_opts.mo_evict_tgt_nids = 1;
5029 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
5031 lmi = server_get_mount(dev);
5033 CERROR("Cannot get mount info for %s!\n", dev);
5036 lsi = s2lsi(lmi->lmi_sb);
5037 /* CMD is supported only in IAM mode */
5039 node_id = simple_strtol(num, NULL, 10);
5040 obd->u.obt.obt_magic = OBT_MAGIC;
5041 if (lsi->lsi_lmd != NULL &&
5042 lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
5043 m->mdt_skip_lfsck = 1;
5046 /* DoM files get IO lock at open by default */
5047 m->mdt_opts.mo_dom_lock = ALWAYS_DOM_LOCK_ON_OPEN;
5048 /* DoM files are read at open and data is packed in the reply */
5049 m->mdt_opts.mo_dom_read_open = 1;
5051 m->mdt_squash.rsi_uid = 0;
5052 m->mdt_squash.rsi_gid = 0;
5053 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
5054 init_rwsem(&m->mdt_squash.rsi_sem);
5055 spin_lock_init(&m->mdt_lock);
5056 m->mdt_enable_remote_dir = 0;
5057 m->mdt_enable_remote_dir_gid = 0;
5059 atomic_set(&m->mdt_mds_mds_conns, 0);
5060 atomic_set(&m->mdt_async_commit_count, 0);
5062 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
5063 m->mdt_lu_dev.ld_obd = obd;
5064 /* Set this lu_device to obd for error handling purposes. */
5065 obd->obd_lu_dev = &m->mdt_lu_dev;
5067 /* init the stack */
5068 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
5070 CERROR("%s: Can't init device stack, rc %d\n",
5071 mdt_obd_name(m), rc);
5076 ss_site = mdt_seq_site(m);
5077 s->ld_seq_site = ss_site;
5080 /* set server index */
5081 ss_site->ss_node_id = node_id;
5083 /* failover is the default
5084 * FIXME: we do not failout mds0/mgs, which may cause some problems.
5085 * assumed whose ss_node_id == 0 XXX
5087 obd->obd_replayable = 1;
5088 /* No connection accepted until configurations will finish */
5089 obd->obd_no_conn = 1;
5091 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
5092 char *str = lustre_cfg_string(cfg, 4);
5093 if (strchr(str, 'n')) {
5094 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
5095 obd->obd_replayable = 0;
5099 rc = mdt_fld_init(env, mdt_obd_name(m), m);
5101 GOTO(err_fini_stack, rc);
5103 rc = mdt_seq_init(env, m);
5105 GOTO(err_fini_fld, rc);
5107 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
5108 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
5109 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
5110 LDLM_NAMESPACE_SERVER,
5111 LDLM_NAMESPACE_GREEDY,
5113 if (m->mdt_namespace == NULL)
5114 GOTO(err_fini_seq, rc = -ENOMEM);
5116 m->mdt_namespace->ns_lvbp = m;
5117 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
5119 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
5120 /* set obd_namespace for compatibility with old code */
5121 obd->obd_namespace = m->mdt_namespace;
5123 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
5124 OBD_FAIL_MDS_ALL_REQUEST_NET,
5125 OBD_FAIL_MDS_ALL_REPLY_NET);
5127 GOTO(err_free_ns, rc);
5129 /* Amount of available space excluded from granting and reserved
5130 * for metadata. It is in percentage and 50% is default value. */
5131 tgd->tgd_reserved_pcnt = 50;
5133 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
5134 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
5136 m->mdt_brw_size = ONE_MB_BRW_SIZE;
5138 rc = mdt_fs_setup(env, m, obd, lsi);
5142 fid.f_seq = FID_SEQ_LOCAL_NAME;
5145 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
5147 GOTO(err_fs_cleanup, rc);
5149 rc = mdt_hsm_cdt_init(m);
5151 CERROR("%s: error initializing coordinator, rc %d\n",
5152 mdt_obd_name(m), rc);
5153 GOTO(err_los_fini, rc);
5156 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
5158 next = m->mdt_child;
5159 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
5161 mntopts = dt_conf->ddp_mntopts;
5163 if (mntopts & MNTOPT_USERXATTR)
5164 m->mdt_opts.mo_user_xattr = 1;
5166 m->mdt_opts.mo_user_xattr = 0;
5168 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
5170 if (mntopts & MNTOPT_ACL)
5171 m->mdt_opts.mo_acl = 1;
5173 m->mdt_opts.mo_acl = 0;
5175 /* XXX: to support suppgid for ACL, we enable identity_upcall
5176 * by default, otherwise, maybe got unexpected -EACCESS. */
5177 if (m->mdt_opts.mo_acl)
5178 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
5180 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
5182 &mdt_identity_upcall_cache_ops);
5183 if (IS_ERR(m->mdt_identity_cache)) {
5184 rc = PTR_ERR(m->mdt_identity_cache);
5185 m->mdt_identity_cache = NULL;
5186 GOTO(err_free_hsm, rc);
5189 rc = mdt_procfs_init(m, dev);
5191 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
5192 GOTO(err_recovery, rc);
5195 rc = mdt_quota_init(env, m, cfg);
5197 GOTO(err_procfs, rc);
5199 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
5200 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
5201 "mdt_ldlm_client", m->mdt_ldlm_client);
5203 ping_evictor_start();
5205 /* recovery will be started upon mdt_prepare()
5206 * when the whole stack is complete and ready
5207 * to serve the requests */
5209 /* Reduce the initial timeout on an MDS because it doesn't need such
5210 * a long timeout as an OST does. Adaptive timeouts will adjust this
5211 * value appropriately. */
5212 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
5213 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
5219 target_recovery_fini(obd);
5220 upcall_cache_cleanup(m->mdt_identity_cache);
5221 m->mdt_identity_cache = NULL;
5223 mdt_hsm_cdt_fini(m);
5225 local_oid_storage_fini(env, m->mdt_los);
5228 mdt_fs_cleanup(env, m);
5230 tgt_fini(env, &m->mdt_lut);
5232 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
5233 obd->obd_namespace = m->mdt_namespace = NULL;
5235 mdt_seq_fini(env, m);
5237 mdt_fld_fini(env, m);
5239 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5242 server_put_mount(dev, true);
5246 /* For interoperability, the left element is old parameter, the right one
5247 * is the new version of the parameter, if some parameter is deprecated,
5248 * the new version should be set as NULL. */
5249 static struct cfg_interop_param mdt_interop_param[] = {
5250 { "mdt.group_upcall", NULL },
5251 { "mdt.quota_type", NULL },
5252 { "mdd.quota_type", NULL },
5253 { "mdt.som", NULL },
5254 { "mdt.rootsquash", "mdt.root_squash" },
5255 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
5259 /* used by MGS to process specific configurations */
5260 static int mdt_process_config(const struct lu_env *env,
5261 struct lu_device *d, struct lustre_cfg *cfg)
5263 struct mdt_device *m = mdt_dev(d);
5264 struct md_device *md_next = m->mdt_child;
5265 struct lu_device *next = md2lu_dev(md_next);
5269 switch (cfg->lcfg_command) {
5271 struct obd_device *obd = d->ld_obd;
5273 /* For interoperability */
5274 struct cfg_interop_param *ptr = NULL;
5275 struct lustre_cfg *old_cfg = NULL;
5278 param = lustre_cfg_string(cfg, 1);
5279 if (param == NULL) {
5280 CERROR("param is empty\n");
5285 ptr = class_find_old_param(param, mdt_interop_param);
5287 if (ptr->new_param == NULL) {
5289 CWARN("For interoperability, skip this %s."
5290 " It is obsolete.\n", ptr->old_param);
5294 CWARN("Found old param %s, changed it to %s.\n",
5295 ptr->old_param, ptr->new_param);
5298 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
5305 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
5307 if (rc > 0 || rc == -ENOSYS) {
5308 /* is it an HSM var ? */
5309 rc = class_process_proc_param(PARAM_HSM,
5310 hsm_cdt_get_proc_vars(),
5312 if (rc > 0 || rc == -ENOSYS)
5313 /* we don't understand; pass it on */
5314 rc = next->ld_ops->ldo_process_config(env, next,
5319 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
5320 cfg->lcfg_buflens));
5324 /* others are passed further */
5325 rc = next->ld_ops->ldo_process_config(env, next, cfg);
5331 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
5332 const struct lu_object_header *hdr,
5333 struct lu_device *d)
5335 struct mdt_object *mo;
5339 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
5341 struct lu_object *o;
5342 struct lu_object_header *h;
5345 h = &mo->mot_header;
5346 lu_object_header_init(h);
5347 lu_object_init(o, h, d);
5348 lu_object_add_top(h, o);
5349 o->lo_ops = &mdt_obj_ops;
5350 spin_lock_init(&mo->mot_write_lock);
5351 mutex_init(&mo->mot_som_mutex);
5352 mutex_init(&mo->mot_lov_mutex);
5353 init_rwsem(&mo->mot_dom_sem);
5354 init_rwsem(&mo->mot_open_sem);
5355 atomic_set(&mo->mot_open_count, 0);
5361 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
5362 const struct lu_object_conf *unused)
5364 struct mdt_device *d = mdt_dev(o->lo_dev);
5365 struct lu_device *under;
5366 struct lu_object *below;
5370 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
5371 PFID(lu_object_fid(o)));
5373 under = &d->mdt_child->md_lu_dev;
5374 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
5375 if (below != NULL) {
5376 lu_object_add(o, below);
5383 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
5385 struct mdt_object *mo = mdt_obj(o);
5386 struct lu_object_header *h;
5390 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
5391 PFID(lu_object_fid(o)));
5393 LASSERT(atomic_read(&mo->mot_open_count) == 0);
5394 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
5397 lu_object_header_fini(h);
5398 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
5403 static int mdt_object_print(const struct lu_env *env, void *cookie,
5404 lu_printer_t p, const struct lu_object *o)
5406 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
5408 return (*p)(env, cookie,
5409 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
5410 mdto, mdto->mot_lov_created ? "lov_created" : "",
5411 mdto->mot_cache_attr ? "cache_attr" : "",
5412 mdto->mot_write_count);
5415 static int mdt_prepare(const struct lu_env *env,
5416 struct lu_device *pdev,
5417 struct lu_device *cdev)
5419 struct mdt_device *mdt = mdt_dev(cdev);
5420 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
5421 struct obd_device *obd = cdev->ld_obd;
5428 rc = next->ld_ops->ldo_prepare(env, cdev, next);
5432 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
5436 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
5440 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
5441 /* The LFSCK instance is registered just now, so it must be there when
5442 * register the namespace to such instance. */
5443 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
5445 if (mdt->mdt_seq_site.ss_node_id == 0) {
5446 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
5447 &mdt->mdt_md_root_fid);
5452 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
5454 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
5455 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
5456 LASSERT(obd->obd_no_conn);
5457 spin_lock(&obd->obd_dev_lock);
5458 obd->obd_no_conn = 0;
5459 spin_unlock(&obd->obd_dev_lock);
5461 if (obd->obd_recovering == 0)
5462 mdt_postrecov(env, mdt);
5467 const struct lu_device_operations mdt_lu_ops = {
5468 .ldo_object_alloc = mdt_object_alloc,
5469 .ldo_process_config = mdt_process_config,
5470 .ldo_prepare = mdt_prepare,
5473 static const struct lu_object_operations mdt_obj_ops = {
5474 .loo_object_init = mdt_object_init,
5475 .loo_object_free = mdt_object_free,
5476 .loo_object_print = mdt_object_print
5479 static int mdt_obd_set_info_async(const struct lu_env *env,
5480 struct obd_export *exp,
5481 __u32 keylen, void *key,
5482 __u32 vallen, void *val,
5483 struct ptlrpc_request_set *set)
5489 if (KEY_IS(KEY_SPTLRPC_CONF)) {
5490 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
5498 * Match client and server connection feature flags.
5500 * Compute the compatibility flags for a connection request based on
5501 * features mutually supported by client and server.
5503 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
5504 * must not be updated here, otherwise a partially initialized value may
5505 * be exposed. After the connection request is successfully processed,
5506 * the top-level MDT connect request handler atomically updates the export
5507 * connect flags from the obd_connect_data::ocd_connect_flags field of the
5508 * reply. \see mdt_connect().
5510 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
5511 * full struct obd_connect_data. So care must be taken when accessing fields
5512 * that are not present in struct obd_connect_data_v1. See LU-16.
5514 * \param exp the obd_export associated with this client/target pair
5515 * \param mdt the target device for the connection
5516 * \param data stores data for this connect request
5519 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
5520 * \retval -EBADE client and server feature requirements are incompatible
5522 static int mdt_connect_internal(const struct lu_env *env,
5523 struct obd_export *exp,
5524 struct mdt_device *mdt,
5525 struct obd_connect_data *data, bool reconnect)
5527 const char *obd_name = mdt_obd_name(mdt);
5528 LASSERT(data != NULL);
5530 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
5532 if (mdt->mdt_bottom->dd_rdonly &&
5533 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5534 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
5537 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
5538 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
5540 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
5542 if (!mdt->mdt_opts.mo_acl)
5543 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
5545 if (!mdt->mdt_opts.mo_user_xattr)
5546 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
5548 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
5549 data->ocd_brw_size = min(data->ocd_brw_size,
5551 if (data->ocd_brw_size == 0) {
5552 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
5553 "ocd_version: %x ocd_grant: %d ocd_index: %u "
5554 "ocd_brw_size unexpectedly zero, network data "
5555 "corruption? Refusing to connect this client\n",
5556 obd_name, exp->exp_client_uuid.uuid,
5557 exp, data->ocd_connect_flags, data->ocd_version,
5558 data->ocd_grant, data->ocd_index);
5563 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
5564 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
5566 /* client is reporting its page size, for future use */
5567 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
5568 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
5569 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
5570 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
5571 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
5572 /* ocd_grant_tax_kb is in 1K byte blocks */
5573 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
5574 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
5577 /* Save connect_data we have so far because tgt_grant_connect()
5578 * uses it to calculate grant, and we want to save the client
5579 * version before it is overwritten by LUSTRE_VERSION_CODE. */
5580 exp->exp_connect_data = *data;
5581 if (OCD_HAS_FLAG(data, GRANT))
5582 tgt_grant_connect(env, exp, data, !reconnect);
5584 if (OCD_HAS_FLAG(data, MAXBYTES))
5585 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
5587 /* NB: Disregard the rule against updating
5588 * exp_connect_data.ocd_connect_flags in this case, since
5589 * tgt_client_new() needs to know if this is a lightweight
5590 * connection, and it is safe to expose this flag before
5591 * connection processing completes. */
5592 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
5593 spin_lock(&exp->exp_lock);
5594 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
5595 spin_unlock(&exp->exp_lock);
5598 data->ocd_version = LUSTRE_VERSION_CODE;
5600 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
5601 CWARN("%s: MDS requires FID support, but client not\n",
5606 if (OCD_HAS_FLAG(data, PINGLESS)) {
5607 if (ptlrpc_pinger_suppress_pings()) {
5608 spin_lock(&exp->exp_obd->obd_dev_lock);
5609 list_del_init(&exp->exp_obd_chain_timed);
5610 spin_unlock(&exp->exp_obd->obd_dev_lock);
5612 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
5616 data->ocd_max_easize = mdt->mdt_max_ea_size;
5618 /* NB: Disregard the rule against updating
5619 * exp_connect_data.ocd_connect_flags in this case, since
5620 * tgt_client_new() needs to know if this is client supports
5621 * multiple modify RPCs, and it is safe to expose this flag before
5622 * connection processing completes. */
5623 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
5624 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
5625 spin_lock(&exp->exp_lock);
5626 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
5627 spin_unlock(&exp->exp_lock);
5630 if (OCD_HAS_FLAG(data, CKSUM)) {
5631 __u32 cksum_types = data->ocd_cksum_types;
5633 /* The client set in ocd_cksum_types the checksum types it
5634 * supports. We have to mask off the algorithms that we don't
5636 data->ocd_cksum_types &=
5637 obd_cksum_types_supported_server(obd_name);
5639 if (unlikely(data->ocd_cksum_types == 0)) {
5640 CERROR("%s: Connect with checksum support but no "
5641 "ocd_cksum_types is set\n",
5642 exp->exp_obd->obd_name);
5646 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
5647 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
5648 cksum_types, data->ocd_cksum_types);
5650 /* This client does not support OBD_CONNECT_CKSUM
5651 * fall back to CRC32 */
5652 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
5653 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
5654 exp->exp_obd->obd_name, obd_export_nid2str(exp));
5660 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
5661 struct mdt_thread_info *info,
5662 struct mdt_file_data *mfd)
5664 struct lu_context ses;
5668 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
5673 lu_context_enter(&ses);
5675 mdt_ucred(info)->uc_valid = UCRED_OLD;
5676 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
5678 lu_context_exit(&ses);
5679 lu_context_fini(&ses);
5685 static int mdt_export_cleanup(struct obd_export *exp)
5687 struct list_head closing_list;
5688 struct mdt_export_data *med = &exp->exp_mdt_data;
5689 struct obd_device *obd = exp->exp_obd;
5690 struct mdt_device *mdt;
5691 struct mdt_thread_info *info;
5693 struct mdt_file_data *mfd, *n;
5697 INIT_LIST_HEAD(&closing_list);
5698 spin_lock(&med->med_open_lock);
5699 while (!list_empty(&med->med_open_head)) {
5700 struct list_head *tmp = med->med_open_head.next;
5701 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
5703 /* Remove mfd handle so it can't be found again.
5704 * We are consuming the mfd_list reference here. */
5705 class_handle_unhash(&mfd->mfd_open_handle);
5706 list_move_tail(&mfd->mfd_list, &closing_list);
5708 spin_unlock(&med->med_open_lock);
5709 mdt = mdt_dev(obd->obd_lu_dev);
5710 LASSERT(mdt != NULL);
5712 rc = lu_env_init(&env, LCT_MD_THREAD);
5716 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5717 LASSERT(info != NULL);
5718 memset(info, 0, sizeof *info);
5719 info->mti_env = &env;
5720 info->mti_mdt = mdt;
5721 info->mti_exp = exp;
5723 if (!list_empty(&closing_list)) {
5724 struct md_attr *ma = &info->mti_attr;
5726 /* Close any open files (which may also cause orphan
5728 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5729 list_del_init(&mfd->mfd_list);
5730 ma->ma_need = ma->ma_valid = 0;
5732 /* This file is being closed due to an eviction, it
5733 * could have been modified and now dirty regarding to
5734 * HSM archive, check this!
5735 * The logic here is to mark a file dirty if there's a
5736 * chance it was dirtied before the client was evicted,
5737 * so that we don't have to wait for a release attempt
5738 * before finding out the file was actually dirty and
5739 * fail the release. Aggressively marking it dirty here
5740 * will cause the policy engine to attempt to
5741 * re-archive it; when rearchiving, we can compare the
5742 * current version to the HSM data_version and make the
5743 * archive request into a noop if it's not actually
5746 if (mfd->mfd_mode & MDS_FMODE_WRITE)
5747 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
5749 /* Don't unlink orphan on failover umount, LU-184 */
5750 if (exp->exp_flags & OBD_OPT_FAILOVER) {
5751 ma->ma_valid = MA_FLAGS;
5752 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
5754 mdt_mfd_close(info, mfd);
5757 info->mti_mdt = NULL;
5758 /* cleanup client slot early */
5759 /* Do not erase record for recoverable client. */
5760 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5761 tgt_client_del(&env, exp);
5767 static inline void mdt_enable_slc(struct mdt_device *mdt)
5769 if (mdt->mdt_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
5770 mdt->mdt_lut.lut_sync_lock_cancel = BLOCKING_SYNC_ON_CANCEL;
5773 static inline void mdt_disable_slc(struct mdt_device *mdt)
5775 if (mdt->mdt_lut.lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL)
5776 mdt->mdt_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
5779 static int mdt_obd_disconnect(struct obd_export *exp)
5786 class_export_get(exp);
5788 if (!(exp->exp_flags & OBD_OPT_FORCE))
5789 tgt_grant_sanity_check(exp->exp_obd, __func__);
5791 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
5792 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
5793 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
5795 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
5796 mdt_disable_slc(mdt);
5799 rc = server_disconnect_export(exp);
5801 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
5803 tgt_grant_discard(exp);
5805 rc = mdt_export_cleanup(exp);
5806 nodemap_del_member(exp);
5807 class_export_put(exp);
5811 /* mds_connect copy */
5812 static int mdt_obd_connect(const struct lu_env *env,
5813 struct obd_export **exp, struct obd_device *obd,
5814 struct obd_uuid *cluuid,
5815 struct obd_connect_data *data,
5818 struct obd_export *lexp;
5819 struct lustre_handle conn = { 0 };
5820 struct mdt_device *mdt;
5822 lnet_nid_t *client_nid = localdata;
5825 LASSERT(env != NULL);
5826 LASSERT(data != NULL);
5828 if (!exp || !obd || !cluuid)
5831 mdt = mdt_dev(obd->obd_lu_dev);
5833 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5834 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
5835 atomic_inc(&mdt->mdt_mds_mds_conns);
5836 mdt_enable_slc(mdt);
5840 * first, check whether the stack is ready to handle requests
5841 * XXX: probably not very appropriate method is used now
5842 * at some point we should find a better one
5844 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
5845 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5846 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5847 rc = obd_get_info(env, mdt->mdt_child_exp,
5848 sizeof(KEY_OSP_CONNECTED),
5849 KEY_OSP_CONNECTED, NULL, NULL);
5852 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5855 rc = class_connect(&conn, obd, cluuid);
5859 lexp = class_conn2export(&conn);
5860 LASSERT(lexp != NULL);
5862 rc = nodemap_add_member(*client_nid, lexp);
5863 if (rc != 0 && rc != -EEXIST)
5866 rc = mdt_connect_internal(env, lexp, mdt, data, false);
5868 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5871 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5872 rc = tgt_client_new(env, lexp);
5874 mdt_export_stats_init(obd, lexp, localdata);
5878 class_disconnect(lexp);
5879 nodemap_del_member(lexp);
5883 /* Because we do not want this export to be evicted by pinger,
5884 * let's not add this export to the timed chain list. */
5885 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
5886 spin_lock(&lexp->exp_obd->obd_dev_lock);
5887 list_del_init(&lexp->exp_obd_chain_timed);
5888 spin_unlock(&lexp->exp_obd->obd_dev_lock);
5895 static int mdt_obd_reconnect(const struct lu_env *env,
5896 struct obd_export *exp, struct obd_device *obd,
5897 struct obd_uuid *cluuid,
5898 struct obd_connect_data *data,
5901 lnet_nid_t *client_nid = localdata;
5905 if (exp == NULL || obd == NULL || cluuid == NULL)
5908 rc = nodemap_add_member(*client_nid, exp);
5909 if (rc != 0 && rc != -EEXIST)
5912 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
5915 mdt_export_stats_init(obd, exp, localdata);
5917 nodemap_del_member(exp);
5922 /* FIXME: Can we avoid using these two interfaces? */
5923 static int mdt_init_export(struct obd_export *exp)
5925 struct mdt_export_data *med = &exp->exp_mdt_data;
5929 INIT_LIST_HEAD(&med->med_open_head);
5930 spin_lock_init(&med->med_open_lock);
5931 spin_lock(&exp->exp_lock);
5932 exp->exp_connecting = 1;
5933 spin_unlock(&exp->exp_lock);
5935 /* self-export doesn't need client data and ldlm initialization */
5936 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5937 &exp->exp_client_uuid)))
5940 rc = tgt_client_alloc(exp);
5944 rc = ldlm_init_export(exp);
5951 tgt_client_free(exp);
5953 CERROR("%s: Failed to initialize export: rc = %d\n",
5954 exp->exp_obd->obd_name, rc);
5958 static int mdt_destroy_export(struct obd_export *exp)
5962 target_destroy_export(exp);
5963 /* destroy can be called from failed obd_setup, so
5964 * checking uuid is safer than obd_self_export */
5965 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5966 &exp->exp_client_uuid)))
5969 ldlm_destroy_export(exp);
5970 tgt_client_free(exp);
5972 LASSERT(list_empty(&exp->exp_outstanding_replies));
5973 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5976 * discard grants once we're sure no more
5977 * interaction with the client is possible
5979 tgt_grant_discard(exp);
5980 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
5981 exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--;
5983 if (!(exp->exp_flags & OBD_OPT_FORCE))
5984 tgt_grant_sanity_check(exp->exp_obd, __func__);
5989 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5990 struct linkea_data *ldata)
5994 LASSERT(ldata->ld_buf->lb_buf != NULL);
5996 if (!mdt_object_exists(mdt_obj))
5999 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6000 ldata->ld_buf, XATTR_NAME_LINK);
6001 if (rc == -ERANGE) {
6002 /* Buf was too small, figure out what we need. */
6003 lu_buf_free(ldata->ld_buf);
6004 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6005 ldata->ld_buf, XATTR_NAME_LINK);
6008 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
6009 if (ldata->ld_buf->lb_buf == NULL)
6011 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6012 ldata->ld_buf, XATTR_NAME_LINK);
6017 return linkea_init_with_rec(ldata);
6021 * Given an MDT object, try to look up the full path to the object.
6022 * Part of the MDT layer implementation of lfs fid2path.
6024 * \param[in] info Per-thread common data shared by MDT level handlers.
6025 * \param[in] obj Object to do path lookup of
6026 * \param[in,out] fp User-provided struct to store path information
6027 * \param[in] root_fid Root FID of current path should reach
6029 * \retval 0 Lookup successful, path information stored in fp
6030 * \retval -EAGAIN Lookup failed, usually because object is being moved
6031 * \retval negative errno if there was a problem
6033 static int mdt_path_current(struct mdt_thread_info *info,
6034 struct mdt_object *obj,
6035 struct getinfo_fid2path *fp,
6036 struct lu_fid *root_fid)
6038 struct mdt_device *mdt = info->mti_mdt;
6039 struct mdt_object *mdt_obj;
6040 struct link_ea_header *leh;
6041 struct link_ea_entry *lee;
6042 struct lu_name *tmpname = &info->mti_name;
6043 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
6044 struct lu_buf *buf = &info->mti_big_buf;
6047 struct linkea_data ldata = { NULL };
6052 /* temp buffer for path element, the buffer will be finally freed
6053 * in mdt_thread_info_fini */
6054 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
6055 if (buf->lb_buf == NULL)
6059 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
6062 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
6064 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
6065 struct lu_buf lmv_buf;
6067 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
6068 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
6069 GOTO(out, rc = -ENOENT);
6071 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
6073 mdt_object_get(info->mti_env, mdt_obj);
6075 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
6076 if (IS_ERR(mdt_obj))
6077 GOTO(out, rc = PTR_ERR(mdt_obj));
6080 if (!mdt_object_exists(mdt_obj)) {
6081 mdt_object_put(info->mti_env, mdt_obj);
6082 GOTO(out, rc = -ENOENT);
6085 if (mdt_object_remote(mdt_obj)) {
6086 mdt_object_put(info->mti_env, mdt_obj);
6087 GOTO(remote_out, rc = -EREMOTE);
6090 rc = mdt_links_read(info, mdt_obj, &ldata);
6092 mdt_object_put(info->mti_env, mdt_obj);
6097 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
6098 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
6099 /* If set, use link #linkno for path lookup, otherwise use
6100 link #0. Only do this for the final path element. */
6101 if (first && fp->gf_linkno < leh->leh_reccount) {
6103 for (count = 0; count < fp->gf_linkno; count++) {
6104 lee = (struct link_ea_entry *)
6105 ((char *)lee + reclen);
6106 linkea_entry_unpack(lee, &reclen, tmpname,
6109 if (fp->gf_linkno < leh->leh_reccount - 1)
6110 /* indicate to user there are more links */
6114 lmv_buf.lb_buf = info->mti_xattr_buf;
6115 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
6116 /* Check if it is slave stripes */
6117 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6118 &lmv_buf, XATTR_NAME_LMV);
6119 mdt_object_put(info->mti_env, mdt_obj);
6121 union lmv_mds_md *lmm = lmv_buf.lb_buf;
6123 /* For slave stripes, get its master */
6124 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
6125 fp->gf_fid = *tmpfid;
6128 } else if (rc < 0 && rc != -ENODATA) {
6134 /* Pack the name in the end of the buffer */
6135 ptr -= tmpname->ln_namelen;
6136 if (ptr - 1 <= fp->gf_u.gf_path)
6137 GOTO(out, rc = -EOVERFLOW);
6138 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
6141 /* keep the last resolved fid to the client, so the
6142 * client will build the left path on another MDT for
6144 fp->gf_fid = *tmpfid;
6150 ptr++; /* skip leading / */
6151 memmove(fp->gf_u.gf_path, ptr,
6152 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
6159 * Given an MDT object, use mdt_path_current to get the path.
6160 * Essentially a wrapper to retry mdt_path_current a set number of times
6161 * if -EAGAIN is returned (usually because an object is being moved).
6163 * Part of the MDT layer implementation of lfs fid2path.
6165 * \param[in] info Per-thread common data shared by mdt level handlers.
6166 * \param[in] obj Object to do path lookup of
6167 * \param[in,out] fp User-provided struct for arguments and to store path
6170 * \retval 0 Lookup successful, path information stored in fp
6171 * \retval negative errno if there was a problem
6173 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
6174 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
6176 struct mdt_device *mdt = info->mti_mdt;
6181 if (fp->gf_pathlen < 3)
6184 if (root_fid == NULL)
6185 root_fid = &mdt->mdt_md_root_fid;
6187 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
6188 fp->gf_u.gf_path[0] = '\0';
6192 /* Retry multiple times in case file is being moved */
6193 while (tries-- && rc == -EAGAIN)
6194 rc = mdt_path_current(info, obj, fp, root_fid);
6200 * Get the full path of the provided FID, as of changelog record recno.
6202 * This checks sanity and looks up object for user provided FID
6203 * before calling the actual path lookup code.
6205 * Part of the MDT layer implementation of lfs fid2path.
6207 * \param[in] info Per-thread common data shared by mdt level handlers.
6208 * \param[in,out] fp User-provided struct for arguments and to store path
6211 * \retval 0 Lookup successful, path information and recno stored in fp
6212 * \retval -ENOENT, object does not exist
6213 * \retval negative errno if there was a problem
6215 static int mdt_fid2path(struct mdt_thread_info *info,
6216 struct lu_fid *root_fid,
6217 struct getinfo_fid2path *fp)
6219 struct mdt_device *mdt = info->mti_mdt;
6220 struct mdt_object *obj;
6224 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
6225 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
6227 if (!fid_is_sane(&fp->gf_fid))
6230 if (!fid_is_namespace_visible(&fp->gf_fid)) {
6231 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
6232 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
6233 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
6237 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
6240 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
6241 PFID(&fp->gf_fid), rc);
6245 if (mdt_object_remote(obj))
6247 else if (!mdt_object_exists(obj))
6253 mdt_object_put(info->mti_env, obj);
6254 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
6255 PFID(&fp->gf_fid), rc);
6259 rc = mdt_path(info, obj, fp, root_fid);
6261 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
6262 PFID(&fp->gf_fid), fp->gf_u.gf_path,
6263 fp->gf_recno, fp->gf_linkno);
6265 mdt_object_put(info->mti_env, obj);
6270 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
6271 void *val, int vallen)
6273 struct getinfo_fid2path *fpout, *fpin;
6274 struct lu_fid *root_fid = NULL;
6277 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
6280 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6281 lustre_swab_fid2path(fpin);
6283 memcpy(fpout, fpin, sizeof(*fpin));
6284 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
6287 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
6288 sizeof(struct lu_fid)) {
6289 /* client sent its root FID, which is normally fileset FID */
6290 root_fid = fpin->gf_u.gf_root_fid;
6291 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6292 lustre_swab_lu_fid(root_fid);
6294 if (root_fid != NULL && !fid_is_sane(root_fid))
6298 rc = mdt_fid2path(info, root_fid, fpout);
6302 int mdt_get_info(struct tgt_session_info *tsi)
6312 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
6314 CDEBUG(D_IOCTL, "No GETINFO key\n");
6315 RETURN(err_serious(-EFAULT));
6317 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
6320 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
6321 if (vallen == NULL) {
6322 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
6323 tgt_name(tsi->tsi_tgt));
6324 RETURN(err_serious(-EFAULT));
6327 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
6329 rc = req_capsule_server_pack(tsi->tsi_pill);
6331 RETURN(err_serious(rc));
6333 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
6334 if (valout == NULL) {
6335 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
6336 tgt_name(tsi->tsi_tgt));
6337 RETURN(err_serious(-EFAULT));
6340 if (KEY_IS(KEY_FID2PATH)) {
6341 struct mdt_thread_info *info = tsi2mdt_info(tsi);
6343 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
6344 mdt_thread_info_fini(info);
6351 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
6353 struct obd_ioctl_data *data = karg;
6356 struct mdt_object *obj;
6357 struct mdt_lock_handle *lh;
6361 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
6362 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
6365 fid = (struct lu_fid *)data->ioc_inlbuf1;
6367 if (!fid_is_sane(fid))
6370 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
6372 lh = &mti->mti_lh[MDT_LH_PARENT];
6373 mdt_lock_reg_init(lh, LCK_CR);
6375 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
6377 RETURN(PTR_ERR(obj));
6379 if (mdt_object_remote(obj)) {
6382 * before calling version get the correct MDS should be
6383 * fid, this is error to find remote object here
6385 CERROR("nonlocal object "DFID"\n", PFID(fid));
6386 } else if (!mdt_object_exists(obj)) {
6387 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
6390 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
6391 *(__u64 *)data->ioc_inlbuf2 = version;
6394 mdt_object_unlock_put(mti, obj, lh, 1);
6398 /* ioctls on obd dev */
6399 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
6400 void *karg, void __user *uarg)
6403 struct obd_device *obd = exp->exp_obd;
6404 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
6405 struct dt_device *dt = mdt->mdt_bottom;
6409 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
6410 rc = lu_env_init(&env, LCT_MD_THREAD);
6416 rc = mdt_device_sync(&env, mdt);
6418 case OBD_IOC_SET_READONLY:
6419 rc = dt_sync(&env, dt);
6421 rc = dt_ro(&env, dt);
6423 case OBD_IOC_ABORT_RECOVERY:
6424 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
6425 obd->obd_abort_recovery = 1;
6426 target_stop_recovery_thread(obd);
6429 case OBD_IOC_CHANGELOG_REG:
6430 case OBD_IOC_CHANGELOG_DEREG:
6431 case OBD_IOC_CHANGELOG_CLEAR:
6432 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env,
6436 case OBD_IOC_START_LFSCK: {
6437 struct md_device *next = mdt->mdt_child;
6438 struct obd_ioctl_data *data = karg;
6439 struct lfsck_start_param lsp;
6441 if (unlikely(data == NULL)) {
6446 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
6447 lsp.lsp_index_valid = 0;
6448 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
6451 case OBD_IOC_STOP_LFSCK: {
6452 struct md_device *next = mdt->mdt_child;
6453 struct obd_ioctl_data *data = karg;
6454 struct lfsck_stop stop;
6456 stop.ls_status = LS_STOPPED;
6457 /* Old lfsck utils may pass NULL @stop. */
6458 if (data->ioc_inlbuf1 == NULL)
6462 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
6464 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
6467 case OBD_IOC_QUERY_LFSCK: {
6468 struct md_device *next = mdt->mdt_child;
6469 struct obd_ioctl_data *data = karg;
6471 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
6475 case OBD_IOC_GET_OBJ_VERSION: {
6476 struct mdt_thread_info *mti;
6477 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6478 memset(mti, 0, sizeof *mti);
6479 mti->mti_env = &env;
6483 rc = mdt_ioc_version_get(mti, karg);
6486 case OBD_IOC_CATLOGLIST: {
6487 struct mdt_thread_info *mti;
6489 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6490 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
6491 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
6492 &mti->mti_tmp_fid1);
6497 CERROR("%s: Not supported cmd = %d, rc = %d\n",
6498 mdt_obd_name(mdt), cmd, rc);
6505 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
6507 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
6511 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
6512 struct lfsck_start_param lsp;
6514 lsp.lsp_start = NULL;
6515 lsp.lsp_index_valid = 0;
6516 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
6517 OBD_IOC_START_LFSCK,
6519 if (rc != 0 && rc != -EALREADY)
6520 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
6521 mdt_obd_name(mdt), rc);
6524 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
6528 static int mdt_obd_postrecov(struct obd_device *obd)
6533 rc = lu_env_init(&env, LCT_MD_THREAD);
6536 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
6541 static struct obd_ops mdt_obd_device_ops = {
6542 .o_owner = THIS_MODULE,
6543 .o_set_info_async = mdt_obd_set_info_async,
6544 .o_connect = mdt_obd_connect,
6545 .o_reconnect = mdt_obd_reconnect,
6546 .o_disconnect = mdt_obd_disconnect,
6547 .o_init_export = mdt_init_export,
6548 .o_destroy_export = mdt_destroy_export,
6549 .o_iocontrol = mdt_iocontrol,
6550 .o_postrecov = mdt_obd_postrecov,
6551 /* Data-on-MDT IO methods */
6552 .o_preprw = mdt_obd_preprw,
6553 .o_commitrw = mdt_obd_commitrw,
6556 static struct lu_device* mdt_device_fini(const struct lu_env *env,
6557 struct lu_device *d)
6559 struct mdt_device *m = mdt_dev(d);
6566 static struct lu_device *mdt_device_free(const struct lu_env *env,
6567 struct lu_device *d)
6569 struct mdt_device *m = mdt_dev(d);
6572 lu_device_fini(&m->mdt_lu_dev);
6578 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
6579 struct lu_device_type *t,
6580 struct lustre_cfg *cfg)
6582 struct lu_device *l;
6583 struct mdt_device *m;
6590 rc = mdt_init0(env, m, t, cfg);
6592 mdt_device_free(env, l);
6597 l = ERR_PTR(-ENOMEM);
6601 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
6602 LU_KEY_INIT(mdt, struct mdt_thread_info);
6604 static void mdt_key_fini(const struct lu_context *ctx,
6605 struct lu_context_key *key, void* data)
6607 struct mdt_thread_info *info = data;
6609 if (info->mti_big_lmm) {
6610 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
6611 info->mti_big_lmm = NULL;
6612 info->mti_big_lmmsize = 0;
6615 if (info->mti_big_acl) {
6616 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
6617 info->mti_big_acl = NULL;
6618 info->mti_big_aclsize = 0;
6624 /* context key: mdt_thread_key */
6625 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
6627 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
6629 return lu_ucred(info->mti_env);
6632 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
6634 return lu_ucred_check(info->mti_env);
6638 * Enable/disable COS (Commit On Sharing).
6640 * Set/Clear the COS flag in mdt options.
6642 * \param mdt mdt device
6643 * \param val 0 disables COS, other values enable COS
6645 void mdt_enable_cos(struct mdt_device *mdt, bool val)
6650 mdt->mdt_opts.mo_cos = val;
6651 rc = lu_env_init(&env, LCT_LOCAL);
6652 if (unlikely(rc != 0)) {
6653 CWARN("%s: lu_env initialization failed, cannot "
6654 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
6657 mdt_device_sync(&env, mdt);
6662 * Check COS (Commit On Sharing) status.
6664 * Return COS flag status.
6666 * \param mdt mdt device
6668 int mdt_cos_is_enabled(struct mdt_device *mdt)
6670 return mdt->mdt_opts.mo_cos != 0;
6673 static struct lu_device_type_operations mdt_device_type_ops = {
6674 .ldto_device_alloc = mdt_device_alloc,
6675 .ldto_device_free = mdt_device_free,
6676 .ldto_device_fini = mdt_device_fini
6679 static struct lu_device_type mdt_device_type = {
6680 .ldt_tags = LU_DEVICE_MD,
6681 .ldt_name = LUSTRE_MDT_NAME,
6682 .ldt_ops = &mdt_device_type_ops,
6683 .ldt_ctx_tags = LCT_MD_THREAD
6686 static int __init mdt_init(void)
6690 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
6691 FID_NOBRACE_LEN + 1);
6692 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
6694 rc = lu_kmem_init(mdt_caches);
6698 rc = mds_mod_init();
6702 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
6703 LUSTRE_MDT_NAME, &mdt_device_type);
6708 lu_kmem_fini(mdt_caches);
6715 static void __exit mdt_exit(void)
6717 class_unregister_type(LUSTRE_MDT_NAME);
6719 lu_kmem_fini(mdt_caches);
6722 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
6723 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
6724 MODULE_VERSION(LUSTRE_VERSION_STRING);
6725 MODULE_LICENSE("GPL");
6727 module_init(mdt_init);
6728 module_exit(mdt_exit);